diff --git a/llama/build-info.cpp b/llama/build-info.cpp index d4ddbffc..52e983a8 100644 --- a/llama/build-info.cpp +++ b/llama/build-info.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/clip.cpp b/llama/clip.cpp index d8a85004..dcc65f02 100644 --- a/llama/clip.cpp +++ b/llama/clip.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -46,6 +46,10 @@ #include "ggml-cann.h" #endif +#ifdef GGML_USE_VULKAN +#include "ggml-vulkan.h" +#endif + #define STB_IMAGE_IMPLEMENTATION #include "stb_image.h" @@ -100,26 +104,28 @@ static std::string format(const char * fmt, ...) { // key constants // -#define KEY_FTYPE "general.file_type" -#define KEY_NAME "general.name" -#define KEY_DESCRIPTION "general.description" -#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" -#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" -#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" -#define KEY_USE_GELU "clip.use_gelu" -#define KEY_N_EMBD "clip.%s.embedding_length" -#define KEY_N_FF "clip.%s.feed_forward_length" -#define KEY_N_BLOCK "clip.%s.block_count" -#define KEY_N_HEAD "clip.%s.attention.head_count" -#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" -#define KEY_PROJ_DIM "clip.%s.projection_dim" -#define KEY_TOKENS "tokenizer.ggml.tokens" -#define KEY_N_POSITIONS "clip.text.context_length" -#define KEY_IMAGE_SIZE "clip.vision.image_size" -#define KEY_PATCH_SIZE "clip.vision.patch_size" -#define KEY_IMAGE_MEAN "clip.vision.image_mean" -#define KEY_IMAGE_STD "clip.vision.image_std" -#define KEY_PROJ_TYPE "clip.projector_type" +#define KEY_FTYPE "general.file_type" +#define KEY_NAME "general.name" +#define KEY_DESCRIPTION "general.description" +#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" +#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" +#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" +#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector" +#define KEY_MINICPMV_VERSION "clip.minicpmv_version" +#define KEY_USE_GELU "clip.use_gelu" +#define KEY_N_EMBD "clip.%s.embedding_length" +#define KEY_N_FF "clip.%s.feed_forward_length" +#define KEY_N_BLOCK "clip.%s.block_count" +#define KEY_N_HEAD "clip.%s.attention.head_count" +#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" +#define KEY_PROJ_DIM "clip.%s.projection_dim" +#define KEY_TOKENS "tokenizer.ggml.tokens" +#define KEY_N_POSITIONS "clip.text.context_length" +#define KEY_IMAGE_SIZE "clip.vision.image_size" +#define KEY_PATCH_SIZE "clip.vision.patch_size" +#define KEY_IMAGE_MEAN "clip.vision.image_mean" +#define KEY_IMAGE_STD "clip.vision.image_std" +#define KEY_PROJ_TYPE "clip.projector_type" #define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type" #define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints" @@ -153,12 +159,20 @@ static std::string format(const char * fmt, ...) { #define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s" #define TN_IMAGE_NEWLINE "model.image_newline" +#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k" +#define TN_MINICPMV_QUERY "resampler.query" +#define TN_MINICPMV_PROJ "resampler.proj.weight" +#define TN_MINICPMV_KV_PROJ "resampler.kv.weight" +#define TN_MINICPMV_ATTN "resampler.attn.%s.%s" +#define TN_MINICPMV_LN "resampler.ln_%s.%s" + enum projector_type { PROJECTOR_TYPE_MLP, PROJECTOR_TYPE_MLP_NORM, PROJECTOR_TYPE_LDP, PROJECTOR_TYPE_LDPV2, + PROJECTOR_TYPE_RESAMPLER, PROJECTOR_TYPE_UNKNOWN, }; @@ -166,6 +180,7 @@ static std::map PROJECTOR_TYPE_NAMES = { { PROJECTOR_TYPE_MLP, "mlp" }, { PROJECTOR_TYPE_LDP, "ldp" }, { PROJECTOR_TYPE_LDPV2, "ldpv2"}, + { PROJECTOR_TYPE_RESAMPLER, "resampler"}, }; @@ -226,17 +241,20 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int } static void replace_all(std::string & s, const std::string & search, const std::string & replace) { - std::string result; - for (size_t pos = 0; ; pos += search.length()) { - auto new_pos = s.find(search, pos); - if (new_pos == std::string::npos) { - result += s.substr(pos, s.size() - pos); - break; - } - result += s.substr(pos, new_pos - pos) + replace; - pos = new_pos; + if (search.empty()) { + return; } - s = std::move(result); + std::string builder; + builder.reserve(s.length()); + size_t pos = 0; + size_t last_pos = 0; + while ((pos = s.find(search, last_pos)) != std::string::npos) { + builder.append(s, last_pos, pos - last_pos); + builder.append(replace); + last_pos = pos + search.length(); + } + builder.append(s, last_pos, std::string::npos); + s = std::move(builder); } static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { @@ -518,12 +536,34 @@ struct clip_vision_model { struct ggml_tensor * mm_model_mlp_2_b; struct ggml_tensor * mm_model_peg_0_w; struct ggml_tensor * mm_model_peg_0_b; + + // MINICPMV projection + struct ggml_tensor * mm_model_pos_embed_k; + struct ggml_tensor * mm_model_query; + struct ggml_tensor * mm_model_proj; + struct ggml_tensor * mm_model_kv_proj; + struct ggml_tensor * mm_model_attn_q_w; + struct ggml_tensor * mm_model_attn_q_b; + struct ggml_tensor * mm_model_attn_k_w; + struct ggml_tensor * mm_model_attn_k_b; + struct ggml_tensor * mm_model_attn_v_w; + struct ggml_tensor * mm_model_attn_v_b; + struct ggml_tensor * mm_model_attn_o_w; + struct ggml_tensor * mm_model_attn_o_b; + struct ggml_tensor * mm_model_ln_q_w; + struct ggml_tensor * mm_model_ln_q_b; + struct ggml_tensor * mm_model_ln_kv_w; + struct ggml_tensor * mm_model_ln_kv_b; + struct ggml_tensor * mm_model_ln_post_w; + struct ggml_tensor * mm_model_ln_post_b; }; struct clip_ctx { bool has_text_encoder = false; bool has_vision_encoder = false; bool has_llava_projector = false; + bool has_minicpmv_projector = false; + int minicpmv_version = 2; struct clip_vision_model vision_model; projector_type proj_type = PROJECTOR_TYPE_MLP; @@ -548,9 +588,11 @@ struct clip_ctx { ggml_backend_t backend = NULL; ggml_gallocr_t compute_alloc = NULL; + + struct clip_image_size * load_image_size; }; -static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) { +static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) { if (!ctx->has_vision_encoder) { LOG_TEE("This gguf file seems to have no vision encoder\n"); return nullptr; @@ -559,20 +601,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 const auto & model = ctx->vision_model; const auto & hparams = model.hparams; - const int image_size = hparams.image_size; + const int image_size = hparams.image_size; + int image_size_width = image_size; + int image_size_height = image_size; + if (ctx->has_minicpmv_projector) { + if (load_image_size == nullptr) { + load_image_size = clip_image_size_init(); + } + LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height); + image_size_width = load_image_size->width; + image_size_height = load_image_size->height; + if (is_inf) { + image_size_width = imgs->data->nx; + image_size_height = imgs->data->ny; + } + } const int patch_size = hparams.patch_size; - const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); - const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side); + const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0); const int hidden_size = hparams.hidden_size; const int n_head = hparams.n_head; const int d_head = hidden_size / n_head; - const int n_layer = hparams.n_layer; + int n_layer = hparams.n_layer; const float eps = hparams.eps; const int batch_size = imgs->size; - if (ctx->has_llava_projector) { + if (ctx->has_llava_projector || ctx->has_minicpmv_projector) { GGML_ASSERT(batch_size == 1); } @@ -585,7 +640,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 struct ggml_context * ctx0 = ggml_init(params); struct ggml_cgraph * gf = ggml_new_graph(ctx0); - struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size); + struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size); ggml_set_name(inp_raw, "inp_raw"); ggml_set_input(inp_raw); @@ -598,19 +653,21 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp)); inp = ggml_add(ctx0, inp, model.patch_bias); } - - // concat class_embeddings and patch_embeddings struct ggml_tensor * embeddings = inp; - if (ctx->has_class_embedding) { - embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size); - ggml_set_name(embeddings, "embeddings"); - ggml_set_input(embeddings); - embeddings = ggml_acc(ctx0, embeddings, model.class_embedding, - embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0); - embeddings = ggml_acc(ctx0, embeddings, inp, - embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]); - } + struct ggml_tensor * pos_embed = nullptr; + if (ctx->has_llava_projector) { + // concat class_embeddings and patch_embeddings + if (ctx->has_class_embedding) { + embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size); + ggml_set_name(embeddings, "embeddings"); + ggml_set_input(embeddings); + embeddings = ggml_acc(ctx0, embeddings, model.class_embedding, + embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0); + embeddings = ggml_acc(ctx0, embeddings, inp, + embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]); + } + } struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions); ggml_set_name(positions, "positions"); @@ -619,6 +676,19 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions)); + if (ctx->has_minicpmv_projector) { + int pos_w = image_size_width/patch_size; + int pos_h = image_size_height/patch_size; + if (ctx->minicpmv_version == 2) { + pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1); + } + else if (ctx->minicpmv_version == 3) { + pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1); + } + ggml_set_name(pos_embed, "pos_embed"); + ggml_set_input(pos_embed); + } + // pre-layernorm if (ctx->has_pre_norm) { embeddings = ggml_norm(ctx0, embeddings, eps); @@ -628,6 +698,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } // loop over layers + if (ctx->has_minicpmv_projector) { + n_layer += 1; + } for (int il = 0; il < n_layer - 1; il++) { struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states @@ -717,7 +790,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } // llava projector - { + if (ctx->has_llava_projector) { embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]); struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches); @@ -738,8 +811,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_gelu(ctx0, embeddings); embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); - - } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + } + else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false); @@ -898,6 +971,75 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 GGML_ABORT("fatal error"); } } + // minicpmv projector + else if (ctx->has_minicpmv_projector) + { + if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { + struct ggml_tensor * q = model.mm_model_query; + { // layernorm + q = ggml_norm(ctx0, q, eps); + q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b); + } + struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings); + { // layernorm + v = ggml_norm(ctx0, v, eps); + v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b); + } + struct ggml_tensor * k; + { // position + // q = ggml_add(ctx0, q, model.mm_model_pos_embed); + k = ggml_add(ctx0, v, pos_embed); + } + + { // attention + int hidden_size = 4096; + const int d_head = 128; + int n_head = hidden_size/d_head; + int num_query = 96; + if (ctx->minicpmv_version == 2) { + hidden_size = 4096; + n_head = hidden_size/d_head; + num_query = 96; + } + else if (ctx->minicpmv_version == 3) { + hidden_size = 3584; + n_head = hidden_size/d_head; + num_query = 64; + } + + struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b); + Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head)); + struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b); + struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b); + // permute + Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size); + Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); + Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size); + K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size); + K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); + K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size); + V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size); + V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); + V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size); + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + KQ = ggml_soft_max_inplace(ctx0, KQ); + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); + KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size); + KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size); + + embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b); + } + { // layernorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b); + } + embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings); + } + else { + GGML_ASSERT(false); + } + } // build the graph ggml_build_forward_expand(gf, embeddings); @@ -1002,7 +1144,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { } } - clip_ctx * new_clip = new clip_ctx; + clip_ctx * new_clip = new clip_ctx{}; // update projector type { @@ -1036,6 +1178,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { LOG_TEE("%s: CLIP using CANN backend\n", __func__); #endif +#ifdef GGML_USE_VULKAN + new_clip->backend = ggml_backend_vk_init(0); + LOG_TEE("%s: CLIP using Vulkan backend\n", __func__); +#endif if (!new_clip->backend) { new_clip->backend = ggml_backend_cpu_init(); @@ -1055,7 +1201,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx); } - GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search + idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ); + if (idx != -1) { + new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx); + } + + idx = gguf_find_key(ctx, KEY_MINICPMV_VERSION); + if (idx != -1) { + new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx); + } + + // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search + GGML_ASSERT(new_clip->has_vision_encoder); GGML_ASSERT(!new_clip->has_text_encoder); @@ -1066,6 +1223,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { LOG_TEE("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder); LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder); LOG_TEE("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector); + LOG_TEE("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector); LOG_TEE("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0); LOG_TEE("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0); } @@ -1307,6 +1465,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight")); vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias")); } + else if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) { + // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD); + vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K); + vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY); + vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ); + vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ); + vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight")); + vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight")); + vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight")); + vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias")); + vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias")); + vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias")); + vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight")); + vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias")); + vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight")); + vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias")); + vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight")); + vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias")); + vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight")); + vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias")); + } else { std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); @@ -1345,7 +1524,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend)); clip_image_f32_batch batch; batch.size = 1; - ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch); + ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false); ggml_gallocr_reserve(new_clip->compute_alloc, gf); size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0); LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0); @@ -1354,6 +1533,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { return new_clip; } +void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) { + ctx_clip->load_image_size = load_image_size; +} + +struct clip_image_size * clip_image_size_init() { + struct clip_image_size * load_image_size = new struct clip_image_size(); + load_image_size->width = 448; + load_image_size->height = 448; + return load_image_size; +} + struct clip_image_u8 * clip_image_u8_init() { return new clip_image_u8(); } @@ -1459,7 +1649,7 @@ static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32* } } -inline float clip(float x, float lower, float upper) { +inline int clip(int x, int lower, int upper) { return std::max(lower, std::min(x, upper)); } @@ -1624,9 +1814,182 @@ static std::vector divide_to_patches_u8(const clip_image_u8 & im return patches; } +static int ensure_divide(int length, int patch_size) { + return std::max(static_cast(std::round(static_cast(length) / patch_size) * patch_size), patch_size); +} + +static std::pair uhd_find_best_resize(std::pair original_size, int scale_resolution, int patch_size, bool allow_upscale = false) { + int width = original_size.first; + int height = original_size.second; + if ((width * height > scale_resolution * scale_resolution) || allow_upscale) { + float r = static_cast(width) / height; + height = static_cast(scale_resolution / std::sqrt(r)); + width = static_cast(height * r); + } + int best_width = ensure_divide(width, patch_size); + int best_height = ensure_divide(height, patch_size); + return std::make_pair(best_width, best_height); +} + +static std::pair uhd_get_refine_size(std::pair original_size, std::pair grid, int scale_resolution, int patch_size, bool allow_upscale = false) { + int width, height; + std::tie(width, height) = original_size; + int grid_x, grid_y; + std::tie(grid_x, grid_y) = grid; + + int refine_width = ensure_divide(width, grid_x); + int refine_height = ensure_divide(height, grid_y); + + int grid_width = refine_width / grid_x; + int grid_height = refine_height / grid_y; + + // auto best_grid_size = find_best_resize(std::make_tuple(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); (old line) + auto best_grid_size = uhd_find_best_resize(std::make_pair(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); // (new line) => fixes conversion for make_tuple to make_pair + int best_grid_width, best_grid_height; + std::tie(best_grid_width, best_grid_height) = best_grid_size; + + // std::pair refine_size = std::make_tuple(best_grid_width * grid_x, best_grid_height * grid_y); (old line) + std::pair refine_size = std::make_pair(best_grid_width * grid_x, best_grid_height * grid_y); // (new line) + return refine_size; +} + +static std::pair uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) { + std::vector candidate_split_grids_nums; + for (int i : {multiple - 1, multiple, multiple + 1}) { + if (i == 1 || i > max_slice_nums) { + continue; + } + candidate_split_grids_nums.push_back(i); + } + + std::vector> candidate_grids; + for (int split_grids_nums : candidate_split_grids_nums) { + int m = 1; + while (m <= split_grids_nums) { + if (split_grids_nums % m == 0) { + candidate_grids.emplace_back(m, split_grids_nums / m); + } + ++m; + } + } + + std::pair best_grid{1, 1}; + float min_error = std::numeric_limits::infinity(); + for (const auto& grid : candidate_grids) { + float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second)); + if (error < min_error) { + best_grid = grid; + min_error = error; + } + } + return best_grid; +} + +// inspired from LLaVA-UHD: +// -> https://arxiv.org/pdf/2403.11703 +// -> https://github.com/thunlp/LLaVA-UHD +// -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118 +static std::vector> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) { + const std::pair original_size={img->nx,img->ny}; + const int original_width = img->nx; + const int original_height = img->ny; + const float log_ratio = log(1.0*original_width/original_height); + const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution); + const int multiple = fmin(ceil(ratio), max_slice_nums); + + std::vector> images; + LOG_TEE("%s: multiple %d\n", __func__, multiple); + images.push_back(std::vector()); + + if (multiple <= 1) { + auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true); + clip_image_u8 * source_image = clip_image_u8_init(); + bicubic_resize(*img, *source_image, best_size.first, best_size.second); + // source_image = image.resize(best_size, Image.Resampling.BICUBIC) + images[images.size()-1].push_back(source_image); + } + else if (multiple > 1) { + auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size); + clip_image_u8 * source_image = clip_image_u8_init(); + bicubic_resize(*img, *source_image, best_size.first, best_size.second); + // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC) + LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second); + images[images.size()-1].push_back(source_image); + + std::pair best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio); + LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second); + + auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true); + clip_image_u8 * refine_image = clip_image_u8_init(); + bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second); + + LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second); + + // split_to_patches + int width = refine_image->nx; + int height = refine_image->ny; + int grid_x = int(width / best_grid.first); + int grid_y = int(height / best_grid.second); + for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){ + images.push_back(std::vector()); + for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){ + clip_image_u8 * patch = clip_image_u8_init(); + patch->nx = grid_x; + patch->ny = grid_y; + patch->buf.resize(3 * patch->nx * patch->ny); + for (int y = patches_i; y < patches_i + grid_y; ++y) { + for (int x = patches_j; x < patches_j + grid_x; ++x) { + const int i = 3 * (y * refine_image->nx + x); + const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j)); + patch->buf[j] = refine_image->buf[i]; + patch->buf[j+1] = refine_image->buf[i+1]; + patch->buf[j+2] = refine_image->buf[i+2]; + } + } + images[images.size()-1].push_back(patch); + } + } + } + return images; +} + +int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) { + const int max_slice_nums=9; + const int scale_resolution=448; + const int original_width = ctx_clip->load_image_size->width; + const int original_height = ctx_clip->load_image_size->height; + const float log_ratio = log(1.0*original_width/original_height); + const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution); + const int multiple = fmin(ceil(ratio), max_slice_nums); + std::pair best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio); + return best_grid.first; +} + // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector // res_imgs memory is being allocated here, previous allocations will be freed if found bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) { + + if(clip_is_minicpmv(ctx)){ + int max_slice_nums = 9; + std::vector> imgs = uhd_slice_image(img, max_slice_nums); + res_imgs->size = 0; + for (size_t i = 0; i < imgs.size(); ++i){ + res_imgs->size += imgs[i].size(); + } + res_imgs->data = new clip_image_f32[res_imgs->size]; + int idx = 0; + for (size_t i = 0; i < imgs.size(); ++i) { + for (size_t j = 0; j < imgs[i].size(); ++j) { + LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny); + clip_image_f32 * res = clip_image_f32_init(); + normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std); + res_imgs->data[idx++] = *res; + clip_image_f32_free(res); + } + } + return true; + } + bool pad_to_square = true; if (!ctx->has_vision_encoder) { LOG_TEE("This gguf file seems to have no vision encoder\n"); @@ -1842,11 +2205,104 @@ int clip_n_patches(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) { n_patches /= 4; + } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { + if (ctx->minicpmv_version == 2) { + n_patches = 96; + } + else if (ctx->minicpmv_version == 3) { + n_patches = 64; + } } return n_patches; } +static std::vector>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector> & pos) { + assert(embed_dim % 2 == 0); + int H = pos.size(); + int W = pos[0].size(); + + std::vector omega(embed_dim / 2); + for (int i = 0; i < embed_dim / 2; ++i) { + omega[i] = 1.0 / pow(10000.0, static_cast(i) / (embed_dim / 2)); + } + + std::vector>> emb(H, std::vector>(W, std::vector(embed_dim))); + for (int h = 0; h < H; ++h) { + for (int w = 0; w < W; ++w) { + for (int d = 0; d < embed_dim / 2; ++d) { + float out_value = pos[h][w] * omega[d]; + emb[h][w][d] = sin(out_value); + emb[h][w][d + embed_dim / 2] = cos(out_value); + } + } + } + + return emb; +} + +static std::vector>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector>> & grid) { + assert(embed_dim % 2 == 0); + std::vector>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2) + std::vector>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2) + + int H = emb_h.size(); + int W = emb_h[0].size(); + std::vector>> emb(H, std::vector>(W, std::vector(embed_dim))); + + for (int h = 0; h < H; ++h) { + for (int w = 0; w < W; ++w) { + for (int d = 0; d < embed_dim / 2; ++d) { + emb[h][w][d] = emb_h[h][w][d]; + emb[h][w][d + embed_dim / 2] = emb_w[h][w][d]; + } + } + } + return emb; +} + +static std::vector> get_2d_sincos_pos_embed(int embed_dim, const std::pair image_size) { + int grid_h_size = image_size.first; + int grid_w_size = image_size.second; + + std::vector grid_h(grid_h_size); + std::vector grid_w(grid_w_size); + + for (int i = 0; i < grid_h_size; ++i) { + grid_h[i] = static_cast(i); + } + for (int i = 0; i < grid_w_size; ++i) { + grid_w[i] = static_cast(i); + } + + std::vector> grid(grid_h_size, std::vector(grid_w_size)); + for (int h = 0; h < grid_h_size; ++h) { + for (int w = 0; w < grid_w_size; ++w) { + grid[h][w] = grid_w[w]; + } + } + std::vector>> grid_2d = {grid, grid}; + for (int h = 0; h < grid_h_size; ++h) { + for (int w = 0; w < grid_w_size; ++w) { + grid_2d[0][h][w] = grid_h[h]; + grid_2d[1][h][w] = grid_w[w]; + } + } + + std::vector>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d); + + int H = image_size.first; + int W = image_size.second; + std::vector> pos_embed_2d(H * W, std::vector(embed_dim)); + for (int h = 0; h < H; ++h) { + for (int w = 0; w < W; ++w) { + pos_embed_2d[w * H + h] = pos_embed_3d[h][w]; + } + } + + return pos_embed_2d; +} + bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) { if (!ctx->has_vision_encoder) { LOG_TEE("This gguf file seems to have no vision encoder\n"); @@ -1869,19 +2325,33 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima if (ctx->has_llava_projector) { GGML_ASSERT(batch_size == 1); // TODO: support multiple images } + if (ctx->has_minicpmv_projector) { + GGML_ASSERT(batch_size == 1); + } // build the inference graph - ggml_cgraph * gf = clip_image_build_graph(ctx, imgs); + ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true); ggml_gallocr_alloc_graph(ctx->compute_alloc, gf); // set inputs const auto & model = ctx->vision_model; const auto & hparams = model.hparams; - const int image_size = hparams.image_size; + const int image_size = hparams.image_size; + int image_size_width = image_size; + int image_size_height = image_size; + if (ctx->has_minicpmv_projector) { + image_size_width = imgs->data[0].nx; + image_size_height = imgs->data[0].ny; + } const int patch_size = hparams.patch_size; - const int num_patches = ((image_size / patch_size) * (image_size / patch_size)); + const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0); + if(ctx->load_image_size==nullptr){ + ctx->load_image_size= clip_image_size_init(); + } + const int pos_w = ctx->load_image_size->width/patch_size; + const int pos_h = ctx->load_image_size->height/patch_size; { struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw"); @@ -1890,7 +2360,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima for (size_t i = 0; i < imgs->size; i++) { const int nx = imgs->data[i].nx; const int ny = imgs->data[i].ny; - GGML_ASSERT(nx == image_size && ny == image_size); + if (!ctx->has_minicpmv_projector) { + GGML_ASSERT(nx == image_size && ny == image_size); + } const int n = nx * ny; @@ -1907,37 +2379,87 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw)); free(data); } + if (ctx->has_minicpmv_projector) { + { + // inspired from siglip: + // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit + // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316 + struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); + int* positions_data = (int*)malloc(ggml_nbytes(positions)); + int bucket_coords_h[70]; + int bucket_coords_w[70]; + for (int i = 0; i < pos_h; i++){ + bucket_coords_h[i] = std::floor(70.0*i/pos_h); + } + for (int i = 0; i < pos_w; i++){ + bucket_coords_w[i] = std::floor(70.0*i/pos_w); + } + for (int i = 0, id = 0; i < pos_h; i++){ + for (int j = 0; j < pos_w; j++){ + positions_data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j]; + } + } + ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); + free(positions_data); + } - { - if (ctx->has_class_embedding) { - struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings"); + { + // inspired from resampler of Qwen-VL: + // -> https://huggingface.co/Qwen/Qwen-VL/tree/main + // -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23 + struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed"); + int embed_dim = 4096; + if (ctx->minicpmv_version == 2) { + embed_dim = 4096; + } + else if (ctx->minicpmv_version == 3) { + embed_dim = 3584; + } + auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h)); - void* zero_mem = malloc(ggml_nbytes(embeddings)); - memset(zero_mem, 0, ggml_nbytes(embeddings)); - ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings)); - free(zero_mem); + float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed)); + for(int i=0;ihas_class_embedding) { + struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings"); - { - struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); - - int* positions_data = (int*)malloc(ggml_nbytes(positions)); - for (int i = 0; i < num_positions; i++) { - positions_data[i] = i; + void* zero_mem = malloc(ggml_nbytes(embeddings)); + memset(zero_mem, 0, ggml_nbytes(embeddings)); + ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings)); + free(zero_mem); + } } - ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); - free(positions_data); - } - { - struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); - int* patches_data = (int*)malloc(ggml_nbytes(patches)); - for (int i = 0; i < num_patches; i++) { - patches_data[i] = i + 1; + { + struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); + + int* positions_data = (int*)malloc(ggml_nbytes(positions)); + for (int i = 0; i < num_positions; i++) { + positions_data[i] = i; + } + ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); + free(positions_data); + } + + { + struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); + int* patches_data = (int*)malloc(ggml_nbytes(patches)); + for (int i = 0; i < num_patches; i++) { + patches_data[i] = i + 1; + } + ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); + free(patches_data); } - ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); - free(patches_data); } if (ggml_backend_is_cpu(ctx->backend)) { @@ -2107,7 +2629,22 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { return ctx->vision_model.mm_3_b->ne[0]; } + if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { + if (ctx->minicpmv_version == 2) { + return 4096; + } + else if (ctx->minicpmv_version == 3) { + return 3584; + } + } std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); } + +int clip_is_minicpmv(const struct clip_ctx * ctx) { + if (ctx->has_minicpmv_projector) { + return ctx->minicpmv_version; + } + return 0; +} diff --git a/llama/clip.h b/llama/clip.h index 84a51470..a064330c 100644 --- a/llama/clip.h +++ b/llama/clip.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -44,14 +44,17 @@ # define CLIP_API #endif -struct clip_ctx; - #ifdef __cplusplus extern "C" { #endif struct clip_ctx; +struct clip_image_size { + int width; + int height; +}; + struct clip_image_u8_batch { struct clip_image_u8 * data; size_t size; @@ -81,6 +84,10 @@ CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx); CLIP_API int clip_n_patches (const struct clip_ctx * ctx); CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx); +CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip); +CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size); + +CLIP_API struct clip_image_size * clip_image_size_init(); CLIP_API struct clip_image_u8 * clip_image_u8_init (); CLIP_API struct clip_image_f32 * clip_image_f32_init(); @@ -104,6 +111,8 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype); +CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx); + #ifdef __cplusplus } #endif diff --git a/llama/common.cpp b/llama/common.cpp index db58b563..6cabaeab 100644 --- a/llama/common.cpp +++ b/llama/common.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -103,6 +103,41 @@ using json = nlohmann::ordered_json; +// +// Environment variable utils +// + +template +static typename std::enable_if::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + target = value ? std::string(value) : target; +} + +template +static typename std::enable_if::value && std::is_integral::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + target = value ? std::stoi(value) : target; +} + +template +static typename std::enable_if::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + target = value ? std::stof(value) : target; +} + +template +static typename std::enable_if::value, void>::type +get_env(std::string name, T & target) { + char * value = std::getenv(name.c_str()); + if (value) { + std::string val(value); + target = val == "1" || val == "true"; + } +} + // // CPU utils // @@ -136,8 +171,34 @@ int32_t cpu_get_num_physical_cores() { if (result == 0) { return num_physical_cores; } -#elif defined(_WIN32) - //TODO: Implement +#elif defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later + // TODO: windows + arm64 + mingw64 + unsigned int n_threads_win = std::thread::hardware_concurrency(); + unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4; + + DWORD buffer_size = 0; + if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) { + if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) { + return default_threads; + } + } + + std::vector buffer(buffer_size); + if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast(buffer.data()), &buffer_size)) { + return default_threads; + } + + int32_t num_physical_cores = 0; + PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast(buffer.data()); + while (buffer_size > 0) { + if (info->Relationship == RelationProcessorCore) { + num_physical_cores += info->Processor.GroupCount; + } + buffer_size -= info->Size; + info = reinterpret_cast(reinterpret_cast(info) + info->Size); + } + + return num_physical_cores > 0 ? num_physical_cores : default_threads; #endif unsigned int n_threads = std::thread::hardware_concurrency(); return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4; @@ -216,16 +277,61 @@ int32_t cpu_get_num_math() { return cpu_get_num_physical_cores(); } +// Helper for setting process priority + +#if defined(_WIN32) + +bool set_process_priority(enum ggml_sched_priority prio) { + if (prio == GGML_SCHED_PRIO_NORMAL) { + return true; + } + + DWORD p = NORMAL_PRIORITY_CLASS; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = NORMAL_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_MEDIUM: p = ABOVE_NORMAL_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_HIGH: p = HIGH_PRIORITY_CLASS; break; + case GGML_SCHED_PRIO_REALTIME: p = REALTIME_PRIORITY_CLASS; break; + } + + if (!SetPriorityClass(GetCurrentProcess(), p)) { + fprintf(stderr, "warn: failed to set process priority class %d : (%d)\n", prio, (int) GetLastError()); + return false; + } + + return true; +} + +#else // MacOS and POSIX +#include +#include + +bool set_process_priority(enum ggml_sched_priority prio) { + if (prio == GGML_SCHED_PRIO_NORMAL) { + return true; + } + + int p = 0; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = 0; break; + case GGML_SCHED_PRIO_MEDIUM: p = -5; break; + case GGML_SCHED_PRIO_HIGH: p = -10; break; + case GGML_SCHED_PRIO_REALTIME: p = -20; break; + } + + if (!setpriority(PRIO_PROCESS, 0, p)) { + fprintf(stderr, "warn: failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno); + return false; + } + return true; +} + +#endif + // // CLI argument parsing // -void gpt_params_handle_hf_token(gpt_params & params) { - if (params.hf_token.empty() && std::getenv("HF_TOKEN")) { - params.hf_token = std::getenv("HF_TOKEN"); - } -} - void gpt_params_handle_model_default(gpt_params & params) { if (!params.hf_repo.empty()) { // short-hand to avoid specifying --hf-file -> default it to --model @@ -248,6 +354,30 @@ void gpt_params_handle_model_default(gpt_params & params) { } } +void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) { + int32_t n_set = 0; + + if (cpuparams.n_threads < 0) { + // Assuming everything about cpuparams is invalid + if (role_model != nullptr) { + cpuparams = *role_model; + } else { + cpuparams.n_threads = cpu_get_num_math(); + } + } + + for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) { + if (cpuparams.cpumask[i]) { + n_set++; + } + } + + if (n_set && n_set < cpuparams.n_threads) { + // Not enough set bits, may experience performance issues. + fprintf(stderr, "warn: Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads); + } +} + bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { bool invalid_param = false; std::string arg; @@ -267,13 +397,20 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { } } + postprocess_cpu_params(params.cpuparams, nullptr); + postprocess_cpu_params(params.cpuparams_batch, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams, ¶ms.cpuparams); + postprocess_cpu_params(params.draft_cpuparams_batch, ¶ms.cpuparams_batch); + if (params.prompt_cache_all && (params.interactive || params.interactive_first)) { throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } gpt_params_handle_model_default(params); - gpt_params_handle_hf_token(params); + if (params.hf_token.empty()) { + get_env("HF_TOKEN", params.hf_token); + } if (params.escape) { string_process_escapes(params.prompt); @@ -293,6 +430,32 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { return true; } +void gpt_params_parse_from_env(gpt_params & params) { + // we only care about server-related params for now + get_env("LLAMA_ARG_MODEL", params.model); + get_env("LLAMA_ARG_MODEL_URL", params.model_url); + get_env("LLAMA_ARG_MODEL_ALIAS", params.model_alias); + get_env("LLAMA_ARG_HF_REPO", params.hf_repo); + get_env("LLAMA_ARG_HF_FILE", params.hf_file); + get_env("LLAMA_ARG_THREADS", params.cpuparams.n_threads); + get_env("LLAMA_ARG_CTX_SIZE", params.n_ctx); + get_env("LLAMA_ARG_N_PARALLEL", params.n_parallel); + get_env("LLAMA_ARG_BATCH", params.n_batch); + get_env("LLAMA_ARG_UBATCH", params.n_ubatch); + get_env("LLAMA_ARG_N_GPU_LAYERS", params.n_gpu_layers); + get_env("LLAMA_ARG_THREADS_HTTP", params.n_threads_http); + get_env("LLAMA_ARG_CHAT_TEMPLATE", params.chat_template); + get_env("LLAMA_ARG_N_PREDICT", params.n_predict); + get_env("LLAMA_ARG_ENDPOINT_METRICS", params.endpoint_metrics); + get_env("LLAMA_ARG_ENDPOINT_SLOTS", params.endpoint_slots); + get_env("LLAMA_ARG_EMBEDDINGS", params.embedding); + get_env("LLAMA_ARG_FLASH_ATTN", params.flash_attn); + get_env("LLAMA_ARG_DEFRAG_THOLD", params.defrag_thold); + get_env("LLAMA_ARG_CONT_BATCHING", params.cont_batching); + get_env("LLAMA_ARG_HOST", params.hostname); + get_env("LLAMA_ARG_PORT", params.port); +} + bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { const auto params_org = params; // the example can modify the default params @@ -311,6 +474,79 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { return true; } +bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) { + size_t dash_loc = range.find('-'); + if (dash_loc == std::string::npos) { + fprintf(stderr, "Format of CPU range is invalid! Expected []-[].\n"); + return false; + } + + size_t start_i; + size_t end_i; + + if (dash_loc == 0) { + start_i = 0; + } else { + start_i = std::stoull(range.substr(0, dash_loc)); + if (start_i >= GGML_MAX_N_THREADS) { + fprintf(stderr, "Start index out of bounds!\n"); + return false; + } + } + + if (dash_loc == range.length() - 1) { + end_i = GGML_MAX_N_THREADS - 1; + } else { + end_i = std::stoull(range.substr(dash_loc + 1)); + if (end_i >= GGML_MAX_N_THREADS) { + fprintf(stderr, "End index out of bounds!\n"); + return false; + } + } + + for (size_t i = start_i; i <= end_i; i++) { + boolmask[i] = true; + } + + return true; +} + +bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREADS]) { + // Discard potential 0x prefix + size_t start_i = 0; + if (mask.length() >= 2 && mask.substr(0, 2) == "0x") { + start_i = 2; + } + + size_t num_digits = mask.length() - start_i; + if (num_digits > 128) num_digits = 128; + + size_t end_i = num_digits + start_i; + + for (size_t i = start_i, n = (num_digits*4 - 1); i < end_i; i++, n-=4) { + char c = mask.at(i); + int8_t id = c; + + if ((c >= '0' && c <= '9')) { + id -= '0'; + } else if (c >= 'a' && c <= 'f') { + id -= 'a' - 10; + } else if (c >= 'A' && c <= 'F') { + id -= 'A' - 10; + } else { + fprintf(stderr, "Invalid hex character '%c' at position %d\n", c, int32_t(i)); + return false; + } + + boolmask[ n ] = boolmask[ n ] || ((id & 8) != 0); + boolmask[n - 1] = boolmask[n - 1] || ((id & 4) != 0); + boolmask[n - 2] = boolmask[n - 2] || ((id & 2) != 0); + boolmask[n - 3] = boolmask[n - 3] || ((id & 1) != 0); + } + + return true; +} + #define CHECK_ARG if (++i >= argc) { invalid_param = true; return true; } bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) { @@ -327,36 +563,142 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa } if (arg == "-t" || arg == "--threads") { CHECK_ARG - params.n_threads = std::stoi(argv[i]); - if (params.n_threads <= 0) { - params.n_threads = std::thread::hardware_concurrency(); + params.cpuparams.n_threads = std::stoi(argv[i]); + if (params.cpuparams.n_threads <= 0) { + params.cpuparams.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-C" || arg == "--cpu-mask") { + CHECK_ARG + std::string mask = argv[i]; + params.cpuparams.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.cpuparams.cpumask); + return true; + } + if (arg == "-Cr" || arg == "--cpu-range") { + CHECK_ARG + std::string range = argv[i]; + params.cpuparams.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.cpuparams.cpumask); + return true; + } + if (arg == "--prio") { + CHECK_ARG + params.cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict") { + CHECK_ARG + params.cpuparams.strict_cpu = std::stoul(argv[i]); + return true; + } + if (arg == "--poll") { + CHECK_ARG + params.cpuparams.poll = std::stoul(argv[i]); + return true; + } if (arg == "-tb" || arg == "--threads-batch") { CHECK_ARG - params.n_threads_batch = std::stoi(argv[i]); - if (params.n_threads_batch <= 0) { - params.n_threads_batch = std::thread::hardware_concurrency(); + params.cpuparams_batch.n_threads = std::stoi(argv[i]); + if (params.cpuparams_batch.n_threads <= 0) { + params.cpuparams_batch.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-Cb" || arg == "--cpu-mask-batch") { + CHECK_ARG + std::string mask = argv[i]; + params.cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.cpuparams_batch.cpumask); + return true; + } + if (arg == "-Crb" || arg == "--cpu-range_batch") { + CHECK_ARG + std::string range = argv[i]; + params.cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.cpuparams_batch.cpumask); + return true; + } + if (arg == "--prio-batch") { + CHECK_ARG + params.cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-batch") { + params.cpuparams_batch.strict_cpu = true; + return true; + } + if (arg == "--poll-batch") { + CHECK_ARG + params.cpuparams_batch.poll = std::stoul(argv[i]); + return true; + } if (arg == "-td" || arg == "--threads-draft") { CHECK_ARG - params.n_threads_draft = std::stoi(argv[i]); - if (params.n_threads_draft <= 0) { - params.n_threads_draft = std::thread::hardware_concurrency(); + params.draft_cpuparams.n_threads = std::stoi(argv[i]); + if (params.draft_cpuparams.n_threads <= 0) { + params.draft_cpuparams.n_threads = std::thread::hardware_concurrency(); } return true; + } + if (arg == "-Cd" || arg == "--cpu-mask-draft") { + CHECK_ARG + std::string mask = argv[i]; + params.draft_cpuparams.mask_valid = true; + invalid_param = !parse_cpu_mask(mask, params.draft_cpuparams.cpumask); + return true; + } + if (arg == "-Crd" || arg == "--cpu-range-draft") { + CHECK_ARG + std::string range = argv[i]; + params.draft_cpuparams.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.draft_cpuparams.cpumask); + return true; + } + if (arg == "--prio-draft") { + CHECK_ARG + params.draft_cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-draft") { + params.draft_cpuparams.strict_cpu = true; + return true; + } + if (arg == "--poll-draft") { + CHECK_ARG + params.draft_cpuparams.poll = std::stoul(argv[i]); + return true; } if (arg == "-tbd" || arg == "--threads-batch-draft") { CHECK_ARG - params.n_threads_batch_draft = std::stoi(argv[i]); - if (params.n_threads_batch_draft <= 0) { - params.n_threads_batch_draft = std::thread::hardware_concurrency(); + params.draft_cpuparams_batch.n_threads = std::stoi(argv[i]); + if (params.draft_cpuparams_batch.n_threads <= 0) { + params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency(); } return true; } + if (arg == "-Crbd" || arg == "--cpu-range-batch-draft") { + CHECK_ARG + std::string range = argv[i]; + params.draft_cpuparams_batch.mask_valid = true; + invalid_param = !parse_cpu_range(range, params.draft_cpuparams_batch.cpumask); + return true; + } + if (arg == "--prio-batch-draft") { + CHECK_ARG + params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]); + return true; + } + if (arg == "--cpu-strict-batch-draft") { + params.draft_cpuparams_batch.strict_cpu = true; + return true; + } + if (arg == "--poll-batch-draft") { + CHECK_ARG + params.draft_cpuparams_batch.poll = std::stoul(argv[i]); + return true; + } if (arg == "-p" || arg == "--prompt") { CHECK_ARG params.prompt = argv[i]; @@ -851,7 +1193,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa } return true; } - if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--gpu-layers-draft") { + if (arg == "-ngld" || arg == "--gpu-layers-draft" || arg == "--n-gpu-layers-draft") { CHECK_ARG params.n_gpu_layers_draft = std::stoi(argv[i]); if (!llama_supports_gpu_offload()) { @@ -1441,11 +1783,40 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "*", " --no-display-prompt", "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" }); options.push_back({ "*", "-co, --color", "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" }); options.push_back({ "*", "-s, --seed SEED", "RNG seed (default: %d, use random seed for < 0)", params.seed }); - options.push_back({ "*", "-t, --threads N", "number of threads to use during generation (default: %d)", params.n_threads }); + options.push_back({ "*", "-t, --threads N", "number of threads to use during generation (default: %d)", params.cpuparams.n_threads }); options.push_back({ "*", "-tb, --threads-batch N", "number of threads to use during batch and prompt processing (default: same as --threads)" }); options.push_back({ "speculative", "-td, --threads-draft N", "number of threads to use during generation (default: same as --threads)" }); - options.push_back({ "speculative", "-tbd, --threads-batch-draft N", - "number of threads to use during batch and prompt processing (default: same as --threads-draft)" }); + options.push_back({ "speculative", "-tbd, --threads-batch-draft N","number of threads to use during batch and prompt processing (default: same as --threads-draft)" }); + +#ifndef GGML_USE_OPENMP + // these options are available only with the internal threadpool + options.push_back({ "*", "-C, --cpu-mask M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")"}); + options.push_back({ "*", "-Cr, --cpu-range lo-hi", "range of CPUs for affinity. Complements --cpu-mask"}); + options.push_back({ "*", " --cpu-strict <0|1>", "use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu}); + options.push_back({ "*", " --priority N", "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority}); + options.push_back({ "*", " --poll <0...100>", "use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll}); + + options.push_back({ "*", "-Cb, --cpu-mask-batch M", "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)"}); + options.push_back({ "*", "-Crb, --cpu-range-batch lo-hi", "ranges of CPUs for affinity. Complements --cpu-mask-batch"}); + options.push_back({ "*", " --cpu-strict-batch <0|1>","use strict CPU placement (default: same as --cpu-strict)"}); + options.push_back({ "*", " --priority-batch N", "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority)"}); + options.push_back({ "*", " --poll-batch <0|1>", "use polling to wait for work (default: same as --poll"}); + + options.push_back({ "speculative", "-Cd, --cpu-mask-draft M", "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)"}); + options.push_back({ "speculative", "-Crd, --cpu-range-draft lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft"}); + options.push_back({ "speculative", " --cpu-strict-draft <0|1>","Use strict CPU placement for draft model (default: same as --cpu-strict)"}); + options.push_back({ "speculative", " --priority-draft N", "Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: same as --priority)"}); + options.push_back({ "speculative", " --poll-draft <0|1>", "Use polling to wait for draft model work (default: same as --poll])"}); + + options.push_back({ "speculative", "-Cbd, --cpu-mask-batch-draft M","Draft model CPU affinity mask. Complements cpu-range-draft-batch (default: same as --cpu-mask-draft)"}); + options.push_back({ "speculative", "-Crbd, --cpu-range-batch-draft lo-hi", + "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)"}); + options.push_back({ "speculative", " --cpu-strict-batch-draft <0|1>", + "Use strict CPU placement for draft model (default: --cpu-strict-draft)"}); + options.push_back({ "speculative", " --priority-batch-draft N","Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority-draft)"}); + options.push_back({ "speculative", " --poll-batch-draft <0|1>","Use polling to wait for draft model work (default: --poll-draft)"}); +#endif // GGML_USE_OPENMP + options.push_back({ "speculative", " --draft N", "number of tokens to draft for speculative decoding (default: %d)", params.n_draft }); options.push_back({ "speculative", "-ps, --p-split N", "speculative decoding split probability (default: %.1f)", (double)params.p_split }); options.push_back({ "*", "-lcs, --lookup-cache-static FNAME", @@ -1717,7 +2088,6 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "export-lora", "-m, --model", "model path from which to load base model (default '%s')", params.model.c_str() }); options.push_back({ "export-lora", " --lora FNAME", "path to LoRA adapter (can be repeated to use multiple adapters)" }); options.push_back({ "export-lora", " --lora-scaled FNAME S", "path to LoRA adapter with user defined scaling S (can be repeated to use multiple adapters)" }); - options.push_back({ "*", "-t, --threads N", "number of threads to use during computation (default: %d)", params.n_threads }); options.push_back({ "export-lora", "-o, --output FNAME", "output file (default: '%s')", params.lora_outfile.c_str() }); printf("usage: %s [options]\n", argv[0]); @@ -1749,11 +2119,17 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param std::string gpt_params_get_system_info(const gpt_params & params) { std::ostringstream os; - os << "system_info: n_threads = " << params.n_threads; - if (params.n_threads_batch != -1) { - os << " (n_threads_batch = " << params.n_threads_batch << ")"; + os << "system_info: n_threads = " << params.cpuparams.n_threads; + if (params.cpuparams_batch.n_threads != -1) { + os << " (n_threads_batch = " << params.cpuparams_batch.n_threads << ")"; } +#if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later + // TODO: windows + arm64 + mingw64 + DWORD logicalProcessorCount = GetActiveProcessorCount(ALL_PROCESSOR_GROUPS); + os << " / " << logicalProcessorCount << " | " << llama_print_system_info(); +#else os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info(); +#endif return os.str(); } @@ -1803,6 +2179,23 @@ std::string string_get_sortable_timestamp() { return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns); } +void string_replace_all(std::string & s, const std::string & search, const std::string & replace) { + if (search.empty()) { + return; + } + std::string builder; + builder.reserve(s.length()); + size_t pos = 0; + size_t last_pos = 0; + while ((pos = s.find(search, last_pos)) != std::string::npos) { + builder.append(s, last_pos, pos - last_pos); + builder.append(replace); + last_pos = pos + search.length(); + } + builder.append(s, last_pos, std::string::npos); + s = std::move(builder); +} + void string_process_escapes(std::string & input) { std::size_t input_len = input.length(); std::size_t output_idx = 0; @@ -2171,7 +2564,9 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) { tmp.clear(); tmp.push_back(decoder_start_token_id); } - llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0)); + if (llama_model_has_decoder(model)) { + llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0)); + } llama_kv_cache_clear(lctx); llama_synchronize(lctx); llama_reset_timings(lctx); @@ -2250,8 +2645,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param cparams.n_seq_max = params.n_parallel; cparams.n_batch = params.n_batch; cparams.n_ubatch = params.n_ubatch; - cparams.n_threads = params.n_threads; - cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; + cparams.n_threads = params.cpuparams.n_threads; + cparams.n_threads_batch = params.cpuparams_batch.n_threads == -1 ? + params.cpuparams.n_threads : params.cpuparams_batch.n_threads; cparams.seed = params.seed; cparams.logits_all = params.logits_all; cparams.embeddings = params.embedding; @@ -2277,6 +2673,22 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param return cparams; } +struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) { + struct ggml_threadpool_params tpp; + + ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults + + if (params.mask_valid) { + std::memcpy(&tpp.cpumask, ¶ms.cpumask, GGML_MAX_N_THREADS); + } + + tpp.prio = params.priority; + tpp.poll = params.poll; + tpp.strict_cpu = params.strict_cpu; + + return tpp; +} + #ifdef LLAMA_USE_CURL static bool starts_with(const std::string & str, const std::string & prefix) { @@ -2715,12 +3127,6 @@ std::string llama_detokenize(llama_context * ctx, const std::vector return text; } -bool llama_should_add_bos_token(const llama_model * model) { - const int add_bos = llama_add_bos_token(model); - - return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM); -} - // // Chat template utils // @@ -3272,7 +3678,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector); fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z); - fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency()); + fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency()); fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k); fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p); fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p); diff --git a/llama/common.h b/llama/common.h index b25dc42e..6ac94243 100644 --- a/llama/common.h +++ b/llama/common.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -93,13 +93,18 @@ enum dimre_method { DIMRE_METHOD_MEAN, }; +struct cpu_params { + int n_threads = -1; + bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask. + bool mask_valid = false; // Default: any CPU + enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime) + bool strict_cpu = false; // Use strict CPU placement + uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling) +}; + struct gpt_params { uint32_t seed = LLAMA_DEFAULT_SEED; // RNG seed - int32_t n_threads = cpu_get_num_math(); - int32_t n_threads_draft = -1; - int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads) - int32_t n_threads_batch_draft = -1; int32_t n_predict = -1; // new tokens to predict int32_t n_ctx = 0; // context size int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS) @@ -126,6 +131,11 @@ struct gpt_params { int32_t yarn_orig_ctx = 0; // YaRN original context length float defrag_thold = -1.0f; // KV cache defragmentation threshold + struct cpu_params cpuparams; + struct cpu_params cpuparams_batch; + struct cpu_params draft_cpuparams; + struct cpu_params draft_cpuparams_batch; + ggml_backend_sched_eval_callback cb_eval = nullptr; void * cb_eval_user_data = nullptr; @@ -230,7 +240,7 @@ struct gpt_params { int32_t port = 8080; // server listens on this network port int32_t timeout_read = 600; // http read timeout in seconds int32_t timeout_write = timeout_read; // http write timeout in seconds - int32_t n_threads_http = -1; // number of threads to process HTTP requests + int n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool) std::string hostname = "127.0.0.1"; std::string public_path = ""; @@ -293,7 +303,7 @@ struct gpt_params { std::string lora_outfile = "ggml-lora-merged-f16.gguf"; }; -void gpt_params_handle_hf_token(gpt_params & params); +void gpt_params_parse_from_env(gpt_params & params); void gpt_params_handle_model_default(gpt_params & params); bool gpt_params_parse_ex (int argc, char ** argv, gpt_params & params); @@ -303,6 +313,11 @@ void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params); std::string gpt_params_get_system_info(const gpt_params & params); +bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]); +bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]); +void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr); +bool set_process_priority(enum ggml_sched_priority prio); + // // String utils // @@ -312,6 +327,8 @@ std::vector string_split(std::string input, char separator); std::string string_strip(const std::string & str); std::string string_get_sortable_timestamp(); +void string_replace_all(std::string & s, const std::string & search, const std::string & replace); + template static std::vector string_split(const std::string & str, char delim) { std::vector values; @@ -351,8 +368,9 @@ struct llama_init_result { struct llama_init_result llama_init_from_gpt_params(gpt_params & params); -struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params); -struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params); +struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params); +struct llama_context_params llama_context_params_from_gpt_params (const gpt_params & params); +struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params); struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params); struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params); @@ -404,10 +422,6 @@ std::string llama_detokenize( const std::vector & tokens, bool special = true); -// Uses the value from the model metadata if possible, otherwise -// defaults to true when model type is SPM, otherwise false. -bool llama_should_add_bos_token(const llama_model * model); - // // Chat template utils // diff --git a/llama/ggml-aarch64.c b/llama/ggml-aarch64.c index ea751c58..5ffad03c 100644 --- a/llama/ggml-aarch64.c +++ b/llama/ggml-aarch64.c @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -42,6 +42,8 @@ #if defined(__GNUC__) #pragma GCC diagnostic ignored "-Woverlength-strings" +#elif defined(_MSC_VER) +#pragma warning(disable: 4244 4267) // possible loss of data #endif #define UNUSED GGML_UNUSED @@ -361,33 +363,18 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds } size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - if (!quant_weights) { - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4); - } - else { - assert(false); - return 0; - } + UNUSED(quant_weights); + return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4); } size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - if (!quant_weights) { - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8); - } - else { - assert(false); - return 0; - } + UNUSED(quant_weights); + return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8); } size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - if (!quant_weights) { - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8); - } - else { - assert(false); - return 0; - } + UNUSED(quant_weights); + return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8); } void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { diff --git a/llama/ggml-aarch64.h b/llama/ggml-aarch64.h index d8f7fd3a..0fc24e80 100644 --- a/llama/ggml-aarch64.h +++ b/llama/ggml-aarch64.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-alloc.c b/llama/ggml-alloc.c index 0bb174f4..edd33746 100644 --- a/llama/ggml-alloc.c +++ b/llama/ggml-alloc.c @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-alloc.h b/llama/ggml-alloc.h index 8c8cb4eb..017d5298 100644 --- a/llama/ggml-alloc.h +++ b/llama/ggml-alloc.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -33,8 +33,8 @@ extern "C" { #endif typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t; -typedef struct ggml_backend_buffer * ggml_backend_buffer_t; -typedef struct ggml_backend * ggml_backend_t; +typedef struct ggml_backend_buffer * ggml_backend_buffer_t; +typedef struct ggml_backend * ggml_backend_t; // Tensor allocator struct ggml_tallocr { diff --git a/llama/ggml-backend-impl.h b/llama/ggml-backend-impl.h index e3a3d335..aeafdbd5 100644 --- a/llama/ggml-backend-impl.h +++ b/llama/ggml-backend-impl.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-backend.c b/llama/ggml-backend.c index bfc6e10a..08af0149 100644 --- a/llama/ggml-backend.c +++ b/llama/ggml-backend.c @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -382,15 +382,10 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b } // an async copy would normally happen after all the queued operations on both backends are completed - // sync src, set_async dst - if (ggml_backend_buffer_is_host(src->buffer)) { - ggml_backend_synchronize(backend_src); - ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src)); - } else { - ggml_backend_synchronize(backend_src); - ggml_backend_tensor_copy(src, dst); - ggml_backend_synchronize(backend_dst); - } + // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy + ggml_backend_synchronize(backend_src); + ggml_backend_synchronize(backend_dst); + ggml_backend_tensor_copy(src, dst); } // events @@ -758,9 +753,11 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) { #endif struct ggml_backend_cpu_context { - int n_threads; - void * work_data; - size_t work_size; + int n_threads; + ggml_threadpool_t threadpool; + + void * work_data; + size_t work_size; ggml_abort_callback abort_callback; void * abort_callback_data; @@ -795,7 +792,7 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu)); - cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads); + cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool); cpu_plan->cgraph = *cgraph; // FIXME: deep copy if (cpu_plan->cplan.work_size > 0) { @@ -832,7 +829,7 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backe GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context; - struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads); + struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool); if (cpu_ctx->work_size < cplan.work_size) { free(cpu_ctx->work_data); @@ -909,6 +906,7 @@ ggml_backend_t ggml_backend_cpu_init(void) { } ctx->n_threads = GGML_DEFAULT_N_THREADS; + ctx->threadpool = NULL; ctx->work_data = NULL; ctx->work_size = 0; ctx->abort_callback = NULL; @@ -939,6 +937,18 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) { ctx->n_threads = n_threads; } +void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) { + GGML_ASSERT(ggml_backend_is_cpu(backend_cpu)); + + struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context; + + if (ctx->threadpool && ctx->threadpool != threadpool) { + // already had a different threadpool, pause/suspend it before switching + ggml_threadpool_pause(ctx->threadpool); + } + ctx->threadpool = threadpool; +} + void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) { GGML_ASSERT(ggml_backend_is_cpu(backend_cpu)); @@ -1054,10 +1064,6 @@ static bool ggml_is_view_op(enum ggml_op op) { #define GGML_SCHED_MAX_BACKENDS 16 #endif -#ifndef GGML_SCHED_MAX_SPLITS -#define GGML_SCHED_MAX_SPLITS 2048 -#endif - #ifndef GGML_SCHED_MAX_SPLIT_INPUTS #define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC #endif @@ -1161,7 +1167,8 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co } #if 0 -static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only +#define GGML_SCHED_MAX_SPLITS_DEBUG 4096 +static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__) #define GET_CAUSE(node) causes[hash_id(node)] #else @@ -1585,7 +1592,6 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg sched->splits = realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split)); GGML_ASSERT(sched->splits != NULL); } - GGML_ASSERT(i_split < GGML_SCHED_MAX_SPLITS); split = &sched->splits[i_split]; split->backend_id = node_backend_id; split->i_start = i; @@ -1813,7 +1819,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s } else { ggml_backend_synchronize(split_backend); } - ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy); + // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events + // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface + if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) { + ggml_backend_synchronize(input_backend); + if (sched->events[split_backend_id][sched->cur_copy] != NULL) { + ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]); + } else { + ggml_backend_synchronize(split_backend); + } + ggml_backend_tensor_copy(input, input_cpy); + } } } @@ -1891,13 +1907,14 @@ ggml_backend_sched_t ggml_backend_sched_new( sched->hv_tensor_backend_ids = malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0])); sched->hv_tensor_copies = malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *)); - const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2; + const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph + const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2; sched->node_backend_ids = calloc(nodes_size, sizeof(sched->node_backend_ids[0])); sched->leaf_backend_ids = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0])); sched->prev_node_backend_ids = calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0])); sched->prev_leaf_backend_ids = calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0])); - sched->context_buffer_size = GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false); + sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false); sched->context_buffer = malloc(sched->context_buffer_size); const int initial_splits_capacity = 16; diff --git a/llama/ggml-backend.h b/llama/ggml-backend.h index 7cf3d939..48385bc4 100644 --- a/llama/ggml-backend.h +++ b/llama/ggml-backend.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -89,6 +89,7 @@ extern "C" { GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + // "offset" refers to the offset of the tensor data for setting/getting data GGML_API GGML_CALL void ggml_backend_tensor_set( struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); @@ -128,6 +129,7 @@ extern "C" { GGML_API GGML_CALL bool ggml_backend_is_cpu (ggml_backend_t backend); GGML_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads); + GGML_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool); GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data); // Create a backend buffer from an existing pointer diff --git a/llama/ggml-common.h b/llama/ggml-common.h index 80c80024..b449d362 100644 --- a/llama/ggml-common.h +++ b/llama/ggml-common.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda.cu b/llama/ggml-cuda.cu index 8b310ae4..0d1f2b53 100644 --- a/llama/ggml-cuda.cu +++ b/llama/ggml-cuda.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -35,8 +35,10 @@ #include "ggml-cuda/binbcast.cuh" #include "ggml-cuda/clamp.cuh" #include "ggml-cuda/concat.cuh" +#include "ggml-cuda/conv-transpose-1d.cuh" #include "ggml-cuda/convert.cuh" #include "ggml-cuda/cpy.cuh" +#include "ggml-cuda/cross-entropy-loss.cuh" #include "ggml-cuda/diagmask.cuh" #include "ggml-cuda/dmmv.cuh" #include "ggml-cuda/fattn.cuh" @@ -55,7 +57,6 @@ #include "ggml-cuda/tsembd.cuh" #include "ggml-cuda/unary.cuh" #include "ggml-cuda/upscale.cuh" -#include "ggml-cuda/conv-transpose-1d.cuh" #include #include @@ -2211,6 +2212,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_ADD: ggml_cuda_op_add(ctx, dst); break; + case GGML_OP_SUB: + ggml_cuda_op_sub(ctx, dst); + break; case GGML_OP_ACC: ggml_cuda_op_acc(ctx, dst); break; @@ -2297,6 +2301,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_SQRT: ggml_cuda_op_sqrt(ctx, dst); break; + case GGML_OP_SIN: + ggml_cuda_op_sin(ctx, dst); + break; + case GGML_OP_COS: + ggml_cuda_op_cos(ctx, dst); + break; case GGML_OP_CLAMP: ggml_cuda_op_clamp(ctx, dst); break; @@ -2333,6 +2343,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_FLASH_ATTN_EXT: ggml_cuda_flash_attn_ext(ctx, dst); break; + case GGML_OP_CROSS_ENTROPY_LOSS: + ggml_cuda_cross_entropy_loss(ctx, dst); + break; default: return false; } @@ -2388,33 +2401,35 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, } GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) { - GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst)); - ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer; ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer; - if (!ggml_backend_buffer_is_cuda(src->buffer)) { + if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) { return false; } - if (!ggml_backend_buffer_is_cuda(dst->buffer)) { + if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) { return false; } - // device -> device + // device -> device copy ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context; ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context; + ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context; + ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context; + + if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) { +#ifndef NDEBUG + GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__); +#endif + return false; + } + if (backend_src != backend_dst) { - ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context; - ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context; - - GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device); - GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device); - // copy on src stream if (cuda_ctx_src->device == cuda_ctx_dst->device) { - CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream())); + CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream())); } else { #ifdef GGML_CUDA_NO_PEER_COPY return false; @@ -2423,7 +2438,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_ #endif } - // record event on src stream + // record event on src stream after the copy if (!cuda_ctx_src->copy_event) { ggml_cuda_set_device(cuda_ctx_src->device); CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming)); @@ -2435,7 +2450,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_ CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0)); } else { // src and dst are on the same backend - CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream())); + CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream())); } return true; } @@ -2638,6 +2653,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device)); for (int j = 0; j < GGML_MAX_SRC; j++) { if (node->src[j] != nullptr) { + assert(node->src[j]->buffer); assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer)); } } @@ -2772,11 +2788,12 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_MUL_MAT_ID: { struct ggml_tensor * a = op->src[0]; - if (op->op == GGML_OP_MUL_MAT) { - struct ggml_tensor * b = op->src[1]; - if (a->ne[3] != b->ne[3]) { - return false; - } + struct ggml_tensor * b = op->src[1]; + if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) { + return false; + } + if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) { + return false; } switch (a->type) { case GGML_TYPE_F32: @@ -2880,12 +2897,15 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons case GGML_OP_TRANSPOSE: case GGML_OP_NORM: case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_RMS_NORM: case GGML_OP_SCALE: case GGML_OP_SQR: case GGML_OP_SQRT: + case GGML_OP_SIN: + case GGML_OP_COS: case GGML_OP_CLAMP: case GGML_OP_CONT: case GGML_OP_DIAG_MASK_INF: @@ -2907,7 +2927,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons return true; case GGML_OP_FLASH_ATTN_EXT: #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) - return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128; + return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128; #else if (op->src[0]->ne[0] == 128) { return true; @@ -2917,6 +2937,8 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons } return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16; + case GGML_OP_CROSS_ENTROPY_LOSS: + return true; #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) default: return false; diff --git a/llama/ggml-cuda.h b/llama/ggml-cuda.h index d5616d8b..a08cd816 100644 --- a/llama/ggml-cuda.h +++ b/llama/ggml-cuda.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/acc.cu b/llama/ggml-cuda/acc.cu index cb7eb35a..7c6d6ea0 100644 --- a/llama/ggml-cuda/acc.cu +++ b/llama/ggml-cuda/acc.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/acc.cuh b/llama/ggml-cuda/acc.cuh index 126a4d6d..23f09703 100644 --- a/llama/ggml-cuda/acc.cuh +++ b/llama/ggml-cuda/acc.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/arange.cu b/llama/ggml-cuda/arange.cu index 8f9b86e3..c36ac823 100644 --- a/llama/ggml-cuda/arange.cu +++ b/llama/ggml-cuda/arange.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/arange.cuh b/llama/ggml-cuda/arange.cuh index 10b0a6fc..186f90c0 100644 --- a/llama/ggml-cuda/arange.cuh +++ b/llama/ggml-cuda/arange.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/argsort.cu b/llama/ggml-cuda/argsort.cu index 796633c6..85f62458 100644 --- a/llama/ggml-cuda/argsort.cu +++ b/llama/ggml-cuda/argsort.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/argsort.cuh b/llama/ggml-cuda/argsort.cuh index 75f434c2..7fc2c9d7 100644 --- a/llama/ggml-cuda/argsort.cuh +++ b/llama/ggml-cuda/argsort.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/binbcast.cu b/llama/ggml-cuda/binbcast.cu index c8262f3d..4dc774d0 100644 --- a/llama/ggml-cuda/binbcast.cu +++ b/llama/ggml-cuda/binbcast.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -35,6 +35,10 @@ static __device__ __forceinline__ float op_add(const float a, const float b) { return a + b; } +static __device__ __forceinline__ float op_sub(const float a, const float b) { + return a - b; +} + static __device__ __forceinline__ float op_mul(const float a, const float b) { return a * b; } @@ -297,6 +301,10 @@ void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { ggml_cuda_op_bin_bcast>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream()); } +void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_bin_bcast>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream()); +} + void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { ggml_cuda_op_bin_bcast>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream()); } diff --git a/llama/ggml-cuda/binbcast.cuh b/llama/ggml-cuda/binbcast.cuh index 74348c78..e433e538 100644 --- a/llama/ggml-cuda/binbcast.cuh +++ b/llama/ggml-cuda/binbcast.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -28,5 +28,6 @@ void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst); +void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/llama/ggml-cuda/clamp.cu b/llama/ggml-cuda/clamp.cu index cf0a60fd..8f5f2eda 100644 --- a/llama/ggml-cuda/clamp.cu +++ b/llama/ggml-cuda/clamp.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/clamp.cuh b/llama/ggml-cuda/clamp.cuh index c7b6a414..e0a3c3e4 100644 --- a/llama/ggml-cuda/clamp.cuh +++ b/llama/ggml-cuda/clamp.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/common.cuh b/llama/ggml-cuda/common.cuh index 37079ab6..bd3838b1 100644 --- a/llama/ggml-cuda/common.cuh +++ b/llama/ggml-cuda/common.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/concat.cu b/llama/ggml-cuda/concat.cu index c73b3311..72c36e40 100644 --- a/llama/ggml-cuda/concat.cu +++ b/llama/ggml-cuda/concat.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/concat.cuh b/llama/ggml-cuda/concat.cuh index 5869b398..c03cd437 100644 --- a/llama/ggml-cuda/concat.cuh +++ b/llama/ggml-cuda/concat.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/conv-transpose-1d.cu b/llama/ggml-cuda/conv-transpose-1d.cu index 4124f559..ab5e6bdf 100644 --- a/llama/ggml-cuda/conv-transpose-1d.cu +++ b/llama/ggml-cuda/conv-transpose-1d.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/conv-transpose-1d.cuh b/llama/ggml-cuda/conv-transpose-1d.cuh index 41990dde..891917b2 100644 --- a/llama/ggml-cuda/conv-transpose-1d.cuh +++ b/llama/ggml-cuda/conv-transpose-1d.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/convert.cu b/llama/ggml-cuda/convert.cu index 52d27ea3..046d5b74 100644 --- a/llama/ggml-cuda/convert.cu +++ b/llama/ggml-cuda/convert.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/convert.cuh b/llama/ggml-cuda/convert.cuh index 40e15280..e3f31c3c 100644 --- a/llama/ggml-cuda/convert.cuh +++ b/llama/ggml-cuda/convert.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/cpy.cu b/llama/ggml-cuda/cpy.cu index a4ee847b..c4334735 100644 --- a/llama/ggml-cuda/cpy.cu +++ b/llama/ggml-cuda/cpy.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/cpy.cuh b/llama/ggml-cuda/cpy.cuh index 96400051..7b140e30 100644 --- a/llama/ggml-cuda/cpy.cuh +++ b/llama/ggml-cuda/cpy.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/cross-entropy-loss.cu b/llama/ggml-cuda/cross-entropy-loss.cu new file mode 100644 index 00000000..55b9416f --- /dev/null +++ b/llama/ggml-cuda/cross-entropy-loss.cu @@ -0,0 +1,132 @@ +/** + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file + * + * MIT License + * + * Copyright (c) 2023-2024 The ggml authors + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "common.cuh" +#include "cross-entropy-loss.cuh" +#include "sumrows.cuh" + +#include +#include + +static __global__ void cross_entropy_loss_f32(const float * logits, const float * labels, float * dst, const int nclasses, const int k) { + const int warp_id = threadIdx.x / WARP_SIZE; + const int lane_id = threadIdx.x % WARP_SIZE; + const int i0 = blockDim.x*blockIdx.x + warp_id*WARP_SIZE; + + const int ne_tmp = WARP_SIZE*nclasses; + + extern __shared__ float tmp_all[]; + float * tmp_logits = tmp_all + (2*warp_id + 0)*ne_tmp; + float * tmp_labels = tmp_all + (2*warp_id + 1)*ne_tmp; + + // Each warp first loads ne_tmp logits/labels into shared memory: + for (int i = lane_id; i < ne_tmp; i += WARP_SIZE) { + const int ig = i0*nclasses + i; // ig == i global + + tmp_logits[i] = ig < k*nclasses ? logits[ig] : 0.0f; + tmp_labels[i] = ig < k*nclasses ? labels[ig] : 0.0f; + } + + // Each thread in the warp then calculates the cross entropy loss for a single row. + // TODO: pad in order to avoid shared memory bank conflicts. + + // Find maximum for softmax: + float max = -INFINITY; + for (int i = 0; i < nclasses; ++i) { + max = fmaxf(max, tmp_logits[lane_id*nclasses + i]); + } + + // Calculate log(softmax(logits)) which is just logits - max: + float sum = 0.0f; + for (int i = 0; i < nclasses; ++i) { + float val = tmp_logits[lane_id*nclasses + i] - max; + sum += expf(val); + tmp_logits[lane_id*nclasses + i] = val; + } + sum = logf(sum); + + // log(exp(logits - max) / sum) = (logits - max) - log(sum) + float loss = 0.0f; + for (int i = 0; i < nclasses; ++i) { + loss += (tmp_logits[lane_id*nclasses + i] - sum) * tmp_labels[lane_id*nclasses + i]; + } + loss = -warp_reduce_sum(loss) / (float)k; + + __syncthreads(); + + if (lane_id == 0) { + tmp_all[warp_id] = loss; + } + + __syncthreads(); + + if (warp_id != 0) { + return; + } + + loss = lane_id < CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE/WARP_SIZE ? tmp_all[lane_id] : 0.0f; + loss = warp_reduce_sum(loss); + + if (lane_id != 0) { + return; + } + + dst[blockIdx.x] = loss; +} + +void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + const float * src0_d = (const float *) src0->data; + const float * src1_d = (const float *) src1->data; + float * dst_d = (float *) dst->data; + + ggml_cuda_pool & pool = ctx.pool(); + cudaStream_t stream = ctx.stream(); + + const dim3 blocks_dim(CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1); + const dim3 blocks_num((nrows + CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE - 1) / CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1); + const int shmem = 2*CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE*ne00*sizeof(float); + + ggml_cuda_pool_alloc dst_tmp(pool, blocks_num.x); + + cross_entropy_loss_f32<<>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows); + + // Combine results from individual blocks: + sum_rows_f32_cuda(dst_tmp.ptr, dst_d, blocks_num.x, 1, stream); +} diff --git a/llama/ggml-cuda/cross-entropy-loss.cuh b/llama/ggml-cuda/cross-entropy-loss.cuh new file mode 100644 index 00000000..7f76ae6c --- /dev/null +++ b/llama/ggml-cuda/cross-entropy-loss.cuh @@ -0,0 +1,31 @@ +/** + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file + * + * MIT License + * + * Copyright (c) 2023-2024 The ggml authors + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "common.cuh" + +#define CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE 256 + +void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/llama/ggml-cuda/dequantize.cuh b/llama/ggml-cuda/dequantize.cuh index 60f0b66e..c464bc6b 100644 --- a/llama/ggml-cuda/dequantize.cuh +++ b/llama/ggml-cuda/dequantize.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/diagmask.cu b/llama/ggml-cuda/diagmask.cu index 1a6bcbc9..43433451 100644 --- a/llama/ggml-cuda/diagmask.cu +++ b/llama/ggml-cuda/diagmask.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/diagmask.cuh b/llama/ggml-cuda/diagmask.cuh index 34f76fe2..ee332322 100644 --- a/llama/ggml-cuda/diagmask.cuh +++ b/llama/ggml-cuda/diagmask.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/dmmv.cu b/llama/ggml-cuda/dmmv.cu index a606493d..0430f1fb 100644 --- a/llama/ggml-cuda/dmmv.cu +++ b/llama/ggml-cuda/dmmv.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/dmmv.cuh b/llama/ggml-cuda/dmmv.cuh index c68aa1d1..4e8c0f37 100644 --- a/llama/ggml-cuda/dmmv.cuh +++ b/llama/ggml-cuda/dmmv.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/fattn-common.cuh b/llama/ggml-cuda/fattn-common.cuh index ac1937a9..21fa5903 100644 --- a/llama/ggml-cuda/fattn-common.cuh +++ b/llama/ggml-cuda/fattn-common.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -48,6 +48,7 @@ typedef void (* fattn_kernel_t)( const float m0, const float m1, const uint32_t n_head_log2, + const float logit_softcap, const int ne00, const int ne01, const int ne02, @@ -683,11 +684,17 @@ void launch_fattn( const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]); const int shmem = 0; - float scale = 1.0f; - float max_bias = 0.0f; + float scale = 1.0f; + float max_bias = 0.0f; + float logit_softcap = 0.0f; - memcpy(&scale, (float *) KQV->op_params + 0, sizeof(float)); - memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float)); + memcpy(&scale, (float *) KQV->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float)); + memcpy(&logit_softcap, (float *) KQV->op_params + 2, sizeof(float)); + + if (logit_softcap != 0.0f) { + scale /= logit_softcap; + } const uint32_t n_head = Q->ne[2]; const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head)); @@ -701,7 +708,7 @@ void launch_fattn( V_data, mask ? ((const char *) mask->data) : nullptr, (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr, - scale, max_bias, m0, m1, n_head_log2, + scale, max_bias, m0, m1, n_head_log2, logit_softcap, Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3], K->ne[0], K->ne[1], K->ne[2], K->ne[3], mask ? mask->ne[1] : 0, mask ? mask->nb[1] : 0, diff --git a/llama/ggml-cuda/fattn-tile-f16.cu b/llama/ggml-cuda/fattn-tile-f16.cu index 4718d05e..c77de58e 100644 --- a/llama/ggml-cuda/fattn-tile-f16.cu +++ b/llama/ggml-cuda/fattn-tile-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -30,7 +30,7 @@ #define FATTN_KQ_STRIDE_TILE_F16 64 -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(nwarps*WARP_SIZE, 1) #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) @@ -46,6 +46,7 @@ static __global__ void flash_attn_tile_ext_f16( const float m0, const float m1, const uint32_t n_head_log2, + const float logit_softcap, const int ne00, const int ne01, const int ne02, @@ -70,6 +71,12 @@ static __global__ void flash_attn_tile_ext_f16( const int ne2, const int ne3) { #ifdef FP16_AVAILABLE + // Skip unused kernel variants for faster compilation: + if (use_logit_softcap && !(D == 128 || D == 256)) { + NO_DEVICE_CODE; + return; + } + //In this kernel Q, K, V are matrices while i, j, k are matrix indices. const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on. @@ -180,7 +187,13 @@ static __global__ void flash_attn_tile_ext_f16( for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) { const int j_KQ = j_KQ_0 + threadIdx.y; - half sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]); + half sum; + if (use_logit_softcap) { + const float2 tmp = __half22float2(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]); + sum = logit_softcap * tanhf(tmp.x + tmp.y); + } else { + sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]); + } sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f); kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum); @@ -296,20 +309,20 @@ static __global__ void flash_attn_tile_ext_f16( #endif // FP16_AVAILABLE } -template +template void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * Q = dst->src[0]; switch (Q->ne[0]) { case 64: { constexpr int D = 64; constexpr int nwarps = 8; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true); } break; case 128: { constexpr int D = 128; constexpr int nwarps = 8; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true); } break; default: { @@ -322,24 +335,45 @@ void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_ten const ggml_tensor * KQV = dst; const ggml_tensor * Q = dst->src[0]; - const int32_t precision = KQV->op_params[2]; + const int32_t precision = KQV->op_params[3]; GGML_ASSERT(precision == GGML_PREC_DEFAULT); + float logit_softcap; + memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); + if (Q->ne[1] <= 16) { constexpr int cols_per_block = 16; constexpr int parallel_blocks = 4; - launch_fattn_tile_f16_64_128(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + launch_fattn_tile_f16_64_128(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + launch_fattn_tile_f16_64_128(ctx, dst); + } return; } if (Q->ne[1] <= 32) { constexpr int cols_per_block = 32; constexpr int parallel_blocks = 4; - launch_fattn_tile_f16_64_128(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + launch_fattn_tile_f16_64_128(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + launch_fattn_tile_f16_64_128(ctx, dst); + } return; } constexpr int cols_per_block = 32; constexpr int parallel_blocks = 1; - launch_fattn_tile_f16_64_128(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + launch_fattn_tile_f16_64_128(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + launch_fattn_tile_f16_64_128(ctx, dst); + } } diff --git a/llama/ggml-cuda/fattn-tile-f16.cuh b/llama/ggml-cuda/fattn-tile-f16.cuh index fd9d0e22..0ac19a67 100644 --- a/llama/ggml-cuda/fattn-tile-f16.cuh +++ b/llama/ggml-cuda/fattn-tile-f16.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/fattn-tile-f32.cu b/llama/ggml-cuda/fattn-tile-f32.cu index 7c8b975c..ad3b1243 100644 --- a/llama/ggml-cuda/fattn-tile-f32.cu +++ b/llama/ggml-cuda/fattn-tile-f32.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -30,7 +30,7 @@ #define FATTN_KQ_STRIDE_TILE_F32 32 -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(nwarps*WARP_SIZE, 1) #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) @@ -46,6 +46,7 @@ static __global__ void flash_attn_tile_ext_f32( const float m0, const float m1, const uint32_t n_head_log2, + const float logit_softcap, const int ne00, const int ne01, const int ne02, @@ -69,6 +70,12 @@ static __global__ void flash_attn_tile_ext_f32( const int ne1, const int ne2, const int ne3) { + // Skip unused kernel variants for faster compilation: + if (use_logit_softcap && !(D == 128 || D == 256)) { + NO_DEVICE_CODE; + return; + } + //In this kernel Q, K, V are matrices while i, j, k are matrix indices. const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on. @@ -177,6 +184,10 @@ static __global__ void flash_attn_tile_ext_f32( for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) { const int j_KQ = j_KQ_0 + threadIdx.y; + if (use_logit_softcap) { + sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] = logit_softcap * tanhf(sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]); + } + sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f; kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]); @@ -293,20 +304,20 @@ static __global__ void flash_attn_tile_ext_f32( } } -template +template void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * Q = dst->src[0]; switch (Q->ne[0]) { case 64: { constexpr int D = 64; constexpr int nwarps = 8; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true); } break; case 128: { constexpr int D = 128; constexpr int nwarps = 8; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true); } break; default: { @@ -316,23 +327,45 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * } void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * KQV = dst; const ggml_tensor * Q = dst->src[0]; + float logit_softcap; + memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); + if (Q->ne[1] <= 16) { constexpr int cols_per_block = 16; constexpr int parallel_blocks = 4; - launch_fattn_tile_f32_64_128(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + launch_fattn_tile_f32_64_128(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + launch_fattn_tile_f32_64_128(ctx, dst); + } return; } if (Q->ne[1] <= 32) { constexpr int cols_per_block = 32; constexpr int parallel_blocks = 4; - launch_fattn_tile_f32_64_128(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + launch_fattn_tile_f32_64_128(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + launch_fattn_tile_f32_64_128(ctx, dst); + } return; } constexpr int cols_per_block = 32; constexpr int parallel_blocks = 1; - launch_fattn_tile_f32_64_128(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + launch_fattn_tile_f32_64_128(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + launch_fattn_tile_f32_64_128(ctx, dst); + } } diff --git a/llama/ggml-cuda/fattn-tile-f32.cuh b/llama/ggml-cuda/fattn-tile-f32.cuh index 67ae64e7..ff8fa4f4 100644 --- a/llama/ggml-cuda/fattn-tile-f32.cuh +++ b/llama/ggml-cuda/fattn-tile-f32.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/fattn-vec-f16.cuh b/llama/ggml-cuda/fattn-vec-f16.cuh index b89d7200..d10f5df8 100644 --- a/llama/ggml-cuda/fattn-vec-f16.cuh +++ b/llama/ggml-cuda/fattn-vec-f16.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -27,7 +27,7 @@ #include "common.cuh" #include "fattn-common.cuh" -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(D, 1) #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) @@ -43,6 +43,7 @@ static __global__ void flash_attn_vec_ext_f16( const float m0, const float m1, const uint32_t n_head_log2, + const float logit_softcap, const int ne00, const int ne01, const int ne02, @@ -67,6 +68,12 @@ static __global__ void flash_attn_vec_ext_f16( const int ne2, const int ne3) { #ifdef FP16_AVAILABLE + // Skip unused kernel variants for faster compilation: + if (use_logit_softcap && !(D == 128 || D == 256)) { + NO_DEVICE_CODE; + return; + } + //In this kernel Q, K, V are matrices while i, j, k are matrix indices. constexpr vec_dot_KQ_f16_t vec_dot_KQ = get_vec_dot_KQ_f16(type_K); @@ -216,6 +223,11 @@ static __global__ void flash_attn_vec_ext_f16( for (int j = 0; j < ncols; ++j) { half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]); sum = warp_reduce_sum(sum); + + if (use_logit_softcap) { + sum = logit_softcap*tanhf(sum); + } + sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f); if (ncols == 1) { @@ -312,10 +324,10 @@ static __global__ void flash_attn_vec_ext_f16( #endif // FP16_AVAILABLE } -template +template void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { constexpr int nwarps = D/WARP_SIZE; - fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16; + fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16; constexpr bool need_f16_K = D != 128; constexpr bool need_f16_V = D != 128 && D != 64; launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V); @@ -323,48 +335,81 @@ void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, template void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - ggml_tensor * KQV = dst; - ggml_tensor * Q = dst->src[0]; - ggml_tensor * K = dst->src[1]; - ggml_tensor * V = dst->src[2]; + const ggml_tensor * KQV = dst; + const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * K = dst->src[1]; + const ggml_tensor * V = dst->src[2]; - const int32_t precision = KQV->op_params[2]; + const int32_t precision = KQV->op_params[3]; GGML_ASSERT(precision == GGML_PREC_DEFAULT); GGML_ASSERT(K->type == type_K); GGML_ASSERT(V->type == type_V); + float logit_softcap; + memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); + if (Q->ne[1] == 1) { constexpr int cols_per_block = 1; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } return; } if (Q->ne[1] == 2) { constexpr int cols_per_block = 2; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } return; } if (Q->ne[1] <= 4) { constexpr int cols_per_block = 4; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } return; } if (Q->ne[1] <= 8) { constexpr int cols_per_block = 8; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } return; } constexpr int cols_per_block = 8; constexpr int parallel_blocks = 1; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + } } #define DECL_FATTN_VEC_F16_CASE(D, type_K, type_V) \ diff --git a/llama/ggml-cuda/fattn-vec-f32.cuh b/llama/ggml-cuda/fattn-vec-f32.cuh index 143d1a28..5f42c314 100644 --- a/llama/ggml-cuda/fattn-vec-f32.cuh +++ b/llama/ggml-cuda/fattn-vec-f32.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -27,7 +27,7 @@ #include "common.cuh" #include "fattn-common.cuh" -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(D, 1) #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) @@ -43,6 +43,7 @@ static __global__ void flash_attn_vec_ext_f32( const float m0, const float m1, const uint32_t n_head_log2, + const float logit_softcap, const int ne00, const int ne01, const int ne02, @@ -66,6 +67,12 @@ static __global__ void flash_attn_vec_ext_f32( const int ne1, const int ne2, const int ne3) { + // Skip unused kernel variants for faster compilation: + if (use_logit_softcap && !(D == 128 || D == 256)) { + NO_DEVICE_CODE; + return; + } + //In this kernel Q, K, V are matrices while i, j, k are matrix indices. constexpr vec_dot_KQ_f32_t vec_dot_KQ = get_vec_dot_KQ_f32(type_K); @@ -206,6 +213,11 @@ static __global__ void flash_attn_vec_ext_f32( for (int j = 0; j < ncols; ++j) { float sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_f2[j], Q_i32[j], Q_ds[j]); sum = warp_reduce_sum(sum); + + if (use_logit_softcap) { + sum = logit_softcap*tanhf(sum); + } + sum += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f; kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum); @@ -293,10 +305,10 @@ static __global__ void flash_attn_vec_ext_f32( } } -template +template void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { constexpr int nwarps = D/WARP_SIZE; - fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32; + fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32; constexpr bool need_f16_K = D != 128; constexpr bool need_f16_V = D != 128 && D != 64; launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V); @@ -304,44 +316,78 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, template void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - ggml_tensor * Q = dst->src[0]; - ggml_tensor * K = dst->src[1]; - ggml_tensor * V = dst->src[2]; + const ggml_tensor * KQV = dst; + const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * K = dst->src[1]; + const ggml_tensor * V = dst->src[2]; GGML_ASSERT(K->type == type_K); GGML_ASSERT(V->type == type_V); + float logit_softcap; + memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); + if (Q->ne[1] == 1) { constexpr int cols_per_block = 1; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } return; } if (Q->ne[1] == 2) { constexpr int cols_per_block = 2; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } return; } if (Q->ne[1] <= 4) { constexpr int cols_per_block = 4; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } return; } if (Q->ne[1] <= 8) { constexpr int cols_per_block = 8; constexpr int parallel_blocks = 4; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } return; } constexpr int cols_per_block = 8; constexpr int parallel_blocks = 1; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } else { + constexpr bool use_logit_softcap = true; + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + } } #define DECL_FATTN_VEC_F32_CASE(D, type_K, type_V) \ diff --git a/llama/ggml-cuda/fattn-wmma-f16.cuh b/llama/ggml-cuda/fattn-wmma-f16.cuh index 1b5701bd..0c34fb93 100644 --- a/llama/ggml-cuda/fattn-wmma-f16.cuh +++ b/llama/ggml-cuda/fattn-wmma-f16.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -32,7 +32,7 @@ #endif // FP16_MMA_AVAILABLE // D == head size, VKQ_stride == num VKQ rows calculated in parallel: -template +template #if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(nwarps*WARP_SIZE, 1) #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) @@ -48,6 +48,7 @@ static __global__ void flash_attn_ext_f16( const float m0, const float m1, const uint32_t n_head_log2, + const float logit_softcap, const int ne00, const int ne01, const int ne02, @@ -72,6 +73,12 @@ static __global__ void flash_attn_ext_f16( const int ne2, const int ne3) { #ifdef FP16_MMA_AVAILABLE + // Skip unused kernel variants for faster compilation: + if (use_logit_softcap && !(D == 128 || D == 256)) { + NO_DEVICE_CODE; + return; + } + //In this kernel Q, K, V are matrices while i, j, k are matrix indices. const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on. @@ -111,6 +118,8 @@ static __global__ void flash_attn_ext_f16( const half slopeh = __float2half(slopef); const half2 slope2 = make_half2(slopef, slopef); + const half2 logit_softcap_2 = make_half2(logit_softcap, logit_softcap); + frag_b Q_b[D/16][ncols/frag_n]; // A single buffer for temporarily holding tiles of KQ and VKQ parts: @@ -220,6 +229,10 @@ static __global__ void flash_attn_ext_f16( const int k = k0 + threadIdx.x; KQ_f_tmp[k0/WARP_SIZE] = KQ_f[j*kqs_padded + k]; + + if (use_logit_softcap) { + KQ_f_tmp[k0/WARP_SIZE] = logit_softcap*tanhf(KQ_f_tmp[k0/WARP_SIZE]); + } } float KQ_max_new = KQ_max_f[j0/nwarps]; @@ -263,6 +276,15 @@ static __global__ void flash_attn_ext_f16( const int k = k0 + threadIdx.x; KQ2_tmp[k0/WARP_SIZE] = KQ2[j*(kqs_padded/2) + k]; + + if (use_logit_softcap) { + // There is no dedicated tangens hyperbolicus function for half2. + KQ2_tmp[k0/WARP_SIZE] = h2exp(KQ2_tmp[k0/WARP_SIZE]*make_half2(2.0f, 2.0f)); + KQ2_tmp[k0/WARP_SIZE] = (KQ2_tmp[k0/WARP_SIZE] - make_half2(1.0f, 1.0f)) + /(KQ2_tmp[k0/WARP_SIZE] + make_half2(1.0f, 1.0f)); + + KQ2_tmp[k0/WARP_SIZE] *= logit_softcap_2; + } } half2 KQ_max_new = KQ_max_h2[j0/nwarps]; @@ -453,7 +475,8 @@ static_assert(get_VKQ_stride( 80, 4, 16) == 16, "Test failed."); template void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * KQV = dst; + const ggml_tensor * Q = dst->src[0]; constexpr int nwarps = 4; @@ -461,20 +484,50 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3]; const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm; + float logit_softcap; + memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); + if (4*blocks_num_pb1 < 2*nsm) { constexpr int parallel_blocks = 4; - fattn_kernel_t fattn_kernel = flash_attn_ext_f16; + fattn_kernel_t fattn_kernel; + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + fattn_kernel = flash_attn_ext_f16< + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + } else { + constexpr bool use_logit_softcap = true; + fattn_kernel = flash_attn_ext_f16< + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + } launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true); return; } if (2*blocks_num_pb1 < 2*nsm) { constexpr int parallel_blocks = 2; - fattn_kernel_t fattn_kernel = flash_attn_ext_f16; + fattn_kernel_t fattn_kernel; + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + fattn_kernel = flash_attn_ext_f16< + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + } else { + constexpr bool use_logit_softcap = true; + fattn_kernel = flash_attn_ext_f16< + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + } launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true); return; } constexpr int parallel_blocks = 1; - fattn_kernel_t fattn_kernel = flash_attn_ext_f16; + fattn_kernel_t fattn_kernel; + if (logit_softcap == 0.0f) { + constexpr bool use_logit_softcap = false; + fattn_kernel = flash_attn_ext_f16< + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + } else { + constexpr bool use_logit_softcap = true; + fattn_kernel = flash_attn_ext_f16< + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + } launch_fattn(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true); } diff --git a/llama/ggml-cuda/fattn.cu b/llama/ggml-cuda/fattn.cu index dc0652d1..5bd553e9 100644 --- a/llama/ggml-cuda/fattn.cu +++ b/llama/ggml-cuda/fattn.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -39,7 +39,7 @@ static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, g const ggml_tensor * KQV = dst; const ggml_tensor * Q = dst->src[0]; - const int32_t precision = KQV->op_params[2]; + const int32_t precision = KQV->op_params[3]; if (precision != GGML_PREC_DEFAULT) { if (Q->ne[1] <= 32 || Q->ne[0] > 128) { @@ -327,7 +327,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst ggml_cuda_set_device(ctx.device); const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc; - const int32_t precision = KQV->op_params[2]; + const int32_t precision = KQV->op_params[3]; // On AMD the tile kernels perform poorly, use the vec kernel instead: if (cc >= CC_OFFSET_AMD) { diff --git a/llama/ggml-cuda/fattn.cuh b/llama/ggml-cuda/fattn.cuh index fab654c7..f502eaaa 100644 --- a/llama/ggml-cuda/fattn.cuh +++ b/llama/ggml-cuda/fattn.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/getrows.cu b/llama/ggml-cuda/getrows.cu index a6c96599..4c627947 100644 --- a/llama/ggml-cuda/getrows.cu +++ b/llama/ggml-cuda/getrows.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/getrows.cuh b/llama/ggml-cuda/getrows.cuh index 3173d5ce..e2bc477e 100644 --- a/llama/ggml-cuda/getrows.cuh +++ b/llama/ggml-cuda/getrows.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/im2col.cu b/llama/ggml-cuda/im2col.cu index 3ee4b485..247cf147 100644 --- a/llama/ggml-cuda/im2col.cu +++ b/llama/ggml-cuda/im2col.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/im2col.cuh b/llama/ggml-cuda/im2col.cuh index 592125e1..df1a8c64 100644 --- a/llama/ggml-cuda/im2col.cuh +++ b/llama/ggml-cuda/im2col.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/mma.cuh b/llama/ggml-cuda/mma.cuh index 552e5082..7f6ea994 100644 --- a/llama/ggml-cuda/mma.cuh +++ b/llama/ggml-cuda/mma.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/mmq.cu b/llama/ggml-cuda/mmq.cu index 2ad38cfd..bdc6a49f 100644 --- a/llama/ggml-cuda/mmq.cu +++ b/llama/ggml-cuda/mmq.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/mmq.cuh b/llama/ggml-cuda/mmq.cuh index fc6ff807..86fa62a1 100644 --- a/llama/ggml-cuda/mmq.cuh +++ b/llama/ggml-cuda/mmq.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/mmvq.cu b/llama/ggml-cuda/mmvq.cu index d6ffc0ec..9c87a602 100644 --- a/llama/ggml-cuda/mmvq.cu +++ b/llama/ggml-cuda/mmvq.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/mmvq.cuh b/llama/ggml-cuda/mmvq.cuh index 5d311870..3da3b904 100644 --- a/llama/ggml-cuda/mmvq.cuh +++ b/llama/ggml-cuda/mmvq.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/norm.cu b/llama/ggml-cuda/norm.cu index dafbaf3f..1f79c462 100644 --- a/llama/ggml-cuda/norm.cu +++ b/llama/ggml-cuda/norm.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/norm.cuh b/llama/ggml-cuda/norm.cuh index 9584aab0..c7975033 100644 --- a/llama/ggml-cuda/norm.cuh +++ b/llama/ggml-cuda/norm.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/pad.cu b/llama/ggml-cuda/pad.cu index 7ddbe97c..dd121289 100644 --- a/llama/ggml-cuda/pad.cu +++ b/llama/ggml-cuda/pad.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/pad.cuh b/llama/ggml-cuda/pad.cuh index 79a760e4..afb93f25 100644 --- a/llama/ggml-cuda/pad.cuh +++ b/llama/ggml-cuda/pad.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/pool2d.cu b/llama/ggml-cuda/pool2d.cu index 6fe947b5..54f6ca5c 100644 --- a/llama/ggml-cuda/pool2d.cu +++ b/llama/ggml-cuda/pool2d.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/pool2d.cuh b/llama/ggml-cuda/pool2d.cuh index b24f68bb..e058cfd0 100644 --- a/llama/ggml-cuda/pool2d.cuh +++ b/llama/ggml-cuda/pool2d.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/quantize.cu b/llama/ggml-cuda/quantize.cu index 3e310e97..fff2d49e 100644 --- a/llama/ggml-cuda/quantize.cu +++ b/llama/ggml-cuda/quantize.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/quantize.cuh b/llama/ggml-cuda/quantize.cuh index 4c046dd8..32296287 100644 --- a/llama/ggml-cuda/quantize.cuh +++ b/llama/ggml-cuda/quantize.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/rope.cu b/llama/ggml-cuda/rope.cu index 11f0d459..4a50372c 100644 --- a/llama/ggml-cuda/rope.cu +++ b/llama/ggml-cuda/rope.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -252,7 +252,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - const bool is_neox = mode & 2; + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; const int32_t * pos = (const int32_t *) src1_d; diff --git a/llama/ggml-cuda/rope.cuh b/llama/ggml-cuda/rope.cuh index 1f3b1c0b..a0cd2f14 100644 --- a/llama/ggml-cuda/rope.cuh +++ b/llama/ggml-cuda/rope.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/scale.cu b/llama/ggml-cuda/scale.cu index 41b72f6e..46498c41 100644 --- a/llama/ggml-cuda/scale.cu +++ b/llama/ggml-cuda/scale.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/scale.cuh b/llama/ggml-cuda/scale.cuh index 508f8a11..7aa65385 100644 --- a/llama/ggml-cuda/scale.cuh +++ b/llama/ggml-cuda/scale.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/softmax.cu b/llama/ggml-cuda/softmax.cu index 5793a3d0..97fcf81b 100644 --- a/llama/ggml-cuda/softmax.cu +++ b/llama/ggml-cuda/softmax.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/softmax.cuh b/llama/ggml-cuda/softmax.cuh index 08f99632..3c408dc6 100644 --- a/llama/ggml-cuda/softmax.cuh +++ b/llama/ggml-cuda/softmax.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/sumrows.cu b/llama/ggml-cuda/sumrows.cu index 4c55bdcb..819103b1 100644 --- a/llama/ggml-cuda/sumrows.cu +++ b/llama/ggml-cuda/sumrows.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -42,7 +42,7 @@ static __global__ void k_sum_rows_f32(const float * x, float * dst, const int nc } } -static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) { +void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) { const dim3 block_dims(WARP_SIZE, 1, 1); const dim3 block_nums(nrows, 1, 1); k_sum_rows_f32<<>>(x, dst, ncols); @@ -58,7 +58,6 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { GGML_ASSERT( dst->type == GGML_TYPE_F32); GGML_ASSERT(ggml_is_contiguous(src0)); - const int64_t ncols = src0->ne[0]; const int64_t nrows = ggml_nrows(src0); diff --git a/llama/ggml-cuda/sumrows.cuh b/llama/ggml-cuda/sumrows.cuh index b3b5cd3f..db47343f 100644 --- a/llama/ggml-cuda/sumrows.cuh +++ b/llama/ggml-cuda/sumrows.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -26,4 +26,6 @@ #include "common.cuh" +void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream); + void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu index 7b9acc47..ac7f1476 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu index 6d2f5b2b..810f9df7 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu index 036e25a4..411d9f8f 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu index cdc846c2..64386f63 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu index 42b9553e..346a5859 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu index 0d79f2ad..22ff28fe 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu index 73e0683c..bf5757a3 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu index e21dc472..50c950d8 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu index 4307b7ef..c1a1d4ff 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu index 8782139f..134a5f5d 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu index c01dc3d5..738fb97f 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu index 2c39ee12..406abd32 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu index ffdb9752..fb3d9019 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu index 48566bc0..0b587c21 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu index adb54260..41287725 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu index 6ba20855..67ef8ae4 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu index e7fa59e8..38d0e15c 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu index e3b85fa4..bf3c378f 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu index 308dbaaa..4b0bf994 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu index 2a804981..347a03d8 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu index ce1a7ab9..b2b6e74f 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu index 42549ea2..8db57b23 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu index a211f509..f46dec7d 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu index 724cf117..f01b195c 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu index 92e6c590..0267b25a 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu index b53d0b40..f96463a2 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu index 0e37890a..7b769933 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu index 725904ab..78b49309 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu index 283daffd..01f7eac0 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu index 0a4b699b..4d357674 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu index 5f00fac6..fd9ff44b 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu index 03e828b4..63ef4c2a 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu index c0ba74e4..625ee2b5 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu index 7c355e7e..b0969686 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu index e589492c..c77b22ba 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu index dc25793c..881243e8 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu index 7c1b8672..803673f9 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu index 5692cf32..062e8e22 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu index 3a6d82a4..9c552c80 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu index 824a9077..ef9a2bc5 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu index 0f8bb517..fc989011 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu index 0677d0a0..42c4755e 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu index ae606891..70d69826 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu index 224235f8..ea8f8aa4 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu index 32a2a026..8f504656 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu index 0aae0de5..ab860347 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu index 4254ae4d..77bff98e 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu index 4c70e71f..235714f0 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu index a90d0f09..6bf113ef 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu index 75a6b6dc..23e0a8c0 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu index 387046e6..4a77ace9 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu index b90165a8..02f804e3 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu index c9ab4302..309a6be6 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu index 05786903..241795bd 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu index 3ae05fe7..50166446 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu index 98738e83..b3affe61 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu index 2a4b87ea..a0e65a0c 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu index 97867cd0..ecaa5d17 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu index 4885065f..d6b289c6 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu index 6be68bd4..36a2a3d4 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu index eccf3e84..96ab4e8c 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu index 3400e197..ae224afe 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu index 60061afb..1ac462ca 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu index 4b1a250c..74d21b53 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu index 03c09c1e..4f146c76 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu index 68d9cb38..6a4170bf 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu index df587ef4..b57023ae 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu index 161acbb9..e84fab88 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu index 5f262c61..3a78463f 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu index 90327e26..75654de9 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu index cb67631e..83fe95c1 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu index e79cac28..13d7e745 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu index 12a2e0ab..59d87603 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu index 3a83cddc..de9f5f40 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu index b902e3dc..4fc0dd24 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu index aaba5d45..84251d1e 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu index 785200ed..36f8c3ac 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu index cbbf0726..384fc92f 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu index 2e902550..fb47bc9f 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu index 3f24e03c..8887b421 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu index 38f42660..19e4e297 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu index 4850c544..e565fdb4 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu index 10a82b72..dcb661ff 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu index ad160596..9a9e6abb 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu index ba59ca88..0e21fd0a 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu index 41fc980a..f699f56b 100644 --- a/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +++ b/llama/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu index 0fdb4990..9b6a210c 100644 --- a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu +++ b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu index b48eb117..74e78f9e 100644 --- a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu +++ b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu index 263d0428..8fc18cef 100644 --- a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu +++ b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu index e2a0252d..d877e1d2 100644 --- a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu +++ b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu index e477e099..987915a4 100644 --- a/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu +++ b/llama/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq1_s.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq1_s.cu index f6be840f..03486aaf 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq1_s.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq1_s.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq2_s.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq2_s.cu index 5e1a54bc..ba63cbac 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq2_s.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq2_s.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu index a2c15684..3c1283c7 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu index c255311c..a2344a5d 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq3_s.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq3_s.cu index 192de6e6..197b619f 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq3_s.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq3_s.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu index bdf93bd6..8edd4ad9 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu index d72ed1e6..a1c3902f 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu b/llama/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu index 479303ce..c421db27 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q2_k.cu b/llama/ggml-cuda/template-instances/mmq-instance-q2_k.cu index fff6937a..726a07a9 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q2_k.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q2_k.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q3_k.cu b/llama/ggml-cuda/template-instances/mmq-instance-q3_k.cu index 7fcee96a..e0781f05 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q3_k.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q3_k.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q4_0.cu b/llama/ggml-cuda/template-instances/mmq-instance-q4_0.cu index 326d23ce..e6d79a89 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q4_0.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q4_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q4_1.cu b/llama/ggml-cuda/template-instances/mmq-instance-q4_1.cu index da40b380..0c60bc56 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q4_1.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q4_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q4_k.cu b/llama/ggml-cuda/template-instances/mmq-instance-q4_k.cu index 41e76496..acd1f1f9 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q4_k.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q4_k.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q5_0.cu b/llama/ggml-cuda/template-instances/mmq-instance-q5_0.cu index dc835624..c301011e 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q5_0.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q5_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q5_1.cu b/llama/ggml-cuda/template-instances/mmq-instance-q5_1.cu index 63622b42..15eb7d24 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q5_1.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q5_1.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q5_k.cu b/llama/ggml-cuda/template-instances/mmq-instance-q5_k.cu index ec34a691..2cc91773 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q5_k.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q5_k.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q6_k.cu b/llama/ggml-cuda/template-instances/mmq-instance-q6_k.cu index baf377ae..ccf7e83c 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q6_k.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q6_k.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/template-instances/mmq-instance-q8_0.cu b/llama/ggml-cuda/template-instances/mmq-instance-q8_0.cu index e5077fc1..276505ba 100644 --- a/llama/ggml-cuda/template-instances/mmq-instance-q8_0.cu +++ b/llama/ggml-cuda/template-instances/mmq-instance-q8_0.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/tsembd.cu b/llama/ggml-cuda/tsembd.cu index b295567e..1b6cfa28 100644 --- a/llama/ggml-cuda/tsembd.cu +++ b/llama/ggml-cuda/tsembd.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/tsembd.cuh b/llama/ggml-cuda/tsembd.cuh index 11334f64..f5eec9b8 100644 --- a/llama/ggml-cuda/tsembd.cuh +++ b/llama/ggml-cuda/tsembd.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/unary.cu b/llama/ggml-cuda/unary.cu index c23e6774..79ba0e57 100644 --- a/llama/ggml-cuda/unary.cu +++ b/llama/ggml-cuda/unary.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -127,6 +127,24 @@ static __global__ void sqrt_f32(const float * x, float * dst, const int k) { dst[i] = sqrtf(x[i]); } +static __global__ void sin_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = sinf(x[i]); +} + +static __global__ void cos_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = cosf(x[i]); +} + static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; gelu_f32<<>>(x, dst, k); @@ -182,6 +200,16 @@ static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_ sqrt_f32<<>>(x, dst, k); } +static void sin_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_SIN_BLOCK_SIZE - 1) / CUDA_SIN_BLOCK_SIZE; + sin_f32<<>>(x, dst, k); +} + +static void cos_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_COS_BLOCK_SIZE - 1) / CUDA_COS_BLOCK_SIZE; + cos_f32<<>>(x, dst, k); +} + void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const float * src0_d = (const float *)src0->data; @@ -338,3 +366,31 @@ void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); } + +void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(ggml_is_contiguous(src0)); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + sin_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} + +void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *)src0->data; + float * dst_d = (float *)dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(ggml_is_contiguous(src0)); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + cos_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +} diff --git a/llama/ggml-cuda/unary.cuh b/llama/ggml-cuda/unary.cuh index ab78ec6a..d60e47e4 100644 --- a/llama/ggml-cuda/unary.cuh +++ b/llama/ggml-cuda/unary.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -35,6 +35,8 @@ #define CUDA_HARDSWISH_BLOCK_SIZE 256 #define CUDA_SQR_BLOCK_SIZE 256 #define CUDA_SQRT_BLOCK_SIZE 256 +#define CUDA_SIN_BLOCK_SIZE 256 +#define CUDA_COS_BLOCK_SIZE 256 void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst); @@ -57,3 +59,7 @@ void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/llama/ggml-cuda/upscale.cu b/llama/ggml-cuda/upscale.cu index 49ebe347..d2f4ce00 100644 --- a/llama/ggml-cuda/upscale.cu +++ b/llama/ggml-cuda/upscale.cu @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/upscale.cuh b/llama/ggml-cuda/upscale.cuh index 73fa0d37..cc4e4aa1 100644 --- a/llama/ggml-cuda/upscale.cuh +++ b/llama/ggml-cuda/upscale.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/vecdotq.cuh b/llama/ggml-cuda/vecdotq.cuh index b0cf4b1a..cc69f02c 100644 --- a/llama/ggml-cuda/vecdotq.cuh +++ b/llama/ggml-cuda/vecdotq.cuh @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/vendors/cuda.h b/llama/ggml-cuda/vendors/cuda.h index 05b8ffef..20b0af80 100644 --- a/llama/ggml-cuda/vendors/cuda.h +++ b/llama/ggml-cuda/vendors/cuda.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/vendors/hip.h b/llama/ggml-cuda/vendors/hip.h index 9f4c538a..f96a6140 100644 --- a/llama/ggml-cuda/vendors/hip.h +++ b/llama/ggml-cuda/vendors/hip.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-cuda/vendors/musa.h b/llama/ggml-cuda/vendors/musa.h index 02f33d27..c0dc04d7 100644 --- a/llama/ggml-cuda/vendors/musa.h +++ b/llama/ggml-cuda/vendors/musa.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-impl.h b/llama/ggml-impl.h index 467c4cc7..2d6cb470 100644 --- a/llama/ggml-impl.h +++ b/llama/ggml-impl.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml-metal.h b/llama/ggml-metal.h index 906ca235..f8a9c3cb 100644 --- a/llama/ggml-metal.h +++ b/llama/ggml-metal.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -76,6 +76,8 @@ GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb); +GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data); + GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void); // helper to check if the device supports a specific family diff --git a/llama/ggml-metal.metal b/llama/ggml-metal.metal index 5e363c19..95f65668 100644 --- a/llama/ggml-metal.metal +++ b/llama/ggml-metal.metal @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -43,7 +43,7 @@ enum ggml_sort_order { GGML_SORT_ORDER_DESC, }; -// general-purpose kernel for addition, multiplication and division of two tensors +// general-purpose kernel for addition, subtraction, multiplication and division of two tensors // pros: works for non-contiguous tensors, supports broadcast across all dims // cons: not very efficient kernel void kernel_add( @@ -96,6 +96,56 @@ kernel void kernel_add( } } +kernel void kernel_sub( + device const char * src0, + device const char * src1, + device char * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant int64_t & ne03, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb03, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant int64_t & ne13, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant uint64_t & nb13, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant int64_t & ne3, + constant uint64_t & nb0, + constant uint64_t & nb1, + constant uint64_t & nb2, + constant uint64_t & nb3, + constant int64_t & offs, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t i03 = tgpig.z; + const int64_t i02 = tgpig.y; + const int64_t i01 = tgpig.x; + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + + device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs; + device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11; + device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + offs; + + for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { + const int i10 = i0 % ne10; + *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) - *((device float *)(src1_ptr + i10*nb10)); + } +} + kernel void kernel_mul( device const char * src0, device const char * src1, @@ -252,6 +302,15 @@ kernel void kernel_add_row( dst[tpig] = src0[tpig] + src1[tpig % nb]; } +kernel void kernel_sub_row( + device const float4 * src0, + device const float4 * src1, + device float4 * dst, + constant uint64_t & nb [[buffer(28)]], + uint tpig[[thread_position_in_grid]]) { + dst[tpig] = src0[tpig] - src1[tpig % nb]; +} + kernel void kernel_mul_row( device const float4 * src0, device const float4 * src1, @@ -384,6 +443,27 @@ kernel void kernel_sqr( dst[tpig] = src0[tpig] * src0[tpig]; } +kernel void kernel_sqrt( + device const float * src0, + device float * dst, + uint tpig[[thread_position_in_grid]]) { + dst[tpig] = sqrt(src0[tpig]); +} + +kernel void kernel_sin( + device const float * src0, + device float * dst, + uint tpig[[thread_position_in_grid]]) { + dst[tpig] = sin(src0[tpig]); +} + +kernel void kernel_cos( + device const float * src0, + device float * dst, + uint tpig[[thread_position_in_grid]]) { + dst[tpig] = cos(src0[tpig]); +} + kernel void kernel_sum_rows( device const float * src0, device float * dst, @@ -693,6 +773,127 @@ kernel void kernel_diag_mask_inf_8( } } +// ref: ggml.c:ggml_compute_forward_ssm_conv_f32 +// TODO: optimize +kernel void kernel_ssm_conv_f32( + device const void * src0, + device const void * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant int64_t & ne0, + constant int64_t & ne1, + constant int64_t & ne2, + constant uint64_t & nb0, + constant uint64_t & nb1, + constant uint64_t & nb2, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t ir = tgpig.x; + const int64_t i2 = tgpig.y; + const int64_t i3 = tgpig.z; + + const int64_t nc = ne10; + const int64_t ncs = ne00; + const int64_t nr = ne01; + const int64_t n_t = ne1; + const int64_t n_s = ne2; + + device const float * s = (device const float *) ((device const char *) src0 + ir*nb01 + i2*nb00 + i3*nb02); + device const float * c = (device const float *) ((device const char *) src1 + ir*nb11); + device float * x = (device float *) ((device char *) dst + ir*nb0 + i2*nb1 + i3*nb2); + + float sumf = 0.0f; + + for (int64_t i0 = 0; i0 < nc; ++i0) { + sumf += s[i0] * c[i0]; + } + + x[0] = sumf; +} + +// ref: ggml.c:ggml_compute_forward_ssm_scan_f32 +// TODO: optimize +kernel void kernel_ssm_scan_f32( + device const void * src0, + device const void * src1, + device const void * src2, + device const void * src3, + device const void * src4, + device const void * src5, + device float * dst, + constant int64_t & d_state, + constant int64_t & d_inner, + constant int64_t & n_seq_tokens, + constant int64_t & n_seqs, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant uint64_t & nb13, + constant uint64_t & nb20, + constant uint64_t & nb21, + constant uint64_t & nb22, + constant uint64_t & nb30, + constant uint64_t & nb31, + constant uint64_t & nb40, + constant uint64_t & nb41, + constant uint64_t & nb42, + constant uint64_t & nb50, + constant uint64_t & nb51, + constant uint64_t & nb52, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int64_t ir = tgpig.x; + const int64_t i3 = tgpig.y; + + const int64_t nc = d_state; + const int64_t nr = d_inner; + const int64_t n_t = n_seq_tokens; + const int64_t n_s = n_seqs; + + for (int64_t i2 = 0; i2 < n_t; ++i2) { + device const float * s0 = (device const float *) ((device const char *) src0 + ir*nb01 + i3*nb02); + device const float * x = (device const float *) ((device const char *) src1 + ir*nb10 + i2*nb11 + i3*nb12); + device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*nb21 + i3*nb22); + device const float * A = (device const float *) ((device const char *) src3 + ir*nb31); + device const float * B = (device const float *) ((device const char *) src4 + i2*nb41 + i3*nb42); + device const float * C = (device const float *) ((device const char *) src5 + i2*nb51 + i3*nb52); + device float * y = (device float *) ((device char *) dst + ir*nb10 + i2*nb11 + i3*nb12); // TODO: do not use src1 strides + device float * s = (device float *) ((device char *) dst + ir*nb01 + i3*nb02 + nb13); + + if (i2 > 0) { + s0 = s; + } + + // i1 == 0 + float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0]; + float x_dt = x[0] * dt_soft_plus; + float sumf = 0.0f; + + for (int64_t i0 = 0; i0 < nc; ++i0) { + int64_t i = i0; + float state = (s0[i] * exp(dt_soft_plus * A[i])) + (B[i0] * x_dt); + sumf += state * C[i0]; + s[i] = state; + } + + y[0] = sumf; + } +} + kernel void kernel_norm( device const void * src0, device float * dst, @@ -2002,6 +2203,7 @@ typedef void (flash_attn_ext_f16_t)( constant float & m0, constant float & m1, constant uint32_t & n_head_log2, + constant float & logit_softcap, threadgroup half * shared, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2040,6 +2242,7 @@ kernel void kernel_flash_attn_ext_f16( constant float & m0, constant float & m1, constant uint32_t & n_head_log2, + constant float & logit_softcap, threadgroup half * shared [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2164,19 +2367,6 @@ kernel void kernel_flash_attn_ext_f16( } simdgroup_store(mqk, ss + 8*cc, TF, 0, false); - - const short tx = tiisg%4; - const short ty = tiisg/4; - - if (mask != q) { - // mqk = mqk*scale + mask*slope - ss[8*cc + ty*TF + 2*tx + 0] = scale*ss[8*cc + ty*TF + 2*tx + 0] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 0]; - ss[8*cc + ty*TF + 2*tx + 1] = scale*ss[8*cc + ty*TF + 2*tx + 1] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 1]; - } else { - // mqk = mqk*scale - ss[8*cc + ty*TF + 2*tx + 0] *= scale; - ss[8*cc + ty*TF + 2*tx + 1] *= scale; - } } } @@ -2188,10 +2378,19 @@ kernel void kernel_flash_attn_ext_f16( float ms[Q]; for (short j = 0; j < Q; ++j) { - const short p = tiisg; - const float m = M[j]; - const float s = ss[j*TF + p]; + + // scale and apply the logitcap / mask + float s = ss[j*TF + tiisg]*scale; + + if (logit_softcap != 0.0f) { + s = logit_softcap*precise::tanh(s); + } + + if (mask != q) { + // mqk = mqk + mask*slope + s += slope*mp[ic + j*nb31/sizeof(half) + tiisg]; + } smax = simd_max(max(smax, s)); M[j] = simd_max(max(M[j], s)); @@ -2202,7 +2401,7 @@ kernel void kernel_flash_attn_ext_f16( S[j] = S[j]*ms[j] + simd_sum(vs); // the P matrix from the paper (Q rows, C columns) - ss[j*TF + p] = vs; + ss[j*TF + tiisg] = vs; } // create a QxQ diagonal matrix for rescaling the output @@ -2371,6 +2570,7 @@ kernel void kernel_flash_attn_ext_vec_f16( constant float & m0, constant float & m1, constant uint32_t & n_head_log2, + constant float & logit_softcap, threadgroup half * shared [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2505,7 +2705,13 @@ kernel void kernel_flash_attn_ext_vec_f16( // mqk = mqk*scale + mask*slope if (tiisg == 0) { - mqk = mqk*scale + ((mask != q) ? ((float4) mp4[ic/4 + cc])*slope : (float4) 0.0f); + mqk *= scale; + + if (logit_softcap != 0.0f) { + mqk = logit_softcap*precise::tanh(mqk); + } + + mqk += (mask != q) ? ((float4) mp4[ic/4 + cc])*slope : (float4) 0.0f; ss4[cc] = mqk; } diff --git a/llama/ggml-metal_darwin_arm64.m b/llama/ggml-metal_darwin_arm64.m index 8d705b9c..952646bd 100644 --- a/llama/ggml-metal_darwin_arm64.m +++ b/llama/ggml-metal_darwin_arm64.m @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -57,6 +57,8 @@ struct ggml_metal_kernel { enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_ADD, GGML_METAL_KERNEL_TYPE_ADD_ROW, + GGML_METAL_KERNEL_TYPE_SUB, + GGML_METAL_KERNEL_TYPE_SUB_ROW, GGML_METAL_KERNEL_TYPE_MUL, GGML_METAL_KERNEL_TYPE_MUL_ROW, GGML_METAL_KERNEL_TYPE_DIV, @@ -108,6 +110,8 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_RMS_NORM, GGML_METAL_KERNEL_TYPE_GROUP_NORM, GGML_METAL_KERNEL_TYPE_NORM, + GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, + GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, @@ -231,12 +235,15 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, GGML_METAL_KERNEL_TYPE_CONCAT, GGML_METAL_KERNEL_TYPE_SQR, + GGML_METAL_KERNEL_TYPE_SQRT, + GGML_METAL_KERNEL_TYPE_SIN, + GGML_METAL_KERNEL_TYPE_COS, GGML_METAL_KERNEL_TYPE_SUM_ROWS, GGML_METAL_KERNEL_TYPE_COUNT }; -struct ggml_metal_context { +struct ggml_backend_metal_context { int n_cb; id device; @@ -250,6 +257,10 @@ struct ggml_metal_context { bool support_simdgroup_mm; bool should_capture_next_compute; + + // abort ggml_metal_graph_compute if callback returns true + ggml_abort_callback abort_callback; + void * abort_callback_data; }; // MSL code @@ -315,7 +326,7 @@ static void * ggml_metal_host_malloc(size_t n) { return data; } -static struct ggml_metal_context * ggml_metal_init(int n_cb) { +static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_LOG_INFO("%s: allocating\n", __func__); #if TARGET_OS_OSX && !GGML_METAL_NDEBUG @@ -332,7 +343,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); // Configure context - struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context)); + struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context)); ctx->device = device; ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); ctx->queue = [ctx->device newCommandQueue]; @@ -513,6 +524,8 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB, sub, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW, sub_row, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); @@ -564,6 +577,8 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, ssm_conv_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, ssm_scan_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction); @@ -687,6 +702,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT, sqrt, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN, sin, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS, cos, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); } @@ -694,7 +712,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { return ctx; } -static void ggml_metal_free(struct ggml_metal_context * ctx) { +static void ggml_metal_free(struct ggml_backend_metal_context * ctx) { GGML_METAL_LOG_INFO("%s: deallocating\n", __func__); for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) { @@ -760,7 +778,7 @@ static id ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs return nil; } -static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) { +static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) { for (size_t i = 0, n = 3; i < n; ++i) { if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) { return false; @@ -787,15 +805,20 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const case GGML_OP_PERMUTE: case GGML_OP_CONCAT: case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_ACC: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_REPEAT: case GGML_OP_SCALE: case GGML_OP_CLAMP: - case GGML_OP_SQR: - case GGML_OP_SUM_ROWS: return true; + case GGML_OP_SQR: + case GGML_OP_SQRT: + case GGML_OP_SIN: + case GGML_OP_COS: + return ggml_is_contiguous(op->src[0]); + case GGML_OP_SUM_ROWS: case GGML_OP_SOFT_MAX: case GGML_OP_RMS_NORM: case GGML_OP_GROUP_NORM: @@ -825,6 +848,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const return false; } return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels + case GGML_OP_SSM_CONV: + case GGML_OP_SSM_SCAN: + return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: return ctx->support_simdgroup_reduction && @@ -871,7 +897,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const } static enum ggml_status ggml_metal_graph_compute( - struct ggml_metal_context * ctx, + struct ggml_backend_metal_context * ctx, struct ggml_cgraph * gf) { @autoreleasepool { @@ -904,8 +930,11 @@ static enum ggml_status ggml_metal_graph_compute( id command_buffer = [ctx->queue commandBufferWithUnretainedReferences]; command_buffer_builder[cb_idx] = command_buffer; - // enqueue the command buffers in order to specify their execution order - [command_buffer enqueue]; + // always enqueue the first two command buffers + // enqueue all of the command buffers if we don't need to abort + if (cb_idx < 2 || ctx->abort_callback == NULL) { + [command_buffer enqueue]; + } } const id *command_buffers = command_buffer_builder; @@ -1069,6 +1098,7 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: { @@ -1092,6 +1122,7 @@ static enum ggml_status ggml_metal_graph_compute( nb = ne00 / 4; switch (dst->op) { case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break; + case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break; case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break; case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break; default: GGML_ABORT("fatal error"); @@ -1101,6 +1132,7 @@ static enum ggml_status ggml_metal_graph_compute( } else { switch (dst->op) { case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break; + case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB].pipeline; break; case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break; case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break; default: GGML_ABORT("fatal error"); @@ -1428,6 +1460,48 @@ static enum ggml_status ggml_metal_graph_compute( const int64_t n = ggml_nelements(dst); + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SQRT: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQRT].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SIN: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIN].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_COS: + { + GGML_ASSERT(ggml_is_contiguous(src0)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_COS].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + + const int64_t n = ggml_nelements(dst); + [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; case GGML_OP_SUM_ROWS: @@ -1557,6 +1631,121 @@ static enum ggml_status ggml_metal_graph_compute( [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } } break; + case GGML_OP_SSM_CONV: + { + GGML_ASSERT(src0t == GGML_TYPE_F32); + GGML_ASSERT(src1t == GGML_TYPE_F32); + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; + [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; + [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; + [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9]; + [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:11]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:12]; + [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13]; + [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14]; + [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:15]; + [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:16]; + [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:17]; + [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:18]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_SSM_SCAN: + { + struct ggml_tensor * src3 = gf->nodes[i]->src[3]; + struct ggml_tensor * src4 = gf->nodes[i]->src[4]; + struct ggml_tensor * src5 = gf->nodes[i]->src[5]; + + GGML_ASSERT(src3); + GGML_ASSERT(src4); + GGML_ASSERT(src5); + + size_t offs_src3 = 0; + size_t offs_src4 = 0; + size_t offs_src5 = 0; + + id id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil; + id id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil; + id id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil; + + const int64_t ne30 = src3->ne[0]; GGML_UNUSED(ne30); + const int64_t ne31 = src3->ne[1]; GGML_UNUSED(ne31); + + const uint64_t nb30 = src3->nb[0]; + const uint64_t nb31 = src3->nb[1]; + + const int64_t ne40 = src4->ne[0]; GGML_UNUSED(ne40); + const int64_t ne41 = src4->ne[1]; GGML_UNUSED(ne41); + const int64_t ne42 = src4->ne[2]; GGML_UNUSED(ne42); + + const uint64_t nb40 = src4->nb[0]; + const uint64_t nb41 = src4->nb[1]; + const uint64_t nb42 = src4->nb[2]; + + const int64_t ne50 = src5->ne[0]; GGML_UNUSED(ne50); + const int64_t ne51 = src5->ne[1]; GGML_UNUSED(ne51); + const int64_t ne52 = src5->ne[2]; GGML_UNUSED(ne52); + + const uint64_t nb50 = src5->nb[0]; + const uint64_t nb51 = src5->nb[1]; + const uint64_t nb52 = src5->nb[2]; + + const int64_t d_state = ne00; + const int64_t d_inner = ne01; + const int64_t n_seq_tokens = ne11; + const int64_t n_seqs = ne02; + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; + [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4]; + [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:6]; + + [encoder setBytes:&d_state length:sizeof(d_state) atIndex:7]; + [encoder setBytes:&d_inner length:sizeof(d_inner) atIndex:8]; + [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9]; + [encoder setBytes:&n_seqs length:sizeof(n_seqs) atIndex:10]; + + [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11]; + [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12]; + [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13]; + [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; + [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; + [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; + [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; + [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18]; + [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19]; + [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20]; + [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21]; + [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22]; + [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23]; + [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24]; + [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25]; + [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26]; + [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27]; + [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28]; + + [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; case GGML_OP_MUL_MAT: { GGML_ASSERT(ne00 == ne10); @@ -2328,7 +2517,7 @@ static enum ggml_status ggml_metal_graph_compute( memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - const bool is_neox = mode & 2; + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; id pipeline = nil; @@ -2639,9 +2828,14 @@ static enum ggml_status ggml_metal_graph_compute( float scale; float max_bias; + float logit_softcap; + memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale)); + memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias)); + memcpy(&logit_softcap, ((int32_t *) dst->op_params) + 2, sizeof(logit_softcap)); - memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale)); - memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias)); + if (logit_softcap != 0.0f) { + scale /= logit_softcap; + } const uint32_t n_head = src0->ne[2]; const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head)); @@ -2692,30 +2886,31 @@ static enum ggml_status ggml_metal_graph_compute( } else { [encoder setBuffer:id_src0 offset:offs_src0 atIndex:3]; } - [encoder setBuffer:id_dst offset:offs_dst atIndex:4]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6]; - [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10]; - [encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:11]; - [encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:12]; - [encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:13]; - [encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:14]; - [encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:15]; - [encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:16]; - [encoder setBytes:&nb21 length:sizeof(uint64_t) atIndex:17]; - [encoder setBytes:&nb22 length:sizeof(uint64_t) atIndex:18]; - [encoder setBytes:&nb23 length:sizeof(uint64_t) atIndex:19]; - [encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:20]; - [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:21]; - [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:22]; - [encoder setBytes:&scale length:sizeof( float) atIndex:23]; - [encoder setBytes:&max_bias length:sizeof( float) atIndex:24]; - [encoder setBytes:&m0 length:sizeof(m0) atIndex:25]; - [encoder setBytes:&m1 length:sizeof(m1) atIndex:26]; - [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:4]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:11]; + [encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:12]; + [encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:14]; + [encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:15]; + [encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:16]; + [encoder setBytes:&nb21 length:sizeof(uint64_t) atIndex:17]; + [encoder setBytes:&nb22 length:sizeof(uint64_t) atIndex:18]; + [encoder setBytes:&nb23 length:sizeof(uint64_t) atIndex:19]; + [encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:20]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:21]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:22]; + [encoder setBytes:&scale length:sizeof( float) atIndex:23]; + [encoder setBytes:&max_bias length:sizeof( float) atIndex:24]; + [encoder setBytes:&m0 length:sizeof(m0) atIndex:25]; + [encoder setBytes:&m1 length:sizeof(m1) atIndex:26]; + [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27]; + [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28]; if (!use_vec_kernel) { // half8x8 kernel @@ -2849,7 +3044,9 @@ static enum ggml_status ggml_metal_graph_compute( [encoder endEncoding]; - [command_buffer commit]; + if (cb_idx < 2 || ctx->abort_callback == NULL) { + [command_buffer commit]; + } }); // Wait for completion and check status of each command buffer @@ -2869,6 +3066,23 @@ static enum ggml_status ggml_metal_graph_compute( return GGML_STATUS_FAILED; } + + id next_buffer = (i + 1 < n_cb ? command_buffers[i + 1] : nil); + if (!next_buffer) { + continue; + } + + bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued); + if (next_queued) { + continue; + } + + if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) { + GGML_METAL_LOG_INFO("%s: command buffer %d aborted", __func__, i); + return GGML_STATUS_ABORTED; + } + + [next_buffer commit]; } if (should_capture) { @@ -3172,7 +3386,7 @@ GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) { } GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) { - struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context; + struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; ggml_metal_free(ctx); free(backend); } @@ -3184,13 +3398,13 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe } GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { - struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context; + struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context; return ggml_metal_graph_compute(metal_ctx, cgraph); } GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { - struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context; + struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context; return ggml_metal_supports_op(metal_ctx, op); } @@ -3235,9 +3449,9 @@ static ggml_guid_t ggml_backend_metal_guid(void) { } ggml_backend_t ggml_backend_metal_init(void) { - struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS); - + struct ggml_backend_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS); if (ctx == NULL) { + GGML_METAL_LOG_ERROR("%s: error: failed to allocate context\n", __func__); return NULL; } @@ -3259,15 +3473,24 @@ bool ggml_backend_is_metal(ggml_backend_t backend) { void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) { GGML_ASSERT(ggml_backend_is_metal(backend)); - struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context; + struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); } +void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) { + GGML_ASSERT(ggml_backend_is_metal(backend)); + + struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; + + ctx->abort_callback = abort_callback; + ctx->abort_callback_data = user_data; +} + bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) { GGML_ASSERT(ggml_backend_is_metal(backend)); - struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context; + struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)]; } @@ -3275,7 +3498,7 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) { void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) { GGML_ASSERT(ggml_backend_is_metal(backend)); - struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context; + struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context; ctx->should_capture_next_compute = true; } diff --git a/llama/ggml-quants.c b/llama/ggml-quants.c index 93552afd..c22ffac0 100644 --- a/llama/ggml-quants.c +++ b/llama/ggml-quants.c @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -3670,7 +3670,7 @@ void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) { quantize_row_q8_K_ref(x, y, k); } -//===================================== Dot ptoducts ================================= +//===================================== Dot products ================================= // // Helper functions diff --git a/llama/ggml-quants.h b/llama/ggml-quants.h index 26d494e9..fa25ebab 100644 --- a/llama/ggml-quants.h +++ b/llama/ggml-quants.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/ggml.c b/llama/ggml.c index d4f84e51..0a64e875 100644 --- a/llama/ggml.c +++ b/llama/ggml.c @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -82,6 +82,9 @@ int ggml_sve_cnt_b = 0; // disable POSIX deprecation warnings // these functions are never going away, anyway #pragma warning(disable: 4996) + +// unreachable code because of multiple instances of code after GGML_ABORT +#pragma warning(disable: 4702) #endif #if defined(_WIN32) @@ -92,23 +95,42 @@ int ggml_sve_cnt_b = 0; #endif #include +#if !defined(__clang__) typedef volatile LONG atomic_int; typedef atomic_int atomic_bool; typedef atomic_int atomic_flag; #define ATOMIC_FLAG_INIT 0 +typedef enum { + memory_order_relaxed, + memory_order_consume, + memory_order_acquire, + memory_order_release, + memory_order_acq_rel, + memory_order_seq_cst +} memory_order; + static void atomic_store(atomic_int * ptr, LONG val) { InterlockedExchange(ptr, val); } +static void atomic_store_explicit(atomic_int * ptr, LONG val, memory_order mo) { + // TODO: add support for explicit memory order + InterlockedExchange(ptr, val); +} static LONG atomic_load(atomic_int * ptr) { return InterlockedCompareExchange(ptr, 0, 0); } +static LONG atomic_load_explicit(atomic_int * ptr, memory_order mo) { + // TODO: add support for explicit memory order + return InterlockedCompareExchange(ptr, 0, 0); +} static LONG atomic_fetch_add(atomic_int * ptr, LONG inc) { return InterlockedExchangeAdd(ptr, inc); } -static LONG atomic_fetch_sub(atomic_int * ptr, LONG dec) { - return atomic_fetch_add(ptr, -(dec)); +static LONG atomic_fetch_add_explicit(atomic_int * ptr, LONG inc, memory_order mo) { + // TODO: add support for explicit memory order + return InterlockedExchangeAdd(ptr, inc); } static atomic_bool atomic_flag_test_and_set(atomic_flag * ptr) { return InterlockedExchange(ptr, 1); @@ -116,6 +138,9 @@ static atomic_bool atomic_flag_test_and_set(atomic_flag * ptr) { static void atomic_flag_clear(atomic_flag * ptr) { InterlockedExchange(ptr, 0); } +#else // clang +#include +#endif typedef HANDLE pthread_t; @@ -144,8 +169,13 @@ static int sched_yield (void) { return 0; } #else + #include #include +#include +#if defined(__FreeBSD__) +#include +#endif typedef void * thread_ret_t; @@ -1891,28 +1921,102 @@ struct ggml_context_container { struct ggml_context context; }; -struct ggml_compute_state_shared { - const struct ggml_cgraph * cgraph; - const struct ggml_cplan * cplan; +// +// Threading defs +// - int n_threads; +typedef pthread_t ggml_thread_t; + +#if defined(_WIN32) + +typedef CONDITION_VARIABLE ggml_cond_t; +typedef SRWLOCK ggml_mutex_t; + +#define ggml_mutex_init(m) InitializeSRWLock(m) +#define ggml_mutex_destroy(m) +#define ggml_mutex_lock(m) AcquireSRWLockExclusive(m) +#define ggml_mutex_unlock(m) ReleaseSRWLockExclusive(m) +#define ggml_mutex_lock_shared(m) AcquireSRWLockShared(m) +#define ggml_mutex_unlock_shared(m) ReleaseSRWLockShared(m) + +#define ggml_cond_init(c) InitializeConditionVariable(c) +#define ggml_cond_destroy(c) +#define ggml_cond_wait(c, m) SleepConditionVariableSRW(c, m, INFINITE, CONDITION_VARIABLE_LOCKMODE_SHARED) +#define ggml_cond_broadcast(c) WakeAllConditionVariable(c) + +#define ggml_thread_create pthread_create +#define ggml_thread_join pthread_join + +#else + +typedef pthread_cond_t ggml_cond_t; +typedef pthread_mutex_t ggml_mutex_t; + +#define ggml_mutex_init(m) pthread_mutex_init(m, NULL) +#define ggml_mutex_destroy(m) pthread_mutex_destroy(m) +#define ggml_mutex_lock(m) pthread_mutex_lock(m) +#define ggml_mutex_unlock(m) pthread_mutex_unlock(m) +#define ggml_mutex_lock_shared(m) pthread_mutex_lock(m) +#define ggml_mutex_unlock_shared(m) pthread_mutex_unlock(m) + +#define ggml_lock_init(x) UNUSED(x) +#define ggml_lock_destroy(x) UNUSED(x) +#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64)) +#define ggml_lock_lock(x) _mm_pause() +#else +#define ggml_lock_lock(x) UNUSED(x) +#endif +#define ggml_lock_unlock(x) UNUSED(x) + +#define GGML_LOCK_INITIALIZER 0 +#define ggml_cond_init(c) pthread_cond_init(c, NULL) +#define ggml_cond_destroy(c) pthread_cond_destroy(c) +#define ggml_cond_wait(c, m) pthread_cond_wait(c, m) +#define ggml_cond_broadcast(c) pthread_cond_broadcast(c) + +#define ggml_thread_create pthread_create +#define ggml_thread_join pthread_join + +#endif + +// Threadpool def +struct ggml_threadpool { + ggml_mutex_t mutex; // mutex for cond.var + ggml_cond_t cond; // cond.var for waiting for new work + + struct ggml_cgraph * cgraph; + struct ggml_cplan * cplan; // synchronization primitives + atomic_int n_graph; // incremented when there is work to be done (i.e each graph) atomic_int n_barrier; atomic_int n_barrier_passed; + atomic_int current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads. - ggml_abort_callback abort_callback; // abort ggml_graph_compute when true - void * abort_callback_data; + // these are atomic as an annotation for thread-sanitizer + atomic_bool stop; // Used for stopping the threadpool altogether + atomic_bool pause; // Used for pausing the threadpool or individual threads - atomic_int current_chunk; // currently processing chunk during mul_mat, shared between all the threads + struct ggml_compute_state * workers; // per thread state + int n_threads_max; // number of threads in the pool + int n_threads_cur; // number of threads used in the current graph + + int32_t prio; // Scheduling priority + uint32_t poll; // Polling level (0 - no polling) enum ggml_status ec; }; +// Per-thread state struct ggml_compute_state { +#ifndef GGML_USE_OPENMP ggml_thread_t thrd; + bool cpumask[GGML_MAX_N_THREADS]; + int last_graph; + bool pending; +#endif + struct ggml_threadpool * threadpool; int ith; - struct ggml_compute_state_shared * shared; }; struct ggml_compute_params { @@ -1923,7 +2027,7 @@ struct ggml_compute_params { size_t wsize; void * wdata; - struct ggml_compute_state_shared * shared; + struct ggml_threadpool * threadpool; }; // @@ -2333,7 +2437,9 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s); } inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } -inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } +inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } +inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); } +inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); } inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); } inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); } inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } @@ -2345,6 +2451,7 @@ inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x // TODO: optimize performance inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } +inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); } static const float GELU_COEF_A = 0.044715f; static const float GELU_QUICK_COEF = -1.702f; @@ -2692,6 +2799,19 @@ static ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, return sum; } +static ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) { + // log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i) + + int i = 0; + ggml_float sum = 0; + for (; i < n; ++i) { + float val = x[i] - max; + y[i] = val; + sum += (ggml_float)expf(val); + } + return sum = (ggml_float)logf(sum); +} + inline static float ggml_silu_backward_f32(float x, float dy) { const float s = 1.0f/(1.0f + expf(-x)); return dy*s*(1.0f + x*(1.0f - s)); @@ -2783,6 +2903,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "SQR", "SQRT", "LOG", + "SIN", + "COS", "SUM", "SUM_ROWS", "MEAN", @@ -2820,9 +2942,11 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CLAMP", "CONV_TRANSPOSE_1D", "IM2COL", + "IM2COL_BACK", "CONV_TRANSPOSE_2D", "POOL_1D", "POOL_2D", + "POOL_2D_BACK", "UPSCALE", "PAD", "ARANGE", @@ -2838,6 +2962,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "WIN_UNPART", "GET_REL_POS", "ADD_REL_POS", + "RWKV_WKV", "UNARY", @@ -2856,7 +2981,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74"); +static_assert(GGML_OP_COUNT == 79, "GGML_OP_COUNT != 79"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -2871,6 +2996,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "x^2", "√x", "log(x)", + "sin(x)", + "cos(x)", "Σx", "Σx_k", "Σx/n", @@ -2908,9 +3035,11 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "clamp(x)", "conv_transpose_1d(x)", "im2col(x)", + "im2col_back(x)", "conv_transpose_2d(x)", "pool_1d(x)", "pool_2d(x)", + "pool_2d_back(x)", "upscale(x)", "pad(x)", "arange(start, stop, step)", @@ -2926,6 +3055,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "win_unpart(x)", "get_rel_pos(x)", "add_rel_pos(x)", + "rwkv_wkv(k, v, r, tf, td, s)", "unary(x)", @@ -2944,7 +3074,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74"); +static_assert(GGML_OP_COUNT == 79, "GGML_OP_COUNT != 79"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -2963,14 +3093,28 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = { "SILU", "HARDSWISH", "HARDSIGMOID", + "EXP", }; -static_assert(GGML_UNARY_OP_COUNT == 13, "GGML_UNARY_OP_COUNT != 13"); +static_assert(GGML_UNARY_OP_COUNT == 14, "GGML_UNARY_OP_COUNT != 14"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN"); +// Helpers for polling loops +#if defined(__aarch64__) && ( defined(__clang__) || defined(__GNUC__) ) +static inline void ggml_thread_cpu_relax(void) { + __asm__ volatile("yield" ::: "memory"); +} +#elif defined(__x86_64__) +static inline void ggml_thread_cpu_relax(void) { + _mm_pause(); +} +#else +static inline void ggml_thread_cpu_relax(void) {;} +#endif + // // NUMA support // @@ -3018,42 +3162,36 @@ inline static void ggml_critical_section_start(void) { } #ifdef GGML_USE_OPENMP -static void ggml_barrier(struct ggml_compute_state_shared * shared) { - if (shared->n_threads == 1) { +static void ggml_barrier(struct ggml_threadpool * threadpool) { + if (threadpool->n_threads_cur == 1) { return; } #pragma omp barrier } #else -static void ggml_barrier(struct ggml_compute_state_shared * shared) { - if (shared->n_threads == 1) { +static void ggml_barrier(struct ggml_threadpool * threadpool) { + if (threadpool->n_threads_cur == 1) { return; } - atomic_int * n_barrier = &shared->n_barrier; - atomic_int * n_barrier_passed = &shared->n_barrier_passed; + atomic_int * n_barrier = &threadpool->n_barrier; + atomic_int * n_barrier_passed = &threadpool->n_barrier_passed; - int n_threads = shared->n_threads; - int passed_old = atomic_load(n_barrier_passed); + int n_threads = threadpool->n_threads_cur; + int passed_old = atomic_load_explicit(n_barrier_passed, memory_order_relaxed); if (atomic_fetch_add(n_barrier, 1) == n_threads - 1) { // last thread atomic_store(n_barrier, 0); - atomic_fetch_add(n_barrier_passed, 1); + atomic_fetch_add_explicit(n_barrier_passed, 1, memory_order_relaxed); } else { // wait for other threads - const int n_spin_before_sleep = 100000; while (true) { - for (int i = 0; i < n_spin_before_sleep; i++) { - if (atomic_load(n_barrier_passed) != passed_old) { - return; - } - #if defined(__SSE3__) - _mm_pause(); - #endif + if (atomic_load_explicit(n_barrier_passed, memory_order_relaxed) != passed_old) { + return; } - sched_yield(); + ggml_thread_cpu_relax(); } } } @@ -3747,7 +3885,8 @@ static struct ggml_tensor * ggml_new_tensor_impl( struct ggml_tensor * view_src, size_t view_offs) { - assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS); + GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT); + GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS); // find the base tensor and absolute offset if (view_src != NULL && view_src->view_src != NULL) { @@ -3789,6 +3928,7 @@ static struct ggml_tensor * ggml_new_tensor_impl( } struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TYPE_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size); + GGML_ASSERT(obj_new); // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here @@ -4508,8 +4648,6 @@ static struct ggml_tensor * ggml_add_impl( bool is_node = false; if (!inplace && (a->grad || b->grad)) { - // TODO: support backward pass for broadcasting - GGML_ASSERT(ggml_are_same_shape(a, b)); is_node = true; } @@ -4683,11 +4821,13 @@ static struct ggml_tensor * ggml_sub_impl( struct ggml_tensor * a, struct ggml_tensor * b, bool inplace) { - GGML_ASSERT(ggml_are_same_shape(a, b)); + GGML_ASSERT(ggml_can_repeat(b, a)); bool is_node = false; if (!inplace && (a->grad || b->grad)) { + // TODO: support backward pass for broadcasting + GGML_ASSERT(ggml_are_same_shape(a, b)); is_node = true; } @@ -4902,6 +5042,72 @@ struct ggml_tensor * ggml_log_inplace( return ggml_log_impl(ctx, a, true); } +// ggml_sin + +static struct ggml_tensor * ggml_sin_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + bool inplace) { + bool is_node = false; + + if (!inplace && (a->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_SIN; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + + return result; +} + +struct ggml_tensor * ggml_sin( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_sin_impl(ctx, a, false); +} + +struct ggml_tensor * ggml_sin_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_sin_impl(ctx, a, true); +} + +// ggml_cos + +static struct ggml_tensor * ggml_cos_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + bool inplace) { + bool is_node = false; + + if (!inplace && (a->grad)) { + is_node = true; + } + + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + result->op = GGML_OP_COS; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + + return result; +} + +struct ggml_tensor * ggml_cos( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_cos_impl(ctx, a, false); +} + +struct ggml_tensor * ggml_cos_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_cos_impl(ctx, a, true); +} + // ggml_sum struct ggml_tensor * ggml_sum( @@ -5291,6 +5497,19 @@ struct ggml_tensor * ggml_hardsigmoid( return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSIGMOID); } +// ggml exp +struct ggml_tensor * ggml_exp( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary(ctx, a, GGML_UNARY_OP_EXP); +} + +struct ggml_tensor * ggml_exp_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXP); +} + // ggml_norm static struct ggml_tensor * ggml_norm_impl( @@ -6749,17 +6968,20 @@ struct ggml_tensor * ggml_im2col( GGML_ASSERT(a->ne[2] == b->ne[2]); } else { GGML_ASSERT(a->ne[1] == b->ne[1]); + GGML_ASSERT(b->ne[3] == 1); } bool is_node = false; - if (a->grad || b->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward + if (/*a->grad ||*/ b->grad) { // a is only used for its shape, not its data is_node = true; } const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0; const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); + GGML_ASSERT((!is_2D || OH > 0) && "b too small compared to a"); + GGML_ASSERT((OW > 0) && "b too small compared to a"); + const int64_t ne[4] = { is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0], OW, @@ -6779,6 +7001,37 @@ struct ggml_tensor * ggml_im2col( return result; } +struct ggml_tensor * ggml_im2col_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int64_t * ne, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1, + bool is_2D) { + + bool is_node = false; + + if (/*a->grad ||*/ b->grad) { // a is only used for its shape, not its data + is_node = true; + } + + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) }; + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_IM2COL_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + result->src[1] = b; + + return result; +} + // a: [OC，IC, KH, KW] // b: [N, IC, IH, IW] // result: [N, OC, OH, OW] @@ -6792,7 +7045,7 @@ struct ggml_tensor * ggml_conv_2d( int p1, int d0, int d1) { - struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, GGML_TYPE_F16); // [N, OH, OW, IC * KH * KW] + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, a->type); // [N, OH, OW, IC * KH * KW] struct ggml_tensor * result = ggml_mul_mat(ctx, @@ -6918,17 +7171,17 @@ struct ggml_tensor * ggml_pool_2d( bool is_node = false; if (a->grad) { - GGML_ABORT("fatal error"); // TODO: implement backward is_node = true; } struct ggml_tensor * result; - const int64_t ne[3] = { + const int64_t ne[4] = { ggml_calc_pool_output_size(a->ne[0], k0, s0, p0), ggml_calc_pool_output_size(a->ne[1], k1, s1, p1), a->ne[2], + a->ne[3], }; - result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne); + result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); int32_t params[] = { op, k0, k1, s0, s1, p0, p1 }; ggml_set_op_params(result, params, sizeof(params)); @@ -6939,6 +7192,37 @@ struct ggml_tensor * ggml_pool_2d( return result; } +struct ggml_tensor * ggml_pool_2d_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * af, + enum ggml_op_pool op, + int k0, + int k1, + int s0, + int s1, + float p0, + float p1) { + + bool is_node = false; + + if (a->grad) { + is_node = true; + } + + struct ggml_tensor * result; + result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, af->ne); + + int32_t params[] = { op, k0, k1, s0, s1, p0, p1 }; + ggml_set_op_params(result, params, sizeof(params)); + + result->op = GGML_OP_POOL_2D_BACK; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = a; + result->src[1] = af; + return result; +} + // ggml_upscale static struct ggml_tensor * ggml_upscale_impl( @@ -7117,7 +7401,8 @@ struct ggml_tensor * ggml_flash_attn_ext( struct ggml_tensor * v, struct ggml_tensor * mask, float scale, - float max_bias) { + float max_bias, + float logit_softcap) { GGML_ASSERT(ggml_can_mul_mat(k, q)); // TODO: check if vT can be multiplied by (k*qT) @@ -7144,7 +7429,7 @@ struct ggml_tensor * ggml_flash_attn_ext( int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - float params[] = { scale, max_bias }; + float params[] = { scale, max_bias, logit_softcap }; ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_FLASH_ATTN_EXT; @@ -7164,7 +7449,7 @@ void ggml_flash_attn_ext_set_prec( const int32_t prec_i32 = (int32_t) prec; - ggml_set_op_params_i32(a, 2, prec_i32); // scale is on first pos, max_bias on second + ggml_set_op_params_i32(a, 3, prec_i32); // scale is on first pos, max_bias on second } // ggml_flash_attn_back @@ -7251,43 +7536,34 @@ struct ggml_tensor * ggml_flash_attn_back( struct ggml_tensor * ggml_ssm_conv( struct ggml_context * ctx, - struct ggml_tensor * s, - struct ggml_tensor * x, - struct ggml_tensor * c, - struct ggml_tensor * sq) { - GGML_ASSERT(ggml_is_3d(s)); - GGML_ASSERT(ggml_is_matrix(x)); + struct ggml_tensor * sx, + struct ggml_tensor * c) { + GGML_ASSERT(ggml_is_3d(sx)); GGML_ASSERT(ggml_is_matrix(c)); - GGML_ASSERT(ggml_is_matrix(sq)); - GGML_ASSERT(sq->type == GGML_TYPE_I32); - const int64_t d_conv = c->ne[0]; - const int64_t d_inner = c->ne[1]; - const int64_t n_tokens = x->ne[1]; - const int64_t n_kv = s->ne[2]; + const int64_t d_conv = c->ne[0]; + const int64_t d_inner = c->ne[1]; + const int64_t n_t = sx->ne[0] - d_conv + 1; // tokens per sequence + const int64_t n_s = sx->ne[2]; - GGML_ASSERT( s->ne[0] == d_conv - 1); - GGML_ASSERT( s->ne[1] == d_inner); - GGML_ASSERT( x->ne[0] == d_inner); - GGML_ASSERT(sq->ne[0] == n_kv); - GGML_ASSERT(sq->ne[1] == n_tokens); + // TODO: maybe support other strides than 1? + GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t); + GGML_ASSERT(sx->ne[1] == d_inner); + GGML_ASSERT(n_t >= 0); bool is_node = false; - if (s->grad || x->grad || c->grad || sq->grad) { + if (sx->grad || c->grad) { GGML_ABORT("fatal error"); // TODO: implement is_node = true; } - // 2-in-1 concatenated x and conv_states, {d_inner, n_tokens} with {d_conv, d_inner, n_kv} - struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, (d_inner*n_tokens) + (d_conv*d_inner*n_kv)); + struct ggml_tensor * result = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_t, n_s); result->op = GGML_OP_SSM_CONV; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = s; - result->src[1] = x; - result->src[2] = c; - result->src[3] = sq; + result->src[0] = sx; + result->src[1] = c; return result; } @@ -7301,39 +7577,42 @@ struct ggml_tensor * ggml_ssm_scan( struct ggml_tensor * dt, struct ggml_tensor * A, struct ggml_tensor * B, - struct ggml_tensor * C, - struct ggml_tensor * sq) { + struct ggml_tensor * C) { GGML_ASSERT(ggml_is_contiguous(s)); GGML_ASSERT(ggml_is_contiguous(x)); GGML_ASSERT(ggml_is_contiguous(dt)); GGML_ASSERT(ggml_is_contiguous(A)); - GGML_ASSERT(sq->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_is_matrix(A)); + GGML_ASSERT(ggml_is_3d(B)); + GGML_ASSERT(ggml_is_3d(s)); GGML_ASSERT(B->nb[0] == ggml_type_size(B->type)); GGML_ASSERT(C->nb[0] == ggml_type_size(C->type)); GGML_ASSERT(ggml_are_same_shape(x, dt)); + GGML_ASSERT(ggml_are_same_shape(B, C)); { - const int64_t d_state = s->ne[0]; - const int64_t d_inner = s->ne[1]; - const int64_t n_tokens = x->ne[1]; + const int64_t d_state = s->ne[0]; + const int64_t d_inner = s->ne[1]; + const int64_t n_seq_tokens = x->ne[1]; + const int64_t n_seqs = x->ne[2]; + GGML_ASSERT(s->ne[2] == n_seqs); GGML_ASSERT(x->ne[0] == d_inner); GGML_ASSERT(A->ne[0] == d_state); GGML_ASSERT(A->ne[1] == d_inner); GGML_ASSERT(B->ne[0] == d_state); - GGML_ASSERT(B->ne[1] == n_tokens); - GGML_ASSERT(C->ne[0] == d_state); - GGML_ASSERT(C->ne[1] == n_tokens); + GGML_ASSERT(B->ne[1] == n_seq_tokens); + GGML_ASSERT(B->ne[2] == n_seqs); } bool is_node = false; - if (s->grad || x->grad || dt->grad || A->grad || B->grad || C->grad || sq->grad) { + if (s->grad || x->grad || dt->grad || A->grad || B->grad || C->grad) { GGML_ABORT("fatal error"); // TODO: implement is_node = true; } - // 2-in-1 concatenated y and ssm_states, {d_inner, n_tokens} with {d_state, d_inner, n_kv} + // concatenated y + ssm_states struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s)); result->op = GGML_OP_SSM_SCAN; @@ -7344,7 +7623,6 @@ struct ggml_tensor * ggml_ssm_scan( result->src[3] = A; result->src[4] = B; result->src[5] = C; - result->src[6] = sq; return result; } @@ -7495,6 +7773,59 @@ struct ggml_tensor * ggml_add_rel_pos_inplace( return ggml_add_rel_pos_impl(ctx, a, pw, ph, true); } +// ggml_rwkv_wkv + +struct ggml_tensor * ggml_rwkv_wkv( + struct ggml_context * ctx, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * r, + struct ggml_tensor * tf, + struct ggml_tensor * td, + struct ggml_tensor * state) { + GGML_ASSERT(ggml_is_contiguous(k)); + GGML_ASSERT(ggml_is_contiguous(v)); + GGML_ASSERT(ggml_is_contiguous(r)); + GGML_ASSERT(ggml_is_contiguous(tf)); + GGML_ASSERT(ggml_is_contiguous(td)); + GGML_ASSERT(ggml_is_contiguous(state)); + + const int64_t S = k->ne[0]; + const int64_t H = k->ne[2]; + const int64_t n_tokens = k->ne[3]; + const int64_t n_seqs = state->ne[1]; + { + GGML_ASSERT(k->ne[1] == 1); + GGML_ASSERT(v->ne[0] == 1 && v->ne[1] == S && v->ne[2] == H && v->ne[3] == n_tokens); + GGML_ASSERT(r->ne[0] == 1 && r->ne[1] == S && r->ne[2] == H && r->ne[3] == n_tokens); + // TODO: RWKV v4 and v5 + GGML_ASSERT(td->ne[0] == 1 && td->ne[1] == S && td->ne[2] == H && td->ne[3] == n_tokens); + GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs); + } + + bool is_node = false; + + if (k->grad || v->grad || r->grad || tf->grad || td->grad || state->grad) { + GGML_ABORT("fatal error"); // TODO: implement backward + is_node = true; + } + + // concat output and new_state + const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + + result->op = GGML_OP_RWKV_WKV; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = k; + result->src[1] = v; + result->src[2] = r; + result->src[3] = tf; + result->src[4] = td; + result->src[5] = state; + + return result; +} + // ggml_unary static struct ggml_tensor * ggml_unary_impl( @@ -10021,7 +10352,7 @@ static void ggml_compute_forward_acc_f32( ((char *) src0->data), ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } const int ith = params->ith; @@ -10126,11 +10457,10 @@ static void ggml_compute_forward_sub_f32( const struct ggml_tensor * src0 = dst->src[0]; const struct ggml_tensor * src1 = dst->src[1]; - if (params->ith != 0) { - return; - } + assert(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); - assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + const int ith = params->ith; + const int nth = params->nth; const int nr = ggml_nrows(src0); @@ -10139,40 +10469,55 @@ static void ggml_compute_forward_sub_f32( GGML_ASSERT( nb0 == sizeof(float)); GGML_ASSERT(nb00 == sizeof(float)); - if (nb10 == sizeof(float)) { - for (int ir = 0; ir < nr; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + // rows per thread + const int dr = (nr + nth - 1)/nth; + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (nb10 == sizeof(float)) { + for (int ir = ir0; ir < ir1; ++ir) { + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + const int64_t nr0 = ne00 / ne10; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); + + for (int64_t r = 0; r < nr0; ++r) { #ifdef GGML_USE_ACCELERATE - vDSP_vsub( - (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, - (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, - (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, - ne0); + vDSP_vsub(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10); #else - ggml_vec_sub_f32(ne0, - (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), - (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), - (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); + ggml_vec_sub_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); #endif - // } - // } + } } } else { // src1 is not contiguous - for (int ir = 0; ir < nr; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + for (int ir = ir0; ir < ir1; ++ir) { + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - for (int i0 = 0; i0 < ne0; i0++) { - float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10); + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + + float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + + for (int64_t i0 = 0; i0 < ne0; ++i0) { + const int64_t i10 = i0 % ne10; + float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); dst_ptr[i0] = src0_ptr[i0] - *src1_ptr; } @@ -10518,6 +10863,96 @@ static void ggml_compute_forward_log( } } +// ggml_compute_forward_sin + +static void ggml_compute_forward_sin_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_sin_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_sin( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sin_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_cos + +static void ggml_compute_forward_cos_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_cos_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_cos( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_cos_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + // ggml_compute_forward_sum static void ggml_compute_forward_sum_f32( @@ -11017,11 +11452,6 @@ static void ggml_compute_forward_concat_f32( GGML_TENSOR_BINARY_OP_LOCALS - // TODO: support for transposed / permuted tensors - GGML_ASSERT(nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - GGML_ASSERT(nb10 == sizeof(float)); - const int32_t dim = ggml_get_op_params_i32(dst, 0); GGML_ASSERT(dim >= 0 && dim < 4); @@ -11795,6 +12225,48 @@ static void ggml_compute_forward_hardsigmoid( } } +static void ggml_compute_forward_exp_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + for (int i = 0; i < n; i++) { + ggml_vec_exp_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_exp( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_exp_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + // ggml_compute_forward_norm @@ -12396,10 +12868,10 @@ UseGgmlGemm1:; if (ith == 0) { // Every thread starts at ith, so the first unprocessed chunk is nth. This save a bit of coordination right at the start. - atomic_store(¶ms->shared->current_chunk, nth); + atomic_store_explicit(¶ms->threadpool->current_chunk, nth, memory_order_relaxed); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); #if GGML_USE_LLAMAFILE if (src1->type != vec_dot_type) { @@ -12507,7 +12979,7 @@ UseGgmlGemm2:; break; } - current_chunk = atomic_fetch_add(¶ms->shared->current_chunk, 1); + current_chunk = atomic_fetch_add_explicit(¶ms->threadpool->current_chunk, 1, memory_order_relaxed); } } @@ -12602,7 +13074,7 @@ static void ggml_compute_forward_mul_mat_id( } } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); // compute each matrix multiplication in sequence for (int cur_a = 0; cur_a < n_as; ++cur_a) { @@ -12756,7 +13228,7 @@ static void ggml_compute_forward_out_prod_f32( if (ith == 0) { ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); // dst[:,:,:,:] = 0 // for i2,i3: @@ -12874,7 +13346,7 @@ static void ggml_compute_forward_out_prod_q_f32( if (ith == 0) { ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); // parallelize by last three dimensions @@ -13060,7 +13532,7 @@ static void ggml_compute_forward_set_f32( ((char *) src0->data), ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } const int ith = params->ith; @@ -13639,7 +14111,7 @@ static void ggml_compute_forward_diag_mask_f32( ((char *) src0->data), ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } // TODO: handle transposed/permuted matrices @@ -14116,7 +14588,7 @@ static void ggml_compute_forward_rope_f32( float corr_dims[2]; ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); - const bool is_neox = mode & 2; + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; const float * freq_factors = NULL; if (src2 != NULL) { @@ -14241,7 +14713,7 @@ static void ggml_compute_forward_rope_f16( float corr_dims[2]; ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); - const bool is_neox = mode & 2; + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; const float * freq_factors = NULL; if (src2 != NULL) { @@ -14415,7 +14887,7 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32( // need to zero dst since we are accumulating into it memset(dst->data, 0, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; @@ -14503,7 +14975,7 @@ static void ggml_compute_forward_conv_transpose_1d_f32( // need to zero dst since we are accumulating into it memset(dst->data, 0, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; @@ -14558,6 +15030,7 @@ static void ggml_compute_forward_conv_transpose_1d( } } +// ggml_compute_forward_im2col_f32 // src0: kernel [OC, IC, KH, KW] // src1: image [N, IC, IH, IW] // dst: result [N, OH, OW, IC*KH*KW] @@ -14568,7 +15041,6 @@ static void ggml_compute_forward_im2col_f32( const struct ggml_tensor * src0 = dst->src[0]; const struct ggml_tensor * src1 = dst->src[1]; - GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); @@ -14599,7 +15071,6 @@ static void ggml_compute_forward_im2col_f32( int ofs0 = is_2D ? nb13 : nb12; int ofs1 = is_2D ? nb12 : nb11; - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb10 == sizeof(float)); // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] @@ -14635,6 +15106,7 @@ static void ggml_compute_forward_im2col_f32( } +// ggml_compute_forward_im2col_f16 // src0: kernel [OC, IC, KH, KW] // src1: image [N, IC, IH, IW] // dst: result [N, OH, OW, IC*KH*KW] @@ -14730,6 +15202,99 @@ static void ggml_compute_forward_im2col( } } +// ggml_compute_forward_im2col_back_f32 + +static void ggml_compute_forward_im2col_back_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t N = is_2D ? ne3 : ne2; + const int64_t IC = is_2D ? ne2 : ne1; + const int64_t IH = is_2D ? ne1 : 1; + const int64_t IW = ne0; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne12 : 1; + const int64_t OW = ne11; + + int ofs0 = is_2D ? nb3 : nb2; + int ofs1 = is_2D ? nb2 : nb1; + + GGML_ASSERT(nb0 == sizeof(float)); + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + float * const wdata = (float *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t iic = ith; iic < IC; iic += nth) { + for (int64_t iih = 0; iih < IH; iih++) { + for (int64_t iiw = 0; iiw < IW; iiw++) { + + // micro kernel + float grad = 0.0f; + for (int64_t ikh = 0; ikh < KH; ikh++) { + for (int64_t ikw = 0; ikw < KW; ikw++) { + // For s0 > 1 some values were skipped over in the forward pass. + // These values have tmpw % s0 != 0 and need to be skipped in the backwards pass as well. + const int64_t tmpw = (iiw + p0 - ikw*d0); + if (tmpw % s0 != 0) { + continue; + } + const int64_t iow = tmpw / s0; + + // Equivalent logic as above except for s1. + int64_t ioh; + if (is_2D) { + const int64_t tmph = iih + p1 - ikh*d1; + + if (tmph % s1 != 0) { + continue; + } + + ioh = tmph / s1; + } else { + ioh = 0; + } + + if (iow < 0 || iow >= OW || ioh < 0 || ioh >= OH) { + continue; + } + + const float * const src_data = (const float *) src1->data + + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + grad += src_data[iic*(KH*KW) + ikh*KW + ikw]; + } + } + float * dst_data = (float *)((char *) wdata + (in*ofs0 + iic*ofs1)); // [IH, IW] + dst_data[iih*IW + iiw] = grad; + } + } + } + } + } +} // ggml_compute_forward_conv_transpose_2d @@ -14790,7 +15355,7 @@ static void ggml_compute_forward_conv_transpose_2d( memset(dst->data, 0, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int32_t stride = ggml_get_op_params_i32(dst, 0); @@ -14972,6 +15537,128 @@ static void ggml_compute_forward_pool_2d( } } +// ggml_compute_forward_pool_2d_back + +static void ggml_compute_forward_pool_2d_back( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src = dst->src[0]; + const struct ggml_tensor * dstf = dst->src[1]; // forward tensor of dst + + assert(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + + if (params->ith != 0) { + return; + } + + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + + char * cdata = (char *) dst->data; + const char * cdataf = (const char *) dstf->data; + const char * const data_end = cdata + ggml_nbytes(dst); + + GGML_ASSERT(params->ith == 0); + memset(cdata, 0, ggml_nbytes(dst)); + + const int64_t px = src->ne[0]; + const int64_t py = src->ne[1]; + const int64_t pa = px * py; + + const float * splane = (const float *) src->data; + + const int ka = k0 * k1; + const int offset0 = -p0; + const int offset1 = -p1; + + while (cdata < data_end) { + for (int oy = 0; oy < py; ++oy) { + const float * const srow = splane + oy * px; + for (int ox = 0; ox < px; ++ox) { + const float grad0 = srow[ox]; + + const int ix = offset0 + ox * s0; + const int iy = offset1 + oy * s1; + + if (op == GGML_OP_POOL_MAX) { + float maxval = -FLT_MAX; + int kxmax = -1; + int kymax = -1; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= dst->ne[1]) { + continue; + } + const void * drowf = (const void *)(cdataf + dst->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= dst->ne[0]) { + continue; + } + + const float val = dst->type == GGML_TYPE_F32 ? + ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]); + if (val <= maxval) { + continue; + } + + maxval = val; + kxmax = kx; + kymax = ky; + } + } + + if (kxmax == -1 || kymax == -1) { + continue; + } + + void * drow = (void *)(cdata + dst->nb[1] * (iy + kymax)); + const int j = ix + kxmax; + if (dst->type == GGML_TYPE_F32) { + ((float *) drow)[j] += grad0; + } else { + ((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j])); + } + } else if (op == GGML_OP_POOL_AVG) { + const float grad = grad0 / ka; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= dst->ne[1]) { + continue; + } + void * drow = (void *)(cdata + dst->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= dst->ne[0]) { + continue; + } + + if (dst->type == GGML_TYPE_F32) { + ((float *) drow)[j] += grad; + } else { + ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad); + } + } + } + } else { + GGML_ASSERT(false); + } + } + } + + cdata += dst->nb[2]; + cdataf += dst->nb[2]; + splane += pa; + } +} + // ggml_compute_forward_upscale static void ggml_compute_forward_upscale_f32( @@ -15305,11 +15992,17 @@ static void ggml_compute_forward_flash_attn_ext_f16( const int ir0 = dr*ith; const int ir1 = MIN(ir0 + dr, nr); - float scale = 1.0f; - float max_bias = 0.0f; + float scale = 1.0f; + float max_bias = 0.0f; + float logit_softcap = 0.0f; - memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); - memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float)); + + if (logit_softcap != 0) { + scale /= logit_softcap; + } const uint32_t n_head = neq2; const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); @@ -15373,7 +16066,13 @@ static void ggml_compute_forward_flash_attn_ext_f16( const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3); kq_vec_dot(D, &s, 0, k_data, 0, Q_q, 0, 1); - s = s*scale + mv; // scale KQ value and apply mask + s = s*scale; // scale KQ value + + if (logit_softcap != 0.0f) { + s = logit_softcap*tanhf(s); + } + + s += mv; // apply mask const float Mold = M; @@ -15382,7 +16081,7 @@ static void ggml_compute_forward_flash_attn_ext_f16( const char * v_data = ((const char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3)); - if (v->type== GGML_TYPE_F16) { + if (v->type == GGML_TYPE_F16) { if (s > M) { // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f M = s; @@ -15449,7 +16148,7 @@ static void ggml_compute_forward_flash_attn_ext( const struct ggml_tensor * v, const struct ggml_tensor * mask, struct ggml_tensor * dst) { - switch (dst->op_params[2]) { + switch (dst->op_params[3]) { case GGML_PREC_DEFAULT: case GGML_PREC_F32: { @@ -15524,7 +16223,7 @@ static void ggml_compute_forward_flash_attn_back_f32( if (ith == 0) { memset(dst->data, 0, nb0*ne0*ne1*ne2*ne3); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); const int64_t elem_q = ggml_nelements(q); const int64_t elem_k = ggml_nelements(k); @@ -15804,27 +16503,22 @@ static void ggml_compute_forward_flash_attn_back( static void ggml_compute_forward_ssm_conv_f32( const struct ggml_compute_params * params, struct ggml_tensor * dst) { - const struct ggml_tensor * src0 = dst->src[0]; // conv_state - const struct ggml_tensor * src1 = dst->src[1]; // x - const struct ggml_tensor * src2 = dst->src[2]; // conv1d.weight - const struct ggml_tensor * src3 = dst->src[3]; // state_seq + const struct ggml_tensor * src0 = dst->src[0]; // conv_x + const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight const int ith = params->ith; const int nth = params->nth; - const int nc = src2->ne[0]; // d_conv - const int nr = src0->ne[1]; // d_inner - const int n_t = src1->ne[1]; // n_tokens - const int n_kv = src0->ne[2]; // max number of sequences in the batch + const int nc = src1->ne[0]; // d_conv + const int ncs = src0->ne[0]; // d_conv - 1 + n_t + const int nr = src0->ne[1]; // d_inner + const int n_t = dst->ne[1]; // tokens per sequence + const int n_s = dst->ne[2]; // number of sequences in the batch - GGML_ASSERT((nr*n_t) + (nc*nr*n_kv) == ggml_nelements(dst)); + GGML_ASSERT( dst->ne[0] == nr); GGML_ASSERT(src0->nb[0] == sizeof(float)); GGML_ASSERT(src1->nb[0] == sizeof(float)); - GGML_ASSERT(src2->nb[0] == sizeof(float)); - GGML_ASSERT(src3->nb[0] == sizeof(int32_t)); GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); - // for use with the destination state offset between sequences - GGML_ASSERT(src2->nb[2] == src2->ne[1]*src2->ne[0]*sizeof(float)); // rows per thread const int dr = (nr + nth - 1)/nth; @@ -15834,76 +16528,29 @@ static void ggml_compute_forward_ssm_conv_f32( const int ir1 = MIN(ir0 + dr, nr); const int ir = ir1 - ir0; - if (n_kv > 1) { - // multiple sequences means it's hard to know when it's the first time a state is read, - // so copy them all over to the destination, just to be sure. - for (int i3 = 0; i3 < n_kv; ++i3) { - float * s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); - float * s = (float *) ((char *) dst->data + ir0*(src2->nb[1]) + i3*(src2->nb[2]) + nr*n_t*sizeof(float)); - // can't use memcpy because of d_conv vs d_conv - 1 + for (int i3 = 0; i3 < n_s; ++i3) { + for (int i2 = 0; i2 < n_t; ++i2) { + // {d_conv - 1 + n_t, d_inner, n_seqs} + // sliding window + const float * s = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i2*(src0->nb[0]) + i3*(src0->nb[2])); // {d_conv, d_inner, n_s} + const float * c = (const float *) ((const char *) src1->data + ir0*(src1->nb[1])); // {d_conv, d_inner} + float * x = (float *) ((char *) dst->data + ir0*(dst->nb[0]) + i2*(dst->nb[1]) + i3*(dst->nb[2])); // {d_inner, n_t, n_s} + + // TODO: transpose the output for smaller strides for big batches? + // d_inner for (int i1 = 0; i1 < ir; ++i1) { - for (int i0 = 0; i0 < nc - 1; ++i0) { - // copy s0 to last (d_conv - 1) columns of s - s[1 + i0 + i1*nc] = s0[i0 + i1*(nc - 1)]; + // rowwise dot product + // NOTE: not using ggml_vec_dot_f32, because its sum is in double precision + float sumf = 0.0f; + + // d_conv + for (int i0 = 0; i0 < nc; ++i0) { + sumf += s[i0 + i1*ncs] * c[i0 + i1*nc]; } + x[i1] = sumf; } } } - - for (int i2 = 0; i2 < n_t; ++i2) { - int32_t * sq = (int32_t *) ((char *) src3->data + i2*(src3->nb[1])); // {n_kv, n_tokens} - float * x = (float *) ((char *) dst->data + ir0*sizeof(float) + i2*(nr*sizeof(float))); // {d_inner, n_tokens} - float * s = (float *) ((char *) dst->data + ir0*(src2->nb[1]) + sq[0]*(src2->nb[2]) + nr*n_t*sizeof(float)); // {d_conv, d_inner, n_kv} - float * s0; // {d_conv - 1, d_inner, n_kv} - float * x0 = (float *) ((char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens} - float * c = (float *) ((char *) src2->data + ir0*(src2->nb[1])); // {d_conv, d_inner} - int ne0s0; - - GGML_ASSERT(0 <= sq[0] && sq[0] < n_kv); - - // avoid needing to copy the state for the first token - if (i2 == 0) { - s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2])); // {d_conv - 1, d_inner, n_kv} - ne0s0 = src0->ne[0]; - } else { - // the source is the last (d_conv - 1) columns of the destination - s0 = s + 1; - ne0s0 = nc; - } - - // d_inner - for (int i1 = 0; i1 < ir; ++i1) { - // shift state left - for (int i0 = 0; i0 < nc - 1; ++i0) { - s[i0 + i1*nc] = s0[i0 + i1*ne0s0]; - } - // insert x on the last column - s[(nc - 1) + i1*nc] = x0[i1]; - } - - // handle copies when there are multiple output states - for (int i3 = 1; i3 < n_kv; ++i3) { - int32_t seq = sq[i3]; - if (0 <= seq && seq < n_kv) { - float * s1 = s + (seq - sq[0])*nc*nr; - memcpy(s1, s, nc*ir*sizeof(float)); - } else { - // stop at negative or too big seq_ids - break; - } - } - - // it seems a little faster when this is separate from the state shift - for (int i1 = 0; i1 < ir; ++i1) { - // rowwise dot product - float sumf = 0.0f; - for (int i0 = 0; i0 < nc; ++i0) { - int i = i0 + i1*nc; - sumf += s[i] * c[i]; - } - x[i1] = sumf; - } - } } static void ggml_compute_forward_ssm_conv( @@ -15932,15 +16579,14 @@ static void ggml_compute_forward_ssm_scan_f32( const struct ggml_tensor * src3 = dst->src[3]; // A const struct ggml_tensor * src4 = dst->src[4]; // B const struct ggml_tensor * src5 = dst->src[5]; // C - const struct ggml_tensor * src6 = dst->src[6]; // sq const int ith = params->ith; const int nth = params->nth; - const int64_t nc = src0->ne[0]; // d_state - const int64_t nr = src0->ne[1]; // d_inner - const int64_t n_t = src1->ne[1]; // number of tokens in the batch - const int64_t n_kv = src0->ne[2]; // max number of sequences in the batch + const int64_t nc = src0->ne[0]; // d_state + const int64_t nr = src0->ne[1]; // d_inner + const int64_t n_t = src1->ne[1]; // number of tokens per sequence + const int64_t n_s = src0->ne[2]; // number of sequences in the batch GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst)); GGML_ASSERT(src0->nb[0] == sizeof(float)); @@ -15949,12 +16595,12 @@ static void ggml_compute_forward_ssm_scan_f32( GGML_ASSERT(src3->nb[0] == sizeof(float)); GGML_ASSERT(src4->nb[0] == sizeof(float)); GGML_ASSERT(src5->nb[0] == sizeof(float)); - // required for the dot product between s and C, and when copying the states + // required for the dot product between s and C GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); // required for per-sequence offsets for states GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float)); - // required to get correct offset for state destination (i.e. src1->nb[2]) - GGML_ASSERT(src1->nb[2] == src1->ne[0]*src1->ne[1]*sizeof(float)); + // required to get correct offset for state destination (i.e. src1->nb[3]) + GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float)); // rows per thread const int dr = (nr + nth - 1)/nth; @@ -15964,64 +16610,36 @@ static void ggml_compute_forward_ssm_scan_f32( const int ir1 = MIN(ir0 + dr, nr); const int ir = ir1 - ir0; - if (n_kv > 1) { - // it's hard to know if the source states have already been copied - // when there are multiple, so copy them already. - for (int i3 = 0; i3 < n_kv; ++i3) { - float * s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); - float * s = (float *) ((char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[2]); - memcpy(s, s0, nc*ir*sizeof(float)); - } - } + for (int i3 = 0; i3 < n_s; ++i3) { + for (int i2 = 0; i2 < n_t; ++i2) { + const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s} + const float * x = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} + const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s} + const float * A = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner} + const float * B = (const float *) ((const char *) src4->data + i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s} + const float * C = (const float *) ((const char *) src5->data + i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s} + float * y = ( float *) (( char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} + float * s = ( float *) (( char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[3]); // {d_state, d_inner, n_s} - for (int i2 = 0; i2 < n_t; ++i2) { - int32_t * sq = (int32_t *) ((char *) src6->data + i2*(src6->nb[1])); // {n_kv, n_tokens} - float * y = (float *) ((char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens} - float * s = (float *) ((char *) dst->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2]) + src1->nb[2]); // {d_state, d_inner, n_kv} - float * s0; - float * x = (float *) ((char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens} - float * dt = (float *) ((char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1])); // {d_inner, n_tokens} - float * A = (float *) ((char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner} - float * B = (float *) ((char *) src4->data + i2*(src4->nb[1])); // {d_state, n_tokens} - float * C = (float *) ((char *) src5->data + i2*(src5->nb[1])); // {d_state, n_tokens} + // use the output as the source for the next token-wise iterations + if (i2 > 0) { s0 = s; } - GGML_ASSERT(0 <= sq[0] && sq[0] < n_kv); - - // avoid needing to copy the state for the first token - if (i2 == 0) { - s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2])); // {d_state, d_inner, n_kv} - } else { - // otherwise the source is the same as the destination - s0 = s; - } - - // d_inner - for (int i1 = 0; i1 < ir; ++i1) { - // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78 - float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1]; - float x_dt = x[i1] * dt_soft_plus; - float sumf = 0.0f; - // d_state - for (int i0 = 0; i0 < nc; ++i0) { - int i = i0 + i1*nc; - // state = prev_state * dA + dB * x - float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt); - // y = rowwise_dotprod(state, C) - sumf += state * C[i0]; - s[i] = state; - } - y[i1] = sumf; - } - - // handle copies when there are multiple output states - for (int i3 = 1; i3 < n_kv; ++i3) { - int32_t seq = sq[i3]; - if (0 <= seq && seq < n_kv) { - float * s1 = s + (seq - sq[0])*nc*nr; - memcpy(s1, s, nc*ir*sizeof(float)); - } else { - // stop at negative or too big seq_ids - break; + // d_inner + for (int i1 = 0; i1 < ir; ++i1) { + // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78 + float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1]; + float x_dt = x[i1] * dt_soft_plus; + float sumf = 0.0f; + // d_state + for (int i0 = 0; i0 < nc; ++i0) { + int i = i0 + i1*nc; + // state = prev_state * dA + dB * x + float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt); + // y = rowwise_dotprod(state, C) + sumf += state * C[i0]; + s[i] = state; + } + y[i1] = sumf; } } } @@ -16227,6 +16845,10 @@ static void ggml_compute_forward_unary( { ggml_compute_forward_hardsigmoid(params, dst); } break; + case GGML_UNARY_OP_EXP: + { + ggml_compute_forward_exp(params, dst); + } break; default: { GGML_ABORT("fatal error"); @@ -16296,7 +16918,7 @@ static void ggml_compute_forward_add_rel_pos_f32( if (params->ith == 0) { memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst)); } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); } // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L357-L359 @@ -16362,6 +16984,96 @@ static void ggml_compute_forward_add_rel_pos( } } +// ggml_compute_forward_rwkv_wkv + +static void ggml_compute_forward_rwkv_wkv_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + const size_t T = dst->src[1]->ne[3]; + const size_t C = dst->ne[0]; + const size_t H = dst->src[1]->ne[2]; + const size_t n_seqs = dst->src[5]->ne[1]; + + float * dst_data = (float *) dst->data; + float * state = ((float *) dst->data) + C * T; + + if (params->ith != 0) { + return; + } + + memset(dst_data, 0, T * C * sizeof(float)); + + float * k = (float *) dst->src[0]->data; + float * v = (float *) dst->src[1]->data; + float * r = (float *) dst->src[2]->data; + float * time_faaaa = (float *) dst->src[3]->data; + float * time_decay = (float *) dst->src[4]->data; + + size_t t_stride = H * (C / H); + + size_t h_stride = C / H; + size_t h_stride_2d = (C / H) * (C / H); + + // basically fused operations: + // dst = r @ (time_faaaa * (k @ v) + state), + // state = time_decay * state + (k @ v), + // recursive through each token + for (size_t t = 0; t < T; t++) { + size_t t_offset = t * t_stride; + size_t state_offset = (C / H) * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset; + + for (size_t h = 0; h < H; h++) { + size_t h_offset = h * h_stride; + size_t t_h_offset = t_offset + h_offset; + size_t h_2d_offset = h * h_stride_2d; + + for (size_t i = 0; i < C / H; i++) { + size_t t_h_i_offset = t_h_offset + i; + size_t h_i_offset = h_offset + i; + size_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float k_val = k[t_h_i_offset]; + float r_val = r[t_h_i_offset]; + float time_faaaa_val = time_faaaa[h_i_offset]; + // RWKV v6: different time_decay for each token. + float time_decay_val = time_decay[t_h_i_offset]; + + for (size_t j = 0; j < C / H; j ++) { + size_t t_h_j_offset = t_h_offset + j; + size_t h_2d_i_j_offset = h_2d_i_offset + j; + + float v_val = v[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + float temp_val = kv_val * time_faaaa_val + prev_state_val; + dst_data[t_h_j_offset] += temp_val * r_val; + state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val; + } + } + } + } +} + +static void ggml_compute_forward_rwkv_wkv( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + const struct ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rwkv_wkv_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + // ggml_compute_forward_map_unary static void ggml_compute_forward_map_unary_f32( @@ -16581,9 +17293,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32( if (ith == 0) { memset(sums, 0, sizeof(float) * (nth + nth * nc)); } - ggml_barrier(params->shared); - - const double eps = 1e-9; + ggml_barrier(params->threadpool); // rows per thread const int dr = (nr + nth - 1)/nth; @@ -16605,20 +17315,15 @@ static void ggml_compute_forward_cross_entropy_loss_f32( } #endif - // soft_max float max = -INFINITY; ggml_vec_max_f32(nc, &max, s0); - ggml_float sum = ggml_vec_soft_max_f32(nc, st, s0, max); - assert(sum > 0.0); - sum = (1.0 - eps) / sum; + ggml_float sum = ggml_vec_log_soft_max_f32(nc, st, s0, max); + assert(sum >= 0.0); - // avoid log(0) by rescaling from [0..1] to [eps..1] - ggml_vec_scale_f32(nc, st, sum); - ggml_vec_add1_f32(nc, st, st, eps); - ggml_vec_log_f32(nc, st, st); + ggml_vec_add1_f32(nc, st, st, -sum); ggml_vec_mul_f32(nc, st, st, s1); - float st_sum = 0; + float st_sum = 0.0f; ggml_vec_sum_f32(nc, &st_sum, st); sums[ith] += st_sum; @@ -16629,7 +17334,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32( } #endif } - ggml_barrier(params->shared); + ggml_barrier(params->threadpool); if (ith == 0) { float * dp = (float *) dst->data; @@ -16675,8 +17380,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( const int64_t ith = params->ith; const int64_t nth = params->nth; - const double eps = 1e-9; - // TODO: handle transposed/permuted matrices const int64_t nc = src0->ne[0]; const int64_t nr = ggml_nrows(src0); @@ -16708,11 +17411,9 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32( ggml_vec_max_f32(nc, &max, s0); ggml_float sum = ggml_vec_soft_max_f32(nc, ds0, s0, max); assert(sum > 0.0); - sum = (1.0 - eps) / sum; + ggml_vec_scale_f32(nc, ds0, 1.0/sum); // grad(src0) = (softmax(src0) - src1) * grad(cross_entropy_loss(src0, src1)) / nr - ggml_vec_scale_f32(nc, ds0, sum); - ggml_vec_add1_f32(nc, ds0, ds0, eps); ggml_vec_sub_f32(nc, ds0, ds0, s1); ggml_vec_scale_f32(nc, ds0, d[0] / (float) nr); @@ -16793,6 +17494,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_log(params, tensor); } break; + case GGML_OP_SIN: + { + ggml_compute_forward_sin(params, tensor); + } break; + case GGML_OP_COS: + { + ggml_compute_forward_cos(params, tensor); + } break; case GGML_OP_SUM: { ggml_compute_forward_sum(params, tensor); @@ -16933,6 +17642,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_im2col(params, tensor); } break; + case GGML_OP_IM2COL_BACK: + { + ggml_compute_forward_im2col_back_f32(params, tensor); + } break; case GGML_OP_CONV_TRANSPOSE_2D: { ggml_compute_forward_conv_transpose_2d(params, tensor); @@ -16945,6 +17658,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_pool_2d(params, tensor); } break; + case GGML_OP_POOL_2D_BACK: + { + ggml_compute_forward_pool_2d_back(params, tensor); + } break; case GGML_OP_UPSCALE: { ggml_compute_forward_upscale(params, tensor); @@ -17008,6 +17725,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_add_rel_pos(params, tensor); } break; + case GGML_OP_RWKV_WKV: + { + ggml_compute_forward_rwkv_wkv(params, tensor); + } break; case GGML_OP_MAP_UNARY: { ggml_unary_op_f32_t fun; @@ -17313,7 +18034,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table); } if (src1->grad) { - src1->grad = ggml_add_or_set(ctx, src1->grad, tensor->grad, zero_table); + if (ggml_are_same_shape(src0, src1)) { + src1->grad = ggml_add_or_set(ctx, src1->grad, tensor->grad, zero_table); + } else { + src1->grad = ggml_add_or_set(ctx, src1->grad, ggml_repeat_back(ctx, tensor->grad, src1), zero_table); + } } } break; case GGML_OP_ADD1: @@ -17439,6 +18164,30 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor zero_table); } } break; + case GGML_OP_SIN: + { + if (src0->grad) { + src0->grad = + ggml_add_or_set(ctx, + src0->grad, + ggml_mul(ctx, + tensor->grad, + ggml_cos(ctx, src0)), + zero_table); + } + } break; + case GGML_OP_COS: + { + if (src0->grad) { + src0->grad = + ggml_sub_or_set(ctx, + src0->grad, + ggml_mul(ctx, + tensor->grad, + ggml_sin(ctx, src0)), + zero_table); + } + } break; case GGML_OP_SUM: { if (src0->grad) { @@ -17886,6 +18635,23 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor GGML_ABORT("fatal error"); // TODO: not implemented } case GGML_OP_IM2COL: + { + if (src1->grad) { + const int32_t s0 = ggml_get_op_params_i32(tensor, 0); + const int32_t s1 = ggml_get_op_params_i32(tensor, 1); + const int32_t p0 = ggml_get_op_params_i32(tensor, 2); + const int32_t p1 = ggml_get_op_params_i32(tensor, 3); + const int32_t d0 = ggml_get_op_params_i32(tensor, 4); + const int32_t d1 = ggml_get_op_params_i32(tensor, 5); + const bool is_2D = ggml_get_op_params_i32(tensor, 6) == 1; + + src1->grad = ggml_add_or_set(ctx, + src1->grad, + ggml_im2col_back(ctx, src0, tensor->grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D), + zero_table); + } + } break; + case GGML_OP_IM2COL_BACK: { GGML_ABORT("fatal error"); // TODO: not implemented } @@ -17898,6 +18664,23 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor GGML_ABORT("fatal error"); // TODO: not implemented } case GGML_OP_POOL_2D: + { + if (src0->grad) { + const enum ggml_op_pool op = ggml_get_op_params_i32(tensor, 0); + const int32_t k0 = ggml_get_op_params_i32(tensor, 1); + const int32_t k1 = ggml_get_op_params_i32(tensor, 2); + const int32_t s0 = ggml_get_op_params_i32(tensor, 3); + const int32_t s1 = ggml_get_op_params_i32(tensor, 4); + const int32_t p0 = ggml_get_op_params_i32(tensor, 5); + const int32_t p1 = ggml_get_op_params_i32(tensor, 6); + + src0->grad = ggml_add_or_set(ctx, + src0->grad, + ggml_pool_2d_back(ctx, tensor->grad, src0, op, k0, k1, s0, s1, p0, p1), + zero_table); + } + } break; + case GGML_OP_POOL_2D_BACK: { GGML_ABORT("fatal error"); // TODO: not implemented } @@ -18063,12 +18846,22 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor zero_table); } } break; + case GGML_UNARY_OP_EXP: + { + if (src0->grad) { + src0->grad = ggml_add_or_set(ctx, + src0->grad, + ggml_mul(ctx, tensor, tensor->grad), + zero_table); + } + } break; default: GGML_ABORT("fatal error"); } } break; case GGML_OP_GET_REL_POS: case GGML_OP_ADD_REL_POS: + case GGML_OP_RWKV_WKV: case GGML_OP_MAP_UNARY: case GGML_OP_MAP_BINARY: case GGML_OP_MAP_CUSTOM1_F32: @@ -18187,6 +18980,7 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep) { GGML_ASSERT(gf->n_nodes > 0); + GGML_ASSERT(gf->grads); // if we are keeping the gradient graph, we have to detach the gradient nodes from the original graph if (keep) { @@ -18370,65 +19164,6 @@ void ggml_graph_clear(struct ggml_cgraph * cgraph) { ggml_hash_set_reset(&cgraph->visited_hash_set); } -// -// thread data -// -// synchronization is done via busy loops -// I tried using spin locks, but not sure how to use them correctly - the things I tried were slower than busy loops -// - -#ifdef __APPLE__ - -//#include -// -//typedef os_unfair_lock ggml_lock_t; -// -//#define ggml_lock_init(x) UNUSED(x) -//#define ggml_lock_destroy(x) UNUSED(x) -//#define ggml_lock_lock os_unfair_lock_lock -//#define ggml_lock_unlock os_unfair_lock_unlock -// -//#define GGML_LOCK_INITIALIZER OS_UNFAIR_LOCK_INIT - -typedef int ggml_lock_t; - -#define ggml_lock_init(x) UNUSED(x) -#define ggml_lock_destroy(x) UNUSED(x) -#define ggml_lock_lock(x) UNUSED(x) -#define ggml_lock_unlock(x) UNUSED(x) - -#define GGML_LOCK_INITIALIZER 0 - -#define ggml_thread_create pthread_create -#define ggml_thread_join pthread_join - -#else - -//typedef pthread_spinlock_t ggml_lock_t; - -//#define ggml_lock_init(x) pthread_spin_init(x, PTHREAD_PROCESS_PRIVATE) -//#define ggml_lock_destroy pthread_spin_destroy -//#define ggml_lock_lock pthread_spin_lock -//#define ggml_lock_unlock pthread_spin_unlock - -typedef int ggml_lock_t; - -#define ggml_lock_init(x) UNUSED(x) -#define ggml_lock_destroy(x) UNUSED(x) -#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64)) -#define ggml_lock_lock(x) _mm_pause() -#else -#define ggml_lock_lock(x) UNUSED(x) -#endif -#define ggml_lock_unlock(x) UNUSED(x) - -#define GGML_LOCK_INITIALIZER 0 - -#define ggml_thread_create pthread_create -#define ggml_thread_join pthread_join - -#endif - // Android's libc implementation "bionic" does not support setting affinity #if defined(__gnu_linux__) static void set_numa_thread_affinity(int thread_n) { @@ -18526,6 +19261,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_OP_SQR: case GGML_OP_SQRT: case GGML_OP_LOG: + case GGML_OP_SIN: + case GGML_OP_COS: case GGML_OP_SUM: case GGML_OP_SUM_ROWS: case GGML_OP_MEAN: @@ -18548,6 +19285,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_UNARY_OP_SIGMOID: case GGML_UNARY_OP_HARDSWISH: case GGML_UNARY_OP_HARDSIGMOID: + case GGML_UNARY_OP_EXP: { n_tasks = 1; } break; @@ -18612,6 +19350,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { n_tasks = MIN(n_threads, ggml_nrows(node->src[0])); } break; case GGML_OP_IM2COL: + case GGML_OP_IM2COL_BACK: case GGML_OP_CONV_TRANSPOSE_1D: case GGML_OP_CONV_TRANSPOSE_2D: { @@ -18619,6 +19358,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { } break; case GGML_OP_POOL_1D: case GGML_OP_POOL_2D: + case GGML_OP_POOL_2D_BACK: { n_tasks = 1; } break; @@ -18637,6 +19377,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_OP_WIN_PART: case GGML_OP_WIN_UNPART: case GGML_OP_GET_REL_POS: + case GGML_OP_RWKV_WKV: case GGML_OP_MAP_UNARY: case GGML_OP_MAP_BINARY: case GGML_OP_MAP_CUSTOM1_F32: @@ -18705,9 +19446,268 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { return n_tasks; } -struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threads) { +static thread_ret_t ggml_graph_compute_secondary_thread(void* data); + +#if defined(_WIN32) +#include "windows.h" + +// TODO: support > 64 CPUs +bool ggml_thread_apply_affinity(bool * mask) { + HANDLE h = GetCurrentThread(); + uint64_t bitmask = 0ULL; + + assert(GGML_MAX_N_THREADS >= 64); + + for (int32_t i = 0; i < 8; i++) { + int32_t idx = i * 8; + uint8_t val = 0; + val |= mask[idx + 0] << 0; + val |= mask[idx + 1] << 1; + val |= mask[idx + 2] << 2; + val |= mask[idx + 3] << 3; + val |= mask[idx + 4] << 4; + val |= mask[idx + 5] << 5; + val |= mask[idx + 6] << 6; + val |= mask[idx + 7] << 7; + bitmask |= (uint64_t)val << idx; + } + + for (int32_t i = 64; i < GGML_MAX_N_THREADS; i++) { + if (mask[i]) { + fprintf(stderr, "warn: setting thread-affinity for > 64 CPUs isn't supported on windows!\n"); + break; + } + } + + DWORD_PTR m = (DWORD_PTR)bitmask; + + m = SetThreadAffinityMask(h, m); + + return m != 0; +} + +static bool ggml_thread_apply_priority(int32_t prio) { + // Note that on Windows the Process Priority Class must be updated in order to set Thread priority. + // This is up to the applications. + DWORD p = THREAD_PRIORITY_NORMAL; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: p = THREAD_PRIORITY_NORMAL; break; + case GGML_SCHED_PRIO_MEDIUM: p = THREAD_PRIORITY_ABOVE_NORMAL; break; + case GGML_SCHED_PRIO_HIGH: p = THREAD_PRIORITY_HIGHEST; break; + case GGML_SCHED_PRIO_REALTIME: p = THREAD_PRIORITY_TIME_CRITICAL; break; + } + + if (prio == GGML_SCHED_PRIO_NORMAL) { + // Keep inherited policy/priority + return true; + } + + if (!SetThreadPriority(GetCurrentThread(), p)) { + fprintf(stderr, "warn: failed to set thread priority %d : (%d)\n", prio, (int) GetLastError()); + return false; + } + + return true; +} + +#elif defined(__APPLE__) +#include +#include + +static bool ggml_thread_apply_affinity(const bool * mask) { + // Not supported on Apple platforms + UNUSED(mask); + return true; +} + +static bool ggml_thread_apply_priority(int32_t prio) { + struct sched_param p; + int32_t policy = SCHED_OTHER; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: policy = SCHED_OTHER; p.sched_priority = 0; break; + case GGML_SCHED_PRIO_MEDIUM: policy = SCHED_FIFO; p.sched_priority = 40; break; + case GGML_SCHED_PRIO_HIGH: policy = SCHED_FIFO; p.sched_priority = 80; break; + case GGML_SCHED_PRIO_REALTIME: policy = SCHED_FIFO; p.sched_priority = 90; break; + } + + if (prio == GGML_SCHED_PRIO_NORMAL) { + // Keep inherited policy/priority + return true; + } + + int32_t err = pthread_setschedparam(pthread_self(), policy, &p); + if (err != 0) { + fprintf(stderr, "warn: failed to set thread priority %d : %s (%d)\n", prio, strerror(err), err); + return false; + } + + return true; +} + +#else // posix? + +static bool ggml_thread_apply_affinity(const bool * mask) { + cpu_set_t cpuset; + int err; + + CPU_ZERO(&cpuset); + + for (uint32_t i = 0; i < GGML_MAX_N_THREADS; i++) { + if (mask[i]) { + GGML_PRINT_DEBUG("Thread %lx: adding %d to cpuset\n", pthread_self(), i); + CPU_SET(i, &cpuset); + } + } + +#ifdef __ANDROID__ + err = sched_setaffinity(0, sizeof(cpuset), &cpuset); + if (err < 0) { + err = errno; + } +#else + err = pthread_setaffinity_np(pthread_self(), sizeof(cpuset), &cpuset); +#endif + if (err != 0) { + fprintf(stderr, "warn: failed to set affinity mask 0x%llx : %s (%d)\n", (unsigned long long)mask, strerror(err), err); + return false; + } + + return true; +} + +static bool ggml_thread_apply_priority(int32_t prio) { + struct sched_param p; + int32_t policy = SCHED_OTHER; + switch (prio) { + case GGML_SCHED_PRIO_NORMAL: policy = SCHED_OTHER; p.sched_priority = 0; break; + case GGML_SCHED_PRIO_MEDIUM: policy = SCHED_FIFO; p.sched_priority = 40; break; + case GGML_SCHED_PRIO_HIGH: policy = SCHED_FIFO; p.sched_priority = 80; break; + case GGML_SCHED_PRIO_REALTIME: policy = SCHED_FIFO; p.sched_priority = 90; break; + } + + if (prio == GGML_SCHED_PRIO_NORMAL) { + // Keep inherited policy/priority + return true; + } + + int32_t err = pthread_setschedparam(pthread_self(), policy, &p); + if (err != 0) { + fprintf(stderr, "warn: failed to set thread priority %d : %s (%d)\n", prio, strerror(err), err); + return false; + } + + return true; +} + +#endif + +static bool ggml_thread_cpumask_is_valid(const bool * mask) { + for (int i = 0; i < GGML_MAX_N_THREADS; i++) { + if (mask[i]) { return true; } + } + return false; +} + +static void ggml_thread_cpumask_next(const bool * global_mask, bool * local_mask, bool strict, int32_t* iter) { + if (!strict) { + memcpy(local_mask, global_mask, GGML_MAX_N_THREADS); + return; + } else { + memset(local_mask, 0, GGML_MAX_N_THREADS); + int32_t base_idx = *iter; + for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) { + int32_t idx = base_idx + i; + if (idx >= GGML_MAX_N_THREADS) { + // Just a cheaper modulo + idx -= GGML_MAX_N_THREADS; + } + if (global_mask[idx]) { + local_mask[idx] = 1; + *iter = idx + 1; + return; + } + } + } +} + +void ggml_threadpool_free(struct ggml_threadpool* threadpool) { + if (!threadpool) return; + +#ifndef GGML_USE_OPENMP + struct ggml_compute_state* workers = threadpool->workers; + const int n_threads = threadpool->n_threads_max; + + ggml_mutex_lock(&threadpool->mutex); + + threadpool->stop = true; + threadpool->pause = false; + + ggml_cond_broadcast(&threadpool->cond); + ggml_mutex_unlock(&threadpool->mutex); + + for (int j = 1; j < n_threads; j++) { + int32_t rc = ggml_thread_join(workers[j].thrd, NULL); + GGML_ASSERT(rc == GGML_EXIT_SUCCESS || rc == GGML_EXIT_ABORTED); + UNUSED(rc); + } + + ggml_mutex_destroy(&threadpool->mutex); + ggml_cond_destroy(&threadpool->cond); +#endif // GGML_USE_OPENMP + + GGML_ALIGNED_FREE(threadpool->workers); + GGML_ALIGNED_FREE(threadpool); +} + +#ifndef GGML_USE_OPENMP +// pause/resume must be called under mutex +static void ggml_threadpool_pause_locked(struct ggml_threadpool * threadpool) { + GGML_PRINT_DEBUG("Pausing threadpool\n"); + threadpool->pause = true; + ggml_cond_broadcast(&threadpool->cond); +} + +static void ggml_threadpool_resume_locked(struct ggml_threadpool * threadpool) { + GGML_PRINT_DEBUG("Resuming threadpool\n"); + threadpool->pause = false; + ggml_cond_broadcast(&threadpool->cond); +} +#endif + +void ggml_threadpool_pause(struct ggml_threadpool * threadpool) { +#ifndef GGML_USE_OPENMP + ggml_mutex_lock(&threadpool->mutex); + if (!threadpool->pause) { + ggml_threadpool_pause_locked(threadpool); + } + ggml_mutex_unlock(&threadpool->mutex); +#else + UNUSED(threadpool); +#endif +} + +void ggml_threadpool_resume(struct ggml_threadpool * threadpool) { +#ifndef GGML_USE_OPENMP + ggml_mutex_lock(&threadpool->mutex); + if (threadpool->pause) { + ggml_threadpool_resume_locked(threadpool); + } + ggml_mutex_unlock(&threadpool->mutex); +#else + UNUSED(threadpool); +#endif +} + +struct ggml_cplan ggml_graph_plan( + const struct ggml_cgraph * cgraph, + int n_threads, + struct ggml_threadpool * threadpool) { + + if (threadpool == NULL) { + GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads); + } if (n_threads <= 0) { - n_threads = GGML_DEFAULT_N_THREADS; + n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS; } size_t work_size = 0; @@ -18863,12 +19863,13 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa } if (work_size > 0) { - work_size += CACHE_LINE_SIZE*(n_threads - 1); + work_size += CACHE_LINE_SIZE*(n_threads); } - cplan.n_threads = MIN(max_tasks, n_threads); - cplan.work_size = work_size; - cplan.work_data = NULL; + cplan.threadpool = threadpool; + cplan.n_threads = MIN(max_tasks, n_threads); + cplan.work_size = work_size; + cplan.work_data = NULL; return cplan; } @@ -18876,17 +19877,17 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa static thread_ret_t ggml_graph_compute_thread(void * data) { struct ggml_compute_state * state = (struct ggml_compute_state *) data; - const struct ggml_cgraph * cgraph = state->shared->cgraph; - const struct ggml_cplan * cplan = state->shared->cplan; + const struct ggml_cgraph * cgraph = state->threadpool->cgraph; + const struct ggml_cplan * cplan = state->threadpool->cplan; set_numa_thread_affinity(state->ith); struct ggml_compute_params params = { - /*.ith =*/ state->ith, - /*.nth =*/ state->shared->n_threads, - /*.wsize =*/ cplan->work_size, - /*.wdata =*/ cplan->work_data, - /*.shared=*/ state->shared, + /*.ith =*/ state->ith, + /*.nth =*/ state->threadpool->n_threads_cur, + /*.wsize =*/ cplan->work_size, + /*.wdata =*/ cplan->work_data, + /*.threadpool=*/ state->threadpool, }; for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) { @@ -18895,12 +19896,12 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { ggml_compute_forward(¶ms, node); if (state->ith == 0 && cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) { - state->shared->ec = GGML_STATUS_ABORTED; + state->threadpool->ec = GGML_STATUS_ABORTED; } - ggml_barrier(state->shared); + ggml_barrier(state->threadpool); - if (state->shared->ec != GGML_STATUS_SUCCESS) { + if (state->threadpool->ec != GGML_STATUS_SUCCESS) { break; } } @@ -18908,24 +19909,243 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { return 0; } +#ifndef GGML_USE_OPENMP + +static inline bool ggml_graph_compute_ready(struct ggml_compute_state * state) { + struct ggml_threadpool * threadpool = state->threadpool; + + if (state->pending || threadpool->stop || threadpool->pause) { return true; } + + // check for new graph/work + int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed); + if (new_graph != state->last_graph) { + state->pending = (state->ith < threadpool->n_threads_cur); + state->last_graph = new_graph; + } + + return state->pending; +} + +static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) { + struct ggml_threadpool * threadpool = state->threadpool; + + // This seems to make 0 ... 100 a decent range for polling level across modern processors. + // Perhaps, we can adjust it dynamically based on load and things. + const uint64_t n_rounds = 1024UL * 128 * threadpool->poll; + + for (uint64_t i=0; !ggml_graph_compute_ready(state) && ipending; +} + +static inline bool ggml_graph_compute_check_for_work(struct ggml_compute_state * state) { + struct ggml_threadpool * threadpool = state->threadpool; + + if (ggml_graph_compute_poll_for_work(state)) { + return state->pending; + } + + ggml_mutex_lock_shared(&threadpool->mutex); + while (!ggml_graph_compute_ready(state)) { + // No new work. Wait for the signal. + GGML_PRINT_DEBUG("thread #%d waiting for work\n", state->ith); + ggml_cond_wait(&threadpool->cond, &threadpool->mutex); + } + ggml_mutex_unlock_shared(&threadpool->mutex); + + return state->pending; +} + +static thread_ret_t ggml_graph_compute_secondary_thread(void* data) { + struct ggml_compute_state * state = (struct ggml_compute_state *) data; + struct ggml_threadpool * threadpool = state->threadpool; + + ggml_thread_apply_priority(threadpool->prio); + if (ggml_thread_cpumask_is_valid(state->cpumask)) { + ggml_thread_apply_affinity(state->cpumask); + } + + while (true) { + // Check if we need to sleep + while (threadpool->pause) { + GGML_PRINT_DEBUG("thread #%d inside pause loop\n", state->ith); + ggml_mutex_lock_shared(&threadpool->mutex); + if (threadpool->pause) { + ggml_cond_wait(&threadpool->cond, &threadpool->mutex); + } + GGML_PRINT_DEBUG("thread #%d resuming after wait\n", state->ith); + ggml_mutex_unlock_shared(&threadpool->mutex); + } + + // This needs to be checked for after the cond_wait + if (threadpool->stop) break; + + // Check if there is new work + // The main thread is the only one that can dispatch new work + + ggml_graph_compute_check_for_work(state); + if (state->pending) { + state->pending = false; + + ggml_graph_compute_thread(state); + } + } + + return (thread_ret_t) 0; +} + +// Start processing new graph +static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool) +{ + // always take the mutex here because the worker threads are doing hybrid poll/wait + + ggml_mutex_lock(&threadpool->mutex); + + atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_relaxed); + + if (threadpool->pause) { + // Update main thread prio and affinity to match the threadpool settings + ggml_thread_apply_priority(threadpool->prio); + if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) { + ggml_thread_apply_affinity(threadpool->workers[0].cpumask); + } + + // resume does cond broadcast + ggml_threadpool_resume_locked(threadpool); + } else { + ggml_cond_broadcast(&threadpool->cond); + } + + ggml_mutex_unlock(&threadpool->mutex); +} + +#endif // GGML_USE_OPENMP + +void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) { + p->n_threads = n_threads; + p->prio = 0; // default priority (usually means normal or inherited) + p->poll = 50; // hybrid-polling enabled + p->strict_cpu = false; // no strict placement (all threads share same cpumask) + p->paused = false; // threads are ready to go + memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited) +} + +struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) { + struct ggml_threadpool_params p; + ggml_threadpool_params_init(&p, n_threads); + return p; +} + +bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) { + if (p0->n_threads != p1->n_threads ) return false; + if (p0->prio != p1->prio ) return false; + if (p0->poll != p1->poll ) return false; + if (p0->strict_cpu != p1->strict_cpu ) return false; + return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0; +} + +static struct ggml_threadpool * ggml_threadpool_new_impl( + struct ggml_threadpool_params * tpp, + struct ggml_cgraph * cgraph, + struct ggml_cplan * cplan) { + + struct ggml_threadpool * threadpool = + GGML_ALIGNED_MALLOC(sizeof(struct ggml_threadpool)); + { + threadpool->cgraph = cgraph; + threadpool->cplan = cplan; + threadpool->n_graph = 0; + threadpool->n_barrier = 0; + threadpool->n_barrier_passed = 0; + threadpool->current_chunk = 0; + threadpool->stop = false; + threadpool->pause = tpp->paused; + threadpool->workers = NULL; + threadpool->n_threads_max = tpp->n_threads; + threadpool->n_threads_cur = tpp->n_threads; + threadpool->poll = tpp->poll; + threadpool->prio = tpp->prio; + threadpool->ec = GGML_STATUS_SUCCESS; + } + + // Allocate and init workers state + const size_t workers_size = sizeof(struct ggml_compute_state) * tpp->n_threads; + struct ggml_compute_state * workers = GGML_ALIGNED_MALLOC(workers_size); + + memset(workers, 0, workers_size); + for (int j = 0; j < tpp->n_threads; j++) { + workers[j].threadpool = threadpool; + workers[j].ith = j; + } + + threadpool->workers = workers; + +#ifndef GGML_USE_OPENMP + ggml_mutex_init(&threadpool->mutex); + ggml_cond_init(&threadpool->cond); + + // Spin the threads for all workers, and update CPU placements. + // Place the main thread last (towards the higher numbered CPU cores). + + int32_t cpumask_iter = 0; + + for (int j = 1; j < tpp->n_threads; j++) { + ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter); + + int32_t rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_secondary_thread, &workers[j]); + GGML_ASSERT(rc == 0); + } + + ggml_thread_cpumask_next(tpp->cpumask, workers[0].cpumask, tpp->strict_cpu, &cpumask_iter); + + if (!threadpool->pause) { + // Update main thread prio and affinity at the start, otherwise we'll do it in resume + ggml_thread_apply_priority(threadpool->prio); + if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) { + ggml_thread_apply_affinity(threadpool->workers[0].cpumask); + } + } +#endif // GGML_USE_OPENMP + + return threadpool; +} + +struct ggml_threadpool * ggml_threadpool_new(struct ggml_threadpool_params * tpp) { + return ggml_threadpool_new_impl(tpp, NULL, NULL); +} + enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { GGML_ASSERT(cplan); GGML_ASSERT(cplan->n_threads > 0); GGML_ASSERT(cplan->work_size == 0 || cplan->work_data != NULL); - int n_threads = cplan->n_threads; + int n_threads = cplan->n_threads; + struct ggml_threadpool * threadpool = cplan->threadpool; - struct ggml_compute_state_shared state_shared = { - /*.cgraph =*/ cgraph, - /*.cgraph_plan =*/ cplan, - /*.n_threads =*/ n_threads, - /*.n_barrier =*/ 0, - /*.n_barrier_passed =*/ 0, - /*.abort_callback =*/ NULL, - /*.abort_callback_data =*/ NULL, - /*.current_chunk =*/ 0, - /*.ec =*/ GGML_STATUS_SUCCESS, - }; + bool disposable_threadpool = false; + + if (threadpool == NULL) { + GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads); + disposable_threadpool = true; + + struct ggml_threadpool_params ttp = ggml_threadpool_params_default(n_threads); + threadpool = ggml_threadpool_new_impl(&ttp, cgraph, cplan); + } else { + // Reset some of the parameters that need resetting + // No worker threads should be accessing the parameters below at this stage + threadpool->cgraph = cgraph; + threadpool->cplan = cplan; + threadpool->n_threads_cur = n_threads; + threadpool->current_chunk = 0; + threadpool->ec = GGML_STATUS_SUCCESS; + } + + if (n_threads > threadpool->n_threads_max) { + GGML_PRINT("WARNING: cplan is requesting more threads than the threadpool contains. Expect a bad time!\n"); + } #ifdef GGML_USE_OPENMP if (n_threads > 1) { @@ -18935,63 +20155,36 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl { // update the number of threads from the actual number of threads that we got from OpenMP n_threads = omp_get_num_threads(); - state_shared.n_threads = n_threads; + threadpool->n_threads_cur = n_threads; } - struct ggml_compute_state worker = { - .thrd = 0, - .ith = omp_get_thread_num(), - .shared = &state_shared, - }; - ggml_graph_compute_thread(&worker); + ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]); } } else { - struct ggml_compute_state worker = { - .thrd = 0, - .ith = 0, - .shared = &state_shared, - }; - ggml_graph_compute_thread(&worker); + ggml_graph_compute_thread(&threadpool->workers[0]); } #else - struct ggml_compute_state * workers = alloca(sizeof(struct ggml_compute_state)*n_threads); + // Kick all threads to start the new graph + ggml_graph_compute_kickoff(threadpool); - for (int j = 0; j < n_threads; ++j) { - workers[j] = (struct ggml_compute_state) { - .thrd = 0, - .ith = j, - .shared = &state_shared, - }; - } - - // create thread pool - for (int j = 1; j < n_threads; ++j) { - const int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]); - GGML_ASSERT(rc == 0); - UNUSED(rc); - } - - // this is a work thread too - ggml_graph_compute_thread(&workers[0]); - - // join or kill thread pool - if (n_threads > 1) { - for (int j = 1; j < n_threads; j++) { - const int rc = ggml_thread_join(workers[j].thrd, NULL); - GGML_ASSERT(rc == 0); - UNUSED(rc); - } - } + // This is a work thread too + ggml_graph_compute_thread(&threadpool->workers[0]); #endif // don't leave affinity set on the main thread clear_numa_thread_affinity(); - return state_shared.ec; + enum ggml_status ret = threadpool->ec; + + if (disposable_threadpool) { + ggml_threadpool_free(threadpool); + } + + return ret; } enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) { - struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads); + struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads, NULL); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size); @@ -19132,9 +20325,11 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { const uint32_t type = tensor->type; const uint32_t op = tensor->op; + const int32_t flags = tensor->flags; fwrite(&type, sizeof(uint32_t), 1, fout); fwrite(&op, sizeof(uint32_t), 1, fout); + fwrite(&flags, sizeof(int32_t), 1, fout); for (int j = 0; j < GGML_MAX_DIMS; ++j) { const uint64_t ne = tensor->ne[j]; @@ -19164,9 +20359,11 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { const uint32_t type = tensor->type; const uint32_t op = tensor->op; + const int32_t flags = tensor->flags; fwrite(&type, sizeof(uint32_t), 1, fout); fwrite(&op, sizeof(uint32_t), 1, fout); + fwrite(&flags, sizeof(int32_t), 1, fout); for (int j = 0; j < GGML_MAX_DIMS; ++j) { const uint64_t ne = tensor->ne[j]; @@ -19225,6 +20422,14 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { } } } + + // dump the data + // TODO: pad this to 32 byte boundary + if ((flags & GGML_TENSOR_FLAG_PARAM)) { + const size_t size = ggml_nbytes(tensor); + + fwrite(tensor->data, sizeof(char), size, fout); + } } } @@ -19338,10 +20543,12 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context * { uint32_t type; uint32_t op; + int32_t flags; for (uint32_t i = 0; i < n_leafs; ++i) { type = *(const uint32_t *) ptr; ptr += sizeof(type); op = *(const uint32_t *) ptr; ptr += sizeof(op); + flags = *(const int32_t *) ptr; ptr += sizeof(flags); int64_t ne[GGML_MAX_DIMS]; size_t nb[GGML_MAX_DIMS]; @@ -19359,20 +20566,19 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context * struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne); - tensor->op = (enum ggml_op) op; + tensor->op = (enum ggml_op) op; + tensor->flags = flags; memcpy(tensor->name, ptr, GGML_MAX_NAME); ptr += GGML_MAX_NAME; memcpy(tensor->op_params, ptr, GGML_MAX_OP_PARAMS); ptr += GGML_MAX_OP_PARAMS; - tensor->data = (void *) ptr; - for (int j = 0; j < GGML_MAX_DIMS; ++j) { tensor->nb[j] = nb[j]; } - result->leafs[i] = tensor; + tensor->data = (void *) ptr; ptr += ggml_nbytes(tensor); - ptr += ggml_nbytes(tensor); + result->leafs[i] = tensor; fprintf(stderr, "%s: loaded leaf %u: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor)); } @@ -19384,10 +20590,12 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context * { uint32_t type; uint32_t op; + int32_t flags; for (uint32_t i = 0; i < n_nodes; ++i) { type = *(const uint32_t *) ptr; ptr += sizeof(type); op = *(const uint32_t *) ptr; ptr += sizeof(op); + flags = *(const int32_t *) ptr; ptr += sizeof(flags); enum ggml_op eop = (enum ggml_op) op; @@ -19477,6 +20685,11 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context * result->nodes[i] = tensor; + // TODO tensor data is be duplicated due to ggml_new_tensor call above + if (flags & GGML_TENSOR_FLAG_PARAM) { + tensor->data = (void *) ptr; ptr += ggml_nbytes(tensor); + } + fprintf(stderr, "%s: loaded node %u: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor)); } } @@ -19745,6 +20958,7 @@ static enum ggml_opt_result ggml_opt_adam( ggml_opt_callback callback, void * callback_data) { GGML_ASSERT(ggml_is_scalar(f)); + GGML_ASSERT(f->type == GGML_TYPE_F32); // these will store the parameters we want to optimize struct ggml_tensor * ps[GGML_MAX_PARAMS]; @@ -19786,7 +21000,7 @@ static enum ggml_opt_result ggml_opt_adam( float * pf = params.past > 0 ? opt->adam.pf->data : NULL; // past function values - struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads); + struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads, NULL); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size); cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; @@ -20133,7 +21347,7 @@ static enum ggml_opt_result ggml_opt_lbfgs( opt->iter = iter; } - struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads); + struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads, NULL); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size); cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs; @@ -20511,6 +21725,8 @@ enum ggml_opt_result ggml_opt( struct ggml_context * ctx, struct ggml_opt_params params, struct ggml_tensor * f) { + GGML_ASSERT(f->grad && "ggml_set_param called for at least one parent tensor."); + bool free_ctx = false; if (ctx == NULL) { struct ggml_init_params params_ctx = { @@ -20565,6 +21781,8 @@ enum ggml_opt_result ggml_opt_resume_g( ggml_opt_callback callback, void * callback_data) { + GGML_ASSERT(f->grad && "ggml_set_param must be called for at least one ancestor"); + // build forward + backward compute graphs enum ggml_opt_result result = GGML_OPT_RESULT_OK; @@ -21151,7 +22369,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p (int64_t) info->ne[2] * (int64_t) info->ne[3]; - if (ne % ggml_blck_size(info->type) != 0) { + if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) { fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n", __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type)); fclose(file); @@ -21652,6 +22870,7 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) { void gguf_add_tensor( struct gguf_context * ctx, const struct ggml_tensor * tensor) { + GGML_ASSERT(tensor); if (gguf_find_tensor(ctx, tensor->name) != -1) { GGML_ABORT("duplicated tensor name"); } diff --git a/llama/ggml.h b/llama/ggml.h index 3e87205b..177b527f 100644 --- a/llama/ggml.h +++ b/llama/ggml.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -246,7 +246,7 @@ #include #define GGML_FILE_MAGIC 0x67676d6c // "ggml" -#define GGML_FILE_VERSION 1 +#define GGML_FILE_VERSION 2 #define GGML_QNT_VERSION 2 // bump this on quantization format changes #define GGML_QNT_VERSION_FACTOR 1000 // do not change this @@ -257,6 +257,8 @@ #define GGML_MAX_SRC 10 #ifndef GGML_MAX_NAME #define GGML_MAX_NAME 64 +#define GGML_MAX_N_THREADS 512 + #endif #define GGML_MAX_OP_PARAMS 64 #define GGML_DEFAULT_N_THREADS 4 @@ -270,6 +272,8 @@ #define GGML_EXIT_SUCCESS 0 #define GGML_EXIT_ABORTED 1 +#define GGML_ROPE_TYPE_NEOX 2 + #define GGUF_MAGIC "GGUF" #define GGUF_VERSION 3 @@ -477,6 +481,8 @@ extern "C" { GGML_OP_SQR, GGML_OP_SQRT, GGML_OP_LOG, + GGML_OP_SIN, + GGML_OP_COS, GGML_OP_SUM, GGML_OP_SUM_ROWS, GGML_OP_MEAN, @@ -514,9 +520,11 @@ extern "C" { GGML_OP_CLAMP, GGML_OP_CONV_TRANSPOSE_1D, GGML_OP_IM2COL, + GGML_OP_IM2COL_BACK, GGML_OP_CONV_TRANSPOSE_2D, GGML_OP_POOL_1D, GGML_OP_POOL_2D, + GGML_OP_POOL_2D_BACK, GGML_OP_UPSCALE, // nearest interpolate GGML_OP_PAD, GGML_OP_ARANGE, @@ -532,6 +540,7 @@ extern "C" { GGML_OP_WIN_UNPART, GGML_OP_GET_REL_POS, GGML_OP_ADD_REL_POS, + GGML_OP_RWKV_WKV, GGML_OP_UNARY, @@ -566,6 +575,7 @@ extern "C" { GGML_UNARY_OP_SILU, GGML_UNARY_OP_HARDSWISH, GGML_UNARY_OP_HARDSIGMOID, + GGML_UNARY_OP_EXP, GGML_UNARY_OP_COUNT, }; @@ -648,6 +658,29 @@ extern "C" { // If it returns true, the computation is aborted typedef bool (*ggml_abort_callback)(void * data); + // Scheduling priorities + enum ggml_sched_priority { + GGML_SCHED_PRIO_NORMAL, + GGML_SCHED_PRIO_MEDIUM, + GGML_SCHED_PRIO_HIGH, + GGML_SCHED_PRIO_REALTIME + }; + + // Threadpool params + // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults + struct ggml_threadpool_params { + bool cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings) + int n_threads; // number of threads + enum ggml_sched_priority prio; // thread priority + uint32_t poll; // polling level (0 - no polling, 100 - aggressive polling) + bool strict_cpu; // strict cpu placement + bool paused; // start in paused state + }; + + struct ggml_threadpool; // forward declaration, see ggml.c + + typedef struct ggml_threadpool * ggml_threadpool_t; + // the compute plan that needs to be prepared for ggml_graph_compute() // since https://github.com/ggerganov/ggml/issues/287 struct ggml_cplan { @@ -655,6 +688,7 @@ extern "C" { uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()` int n_threads; + struct ggml_threadpool * threadpool; // abort ggml_graph_compute when true ggml_abort_callback abort_callback; @@ -993,6 +1027,22 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + GGML_API struct ggml_tensor * ggml_sin( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_sin_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_cos( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_cos_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + // return scalar GGML_API struct ggml_tensor * ggml_sum( struct ggml_context * ctx, @@ -1143,6 +1193,14 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + GGML_API struct ggml_tensor * ggml_exp( + struct ggml_context * ctx, + struct ggml_tensor * a); + + GGML_API struct ggml_tensor * ggml_exp_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a); + // normalize along rows GGML_API struct ggml_tensor * ggml_norm( struct ggml_context * ctx, @@ -1479,8 +1537,8 @@ extern "C" { struct ggml_tensor * b); // rotary position embedding - // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED) - // if mode & 2 == 1, GPT-NeoX style + // if (mode & 1) - skip n_past elements (NOT SUPPORTED) + // if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style // // b is an int32 vector with size a->ne[2], it contains the positions GGML_API struct ggml_tensor * ggml_rope( @@ -1590,34 +1648,49 @@ extern "C" { float min, float max); + // im2col + // converts data into a format that effectively results in a convolution when combined with matrix multiplication GGML_API struct ggml_tensor * ggml_im2col( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1, - bool is_2D, - enum ggml_type dst_type); + struct ggml_tensor * a, // convolution kernel + struct ggml_tensor * b, // data + int s0, // stride dimension 0 + int s1, // stride dimension 1 + int p0, // padding dimension 0 + int p1, // padding dimension 1 + int d0, // dilation dimension 0 + int d1, // dilation dimension 1 + bool is_2D, + enum ggml_type dst_type); + + GGML_API struct ggml_tensor * ggml_im2col_back( + struct ggml_context * ctx, + struct ggml_tensor * a, // convolution kernel + struct ggml_tensor * b, // gradient of im2col output + int64_t * ne, // shape of im2col input + int s0, // stride dimension 0 + int s1, // stride dimension 1 + int p0, // padding dimension 0 + int p1, // padding dimension 1 + int d0, // dilation dimension 0 + int d1, // dilation dimension 1 + bool is_2D); GGML_API struct ggml_tensor * ggml_conv_depthwise_2d( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1); + struct ggml_tensor * a, // convolution kernel + struct ggml_tensor * b, // data + int s0, // stride dimension 0 + int s1, // stride dimension 1 + int p0, // padding dimension 0 + int p1, // padding dimension 1 + int d0, // dilation dimension 0 + int d1); // dilation dimension 1 GGML_API struct ggml_tensor * ggml_conv_1d( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, + struct ggml_tensor * a, // convolution kernel + struct ggml_tensor * b, // data int s0, // stride int p0, // padding int d0); // dilation @@ -1626,29 +1699,29 @@ extern "C" { // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d) GGML_API struct ggml_tensor* ggml_conv_1d_ph( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s, - int d); + struct ggml_tensor * a, // convolution kernel + struct ggml_tensor * b, // data + int s, // stride + int d); // dilation GGML_API struct ggml_tensor * ggml_conv_transpose_1d( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int p0, - int d0); + struct ggml_tensor * a, // convolution kernel + struct ggml_tensor * b, // data + int s0, // stride + int p0, // padding + int d0); // dilation GGML_API struct ggml_tensor * ggml_conv_2d( struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1); + struct ggml_tensor * a, // convolution kernel + struct ggml_tensor * b, // data + int s0, // stride dimension 0 + int s1, // stride dimension 1 + int p0, // padding dimension 0 + int p1, // padding dimension 1 + int d0, // dilation dimension 0 + int d1); // dilation dimension 1 // kernel size is a->ne[0] x a->ne[1] @@ -1710,6 +1783,18 @@ extern "C" { float p0, float p1); + GGML_API struct ggml_tensor * ggml_pool_2d_back( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * af, // "a"/input used in forward pass + enum ggml_op_pool op, + int k0, + int k1, + int s0, + int s1, + float p0, + float p1); + // nearest interpolate // multiplies ne0 and ne1 by scale factor // used in stable-diffusion @@ -1784,7 +1869,8 @@ extern "C" { struct ggml_tensor * v, struct ggml_tensor * mask, float scale, - float max_bias); + float max_bias, + float logit_softcap); GGML_API void ggml_flash_attn_ext_set_prec( struct ggml_tensor * a, @@ -1801,10 +1887,8 @@ extern "C" { GGML_API struct ggml_tensor * ggml_ssm_conv( struct ggml_context * ctx, - struct ggml_tensor * s, - struct ggml_tensor * x, - struct ggml_tensor * c, - struct ggml_tensor * sq); + struct ggml_tensor * sx, + struct ggml_tensor * c); GGML_API struct ggml_tensor * ggml_ssm_scan( struct ggml_context * ctx, @@ -1813,8 +1897,7 @@ extern "C" { struct ggml_tensor * dt, struct ggml_tensor * A, struct ggml_tensor * B, - struct ggml_tensor * C, - struct ggml_tensor * sq); + struct ggml_tensor * C); // partition into non-overlapping windows with padding if needed // example: @@ -1866,6 +1949,15 @@ extern "C" { struct ggml_tensor * pw, struct ggml_tensor * ph); + GGML_API struct ggml_tensor * ggml_rwkv_wkv( + struct ggml_context * ctx, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * r, + struct ggml_tensor * tf, + struct ggml_tensor * td, + struct ggml_tensor * state); + // custom operators typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *); @@ -2036,10 +2128,23 @@ extern "C" { GGML_API size_t ggml_graph_overhead(void); GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads); + GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads); + GGML_API void ggml_threadpool_params_init (struct ggml_threadpool_params *p, int n_threads); + GGML_API bool ggml_threadpool_params_match (const struct ggml_threadpool_params *p0, const struct ggml_threadpool_params *p1); + GGML_API struct ggml_threadpool* ggml_threadpool_new (struct ggml_threadpool_params * params); + GGML_API void ggml_threadpool_free (struct ggml_threadpool * threadpool); + GGML_API int ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool); + GGML_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool); + GGML_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool); + // ggml_graph_plan() has to be called before ggml_graph_compute() // when plan.work_size > 0, caller must allocate memory for plan.work_data - GGML_API struct ggml_cplan ggml_graph_plan (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/); - GGML_API enum ggml_status ggml_graph_compute( struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); + GGML_API struct ggml_cplan ggml_graph_plan( + const struct ggml_cgraph * cgraph, + int n_threads, /* = GGML_DEFAULT_N_THREADS */ + struct ggml_threadpool * threadpool /* = NULL */ ); + GGML_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); + // same as ggml_graph_compute() but the work data is allocated as a part of the context // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data GGML_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads); diff --git a/llama/grammar-parser.cpp b/llama/grammar-parser.cpp index b20310d0..979e681a 100644 --- a/llama/grammar-parser.cpp +++ b/llama/grammar-parser.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -395,6 +395,9 @@ namespace grammar_parser { } // Validate the state to ensure that all rules are defined for (const auto & rule : state.rules) { + if (rule.empty()) { + throw std::runtime_error("Undefined rule"); + } for (const auto & elem : rule) { if (elem.type == LLAMA_GRETYPE_RULE_REF) { // Ensure that the rule at that location exists diff --git a/llama/grammar-parser.h b/llama/grammar-parser.h index 1c4e3aff..51f98664 100644 --- a/llama/grammar-parser.h +++ b/llama/grammar-parser.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/json-schema-to-grammar.cpp b/llama/json-schema-to-grammar.cpp index df453bac..9623d710 100644 --- a/llama/json-schema-to-grammar.cpp +++ b/llama/json-schema-to-grammar.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/json-schema-to-grammar.h b/llama/json-schema-to-grammar.h index 0b2fb0d7..01f56761 100644 --- a/llama/json-schema-to-grammar.h +++ b/llama/json-schema-to-grammar.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/llama-grammar.cpp b/llama/llama-grammar.cpp index 422578cc..a106428d 100644 --- a/llama/llama-grammar.cpp +++ b/llama/llama-grammar.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/llama-grammar.h b/llama/llama-grammar.h index 8d4d342a..f16cf336 100644 --- a/llama/llama-grammar.h +++ b/llama/llama-grammar.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/llama-impl.h b/llama/llama-impl.h index f34ff1d8..3d127b37 100644 --- a/llama/llama-impl.h +++ b/llama/llama-impl.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -50,3 +50,24 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void * #define LLAMA_LOG_INFO(...) llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__) #define LLAMA_LOG_WARN(...) llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__) #define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) + +// +// helpers +// + +static void replace_all(std::string & s, const std::string & search, const std::string & replace) { + if (search.empty()) { + return; + } + std::string builder; + builder.reserve(s.length()); + size_t pos = 0; + size_t last_pos = 0; + while ((pos = s.find(search, last_pos)) != std::string::npos) { + builder.append(s, last_pos, pos - last_pos); + builder.append(replace); + last_pos = pos + search.length(); + } + builder.append(s, last_pos, std::string::npos); + s = std::move(builder); +} diff --git a/llama/llama-sampling.cpp b/llama/llama-sampling.cpp index fdb3f079..8cc360d8 100644 --- a/llama/llama-sampling.cpp +++ b/llama/llama-sampling.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -111,14 +111,14 @@ void llama_sample_top_k_impl(struct llama_sampling * smpl, llama_token_data_arra constexpr float bucket_low = -10.0f; constexpr float bucket_high = 10.0f; constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low); - constexpr float bucker_inter = -bucket_low * bucket_scale; + constexpr float bucket_inter = -bucket_low * bucket_scale; std::vector bucket_idx(candidates->size); std::vector histo(nbuckets, 0); for (int i = 0; i < (int)candidates->size; ++i) { const float val = candidates->data[i].logit; - int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low); + int ib = int(bucket_scale * val + bucket_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low); ib = std::max(0, std::min(nbuckets-1, ib)); bucket_idx[i] = ib; ++histo[ib]; diff --git a/llama/llama-sampling.h b/llama/llama-sampling.h index 9c44e9b8..614987c7 100644 --- a/llama/llama-sampling.h +++ b/llama/llama-sampling.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/llama-vocab.cpp b/llama/llama-vocab.cpp index 154ab4dc..4bc61196 100644 --- a/llama/llama-vocab.cpp +++ b/llama/llama-vocab.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -42,20 +42,6 @@ // helpers // -static void replace_all(std::string & s, const std::string & search, const std::string & replace) { - std::string result; - for (size_t pos = 0; ; pos += search.length()) { - auto new_pos = s.find(search, pos); - if (new_pos == std::string::npos) { - result += s.substr(pos, s.size() - pos); - break; - } - result += s.substr(pos, new_pos - pos) + replace; - pos = new_pos; - } - s = std::move(result); -} - LLAMA_ATTRIBUTE_FORMAT(1, 2) static std::string format(const char * fmt, ...) { va_list ap; @@ -98,17 +84,17 @@ struct naive_trie { auto res = children.find(c); if (res != children.end()) { return res->second.get_longest_prefix(key, len, offset + 1); - } else { - return std::make_pair(key, offset); } + + return std::make_pair(key, offset); } - struct naive_trie * traverse(const char c) { + const struct naive_trie * traverse(const char c) const { auto res = children.find(c); if (res != children.end()) { return &res->second; - } else { - return NULL; } + + return NULL; } std::map children; bool has_value; @@ -361,6 +347,21 @@ private: // TODO: there are a lot of common parts between spm and bpe tokenizers, should be refactored and reused +template, typename Compare = std::less> +class llama_priority_queue : public std::priority_queue { +public: + using std::priority_queue::priority_queue; + + T pop_move() { + T item = std::move(this->c.front()); + std::pop_heap(this->c.begin(), this->c.end(), this->comp); + this->c.pop_back(); + return item; + } + + void pop() = delete; +}; + struct llm_bigram_bpe { struct comparator { bool operator()(const llm_bigram_bpe & l, const llm_bigram_bpe & r) const { @@ -369,7 +370,7 @@ struct llm_bigram_bpe { }; using queue_storage = std::vector; - using queue = std::priority_queue; + using queue = llama_priority_queue; llm_symbol::index left; llm_symbol::index right; std::string text; @@ -428,6 +429,7 @@ struct llm_tokenizer_bpe { case LLAMA_VOCAB_PRE_TYPE_COMMAND_R: case LLAMA_VOCAB_PRE_TYPE_SMOLLM: case LLAMA_VOCAB_PRE_TYPE_CODESHELL: + case LLAMA_VOCAB_PRE_TYPE_EXAONE: regex_exprs = { "\\p{N}", "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", @@ -450,6 +452,8 @@ struct llm_tokenizer_bpe { }; break; case LLAMA_VOCAB_PRE_TYPE_PORO: + case LLAMA_VOCAB_PRE_TYPE_BLOOM: + case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH: regex_exprs = { " ?[^(\\s|.,!?…。，、।۔،)]+", }; @@ -557,8 +561,7 @@ struct llm_tokenizer_bpe { // build token(s) while (!work_queue.empty()) { - auto bigram = work_queue.top(); - work_queue.pop(); + auto bigram = work_queue.pop_move(); auto & left_symbol = symbols[bigram.left]; auto & right_symbol = symbols[bigram.right]; @@ -866,7 +869,7 @@ struct llm_tokenizer_ugm { // traverse the token matcher trie to find a matching token bool single_codepoint_token_found = false; const struct best_tokenization & current_best = tokenization_results[input_offset]; - struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]); + const struct naive_trie * node = token_matcher.traverse(normalized[prefix_offset++]); while (prefix_offset <= input_len && node != NULL) { // check if we found valid token in prefix @@ -986,7 +989,7 @@ private: /* * This structure is a view wrapper for XOR-compressed double array (XCDA) * See Shunsuke Kanda (2018). Space- and Time-Efficient String Dictionaries. - * Eeach bit-packed entry contains: + * Each bit-packed entry contains: * - BASE array value in bits 10-30 * - LCHECK array value in bits 0-7 * - LEAF array value in bit 9 @@ -1120,6 +1123,111 @@ private: struct naive_trie token_matcher; }; +// +// RWKV tokenizer +// + +static std::vector llama_unescape_rwkv_token(const std::string & escaped) { + std::vector output; + output.reserve(escaped.size()); + + // Parser state + bool escaping = false; + uint8_t hex_remaining = 0; + uint8_t hex_acc = 0; + + // Step through characters, performing parsing + for (const char & c : escaped) { + // If we're parsing a hex code, interpret the next character + if (hex_remaining != 0) { + uint8_t value = (c >= 'a') ? (c - 'a' + 10) : (c - '0'); + hex_acc = (hex_acc << 4) + value; + + hex_remaining -= 1; + if (hex_remaining == 0) { + output.push_back(hex_acc); + hex_acc = 0; + } + + continue; + } + + // If we got an escape character, interpret it + if (escaping) { + if (c == 't') { + output.push_back('\t'); + } else if (c == 'n') { + output.push_back('\n'); + } else if (c == 'r') { + output.push_back('\r'); + } else if (c == 'x') { + hex_remaining = 2; + } else { + output.push_back(c); + } + + escaping = false; + continue; + } + + if (c == '\\') { + escaping = true; + continue; + } + + output.push_back(c); + } + + return output; +} + +struct llm_tokenizer_rwkv { + llm_tokenizer_rwkv(const llama_vocab & vocab): vocab(vocab) { + // RWKV supports arbitrary byte tokens, but the vocab struct only supports string tokens. + // For now, we decode the vocab here into the lookup we'll use for tokenization. + + // build trie + for (unsigned int id = 0; id < vocab.id_to_token.size(); ++id) { + const auto & token = vocab.id_to_token[id]; + const auto data = llama_unescape_rwkv_token(token.text); + token_matcher.insert((const char *) data.data(), data.size(), id); + } + } + + void tokenize(const std::string & text, std::vector & output) { + uint32_t position = 0; + + while (position < text.size()) { + const struct naive_trie * node = token_matcher.traverse(text[position]); + if (node == NULL) { + // no matching token found, add unknown token + output.push_back(vocab.special_unk_id); + position += 1; + continue; + } + + // traverse the trie to find the longest matching token + uint32_t token_id = 0; + uint32_t token_length = 0; + while (node != NULL) { + if (node->has_value) { + token_id = node->value; + token_length = position + 1; + } + node = node->traverse(text[++position]); + } + + // add the longest matching token + output.push_back(token_id); + position = token_length; + } + } + + const llama_vocab & vocab; + + struct naive_trie token_matcher; +}; + // // (de-) tokenize // @@ -1424,6 +1532,23 @@ std::vector llama_tokenize_internal(const llama_vocab & vocab, output.push_back(vocab.special_eos_id); } } break; + case LLAMA_VOCAB_TYPE_RWKV: + { + for (const auto & fragment : fragment_buffer) { + if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + +#ifdef PRETOKENIZERDEBUG + LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); +#endif + + llm_tokenizer_rwkv tokenizer(vocab); + tokenizer.tokenize(raw_text, output); + } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) + output.push_back(fragment.token); + } + } + } break; case LLAMA_VOCAB_TYPE_NONE: GGML_ABORT("fatal error"); } @@ -1506,11 +1631,11 @@ llama_token llama_token_pad_impl(const struct llama_vocab & vocab) { return vocab.special_pad_id; } -int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab) { +bool llama_add_bos_token_impl(const struct llama_vocab & vocab) { return vocab.tokenizer_add_bos; } -int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab) { +bool llama_add_eos_token_impl(const struct llama_vocab & vocab) { return vocab.tokenizer_add_eos; } @@ -1639,6 +1764,17 @@ int32_t llama_token_to_piece_impl(const struct llama_vocab & vocab, llama_token } break; } + case LLAMA_VOCAB_TYPE_RWKV: { + std::vector result = llama_unescape_rwkv_token(token_text); + + // If we don't have enough space, return an error + if (result.size() > (size_t)length) { + return -(int)result.size(); + } + + memcpy(buf, result.data(), result.size()); + return (int)result.size(); + } default: GGML_ABORT("fatal error"); } diff --git a/llama/llama-vocab.h b/llama/llama-vocab.h index 2b7f5d0c..3b935909 100644 --- a/llama/llama-vocab.h +++ b/llama/llama-vocab.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -121,8 +121,8 @@ llama_token llama_token_sep_impl(const struct llama_vocab & vocab); llama_token llama_token_nl_impl (const struct llama_vocab & vocab); llama_token llama_token_pad_impl(const struct llama_vocab & vocab); -int32_t llama_add_bos_token_impl(const struct llama_vocab & vocab); -int32_t llama_add_eos_token_impl(const struct llama_vocab & vocab); +bool llama_add_bos_token_impl(const struct llama_vocab & vocab); +bool llama_add_eos_token_impl(const struct llama_vocab & vocab); llama_token llama_token_prefix_impl(const struct llama_vocab & vocab); llama_token llama_token_middle_impl(const struct llama_vocab & vocab); diff --git a/llama/llama.cpp b/llama/llama.cpp index 4180bffe..0f666c7f 100644 --- a/llama/llama.cpp +++ b/llama/llama.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -147,17 +147,6 @@ static std::string trim(const std::string & str) { return str.substr(start, end - start); } -static void replace_all(std::string & s, const std::string & search, const std::string & replace) { - if (search.empty()) { - return; // Avoid infinite loop if 'search' is an empty string - } - size_t pos = 0; - while ((pos = s.find(search, pos)) != std::string::npos) { - s.replace(pos, search.length(), replace); - pos += replace.length(); - } -} - static bool is_float_close(float a, float b, float abs_tol) { // Check for non-negative tolerance if (abs_tol < 0.0) { @@ -245,7 +234,11 @@ enum llm_arch { LLM_ARCH_CHATGLM, LLM_ARCH_BITNET, LLM_ARCH_T5, + LLM_ARCH_T5ENCODER, LLM_ARCH_JAIS, + LLM_ARCH_NEMOTRON, + LLM_ARCH_EXAONE, + LLM_ARCH_RWKV6, LLM_ARCH_UNKNOWN, }; @@ -289,7 +282,11 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_CHATGLM, "chatglm" }, { LLM_ARCH_BITNET, "bitnet" }, { LLM_ARCH_T5, "t5" }, + { LLM_ARCH_T5ENCODER, "t5encoder" }, { LLM_ARCH_JAIS, "jais" }, + { LLM_ARCH_NEMOTRON, "nemotron" }, + { LLM_ARCH_EXAONE, "exaone" }, + { LLM_ARCH_RWKV6, "rwkv6" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -326,6 +323,9 @@ enum llm_kv { LLM_KV_DECODER_START_TOKEN_ID, LLM_KV_ATTN_LOGIT_SOFTCAPPING, LLM_KV_FINAL_LOGIT_SOFTCAPPING, + LLM_KV_RESCALE_EVERY_N_LAYERS, + LLM_KV_TIME_MIX_EXTRA_DIM, + LLM_KV_TIME_DECAY_EXTRA_DIM, LLM_KV_ATTENTION_HEAD_COUNT, LLM_KV_ATTENTION_HEAD_COUNT_KV, @@ -359,6 +359,9 @@ enum llm_kv { LLM_KV_SSM_CONV_KERNEL, LLM_KV_SSM_STATE_SIZE, LLM_KV_SSM_TIME_STEP_RANK, + LLM_KV_SSM_DT_B_C_RMS, + + LLM_KV_WKV_HEAD_SIZE, LLM_KV_TOKENIZER_MODEL, LLM_KV_TOKENIZER_PRE, @@ -419,11 +422,14 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" }, { LLM_KV_EXPERT_SHARED_COUNT, "%s.expert_shared_count" }, { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" }, - { LLM_KV_POOLING_TYPE , "%s.pooling_type" }, + { LLM_KV_POOLING_TYPE, "%s.pooling_type" }, { LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" }, { LLM_KV_ATTN_LOGIT_SOFTCAPPING, "%s.attn_logit_softcapping" }, { LLM_KV_FINAL_LOGIT_SOFTCAPPING, "%s.final_logit_softcapping" }, + { LLM_KV_RESCALE_EVERY_N_LAYERS, "%s.rescale_every_n_layers" }, + { LLM_KV_TIME_MIX_EXTRA_DIM, "%s.time_mix_extra_dim" }, + { LLM_KV_TIME_DECAY_EXTRA_DIM, "%s.time_decay_extra_dim" }, { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, @@ -457,6 +463,9 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" }, { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" }, { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" }, + { LLM_KV_SSM_DT_B_C_RMS, "%s.ssm.dt_b_c_rms" }, + + { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" }, { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, { LLM_KV_TOKENIZER_PRE, "tokenizer.ggml.pre" }, @@ -547,6 +556,29 @@ enum llm_tensor { LLM_TENSOR_SSM_A, LLM_TENSOR_SSM_D, LLM_TENSOR_SSM_OUT, + LLM_TENSOR_TIME_MIX_W1, + LLM_TENSOR_TIME_MIX_W2, + LLM_TENSOR_TIME_MIX_LERP_X, + LLM_TENSOR_TIME_MIX_LERP_W, + LLM_TENSOR_TIME_MIX_LERP_K, + LLM_TENSOR_TIME_MIX_LERP_V, + LLM_TENSOR_TIME_MIX_LERP_R, + LLM_TENSOR_TIME_MIX_LERP_G, + LLM_TENSOR_TIME_MIX_FIRST, + LLM_TENSOR_TIME_MIX_DECAY, + LLM_TENSOR_TIME_MIX_DECAY_W1, + LLM_TENSOR_TIME_MIX_DECAY_W2, + LLM_TENSOR_TIME_MIX_KEY, + LLM_TENSOR_TIME_MIX_VALUE, + LLM_TENSOR_TIME_MIX_RECEPTANCE, + LLM_TENSOR_TIME_MIX_GATE, + LLM_TENSOR_TIME_MIX_LN, + LLM_TENSOR_TIME_MIX_OUTPUT, + LLM_TENSOR_CHANNEL_MIX_LERP_K, + LLM_TENSOR_CHANNEL_MIX_LERP_R, + LLM_TENSOR_CHANNEL_MIX_KEY, + LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, + LLM_TENSOR_CHANNEL_MIX_VALUE, LLM_TENSOR_ATTN_Q_A, LLM_TENSOR_ATTN_Q_B, LLM_TENSOR_ATTN_KV_A_MQA, @@ -1298,6 +1330,24 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_T5ENCODER, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" }, + { LLM_TENSOR_ENC_ATTN_NORM, "enc.blk.%d.attn_norm" }, + { LLM_TENSOR_ENC_ATTN_Q, "enc.blk.%d.attn_q" }, + { LLM_TENSOR_ENC_ATTN_K, "enc.blk.%d.attn_k" }, + { LLM_TENSOR_ENC_ATTN_V, "enc.blk.%d.attn_v" }, + { LLM_TENSOR_ENC_ATTN_OUT, "enc.blk.%d.attn_o" }, + { LLM_TENSOR_ENC_ATTN_REL_B, "enc.blk.%d.attn_rel_b" }, + { LLM_TENSOR_ENC_FFN_NORM, "enc.blk.%d.ffn_norm" }, + { LLM_TENSOR_ENC_FFN_GATE, "enc.blk.%d.ffn_gate" }, + { LLM_TENSOR_ENC_FFN_DOWN, "enc.blk.%d.ffn_down" }, + { LLM_TENSOR_ENC_FFN_UP, "enc.blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_JAIS, { @@ -1313,6 +1363,77 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, }, }, + { + LLM_ARCH_NEMOTRON, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_EXAONE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_RWKV6, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, + { LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" }, + { LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" }, + { LLM_TENSOR_TIME_MIX_LERP_X, "blk.%d.time_mix_lerp_x" }, + { LLM_TENSOR_TIME_MIX_LERP_W, "blk.%d.time_mix_lerp_w" }, + { LLM_TENSOR_TIME_MIX_LERP_K, "blk.%d.time_mix_lerp_k" }, + { LLM_TENSOR_TIME_MIX_LERP_V, "blk.%d.time_mix_lerp_v" }, + { LLM_TENSOR_TIME_MIX_LERP_R, "blk.%d.time_mix_lerp_r" }, + { LLM_TENSOR_TIME_MIX_LERP_G, "blk.%d.time_mix_lerp_g" }, + { LLM_TENSOR_TIME_MIX_FIRST, "blk.%d.time_mix_first" }, + { LLM_TENSOR_TIME_MIX_DECAY, "blk.%d.time_mix_decay" }, + { LLM_TENSOR_TIME_MIX_DECAY_W1, "blk.%d.time_mix_decay_w1" }, + { LLM_TENSOR_TIME_MIX_DECAY_W2, "blk.%d.time_mix_decay_w2" }, + { LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" }, + { LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" }, + { LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" }, + { LLM_TENSOR_TIME_MIX_GATE, "blk.%d.time_mix_gate" }, + { LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" }, + { LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_output" }, + { LLM_TENSOR_CHANNEL_MIX_LERP_K, "blk.%d.channel_mix_lerp_k" }, + { LLM_TENSOR_CHANNEL_MIX_LERP_R, "blk.%d.channel_mix_lerp_r" }, + { LLM_TENSOR_CHANNEL_MIX_KEY, "blk.%d.channel_mix_key" }, + { LLM_TENSOR_CHANNEL_MIX_VALUE, "blk.%d.channel_mix_value" }, + { LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, "blk.%d.channel_mix_receptance" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -2125,6 +2246,7 @@ enum e_model { MODEL_1B, MODEL_1_3B, MODEL_1_4B, + MODEL_1_6B, MODEL_2B, MODEL_2_8B, MODEL_3B, @@ -2202,6 +2324,12 @@ struct llama_hparams { float f_attn_logit_softcapping = 50.0f; float f_final_logit_softcapping = 30.0f; + // for RWKV + uint32_t rescale_every_n_layers = 0; + uint32_t time_mix_extra_dim = 0; + uint32_t time_decay_extra_dim = 0; + uint32_t wkv_head_size = 0; + float rope_attn_factor = 1.0f; float rope_freq_base_train; float rope_freq_scale_train; @@ -2213,6 +2341,7 @@ struct llama_hparams { uint32_t ssm_d_inner = 0; uint32_t ssm_d_state = 0; uint32_t ssm_dt_rank = 0; + bool ssm_dt_b_c_rms = false; float f_clamp_kqv = 0.0f; float f_max_alibi_bias = 0.0f; @@ -2262,6 +2391,12 @@ struct llama_hparams { if (this->ssm_d_inner != other.ssm_d_inner) return true; if (this->ssm_d_state != other.ssm_d_state) return true; if (this->ssm_dt_rank != other.ssm_dt_rank) return true; + if (this->ssm_dt_b_c_rms != other.ssm_dt_b_c_rms) return true; + + if (this->rescale_every_n_layers != other.rescale_every_n_layers) return true; + if (this->time_mix_extra_dim != other.time_mix_extra_dim) return true; + if (this->time_decay_extra_dim != other.time_decay_extra_dim) return true; + if (this->wkv_head_size != other.wkv_head_size) return true; if (this->dec_start_token_id != other.dec_start_token_id) return true; @@ -2326,15 +2461,25 @@ struct llama_hparams { } uint32_t n_embd_k_s() const { // dimension of the rolling state embeddings - // corresponds to Mamba's conv_states size - // TODO: maybe support other convolution strides than 1 - // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed - return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner; + // corresponds to Mamba's conv_states size or RWKV's token_shift states size + if (wkv_head_size != 0) { + // for RWKV models + return 2 * n_embd; + } else { + // TODO: maybe support other convolution strides than 1 + // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed + return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner; + } } uint32_t n_embd_v_s() const { // dimension of the recurrent state embeddings - // corresponds to Mamba's ssm_states size - return ssm_d_state * ssm_d_inner; + if (wkv_head_size != 0) { + // corresponds to RWKV's wkv_states size + return n_embd * wkv_head_size; + } else { + // corresponds to Mamba's ssm_states size + return ssm_d_state * ssm_d_inner; + } } }; @@ -2345,8 +2490,8 @@ struct llama_cparams { uint32_t n_batch; uint32_t n_ubatch; uint32_t n_seq_max; - uint32_t n_threads; // number of threads to use for generation - uint32_t n_threads_batch; // number of threads to use for batch processing + int n_threads; // number of threads to use for generation + int n_threads_batch; // number of threads to use for batch processing float rope_freq_base; float rope_freq_scale; @@ -2473,6 +2618,36 @@ struct llama_layer { struct ggml_tensor * ssm_conv1d_b; struct ggml_tensor * ssm_dt_b; + // rwkv + struct ggml_tensor * time_mix_w1; + struct ggml_tensor * time_mix_w2; + struct ggml_tensor * time_mix_lerp_x; + struct ggml_tensor * time_mix_lerp_w; + struct ggml_tensor * time_mix_lerp_k; + struct ggml_tensor * time_mix_lerp_v; + struct ggml_tensor * time_mix_lerp_r; + struct ggml_tensor * time_mix_lerp_g; + + struct ggml_tensor * time_mix_first; + struct ggml_tensor * time_mix_decay; + struct ggml_tensor * time_mix_decay_w1; + struct ggml_tensor * time_mix_decay_w2; + struct ggml_tensor * time_mix_key; + struct ggml_tensor * time_mix_value; + struct ggml_tensor * time_mix_receptance; + struct ggml_tensor * time_mix_gate; + + struct ggml_tensor * time_mix_ln; + struct ggml_tensor * time_mix_ln_b; + struct ggml_tensor * time_mix_output; + + struct ggml_tensor * channel_mix_lerp_k; + struct ggml_tensor * channel_mix_lerp_r; + + struct ggml_tensor * channel_mix_key; + struct ggml_tensor * channel_mix_receptance; + struct ggml_tensor * channel_mix_value; + // long rope factors struct ggml_tensor * rope_long = nullptr; struct ggml_tensor * rope_short = nullptr; @@ -2488,10 +2663,29 @@ struct llama_layer { struct ggml_tensor * ffn_down_scale; }; +// very similar to llama_batch, +// but has more metadata about sequences +struct llama_ubatch { + bool equal_seqs; + // TODO: whole_seqs for embeddings? + + uint32_t n_tokens; // total tokens (n_seq_tokens * n_seqs) + uint32_t n_seq_tokens; // tokens per sequence + uint32_t n_seqs; + + llama_token * token; // [n_tokens] + float * embd; // [n_embd, n_tokens] + llama_pos * pos; // [n_tokens] + int32_t * n_seq_id; // [n_seqs] + llama_seq_id ** seq_id; // [n_seqs] + int8_t * output; // [n_tokens] +}; + struct llama_kv_cell { llama_pos pos = -1; llama_pos delta = 0; - int32_t src = 0; // used by recurrent state models to copy states + int32_t src = -1; // used by recurrent state models to copy states + int32_t tail = -1; std::set seq_id; @@ -2512,7 +2706,6 @@ struct llama_kv_cell { struct llama_kv_cache { bool has_shift = false; bool do_defrag = false; - bool do_copy = false; bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token bool v_trans = true; // the value tensor is transposed @@ -2675,6 +2868,340 @@ struct llama_model { } }; +struct llama_sbatch_seq { + int32_t n_seq_id; + llama_seq_id * seq_id; + size_t offset; + size_t length; + + // helper for smoother batch API transition -- can be deprecated in the future + llama_seq_id all_seq_id; // used if seq_id == NULL +}; + +// sequence-length-aware batch splitting +struct llama_sbatch { + // tokens left in this batch + size_t n_tokens; + + size_t n_embd; + + bool logits_all; // TODO: remove once lctx.logits_all is removed too + + // sorted indices into the batch + std::vector ids; + // batch indices of the output + std::vector out_ids; + std::vector seq; + const llama_batch * batch = nullptr; + + // buffers for the ubatch + std::vector ubatch_token; + std::vector ubatch_embd; + std::vector ubatch_pos; + std::vector ubatch_n_seq_id; + std::vector ubatch_seq_id; + std::vector ubatch_output; + + llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false) { + // clear empty sequences + // the previous ubatch is assumed to be gone, + // so nothing should refer to values in these sequences anymore. + for (size_t i = seq.size(); i-- > 0;) { + if (seq[i].length == 0) { + seq.pop_back(); + } else { + break; + } + } + ubatch_token.resize(!has_embd ? n_ubatch : 0); + ubatch_embd.resize(has_embd ? n_embd * n_ubatch : 0); + ubatch_pos.resize(n_ubatch); + ubatch_n_seq_id.resize(n_ubatch); + ubatch_seq_id.resize(n_ubatch); + ubatch_output.resize(n_ubatch); + llama_ubatch ubatch = { + /*equal_seqs =*/ true, + /*n_tokens =*/ 0, + /*n_seq_tokens =*/ 0, + /*n_seqs =*/ 0, + /*token =*/ !has_embd ? ubatch_token.data() : nullptr, + /*embd =*/ has_embd ? ubatch_embd.data() : nullptr, + /*pos =*/ ubatch_pos.data(), + /*n_seq_id =*/ ubatch_n_seq_id.data(), + /*seq_id =*/ ubatch_seq_id.data(), + /*output =*/ ubatch_output.data(), + }; + return ubatch; + } + + void add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length) { + GGML_ASSERT(batch != nullptr); + GGML_ASSERT(length <= seq.length); + // Can only add sequences of equal lengths to a batch, + // otherwise it isn't clear to which sequence a token belongs + GGML_ASSERT(seq.n_seq_id == 0 || ubatch.n_seqs == 0 || length == (size_t) ubatch.n_tokens / ubatch.n_seqs); + GGML_ASSERT((seq.n_seq_id != 0) == ubatch.equal_seqs); + // NOTE: loops are separated for cache-friendliness + if (batch->token) { + if (ubatch.equal_seqs) { + for (size_t i = 0; i < length; ++i) { + ubatch.token[ubatch.n_tokens + i] = batch->token[ids[seq.offset + i]]; + } + } else { + // simple split + ubatch.token = batch->token + seq.offset; + } + } else { + ubatch.token = nullptr; + } + if (batch->embd) { + if (ubatch.equal_seqs) { + for (size_t i = 0; i < length; ++i) { + memcpy( + ubatch.embd + n_embd * (ubatch.n_tokens + i), + batch->embd + n_embd * ids[seq.offset + i], + n_embd * sizeof(float) + ); + } + } else { + // simple split + ubatch.embd = batch->embd + (n_embd * seq.offset); + } + } else { + ubatch.embd = nullptr; + } + // from here on, the else branches are deprecated; + // they are helpers for smoother batch API transition + if (batch->pos) { + if (ubatch.equal_seqs) { + for (size_t i = 0; i < length; ++i) { + ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]]; + } + } else { + // simple split + ubatch.pos = batch->pos + seq.offset; + } + } else { + for (size_t i = 0; i < length; ++i) { + llama_pos bi = ids[seq.offset + i]; + ubatch.pos[ubatch.n_tokens + i] = batch->all_pos_0 + (bi * batch->all_pos_1); + } + } + if (ubatch.equal_seqs) { + ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id; + if (seq.seq_id) { + ubatch.seq_id[ubatch.n_seqs] = seq.seq_id; + } else { + GGML_ASSERT(seq.n_seq_id == 1); + ubatch.seq_id[ubatch.n_seqs] = &seq.all_seq_id; + } + } else { + // simple split + if (batch->n_seq_id) { + for (size_t i = 0; i < length; ++i) { + ubatch.n_seq_id = batch->n_seq_id + seq.offset; + } + } else { + for (size_t i = 0; i < length; ++i) { + ubatch.n_seq_id[ubatch.n_seqs + i] = 1; + } + } + if (batch->seq_id) { + for (size_t i = 0; i < length; ++i) { + ubatch.seq_id = batch->seq_id + seq.offset; + } + } else { + for (size_t i = 0; i < length; ++i) { + ubatch.seq_id[ubatch.n_seqs + i] = &seq.all_seq_id; + } + } + } + if (logits_all) { + for (size_t i = 0; i < length; ++i) { + ubatch.output[ubatch.n_tokens + i] = 1; + out_ids.push_back(ids[seq.offset + i]); + } + } else if (batch->logits) { + if (ubatch.equal_seqs) { + for (size_t i = 0; i < length; ++i) { + size_t id = ids[seq.offset + i]; + int8_t is_output = batch->logits[id]; + ubatch.output[ubatch.n_tokens + i] = is_output; + if (is_output) { out_ids.push_back(id); } + } + } else { + // simple split + ubatch.output = batch->logits + seq.offset; + for (size_t i = 0; i < length; ++i) { + if (ubatch.output[i] != 0) { out_ids.push_back(seq.offset + i); } + } + } + } else { + // only get last output + for (size_t i = 0; i < length; ++i) { + size_t id = ids[seq.offset + i]; + int8_t is_last = id == ids.size() - 1; + ubatch.output[ubatch.n_tokens + i] = is_last; + if (is_last) { out_ids.push_back(id); } + } + } + if (ubatch.n_tokens == 0 && ubatch.n_seqs == 0) { + ubatch.n_seq_tokens = ubatch.equal_seqs ? length : 1; + } + ubatch.n_tokens += length; + ubatch.n_seqs += ubatch.equal_seqs ? 1 : length; // virtual sequences for simple splits + seq.offset += length; + seq.length -= length; + n_tokens -= length; + GGML_ASSERT(ubatch.n_tokens == ubatch.n_seq_tokens * ubatch.n_seqs); + } + + // simple split, unknown number of sequences of unequal lengths + llama_ubatch split_simple(size_t n_ubatch) { + n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; + llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); + ubatch.equal_seqs = false; + if (!seq.empty()) { + llama_sbatch_seq & s = seq[0]; + size_t length = s.length < n_ubatch ? s.length : n_ubatch; + GGML_ASSERT(seq.size() == 1 && s.n_seq_id == 0); // don't mix with other splits + add_seq_to_ubatch(ubatch, s, length); + } + return ubatch; + } + + // make batches of equal-length sequences + llama_ubatch split_equal(size_t n_ubatch) { + n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; + llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); + if (!seq.empty()) { + size_t length = 0; + size_t n_tokens_in_ubatch = 0; + GGML_ASSERT(seq[0].n_seq_id > 0); // should not be mixed with simple splits + // smallest first, because it's easier to split this way; + // starting from the end to pop in constant time. + for (size_t i = seq.size(); i-- > 0;) { + llama_sbatch_seq & s = seq[i]; + GGML_ASSERT(s.length > 0); + if (length == 0) { + length = s.length < n_ubatch ? s.length : n_ubatch; + } + add_seq_to_ubatch(ubatch, s, length); + n_tokens_in_ubatch += length; + // shared prompts can't be mixed with any of their sequences, + // so it's safer to compute them in their own ubatch + if (s.n_seq_id > 1) { break; } + // stop when there isn't enough space for another sequence + if (length + n_tokens_in_ubatch > n_ubatch) { break; } + } + } + return ubatch; + } + + // sequence-wise split + llama_ubatch split_seq(size_t n_ubatch) { + n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; + llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); + if (!seq.empty()) { + llama_sbatch_seq & s = seq[seq.size() - 1]; + size_t length = s.length < n_ubatch ? s.length : n_ubatch; + GGML_ASSERT(s.n_seq_id > 0); // should not be mixed with simple splits + add_seq_to_ubatch(ubatch, s, length); + } + return ubatch; + } + + void from_batch(const llama_batch & batch, const size_t n_embd, const bool simple_split = false, const bool logits_all = false) { + GGML_ASSERT(batch.n_tokens >= 0); + this->batch = &batch; + this->n_embd = n_embd; + this->logits_all = logits_all; + + n_tokens = batch.n_tokens; + ids.resize(n_tokens); + out_ids.clear(); + // TODO: reserve out_ids and seq + + for (size_t i = 0; i < n_tokens; ++i) { + ids[i] = i; + } + if (simple_split) { + seq.resize(1); + llama_sbatch_seq & s = seq[0]; + s.n_seq_id = 0; + s.seq_id = nullptr; + s.offset = 0; + s.length = n_tokens; + s.all_seq_id = batch.all_seq_id; + return; + } + std::sort(ids.begin(), ids.end(), + [&batch](size_t a, size_t b) { + int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1; + int32_t n_seq_b = batch.n_seq_id ? batch.n_seq_id[b] : 1; + // sort by seq_id, then by pos + if (n_seq_a == n_seq_b) { + if (batch.seq_id) { + for (int32_t i = 0; i < n_seq_a; ++i) { + llama_seq_id seq_id_a = batch.seq_id[a][i]; + llama_seq_id seq_id_b = batch.seq_id[b][i]; + // smaller seq_ids go first + if (seq_id_a != seq_id_b) { + return seq_id_a < seq_id_b; + } + } + } + // when all else is equal, sort by pos + if (batch.pos) { + return batch.pos[a] < batch.pos[b]; + } + // no pos, sort by id (assuming batch.all_pos_1 is positive) + return a < b; + } + // shared prompts go first + return n_seq_a > n_seq_b; + } + ); + // init seq + llama_sbatch_seq * last_seq = nullptr; + + if (batch.n_seq_id != nullptr && batch.seq_id != nullptr) { + for (size_t i = 0; i < n_tokens; ++i) { + const size_t bi = ids[i]; + const int32_t n_seqs = batch.n_seq_id[bi]; + llama_seq_id * seq_ids = batch.seq_id[bi]; + if (last_seq != nullptr) { + bool same = n_seqs == last_seq->n_seq_id; + for (int32_t j = 0; same && j < n_seqs; ++j) { + if (seq_ids[j] != last_seq->seq_id[j]) { + same = false; + } + } + if (same) { + last_seq->length += 1; + continue; + } + } + llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1, batch.all_seq_id}; + seq.push_back(new_seq); + last_seq = &seq.back(); + } + } else { + llama_sbatch_seq new_seq = {1, nullptr, 0, n_tokens, batch.all_seq_id}; + seq.push_back(new_seq); + } + // keep shared prompts first at the end, then sort by length descending. + std::sort(seq.begin(), seq.end(), + [](llama_sbatch_seq & a, llama_sbatch_seq & b) { + if (a.n_seq_id == b.n_seq_id) { + return a.length > b.length; + } + return a.n_seq_id < b.n_seq_id; + } + ); + } +}; + struct llama_context { llama_context(const llama_model & model) : model(model) @@ -2696,6 +3223,7 @@ struct llama_context { struct llama_cparams cparams; struct llama_sampling sampling; + struct llama_sbatch sbatch; struct llama_kv_cache kv_self; struct llama_control_vector cvec; @@ -2710,6 +3238,9 @@ struct llama_context { #endif ggml_backend_t backend_cpu = nullptr; + ggml_threadpool_t threadpool = nullptr; + ggml_threadpool_t threadpool_batch = nullptr; + bool has_evaluated_once = false; int64_t t_start_us; @@ -2956,8 +3487,7 @@ static bool llama_kv_cache_init( cache.has_shift = false; - // TODO: find a nicer way to add other recurrent model architectures - cache.recurrent = model.arch == LLM_ARCH_MAMBA; + cache.recurrent = llama_model_is_recurrent(&model); cache.v_trans = !cache.recurrent && !cparams.flash_attn; cache.head = 0; @@ -2970,13 +3500,6 @@ static bool llama_kv_cache_init( cache.cells.clear(); cache.cells.resize(kv_size); - if (cache.recurrent) { - // init state copy sources - for (uint32_t i = 0; i < cache.size; ++i) { - cache.cells[i].src = i; - } - } - // count used buffer types std::map buft_layer_count; if (offload) { @@ -3044,45 +3567,161 @@ static bool llama_kv_cache_init( // to the first cell of the slot. static bool llama_kv_cache_find_slot( struct llama_kv_cache & cache, - const struct llama_batch & batch) { + const struct llama_ubatch & batch) { const uint32_t n_tokens = batch.n_tokens; + const uint32_t n_seqs = batch.n_seqs; + const uint32_t n_seq_tokens = batch.n_seq_tokens; if (cache.recurrent) { - // For recurrent state architectures (like Mamba), - // each KV cache cell can store the state for a whole sequence. + // For recurrent state architectures (like Mamba or RWKV), + // each cache cell can store the state for a whole sequence. + // A slot should be always be contiguous. - llama_seq_id min = cache.size - 1; - llama_seq_id max = 0; + // can only process batches with an equal number of new tokens in each sequence + GGML_ASSERT(batch.equal_seqs); - for (uint32_t i = 0; i < n_tokens; ++i) { - for (int32_t j = 0; j < batch.n_seq_id[i]; ++j) { - llama_seq_id seq_id = batch.seq_id[i][j]; - // make sure it's a valid seq_id - if ((uint32_t) seq_id < cache.size) { - if (seq_id > max) { - max = seq_id; - } - if (seq_id < min) { - min = seq_id; - } - // Assuming the tokens are in-order - if (batch.pos[i] != cache.cells[seq_id].pos + 1) { - // What should happen when the pos backtracks or skips a value? - // Clearing the state mid-batch would require special-casing which isn't done. - LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d\n", - __func__, batch.pos[i], cache.cells[seq_id].pos, seq_id); - } - if (cache.cells[seq_id].pos < 0 && 0 <= batch.pos[i]) { - cache.used += 1; - } - cache.cells[seq_id].pos = batch.pos[i]; - // NOTE: seq_ids are not inserted here; they are handled when the input tensors are set - } else { + int32_t min = cache.size - 1; + int32_t max = 0; + + // everything should fit if all seq_ids are smaller than the max + for (uint32_t s = 0; s < n_seqs; ++s) { + const uint32_t n_seq_id = batch.n_seq_id[s]; + for (uint32_t j = 0; j < n_seq_id; ++j) { + const llama_seq_id seq_id = batch.seq_id[s][j]; + + if (seq_id < 0 || (uint32_t) seq_id >= cache.size) { // too big seq_id - // TODO: would it be possible to resize the KV cache size instead? - LLAMA_LOG_ERROR("%s: seq_id=%d >= kv_size=%d Try using a bigger --parallel value\n", __func__, seq_id, cache.size); + // TODO: would it be possible to resize the cache instead? + LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, cache.size); return false; } + if (j > 0) { + llama_kv_cell & seq = cache.cells[seq_id]; + if (seq.tail >= 0) { + llama_kv_cell & cell = cache.cells[seq.tail]; + // clear cells from seq_ids that become shared + // (should not normally happen, but let's handle it anyway) + cell.seq_id.erase(seq_id); + seq.tail = -1; + if (cell.seq_id.empty()) { + cell.pos = -1; + cell.src = -1; + cache.used -= 1; + } + } + } + } + } + +#ifndef NDEBUG + { + std::vector tails_verif; + tails_verif.assign(cache.size, -1); + for (uint32_t i = 0; i < cache.size; ++i) { + llama_kv_cell & cell = cache.cells[i]; + for (llama_seq_id seq_id : cell.seq_id) { + if (tails_verif[seq_id] != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]); + } + tails_verif[seq_id] = i; + } + } + for (uint32_t i = 0; i < cache.size; ++i) { + if (tails_verif[i] != cache.cells[i].tail) { + LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cache.cells[i].tail, tails_verif[i]); + } + } + } +#endif + + // find next empty cell + uint32_t next_empty_cell = cache.head; + + for (uint32_t i = 0; i < cache.size; ++i) { + if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; } + llama_kv_cell & cell = cache.cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + + // find usable cell range + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = batch.seq_id[s][0]; + llama_kv_cell & seq_meta = cache.cells[seq_id]; + bool has_cell = false; + if (seq_meta.tail >= 0) { + llama_kv_cell & cell = cache.cells[seq_meta.tail]; + GGML_ASSERT(cell.has_seq_id(seq_id)); + // does this seq_id "own" the cell? + if (cell.seq_id.size() == 1) { has_cell = true; } + } + if (!has_cell) { + llama_kv_cell & empty_cell = cache.cells[next_empty_cell]; + GGML_ASSERT(empty_cell.is_empty()); + // copy old tail into the empty cell + if (seq_meta.tail >= 0) { + llama_kv_cell & orig_cell = cache.cells[seq_meta.tail]; + empty_cell.pos = orig_cell.pos; + empty_cell.src = orig_cell.src; + orig_cell.seq_id.erase(seq_id); + empty_cell.seq_id.insert(seq_id); // will be overwritten + } + seq_meta.tail = next_empty_cell; + // find next empty cell + if (s + 1 < n_seqs) { + next_empty_cell += 1; + for (uint32_t i = 0; i < cache.size; ++i) { + if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; } + llama_kv_cell & cell = cache.cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + } + } + if (min > seq_meta.tail) { min = seq_meta.tail; } + if (max < seq_meta.tail) { max = seq_meta.tail; } + } + + // gather and re-order + for (uint32_t s = 0; s < n_seqs; ++s) { + int32_t dst_id = s + min; + int32_t src_id = cache.cells[batch.seq_id[s][0]].tail; + if (dst_id != src_id) { + llama_kv_cell & dst_cell = cache.cells[dst_id]; + llama_kv_cell & src_cell = cache.cells[src_id]; + + std::swap(dst_cell.pos, src_cell.pos); + std::swap(dst_cell.src, src_cell.src); + std::swap(dst_cell.seq_id, src_cell.seq_id); + + // swap tails (assuming they NEVER overlap) + for (const llama_seq_id seq_id : src_cell.seq_id) { + cache.cells[seq_id].tail = src_id; + } + for (const llama_seq_id seq_id : dst_cell.seq_id) { + cache.cells[seq_id].tail = dst_id; + } + } + } + + // update the pos of the used seqs + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_pos last_pos = batch.pos[n_seq_tokens * s + n_seq_tokens - 1]; + int32_t cell_id = s + min; + llama_kv_cell & cell = cache.cells[cell_id]; + + if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) { + // What should happen when the pos backtracks or skips a value? + // Clearing the state mid-batch would require special-casing which isn't done. + LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", + __func__, last_pos, cell.pos, batch.seq_id[s][0], n_seq_tokens); + } + cell.pos = last_pos; + cell.seq_id.clear(); + for (int32_t j = 0; j < batch.n_seq_id[s]; ++j) { + const llama_seq_id seq_id = batch.seq_id[s][j]; + cell.seq_id.insert(seq_id); + cache.cells[seq_id].tail = cell_id; } } @@ -3091,7 +3730,7 @@ static bool llama_kv_cache_find_slot( cache.n = max - min + 1; // sanity check - return max >= min; + return cache.n >= n_seqs; } // otherwise, one cell per token. @@ -3129,11 +3768,14 @@ static bool llama_kv_cache_find_slot( } } - for (uint32_t i = 0; i < n_tokens; i++) { - cache.cells[cache.head + i].pos = batch.pos[i]; + for (uint32_t s = 0; s < n_seqs; s++) { + for (uint32_t i = 0; i < n_seq_tokens; ++i) { + uint32_t k = s*n_seq_tokens + i; + cache.cells[cache.head + k].pos = batch.pos[k]; - for (int32_t j = 0; j < batch.n_seq_id[i]; j++) { - cache.cells[cache.head + i].seq_id.insert(batch.seq_id[i][j]); + for (int32_t j = 0; j < batch.n_seq_id[s]; j++) { + cache.cells[cache.head + k].seq_id.insert(batch.seq_id[s][j]); + } } } @@ -3159,6 +3801,8 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { for (int32_t i = 0; i < (int32_t) cache.size; ++i) { cache.cells[i].pos = -1; cache.cells[i].seq_id.clear(); + cache.cells[i].src = -1; + cache.cells[i].tail = -1; } cache.head = 0; cache.used = 0; @@ -3178,16 +3822,24 @@ static bool llama_kv_cache_seq_rm( if (p0 < 0) p0 = 0; if (p1 < 0) p1 = std::numeric_limits::max(); - // models like Mamba can't have a state partially erased + // models like Mamba or RWKV can't have a state partially erased if (cache.recurrent) { if (seq_id >= (int64_t) cache.size) { // could be fatal return false; } if (0 <= seq_id) { - // partial intersection is invalid - if ((0 < p0 && p0 <= cache.cells[seq_id].pos) || (0 < p1 && p1 <= cache.cells[seq_id].pos)) { - return false; + int32_t & tail_id = cache.cells[seq_id].tail; + if (tail_id >= 0) { + const llama_kv_cell & cell = cache.cells[tail_id]; + // partial intersection is invalid + if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) { + return false; + } + // invalidate tails which will be cleared + if (p0 <= cell.pos && cell.pos < p1) { + tail_id = -1; + } } } else { // seq_id is negative, then the range should include everything or nothing @@ -3211,6 +3863,7 @@ static bool llama_kv_cache_seq_rm( if (cache.cells[i].pos >= 0) cache.used--; cache.cells[i].pos = -1; + cache.cells[i].src = -1; if (new_head == cache.size) new_head = i; } } @@ -3233,23 +3886,29 @@ static void llama_kv_cache_seq_cp( if (cache.recurrent) { if ((uint32_t) seq_id_dst < cache.size && (uint32_t) seq_id_src < cache.size) { - seq_id_src = cache.cells[seq_id_src].src; - GGML_ASSERT((uint32_t) seq_id_src < cache.size); - // intent to "copy from" - // supports copy chains thanks to taking the source of the source - cache.cells[seq_id_dst].src = seq_id_src; + llama_kv_cell & tail_src = cache.cells[seq_id_src]; + llama_kv_cell & tail_dst = cache.cells[seq_id_dst]; + if (tail_dst.tail >= 0) { + // clear destination seq_id if it wasn't empty + llama_kv_cell & cell_dst = cache.cells[tail_dst.tail]; - // preserve the "keep or clear" status of the copied sequence - if (cache.cells[seq_id_src].has_seq_id(seq_id_src)) { - cache.cells[seq_id_dst].seq_id.insert(seq_id_dst); - } else { - cache.cells[seq_id_dst].seq_id.erase(seq_id_dst); + cell_dst.seq_id.erase(seq_id_dst); + tail_dst.tail = -1; + if (cell_dst.seq_id.empty()) { + cell_dst.pos = -1; + cell_dst.delta = -1; + cell_dst.src = -1; + cache.used -= 1; + } } + if (tail_src.tail >= 0) { + llama_kv_cell & cell_src = cache.cells[tail_src.tail]; - cache.do_copy = true; - - cache.cells[seq_id_dst].pos = cache.cells[seq_id_src].pos; + cell_src.seq_id.insert(seq_id_dst); + tail_dst.tail = tail_src.tail; + } } + return; } // otherwise, this is the KV cache of a Transformer-like model @@ -3267,9 +3926,13 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id uint32_t new_head = cache.size; for (uint32_t i = 0; i < cache.size; ++i) { + if (cache.recurrent && (llama_seq_id) i != seq_id) { + cache.cells[i].tail = -1; + } if (!cache.cells[i].has_seq_id(seq_id)) { if (cache.cells[i].pos >= 0) cache.used--; cache.cells[i].pos = -1; + cache.cells[i].src = -1; cache.cells[i].seq_id.clear(); if (new_head == cache.size) new_head = i; } else { @@ -3296,11 +3959,14 @@ static void llama_kv_cache_seq_add( if (p0 == p1) return; if (cache.recurrent) { - // for Mamba-like models, only the pos needs to be shifted + // for Mamba-like or RWKV models, only the pos needs to be shifted if (0 <= seq_id && seq_id < (int64_t) cache.size) { - llama_kv_cell & cell = cache.cells[seq_id]; - if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { - cell.pos += delta; + const int32_t tail_id = cache.cells[seq_id].tail; + if (tail_id >= 0) { + llama_kv_cell & cell = cache.cells[tail_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos += delta; + } } } return; @@ -3342,11 +4008,14 @@ static void llama_kv_cache_seq_div( if (p0 == p1) return; if (cache.recurrent) { - // for Mamba-like models, only the pos needs to be changed + // for Mamba-like or RWKV models, only the pos needs to be changed if (0 <= seq_id && seq_id < (int64_t) cache.size) { - llama_kv_cell & cell = cache.cells[seq_id]; - if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { - cell.pos /= d; + const int32_t tail_id = cache.cells[seq_id].tail; + if (tail_id >= 0) { + llama_kv_cell & cell = cache.cells[tail_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos /= d; + } } } return; @@ -3378,7 +4047,9 @@ static llama_pos llama_kv_cache_seq_pos_max(struct llama_kv_cache & cache, llama } static void llama_kv_cache_defrag(struct llama_kv_cache & cache) { - cache.do_defrag = true; + if (!cache.recurrent) { + cache.do_defrag = true; + } } static uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) { @@ -3592,13 +4263,8 @@ namespace GGUFMeta { using llama_buf_map = std::unordered_map; -// TODO: update when needed or think of some clever automatic way to do this -static size_t llama_model_max_nodes(const llama_model & /*model*/) { - //if (model.arch == LLM_ARCH_LLAMA && model.hparams.n_layer > ??) { // llama-3 405B - // return 32768; - //} - - return 8192; +static size_t llama_model_max_nodes(const llama_model & model) { + return std::max(8192, model.tensors_by_name.size()*5); } struct llama_model_loader { @@ -4533,6 +5199,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_1B: return "1B"; case MODEL_1_3B: return "1.3B"; case MODEL_1_4B: return "1.4B"; + case MODEL_1_6B: return "1.6B"; case MODEL_2B: return "2B"; case MODEL_2_8B: return "2.8B"; case MODEL_3B: return "3B"; @@ -4579,6 +5246,7 @@ static const char * llama_model_vocab_type_name(enum llama_vocab_type type){ case LLAMA_VOCAB_TYPE_BPE: return "BPE"; case LLAMA_VOCAB_TYPE_WPM: return "WPM"; case LLAMA_VOCAB_TYPE_UGM: return "UGM"; + case LLAMA_VOCAB_TYPE_RWKV: return "RWKV"; default: return "unknown"; } } @@ -4918,7 +5586,6 @@ static void llm_load_hparams( } break; case LLM_ARCH_PHI3: { - ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { @@ -4927,6 +5594,22 @@ static void llm_load_hparams( case 40: model.type = e_model::MODEL_14B; break; default: model.type = e_model::MODEL_UNKNOWN; } + + // for backward compatibility ; see: https://github.com/ggerganov/llama.cpp/pull/8931 + if ((hparams.n_layer == 32 || hparams.n_layer == 40) && hparams.n_ctx_train == 4096) { + // default value for Phi-3-mini-4k-instruct and Phi-3-medium-4k-instruct + hparams.n_swa = 2047; + } else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) { + // default value for Phi-3-mini-128k-instruct + hparams.n_swa = 262144; + } else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) { + // default value for Phi-3-medium-128k-instruct + hparams.n_swa = 131072; + } + bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); + if (!found_swa && hparams.n_swa == 0) { + throw std::runtime_error("invalid value for sliding_window"); + } } break; case LLM_ARCH_PLAMO: { @@ -5018,6 +5701,7 @@ static void llm_load_hparams( ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_DT_B_C_RMS, hparams.ssm_dt_b_c_rms, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -5224,6 +5908,12 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_T5ENCODER: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts); + model.type = e_model::MODEL_UNKNOWN; + } break; case LLM_ARCH_JAIS: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -5236,6 +5926,43 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_NEMOTRON: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_4B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_EXAONE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_8B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + case LLM_ARCH_RWKV6: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + ml.get_key(LLM_KV_WKV_HEAD_SIZE, hparams.wkv_head_size); + ml.get_key(LLM_KV_TIME_MIX_EXTRA_DIM, hparams.time_mix_extra_dim); + ml.get_key(LLM_KV_TIME_DECAY_EXTRA_DIM, hparams.time_decay_extra_dim); + ml.get_key(LLM_KV_RESCALE_EVERY_N_LAYERS, hparams.rescale_every_n_layers, false); + + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1_6B; break; + case 32: + switch (hparams.n_embd) { + case 2560: model.type = e_model::MODEL_3B; break; + case 4096: model.type = e_model::MODEL_7B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 61: model.type = e_model::MODEL_14B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -5365,6 +6092,15 @@ static void llm_load_vocab( } #endif } + } else if (tokenizer_model == "rwkv") { + vocab.type = LLAMA_VOCAB_TYPE_RWKV; + + // default special tokens + vocab.special_bos_id = -1; + vocab.special_eos_id = -1; + vocab.special_unk_id = -1; + vocab.special_sep_id = -1; + vocab.special_pad_id = -1; } else { throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str())); } @@ -5459,6 +6195,15 @@ static void llm_load_vocab( } else if ( tokenizer_pre == "codeshell") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL; + } else if ( + tokenizer_pre == "bloom") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_BLOOM; + } else if ( + tokenizer_pre == "gpt3-finnish") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH; + } else if ( + tokenizer_pre == "exaone") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE; } else { LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__); vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; @@ -5479,6 +6224,12 @@ static void llm_load_vocab( vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; vocab.tokenizer_add_bos = false; vocab.tokenizer_add_eos = true; + } else if (vocab.type == LLAMA_VOCAB_TYPE_RWKV) { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + vocab.tokenizer_add_space_prefix = false; + vocab.tokenizer_clean_spaces = false; + vocab.tokenizer_add_bos = false; + vocab.tokenizer_add_eos = false; } else { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; } @@ -5583,6 +6334,10 @@ static void llm_load_vocab( } } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) { vocab.linefeed_id = vocab.special_pad_id; + } else if (vocab.type == LLAMA_VOCAB_TYPE_RWKV) { + const std::vector ids = llama_tokenize_internal(vocab, "\n", false); + GGML_ASSERT(!ids.empty() && "model vocab missing newline token"); + vocab.linefeed_id = ids[0]; } else { const std::vector ids = llama_tokenize_internal(vocab, "\xC4\x8A", false); // U+010A GGML_ASSERT(!ids.empty() && "model vocab missing newline token"); @@ -5833,6 +6588,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); + LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms = %d\n", __func__, hparams.ssm_dt_b_c_rms); } LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); @@ -6033,6 +6789,7 @@ static bool llm_load_tensors( const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_vocab = hparams.n_vocab; const int64_t n_vocab_type = hparams.n_vocab_type; + const int64_t n_rot = hparams.n_rot; const int64_t n_expert = hparams.n_expert; const int64_t n_expert_used = hparams.n_expert_used; const int64_t n_ctx_train = hparams.n_ctx_train; @@ -6090,7 +6847,7 @@ static bool llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); - layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_embd/n_head/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); if (n_expert == 0) { layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); @@ -6098,9 +6855,9 @@ static bool llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); // optional MLP bias - layer.ffn_gate_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.ffn_down_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.ffn_up_b = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_gate_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); } else { layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}); @@ -6424,7 +7181,7 @@ static bool llm_load_tensors( layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); //output_dens - layer.bo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); //output_dens + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); //output_dens layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}); @@ -7450,6 +8207,42 @@ static bool llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_DEC_FFN_UP, "weight", i), {n_embd, n_ff}); } } break; + case LLM_ARCH_T5ENCODER: + { + const auto n_rel_attn_bkts = hparams.n_rel_attn_bkts; + + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm_enc = ml.create_tensor(ctx_output, tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); + } + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_rel_b_enc = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ENC_ATTN_REL_B, "weight", i), {n_head, n_rel_attn_bkts}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.wq_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wk_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_ATTN_OUT, "weight", i), {n_embd_v_gqa, n_embd}); + + layer.ffn_norm_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_gate_enc = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ENC_FFN_GATE, "weight", i), {n_embd, n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_down_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up_enc = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ENC_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; case LLM_ARCH_JAIS: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -7507,8 +8300,8 @@ static bool llm_load_tensors( layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); - layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + (hparams.n_embd_head_k << 2)}); - layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + (hparams.n_embd_head_k << 2)}); + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); @@ -7519,6 +8312,140 @@ static bool llm_load_tensors( layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); } } break; + case LLM_ARCH_NEMOTRON: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + // optional bias tensors + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + + // optional MLP bias + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, llama_model_loader::TENSOR_NOT_REQUIRED); + } + } break; + case LLM_ARCH_EXAONE: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0)); + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; + case LLM_ARCH_RWKV6: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // Block 0, LN0 + model.tok_norm = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); + model.tok_norm_b = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}); + + // output + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + + const int time_mix_extra_dim = hparams.time_mix_extra_dim; + const int time_decay_extra_dim = hparams.time_decay_extra_dim; + const int head_size = hparams.wkv_head_size; + const int attn_hidden_size = n_embd; + const int ffn_size = hparams.n_ff_arr[0]; + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.attn_norm_2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}); + layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}); + + layer.time_mix_w1 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_W1, "weight", i), {n_embd, time_mix_extra_dim * 5}); + layer.time_mix_w2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {time_mix_extra_dim, n_embd, 5}); + + layer.time_mix_lerp_x = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}); + layer.time_mix_lerp_w = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}); + layer.time_mix_lerp_k = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}); + layer.time_mix_lerp_v = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}); + layer.time_mix_lerp_r = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}); + layer.time_mix_lerp_g = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}); + + layer.time_mix_first = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}); + layer.time_mix_decay = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_DECAY, "weight", i), {n_embd}); + layer.time_mix_decay_w1 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_DECAY_W1, "weight", i), {n_embd, time_decay_extra_dim}); + layer.time_mix_decay_w2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_DECAY_W2, "weight", i), {time_decay_extra_dim, attn_hidden_size}); + layer.time_mix_key = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_KEY, "weight", i), {attn_hidden_size, n_embd}); + layer.time_mix_value = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}); + layer.time_mix_receptance = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}); + layer.time_mix_gate = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_GATE, "weight", i), {attn_hidden_size, n_embd}); + + layer.time_mix_ln = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}); + layer.time_mix_ln_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}); + layer.time_mix_output = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}); + + layer.channel_mix_lerp_k = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_CHANNEL_MIX_LERP_K, "weight", i), {n_embd, 1, 1}); + layer.channel_mix_lerp_r = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_CHANNEL_MIX_LERP_R, "weight", i), {n_embd, 1, 1}); + + layer.channel_mix_key = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_CHANNEL_MIX_KEY, "weight", i), {n_embd, ffn_size}); + layer.channel_mix_value = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_CHANNEL_MIX_VALUE, "weight", i), {ffn_size, n_embd}); + layer.channel_mix_receptance = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_CHANNEL_MIX_RECEPTANCE, "weight", i), {n_embd, n_embd}); + } + + } break; default: throw std::runtime_error("unknown architecture"); } @@ -7760,7 +8687,7 @@ static struct ggml_tensor * llm_build_inp_embd( struct ggml_context * ctx, struct llama_context & lctx, const llama_hparams & hparams, - const llama_batch & batch, + const llama_ubatch & batch, struct ggml_tensor * tok_embd, const llm_build_cb & cb) { const int64_t n_embd = hparams.n_embd; @@ -7803,8 +8730,7 @@ static void llm_build_kv_store( GGML_ASSERT(kv.size == n_ctx); - struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa, - (ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa))*kv_head); + struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa, ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa)*kv_head); cb(k_cache_view, "k_cache_view", il); // note: storing RoPE-ed version of K in the KV cache @@ -7815,8 +8741,7 @@ static void llm_build_kv_store( struct ggml_tensor * v_cache_view = nullptr; if (cparams.flash_attn) { - v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa, - (kv_head)*ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa)); + v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa, ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa)*kv_head); } else { // note: the V cache is transposed when not using flash attention v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa, @@ -8194,9 +9119,10 @@ static struct ggml_tensor * llm_build_kqv( 0); cb(v, "v", il); - cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias); + cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias, + hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f); - if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) { + if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_GEMMA2) { ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); } @@ -8205,7 +9131,7 @@ static struct ggml_tensor * llm_build_kqv( struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); cb(kq, "kq", il); - if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) { + if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 || model.arch == LLM_ARCH_NEMOTRON || model.arch == LLM_ARCH_CHATGLM) { // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847 ggml_mul_mat_set_prec(kq, GGML_PREC_F32); @@ -8302,19 +9228,351 @@ static struct ggml_tensor * llm_build_kv( struct ggml_tensor * cur; - cur = llm_build_kqv(ctx, lctx, kv, graph, wo, wo_b, - q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il); + cur = llm_build_kqv(ctx, lctx, kv, graph, wo, wo_b, q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il); cb(cur, "kqv_out", il); return cur; } +static struct ggml_tensor * llm_build_copy_mask_state( + struct ggml_context * ctx, + struct ggml_cgraph * graph, + struct ggml_tensor * s, + struct ggml_tensor * state_copy, + struct ggml_tensor * state_mask, + int32_t n_state, + int32_t kv_size, + int32_t kv_head, + int32_t n_kv, + int32_t n_seqs) { + struct ggml_tensor * states = ggml_reshape_2d(ctx, s, n_state, kv_size); + + // copy states + // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv + // this shrinks the tensors's ne[1] to n_kv + states = ggml_get_rows(ctx, states, state_copy); + + // clear states of sequences which are starting at the beginning of this batch + // FIXME: zero-out NANs? + states = ggml_mul(ctx, states, state_mask); + + // copy states which won't be changed further (between n_seqs and n_rs) + ggml_build_forward_expand(graph, + ggml_cpy(ctx, + ggml_view_1d(ctx, states, n_state*(n_kv - n_seqs), n_seqs*n_state*ggml_element_size(states)), + ggml_view_1d(ctx, s, n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*ggml_element_size(s)))); + + // the part of the states that will be used and modified + return ggml_view_2d(ctx, states, n_state, n_seqs, states->nb[1], 0); +} + +// TODO: split +static struct ggml_tensor * llm_build_mamba( + struct ggml_context * ctx, + struct llama_context & lctx, + const llama_ubatch & batch, + struct ggml_cgraph * graph, + struct ggml_tensor * cur, + struct ggml_tensor * state_copy, + struct ggml_tensor * state_mask, + int32_t kv_head, + int32_t n_kv, + const llm_build_cb & cb, + int il) { + const llama_model & model = lctx.model; + const llama_hparams & hparams = model.hparams; + const llama_kv_cache & kv = lctx.kv_self; + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t dt_rank = hparams.ssm_dt_rank; + const int64_t n_seqs = batch.n_seqs; + // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers) + const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms; + // Use the same RMS norm as the final layer norm + const float norm_rms_eps = hparams.f_norm_rms_eps; + + const int64_t n_seq_tokens = batch.n_seq_tokens; + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(batch.equal_seqs); + GGML_ASSERT(batch.n_tokens == n_seq_tokens * n_seqs); + + struct ggml_tensor * conv_states_all = kv.k_l[il]; + struct ggml_tensor * ssm_states_all = kv.v_l[il]; + + // (ab)using the KV cache to store the states + struct ggml_tensor * conv = llm_build_copy_mask_state(ctx, + graph, conv_states_all, state_copy, state_mask, + hparams.n_embd_k_s(), kv.size, kv_head, n_kv, n_seqs); + conv = ggml_reshape_3d(ctx, conv, d_conv - 1, d_inner, n_seqs); + struct ggml_tensor * ssm = llm_build_copy_mask_state(ctx, + graph, ssm_states_all, state_copy, state_mask, + hparams.n_embd_v_s(), kv.size, kv_head, n_kv, n_seqs); + ssm = ggml_reshape_3d(ctx, ssm, d_state, d_inner, n_seqs); + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx, cur, cur->ne[0], n_seq_tokens, n_seqs); + + // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} + struct ggml_tensor * xz = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_in, cur); + // split the above in two + // => {d_inner, n_seq_tokens, n_seqs} + struct ggml_tensor * x = ggml_view_3d(ctx, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0); + struct ggml_tensor * z = ggml_view_3d(ctx, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner*ggml_element_size(xz)); + + // conv + { + // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} + struct ggml_tensor * conv_x = ggml_concat(ctx, conv, ggml_transpose(ctx, x), 0); + + // copy last (d_conv - 1) columns back into the state cache + struct ggml_tensor * last_conv = ggml_view_3d(ctx, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); + + ggml_build_forward_expand(graph, + ggml_cpy(ctx, last_conv, + ggml_view_1d(ctx, conv_states_all, + (d_conv - 1)*(d_inner)*(n_seqs), + kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all)))); + + // 1D convolution + // The equivalent is to make a self-overlapping view of conv_x + // over d_conv columns at each stride in the 3rd dimension, + // then element-wise multiply that with the conv1d weight, + // then sum the elements of each row, + // (the last two steps are a dot product over rows (also doable with mul_mat)) + // then permute away the ne[0] dimension, + // and then you're left with the resulting x tensor. + // For simultaneous sequences, all sequences need to have the same length. + x = ggml_ssm_conv(ctx, conv_x, model.layers[il].ssm_conv1d); + + // bias + x = ggml_add(ctx, x, model.layers[il].ssm_conv1d_b); + + x = ggml_silu(ctx, x); + } + + // ssm + { + // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} + struct ggml_tensor * x_db = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_x, x); + // split + struct ggml_tensor * dt = ggml_view_3d(ctx, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0); + struct ggml_tensor * B = ggml_view_3d(ctx, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank); + struct ggml_tensor * C = ggml_view_3d(ctx, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state)); + + // Some Mamba variants (e.g. FalconMamba) apply RMS norm in B, C & Dt layers + if (ssm_dt_b_c_rms) { + dt = ggml_rms_norm(ctx, dt, norm_rms_eps); + B = ggml_rms_norm(ctx, B, norm_rms_eps); + C = ggml_rms_norm(ctx, C, norm_rms_eps); + } + + // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} + dt = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_dt, dt); + dt = ggml_add(ctx, dt, model.layers[il].ssm_dt_b); + + // Custom operator to optimize the parallel associative scan + // as described in the Annex D of the Mamba paper. + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + struct ggml_tensor * y_ssm = ggml_ssm_scan(ctx, ssm, x, dt, model.layers[il].ssm_a, B, C); + + // store last states + ggml_build_forward_expand(graph, + ggml_cpy(ctx, + ggml_view_1d(ctx, y_ssm, d_state*d_inner*n_seqs, x->nb[3]), + ggml_view_1d(ctx, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); + + struct ggml_tensor * y = ggml_view_3d(ctx, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[1], x->nb[2], 0); + + // TODO: skip computing output earlier for unused tokens + + // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs} + y = ggml_add(ctx, y, ggml_mul(ctx, x, model.layers[il].ssm_d)); + y = ggml_mul(ctx, y, ggml_silu(ctx, ggml_cont(ctx, z))); + + // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} + cur = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_out, y); + } + + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + cur = ggml_reshape_2d(ctx, cur, cur->ne[0], n_seq_tokens * n_seqs); + cb(cur, "mamba_out", il); + + return cur; +} + +static struct ggml_tensor * llm_build_rwkv6_time_mix( + struct llama_context & lctx, + struct ggml_context * ctx, + const struct llama_layer * layer, + struct ggml_tensor * cur, + struct ggml_tensor * x_prev, + struct ggml_tensor ** wkv_state) { + size_t n_embed = cur->ne[0]; + size_t n_seq_tokens = cur->ne[1]; + size_t n_seqs = cur->ne[2]; + + size_t head_size = layer->time_mix_first->ne[0]; + size_t head_count = layer->time_mix_first->ne[1]; + + size_t n_tokens = n_seqs * n_seq_tokens; + + struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur); + + sx = ggml_reshape_2d(ctx, sx, n_embed, n_tokens); + cur = ggml_reshape_2d(ctx, cur, n_embed, n_tokens); + + struct ggml_tensor * xxx = ggml_add(ctx, ggml_mul(ctx, sx, layer->time_mix_lerp_x), cur); + + xxx = ggml_reshape_4d( + ctx, + ggml_tanh( + ctx, + ggml_mul_mat(ctx, layer->time_mix_w1, xxx) + ), + layer->time_mix_w1->ne[1] / 5, 1, 5, n_tokens + ); + + xxx = ggml_cont(ctx, ggml_permute(ctx, xxx, 0, 1, 3, 2)); + + xxx = ggml_mul_mat( + ctx, + ggml_reshape_4d( + ctx, + layer->time_mix_w2, + layer->time_mix_w2->ne[0], layer->time_mix_w2->ne[1], 1, 5 + ), + xxx + ); + + struct ggml_tensor *mw = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], 0); + struct ggml_tensor *mk = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * sizeof(float)); + struct ggml_tensor *mv = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 2 * sizeof(float)); + struct ggml_tensor *mr = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 3 * sizeof(float)); + struct ggml_tensor *mg = ggml_view_2d(ctx, xxx, n_embed, n_tokens, xxx->nb[1], n_embed * n_tokens * 4 * sizeof(float)); + + struct ggml_tensor * xw = ggml_add( + ctx, + ggml_mul( + ctx, + ggml_add(ctx, mw, layer->time_mix_lerp_w), + sx + ), + cur + ); + + struct ggml_tensor * xk = ggml_add( + ctx, + ggml_mul( + ctx, + ggml_add(ctx, mk, layer->time_mix_lerp_k), + sx + ), + cur + ); + + struct ggml_tensor * xv = ggml_add( + ctx, + ggml_mul( + ctx, + ggml_add(ctx, mv, layer->time_mix_lerp_v), + sx + ), + cur + ); + + struct ggml_tensor * xr = ggml_add( + ctx, + ggml_mul( + ctx, + ggml_add(ctx, mr, layer->time_mix_lerp_r), + sx + ), + cur + ); + + struct ggml_tensor * xg = ggml_add( + ctx, + ggml_mul( + ctx, + ggml_add(ctx, mg, layer->time_mix_lerp_g), + sx + ), + cur + ); + + struct ggml_tensor * r = ggml_reshape_4d(ctx, llm_build_lora_mm(lctx, ctx, layer->time_mix_receptance, xr), head_size, 1, head_count, n_tokens); + struct ggml_tensor * k = ggml_reshape_4d(ctx, llm_build_lora_mm(lctx, ctx, layer->time_mix_key, xk), 1, head_size, head_count, n_tokens); + struct ggml_tensor * v = ggml_reshape_4d(ctx, llm_build_lora_mm(lctx, ctx, layer->time_mix_value, xv), head_size, 1, head_count, n_tokens); + struct ggml_tensor * g = ggml_silu( + ctx, + llm_build_lora_mm(lctx, ctx, layer->time_mix_gate, xg) + ); + + struct ggml_tensor * w = ggml_mul_mat( + ctx, + layer->time_mix_decay_w2, + ggml_tanh( + ctx, + ggml_mul_mat(ctx, layer->time_mix_decay_w1, xw) + ) + ); + + w = ggml_add(ctx, w, ggml_reshape_1d(ctx, layer->time_mix_decay, n_embed)); + w = ggml_exp(ctx, ggml_neg(ctx, ggml_exp(ctx, w))); + w = ggml_reshape_4d(ctx, w, 1, head_size, head_count, n_tokens); + + k = ggml_transpose(ctx, k); + v = ggml_transpose(ctx, v); + r = ggml_transpose(ctx, r); + + struct ggml_tensor * wkv_output = ggml_rwkv_wkv(ctx, k, v, r, layer->time_mix_first, w, *wkv_state); + cur = ggml_view_1d(ctx, wkv_output, n_embed * n_tokens, 0); + *wkv_state = ggml_view_1d(ctx, wkv_output, n_embed * head_size * n_seqs, n_embed * n_tokens * sizeof(float)); + + // group norm with head_count groups + cur = ggml_reshape_3d(ctx, cur, n_embed / head_count, head_count, n_tokens); + cur = ggml_norm(ctx, cur, 64e-5f); + + // Convert back to regular vectors. + cur = ggml_reshape_2d(ctx, cur, n_embed, n_tokens); + cur = ggml_add(ctx, ggml_mul(ctx, cur, layer->time_mix_ln), layer->time_mix_ln_b); + + cur = ggml_mul(ctx, cur, g); + cur = llm_build_lora_mm(lctx, ctx, layer->time_mix_output, cur); + + return ggml_reshape_3d(ctx, cur, n_embed, n_seq_tokens, n_seqs); +} + +static struct ggml_tensor * llm_build_rwkv6_channel_mix( + struct llama_context & lctx, + struct ggml_context * ctx, + const struct llama_layer * layer, + struct ggml_tensor * cur, + struct ggml_tensor * x_prev) { + struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur); + struct ggml_tensor * xk = ggml_add(ctx, ggml_mul(ctx, sx, layer->channel_mix_lerp_k), cur); + struct ggml_tensor * xr = ggml_add(ctx, ggml_mul(ctx, sx, layer->channel_mix_lerp_r), cur); + + struct ggml_tensor * r = ggml_sigmoid(ctx, llm_build_lora_mm(lctx, ctx, layer->channel_mix_receptance, xr)); + struct ggml_tensor * k = ggml_sqr( + ctx, + ggml_relu( + ctx, + llm_build_lora_mm(lctx, ctx, layer->channel_mix_key, xk) + ) + ); + + return ggml_mul(ctx, r, llm_build_lora_mm(lctx, ctx, layer->channel_mix_value, k)); +} + struct llm_build_context { const llama_model & model; llama_context & lctx; const llama_hparams & hparams; const llama_cparams & cparams; - const llama_batch & batch; + const llama_ubatch & batch; const llama_kv_cache & kv_self; const int64_t n_embd; @@ -8360,7 +9618,7 @@ struct llm_build_context { // TODO: consider making the entire interface noexcept llm_build_context( llama_context & lctx, - const llama_batch & batch, + const llama_ubatch & batch, const llm_build_cb & cb, bool worst_case) : model (lctx.model), @@ -8467,29 +9725,6 @@ struct llm_build_context { return gf; } - struct ggml_cgraph * build_s_copy() { - struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); - - GGML_ASSERT(kv_self.recurrent); - - struct ggml_tensor * state_copy = build_inp_s_copy(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * conv_states = ggml_reshape_2d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s(), kv_self.size); - struct ggml_tensor * ssm_states = ggml_reshape_2d(ctx0, kv_self.v_l[il], hparams.n_embd_v_s(), kv_self.size); - - conv_states = ggml_get_rows(ctx0, conv_states, state_copy); - ssm_states = ggml_get_rows(ctx0, ssm_states, state_copy); - - // TODO: name the intermediate tensors with cb() - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, conv_states, kv_self.k_l[il])); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, ssm_states, kv_self.v_l[il])); - } - - return gf; - } - struct ggml_cgraph * build_defrag(const std::vector & ids) { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); @@ -8624,7 +9859,7 @@ struct llm_build_context { } struct ggml_tensor * build_inp_s_copy() { - lctx.inp_s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, kv_self.size); + lctx.inp_s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv); cb(lctx.inp_s_copy, "inp_s_copy", -1); ggml_set_input(lctx.inp_s_copy); return lctx.inp_s_copy; @@ -8637,13 +9872,6 @@ struct llm_build_context { return lctx.inp_s_mask; } - struct ggml_tensor * build_inp_s_seq() { - lctx.inp_s_seq = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens); - cb(lctx.inp_s_seq, "inp_s_seq", -1); - ggml_set_input(lctx.inp_s_seq); - return lctx.inp_s_seq; - } - struct ggml_cgraph * append_pooling(struct ggml_cgraph * gf) { // find result_norm tensor for input struct ggml_tensor * inp = nullptr; @@ -11973,125 +13201,31 @@ struct llm_build_context { struct ggml_cgraph * build_mamba() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); - const int64_t d_model = n_embd; - const int64_t d_conv = hparams.ssm_d_conv; - const int64_t d_inner = hparams.ssm_d_inner; - GGML_ASSERT(2 * d_model == d_inner); - const int64_t d_state = hparams.ssm_d_state; - const int64_t dt_rank = hparams.ssm_dt_rank; - struct ggml_tensor * cur; struct ggml_tensor * inpL; // {n_embd, n_tokens} inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + struct ggml_tensor * state_copy = build_inp_s_copy(); struct ggml_tensor * state_mask = build_inp_s_mask(); - struct ggml_tensor * state_seq = build_inp_s_seq(); for (int il = 0; il < n_layer; ++il) { - // (ab)using the KV cache to store the states - struct ggml_tensor * conv_states = ggml_reshape_2d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s(), kv_self.size); - struct ggml_tensor * ssm_states = ggml_reshape_2d(ctx0, kv_self.v_l[il], hparams.n_embd_v_s(), kv_self.size); - - // clear states of sequences which are starting at the beginning of this batch - { - conv_states = ggml_mul(ctx0, - ggml_view_2d(ctx0, conv_states, conv_states->ne[0], n_kv, conv_states->nb[1], kv_head*conv_states->nb[1]), - state_mask); - ssm_states = ggml_mul(ctx0, - ggml_view_2d(ctx0, ssm_states, ssm_states->ne[0], n_kv, ssm_states->nb[1], kv_head*ssm_states->nb[1]), - state_mask); - } - - conv_states = ggml_reshape_3d(ctx0, conv_states, d_conv - 1, d_inner, n_kv); - ssm_states = ggml_reshape_3d(ctx0, ssm_states, d_state, d_inner, n_kv); - // norm cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, cb, il); cb(cur, "attn_norm", il); - // {n_embd, 2*d_inner} * {n_embd, n_tokens} => {2*d_inner, n_tokens} - struct ggml_tensor * xz = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_in, cur); - // split the above in two - // => {d_inner, n_tokens} - struct ggml_tensor * x = ggml_view_2d(ctx0, xz, d_inner, xz->ne[1], xz->nb[1], 0); - struct ggml_tensor * z = ggml_view_2d(ctx0, xz, d_inner, xz->ne[1], xz->nb[1], ggml_element_size(xz)*d_inner); + cur = llm_build_mamba(ctx0, lctx, batch, gf, cur, + state_copy, state_mask, + kv_head, n_kv, cb, il); - // conv - { - // Custom operator which is needed only to ease simultaneous sequence processing. - // For a single sequence, the equivalent is to concatenate the columns of conv_states and x, - // then make a self-overlapping view of that over d_conv columns at each stride in the 3rd dimension, - // then element-wise multiply that with the conv1d weigth, - // then sum the elements of each row, - // (the last two steps are a dot product over rows (also doable with mul_mat)) - // then permute away the ne[0] dimension, - // and then you're left with the resulting x tensor. - // The new conv_states is the last (d_conv - 1) columns - // of the last 3rd dimensional "layer" of the self-overlapping view. - // For simultaneous sequences, it's more complicated. - struct ggml_tensor * x_conv = ggml_ssm_conv(ctx0, conv_states, x, model.layers[il].ssm_conv1d, state_seq); - - // store last (d_conv - 1) columns of the conv_state part of x_conv back into the KV cache - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, - ggml_view_2d(ctx0, x_conv, d_conv - 1, d_inner*n_kv, d_conv*ggml_element_size(x_conv), (1+d_inner*n_tokens)*ggml_element_size(x_conv)), - ggml_view_1d(ctx0, kv_self.k_l[il], (d_conv - 1)*(d_inner)*(n_kv), kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(x_conv)))); - - // extract x from x_conv - x = ggml_view_2d(ctx0, x_conv, d_inner, n_tokens, d_inner*ggml_element_size(x_conv), 0); - - // bias - x = ggml_add(ctx0, x, model.layers[il].ssm_conv1d_b); - - x = ggml_silu(ctx0, x); - } - - // ssm - { - // {d_inner, dt_rank + 2*d_state} * {d_inner, n_tokens} => {dt_rank + 2*d_state, n_tokens} - struct ggml_tensor * x_db = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_x, x); - // split - struct ggml_tensor * dt = ggml_view_2d(ctx0, x_db, dt_rank, n_tokens, x_db->nb[1], 0); - struct ggml_tensor * B = ggml_view_2d(ctx0, x_db, d_state, n_tokens, x_db->nb[1], ggml_element_size(x_db)*dt_rank); - struct ggml_tensor * C = ggml_view_2d(ctx0, x_db, d_state, n_tokens, x_db->nb[1], ggml_element_size(x_db)*(dt_rank+d_state)); - - // {dt_rank, d_inner} * {dt_rank, n_tokens} => {d_inner, n_tokens} - dt = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_dt, dt); - dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b); - - // Custom operator to optimize the parallel associative scan - // as described in the Annex D of the Mamba paper. - // => {d_inner, n_tokens} and {d_state, d_inner, n_kv} combined, - // because only a single tensor can be returned. - struct ggml_tensor * y_ssm_states = ggml_ssm_scan(ctx0, ssm_states, x, dt, model.layers[il].ssm_a, B, C, state_seq); - - // store last states (the second part of y_ssm_states) - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, - ggml_view_1d(ctx0, y_ssm_states, d_state*d_inner*n_kv, d_inner*n_tokens*ggml_element_size(y_ssm_states)), - ggml_view_1d(ctx0, kv_self.v_l[il], d_state*d_inner*n_kv, kv_head*d_state*d_inner*ggml_element_size(ssm_states)))); - - struct ggml_tensor * y = ggml_view_2d(ctx0, y_ssm_states, d_inner, n_tokens, d_inner*ggml_element_size(y_ssm_states), 0); - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - x = ggml_get_rows(ctx0, x, inp_out_ids); - y = ggml_get_rows(ctx0, y, inp_out_ids); - z = ggml_get_rows(ctx0, z, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // {d_inner, n_tokens} * {d_inner} => {d_inner, n_tokens} - y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d)); - y = ggml_mul(ctx0, y, ggml_silu(ctx0, z)); - - // {d_inner, n_embd} * {d_inner, n_tokens} => {n_embd, n_tokens} - cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].ssm_out, y); + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } // residual @@ -13164,7 +14298,7 @@ struct llm_build_context { return gf; } - struct ggml_cgraph * build_t5() { + struct ggml_cgraph * build_t5_encoder() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); // mutable variable, needed during the last layer of the computation to skip unused tokens @@ -13179,303 +14313,323 @@ struct llm_build_context { inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); - if (lctx.is_encoding) { - struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false); + GGML_ASSERT(lctx.is_encoding); + struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false); + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false); - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm_enc, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm_enc, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); - // self-attention - { - struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_enc, cur); - cb(Qcur, "Qcur", il); + // self-attention + { + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_enc, cur); + cb(Qcur, "Qcur", il); - struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_enc, cur); - cb(Kcur, "Kcur", il); + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_enc, cur); + cb(Kcur, "Kcur", il); - struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_enc, cur); - cb(Vcur, "Vcur", il); + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_enc, cur); + cb(Vcur, "Vcur", il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - cb(kq, "kq", il); + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); - struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; - struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b); - struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); - cb(kq_b, "kq_b", il); + struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; + struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b); + struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); + cb(kq_b, "kq_b", il); - kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); + kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); - struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); - cb(v, "v", il); + struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens))); + cb(v, "v", il); - struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq); - cb(kqv, "kqv", il); + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq); + cb(kqv, "kqv", il); - struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); - cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - cb(cur, "kqv_merged_cont", il); + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); - ggml_build_forward_expand(gf, cur); + ggml_build_forward_expand(gf, cur); - cur = ggml_mul_mat(ctx0, model.layers[il].wo_enc, cur); - cb(cur, "kqv_out", il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm_enc, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - // T5 uses relu, flan-T5 uses gelu-gated - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up_enc, NULL, NULL, - model.layers[il].ffn_gate_enc, NULL, NULL, - model.layers[il].ffn_down_enc, NULL, NULL, - NULL, - model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, - model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, - cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); - if (layer_dir != nullptr) { - cur = ggml_add(ctx0, cur, layer_dir); - } - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; + cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_enc, cur); + cb(cur, "kqv_out", il); } - cur = inpL; - cb(cur, "result_embd", -1); - - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm_enc, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); - } else { - GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first"); - - struct ggml_tensor * embd_enc = llm_build_inp_embd_enc(); - struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true); - - struct ggml_tensor * KQ_mask_dec = build_inp_KQ_mask(); - struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = llm_build_norm(ctx0, inpL, hparams, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm", il); - - // self-attention - { - struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); - - struct ggml_tensor * k = - ggml_view_3d(ctx0, kv_self.k_l[il], - n_embd_head_k, n_kv, n_head_kv, - ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), - ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), - 0); - cb(k, "k", il); - - struct ggml_tensor * v = - ggml_view_3d(ctx0, kv_self.v_l[il], - n_kv, n_embd_head_v, n_head_kv, - ggml_element_size(kv_self.v_l[il])*n_ctx, - ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v, - 0); - cb(v, "v", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - cb(kq, "kq", il); - - struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; - struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b); - struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); - cb(kq_b, "kq_b", il); - - kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - cb(cur, "kqv_merged_cont", il); - - ggml_build_forward_expand(gf, cur); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); - cb(cur, "kqv_out", il); - } - - cur = ggml_add(ctx0, cur, inpSA); - cb(cur, "cross_inp", il); - - struct ggml_tensor * inpCA = cur; - - // norm - cur = llm_build_norm(ctx0, cur, hparams, - model.layers[il].attn_norm_cross, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "attn_norm_cross", il); - - // cross-attention - { - struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq_cross, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk_cross, embd_enc); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv_cross, embd_enc); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); - - struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); - - struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - cb(kq, "kq", il); - - kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); - cb(kq, "kq_soft_max_ext", il); - - struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); - cb(v, "v", il); - - struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); - cb(kqv, "kqv", il); - - struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - cb(kqv_merged, "kqv_merged", il); - - cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - cb(cur, "kqv_merged_cont", il); - - ggml_build_forward_expand(gf, cur); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo_cross, cur); - cb(cur, "kqv_out", il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - n_tokens = n_outputs; - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); - } - - struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = llm_build_norm(ctx0, ffn_inp, hparams, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, cb, il); - cb(cur, "ffn_norm", il); - - // T5 uses relu, flan-T5 uses gelu-gated - cur = llm_build_ffn(ctx0, lctx, cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, - model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, - cb, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); - if (layer_dir != nullptr) { - cur = ggml_add(ctx0, cur, layer_dir); - } - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } - cur = inpL; - cb(cur, "result_embd", -1); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); - cur = llm_build_norm(ctx0, cur, hparams, - model.output_norm, NULL, - LLM_NORM_RMS, cb, -1); - cb(cur, "result_norm", -1); + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm_enc, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); - // lm_head - cur = ggml_mul_mat(ctx0, model.output, cur); - cb(cur, "result_output", -1); + // T5 uses relu, flan-T5 uses gelu-gated + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up_enc, NULL, NULL, + model.layers[il].ffn_gate_enc, NULL, NULL, + model.layers[il].ffn_down_enc, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; } + cur = inpL; + cb(cur, "result_embd", -1); + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm_enc, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_t5_decoder() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + GGML_ASSERT(!lctx.is_encoding); + GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first"); + + struct ggml_tensor * embd_enc = llm_build_inp_embd_enc(); + struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true); + + struct ggml_tensor * KQ_mask_dec = build_inp_KQ_mask(); + struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il); + + struct ggml_tensor * k = + ggml_view_3d(ctx0, kv_self.k_l[il], + n_embd_head_k, n_kv, n_head_kv, + ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k), + 0); + cb(k, "k", il); + + struct ggml_tensor * v = + ggml_view_3d(ctx0, kv_self.v_l[il], + n_kv, n_embd_head_v, n_head_kv, + ggml_element_size(kv_self.v_l[il])*n_ctx, + ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v, + 0); + cb(v, "v", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; + struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b); + struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias); + cb(kq_b, "kq_b", il); + + kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur); + cb(cur, "kqv_out", il); + } + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "cross_inp", il); + + struct ggml_tensor * inpCA = cur; + + // norm + cur = llm_build_norm(ctx0, cur, hparams, + model.layers[il].attn_norm_cross, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm_cross", il); + + // cross-attention + { + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_cross, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_cross, embd_enc); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_cross, embd_enc); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); + + struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + cb(kq, "kq", il); + + kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); + cb(kq, "kq_soft_max_ext", il); + + struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); + cb(v, "v", il); + + struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); + cb(kqv, "kqv", il); + + struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + cb(cur, "kqv_merged_cont", il); + + ggml_build_forward_expand(gf, cur); + + cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_cross, cur); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + ggml_tensor * layer_dir = lctx.cvec.tensor_for(il); + if (layer_dir != nullptr) { + cur = ggml_add(ctx0, cur, layer_dir); + } + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + cb(cur, "result_output", -1); + ggml_build_forward_expand(gf, cur); return gf; @@ -13686,11 +14840,370 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_nemotron() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + //GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, lctx, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_exaone() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + // mutable variable, needed during the last layer of the computation to skip unused tokens + int32_t n_tokens = this->n_tokens; + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = build_inp_pos(); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + struct ggml_tensor * rope_factors = build_rope_factors(il); + + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_ext( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, lctx, kv_self, gf, + model.layers[il].wo, model.layers[il].bo, + Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + n_tokens = n_outputs; + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, lctx, cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + ggml_cgraph * build_rwkv6() { + ggml_cgraph *gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false); + + // Token shift state dimensions should be 2 * n_emb + GGML_ASSERT(n_embd == hparams.n_embd_k_s() / 2); + + const int64_t n_seqs = batch.n_seqs; + const int64_t n_seq_tokens = batch.n_seq_tokens; + const int64_t n_tokens = batch.n_tokens; + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(batch.equal_seqs); + GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + struct ggml_tensor * state_copy = build_inp_s_copy(); + struct ggml_tensor * state_mask = build_inp_s_mask(); + + inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); + inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1); + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + + // (ab)using the KV cache to store the states + struct ggml_tensor * token_shift = llm_build_copy_mask_state(ctx0, + gf, kv_self.k_l[il], state_copy, state_mask, + hparams.n_embd_k_s(), kv_self.size, kv_head, n_kv, n_seqs); + struct ggml_tensor * wkv_states = llm_build_copy_mask_state(ctx0, + gf, kv_self.v_l[il], state_copy, state_mask, + hparams.n_embd_v_s(), kv_self.size, kv_head, n_kv, n_seqs); + + cur = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + token_shift = ggml_reshape_3d(ctx0, token_shift, n_embd, 2, n_seqs); + + struct ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); + struct ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); + + struct ggml_tensor * x_norm_att = llm_build_norm(ctx0, cur, hparams, layer->attn_norm, layer->attn_norm_b, LLM_NORM, cb, il); + struct ggml_tensor * x_prev = ggml_concat( + ctx0, + att_shift, + ggml_view_3d(ctx0, x_norm_att, n_embd, n_seq_tokens - 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], 0), + 1 + ); + + cur = ggml_add(ctx0, cur, llm_build_rwkv6_time_mix(lctx, ctx0, layer, x_norm_att, x_prev, &wkv_states)); + ggml_build_forward_expand(gf, cur); + ggml_build_forward_expand( + gf, + ggml_cpy( + ctx0, + wkv_states, + ggml_view_1d( + ctx0, + kv_self.v_l[il], + hparams.n_embd_v_s() * n_seqs, + hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il]) + ) + ) + ); + + struct ggml_tensor * x_norm_ffn = llm_build_norm(ctx0, cur, hparams, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, cb, il); + x_prev = ggml_concat( + ctx0, + ffn_shift, + ggml_view_3d(ctx0, x_norm_ffn, n_embd, n_seq_tokens - 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], 0), + 1 + ); + cur = ggml_add(ctx0, cur, llm_build_rwkv6_channel_mix(lctx, ctx0, layer, x_norm_ffn, x_prev)); + ggml_build_forward_expand(gf, cur); + + struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att)); + struct ggml_tensor * last_norm_ffn = ggml_view_3d(ctx0, x_norm_ffn, n_embd, 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_ffn)); + + token_shift = ggml_concat(ctx0, last_norm_att, last_norm_ffn, 1); + + ggml_build_forward_expand( + gf, + ggml_cpy( + ctx0, + ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * 2, 0), + ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il])) + ) + ); + + if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) { + cur = ggml_scale(ctx0, cur, 0.5F); + } + + cur = lctx.cvec.apply_to(ctx0, cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + + cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, model.output_norm_b, LLM_NORM, cb, -1); + cur = llm_build_lora_mm(lctx, ctx0, model.output, cur); + + cb(cur, "result_output", -1); + ggml_build_forward_expand(gf, cur); + + return gf; + } }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { - llama_batch dummy; - dummy.n_tokens = 0; + llama_ubatch dummy = {}; + dummy.equal_seqs = true; llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; @@ -13706,8 +15219,8 @@ static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const } static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) { - llama_batch dummy; - dummy.n_tokens = 0; + llama_ubatch dummy = {}; + dummy.equal_seqs = true; llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; @@ -13722,26 +15235,9 @@ static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) { return result; } -static struct ggml_cgraph * llama_build_graph_s_copy(llama_context & lctx) { - llama_batch dummy; - dummy.n_tokens = 0; - - llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; - - struct llm_build_context llm(lctx, dummy, cb, false); - - llm.init(); - - struct ggml_cgraph * result = llm.build_s_copy(); - - llm.free(); - - return result; -} - static struct ggml_cgraph * llama_build_graph( llama_context & lctx, - const llama_batch & batch, + const llama_ubatch & batch, bool worst_case) { const auto & model = lctx.model; @@ -13927,12 +15423,32 @@ static struct ggml_cgraph * llama_build_graph( } break; case LLM_ARCH_T5: { - result = llm.build_t5(); + if (lctx.is_encoding) { + result = llm.build_t5_encoder(); + } else { + result = llm.build_t5_decoder(); + } + } break; + case LLM_ARCH_T5ENCODER: + { + result = llm.build_t5_encoder(); } break; case LLM_ARCH_JAIS: { result = llm.build_jais(); } break; + case LLM_ARCH_NEMOTRON: + { + result = llm.build_nemotron(); + } break; + case LLM_ARCH_EXAONE: + { + result = llm.build_exaone(); + } break; + case LLM_ARCH_RWKV6: + { + result = llm.build_rwkv6(); + } break; default: GGML_ABORT("fatal error"); } @@ -13995,7 +15511,7 @@ static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t return relative_bucket; } -static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { +static void llama_set_inputs(llama_context & lctx, const llama_ubatch & batch) { // // set input data // @@ -14034,10 +15550,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { for (int i = 0; i < n_tokens; ++i) { data[i] = i; } - } else if (batch.logits) { + } else if (batch.output) { int32_t n_outputs = 0; for (int i = 0; i < n_tokens; ++i) { - if (batch.logits[i]) { + if (batch.output[i]) { data[n_outputs++] = i; } } @@ -14061,8 +15577,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { if (lctx.inp_KQ_mask || lctx.inp_KQ_mask_swa) { // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache. if (cparams.causal_attn && !lctx.is_encoding) { - const int64_t n_kv = kv_self.n; - const int64_t n_tokens = batch.n_tokens; + const int64_t n_kv = kv_self.n; + const int64_t n_tokens = batch.n_tokens; + const int64_t n_seq_tokens = batch.n_seq_tokens; + const int64_t n_seqs = batch.n_seqs; float * data = nullptr; @@ -14082,32 +15600,35 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { // of the correct sequence for each token of the batch. // It's assumed that if a token in the batch has multiple sequences, they are equivalent. for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = batch.seq_id[s][0]; - for (int i = 0; i < n_kv; ++i) { - float f; - if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { - f = -INFINITY; - } else { - if (hparams.use_alibi) { - f = -std::abs(lctx.kv_self.cells[i].pos - pos); - } else { - f = 0.0f; - } - } + for (int j = 0; j < n_seq_tokens; ++j) { + const llama_pos pos = batch.pos[s*n_seq_tokens + j]; - if (data) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = f; - } - - // may need to cut off old tokens for sliding window - if (data_swa) { - if (pos - lctx.kv_self.cells[i].pos >= (int32_t)hparams.n_swa) { + for (int i = 0; i < n_kv; ++i) { + float f; + if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { f = -INFINITY; + } else { + if (hparams.use_alibi) { + f = -std::abs(kv_self.cells[i].pos - pos); + } else { + f = 0.0f; + } + } + + if (data) { + data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; + } + + // may need to cut off old tokens for sliding window + if (data_swa) { + if (pos - kv_self.cells[i].pos >= (int32_t)hparams.n_swa) { + f = -INFINITY; + } + data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; } - data_swa[h*(n_kv*n_tokens) + j*n_kv + i] = f; } } } @@ -14129,8 +15650,10 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } } else { + const int64_t n_tokens = batch.n_tokens; + const int64_t n_seq_tokens = batch.n_seq_tokens; + const int64_t n_seqs = batch.n_seqs; // when using kv cache, the mask needs to match the kv cache size - const int64_t n_tokens = batch.n_tokens; const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); @@ -14138,27 +15661,35 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { float * data = (float *) lctx.inp_KQ_mask->data; for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_seq_id seq_id = batch.seq_id[j][0]; + for (int s1 = 0; s1 < n_seqs; ++s1) { + const llama_seq_id seq_id = batch.seq_id[s1][0]; - for (int i = 0; i < n_tokens; ++i) { - float f = -INFINITY; - for (int s = 0; s < batch.n_seq_id[i]; ++s) { - if (batch.seq_id[i][s] == seq_id) { - if (hparams.use_alibi) { - f = -std::abs(batch.pos[i] - batch.pos[j]); - } else { - f = 0.0f; + for (int j = 0; j < n_seq_tokens; ++j) { + const int32_t tj = s1*n_seq_tokens + j; + + for (int s0 = 0; s0 < n_seqs; ++s0) { + for (int i = 0; i < n_seq_tokens; ++i) { + const int32_t ti = s0*n_seq_tokens + i; + float f = -INFINITY; + + for (int s = 0; s < batch.n_seq_id[s0]; ++s) { + if (batch.seq_id[s0][s] == seq_id) { + if (hparams.use_alibi) { + f = -std::abs(batch.pos[ti] - batch.pos[tj]); + } else { + f = 0.0f; + } + break; + } } - break; + + data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; } } - data[h*(n_tokens*n_tokens) + j*n_stride + i] = f; - } - - for (int i = n_tokens; i < n_stride; ++i) { - data[h*(n_tokens*n_tokens) + j*n_stride + i] = -INFINITY; + for (int i = n_tokens; i < n_stride; ++i) { + data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; + } } } } @@ -14166,7 +15697,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { - const int64_t n_tokens = batch.n_tokens; + const int64_t n_tokens = batch.n_tokens; + const int64_t n_seq_tokens = batch.n_seq_tokens; + const int64_t n_seqs = batch.n_seqs; GGML_ASSERT(lctx.inp_mean); GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer)); @@ -14175,12 +15708,14 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean)); std::vector sum(n_tokens, 0); - for (int i = 0; i < n_tokens; ++i) { - const llama_seq_id seq_id = batch.seq_id[i][0]; + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = batch.seq_id[s][0]; + + // TODO: adapt limits to n_seqs when batch.equal_seqs is true GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); - sum[seq_id] += 1; + sum[seq_id] += batch.n_seq_tokens; } std::vector div(n_tokens, 0.0f); @@ -14191,14 +15726,19 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } - for (int i = 0; i < n_tokens; ++i) { - const llama_seq_id seq_id = batch.seq_id[i][0]; - data[seq_id*n_tokens + i] = div[seq_id]; + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = batch.seq_id[s][0]; + + for (int i = 0; i < n_seq_tokens; ++i) { + data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id]; + } } } if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_CLS) { - const int64_t n_tokens = batch.n_tokens; + const int64_t n_tokens = batch.n_tokens; + const int64_t n_seq_tokens = batch.n_seq_tokens; + const int64_t n_seqs = batch.n_seqs; GGML_ASSERT(lctx.inp_cls); GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer)); @@ -14206,20 +15746,26 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { uint32_t * data = (uint32_t *) lctx.inp_cls->data; memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls)); - for (int i = 0; i < n_tokens; ++i) { - const llama_seq_id seq_id = batch.seq_id[i][0]; - const llama_pos pos = batch.pos[i]; + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = batch.seq_id[s][0]; + // TODO: adapt limits to n_seqs when batch.equal_seqs is true GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS"); - if (pos == 0) { - data[seq_id] = i; + for (int i = 0; i < n_seq_tokens; ++i) { + const llama_pos pos = batch.pos[s*n_seq_tokens + i]; + + if (pos == 0) { + data[seq_id] = s*n_seq_tokens + i; + } } } } if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { - const int64_t n_tokens = batch.n_tokens; + const int64_t n_tokens = batch.n_tokens; + const int64_t n_seq_tokens = batch.n_seq_tokens; + const int64_t n_seqs = batch.n_seqs; GGML_ASSERT(lctx.inp_cls); GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer)); @@ -14230,15 +15776,19 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { std::vector last_pos(n_tokens, -1); std::vector last_row(n_tokens, -1); - for (int i = 0; i < n_tokens; ++i) { - const llama_seq_id seq_id = batch.seq_id[i][0]; - const llama_pos pos = batch.pos[i]; + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = batch.seq_id[s][0]; + // TODO: adapt limits to n_seqs when batch.equal_seqs is true GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST"); - if (pos >= last_pos[seq_id]) { - last_pos[seq_id] = pos; - last_row[seq_id] = i; + for (int i = 0; i < n_seq_tokens; ++i) { + const llama_pos pos = batch.pos[s*n_seq_tokens + i]; + + if (pos >= last_pos[seq_id]) { + last_pos[seq_id] = pos; + last_row[seq_id] = s*n_seq_tokens + i; + } } } @@ -14256,41 +15806,39 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_mask->buffer)); float * data = (float *) lctx.inp_s_mask->data; - // states which are not affected by the current batch are left untouched + // clear unused states for (int i = 0; i < n_kv; ++i) { - llama_seq_id seq_id = i + lctx.kv_self.head; - llama_kv_cell & kv_cell = lctx.kv_self.cells[seq_id]; - bool has_self_seq = kv_cell.has_seq_id(seq_id); + uint32_t cell_id = i + kv_self.head; + llama_kv_cell & kv_cell = lctx.kv_self.cells[cell_id]; - data[i] = (float) has_self_seq; + data[i] = (float) (kv_cell.src >= 0); - // ensure current sequences will be kept - if (!has_self_seq && kv_cell.pos >= 0) { - kv_cell.seq_id.insert(seq_id); + // only clear once + if (kv_cell.src < 0) { + kv_cell.src = cell_id; } } } - // For Mamba (and other recurrent architectures), - // update the correct state(s)/sequence(s) for each token of the batch. - // Like with the KQ_mask, if a token in the batch has multiple sequences, - // they are assumed to be equivalent (not here, but in ggml_ssm_scan and ggml_ssm_conv). - if (lctx.inp_s_seq) { - const int64_t n_tokens = batch.n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_seq->buffer)); - int32_t * data = (int32_t *) lctx.inp_s_seq->data; + if (lctx.inp_s_copy) { + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_copy->buffer)); + int32_t * data = (int32_t *) lctx.inp_s_copy->data; - for (int j = 0; j < n_tokens; ++j) { - const int32_t n_seq = batch.n_seq_id[j]; - GGML_ASSERT(0 < n_seq); // a token should be part of at least 1 sequence + // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n + for (uint32_t i = 0; i < n_kv; ++i) { + const uint32_t cell_id = i + kv_self.head; + llama_kv_cell & kv_cell = lctx.kv_self.cells[cell_id]; - for (int i = 0; i < n_kv; ++i) { - if (i < n_seq) { - // for this type of model, the head is the minimum seq_id of the batch - data[j*n_kv + i] = batch.seq_id[j][i] - kv_self.head; - } else { - data[j*n_kv + i] = -1; - } + // prevent out-of-bound sources + if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self.size) { + kv_cell.src = cell_id; + } + + data[i] = kv_cell.src; + + // ensure copy only happens once + if (kv_cell.src != (int32_t) cell_id) { + kv_cell.src = cell_id; } } } @@ -14300,6 +15848,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { const int64_t n_tokens = batch.n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer)); + GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing int32_t * data = (int32_t *) lctx.inp_pos_bucket->data; @@ -14335,6 +15884,7 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { const int64_t n_tokens = batch.n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer)); + GGML_ASSERT(!batch.equal_seqs); // TODO: use batch.n_seqs instead of failing float * data = (float *) lctx.inp_KQ_mask_cross->data; @@ -14375,7 +15925,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { // TODO: use a per-batch flag for logits presence instead const bool has_logits = cparams.causal_attn; - const bool has_embd = lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE)); + const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0; const size_t embd_size = has_embd ? n_embd*n_outputs_max : 0; @@ -14428,11 +15978,49 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { return n_outputs_max; } +// make the outputs have the same order they had in the user-provided batch +static void llama_output_reorder(struct llama_context * ctx) { + std::vector & out_ids = ctx->sbatch.out_ids; + if (!out_ids.empty()) { + uint32_t n_vocab = ctx->model.hparams.n_vocab; + uint32_t n_embd = ctx->model.hparams.n_embd; + int32_t n_outputs = ctx->n_outputs; + GGML_ASSERT((size_t) n_outputs == out_ids.size()); + // TODO: is there something more efficient which also minimizes swaps? + // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort) + for (int32_t i = 0; i < n_outputs - 1; ++i) { + int32_t j_min = i; + for (int32_t j = i + 1; j < n_outputs; ++j) { + if (out_ids[j] < out_ids[j_min]) { + j_min = j; + } + } + if (j_min == i) { continue; } + std::swap(out_ids[i], out_ids[j_min]); + if (ctx->logits_size > 0) { + for (uint32_t k = 0; k < n_vocab; k++) { + std::swap(ctx->logits[i*n_vocab + k], ctx->logits[j_min*n_vocab + k]); + } + } + if (ctx->embd_size > 0) { + for (uint32_t k = 0; k < n_embd; k++) { + std::swap(ctx->embd[i*n_embd + k], ctx->embd[j_min*n_embd + k]); + } + } + } + std::fill(ctx->output_ids.begin(), ctx->output_ids.end(), -1); + for (int32_t i = 0; i < n_outputs; ++i) { + ctx->output_ids[out_ids[i]] = i; + } + out_ids.clear(); + } +} static void llama_graph_compute( - llama_context & lctx, - ggml_cgraph * gf, - int n_threads) { + llama_context & lctx, + ggml_cgraph * gf, + int n_threads, + ggml_threadpool * threadpool) { #ifdef GGML_USE_METAL if (ggml_backend_is_metal(lctx.backend_metal)) { ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads); @@ -14441,6 +16029,7 @@ static void llama_graph_compute( if (lctx.backend_cpu != nullptr) { ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads); + ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool); ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data); } #ifdef GGML_USE_BLAS @@ -14500,15 +16089,11 @@ static int llama_decode_internal( const auto n_ubatch = cparams.n_ubatch; - // TODO: simplify or deprecate - std::vector pos; - std::vector n_seq_id; - std::vector seq_id_arr; - std::vector> seq_id; - // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE; + lctx.embd_seq.clear(); + // count outputs if (batch_all.logits && !embd_pooled) { for (uint32_t i = 0; i < n_tokens_all; ++i) { @@ -14521,55 +16106,42 @@ static int llama_decode_internal( n_outputs = 1; } + lctx.sbatch.from_batch(batch_all, n_embd, + /* simple_split */ !kv_self.recurrent, + /* logits_all */ n_outputs == n_tokens_all); + // reserve output buffer if (llama_output_reserve(lctx, n_outputs) < n_outputs) { LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_outputs); return -2; }; - // set output mappings - if (batch_all.logits) { - int32_t i_logits = 0; - for (uint32_t i = 0; i < n_tokens_all; ++i) { - if (batch_all.logits[i]) { - lctx.output_ids[i] = i_logits++; + while (lctx.sbatch.n_tokens > 0) { + llama_ubatch ubatch; + if (kv_self.recurrent) { + if (embd_pooled) { + // Pooled embeddings cannot be split across ubatches (yet) + ubatch = lctx.sbatch.split_seq(n_ubatch); + } else { + // recurrent model architectures are easier to implement + // with equal-length sequences + ubatch = lctx.sbatch.split_equal(n_ubatch); } + } else { + ubatch = lctx.sbatch.split_simple(n_ubatch); } - } else { - for (uint32_t i = 0; i < n_outputs; ++i) { - lctx.output_ids[i] = i; - } - } - - for (uint32_t cur_token = 0; cur_token < n_tokens_all; cur_token += n_ubatch) { - const uint32_t n_tokens = std::min(n_ubatch, n_tokens_all - cur_token); - llama_batch u_batch = { - /* .n_tokens = */ (int32_t) n_tokens, - /* .token = */ batch_all.token ? batch_all.token + cur_token : nullptr, - /* .embd = */ batch_all.embd ? batch_all.embd + cur_token*n_embd : nullptr, - /* .pos = */ batch_all.pos ? batch_all.pos + cur_token : nullptr, - /* .n_seq_id = */ batch_all.n_seq_id ? batch_all.n_seq_id + cur_token : nullptr, - /* .seq_id = */ batch_all.seq_id ? batch_all.seq_id + cur_token : nullptr, - /* .logits = */ batch_all.logits ? batch_all.logits + cur_token : nullptr, - /* .all_pos_0 = */ batch_all.all_pos_0 + (llama_pos) cur_token*batch_all.all_pos_1, - /* .all_pos_1 = */ batch_all.all_pos_1, - /* .all_seq_id = */ batch_all.all_seq_id, - }; + const uint32_t n_tokens = ubatch.n_tokens; // count the outputs in this u_batch { int32_t n_outputs_new = 0; - if (u_batch.logits && !embd_pooled) { - for (uint32_t i = 0; i < n_tokens; i++) { - n_outputs_new += u_batch.logits[i] != 0; - } - } else if (n_outputs == n_tokens_all) { + if (n_outputs == n_tokens_all) { n_outputs_new = n_tokens; } else { - // keep last output only - if (cur_token + n_tokens >= n_tokens_all) { - n_outputs_new = 1; + GGML_ASSERT(ubatch.output); + for (uint32_t i = 0; i < n_tokens; i++) { + n_outputs_new += (int32_t) (ubatch.output[i] != 0); } } @@ -14578,34 +16150,10 @@ static int llama_decode_internal( } int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch; + GGML_ASSERT(n_threads > 0); - // helpers for smoother batch API transition - // after deprecating the llama_eval calls, these will be removed - if (u_batch.pos == nullptr) { - pos.resize(n_tokens); - for (uint32_t i = 0; i < n_tokens; i++) { - pos[i] = u_batch.all_pos_0 + i*u_batch.all_pos_1; - } - - u_batch.pos = pos.data(); - } - - if (u_batch.seq_id == nullptr) { - n_seq_id.resize(n_tokens); - seq_id.resize(n_tokens); - seq_id_arr.resize(n_tokens); - for (uint32_t i = 0; i < n_tokens; i++) { - n_seq_id[i] = 1; - seq_id[i].resize(1); - seq_id[i][0] = u_batch.all_seq_id; - seq_id_arr[i] = seq_id[i].data(); - } - - u_batch.n_seq_id = n_seq_id.data(); - u_batch.seq_id = seq_id_arr.data(); - } - // non-causal masks do not use the KV cache if (hparams.causal_attn) { llama_kv_cache_update(&lctx); @@ -14616,7 +16164,7 @@ static int llama_decode_internal( kv_self.head = 0; } - if (!llama_kv_cache_find_slot(kv_self, u_batch)) { + if (!llama_kv_cache_find_slot(kv_self, ubatch)) { return 1; } @@ -14635,7 +16183,7 @@ static int llama_decode_internal( ggml_backend_sched_reset(lctx.sched); ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); - ggml_cgraph * gf = llama_build_graph(lctx, u_batch, false); + ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false); // the output is always the last tensor in the graph struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1]; @@ -14654,8 +16202,7 @@ static int llama_decode_internal( break; } } - GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor"); - } else { + } else { embd = nullptr; // do not extract embeddings when not needed GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor"); } @@ -14663,14 +16210,13 @@ static int llama_decode_internal( if (!cparams.causal_attn) { res = nullptr; // do not extract logits when not needed } - // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); ggml_backend_sched_alloc_graph(lctx.sched, gf); - llama_set_inputs(lctx, u_batch); + llama_set_inputs(lctx, ubatch); - llama_graph_compute(lctx, gf, n_threads); + llama_graph_compute(lctx, gf, n_threads, threadpool); // update the kv ring buffer { @@ -14726,12 +16272,11 @@ static int llama_decode_internal( case LLAMA_POOLING_TYPE_CLS: case LLAMA_POOLING_TYPE_LAST: { - // extract sequence embeddings + // extract sequence embeddings (cleared before processing each batch) auto & embd_seq_out = lctx.embd_seq; - embd_seq_out.clear(); - for (uint32_t i = 0; i < n_tokens; i++) { - const llama_seq_id seq_id = u_batch.seq_id[i][0]; + for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { continue; } @@ -14748,6 +16293,25 @@ static int llama_decode_internal( n_outputs_prev += lctx.n_outputs; } + // set output mappings + { + bool sorted_output = true; + + GGML_ASSERT(lctx.sbatch.out_ids.size() == n_outputs); + + for (size_t i = 0; i < n_outputs; ++i) { + size_t out_id = lctx.sbatch.out_ids[i]; + lctx.output_ids[out_id] = i; + if (out_id != i) { + sorted_output = false; + } + } + + if (sorted_output) { + lctx.sbatch.out_ids.clear(); + } + } + // set to total number of outputs in the batch, for use in llama_get_logits_ith lctx.n_outputs = n_outputs; @@ -14812,11 +16376,9 @@ static int llama_encode_internal( const int64_t n_embd = hparams.n_embd; - // TODO: simplify or deprecate - std::vector pos; - std::vector n_seq_id; - std::vector seq_id_arr; - std::vector> seq_id; + lctx.sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true); + + const llama_ubatch ubatch = lctx.sbatch.split_simple(n_tokens); // reserve output buffer if (llama_output_reserve(lctx, n_tokens) < n_tokens) { @@ -14831,70 +16393,98 @@ static int llama_encode_internal( lctx.inp_embd_enc = NULL; lctx.n_outputs = n_tokens; - const int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch; + ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch; + GGML_ASSERT(n_threads > 0); - // helpers for smoother batch API transition - // after deprecating the llama_eval calls, these will be removed - if (batch.pos == nullptr) { - pos.resize(n_tokens); - for (uint32_t i = 0; i < n_tokens; i++) { - pos[i] = batch.all_pos_0 + i*batch.all_pos_1; - } - - batch.pos = pos.data(); - } - - if (batch.seq_id == nullptr) { - n_seq_id.resize(n_tokens); - seq_id.resize(n_tokens); - seq_id_arr.resize(n_tokens); - for (uint32_t i = 0; i < n_tokens; i++) { - n_seq_id[i] = 1; - seq_id[i].resize(1); - seq_id[i][0] = batch.all_seq_id; - seq_id_arr[i] = seq_id[i].data(); - } - - batch.n_seq_id = n_seq_id.data(); - batch.seq_id = seq_id_arr.data(); - } - ggml_backend_sched_reset(lctx.sched); ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); - ggml_cgraph * gf = llama_build_graph(lctx, batch, false); + ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false); // the output embeddings after the final encoder normalization - struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 1]; + struct ggml_tensor * embd = nullptr; - GGML_ASSERT(strcmp(embd->name, "result_norm") == 0); + // there are two cases here + if (llama_model_has_decoder(&lctx.model)) { + // first case is an encoder-decoder T5 model where embeddings are passed to decoder + embd = gf->nodes[gf->n_nodes - 1]; + GGML_ASSERT(strcmp(embd->name, "result_norm") == 0 && "missing result_output tensor"); + } else { + // second case is an encoder-only T5 model + if (cparams.embeddings) { + // only output embeddings if required + embd = gf->nodes[gf->n_nodes - 1]; + if (strcmp(embd->name, "result_embd_pooled") != 0) { + embd = gf->nodes[gf->n_nodes - 2]; + } + GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor"); + } + } ggml_backend_sched_alloc_graph(lctx.sched, gf); - llama_set_inputs(lctx, batch); + llama_set_inputs(lctx, ubatch); - llama_graph_compute(lctx, gf, n_threads); + llama_graph_compute(lctx, gf, n_threads, threadpool); // extract embeddings if (embd) { ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd); GGML_ASSERT(backend_embd != nullptr); - // extract token embeddings - GGML_ASSERT(lctx.embd != nullptr); + if (llama_model_has_decoder(&lctx.model)) { + lctx.embd_enc.resize(n_tokens*n_embd); + float * embd_out = lctx.embd_enc.data(); - lctx.embd_enc.resize(n_tokens*n_embd); - float * embd_out = lctx.embd_enc.data(); + ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float)); + GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits - ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float)); + // remember the sequence ids used during the encoding - needed for cross attention later + lctx.seq_ids_enc.resize(n_tokens); + for (uint32_t i = 0; i < n_tokens; i++) { + for (int s = 0; s < ubatch.n_seq_id[i]; s++) { + llama_seq_id seq_id = ubatch.seq_id[i][s]; + lctx.seq_ids_enc[i].insert(seq_id); + } + } + } else { + GGML_ASSERT(lctx.embd != nullptr); - // remember the sequence ids used during the encoding - needed for cross attention later - lctx.seq_ids_enc.resize(n_tokens); - for (uint32_t i = 0; i < n_tokens; i++) { - for (int s = 0; s < batch.n_seq_id[i]; s++) { - llama_seq_id seq_id = batch.seq_id[i][s]; - lctx.seq_ids_enc[i].insert(seq_id); + switch (cparams.pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + // extract token embeddings + GGML_ASSERT(lctx.embd != nullptr); + float * embd_out = lctx.embd; + + GGML_ASSERT(n_tokens*n_embd <= (int64_t) lctx.embd_size); + ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float)); + } break; + case LLAMA_POOLING_TYPE_MEAN: + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_LAST: + { + // extract sequence embeddings + auto & embd_seq_out = lctx.embd_seq; + embd_seq_out.clear(); + + GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits + + for (uint32_t i = 0; i < n_tokens; i++) { + const llama_seq_id seq_id = ubatch.seq_id[i][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(n_embd); + ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_UNSPECIFIED: + { + GGML_ABORT("unknown pooling type"); + } } } } @@ -15119,7 +16709,7 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) { ggml_cgraph * gf = llama_build_graph_defrag(lctx, ids); - llama_graph_compute(lctx, gf, lctx.cparams.n_threads); + llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool); #endif //const int64_t t_end = ggml_time_us(); @@ -15145,7 +16735,7 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) { llama_set_k_shift(lctx); - llama_graph_compute(lctx, gf, lctx.cparams.n_threads); + llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool); need_reserve = true; } @@ -15161,32 +16751,6 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) { } } - if (lctx.kv_self.recurrent && lctx.kv_self.do_copy) { - { - ggml_backend_sched_reset(lctx.sched); - - ggml_cgraph * gf = llama_build_graph_s_copy(lctx); - - ggml_backend_sched_alloc_graph(lctx.sched, gf); - - llama_set_s_copy(lctx); - - llama_graph_compute(lctx, gf, lctx.cparams.n_threads); - - need_reserve = true; - } - - { - auto & kv_self = lctx.kv_self; - - kv_self.do_copy = false; - - for (uint32_t i = 0; i < kv_self.size; ++i) { - kv_self.cells[i].src = i; - } - } - } - // defragment the KV cache if needed if (lctx.kv_self.do_defrag) { llama_kv_cache_defrag_internal(lctx); @@ -15200,10 +16764,11 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) { if (need_reserve) { // TODO: extract to a function // build worst-case graph - int n_tokens = (int)std::min(lctx.cparams.n_ctx, lctx.cparams.n_ubatch); - int n_past = lctx.cparams.n_ctx - n_tokens; + uint32_t n_seqs = 1; // TODO: worst-case number of sequences + uint32_t n_tokens = std::min(lctx.cparams.n_ctx, lctx.cparams.n_ubatch); llama_token token = llama_token_bos(&lctx.model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph - ggml_cgraph * gf = llama_build_graph(lctx, llama_batch_get_one(&token, n_tokens, n_past, 0), true); + llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + ggml_cgraph * gf = llama_build_graph(lctx, ubatch, true); // initialize scheduler with the worst-case graph ggml_backend_sched_reset(lctx.sched); @@ -15330,7 +16895,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) { if (n_expert > 1) { - // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly + // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but occasionally randomly // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work // for getting the current layer as I initially thought, and we need to resort to parsing the // tensor name. @@ -15595,6 +17160,9 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n case GGML_TYPE_Q6_K: new_type = GGML_TYPE_Q8_0; break; default: throw std::runtime_error("\nUnsupported tensor size encountered\n"); } + if (tensor->ne[0] % ggml_blck_size(new_type) != 0) { + new_type = GGML_TYPE_F16; + } LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type)); ++qs.n_fallback; } @@ -15786,7 +17354,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // TODO: avoid hardcoded tensor names - use the TN_* constants if (name.find("attn_v.weight") != std::string::npos || - name.find("attn_qkv.weight") != std::string::npos) { + name.find("attn_qkv.weight") != std::string::npos || + name.find("attn_kv_b.weight")!= std::string::npos) { ++qs.n_attention_wv; } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) { qs.has_output = true; @@ -15796,12 +17365,15 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer; // sanity checks - // - // - qs.n_attention_wv == 0 for Mamba models - // - qs.n_attention_wv == model.hparams.n_layer for Transformer models - // - qs.n_attention_wv == 3 * model.hparams.n_layer for Encoder-Decoder models - // - GGML_ASSERT((qs.n_attention_wv == 0 || qs.n_attention_wv == (int)model.hparams.n_layer || qs.n_attention_wv == 3 * (int)model.hparams.n_layer) && "n_attention_wv is unexpected"); + { + const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin(); + // attention layers have a non-zero number of kv heads + int32_t n_attn_layer = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0); + if (llama_model_has_encoder(&model)) { + n_attn_layer *= 3; + } + GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected"); + } size_t total_size_org = 0; size_t total_size_new = 0; @@ -15923,8 +17495,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // do not quantize Mamba's small yet 2D weights // NOTE: can't use LLM_TN here because the layer number is not known quantize &= name.find("ssm_conv1d.weight") == std::string::npos; - quantize &= name.find("ssm_x.weight") == std::string::npos; - quantize &= name.find("ssm_dt.weight") == std::string::npos; + + // do not quantize RWKV's time_mix_first tensors + quantize &= name.find("time_mix_first.weight") == std::string::npos; + quantize &= name.find("time_mix_w1.weight") == std::string::npos; + quantize &= name.find("time_mix_w2.weight") == std::string::npos; // do not quantize relative position bias (T5) quantize &= name.find("attn_rel_b.weight") == std::string::npos; @@ -16413,6 +17988,19 @@ void llama_numa_init(enum ggml_numa_strategy numa) { } } +void llama_attach_threadpool( + struct llama_context * ctx, + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch) { + ctx->threadpool = threadpool; + ctx->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool; +} + +void llama_detach_threadpool(struct llama_context * ctx) { + ctx->threadpool = nullptr; + ctx->threadpool_batch = nullptr; +} + void llama_backend_free(void) { ggml_quantize_free(); } @@ -16498,12 +18086,6 @@ struct llama_context * llama_new_context_with_model( params.flash_attn = false; } - if (params.flash_attn && model->hparams.attn_soft_cap) { - LLAMA_LOG_WARN("%s: flash_attn is not compatible with attn_soft_cap - forcing off\n", __func__); - params.flash_attn = false; - } - - if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) { LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__); params.flash_attn = false; @@ -16604,13 +18186,15 @@ struct llama_context * llama_new_context_with_model( ctx->sampling.rng = std::mt19937(params.seed); ctx->logits_all = params.logits_all; + // build worst-case graph for encoder if a model contains encoder + ctx->is_encoding = llama_model_has_encoder(model); uint32_t kv_size = cparams.n_ctx; ggml_type type_k = params.type_k; ggml_type type_v = params.type_v; // Mamba only needs a constant number of KV cache cells per sequence - if (model->arch == LLM_ARCH_MAMBA) { + if (llama_model_is_recurrent(model)) { // Mamba needs at least as many KV cells as there are sequences kept at any time kv_size = std::max((uint32_t) 1, params.n_seq_max); // it's probably best to keep as much precision as possible for the states @@ -16842,10 +18426,11 @@ struct llama_context * llama_new_context_with_model( } // build worst-case graph - int n_tokens = (int)std::min(cparams.n_ctx, cparams.n_ubatch); - int n_past = cparams.n_ctx - n_tokens; + uint32_t n_seqs = 1; // TODO: worst-case number of sequences + uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); llama_token token = llama_token_bos(&ctx->model); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph - ggml_cgraph * gf = llama_build_graph(*ctx, llama_batch_get_one(&token, n_tokens, n_past, 0), true); + llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + ggml_cgraph * gf = llama_build_graph(*ctx, ubatch, true); // initialize scheduler with the worst-case graph if (!ggml_backend_sched_reserve(ctx->sched, gf)) { @@ -16918,7 +18503,9 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_MAMBA: case LLM_ARCH_JINA_BERT_V2: case LLM_ARCH_T5: + case LLM_ARCH_T5ENCODER: case LLM_ARCH_JAIS: + case LLM_ARCH_RWKV6: return LLAMA_ROPE_TYPE_NONE; // use what we call a normal RoPE, operating on pairs of consecutive head values @@ -16956,6 +18543,8 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: case LLM_ARCH_CODESHELL: + case LLM_ARCH_NEMOTRON: + case LLM_ARCH_EXAONE: return LLAMA_ROPE_TYPE_NEOX; // all model arches should be listed explicitly here @@ -17065,8 +18654,16 @@ struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const ch bool llama_model_has_encoder(const struct llama_model * model) { switch (model->arch) { - case LLM_ARCH_T5: return true; - default: return false; + case LLM_ARCH_T5: return true; + case LLM_ARCH_T5ENCODER: return true; + default: return false; + } +} + +bool llama_model_has_decoder(const struct llama_model * model) { + switch (model->arch) { + case LLM_ARCH_T5ENCODER: return false; + default: return true; } } @@ -17074,6 +18671,14 @@ llama_token llama_model_decoder_start_token(const struct llama_model * model) { return model->hparams.dec_start_token_id; } +bool llama_model_is_recurrent(const struct llama_model * model) { + switch (model->arch) { + case LLM_ARCH_MAMBA: return true; + case LLM_ARCH_RWKV6: return true; + default: return false; + } +} + uint32_t llama_model_quantize( const char * fname_inp, const char * fname_out, @@ -17369,6 +18974,7 @@ bool llama_save_session_file(struct llama_context * ctx, const char * path_sessi // TODO: replace all non-fatal assertions with returned errors or exceptions struct llama_data_write { virtual void write(const void * src, size_t size) = 0; + virtual void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) = 0; virtual size_t get_size_written() = 0; virtual ~llama_data_write() = default; @@ -17394,7 +19000,9 @@ struct llama_data_write { write_string(rng_str); } - void write_output_ids(const struct llama_context * ctx) { + void write_output_ids(struct llama_context * ctx) { + llama_output_reorder(ctx); + const uint32_t n_outputs = ctx->n_outputs; std::vector output_pos; @@ -17491,9 +19099,8 @@ struct llama_data_write { // Read each range of cells of k_size length each into tmp_buf and write out for (const auto & range : cell_ranges) { const size_t range_size = range.second - range.first; - tmp_buf.resize(range_size * k_size_row); - ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), range.first * k_size_row, range_size * k_size_row); - write(tmp_buf.data(), tmp_buf.size()); + const size_t buf_size = range_size * k_size_row; + write_tensor_data(kv_self.k_l[il], range.first * k_size_row, buf_size); } } @@ -17512,9 +19119,8 @@ struct llama_data_write { // Read each range of cells of v_size length each into tmp_buf and write out for (const auto & range : cell_ranges) { const size_t range_size = range.second - range.first; - tmp_buf.resize(range_size * v_size_row); - ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), range.first * v_size_row, range_size * v_size_row); - write(tmp_buf.data(), tmp_buf.size()); + const size_t buf_size = range_size * v_size_row; + write_tensor_data(kv_self.v_l[il], range.first * v_size_row, buf_size); } } } else { @@ -17540,9 +19146,8 @@ struct llama_data_write { for (const auto & range : cell_ranges) { const size_t range_size = range.second - range.first; const size_t src_offset = (range.first + j * kv_size) * v_size_el; - tmp_buf.resize(range_size * v_size_el); - ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), src_offset, tmp_buf.size()); - write(tmp_buf.data(), tmp_buf.size()); + const size_t buf_size = range_size * v_size_el; + write_tensor_data(kv_self.v_l[il], src_offset, buf_size); } } } @@ -17685,8 +19290,11 @@ struct llama_data_read { llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); - llama_batch batch = llama_batch_init(cell_count, 0, 1); + llama_ubatch batch = ctx->sbatch.reserve_ubatch(cell_count, /* has_embd */ false); batch.n_tokens = cell_count; + batch.n_seq_tokens = cell_count; + batch.n_seqs = 1; + for (uint32_t i = 0; i < cell_count; ++i) { llama_pos pos; uint32_t n_seq_id; @@ -17700,11 +19308,10 @@ struct llama_data_read { } batch.pos[i] = pos; - batch.n_seq_id[i] = 1; - batch.seq_id[i][0] = dest_seq_id; } + batch.n_seq_id[0] = 1; + batch.seq_id[0] = &dest_seq_id; if (!llama_kv_cache_find_slot(kv_self, batch)) { - llama_batch_free(batch); LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); return false; } @@ -17716,9 +19323,6 @@ struct llama_data_read { GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].pos == batch.pos[cell_count - 1]); GGML_ASSERT(kv_self.cells[kv_self.head].has_seq_id(dest_seq_id)); GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].has_seq_id(dest_seq_id)); - - // Cleanup - llama_batch_free(batch); } else { // whole KV cache restore @@ -17750,6 +19354,15 @@ struct llama_data_read { } cell.seq_id.insert(seq_id); + + if (kv_self.recurrent) { + int32_t & tail = kv_self.cells[seq_id].tail; + if (tail != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tail); + return false; + } + tail = i; + } } } @@ -17757,6 +19370,14 @@ struct llama_data_read { kv_self.used = cell_count; } + if (kv_self.recurrent) { + for (uint32_t i = 0; i < cell_count; ++i) { + uint32_t cell_id = kv_self.head + i; + // make sure the recurrent states will keep their restored state + kv_self.cells[cell_id].src = cell_id; + } + } + return true; } @@ -17901,12 +19522,14 @@ struct llama_data_write_dummy : llama_data_write { llama_data_write_dummy() {} - // TODO: avoid unnecessary calls to ggml_backend_tensor_get in a dummy context - void write(const void * /* src */, size_t size) override { size_written += size; } + void write_tensor_data(const struct ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override { + size_written += size; + } + size_t get_size_written() override { return size_written; } @@ -17929,6 +19552,16 @@ struct llama_data_write_buffer : llama_data_write { buf_size -= size; } + void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override { + if (size > buf_size) { + throw std::runtime_error("unexpectedly reached end of buffer"); + } + ggml_backend_tensor_get(tensor, ptr, offset, size); + ptr += size; + size_written += size; + buf_size -= size; + } + size_t get_size_written() override { return size_written; } @@ -17964,6 +19597,7 @@ struct llama_data_read_buffer : llama_data_read { struct llama_data_write_file : llama_data_write { llama_file * file; size_t size_written = 0; + std::vector temp_buffer; llama_data_write_file(llama_file * f) : file(f) {} @@ -17972,6 +19606,12 @@ struct llama_data_write_file : llama_data_write { size_written += size; } + void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override { + temp_buffer.resize(size); + ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size); + write(temp_buffer.data(), temp_buffer.size()); + } + size_t get_size_written() override { return size_written; } @@ -18279,16 +19919,16 @@ size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepa } } -void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch) { +void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) { ctx->cparams.n_threads = n_threads; ctx->cparams.n_threads_batch = n_threads_batch; } -uint32_t llama_n_threads(struct llama_context * ctx) { +int32_t llama_n_threads(struct llama_context * ctx) { return ctx->cparams.n_threads; } -uint32_t llama_n_threads_batch(struct llama_context * ctx) { +int32_t llama_n_threads_batch(struct llama_context * ctx) { return ctx->cparams.n_threads_batch; } @@ -18325,7 +19965,18 @@ struct llama_batch llama_batch_get_one( } struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) { - llama_batch batch = { 0, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, 0, 0, }; + llama_batch batch = { + /*n_tokens =*/ 0, + /*tokens =*/ nullptr, + /*embd =*/ nullptr, + /*pos =*/ nullptr, + /*n_seq_id =*/ nullptr, + /*seq_id =*/ nullptr, + /*logits =*/ nullptr, + /*all_pos_0 =*/ 0, + /*all_pos_1 =*/ 0, + /*all_seq_id =*/ 0, + }; if (embd) { batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd); @@ -18411,6 +20062,10 @@ void llama_synchronize(struct llama_context * ctx) { float * llama_get_logits(struct llama_context * ctx) { llama_synchronize(ctx); + // reorder logits for backward compatibility + // TODO: maybe deprecate this + llama_output_reorder(ctx); + return ctx->logits; } @@ -18455,6 +20110,10 @@ float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { float * llama_get_embeddings(struct llama_context * ctx) { llama_synchronize(ctx); + // reorder embeddings for backward compatibility + // TODO: maybe deprecate this + llama_output_reorder(ctx); + return ctx->embd; } @@ -18556,11 +20215,11 @@ llama_token llama_token_pad(const struct llama_model * model) { return llama_token_pad_impl(model->vocab); } -int32_t llama_add_bos_token(const struct llama_model * model) { +bool llama_add_bos_token(const struct llama_model * model) { return llama_add_bos_token_impl(model->vocab); } -int32_t llama_add_eos_token(const struct llama_model * model) { +bool llama_add_eos_token(const struct llama_model * model) { return llama_add_eos_token_impl(model->vocab); } @@ -18861,6 +20520,22 @@ static int32_t llama_chat_apply_template_internal( if (add_ass) { ss << "Assistant:"; } + } else if (tmpl == "exaone3" || (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]"))) { + // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb + // EXAONE-3.0-7.8B-Instruct + for (auto message : chat) { + std::string role(message->role); + if (role == "system") { + ss << "[|system|]" << trim(message->content) << "[|endofturn|]\n"; + } else if (role == "user") { + ss << "[|user|]" << trim(message->content) << "\n"; + } else if (role == "assistant") { + ss << "[|assistant|]" << trim(message->content) << "[|endofturn|]\n"; + } + } + if (add_ass) { + ss << "[|assistant|]"; + } } else { // template not supported return -1; diff --git a/llama/llama.go b/llama/llama.go index 132d7b33..c8999566 100644 --- a/llama/llama.go +++ b/llama/llama.go @@ -70,11 +70,10 @@ type ContextParams struct { func NewContextParams(numCtx int, threads int, flashAttention bool) ContextParams { params := C.llama_context_default_params() params.n_ctx = C.uint(numCtx) - params.n_threads = C.uint(runtime.NumCPU()) + params.n_threads = C.int(threads) params.n_threads_batch = params.n_threads params.embeddings = C.bool(true) params.flash_attn = C.bool(flashAttention) - params.n_threads = C.uint(threads) return ContextParams{c: params} } @@ -213,14 +212,8 @@ func (m *Model) TokenIsEog(token int) bool { return bool(C.llama_token_is_eog(m.c, C.llama_token(token))) } -func (m *Model) ShouldAddBOSToken() bool { - addBos := int(C.llama_add_bos_token(m.c)) - - if addBos != -1 { - return addBos != 0 - } else { - return C.llama_vocab_type(m.c) == C.LLAMA_VOCAB_TYPE_SPM - } +func (m *Model) AddBOSToken() bool { + return bool(C.llama_add_bos_token(m.c)) } func (m *Model) ApplyLoraFromFile(context *Context, loraPath string, scale float32, threads int) error { diff --git a/llama/llama.h b/llama/llama.h index cbd8eb7c..b0787fa2 100644 --- a/llama/llama.h +++ b/llama/llama.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -92,6 +92,7 @@ extern "C" { LLAMA_VOCAB_TYPE_BPE = 2, // GPT-2 tokenizer based on byte-level BPE LLAMA_VOCAB_TYPE_WPM = 3, // BERT tokenizer based on WordPiece LLAMA_VOCAB_TYPE_UGM = 4, // T5 tokenizer based on Unigram + LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization }; // pre-tokenization types @@ -119,15 +120,15 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_TEKKEN = 20, LLAMA_VOCAB_PRE_TYPE_SMOLLM = 21, LLAMA_VOCAB_PRE_TYPE_CODESHELL = 22, + LLAMA_VOCAB_PRE_TYPE_BLOOM = 23, + LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH = 24, + LLAMA_VOCAB_PRE_TYPE_EXAONE = 25, }; - // note: these values should be synchronized with ggml_rope - // TODO: maybe move this enum to ggml.h (ggml_rope_type) enum llama_rope_type { LLAMA_ROPE_TYPE_NONE = -1, - LLAMA_ROPE_TYPE_NORM = 0, - LLAMA_ROPE_TYPE_NEOX = 2, - LLAMA_ROPE_TYPE_GLM = 4, + LLAMA_ROPE_TYPE_NORM = 0, + LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX, }; enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file @@ -293,9 +294,9 @@ extern "C" { enum llama_split_mode split_mode; // how to split the model across multiple GPUs // main_gpu interpretation depends on split_mode: - // LLAMA_SPLIT_NONE: the GPU that is used for the entire model - // LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results - // LLAMA_SPLIT_LAYER: ignored + // LLAMA_SPLIT_MODE_NONE: the GPU that is used for the entire model + // LLAMA_SPLIT_MODE_ROW: the GPU that is used for small tensors and intermediate results + // LLAMA_SPLIT_MODE_LAYER: ignored int32_t main_gpu; // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices() @@ -330,8 +331,8 @@ extern "C" { uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode uint32_t n_ubatch; // physical maximum batch size uint32_t n_seq_max; // max number of sequences (i.e. distinct states for recurrent models) - uint32_t n_threads; // number of threads to use for generation - uint32_t n_threads_batch; // number of threads to use for batch processing + int32_t n_threads; // number of threads to use for generation + int32_t n_threads_batch; // number of threads to use for batch processing enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id @@ -371,7 +372,7 @@ extern "C" { int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() enum llama_ftype ftype; // quantize to this llama_ftype enum ggml_type output_tensor_type; // output tensor type - enum ggml_type token_embedding_type; // itoken embeddings tensor type + enum ggml_type token_embedding_type; // token embeddings tensor type bool allow_requantize; // allow quantizing non-f32/f16 tensors bool quantize_output_tensor; // quantize output.weight bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored @@ -454,6 +455,13 @@ extern "C" { //optional: LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); + // Optional: an auto threadpool gets created in ggml if not passed explicitly + LLAMA_API void llama_attach_threadpool( + struct llama_context * ctx, + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch); + LLAMA_API void llama_detach_threadpool(struct llama_context * ctx); + // Call once at the end of the program - currently only used for MPI LLAMA_API void llama_backend_free(void); @@ -530,10 +538,16 @@ extern "C" { // Returns true if the model contains an encoder that requires llama_encode() call LLAMA_API bool llama_model_has_encoder(const struct llama_model * model); + // Returns true if the model contains a decoder that requires llama_decode() call + LLAMA_API bool llama_model_has_decoder(const struct llama_model * model); + // For encoder-decoder models, this function returns id of the token that must be provided // to the decoder to start generating output sequence. For other models, it returns -1. LLAMA_API llama_token llama_model_decoder_start_token(const struct llama_model * model); + // Returns true if the model is recurrent (like Mamba, RWKV, etc.) + LLAMA_API bool llama_model_is_recurrent(const struct llama_model * model); + // Returns 0 on success LLAMA_API uint32_t llama_model_quantize( const char * fname_inp, @@ -857,13 +871,13 @@ extern "C" { // Set the number of threads used for decoding // n_threads is the number of threads used for generation (single token) // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) - LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); + LLAMA_API void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch); // Get the number of threads used for generation of a single token. - LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx); + LLAMA_API int32_t llama_n_threads(struct llama_context * ctx); // Get the number of threads used for prompt and batch processing (multiple token). - LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx); + LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx); // Set whether the model is in embeddings mode or not // If true, embeddings will be returned but logits will not @@ -938,11 +952,8 @@ extern "C" { LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line LLAMA_API llama_token llama_token_pad(const struct llama_model * model); // padding - // Returns -1 if unknown, 1 for true or 0 for false. - LLAMA_API int32_t llama_add_bos_token(const struct llama_model * model); - - // Returns -1 if unknown, 1 for true or 0 for false. - LLAMA_API int32_t llama_add_eos_token(const struct llama_model * model); + LLAMA_API bool llama_add_bos_token(const struct llama_model * model); + LLAMA_API bool llama_add_eos_token(const struct llama_model * model); // Codellama infill tokens LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix diff --git a/llama/llava.cpp b/llama/llava.cpp index d54175b3..128e46b9 100644 --- a/llama/llava.cpp +++ b/llama/llava.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -228,6 +228,33 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector return true; } +static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) { + int width = image->nx; + int height = image->ny; + int num_patches = (height / patch_size) * (width / patch_size); + clip_image_f32 * patch = clip_image_f32_init(); + patch->nx = patch_size * num_patches; + patch->ny = patch_size; + patch->buf.resize(3 * patch->nx * patch->ny); + + int patch_index = 0; + + for (int i = 0; i < height; i += patch_size) { + for (int j = 0; j < width; j += patch_size) { + for (int pi = 0; pi < patch_size; ++pi) { + for (int pj = 0; pj < patch_size; ++pj) { + int input_index = ((i + pi) * width + (j + pj)) * 3; + int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3; + patch->buf[output_index] = image->buf[input_index]; + patch->buf[output_index+1] = image->buf[input_index+1]; + patch->buf[output_index+2] = image->buf[input_index+2]; + } + } + patch_index++; + } + } + return patch; +} static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) { // std::vector img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336 @@ -244,7 +271,51 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip); - if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { + if (clip_is_minicpmv(ctx_clip)) { + std::vector image_embd_v; + image_embd_v.resize(img_res_v.size); + struct clip_image_size * load_image_size = clip_image_size_init(); + for (size_t i = 0; i < img_res_v.size; i++) { + const int64_t t_img_enc_step_start_us = ggml_time_us(); + image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); + int patch_size=14; + load_image_size->width = img_res_v.data[i].nx; + load_image_size->height = img_res_v.data[i].ny; + clip_add_load_image_size(ctx_clip, load_image_size); + bool encoded = false; + int has_minicpmv_projector = clip_is_minicpmv(ctx_clip); + if (has_minicpmv_projector == 2) { + encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]); + } + else if (has_minicpmv_projector == 3) { + encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); + } + if (!encoded) { + LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size); + return false; + } + const int64_t t_img_enc_steop_batch_us = ggml_time_us(); + LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0); + } + const int64_t t_img_enc_batch_us = ggml_time_us(); + LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); + + int n_img_pos_out = 0; + for (size_t i = 0; i < image_embd_v.size(); i++) { + std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip)); + n_img_pos_out += clip_n_patches(ctx_clip); + } + *n_img_pos = n_img_pos_out; + for (size_t i = 0; i < image_embd_v.size(); i++) { + free(image_embd_v[i]); + } + image_embd_v.clear(); + load_image_size->width = img->nx; + load_image_size->height = img->ny; + clip_add_load_image_size(ctx_clip, load_image_size); + LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height); + } + else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { // flat / default llava-1.5 type embedding *n_img_pos = clip_n_patches(ctx_clip); bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096 @@ -254,7 +325,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli return false; } - } else { + } + else { // spatial_unpad llava-1.6 type embedding // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working std::vector image_embd_v; @@ -323,7 +395,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * } bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) { - float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model + int num_max_patches = 6; + if (clip_is_minicpmv(ctx_clip)) { + num_max_patches = 10; + } + float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model if (!image_embd) { LOG_TEE("Unable to allocate memory for image embeddings\n"); return false; diff --git a/llama/llava.h b/llama/llava.h index 8d1ed60f..859baf37 100644 --- a/llama/llava.h +++ b/llama/llava.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * @@ -43,12 +43,11 @@ # define LLAVA_API #endif -struct clip_ctx; - #ifdef __cplusplus extern "C" { #endif +struct clip_ctx; struct llava_image_embed { float * embed; int n_image_pos; @@ -63,8 +62,8 @@ LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length); /** build an image embed from a path to an image filename */ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path); -LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); /** free an embedding made with llava_image_embed_make_* */ +LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); diff --git a/llama/log.h b/llama/log.h index e600939c..cf552678 100644 --- a/llama/log.h +++ b/llama/log.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/patches/02-llamacpp.diff b/llama/patches/02-pretokenizer.diff similarity index 87% rename from llama/patches/02-llamacpp.diff rename to llama/patches/02-pretokenizer.diff index 0d40fc3c..351bcaef 100644 --- a/llama/patches/02-llamacpp.diff +++ b/llama/patches/02-pretokenizer.diff @@ -1,8 +1,8 @@ diff --git a/src/llama.cpp b/src/llama.cpp -index a207451f..2ddf431d 100644 +index 88355971..dd7d41ed 100644 --- a/src/llama.cpp +++ b/src/llama.cpp -@@ -5347,16 +5347,7 @@ static void llm_load_vocab( +@@ -6083,16 +6083,7 @@ static void llm_load_vocab( if (vocab.type == LLAMA_VOCAB_TYPE_BPE) { vocab.tokenizer_add_space_prefix = false; vocab.tokenizer_clean_spaces = true; @@ -20,9 +20,9 @@ index a207451f..2ddf431d 100644 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; } else if ( tokenizer_pre == "llama3" || -@@ -5443,7 +5434,8 @@ static void llm_load_vocab( - tokenizer_pre == "codeshell") { - vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL; +@@ -6188,7 +6179,8 @@ static void llm_load_vocab( + tokenizer_pre == "exaone") { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE; } else { - throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); + LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__); diff --git a/llama/patches/06-embeddings.diff b/llama/patches/05-embeddings.diff similarity index 61% rename from llama/patches/06-embeddings.diff rename to llama/patches/05-embeddings.diff index a84e3b06..f3c071cb 100644 --- a/llama/patches/06-embeddings.diff +++ b/llama/patches/05-embeddings.diff @@ -1,37 +1,36 @@ diff --git a/src/llama.cpp b/src/llama.cpp -index 1fe2b9f7..a43312a7 100644 +index 88355971..d7db689b 100644 --- a/src/llama.cpp +++ b/src/llama.cpp -@@ -13689,7 +13689,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { +@@ -15906,7 +15906,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) { const auto n_embd = hparams.n_embd; // TODO: use a per-batch flag for logits presence instead - const bool has_logits = !cparams.embeddings; + const bool has_logits = cparams.causal_attn; - const bool has_embd = lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE)); + const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0; -@@ -13959,17 +13959,25 @@ static int llama_decode_internal( +@@ -16175,20 +16175,23 @@ static int llama_decode_internal( // no output res = nullptr; embd = nullptr; - } else if (cparams.embeddings) { -- res = nullptr; // do not extract logits for embedding case -- embd = gf->nodes[gf->n_nodes - 1]; -- if (strcmp(embd->name, "result_embd_pooled") != 0) { -- embd = gf->nodes[gf->n_nodes - 2]; +- res = nullptr; // do not extract logits for embedding case +- embd = nullptr; + } + + if (cparams.embeddings) { -+ for (int i = gf->n_nodes - 1; i >= 0; --i) { + for (int i = gf->n_nodes - 1; i >= 0; --i) { +- if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) { +- embd = gf->nodes[i]; + embd = gf->nodes[i]; + if (strcmp(embd->name, "result_embd_pooled") == 0) { -+ break; -+ } + break; + } } - GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor"); -- } else { -+ } else { +- GGML_ASSERT(embd != nullptr && "missing embeddings tensor"); + } else { embd = nullptr; // do not extract embeddings when not needed GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor"); } @@ -39,7 +38,6 @@ index 1fe2b9f7..a43312a7 100644 + if (!cparams.causal_attn) { + res = nullptr; // do not extract logits when not needed + } -+ // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); ggml_backend_sched_alloc_graph(lctx.sched, gf); diff --git a/llama/runner/runner.go b/llama/runner/runner.go index c17e1ebc..fb4f06fd 100644 --- a/llama/runner/runner.go +++ b/llama/runner/runner.go @@ -675,7 +675,7 @@ func main() { } } - if server.model.ShouldAddBOSToken() { + if server.model.AddBOSToken() { server.bosToken = 1 } diff --git a/llama/sampling.cpp b/llama/sampling.cpp index 4c947815..604e99c9 100644 --- a/llama/sampling.cpp +++ b/llama/sampling.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/sampling.h b/llama/sampling.h index b1c82a5c..b5cb7982 100644 --- a/llama/sampling.h +++ b/llama/sampling.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/sgemm.cpp b/llama/sgemm.cpp index 6626ceb2..f0988ba7 100644 --- a/llama/sgemm.cpp +++ b/llama/sgemm.cpp @@ -606,17 +606,29 @@ class tinyBLAS_Q0_AVX { case 0x44: mc = 4; nc = 4; +#if defined(__AVX2__) && defined(__F16C__) + gemm4xN<4>(m0, m, n0, n); +#else gemm<4, 4>(m0, m, n0, n); +#endif break; case 0x43: mc = 4; nc = 3; +#if defined(__AVX2__) && defined(__F16C__) + gemm4xN<3>(m0, m, n0, n); +#else gemm<4, 3>(m0, m, n0, n); +#endif break; case 0x34: mc = 3; nc = 4; +#if defined(__AVX2__) && defined(__F16C__) + gemmMx4<3>(m0, m, n0, n); +#else gemm<3, 4>(m0, m, n0, n); +#endif break; case 0x33: mc = 3; @@ -626,12 +638,20 @@ class tinyBLAS_Q0_AVX { case 0x42: mc = 4; nc = 2; +#if defined(__AVX2__) && defined(__F16C__) + gemm4xN<2>(m0, m, n0, n); +#else gemm<4, 2>(m0, m, n0, n); +#endif break; case 0x24: mc = 2; nc = 4; +#if defined(__AVX2__) && defined(__F16C__) + gemmMx4<2>(m0, m, n0, n); +#else gemm<2, 4>(m0, m, n0, n); +#endif break; #else case 0x44: @@ -639,13 +659,21 @@ class tinyBLAS_Q0_AVX { case 0x42: mc = 4; nc = 2; +#if defined(__AVX2__) && defined(__F16C__) + gemm4xN<2>(m0, m, n0, n); +#else gemm<4, 2>(m0, m, n0, n); +#endif break; case 0x34: case 0x24: mc = 2; nc = 4; +#if defined(__AVX2__) && defined(__F16C__) + gemmMx4<2>(m0, m, n0, n); +#else gemm<2, 4>(m0, m, n0, n); +#endif break; case 0x33: #endif @@ -662,7 +690,11 @@ class tinyBLAS_Q0_AVX { case 0x41: mc = 4; nc = 1; +#if defined(__AVX2__) && defined(__F16C__) + gemm4xN<1>(m0, m, n0, n); +#else gemm<4, 1>(m0, m, n0, n); +#endif break; case 0x22: mc = 2; @@ -672,7 +704,11 @@ class tinyBLAS_Q0_AVX { case 0x14: mc = 1; nc = 4; +#if defined(__AVX2__) && defined(__F16C__) + gemmMx4<1>(m0, m, n0, n); +#else gemm<1, 4>(m0, m, n0, n); +#endif break; case 0x31: mc = 3; @@ -708,6 +744,119 @@ class tinyBLAS_Q0_AVX { mnpack(m0, m, np, n); } +#if defined(__AVX2__) && defined(__F16C__) +// Templated functions for gemm of dimensions 4xN + template + NOINLINE void gemm4xN(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / 4; + int64_t xtiles = (n - n0) / RN; + int64_t tiles = xtiles * ytiles; + int64_t duty = (tiles + nth - 1) / nth; + int64_t start = duty * ith; + int64_t end = start + duty; + if (end > tiles) + end = tiles; + for (int64_t job = start; job < end; ++job) { + int64_t ii = m0 + job / xtiles * 4; + int64_t jj = n0 + job % xtiles * RN; + __m256 Cv[RN][4] = {}; + for (int64_t l = 0; l < k; ++l) { + uint64_t a_delta = ((uint64_t)A[lda * (ii + 3) + l].d << 48) | ((uint64_t)A[lda * (ii + 2) + l].d << 32) | ((uint64_t)A[lda * (ii + 1) + l].d << 16) | (A[lda * (ii + 0) + l].d); + // Convert delta values for four blocks to float values + __m128 da = _mm_cvtph_ps(_mm_set_epi64x(0, a_delta)); + __m256i avec0 = load(A + lda * (ii + 0) + l); + __m256i avec1 = load(A + lda * (ii + 1) + l); + __m256i avec2 = load(A + lda * (ii + 2) + l); + __m256i avec3 = load(A + lda * (ii + 3) + l); + for (int64_t j = 0; j < RN; ++j) { + __m128 db = _mm_set1_ps(unhalf(B[ldb * (jj + j) + l].d)); + // Computation of product of delta values for four blocks and replicate it across 256 bit lane + __m256 dvec = _mm256_castps128_ps256(_mm_mul_ps(da, db)); + dvec = _mm256_permute2f128_ps(dvec ,dvec, 0); + // Computation of dot product and multiplication with appropriate delta value products + Cv[j][0] = madd(_mm256_shuffle_ps(dvec, dvec, 0), + updot(_mm256_sign_epi8(avec0, avec0), + _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec0)), + Cv[j][0]); + Cv[j][1] = madd(_mm256_shuffle_ps(dvec, dvec, 85), + updot(_mm256_sign_epi8(avec1, avec1), + _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec1)), + Cv[j][1]); + Cv[j][2] = madd(_mm256_shuffle_ps(dvec, dvec, 170), + updot(_mm256_sign_epi8(avec2, avec2), + _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec2)), + Cv[j][2]); + Cv[j][3] = madd(_mm256_shuffle_ps(dvec, dvec, 255), + updot(_mm256_sign_epi8(avec3, avec3), + _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec3)), + Cv[j][3]); + } + } + + for (int64_t j = 0; j < RN; ++j) + for (int64_t i = 0; i < 4; ++i) + C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]); + } + } + + // Templated functions for gemm of dimensions Mx4 + template + NOINLINE void gemmMx4(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / RM; + int64_t xtiles = (n - n0) / 4; + int64_t tiles = xtiles * ytiles; + int64_t duty = (tiles + nth - 1) / nth; + int64_t start = duty * ith; + int64_t end = start + duty; + if (end > tiles) + end = tiles; + for (int64_t job = start; job < end; ++job) { + int64_t ii = m0 + job / xtiles * RM; + int64_t jj = n0 + job % xtiles * 4; + __m256 Cv[4][RM] = {}; + for (int64_t l = 0; l < k; ++l) { + uint64_t b_delta = ((uint64_t)B[ldb * (jj + 3) + l].d << 48) | ((uint64_t)B[ldb * (jj + 2) + l].d << 32) | ((uint64_t)B[ldb * (jj + 1) + l].d << 16) | (B[ldb * (jj + 0) + l].d); + // Convert delta values for four blocks to float values + __m128 db = _mm_cvtph_ps(_mm_set_epi64x(0, b_delta)); + __m256i bvec0 = load(B + ldb * (jj + 0) + l); + __m256i bvec1 = load(B + ldb * (jj + 1) + l); + __m256i bvec2 = load(B + ldb * (jj + 2) + l); + __m256i bvec3 = load(B + ldb * (jj + 3) + l); + for (int64_t i = 0; i < RM; ++i) { + __m128 da = _mm_set1_ps(unhalf((A[lda * (ii + i) + l].d))); + // Computation of product of delta values for four blocks and replicate it across 256 bit lane + __m256 dvec = _mm256_castps128_ps256(_mm_mul_ps(da, db)); + dvec = _mm256_permute2f128_ps(dvec ,dvec, 0); + // Computation of dot product and multiplication with appropriate delta value products + Cv[0][i] = madd(_mm256_shuffle_ps(dvec, dvec, 0), + updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l), + load(A + lda * (ii + i) + l)), + _mm256_sign_epi8(bvec0, load(A + lda * (ii + i) + l))), + Cv[0][i]); + Cv[1][i] = madd(_mm256_shuffle_ps(dvec, dvec, 85), + updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l), + load(A + lda * (ii + i) + l)), + _mm256_sign_epi8(bvec1, load(A + lda * (ii + i) + l))), + Cv[1][i]); + Cv[2][i] = madd(_mm256_shuffle_ps(dvec, dvec, 170), + updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l), + load(A + lda * (ii + i) + l)), + _mm256_sign_epi8(bvec2, load(A + lda * (ii + i) + l))), + Cv[2][i]); + Cv[3][i] = madd(_mm256_shuffle_ps(dvec, dvec, 255), + updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l), + load(A + lda * (ii + i) + l)), + _mm256_sign_epi8(bvec3, load(A + lda * (ii + i) + l))), + Cv[3][i]); + } + } + for (int64_t j = 0; j < 4; ++j) + for (int64_t i = 0; i < RM; ++i) + C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]); + } + } +#endif + template NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) { int64_t ytiles = (m - m0) / RM; diff --git a/llama/stb_image.h b/llama/stb_image.h index 4766d7e6..9eedabed 100644 --- a/llama/stb_image.h +++ b/llama/stb_image.h @@ -1,4 +1,4 @@ -/* stb_image - v2.28 - public domain image loader - http://nothings.org/stb +/* stb_image - v2.30 - public domain image loader - http://nothings.org/stb no warranty implied; use at your own risk Do this: @@ -48,6 +48,8 @@ LICENSE RECENT REVISION HISTORY: + 2.30 (2024-05-31) avoid erroneous gcc warning + 2.29 (2023-05-xx) optimizations 2.28 (2023-01-29) many error fixes, security errors, just tons of stuff 2.27 (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes 2.26 (2020-07-13) many minor fixes @@ -371,13 +373,14 @@ RECENT REVISION HISTORY: #define STBI_VERSION 1 -enum { - STBI_default = 0, // only used for desired_channels +enum +{ + STBI_default = 0, // only used for desired_channels - STBI_grey = 1, - STBI_grey_alpha = 2, - STBI_rgb = 3, - STBI_rgb_alpha = 4 + STBI_grey = 1, + STBI_grey_alpha = 2, + STBI_rgb = 3, + STBI_rgb_alpha = 4 }; #include @@ -405,11 +408,11 @@ extern "C" { // load image by filename, open file, or memory buffer // -typedef struct { - int (*read)(void * user, char * data, - int size); // fill 'data' with 'size' bytes. return number of bytes actually read - void (*skip)(void * user, int n); // skip the next 'n' bytes, or 'unget' the last -n bytes if negative - int (*eof)(void * user); // returns nonzero if we are at end of file/data +typedef struct +{ + int (*read) (void *user,char *data,int size); // fill 'data' with 'size' bytes. return number of bytes actually read + void (*skip) (void *user,int n); // skip the next 'n' bytes, or 'unget' the last -n bytes if negative + int (*eof) (void *user); // returns nonzero if we are at end of file/data } stbi_io_callbacks; //////////////////////////////////// @@ -417,24 +420,21 @@ typedef struct { // 8-bits-per-channel interface // -STBIDEF stbi_uc * stbi_load_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels); -STBIDEF stbi_uc * stbi_load_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, - int * channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_memory (stbi_uc const *buffer, int len , int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk , void *user, int *x, int *y, int *channels_in_file, int desired_channels); #ifndef STBI_NO_STDIO -STBIDEF stbi_uc * stbi_load(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels); -STBIDEF stbi_uc * stbi_load_from_file(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_uc *stbi_load_from_file (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); // for stbi_load_from_file, file pointer is left pointing immediately after image #endif #ifndef STBI_NO_GIF -STBIDEF stbi_uc * stbi_load_gif_from_memory(stbi_uc const * buffer, int len, int ** delays, int * x, int * y, int * z, - int * comp, int req_comp); +STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp); #endif #ifdef STBI_WINDOWS_UTF8 -STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wchar_t * input); +STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input); #endif //////////////////////////////////// @@ -442,14 +442,12 @@ STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wc // 16-bits-per-channel interface // -STBIDEF stbi_us * stbi_load_16_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels); -STBIDEF stbi_us * stbi_load_16_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, - int * channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_16_from_memory (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels); #ifndef STBI_NO_STDIO -STBIDEF stbi_us * stbi_load_16(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels); -STBIDEF stbi_us * stbi_load_from_file_16(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_16 (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); +STBIDEF stbi_us *stbi_load_from_file_16(FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); #endif //////////////////////////////////// @@ -457,55 +455,56 @@ STBIDEF stbi_us * stbi_load_from_file_16(FILE * f, int * x, int * y, int * chann // float-per-channel interface // #ifndef STBI_NO_LINEAR -STBIDEF float * stbi_loadf_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels); -STBIDEF float * stbi_loadf_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * channels_in_file, - int desired_channels); + STBIDEF float *stbi_loadf_from_memory (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels); + STBIDEF float *stbi_loadf_from_callbacks (stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels); -#ifndef STBI_NO_STDIO -STBIDEF float * stbi_loadf(char const * filename, int * x, int * y, int * channels_in_file, int desired_channels); -STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * channels_in_file, int desired_channels); -#endif + #ifndef STBI_NO_STDIO + STBIDEF float *stbi_loadf (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels); + STBIDEF float *stbi_loadf_from_file (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels); + #endif #endif #ifndef STBI_NO_HDR -STBIDEF void stbi_hdr_to_ldr_gamma(float gamma); -STBIDEF void stbi_hdr_to_ldr_scale(float scale); + STBIDEF void stbi_hdr_to_ldr_gamma(float gamma); + STBIDEF void stbi_hdr_to_ldr_scale(float scale); #endif // STBI_NO_HDR #ifndef STBI_NO_LINEAR -STBIDEF void stbi_ldr_to_hdr_gamma(float gamma); -STBIDEF void stbi_ldr_to_hdr_scale(float scale); + STBIDEF void stbi_ldr_to_hdr_gamma(float gamma); + STBIDEF void stbi_ldr_to_hdr_scale(float scale); #endif // STBI_NO_LINEAR // stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR -STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const * clbk, void * user); -STBIDEF int stbi_is_hdr_from_memory(stbi_uc const * buffer, int len); +STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user); +STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len); #ifndef STBI_NO_STDIO -STBIDEF int stbi_is_hdr(char const * filename); -STBIDEF int stbi_is_hdr_from_file(FILE * f); +STBIDEF int stbi_is_hdr (char const *filename); +STBIDEF int stbi_is_hdr_from_file(FILE *f); #endif // STBI_NO_STDIO + // get a VERY brief reason for failure // on most compilers (and ALL modern mainstream compilers) this is threadsafe -STBIDEF const char * stbi_failure_reason(void); +STBIDEF const char *stbi_failure_reason (void); // free the loaded image -- this is just free() -STBIDEF void stbi_image_free(void * retval_from_stbi_load); +STBIDEF void stbi_image_free (void *retval_from_stbi_load); // get image dimensions & components without fully decoding -STBIDEF int stbi_info_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp); -STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp); -STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const * buffer, int len); -STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * clbk, void * user); +STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp); +STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp); +STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len); +STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *clbk, void *user); #ifndef STBI_NO_STDIO -STBIDEF int stbi_info(char const * filename, int * x, int * y, int * comp); -STBIDEF int stbi_info_from_file(FILE * f, int * x, int * y, int * comp); -STBIDEF int stbi_is_16_bit(char const * filename); -STBIDEF int stbi_is_16_bit_from_file(FILE * f); +STBIDEF int stbi_info (char const *filename, int *x, int *y, int *comp); +STBIDEF int stbi_info_from_file (FILE *f, int *x, int *y, int *comp); +STBIDEF int stbi_is_16_bit (char const *filename); +STBIDEF int stbi_is_16_bit_from_file(FILE *f); #endif + + // for image formats that explicitly notate that they have premultiplied alpha, // we just return the colors as stored in the file. set this flag to force // unpremultiplication. results are undefined if the unpremultiply overflow. @@ -527,14 +526,14 @@ STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_fli // ZLIB client - used by PNG, available for other purposes -STBIDEF char * stbi_zlib_decode_malloc_guesssize(const char * buffer, int len, int initial_size, int * outlen); -STBIDEF char * stbi_zlib_decode_malloc_guesssize_headerflag(const char * buffer, int len, int initial_size, int * outlen, - int parse_header); -STBIDEF char * stbi_zlib_decode_malloc(const char * buffer, int len, int * outlen); -STBIDEF int stbi_zlib_decode_buffer(char * obuffer, int olen, const char * ibuffer, int ilen); +STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen); +STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header); +STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen); +STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen); + +STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen); +STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen); -STBIDEF char * stbi_zlib_decode_noheader_malloc(const char * buffer, int len, int * outlen); -STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const char * ibuffer, int ilen); #ifdef __cplusplus } @@ -547,50 +546,52 @@ STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const cha #ifdef STB_IMAGE_IMPLEMENTATION -#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) || defined(STBI_ONLY_TGA) || \ - defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || \ - defined(STBI_ONLY_PNM) || defined(STBI_ONLY_ZLIB) -#ifndef STBI_ONLY_JPEG -#define STBI_NO_JPEG -#endif -#ifndef STBI_ONLY_PNG -#define STBI_NO_PNG -#endif -#ifndef STBI_ONLY_BMP -#define STBI_NO_BMP -#endif -#ifndef STBI_ONLY_PSD -#define STBI_NO_PSD -#endif -#ifndef STBI_ONLY_TGA -#define STBI_NO_TGA -#endif -#ifndef STBI_ONLY_GIF -#define STBI_NO_GIF -#endif -#ifndef STBI_ONLY_HDR -#define STBI_NO_HDR -#endif -#ifndef STBI_ONLY_PIC -#define STBI_NO_PIC -#endif -#ifndef STBI_ONLY_PNM -#define STBI_NO_PNM -#endif +#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \ + || defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \ + || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \ + || defined(STBI_ONLY_ZLIB) + #ifndef STBI_ONLY_JPEG + #define STBI_NO_JPEG + #endif + #ifndef STBI_ONLY_PNG + #define STBI_NO_PNG + #endif + #ifndef STBI_ONLY_BMP + #define STBI_NO_BMP + #endif + #ifndef STBI_ONLY_PSD + #define STBI_NO_PSD + #endif + #ifndef STBI_ONLY_TGA + #define STBI_NO_TGA + #endif + #ifndef STBI_ONLY_GIF + #define STBI_NO_GIF + #endif + #ifndef STBI_ONLY_HDR + #define STBI_NO_HDR + #endif + #ifndef STBI_ONLY_PIC + #define STBI_NO_PIC + #endif + #ifndef STBI_ONLY_PNM + #define STBI_NO_PNM + #endif #endif #if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB) #define STBI_NO_ZLIB #endif -#include + #include #include // ptrdiff_t on osx #include #include +#include #if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) -#include // ldexp, pow +#include // ldexp, pow #endif #ifndef STBI_NO_STDIO @@ -608,54 +609,55 @@ STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const cha #define STBI_EXTERN extern #endif + #ifndef _MSC_VER -#ifdef __cplusplus -#define stbi_inline inline + #ifdef __cplusplus + #define stbi_inline inline + #else + #define stbi_inline + #endif #else -#define stbi_inline -#endif -#else -#define stbi_inline __forceinline + #define stbi_inline __forceinline #endif #ifndef STBI_NO_THREAD_LOCALS -#if defined(__cplusplus) && __cplusplus >= 201103L -#define STBI_THREAD_LOCAL thread_local -#elif defined(__GNUC__) && __GNUC__ < 5 -#define STBI_THREAD_LOCAL __thread -#elif defined(_MSC_VER) -#define STBI_THREAD_LOCAL __declspec(thread) -#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__) -#define STBI_THREAD_LOCAL _Thread_local -#endif + #if defined(__cplusplus) && __cplusplus >= 201103L + #define STBI_THREAD_LOCAL thread_local + #elif defined(__GNUC__) && __GNUC__ < 5 + #define STBI_THREAD_LOCAL __thread + #elif defined(_MSC_VER) + #define STBI_THREAD_LOCAL __declspec(thread) + #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__) + #define STBI_THREAD_LOCAL _Thread_local + #endif -#ifndef STBI_THREAD_LOCAL -#if defined(__GNUC__) -#define STBI_THREAD_LOCAL __thread -#endif -#endif + #ifndef STBI_THREAD_LOCAL + #if defined(__GNUC__) + #define STBI_THREAD_LOCAL __thread + #endif + #endif #endif #if defined(_MSC_VER) || defined(__SYMBIAN32__) typedef unsigned short stbi__uint16; -typedef signed short stbi__int16; -typedef unsigned int stbi__uint32; -typedef signed int stbi__int32; +typedef signed short stbi__int16; +typedef unsigned int stbi__uint32; +typedef signed int stbi__int32; #else #include typedef uint16_t stbi__uint16; -typedef int16_t stbi__int16; +typedef int16_t stbi__int16; typedef uint32_t stbi__uint32; -typedef int32_t stbi__int32; +typedef int32_t stbi__int32; #endif // should produce compiler error if size is wrong -typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; +typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1]; #ifdef _MSC_VER -#define STBI_NOTUSED(v) (void)(v) +#define STBI_NOTUSED(v) (void)(v) #else -#define STBI_NOTUSED(v) (void)sizeof(v) +#define STBI_NOTUSED(v) (void)sizeof(v) #endif #ifdef _MSC_VER @@ -663,9 +665,9 @@ typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; #endif #ifdef STBI_HAS_LROTL -#define stbi_lrot(x, y) _lrotl(x, y) + #define stbi_lrot(x,y) _lrotl(x,y) #else -#define stbi_lrot(x, y) (((x) << (y)) | ((x) >> (-(y)&31))) + #define stbi_lrot(x,y) (((x) << (y)) | ((x) >> (-(y) & 31))) #endif #if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED)) @@ -677,13 +679,13 @@ typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; #endif #ifndef STBI_MALLOC -#define STBI_MALLOC(sz) malloc(sz) -#define STBI_REALLOC(p, newsz) realloc(p, newsz) -#define STBI_FREE(p) free(p) +#define STBI_MALLOC(sz) malloc(sz) +#define STBI_REALLOC(p,newsz) realloc(p,newsz) +#define STBI_FREE(p) free(p) #endif #ifndef STBI_REALLOC_SIZED -#define STBI_REALLOC_SIZED(p, oldsz, newsz) STBI_REALLOC(p, newsz) +#define STBI_REALLOC_SIZED(p,oldsz,newsz) STBI_REALLOC(p,newsz) #endif // x86/x64 detection @@ -725,31 +727,34 @@ typedef unsigned char validate_uint32[sizeof(stbi__uint32) == 4 ? 1 : -1]; #ifdef _MSC_VER -#if _MSC_VER >= 1400 // not VC6 -#include // __cpuid -static int stbi__cpuid3(void) { - int info[4]; - __cpuid(info, 1); - return info[3]; +#if _MSC_VER >= 1400 // not VC6 +#include // __cpuid +static int stbi__cpuid3(void) +{ + int info[4]; + __cpuid(info,1); + return info[3]; } #else -static int stbi__cpuid3(void) { - int res; - __asm { +static int stbi__cpuid3(void) +{ + int res; + __asm { mov eax,1 cpuid mov res,edx - } - return res; + } + return res; } #endif #define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name #if !defined(STBI_NO_JPEG) && defined(STBI_SSE2) -static int stbi__sse2_available(void) { - int info3 = stbi__cpuid3(); - return ((info3 >> 26) & 1) != 0; +static int stbi__sse2_available(void) +{ + int info3 = stbi__cpuid3(); + return ((info3 >> 26) & 1) != 0; } #endif @@ -757,11 +762,12 @@ static int stbi__sse2_available(void) { #define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16))) #if !defined(STBI_NO_JPEG) && defined(STBI_SSE2) -static int stbi__sse2_available(void) { - // If we're even attempting to compile this on GCC/Clang, that means - // -msse2 is on, which means the compiler is allowed to use SSE2 - // instructions at will, and so are we. - return 1; +static int stbi__sse2_available(void) +{ + // If we're even attempting to compile this on GCC/Clang, that means + // -msse2 is on, which means the compiler is allowed to use SSE2 + // instructions at will, and so are we. + return 1; } #endif @@ -796,162 +802,190 @@ static int stbi__sse2_available(void) { // stbi__context structure is our basic context used by all images, so it // contains all the IO context, plus some basic image information -typedef struct { - stbi__uint32 img_x, img_y; - int img_n, img_out_n; +typedef struct +{ + stbi__uint32 img_x, img_y; + int img_n, img_out_n; - stbi_io_callbacks io; - void * io_user_data; + stbi_io_callbacks io; + void *io_user_data; - int read_from_callbacks; - int buflen; - stbi_uc buffer_start[128]; - int callback_already_read; + int read_from_callbacks; + int buflen; + stbi_uc buffer_start[128]; + int callback_already_read; - stbi_uc *img_buffer, *img_buffer_end; - stbi_uc *img_buffer_original, *img_buffer_original_end; + stbi_uc *img_buffer, *img_buffer_end; + stbi_uc *img_buffer_original, *img_buffer_original_end; } stbi__context; -static void stbi__refill_buffer(stbi__context * s); + +static void stbi__refill_buffer(stbi__context *s); // initialize a memory-decode context -static void stbi__start_mem(stbi__context * s, stbi_uc const * buffer, int len) { - s->io.read = NULL; - s->read_from_callbacks = 0; - s->callback_already_read = 0; - s->img_buffer = s->img_buffer_original = (stbi_uc *)buffer; - s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *)buffer + len; +static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len) +{ + s->io.read = NULL; + s->read_from_callbacks = 0; + s->callback_already_read = 0; + s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer; + s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len; } // initialize a callback-based context -static void stbi__start_callbacks(stbi__context * s, stbi_io_callbacks * c, void * user) { - s->io = *c; - s->io_user_data = user; - s->buflen = sizeof(s->buffer_start); - s->read_from_callbacks = 1; - s->callback_already_read = 0; - s->img_buffer = s->img_buffer_original = s->buffer_start; - stbi__refill_buffer(s); - s->img_buffer_original_end = s->img_buffer_end; +static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user) +{ + s->io = *c; + s->io_user_data = user; + s->buflen = sizeof(s->buffer_start); + s->read_from_callbacks = 1; + s->callback_already_read = 0; + s->img_buffer = s->img_buffer_original = s->buffer_start; + stbi__refill_buffer(s); + s->img_buffer_original_end = s->img_buffer_end; } #ifndef STBI_NO_STDIO -static int stbi__stdio_read(void * user, char * data, int size) { return (int)fread(data, 1, size, (FILE *)user); } - -static void stbi__stdio_skip(void * user, int n) { - int ch; - fseek((FILE *)user, n, SEEK_CUR); - ch = fgetc((FILE *)user); /* have to read a byte to reset feof()'s flag */ - if (ch != EOF) { - ungetc(ch, (FILE *)user); /* push byte back onto stream if valid. */ - } +static int stbi__stdio_read(void *user, char *data, int size) +{ + return (int) fread(data,1,size,(FILE*) user); } -static int stbi__stdio_eof(void * user) { return feof((FILE *)user) || ferror((FILE *)user); } +static void stbi__stdio_skip(void *user, int n) +{ + int ch; + fseek((FILE*) user, n, SEEK_CUR); + ch = fgetc((FILE*) user); /* have to read a byte to reset feof()'s flag */ + if (ch != EOF) { + ungetc(ch, (FILE *) user); /* push byte back onto stream if valid. */ + } +} -static stbi_io_callbacks stbi__stdio_callbacks = { - stbi__stdio_read, - stbi__stdio_skip, - stbi__stdio_eof, +static int stbi__stdio_eof(void *user) +{ + return feof((FILE*) user) || ferror((FILE *) user); +} + +static stbi_io_callbacks stbi__stdio_callbacks = +{ + stbi__stdio_read, + stbi__stdio_skip, + stbi__stdio_eof, }; -static void stbi__start_file(stbi__context * s, FILE * f) { stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *)f); } +static void stbi__start_file(stbi__context *s, FILE *f) +{ + stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f); +} -// static void stop_file(stbi__context *s) { } +//static void stop_file(stbi__context *s) { } #endif // !STBI_NO_STDIO -static void stbi__rewind(stbi__context * s) { - // conceptually rewind SHOULD rewind to the beginning of the stream, - // but we just rewind to the beginning of the initial buffer, because - // we only use it after doing 'test', which only ever looks at at most 92 bytes - s->img_buffer = s->img_buffer_original; - s->img_buffer_end = s->img_buffer_original_end; +static void stbi__rewind(stbi__context *s) +{ + // conceptually rewind SHOULD rewind to the beginning of the stream, + // but we just rewind to the beginning of the initial buffer, because + // we only use it after doing 'test', which only ever looks at at most 92 bytes + s->img_buffer = s->img_buffer_original; + s->img_buffer_end = s->img_buffer_original_end; } -enum { STBI_ORDER_RGB, STBI_ORDER_BGR }; +enum +{ + STBI_ORDER_RGB, + STBI_ORDER_BGR +}; -typedef struct { - int bits_per_channel; - int num_channels; - int channel_order; +typedef struct +{ + int bits_per_channel; + int num_channels; + int channel_order; } stbi__result_info; #ifndef STBI_NO_JPEG -static int stbi__jpeg_test(stbi__context * s); -static void * stbi__jpeg_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__jpeg_test(stbi__context *s); +static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PNG -static int stbi__png_test(stbi__context * s); -static void * stbi__png_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__png_info(stbi__context * s, int * x, int * y, int * comp); -static int stbi__png_is16(stbi__context * s); +static int stbi__png_test(stbi__context *s); +static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp); +static int stbi__png_is16(stbi__context *s); #endif #ifndef STBI_NO_BMP -static int stbi__bmp_test(stbi__context * s); -static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__bmp_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__bmp_test(stbi__context *s); +static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_TGA -static int stbi__tga_test(stbi__context * s); -static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__tga_test(stbi__context *s); +static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PSD -static int stbi__psd_test(stbi__context * s); -static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc); -static int stbi__psd_info(stbi__context * s, int * x, int * y, int * comp); -static int stbi__psd_is16(stbi__context * s); +static int stbi__psd_test(stbi__context *s); +static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc); +static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp); +static int stbi__psd_is16(stbi__context *s); #endif #ifndef STBI_NO_HDR -static int stbi__hdr_test(stbi__context * s); -static float * stbi__hdr_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__hdr_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__hdr_test(stbi__context *s); +static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PIC -static int stbi__pic_test(stbi__context * s); -static void * stbi__pic_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__pic_test(stbi__context *s); +static void *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_GIF -static int stbi__gif_test(stbi__context * s); -static void * stbi__gif_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static void * stbi__load_gif_main(stbi__context * s, int ** delays, int * x, int * y, int * z, int * comp, int req_comp); -static int stbi__gif_info(stbi__context * s, int * x, int * y, int * comp); +static int stbi__gif_test(stbi__context *s); +static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp); +static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp); #endif #ifndef STBI_NO_PNM -static int stbi__pnm_test(stbi__context * s); -static void * stbi__pnm_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri); -static int stbi__pnm_info(stbi__context * s, int * x, int * y, int * comp); -static int stbi__pnm_is16(stbi__context * s); +static int stbi__pnm_test(stbi__context *s); +static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri); +static int stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp); +static int stbi__pnm_is16(stbi__context *s); #endif static #ifdef STBI_THREAD_LOCAL - STBI_THREAD_LOCAL +STBI_THREAD_LOCAL #endif - const char * stbi__g_failure_reason; +const char *stbi__g_failure_reason; -STBIDEF const char * stbi_failure_reason(void) { return stbi__g_failure_reason; } +STBIDEF const char *stbi_failure_reason(void) +{ + return stbi__g_failure_reason; +} #ifndef STBI_NO_FAILURE_STRINGS -static int stbi__err(const char * str) { - stbi__g_failure_reason = str; - return 0; +static int stbi__err(const char *str) +{ + stbi__g_failure_reason = str; + return 0; } #endif -static void * stbi__malloc(size_t size) { return STBI_MALLOC(size); } +static void *stbi__malloc(size_t size) +{ + return STBI_MALLOC(size); +} // stb_image uses ints pervasively, including for offset calculations. // therefore the largest decoded image size we can support with the @@ -965,88 +999,88 @@ static void * stbi__malloc(size_t size) { return STBI_MALLOC(size); } // return 1 if the sum is valid, 0 on overflow. // negative terms are considered invalid. -static int stbi__addsizes_valid(int a, int b) { - if (b < 0) - return 0; - // now 0 <= b <= INT_MAX, hence also - // 0 <= INT_MAX - b <= INTMAX. - // And "a + b <= INT_MAX" (which might overflow) is the - // same as a <= INT_MAX - b (no overflow) - return a <= INT_MAX - b; +static int stbi__addsizes_valid(int a, int b) +{ + if (b < 0) return 0; + // now 0 <= b <= INT_MAX, hence also + // 0 <= INT_MAX - b <= INTMAX. + // And "a + b <= INT_MAX" (which might overflow) is the + // same as a <= INT_MAX - b (no overflow) + return a <= INT_MAX - b; } // returns 1 if the product is valid, 0 on overflow. // negative factors are considered invalid. -static int stbi__mul2sizes_valid(int a, int b) { - if (a < 0 || b < 0) - return 0; - if (b == 0) - return 1; // mul-by-0 is always safe - // portable way to check for no overflows in a*b - return a <= INT_MAX / b; +static int stbi__mul2sizes_valid(int a, int b) +{ + if (a < 0 || b < 0) return 0; + if (b == 0) return 1; // mul-by-0 is always safe + // portable way to check for no overflows in a*b + return a <= INT_MAX/b; } #if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR) // returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow -static int stbi__mad2sizes_valid(int a, int b, int add) { - return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a * b, add); +static int stbi__mad2sizes_valid(int a, int b, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a*b, add); } #endif // returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow -static int stbi__mad3sizes_valid(int a, int b, int c, int add) { - return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a * b, c) && stbi__addsizes_valid(a * b * c, add); +static int stbi__mad3sizes_valid(int a, int b, int c, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) && + stbi__addsizes_valid(a*b*c, add); } // returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow #if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM) -static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add) { - return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a * b, c) && stbi__mul2sizes_valid(a * b * c, d) && - stbi__addsizes_valid(a * b * c * d, add); +static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add) +{ + return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) && + stbi__mul2sizes_valid(a*b*c, d) && stbi__addsizes_valid(a*b*c*d, add); } #endif #if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR) // mallocs with size overflow checking -static void * stbi__malloc_mad2(int a, int b, int add) { - if (!stbi__mad2sizes_valid(a, b, add)) - return NULL; - return stbi__malloc(a * b + add); +static void *stbi__malloc_mad2(int a, int b, int add) +{ + if (!stbi__mad2sizes_valid(a, b, add)) return NULL; + return stbi__malloc(a*b + add); } #endif -static void * stbi__malloc_mad3(int a, int b, int c, int add) { - if (!stbi__mad3sizes_valid(a, b, c, add)) - return NULL; - return stbi__malloc(a * b * c + add); +static void *stbi__malloc_mad3(int a, int b, int c, int add) +{ + if (!stbi__mad3sizes_valid(a, b, c, add)) return NULL; + return stbi__malloc(a*b*c + add); } #if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM) -static void * stbi__malloc_mad4(int a, int b, int c, int d, int add) { - if (!stbi__mad4sizes_valid(a, b, c, d, add)) - return NULL; - return stbi__malloc(a * b * c * d + add); +static void *stbi__malloc_mad4(int a, int b, int c, int d, int add) +{ + if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL; + return stbi__malloc(a*b*c*d + add); } #endif // returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow. -static int stbi__addints_valid(int a, int b) { - if ((a >= 0) != (b >= 0)) - return 1; // a and b have different signs, so no overflow - if (a < 0 && b < 0) - return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0. - return a <= INT_MAX - b; +static int stbi__addints_valid(int a, int b) +{ + if ((a >= 0) != (b >= 0)) return 1; // a and b have different signs, so no overflow + if (a < 0 && b < 0) return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0. + return a <= INT_MAX - b; } -// returns 1 if the product of two signed shorts is valid, 0 on overflow. -static int stbi__mul2shorts_valid(short a, short b) { - if (b == 0 || b == -1) - return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow - if ((a >= 0) == (b >= 0)) - return a <= SHRT_MAX / b; // product is positive, so similar to mul2sizes_valid - if (b < 0) - return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN - return a >= SHRT_MIN / b; +// returns 1 if the product of two ints fits in a signed short, 0 on overflow. +static int stbi__mul2shorts_valid(int a, int b) +{ + if (b == 0 || b == -1) return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow + if ((a >= 0) == (b >= 0)) return a <= SHRT_MAX/b; // product is positive, so similar to mul2sizes_valid + if (b < 0) return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN + return a >= SHRT_MIN / b; } // stbi__err - error @@ -1054,411 +1088,423 @@ static int stbi__mul2shorts_valid(short a, short b) { // stbi__errpuc - error returning pointer to unsigned char #ifdef STBI_NO_FAILURE_STRINGS -#define stbi__err(x, y) 0 + #define stbi__err(x,y) 0 #elif defined(STBI_FAILURE_USERMSG) -#define stbi__err(x, y) stbi__err(y) + #define stbi__err(x,y) stbi__err(y) #else -#define stbi__err(x, y) stbi__err(x) + #define stbi__err(x,y) stbi__err(x) #endif -#define stbi__errpf(x, y) ((float *)(size_t)(stbi__err(x, y) ? NULL : NULL)) -#define stbi__errpuc(x, y) ((unsigned char *)(size_t)(stbi__err(x, y) ? NULL : NULL)) +#define stbi__errpf(x,y) ((float *)(size_t) (stbi__err(x,y)?NULL:NULL)) +#define stbi__errpuc(x,y) ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL)) -STBIDEF void stbi_image_free(void * retval_from_stbi_load) { STBI_FREE(retval_from_stbi_load); } +STBIDEF void stbi_image_free(void *retval_from_stbi_load) +{ + STBI_FREE(retval_from_stbi_load); +} #ifndef STBI_NO_LINEAR -static float * stbi__ldr_to_hdr(stbi_uc * data, int x, int y, int comp); +static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp); #endif #ifndef STBI_NO_HDR -static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp); +static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp); #endif static int stbi__vertically_flip_on_load_global = 0; -STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip) { - stbi__vertically_flip_on_load_global = flag_true_if_should_flip; +STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip) +{ + stbi__vertically_flip_on_load_global = flag_true_if_should_flip; } #ifndef STBI_THREAD_LOCAL -#define stbi__vertically_flip_on_load stbi__vertically_flip_on_load_global +#define stbi__vertically_flip_on_load stbi__vertically_flip_on_load_global #else static STBI_THREAD_LOCAL int stbi__vertically_flip_on_load_local, stbi__vertically_flip_on_load_set; -STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip) { - stbi__vertically_flip_on_load_local = flag_true_if_should_flip; - stbi__vertically_flip_on_load_set = 1; +STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip) +{ + stbi__vertically_flip_on_load_local = flag_true_if_should_flip; + stbi__vertically_flip_on_load_set = 1; } -#define stbi__vertically_flip_on_load \ - (stbi__vertically_flip_on_load_set ? stbi__vertically_flip_on_load_local : stbi__vertically_flip_on_load_global) +#define stbi__vertically_flip_on_load (stbi__vertically_flip_on_load_set \ + ? stbi__vertically_flip_on_load_local \ + : stbi__vertically_flip_on_load_global) #endif // STBI_THREAD_LOCAL -static void * stbi__load_main(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc) { - memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields - ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed - ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order - ri->num_channels = 0; +static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc) +{ + memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields + ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed + ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order + ri->num_channels = 0; -// test the formats with a very explicit header first (at least a FOURCC -// or distinctive magic number first) -#ifndef STBI_NO_PNG - if (stbi__png_test(s)) - return stbi__png_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_BMP - if (stbi__bmp_test(s)) - return stbi__bmp_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_GIF - if (stbi__gif_test(s)) - return stbi__gif_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_PSD - if (stbi__psd_test(s)) - return stbi__psd_load(s, x, y, comp, req_comp, ri, bpc); -#else - STBI_NOTUSED(bpc); -#endif -#ifndef STBI_NO_PIC - if (stbi__pic_test(s)) - return stbi__pic_load(s, x, y, comp, req_comp, ri); -#endif + // test the formats with a very explicit header first (at least a FOURCC + // or distinctive magic number first) + #ifndef STBI_NO_PNG + if (stbi__png_test(s)) return stbi__png_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_BMP + if (stbi__bmp_test(s)) return stbi__bmp_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_GIF + if (stbi__gif_test(s)) return stbi__gif_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_PSD + if (stbi__psd_test(s)) return stbi__psd_load(s,x,y,comp,req_comp, ri, bpc); + #else + STBI_NOTUSED(bpc); + #endif + #ifndef STBI_NO_PIC + if (stbi__pic_test(s)) return stbi__pic_load(s,x,y,comp,req_comp, ri); + #endif -// then the formats that can end up attempting to load with just 1 or 2 -// bytes matching expectations; these are prone to false positives, so -// try them later -#ifndef STBI_NO_JPEG - if (stbi__jpeg_test(s)) - return stbi__jpeg_load(s, x, y, comp, req_comp, ri); -#endif -#ifndef STBI_NO_PNM - if (stbi__pnm_test(s)) - return stbi__pnm_load(s, x, y, comp, req_comp, ri); -#endif + // then the formats that can end up attempting to load with just 1 or 2 + // bytes matching expectations; these are prone to false positives, so + // try them later + #ifndef STBI_NO_JPEG + if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri); + #endif + #ifndef STBI_NO_PNM + if (stbi__pnm_test(s)) return stbi__pnm_load(s,x,y,comp,req_comp, ri); + #endif -#ifndef STBI_NO_HDR - if (stbi__hdr_test(s)) { - float * hdr = stbi__hdr_load(s, x, y, comp, req_comp, ri); - return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp); - } -#endif + #ifndef STBI_NO_HDR + if (stbi__hdr_test(s)) { + float *hdr = stbi__hdr_load(s, x,y,comp,req_comp, ri); + return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp); + } + #endif -#ifndef STBI_NO_TGA - // test tga last because it's a crappy test! - if (stbi__tga_test(s)) - return stbi__tga_load(s, x, y, comp, req_comp, ri); -#endif + #ifndef STBI_NO_TGA + // test tga last because it's a crappy test! + if (stbi__tga_test(s)) + return stbi__tga_load(s,x,y,comp,req_comp, ri); + #endif - return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt"); + return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt"); } -static stbi_uc * stbi__convert_16_to_8(stbi__uint16 * orig, int w, int h, int channels) { - int i; - int img_len = w * h * channels; - stbi_uc * reduced; +static stbi_uc *stbi__convert_16_to_8(stbi__uint16 *orig, int w, int h, int channels) +{ + int i; + int img_len = w * h * channels; + stbi_uc *reduced; - reduced = (stbi_uc *)stbi__malloc(img_len); - if (reduced == NULL) - return stbi__errpuc("outofmem", "Out of memory"); + reduced = (stbi_uc *) stbi__malloc(img_len); + if (reduced == NULL) return stbi__errpuc("outofmem", "Out of memory"); - for (i = 0; i < img_len; ++i) - reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling + for (i = 0; i < img_len; ++i) + reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling - STBI_FREE(orig); - return reduced; + STBI_FREE(orig); + return reduced; } -static stbi__uint16 * stbi__convert_8_to_16(stbi_uc * orig, int w, int h, int channels) { - int i; - int img_len = w * h * channels; - stbi__uint16 * enlarged; +static stbi__uint16 *stbi__convert_8_to_16(stbi_uc *orig, int w, int h, int channels) +{ + int i; + int img_len = w * h * channels; + stbi__uint16 *enlarged; - enlarged = (stbi__uint16 *)stbi__malloc(img_len * 2); - if (enlarged == NULL) - return (stbi__uint16 *)stbi__errpuc("outofmem", "Out of memory"); + enlarged = (stbi__uint16 *) stbi__malloc(img_len*2); + if (enlarged == NULL) return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory"); - for (i = 0; i < img_len; ++i) - enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff + for (i = 0; i < img_len; ++i) + enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff - STBI_FREE(orig); - return enlarged; + STBI_FREE(orig); + return enlarged; } -static void stbi__vertical_flip(void * image, int w, int h, int bytes_per_pixel) { - int row; - size_t bytes_per_row = (size_t)w * bytes_per_pixel; - stbi_uc temp[2048]; - stbi_uc * bytes = (stbi_uc *)image; +static void stbi__vertical_flip(void *image, int w, int h, int bytes_per_pixel) +{ + int row; + size_t bytes_per_row = (size_t)w * bytes_per_pixel; + stbi_uc temp[2048]; + stbi_uc *bytes = (stbi_uc *)image; - for (row = 0; row < (h >> 1); row++) { - stbi_uc * row0 = bytes + row * bytes_per_row; - stbi_uc * row1 = bytes + (h - row - 1) * bytes_per_row; - // swap row0 with row1 - size_t bytes_left = bytes_per_row; - while (bytes_left) { - size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp); - memcpy(temp, row0, bytes_copy); - memcpy(row0, row1, bytes_copy); - memcpy(row1, temp, bytes_copy); - row0 += bytes_copy; - row1 += bytes_copy; - bytes_left -= bytes_copy; - } - } + for (row = 0; row < (h>>1); row++) { + stbi_uc *row0 = bytes + row*bytes_per_row; + stbi_uc *row1 = bytes + (h - row - 1)*bytes_per_row; + // swap row0 with row1 + size_t bytes_left = bytes_per_row; + while (bytes_left) { + size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp); + memcpy(temp, row0, bytes_copy); + memcpy(row0, row1, bytes_copy); + memcpy(row1, temp, bytes_copy); + row0 += bytes_copy; + row1 += bytes_copy; + bytes_left -= bytes_copy; + } + } } #ifndef STBI_NO_GIF -static void stbi__vertical_flip_slices(void * image, int w, int h, int z, int bytes_per_pixel) { - int slice; - int slice_size = w * h * bytes_per_pixel; +static void stbi__vertical_flip_slices(void *image, int w, int h, int z, int bytes_per_pixel) +{ + int slice; + int slice_size = w * h * bytes_per_pixel; - stbi_uc * bytes = (stbi_uc *)image; - for (slice = 0; slice < z; ++slice) { - stbi__vertical_flip(bytes, w, h, bytes_per_pixel); - bytes += slice_size; - } + stbi_uc *bytes = (stbi_uc *)image; + for (slice = 0; slice < z; ++slice) { + stbi__vertical_flip(bytes, w, h, bytes_per_pixel); + bytes += slice_size; + } } #endif -static unsigned char * stbi__load_and_postprocess_8bit(stbi__context * s, int * x, int * y, int * comp, int req_comp) { - stbi__result_info ri; - void * result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8); +static unsigned char *stbi__load_and_postprocess_8bit(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + stbi__result_info ri; + void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8); - if (result == NULL) - return NULL; + if (result == NULL) + return NULL; - // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. - STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); + // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. + STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); - if (ri.bits_per_channel != 8) { - result = stbi__convert_16_to_8((stbi__uint16 *)result, *x, *y, req_comp == 0 ? *comp : req_comp); - ri.bits_per_channel = 8; - } + if (ri.bits_per_channel != 8) { + result = stbi__convert_16_to_8((stbi__uint16 *) result, *x, *y, req_comp == 0 ? *comp : req_comp); + ri.bits_per_channel = 8; + } - // @TODO: move stbi__convert_format to here + // @TODO: move stbi__convert_format to here - if (stbi__vertically_flip_on_load) { - int channels = req_comp ? req_comp : *comp; - stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc)); - } + if (stbi__vertically_flip_on_load) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc)); + } - return (unsigned char *)result; + return (unsigned char *) result; } -static stbi__uint16 * stbi__load_and_postprocess_16bit(stbi__context * s, int * x, int * y, int * comp, int req_comp) { - stbi__result_info ri; - void * result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16); +static stbi__uint16 *stbi__load_and_postprocess_16bit(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + stbi__result_info ri; + void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16); - if (result == NULL) - return NULL; + if (result == NULL) + return NULL; - // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. - STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); + // it is the responsibility of the loaders to make sure we get either 8 or 16 bit. + STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16); - if (ri.bits_per_channel != 16) { - result = stbi__convert_8_to_16((stbi_uc *)result, *x, *y, req_comp == 0 ? *comp : req_comp); - ri.bits_per_channel = 16; - } + if (ri.bits_per_channel != 16) { + result = stbi__convert_8_to_16((stbi_uc *) result, *x, *y, req_comp == 0 ? *comp : req_comp); + ri.bits_per_channel = 16; + } - // @TODO: move stbi__convert_format16 to here - // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision + // @TODO: move stbi__convert_format16 to here + // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision - if (stbi__vertically_flip_on_load) { - int channels = req_comp ? req_comp : *comp; - stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16)); - } + if (stbi__vertically_flip_on_load) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16)); + } - return (stbi__uint16 *)result; + return (stbi__uint16 *) result; } #if !defined(STBI_NO_HDR) && !defined(STBI_NO_LINEAR) -static void stbi__float_postprocess(float * result, int * x, int * y, int * comp, int req_comp) { - if (stbi__vertically_flip_on_load && result != NULL) { - int channels = req_comp ? req_comp : *comp; - stbi__vertical_flip(result, *x, *y, channels * sizeof(float)); - } +static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp) +{ + if (stbi__vertically_flip_on_load && result != NULL) { + int channels = req_comp ? req_comp : *comp; + stbi__vertical_flip(result, *x, *y, channels * sizeof(float)); + } } #endif #ifndef STBI_NO_STDIO #if defined(_WIN32) && defined(STBI_WINDOWS_UTF8) -STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char * str, - int cbmb, wchar_t * widestr, int cchwide); -STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, - const wchar_t * widestr, int cchwide, char * str, int cbmb, - const char * defchar, int * used_default); +STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide); +STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default); #endif #if defined(_WIN32) && defined(STBI_WINDOWS_UTF8) -STBIDEF int stbi_convert_wchar_to_utf8(char * buffer, size_t bufferlen, const wchar_t * input) { - return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int)bufferlen, NULL, NULL); +STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input) +{ + return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL); } #endif -static FILE * stbi__fopen(char const * filename, char const * mode) { - FILE * f; +static FILE *stbi__fopen(char const *filename, char const *mode) +{ + FILE *f; #if defined(_WIN32) && defined(STBI_WINDOWS_UTF8) - wchar_t wMode[64]; - wchar_t wFilename[1024]; - if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename) / sizeof(*wFilename))) - return 0; + wchar_t wMode[64]; + wchar_t wFilename[1024]; + if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename))) + return 0; - if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode) / sizeof(*wMode))) - return 0; + if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode))) + return 0; #if defined(_MSC_VER) && _MSC_VER >= 1400 - if (0 != _wfopen_s(&f, wFilename, wMode)) - f = 0; + if (0 != _wfopen_s(&f, wFilename, wMode)) + f = 0; #else - f = _wfopen(wFilename, wMode); + f = _wfopen(wFilename, wMode); #endif #elif defined(_MSC_VER) && _MSC_VER >= 1400 - if (0 != fopen_s(&f, filename, mode)) - f = 0; + if (0 != fopen_s(&f, filename, mode)) + f=0; #else - f = fopen(filename, mode); + f = fopen(filename, mode); #endif - return f; + return f; } -STBIDEF stbi_uc * stbi_load(char const * filename, int * x, int * y, int * comp, int req_comp) { - FILE * f = stbi__fopen(filename, "rb"); - unsigned char * result; - if (!f) - return stbi__errpuc("can't fopen", "Unable to open file"); - result = stbi_load_from_file(f, x, y, comp, req_comp); - fclose(f); - return result; + +STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + FILE *f = stbi__fopen(filename, "rb"); + unsigned char *result; + if (!f) return stbi__errpuc("can't fopen", "Unable to open file"); + result = stbi_load_from_file(f,x,y,comp,req_comp); + fclose(f); + return result; } -STBIDEF stbi_uc * stbi_load_from_file(FILE * f, int * x, int * y, int * comp, int req_comp) { - unsigned char * result; - stbi__context s; - stbi__start_file(&s, f); - result = stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp); - if (result) { - // need to 'unget' all the characters in the IO buffer - fseek(f, -(int)(s.img_buffer_end - s.img_buffer), SEEK_CUR); - } - return result; +STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + unsigned char *result; + stbi__context s; + stbi__start_file(&s,f); + result = stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); + if (result) { + // need to 'unget' all the characters in the IO buffer + fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR); + } + return result; } -STBIDEF stbi__uint16 * stbi_load_from_file_16(FILE * f, int * x, int * y, int * comp, int req_comp) { - stbi__uint16 * result; - stbi__context s; - stbi__start_file(&s, f); - result = stbi__load_and_postprocess_16bit(&s, x, y, comp, req_comp); - if (result) { - // need to 'unget' all the characters in the IO buffer - fseek(f, -(int)(s.img_buffer_end - s.img_buffer), SEEK_CUR); - } - return result; +STBIDEF stbi__uint16 *stbi_load_from_file_16(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + stbi__uint16 *result; + stbi__context s; + stbi__start_file(&s,f); + result = stbi__load_and_postprocess_16bit(&s,x,y,comp,req_comp); + if (result) { + // need to 'unget' all the characters in the IO buffer + fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR); + } + return result; } -STBIDEF stbi_us * stbi_load_16(char const * filename, int * x, int * y, int * comp, int req_comp) { - FILE * f = stbi__fopen(filename, "rb"); - stbi__uint16 * result; - if (!f) - return (stbi_us *)stbi__errpuc("can't fopen", "Unable to open file"); - result = stbi_load_from_file_16(f, x, y, comp, req_comp); - fclose(f); - return result; +STBIDEF stbi_us *stbi_load_16(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + FILE *f = stbi__fopen(filename, "rb"); + stbi__uint16 *result; + if (!f) return (stbi_us *) stbi__errpuc("can't fopen", "Unable to open file"); + result = stbi_load_from_file_16(f,x,y,comp,req_comp); + fclose(f); + return result; } -#endif //! STBI_NO_STDIO -STBIDEF stbi_us * stbi_load_16_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * channels_in_file, - int desired_channels) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__load_and_postprocess_16bit(&s, x, y, channels_in_file, desired_channels); +#endif //!STBI_NO_STDIO + +STBIDEF stbi_us *stbi_load_16_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels); } -STBIDEF stbi_us * stbi_load_16_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, - int * channels_in_file, int desired_channels) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__load_and_postprocess_16bit(&s, x, y, channels_in_file, desired_channels); +STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); + return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels); } -STBIDEF stbi_uc * stbi_load_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp, int req_comp) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp); +STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); } -STBIDEF stbi_uc * stbi_load_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp, - int req_comp) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__load_and_postprocess_8bit(&s, x, y, comp, req_comp); +STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp); } #ifndef STBI_NO_GIF -STBIDEF stbi_uc * stbi_load_gif_from_memory(stbi_uc const * buffer, int len, int ** delays, int * x, int * y, int * z, - int * comp, int req_comp) { - unsigned char * result; - stbi__context s; - stbi__start_mem(&s, buffer, len); +STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp) +{ + unsigned char *result; + stbi__context s; + stbi__start_mem(&s,buffer,len); - result = (unsigned char *)stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp); - if (stbi__vertically_flip_on_load) { - stbi__vertical_flip_slices(result, *x, *y, *z, *comp); - } + result = (unsigned char*) stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp); + if (stbi__vertically_flip_on_load) { + stbi__vertical_flip_slices( result, *x, *y, *z, *comp ); + } - return result; + return result; } #endif #ifndef STBI_NO_LINEAR -static float * stbi__loadf_main(stbi__context * s, int * x, int * y, int * comp, int req_comp) { - unsigned char * data; -#ifndef STBI_NO_HDR - if (stbi__hdr_test(s)) { - stbi__result_info ri; - float * hdr_data = stbi__hdr_load(s, x, y, comp, req_comp, &ri); - if (hdr_data) - stbi__float_postprocess(hdr_data, x, y, comp, req_comp); - return hdr_data; - } -#endif - data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp); - if (data) - return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp); - return stbi__errpf("unknown image type", "Image not of any known type, or corrupt"); +static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp) +{ + unsigned char *data; + #ifndef STBI_NO_HDR + if (stbi__hdr_test(s)) { + stbi__result_info ri; + float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp, &ri); + if (hdr_data) + stbi__float_postprocess(hdr_data,x,y,comp,req_comp); + return hdr_data; + } + #endif + data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp); + if (data) + return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp); + return stbi__errpf("unknown image type", "Image not of any known type, or corrupt"); } -STBIDEF float * stbi_loadf_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp, int req_comp) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__loadf_main(&s, x, y, comp, req_comp); +STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__loadf_main(&s,x,y,comp,req_comp); } -STBIDEF float * stbi_loadf_from_callbacks(stbi_io_callbacks const * clbk, void * user, int * x, int * y, int * comp, - int req_comp) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__loadf_main(&s, x, y, comp, req_comp); +STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__loadf_main(&s,x,y,comp,req_comp); } #ifndef STBI_NO_STDIO -STBIDEF float * stbi_loadf(char const * filename, int * x, int * y, int * comp, int req_comp) { - float * result; - FILE * f = stbi__fopen(filename, "rb"); - if (!f) - return stbi__errpf("can't fopen", "Unable to open file"); - result = stbi_loadf_from_file(f, x, y, comp, req_comp); - fclose(f); - return result; +STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp) +{ + float *result; + FILE *f = stbi__fopen(filename, "rb"); + if (!f) return stbi__errpf("can't fopen", "Unable to open file"); + result = stbi_loadf_from_file(f,x,y,comp,req_comp); + fclose(f); + return result; } -STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * comp, int req_comp) { - stbi__context s; - stbi__start_file(&s, f); - return stbi__loadf_main(&s, x, y, comp, req_comp); +STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp) +{ + stbi__context s; + stbi__start_file(&s,f); + return stbi__loadf_main(&s,x,y,comp,req_comp); } #endif // !STBI_NO_STDIO @@ -1468,208 +1514,222 @@ STBIDEF float * stbi_loadf_from_file(FILE * f, int * x, int * y, int * comp, int // defined, for API simplicity; if STBI_NO_LINEAR is defined, it always // reports false! -STBIDEF int stbi_is_hdr_from_memory(stbi_uc const * buffer, int len) { -#ifndef STBI_NO_HDR - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__hdr_test(&s); -#else - STBI_NOTUSED(buffer); - STBI_NOTUSED(len); - return 0; -#endif +STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len) +{ + #ifndef STBI_NO_HDR + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__hdr_test(&s); + #else + STBI_NOTUSED(buffer); + STBI_NOTUSED(len); + return 0; + #endif } #ifndef STBI_NO_STDIO -STBIDEF int stbi_is_hdr(char const * filename) { - FILE * f = stbi__fopen(filename, "rb"); - int result = 0; - if (f) { - result = stbi_is_hdr_from_file(f); - fclose(f); - } - return result; +STBIDEF int stbi_is_hdr (char const *filename) +{ + FILE *f = stbi__fopen(filename, "rb"); + int result=0; + if (f) { + result = stbi_is_hdr_from_file(f); + fclose(f); + } + return result; } -STBIDEF int stbi_is_hdr_from_file(FILE * f) { -#ifndef STBI_NO_HDR - long pos = ftell(f); - int res; - stbi__context s; - stbi__start_file(&s, f); - res = stbi__hdr_test(&s); - fseek(f, pos, SEEK_SET); - return res; -#else - STBI_NOTUSED(f); - return 0; -#endif +STBIDEF int stbi_is_hdr_from_file(FILE *f) +{ + #ifndef STBI_NO_HDR + long pos = ftell(f); + int res; + stbi__context s; + stbi__start_file(&s,f); + res = stbi__hdr_test(&s); + fseek(f, pos, SEEK_SET); + return res; + #else + STBI_NOTUSED(f); + return 0; + #endif } #endif // !STBI_NO_STDIO -STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const * clbk, void * user) { -#ifndef STBI_NO_HDR - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user); - return stbi__hdr_test(&s); -#else - STBI_NOTUSED(clbk); - STBI_NOTUSED(user); - return 0; -#endif +STBIDEF int stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user) +{ + #ifndef STBI_NO_HDR + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user); + return stbi__hdr_test(&s); + #else + STBI_NOTUSED(clbk); + STBI_NOTUSED(user); + return 0; + #endif } #ifndef STBI_NO_LINEAR -static float stbi__l2h_gamma = 2.2f, stbi__l2h_scale = 1.0f; +static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f; -STBIDEF void stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; } -STBIDEF void stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; } +STBIDEF void stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; } +STBIDEF void stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; } #endif -static float stbi__h2l_gamma_i = 1.0f / 2.2f, stbi__h2l_scale_i = 1.0f; +static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f; + +STBIDEF void stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; } +STBIDEF void stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; } -STBIDEF void stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1 / gamma; } -STBIDEF void stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1 / scale; } ////////////////////////////////////////////////////////////////////////////// // // Common code used by all image loaders // -enum { STBI__SCAN_load = 0, STBI__SCAN_type, STBI__SCAN_header }; +enum +{ + STBI__SCAN_load=0, + STBI__SCAN_type, + STBI__SCAN_header +}; -static void stbi__refill_buffer(stbi__context * s) { - int n = (s->io.read)(s->io_user_data, (char *)s->buffer_start, s->buflen); - s->callback_already_read += (int)(s->img_buffer - s->img_buffer_original); - if (n == 0) { - // at end of file, treat same as if from memory, but need to handle case - // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file - s->read_from_callbacks = 0; - s->img_buffer = s->buffer_start; - s->img_buffer_end = s->buffer_start + 1; - *s->img_buffer = 0; - } else { - s->img_buffer = s->buffer_start; - s->img_buffer_end = s->buffer_start + n; - } +static void stbi__refill_buffer(stbi__context *s) +{ + int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen); + s->callback_already_read += (int) (s->img_buffer - s->img_buffer_original); + if (n == 0) { + // at end of file, treat same as if from memory, but need to handle case + // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file + s->read_from_callbacks = 0; + s->img_buffer = s->buffer_start; + s->img_buffer_end = s->buffer_start+1; + *s->img_buffer = 0; + } else { + s->img_buffer = s->buffer_start; + s->img_buffer_end = s->buffer_start + n; + } } -stbi_inline static stbi_uc stbi__get8(stbi__context * s) { - if (s->img_buffer < s->img_buffer_end) - return *s->img_buffer++; - if (s->read_from_callbacks) { - stbi__refill_buffer(s); - return *s->img_buffer++; - } - return 0; +stbi_inline static stbi_uc stbi__get8(stbi__context *s) +{ + if (s->img_buffer < s->img_buffer_end) + return *s->img_buffer++; + if (s->read_from_callbacks) { + stbi__refill_buffer(s); + return *s->img_buffer++; + } + return 0; } #if defined(STBI_NO_JPEG) && defined(STBI_NO_HDR) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) // nothing #else -stbi_inline static int stbi__at_eof(stbi__context * s) { - if (s->io.read) { - if (!(s->io.eof)(s->io_user_data)) - return 0; - // if feof() is true, check if buffer = end - // special case: we've only got the special 0 character at the end - if (s->read_from_callbacks == 0) - return 1; - } +stbi_inline static int stbi__at_eof(stbi__context *s) +{ + if (s->io.read) { + if (!(s->io.eof)(s->io_user_data)) return 0; + // if feof() is true, check if buffer = end + // special case: we've only got the special 0 character at the end + if (s->read_from_callbacks == 0) return 1; + } - return s->img_buffer >= s->img_buffer_end; + return s->img_buffer >= s->img_buffer_end; } #endif -#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && \ - defined(STBI_NO_GIF) && defined(STBI_NO_PIC) +#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) // nothing #else -static void stbi__skip(stbi__context * s, int n) { - if (n == 0) - return; // already there! - if (n < 0) { - s->img_buffer = s->img_buffer_end; - return; - } - if (s->io.read) { - int blen = (int)(s->img_buffer_end - s->img_buffer); - if (blen < n) { - s->img_buffer = s->img_buffer_end; - (s->io.skip)(s->io_user_data, n - blen); - return; - } - } - s->img_buffer += n; +static void stbi__skip(stbi__context *s, int n) +{ + if (n == 0) return; // already there! + if (n < 0) { + s->img_buffer = s->img_buffer_end; + return; + } + if (s->io.read) { + int blen = (int) (s->img_buffer_end - s->img_buffer); + if (blen < n) { + s->img_buffer = s->img_buffer_end; + (s->io.skip)(s->io_user_data, n - blen); + return; + } + } + s->img_buffer += n; } #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_TGA) && defined(STBI_NO_HDR) && defined(STBI_NO_PNM) // nothing #else -static int stbi__getn(stbi__context * s, stbi_uc * buffer, int n) { - if (s->io.read) { - int blen = (int)(s->img_buffer_end - s->img_buffer); - if (blen < n) { - int res, count; +static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n) +{ + if (s->io.read) { + int blen = (int) (s->img_buffer_end - s->img_buffer); + if (blen < n) { + int res, count; - memcpy(buffer, s->img_buffer, blen); + memcpy(buffer, s->img_buffer, blen); - count = (s->io.read)(s->io_user_data, (char *)buffer + blen, n - blen); - res = (count == (n - blen)); - s->img_buffer = s->img_buffer_end; - return res; - } - } + count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen); + res = (count == (n-blen)); + s->img_buffer = s->img_buffer_end; + return res; + } + } - if (s->img_buffer + n <= s->img_buffer_end) { - memcpy(buffer, s->img_buffer, n); - s->img_buffer += n; - return 1; - } else - return 0; + if (s->img_buffer+n <= s->img_buffer_end) { + memcpy(buffer, s->img_buffer, n); + s->img_buffer += n; + return 1; + } else + return 0; } #endif #if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC) // nothing #else -static int stbi__get16be(stbi__context * s) { - int z = stbi__get8(s); - return (z << 8) + stbi__get8(s); +static int stbi__get16be(stbi__context *s) +{ + int z = stbi__get8(s); + return (z << 8) + stbi__get8(s); } #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC) // nothing #else -static stbi__uint32 stbi__get32be(stbi__context * s) { - stbi__uint32 z = stbi__get16be(s); - return (z << 16) + stbi__get16be(s); +static stbi__uint32 stbi__get32be(stbi__context *s) +{ + stbi__uint32 z = stbi__get16be(s); + return (z << 16) + stbi__get16be(s); } #endif #if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) // nothing #else -static int stbi__get16le(stbi__context * s) { - int z = stbi__get8(s); - return z + (stbi__get8(s) << 8); +static int stbi__get16le(stbi__context *s) +{ + int z = stbi__get8(s); + return z + (stbi__get8(s) << 8); } #endif #ifndef STBI_NO_BMP -static stbi__uint32 stbi__get32le(stbi__context * s) { - stbi__uint32 z = stbi__get16le(s); - z += (stbi__uint32)stbi__get16le(s) << 16; - return z; +static stbi__uint32 stbi__get32le(stbi__context *s) +{ + stbi__uint32 z = stbi__get16le(s); + z += (stbi__uint32)stbi__get16le(s) << 16; + return z; } #endif -#define STBI__BYTECAST(x) ((stbi_uc)((x)&255)) // truncate int to byte without warnings +#define STBI__BYTECAST(x) ((stbi_uc) ((x) & 255)) // truncate int to byte without warnings -#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && \ - defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) +#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) // nothing #else ////////////////////////////////////////////////////////////////////////////// @@ -1683,264 +1743,169 @@ static stbi__uint32 stbi__get32le(stbi__context * s) { // assume data buffer is malloced, so malloc a new one and free that one // only failure mode is malloc failing -static stbi_uc stbi__compute_y(int r, int g, int b) { return (stbi_uc)(((r * 77) + (g * 150) + (29 * b)) >> 8); } +static stbi_uc stbi__compute_y(int r, int g, int b) +{ + return (stbi_uc) (((r*77) + (g*150) + (29*b)) >> 8); +} #endif -#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && \ - defined(STBI_NO_PIC) && defined(STBI_NO_PNM) +#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM) // nothing #else -static unsigned char * stbi__convert_format(unsigned char * data, int img_n, int req_comp, unsigned int x, unsigned int y) { - int i, j; - unsigned char * good; +static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y) +{ + int i,j; + unsigned char *good; - if (req_comp == img_n) - return data; - STBI_ASSERT(req_comp >= 1 && req_comp <= 4); + if (req_comp == img_n) return data; + STBI_ASSERT(req_comp >= 1 && req_comp <= 4); - good = (unsigned char *)stbi__malloc_mad3(req_comp, x, y, 0); - if (good == NULL) { - STBI_FREE(data); - return stbi__errpuc("outofmem", "Out of memory"); - } + good = (unsigned char *) stbi__malloc_mad3(req_comp, x, y, 0); + if (good == NULL) { + STBI_FREE(data); + return stbi__errpuc("outofmem", "Out of memory"); + } - for (j = 0; j < (int)y; ++j) { - unsigned char * src = data + j * x * img_n; - unsigned char * dest = good + j * x * req_comp; + for (j=0; j < (int) y; ++j) { + unsigned char *src = data + j * x * img_n ; + unsigned char *dest = good + j * x * req_comp; -#define STBI__COMBO(a, b) ((a)*8 + (b)) -#define STBI__CASE(a, b) \ - case STBI__COMBO(a, b): \ - for (i = x - 1; i >= 0; --i, src += a, dest += b) - // convert source image with img_n components to one with req_comp components; - // avoid switch per pixel, so use switch per scanline and massive macros - switch (STBI__COMBO(img_n, req_comp)) { - STBI__CASE(1, 2) { - dest[0] = src[0]; - dest[1] = 255; - } - break; - STBI__CASE(1, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(1, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = 255; - } - break; - STBI__CASE(2, 1) { dest[0] = src[0]; } - break; - STBI__CASE(2, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(2, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = src[1]; - } - break; - STBI__CASE(3, 4) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - dest[3] = 255; - } - break; - STBI__CASE(3, 1) { dest[0] = stbi__compute_y(src[0], src[1], src[2]); } - break; - STBI__CASE(3, 2) { - dest[0] = stbi__compute_y(src[0], src[1], src[2]); - dest[1] = 255; - } - break; - STBI__CASE(4, 1) { dest[0] = stbi__compute_y(src[0], src[1], src[2]); } - break; - STBI__CASE(4, 2) { - dest[0] = stbi__compute_y(src[0], src[1], src[2]); - dest[1] = src[3]; - } - break; - STBI__CASE(4, 3) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - } - break; - default: - STBI_ASSERT(0); - STBI_FREE(data); - STBI_FREE(good); - return stbi__errpuc("unsupported", "Unsupported format conversion"); - } -#undef STBI__CASE - } + #define STBI__COMBO(a,b) ((a)*8+(b)) + #define STBI__CASE(a,b) case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b) + // convert source image with img_n components to one with req_comp components; + // avoid switch per pixel, so use switch per scanline and massive macros + switch (STBI__COMBO(img_n, req_comp)) { + STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=255; } break; + STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=255; } break; + STBI__CASE(2,1) { dest[0]=src[0]; } break; + STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1]; } break; + STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=255; } break; + STBI__CASE(3,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); } break; + STBI__CASE(3,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = 255; } break; + STBI__CASE(4,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); } break; + STBI__CASE(4,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = src[3]; } break; + STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2]; } break; + default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return stbi__errpuc("unsupported", "Unsupported format conversion"); + } + #undef STBI__CASE + } - STBI_FREE(data); - return good; + STBI_FREE(data); + return good; } #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) // nothing #else -static stbi__uint16 stbi__compute_y_16(int r, int g, int b) { return (stbi__uint16)(((r * 77) + (g * 150) + (29 * b)) >> 8); } +static stbi__uint16 stbi__compute_y_16(int r, int g, int b) +{ + return (stbi__uint16) (((r*77) + (g*150) + (29*b)) >> 8); +} #endif #if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) // nothing #else -static stbi__uint16 * stbi__convert_format16(stbi__uint16 * data, int img_n, int req_comp, unsigned int x, unsigned int y) { - int i, j; - stbi__uint16 * good; +static stbi__uint16 *stbi__convert_format16(stbi__uint16 *data, int img_n, int req_comp, unsigned int x, unsigned int y) +{ + int i,j; + stbi__uint16 *good; - if (req_comp == img_n) - return data; - STBI_ASSERT(req_comp >= 1 && req_comp <= 4); + if (req_comp == img_n) return data; + STBI_ASSERT(req_comp >= 1 && req_comp <= 4); - good = (stbi__uint16 *)stbi__malloc(req_comp * x * y * 2); - if (good == NULL) { - STBI_FREE(data); - return (stbi__uint16 *)stbi__errpuc("outofmem", "Out of memory"); - } + good = (stbi__uint16 *) stbi__malloc(req_comp * x * y * 2); + if (good == NULL) { + STBI_FREE(data); + return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory"); + } - for (j = 0; j < (int)y; ++j) { - stbi__uint16 * src = data + j * x * img_n; - stbi__uint16 * dest = good + j * x * req_comp; + for (j=0; j < (int) y; ++j) { + stbi__uint16 *src = data + j * x * img_n ; + stbi__uint16 *dest = good + j * x * req_comp; -#define STBI__COMBO(a, b) ((a)*8 + (b)) -#define STBI__CASE(a, b) \ - case STBI__COMBO(a, b): \ - for (i = x - 1; i >= 0; --i, src += a, dest += b) - // convert source image with img_n components to one with req_comp components; - // avoid switch per pixel, so use switch per scanline and massive macros - switch (STBI__COMBO(img_n, req_comp)) { - STBI__CASE(1, 2) { - dest[0] = src[0]; - dest[1] = 0xffff; - } - break; - STBI__CASE(1, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(1, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = 0xffff; - } - break; - STBI__CASE(2, 1) { dest[0] = src[0]; } - break; - STBI__CASE(2, 3) { dest[0] = dest[1] = dest[2] = src[0]; } - break; - STBI__CASE(2, 4) { - dest[0] = dest[1] = dest[2] = src[0]; - dest[3] = src[1]; - } - break; - STBI__CASE(3, 4) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - dest[3] = 0xffff; - } - break; - STBI__CASE(3, 1) { dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); } - break; - STBI__CASE(3, 2) { - dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); - dest[1] = 0xffff; - } - break; - STBI__CASE(4, 1) { dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); } - break; - STBI__CASE(4, 2) { - dest[0] = stbi__compute_y_16(src[0], src[1], src[2]); - dest[1] = src[3]; - } - break; - STBI__CASE(4, 3) { - dest[0] = src[0]; - dest[1] = src[1]; - dest[2] = src[2]; - } - break; - default: - STBI_ASSERT(0); - STBI_FREE(data); - STBI_FREE(good); - return (stbi__uint16 *)stbi__errpuc("unsupported", "Unsupported format conversion"); - } -#undef STBI__CASE - } + #define STBI__COMBO(a,b) ((a)*8+(b)) + #define STBI__CASE(a,b) case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b) + // convert source image with img_n components to one with req_comp components; + // avoid switch per pixel, so use switch per scanline and massive macros + switch (STBI__COMBO(img_n, req_comp)) { + STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=0xffff; } break; + STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=0xffff; } break; + STBI__CASE(2,1) { dest[0]=src[0]; } break; + STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0]; } break; + STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1]; } break; + STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=0xffff; } break; + STBI__CASE(3,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); } break; + STBI__CASE(3,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = 0xffff; } break; + STBI__CASE(4,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); } break; + STBI__CASE(4,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = src[3]; } break; + STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2]; } break; + default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return (stbi__uint16*) stbi__errpuc("unsupported", "Unsupported format conversion"); + } + #undef STBI__CASE + } - STBI_FREE(data); - return good; + STBI_FREE(data); + return good; } #endif #ifndef STBI_NO_LINEAR -static float * stbi__ldr_to_hdr(stbi_uc * data, int x, int y, int comp) { - int i, k, n; - float * output; - if (!data) - return NULL; - output = (float *)stbi__malloc_mad4(x, y, comp, sizeof(float), 0); - if (output == NULL) { - STBI_FREE(data); - return stbi__errpf("outofmem", "Out of memory"); - } - // compute number of non-alpha components - if (comp & 1) - n = comp; - else - n = comp - 1; - for (i = 0; i < x * y; ++i) { - for (k = 0; k < n; ++k) { - output[i * comp + k] = (float)(pow(data[i * comp + k] / 255.0f, stbi__l2h_gamma) * stbi__l2h_scale); - } - } - if (n < comp) { - for (i = 0; i < x * y; ++i) { - output[i * comp + n] = data[i * comp + n] / 255.0f; - } - } - STBI_FREE(data); - return output; +static float *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp) +{ + int i,k,n; + float *output; + if (!data) return NULL; + output = (float *) stbi__malloc_mad4(x, y, comp, sizeof(float), 0); + if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); } + // compute number of non-alpha components + if (comp & 1) n = comp; else n = comp-1; + for (i=0; i < x*y; ++i) { + for (k=0; k < n; ++k) { + output[i*comp + k] = (float) (pow(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale); + } + } + if (n < comp) { + for (i=0; i < x*y; ++i) { + output[i*comp + n] = data[i*comp + n]/255.0f; + } + } + STBI_FREE(data); + return output; } #endif #ifndef STBI_NO_HDR -#define stbi__float2int(x) ((int)(x)) -static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp) { - int i, k, n; - stbi_uc * output; - if (!data) - return NULL; - output = (stbi_uc *)stbi__malloc_mad3(x, y, comp, 0); - if (output == NULL) { - STBI_FREE(data); - return stbi__errpuc("outofmem", "Out of memory"); - } - // compute number of non-alpha components - if (comp & 1) - n = comp; - else - n = comp - 1; - for (i = 0; i < x * y; ++i) { - for (k = 0; k < n; ++k) { - float z = (float)pow(data[i * comp + k] * stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f; - if (z < 0) - z = 0; - if (z > 255) - z = 255; - output[i * comp + k] = (stbi_uc)stbi__float2int(z); - } - if (k < comp) { - float z = data[i * comp + k] * 255 + 0.5f; - if (z < 0) - z = 0; - if (z > 255) - z = 255; - output[i * comp + k] = (stbi_uc)stbi__float2int(z); - } - } - STBI_FREE(data); - return output; +#define stbi__float2int(x) ((int) (x)) +static stbi_uc *stbi__hdr_to_ldr(float *data, int x, int y, int comp) +{ + int i,k,n; + stbi_uc *output; + if (!data) return NULL; + output = (stbi_uc *) stbi__malloc_mad3(x, y, comp, 0); + if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); } + // compute number of non-alpha components + if (comp & 1) n = comp; else n = comp-1; + for (i=0; i < x*y; ++i) { + for (k=0; k < n; ++k) { + float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f; + if (z < 0) z = 0; + if (z > 255) z = 255; + output[i*comp + k] = (stbi_uc) stbi__float2int(z); + } + if (k < comp) { + float z = data[i*comp+k] * 255 + 0.5f; + if (z < 0) z = 0; + if (z > 255) z = 255; + output[i*comp + k] = (stbi_uc) stbi__float2int(z); + } + } + STBI_FREE(data); + return output; } #endif @@ -1968,783 +1933,763 @@ static stbi_uc * stbi__hdr_to_ldr(float * data, int x, int y, int comp) { #ifndef STBI_NO_JPEG // huffman decoding acceleration -#define FAST_BITS 9 // larger handles more cases; smaller stomps less cache +#define FAST_BITS 9 // larger handles more cases; smaller stomps less cache -typedef struct { - stbi_uc fast[1 << FAST_BITS]; - // weirdly, repacking this into AoS is a 10% speed loss, instead of a win - stbi__uint16 code[256]; - stbi_uc values[256]; - stbi_uc size[257]; - unsigned int maxcode[18]; - int delta[17]; // old 'firstsymbol' - old 'firstcode' +typedef struct +{ + stbi_uc fast[1 << FAST_BITS]; + // weirdly, repacking this into AoS is a 10% speed loss, instead of a win + stbi__uint16 code[256]; + stbi_uc values[256]; + stbi_uc size[257]; + unsigned int maxcode[18]; + int delta[17]; // old 'firstsymbol' - old 'firstcode' } stbi__huffman; -typedef struct { - stbi__context * s; - stbi__huffman huff_dc[4]; - stbi__huffman huff_ac[4]; - stbi__uint16 dequant[4][64]; - stbi__int16 fast_ac[4][1 << FAST_BITS]; +typedef struct +{ + stbi__context *s; + stbi__huffman huff_dc[4]; + stbi__huffman huff_ac[4]; + stbi__uint16 dequant[4][64]; + stbi__int16 fast_ac[4][1 << FAST_BITS]; - // sizes for components, interleaved MCUs - int img_h_max, img_v_max; - int img_mcu_x, img_mcu_y; - int img_mcu_w, img_mcu_h; +// sizes for components, interleaved MCUs + int img_h_max, img_v_max; + int img_mcu_x, img_mcu_y; + int img_mcu_w, img_mcu_h; - // definition of jpeg image component - struct { - int id; - int h, v; - int tq; - int hd, ha; - int dc_pred; +// definition of jpeg image component + struct + { + int id; + int h,v; + int tq; + int hd,ha; + int dc_pred; - int x, y, w2, h2; - stbi_uc * data; - void *raw_data, *raw_coeff; - stbi_uc * linebuf; - short * coeff; // progressive only - int coeff_w, coeff_h; // number of 8x8 coefficient blocks - } img_comp[4]; + int x,y,w2,h2; + stbi_uc *data; + void *raw_data, *raw_coeff; + stbi_uc *linebuf; + short *coeff; // progressive only + int coeff_w, coeff_h; // number of 8x8 coefficient blocks + } img_comp[4]; - stbi__uint32 code_buffer; // jpeg entropy-coded buffer - int code_bits; // number of valid bits - unsigned char marker; // marker seen while filling entropy buffer - int nomore; // flag if we saw a marker so must stop + stbi__uint32 code_buffer; // jpeg entropy-coded buffer + int code_bits; // number of valid bits + unsigned char marker; // marker seen while filling entropy buffer + int nomore; // flag if we saw a marker so must stop - int progressive; - int spec_start; - int spec_end; - int succ_high; - int succ_low; - int eob_run; - int jfif; - int app14_color_transform; // Adobe APP14 tag - int rgb; + int progressive; + int spec_start; + int spec_end; + int succ_high; + int succ_low; + int eob_run; + int jfif; + int app14_color_transform; // Adobe APP14 tag + int rgb; - int scan_n, order[4]; - int restart_interval, todo; + int scan_n, order[4]; + int restart_interval, todo; - // kernels - void (*idct_block_kernel)(stbi_uc * out, int out_stride, short data[64]); - void (*YCbCr_to_RGB_kernel)(stbi_uc * out, const stbi_uc * y, const stbi_uc * pcb, const stbi_uc * pcr, int count, - int step); - stbi_uc * (*resample_row_hv_2_kernel)(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs); +// kernels + void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]); + void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step); + stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs); } stbi__jpeg; -static int stbi__build_huffman(stbi__huffman * h, int * count) { - int i, j, k = 0; - unsigned int code; - // build size list for each symbol (from JPEG spec) - for (i = 0; i < 16; ++i) { - for (j = 0; j < count[i]; ++j) { - h->size[k++] = (stbi_uc)(i + 1); - if (k >= 257) - return stbi__err("bad size list", "Corrupt JPEG"); - } - } - h->size[k] = 0; +static int stbi__build_huffman(stbi__huffman *h, int *count) +{ + int i,j,k=0; + unsigned int code; + // build size list for each symbol (from JPEG spec) + for (i=0; i < 16; ++i) { + for (j=0; j < count[i]; ++j) { + h->size[k++] = (stbi_uc) (i+1); + if(k >= 257) return stbi__err("bad size list","Corrupt JPEG"); + } + } + h->size[k] = 0; - // compute actual symbols (from jpeg spec) - code = 0; - k = 0; - for (j = 1; j <= 16; ++j) { - // compute delta to add to code to compute symbol id - h->delta[j] = k - code; - if (h->size[k] == j) { - while (h->size[k] == j) - h->code[k++] = (stbi__uint16)(code++); - if (code - 1 >= (1u << j)) - return stbi__err("bad code lengths", "Corrupt JPEG"); - } - // compute largest code + 1 for this size, preshifted as needed later - h->maxcode[j] = code << (16 - j); - code <<= 1; - } - h->maxcode[j] = 0xffffffff; + // compute actual symbols (from jpeg spec) + code = 0; + k = 0; + for(j=1; j <= 16; ++j) { + // compute delta to add to code to compute symbol id + h->delta[j] = k - code; + if (h->size[k] == j) { + while (h->size[k] == j) + h->code[k++] = (stbi__uint16) (code++); + if (code-1 >= (1u << j)) return stbi__err("bad code lengths","Corrupt JPEG"); + } + // compute largest code + 1 for this size, preshifted as needed later + h->maxcode[j] = code << (16-j); + code <<= 1; + } + h->maxcode[j] = 0xffffffff; - // build non-spec acceleration table; 255 is flag for not-accelerated - memset(h->fast, 255, 1 << FAST_BITS); - for (i = 0; i < k; ++i) { - int s = h->size[i]; - if (s <= FAST_BITS) { - int c = h->code[i] << (FAST_BITS - s); - int m = 1 << (FAST_BITS - s); - for (j = 0; j < m; ++j) { - h->fast[c + j] = (stbi_uc)i; - } - } - } - return 1; + // build non-spec acceleration table; 255 is flag for not-accelerated + memset(h->fast, 255, 1 << FAST_BITS); + for (i=0; i < k; ++i) { + int s = h->size[i]; + if (s <= FAST_BITS) { + int c = h->code[i] << (FAST_BITS-s); + int m = 1 << (FAST_BITS-s); + for (j=0; j < m; ++j) { + h->fast[c+j] = (stbi_uc) i; + } + } + } + return 1; } // build a table that decodes both magnitude and value of small ACs in // one go. -static void stbi__build_fast_ac(stbi__int16 * fast_ac, stbi__huffman * h) { - int i; - for (i = 0; i < (1 << FAST_BITS); ++i) { - stbi_uc fast = h->fast[i]; - fast_ac[i] = 0; - if (fast < 255) { - int rs = h->values[fast]; - int run = (rs >> 4) & 15; - int magbits = rs & 15; - int len = h->size[fast]; +static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h) +{ + int i; + for (i=0; i < (1 << FAST_BITS); ++i) { + stbi_uc fast = h->fast[i]; + fast_ac[i] = 0; + if (fast < 255) { + int rs = h->values[fast]; + int run = (rs >> 4) & 15; + int magbits = rs & 15; + int len = h->size[fast]; - if (magbits && len + magbits <= FAST_BITS) { - // magnitude code followed by receive_extend code - int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits); - int m = 1 << (magbits - 1); - if (k < m) - k += (~0U << magbits) + 1; - // if the result is small enough, we can fit it in fast_ac table - if (k >= -128 && k <= 127) - fast_ac[i] = (stbi__int16)((k * 256) + (run * 16) + (len + magbits)); - } - } - } + if (magbits && len + magbits <= FAST_BITS) { + // magnitude code followed by receive_extend code + int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits); + int m = 1 << (magbits - 1); + if (k < m) k += (~0U << magbits) + 1; + // if the result is small enough, we can fit it in fast_ac table + if (k >= -128 && k <= 127) + fast_ac[i] = (stbi__int16) ((k * 256) + (run * 16) + (len + magbits)); + } + } + } } -static void stbi__grow_buffer_unsafe(stbi__jpeg * j) { - do { - unsigned int b = j->nomore ? 0 : stbi__get8(j->s); - if (b == 0xff) { - int c = stbi__get8(j->s); - while (c == 0xff) - c = stbi__get8(j->s); // consume fill bytes - if (c != 0) { - j->marker = (unsigned char)c; - j->nomore = 1; - return; - } - } - j->code_buffer |= b << (24 - j->code_bits); - j->code_bits += 8; - } while (j->code_bits <= 24); +static void stbi__grow_buffer_unsafe(stbi__jpeg *j) +{ + do { + unsigned int b = j->nomore ? 0 : stbi__get8(j->s); + if (b == 0xff) { + int c = stbi__get8(j->s); + while (c == 0xff) c = stbi__get8(j->s); // consume fill bytes + if (c != 0) { + j->marker = (unsigned char) c; + j->nomore = 1; + return; + } + } + j->code_buffer |= b << (24 - j->code_bits); + j->code_bits += 8; + } while (j->code_bits <= 24); } // (1 << n) - 1 -static const stbi__uint32 stbi__bmask[17] = {0, 1, 3, 7, 15, 31, 63, 127, 255, - 511, 1023, 2047, 4095, 8191, 16383, 32767, 65535}; +static const stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535}; // decode a jpeg huffman value from the bitstream -stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg * j, stbi__huffman * h) { - unsigned int temp; - int c, k; +stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h) +{ + unsigned int temp; + int c,k; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); - // look at the top FAST_BITS and determine what symbol ID it is, - // if the code is <= FAST_BITS - c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1); - k = h->fast[c]; - if (k < 255) { - int s = h->size[k]; - if (s > j->code_bits) - return -1; - j->code_buffer <<= s; - j->code_bits -= s; - return h->values[k]; - } + // look at the top FAST_BITS and determine what symbol ID it is, + // if the code is <= FAST_BITS + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + k = h->fast[c]; + if (k < 255) { + int s = h->size[k]; + if (s > j->code_bits) + return -1; + j->code_buffer <<= s; + j->code_bits -= s; + return h->values[k]; + } - // naive test is to shift the code_buffer down so k bits are - // valid, then test against maxcode. To speed this up, we've - // preshifted maxcode left so that it has (16-k) 0s at the - // end; in other words, regardless of the number of bits, it - // wants to be compared against something shifted to have 16; - // that way we don't need to shift inside the loop. - temp = j->code_buffer >> 16; - for (k = FAST_BITS + 1;; ++k) - if (temp < h->maxcode[k]) - break; - if (k == 17) { - // error! code not found - j->code_bits -= 16; - return -1; - } + // naive test is to shift the code_buffer down so k bits are + // valid, then test against maxcode. To speed this up, we've + // preshifted maxcode left so that it has (16-k) 0s at the + // end; in other words, regardless of the number of bits, it + // wants to be compared against something shifted to have 16; + // that way we don't need to shift inside the loop. + temp = j->code_buffer >> 16; + for (k=FAST_BITS+1 ; ; ++k) + if (temp < h->maxcode[k]) + break; + if (k == 17) { + // error! code not found + j->code_bits -= 16; + return -1; + } - if (k > j->code_bits) - return -1; + if (k > j->code_bits) + return -1; - // convert the huffman code to the symbol id - c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k]; - if (c < 0 || c >= 256) // symbol id out of bounds! - return -1; - STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]); + // convert the huffman code to the symbol id + c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k]; + if(c < 0 || c >= 256) // symbol id out of bounds! + return -1; + STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]); - // convert the id to a symbol - j->code_bits -= k; - j->code_buffer <<= k; - return h->values[c]; + // convert the id to a symbol + j->code_bits -= k; + j->code_buffer <<= k; + return h->values[c]; } // bias[n] = (-1<code_bits < n) - stbi__grow_buffer_unsafe(j); - if (j->code_bits < n) - return 0; // ran out of bits from stream, return 0s intead of continuing +stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n) +{ + unsigned int k; + int sgn; + if (j->code_bits < n) stbi__grow_buffer_unsafe(j); + if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing - sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative) - k = stbi_lrot(j->code_buffer, n); - j->code_buffer = k & ~stbi__bmask[n]; - k &= stbi__bmask[n]; - j->code_bits -= n; - return k + (stbi__jbias[n] & (sgn - 1)); + sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative) + k = stbi_lrot(j->code_buffer, n); + j->code_buffer = k & ~stbi__bmask[n]; + k &= stbi__bmask[n]; + j->code_bits -= n; + return k + (stbi__jbias[n] & (sgn - 1)); } // get some unsigned bits -stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg * j, int n) { - unsigned int k; - if (j->code_bits < n) - stbi__grow_buffer_unsafe(j); - if (j->code_bits < n) - return 0; // ran out of bits from stream, return 0s intead of continuing - k = stbi_lrot(j->code_buffer, n); - j->code_buffer = k & ~stbi__bmask[n]; - k &= stbi__bmask[n]; - j->code_bits -= n; - return k; +stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n) +{ + unsigned int k; + if (j->code_bits < n) stbi__grow_buffer_unsafe(j); + if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing + k = stbi_lrot(j->code_buffer, n); + j->code_buffer = k & ~stbi__bmask[n]; + k &= stbi__bmask[n]; + j->code_bits -= n; + return k; } -stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg * j) { - unsigned int k; - if (j->code_bits < 1) - stbi__grow_buffer_unsafe(j); - if (j->code_bits < 1) - return 0; // ran out of bits from stream, return 0s intead of continuing - k = j->code_buffer; - j->code_buffer <<= 1; - --j->code_bits; - return k & 0x80000000; +stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j) +{ + unsigned int k; + if (j->code_bits < 1) stbi__grow_buffer_unsafe(j); + if (j->code_bits < 1) return 0; // ran out of bits from stream, return 0s intead of continuing + k = j->code_buffer; + j->code_buffer <<= 1; + --j->code_bits; + return k & 0x80000000; } // given a value that's at position X in the zigzag stream, // where does it appear in the 8x8 matrix coded as row-major? -static const stbi_uc stbi__jpeg_dezigzag[64 + 15] = { - 0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, 27, 20, 13, 6, 7, 14, 21, 28, 35, - 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, 53, 60, 61, 54, 47, 55, 62, 63, - // let corrupt input sample past end - 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63}; +static const stbi_uc stbi__jpeg_dezigzag[64+15] = +{ + 0, 1, 8, 16, 9, 2, 3, 10, + 17, 24, 32, 25, 18, 11, 4, 5, + 12, 19, 26, 33, 40, 48, 41, 34, + 27, 20, 13, 6, 7, 14, 21, 28, + 35, 42, 49, 56, 57, 50, 43, 36, + 29, 22, 15, 23, 30, 37, 44, 51, + 58, 59, 52, 45, 38, 31, 39, 46, + 53, 60, 61, 54, 47, 55, 62, 63, + // let corrupt input sample past end + 63, 63, 63, 63, 63, 63, 63, 63, + 63, 63, 63, 63, 63, 63, 63 +}; // decode one 64-entry block-- -static int stbi__jpeg_decode_block(stbi__jpeg * j, short data[64], stbi__huffman * hdc, stbi__huffman * hac, stbi__int16 * fac, - int b, stbi__uint16 * dequant) { - int diff, dc, k; - int t; +static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi__uint16 *dequant) +{ + int diff,dc,k; + int t; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); - t = stbi__jpeg_huff_decode(j, hdc); - if (t < 0 || t > 15) - return stbi__err("bad huffman code", "Corrupt JPEG"); + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + t = stbi__jpeg_huff_decode(j, hdc); + if (t < 0 || t > 15) return stbi__err("bad huffman code","Corrupt JPEG"); - // 0 all the ac values now so we can do it 32-bits at a time - memset(data, 0, 64 * sizeof(data[0])); + // 0 all the ac values now so we can do it 32-bits at a time + memset(data,0,64*sizeof(data[0])); - diff = t ? stbi__extend_receive(j, t) : 0; - if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) - return stbi__err("bad delta", "Corrupt JPEG"); - dc = j->img_comp[b].dc_pred + diff; - j->img_comp[b].dc_pred = dc; - if (!stbi__mul2shorts_valid(dc, dequant[0])) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - data[0] = (short)(dc * dequant[0]); + diff = t ? stbi__extend_receive(j, t) : 0; + if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta","Corrupt JPEG"); + dc = j->img_comp[b].dc_pred + diff; + j->img_comp[b].dc_pred = dc; + if (!stbi__mul2shorts_valid(dc, dequant[0])) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + data[0] = (short) (dc * dequant[0]); - // decode AC components, see JPEG spec - k = 1; - do { - unsigned int zig; - int c, r, s; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); - c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1); - r = fac[c]; - if (r) { // fast-AC path - k += (r >> 4) & 15; // run - s = r & 15; // combined length - if (s > j->code_bits) - return stbi__err("bad huffman code", "Combined length longer than code bits available"); - j->code_buffer <<= s; - j->code_bits -= s; + // decode AC components, see JPEG spec + k = 1; + do { + unsigned int zig; + int c,r,s; + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + r = fac[c]; + if (r) { // fast-AC path + k += (r >> 4) & 15; // run + s = r & 15; // combined length + if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available"); + j->code_buffer <<= s; + j->code_bits -= s; + // decode into unzigzag'd location + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) ((r >> 8) * dequant[zig]); + } else { + int rs = stbi__jpeg_huff_decode(j, hac); + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (rs != 0xf0) break; // end block + k += 16; + } else { + k += r; // decode into unzigzag'd location zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)((r >> 8) * dequant[zig]); - } else { - int rs = stbi__jpeg_huff_decode(j, hac); - if (rs < 0) - return stbi__err("bad huffman code", "Corrupt JPEG"); - s = rs & 15; - r = rs >> 4; - if (s == 0) { - if (rs != 0xf0) - break; // end block - k += 16; - } else { - k += r; - // decode into unzigzag'd location - zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)(stbi__extend_receive(j, s) * dequant[zig]); - } - } - } while (k < 64); - return 1; + data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]); + } + } + } while (k < 64); + return 1; } -static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg * j, short data[64], stbi__huffman * hdc, int b) { - int diff, dc; - int t; - if (j->spec_end != 0) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); +static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b) +{ + int diff,dc; + int t; + if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); - if (j->succ_high == 0) { - // first scan for DC coefficient, must be first - memset(data, 0, 64 * sizeof(data[0])); // 0 all the ac values now - t = stbi__jpeg_huff_decode(j, hdc); - if (t < 0 || t > 15) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - diff = t ? stbi__extend_receive(j, t) : 0; + if (j->succ_high == 0) { + // first scan for DC coefficient, must be first + memset(data,0,64*sizeof(data[0])); // 0 all the ac values now + t = stbi__jpeg_huff_decode(j, hdc); + if (t < 0 || t > 15) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + diff = t ? stbi__extend_receive(j, t) : 0; - if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) - return stbi__err("bad delta", "Corrupt JPEG"); - dc = j->img_comp[b].dc_pred + diff; - j->img_comp[b].dc_pred = dc; - if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - data[0] = (short)(dc * (1 << j->succ_low)); - } else { - // refinement scan for DC coefficient - if (stbi__jpeg_get_bit(j)) - data[0] += (short)(1 << j->succ_low); - } - return 1; + if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta", "Corrupt JPEG"); + dc = j->img_comp[b].dc_pred + diff; + j->img_comp[b].dc_pred = dc; + if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); + data[0] = (short) (dc * (1 << j->succ_low)); + } else { + // refinement scan for DC coefficient + if (stbi__jpeg_get_bit(j)) + data[0] += (short) (1 << j->succ_low); + } + return 1; } // @OPTIMIZE: store non-zigzagged during the decode passes, // and only de-zigzag when dequantizing -static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg * j, short data[64], stbi__huffman * hac, stbi__int16 * fac) { - int k; - if (j->spec_start == 0) - return stbi__err("can't merge dc and ac", "Corrupt JPEG"); +static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac) +{ + int k; + if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG"); - if (j->succ_high == 0) { - int shift = j->succ_low; + if (j->succ_high == 0) { + int shift = j->succ_low; - if (j->eob_run) { - --j->eob_run; - return 1; - } + if (j->eob_run) { + --j->eob_run; + return 1; + } - k = j->spec_start; - do { - unsigned int zig; - int c, r, s; - if (j->code_bits < 16) - stbi__grow_buffer_unsafe(j); - c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS) - 1); - r = fac[c]; - if (r) { // fast-AC path - k += (r >> 4) & 15; // run - s = r & 15; // combined length - if (s > j->code_bits) - return stbi__err("bad huffman code", "Combined length longer than code bits available"); - j->code_buffer <<= s; - j->code_bits -= s; - zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)((r >> 8) * (1 << shift)); + k = j->spec_start; + do { + unsigned int zig; + int c,r,s; + if (j->code_bits < 16) stbi__grow_buffer_unsafe(j); + c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1); + r = fac[c]; + if (r) { // fast-AC path + k += (r >> 4) & 15; // run + s = r & 15; // combined length + if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available"); + j->code_buffer <<= s; + j->code_bits -= s; + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) ((r >> 8) * (1 << shift)); + } else { + int rs = stbi__jpeg_huff_decode(j, hac); + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (r < 15) { + j->eob_run = (1 << r); + if (r) + j->eob_run += stbi__jpeg_get_bits(j, r); + --j->eob_run; + break; + } + k += 16; } else { - int rs = stbi__jpeg_huff_decode(j, hac); - if (rs < 0) - return stbi__err("bad huffman code", "Corrupt JPEG"); - s = rs & 15; - r = rs >> 4; - if (s == 0) { - if (r < 15) { - j->eob_run = (1 << r); - if (r) - j->eob_run += stbi__jpeg_get_bits(j, r); - --j->eob_run; - break; - } - k += 16; - } else { - k += r; - zig = stbi__jpeg_dezigzag[k++]; - data[zig] = (short)(stbi__extend_receive(j, s) * (1 << shift)); - } + k += r; + zig = stbi__jpeg_dezigzag[k++]; + data[zig] = (short) (stbi__extend_receive(j,s) * (1 << shift)); } - } while (k <= j->spec_end); - } else { - // refinement scan for these AC coefficients + } + } while (k <= j->spec_end); + } else { + // refinement scan for these AC coefficients - short bit = (short)(1 << j->succ_low); + short bit = (short) (1 << j->succ_low); - if (j->eob_run) { - --j->eob_run; - for (k = j->spec_start; k <= j->spec_end; ++k) { - short * p = &data[stbi__jpeg_dezigzag[k]]; - if (*p != 0) - if (stbi__jpeg_get_bit(j)) - if ((*p & bit) == 0) { - if (*p > 0) - *p += bit; - else - *p -= bit; - } + if (j->eob_run) { + --j->eob_run; + for (k = j->spec_start; k <= j->spec_end; ++k) { + short *p = &data[stbi__jpeg_dezigzag[k]]; + if (*p != 0) + if (stbi__jpeg_get_bit(j)) + if ((*p & bit)==0) { + if (*p > 0) + *p += bit; + else + *p -= bit; + } + } + } else { + k = j->spec_start; + do { + int r,s; + int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh + if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG"); + s = rs & 15; + r = rs >> 4; + if (s == 0) { + if (r < 15) { + j->eob_run = (1 << r) - 1; + if (r) + j->eob_run += stbi__jpeg_get_bits(j, r); + r = 64; // force end of block + } else { + // r=15 s=0 should write 16 0s, so we just do + // a run of 15 0s and then write s (which is 0), + // so we don't have to do anything special here + } + } else { + if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG"); + // sign bit + if (stbi__jpeg_get_bit(j)) + s = bit; + else + s = -bit; } - } else { - k = j->spec_start; - do { - int r, s; - int rs = stbi__jpeg_huff_decode( - j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh - if (rs < 0) - return stbi__err("bad huffman code", "Corrupt JPEG"); - s = rs & 15; - r = rs >> 4; - if (s == 0) { - if (r < 15) { - j->eob_run = (1 << r) - 1; - if (r) - j->eob_run += stbi__jpeg_get_bits(j, r); - r = 64; // force end of block - } else { - // r=15 s=0 should write 16 0s, so we just do - // a run of 15 0s and then write s (which is 0), - // so we don't have to do anything special here - } - } else { - if (s != 1) - return stbi__err("bad huffman code", "Corrupt JPEG"); - // sign bit - if (stbi__jpeg_get_bit(j)) - s = bit; - else - s = -bit; - } - // advance by r - while (k <= j->spec_end) { - short * p = &data[stbi__jpeg_dezigzag[k++]]; - if (*p != 0) { - if (stbi__jpeg_get_bit(j)) - if ((*p & bit) == 0) { - if (*p > 0) - *p += bit; - else - *p -= bit; - } - } else { - if (r == 0) { - *p = (short)s; - break; - } - --r; - } - } - } while (k <= j->spec_end); - } - } - return 1; + // advance by r + while (k <= j->spec_end) { + short *p = &data[stbi__jpeg_dezigzag[k++]]; + if (*p != 0) { + if (stbi__jpeg_get_bit(j)) + if ((*p & bit)==0) { + if (*p > 0) + *p += bit; + else + *p -= bit; + } + } else { + if (r == 0) { + *p = (short) s; + break; + } + --r; + } + } + } while (k <= j->spec_end); + } + } + return 1; } // take a -128..127 value and stbi__clamp it and convert to 0..255 -stbi_inline static stbi_uc stbi__clamp(int x) { - // trick to use a single test to catch both cases - if ((unsigned int)x > 255) { - if (x < 0) - return 0; - if (x > 255) - return 255; - } - return (stbi_uc)x; +stbi_inline static stbi_uc stbi__clamp(int x) +{ + // trick to use a single test to catch both cases + if ((unsigned int) x > 255) { + if (x < 0) return 0; + if (x > 255) return 255; + } + return (stbi_uc) x; } -#define stbi__f2f(x) ((int)(((x)*4096 + 0.5))) -#define stbi__fsh(x) ((x)*4096) +#define stbi__f2f(x) ((int) (((x) * 4096 + 0.5))) +#define stbi__fsh(x) ((x) * 4096) // derived from jidctint -- DCT_ISLOW -#define STBI__IDCT_1D(s0, s1, s2, s3, s4, s5, s6, s7) \ - int t0, t1, t2, t3, p1, p2, p3, p4, p5, x0, x1, x2, x3; \ - p2 = s2; \ - p3 = s6; \ - p1 = (p2 + p3) * stbi__f2f(0.5411961f); \ - t2 = p1 + p3 * stbi__f2f(-1.847759065f); \ - t3 = p1 + p2 * stbi__f2f(0.765366865f); \ - p2 = s0; \ - p3 = s4; \ - t0 = stbi__fsh(p2 + p3); \ - t1 = stbi__fsh(p2 - p3); \ - x0 = t0 + t3; \ - x3 = t0 - t3; \ - x1 = t1 + t2; \ - x2 = t1 - t2; \ - t0 = s7; \ - t1 = s5; \ - t2 = s3; \ - t3 = s1; \ - p3 = t0 + t2; \ - p4 = t1 + t3; \ - p1 = t0 + t3; \ - p2 = t1 + t2; \ - p5 = (p3 + p4) * stbi__f2f(1.175875602f); \ - t0 = t0 * stbi__f2f(0.298631336f); \ - t1 = t1 * stbi__f2f(2.053119869f); \ - t2 = t2 * stbi__f2f(3.072711026f); \ - t3 = t3 * stbi__f2f(1.501321110f); \ - p1 = p5 + p1 * stbi__f2f(-0.899976223f); \ - p2 = p5 + p2 * stbi__f2f(-2.562915447f); \ - p3 = p3 * stbi__f2f(-1.961570560f); \ - p4 = p4 * stbi__f2f(-0.390180644f); \ - t3 += p1 + p4; \ - t2 += p2 + p3; \ - t1 += p2 + p4; \ - t0 += p1 + p3; +#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \ + int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \ + p2 = s2; \ + p3 = s6; \ + p1 = (p2+p3) * stbi__f2f(0.5411961f); \ + t2 = p1 + p3*stbi__f2f(-1.847759065f); \ + t3 = p1 + p2*stbi__f2f( 0.765366865f); \ + p2 = s0; \ + p3 = s4; \ + t0 = stbi__fsh(p2+p3); \ + t1 = stbi__fsh(p2-p3); \ + x0 = t0+t3; \ + x3 = t0-t3; \ + x1 = t1+t2; \ + x2 = t1-t2; \ + t0 = s7; \ + t1 = s5; \ + t2 = s3; \ + t3 = s1; \ + p3 = t0+t2; \ + p4 = t1+t3; \ + p1 = t0+t3; \ + p2 = t1+t2; \ + p5 = (p3+p4)*stbi__f2f( 1.175875602f); \ + t0 = t0*stbi__f2f( 0.298631336f); \ + t1 = t1*stbi__f2f( 2.053119869f); \ + t2 = t2*stbi__f2f( 3.072711026f); \ + t3 = t3*stbi__f2f( 1.501321110f); \ + p1 = p5 + p1*stbi__f2f(-0.899976223f); \ + p2 = p5 + p2*stbi__f2f(-2.562915447f); \ + p3 = p3*stbi__f2f(-1.961570560f); \ + p4 = p4*stbi__f2f(-0.390180644f); \ + t3 += p1+p4; \ + t2 += p2+p3; \ + t1 += p2+p4; \ + t0 += p1+p3; -static void stbi__idct_block(stbi_uc * out, int out_stride, short data[64]) { - int i, val[64], *v = val; - stbi_uc * o; - short * d = data; +static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64]) +{ + int i,val[64],*v=val; + stbi_uc *o; + short *d = data; - // columns - for (i = 0; i < 8; ++i, ++d, ++v) { - // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing - if (d[8] == 0 && d[16] == 0 && d[24] == 0 && d[32] == 0 && d[40] == 0 && d[48] == 0 && d[56] == 0) { - // no shortcut 0 seconds - // (1|2|3|4|5|6|7)==0 0 seconds - // all separate -0.047 seconds - // 1 && 2|3 && 4|5 && 6|7: -0.047 seconds - int dcterm = d[0] * 4; - v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm; - } else { - STBI__IDCT_1D(d[0], d[8], d[16], d[24], d[32], d[40], d[48], d[56]) - // constants scaled things up by 1<<12; let's bring them back - // down, but keep 2 extra bits of precision - x0 += 512; - x1 += 512; - x2 += 512; - x3 += 512; - v[0] = (x0 + t3) >> 10; - v[56] = (x0 - t3) >> 10; - v[8] = (x1 + t2) >> 10; - v[48] = (x1 - t2) >> 10; - v[16] = (x2 + t1) >> 10; - v[40] = (x2 - t1) >> 10; - v[24] = (x3 + t0) >> 10; - v[32] = (x3 - t0) >> 10; - } - } + // columns + for (i=0; i < 8; ++i,++d, ++v) { + // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing + if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0 + && d[40]==0 && d[48]==0 && d[56]==0) { + // no shortcut 0 seconds + // (1|2|3|4|5|6|7)==0 0 seconds + // all separate -0.047 seconds + // 1 && 2|3 && 4|5 && 6|7: -0.047 seconds + int dcterm = d[0]*4; + v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm; + } else { + STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56]) + // constants scaled things up by 1<<12; let's bring them back + // down, but keep 2 extra bits of precision + x0 += 512; x1 += 512; x2 += 512; x3 += 512; + v[ 0] = (x0+t3) >> 10; + v[56] = (x0-t3) >> 10; + v[ 8] = (x1+t2) >> 10; + v[48] = (x1-t2) >> 10; + v[16] = (x2+t1) >> 10; + v[40] = (x2-t1) >> 10; + v[24] = (x3+t0) >> 10; + v[32] = (x3-t0) >> 10; + } + } - for (i = 0, v = val, o = out; i < 8; ++i, v += 8, o += out_stride) { - // no fast case since the first 1D IDCT spread components out - STBI__IDCT_1D(v[0], v[1], v[2], v[3], v[4], v[5], v[6], v[7]) - // constants scaled things up by 1<<12, plus we had 1<<2 from first - // loop, plus horizontal and vertical each scale by sqrt(8) so together - // we've got an extra 1<<3, so 1<<17 total we need to remove. - // so we want to round that, which means adding 0.5 * 1<<17, - // aka 65536. Also, we'll end up with -128 to 127 that we want - // to encode as 0..255 by adding 128, so we'll add that before the shift - x0 += 65536 + (128 << 17); - x1 += 65536 + (128 << 17); - x2 += 65536 + (128 << 17); - x3 += 65536 + (128 << 17); - // tried computing the shifts into temps, or'ing the temps to see - // if any were out of range, but that was slower - o[0] = stbi__clamp((x0 + t3) >> 17); - o[7] = stbi__clamp((x0 - t3) >> 17); - o[1] = stbi__clamp((x1 + t2) >> 17); - o[6] = stbi__clamp((x1 - t2) >> 17); - o[2] = stbi__clamp((x2 + t1) >> 17); - o[5] = stbi__clamp((x2 - t1) >> 17); - o[3] = stbi__clamp((x3 + t0) >> 17); - o[4] = stbi__clamp((x3 - t0) >> 17); - } + for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) { + // no fast case since the first 1D IDCT spread components out + STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7]) + // constants scaled things up by 1<<12, plus we had 1<<2 from first + // loop, plus horizontal and vertical each scale by sqrt(8) so together + // we've got an extra 1<<3, so 1<<17 total we need to remove. + // so we want to round that, which means adding 0.5 * 1<<17, + // aka 65536. Also, we'll end up with -128 to 127 that we want + // to encode as 0..255 by adding 128, so we'll add that before the shift + x0 += 65536 + (128<<17); + x1 += 65536 + (128<<17); + x2 += 65536 + (128<<17); + x3 += 65536 + (128<<17); + // tried computing the shifts into temps, or'ing the temps to see + // if any were out of range, but that was slower + o[0] = stbi__clamp((x0+t3) >> 17); + o[7] = stbi__clamp((x0-t3) >> 17); + o[1] = stbi__clamp((x1+t2) >> 17); + o[6] = stbi__clamp((x1-t2) >> 17); + o[2] = stbi__clamp((x2+t1) >> 17); + o[5] = stbi__clamp((x2-t1) >> 17); + o[3] = stbi__clamp((x3+t0) >> 17); + o[4] = stbi__clamp((x3-t0) >> 17); + } } #ifdef STBI_SSE2 // sse2 integer IDCT. not the fastest possible implementation but it // produces bit-identical results to the generic C version so it's // fully "transparent". -static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { - // This is constructed to match our regular (generic) integer IDCT exactly. - __m128i row0, row1, row2, row3, row4, row5, row6, row7; - __m128i tmp; +static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64]) +{ + // This is constructed to match our regular (generic) integer IDCT exactly. + __m128i row0, row1, row2, row3, row4, row5, row6, row7; + __m128i tmp; -// dot product constant: even elems=x, odd elems=y -#define dct_const(x, y) _mm_setr_epi16((x), (y), (x), (y), (x), (y), (x), (y)) + // dot product constant: even elems=x, odd elems=y + #define dct_const(x,y) _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y)) -// out(0) = c0[even]*x + c0[odd]*y (c0, x, y 16-bit, out 32-bit) -// out(1) = c1[even]*x + c1[odd]*y -#define dct_rot(out0, out1, x, y, c0, c1) \ - __m128i c0##lo = _mm_unpacklo_epi16((x), (y)); \ - __m128i c0##hi = _mm_unpackhi_epi16((x), (y)); \ - __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \ - __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \ - __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \ - __m128i out1##_h = _mm_madd_epi16(c0##hi, c1) + // out(0) = c0[even]*x + c0[odd]*y (c0, x, y 16-bit, out 32-bit) + // out(1) = c1[even]*x + c1[odd]*y + #define dct_rot(out0,out1, x,y,c0,c1) \ + __m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \ + __m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \ + __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \ + __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \ + __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \ + __m128i out1##_h = _mm_madd_epi16(c0##hi, c1) -// out = in << 12 (in 16-bit, out 32-bit) -#define dct_widen(out, in) \ - __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \ - __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4) + // out = in << 12 (in 16-bit, out 32-bit) + #define dct_widen(out, in) \ + __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \ + __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4) -// wide add -#define dct_wadd(out, a, b) \ - __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \ - __m128i out##_h = _mm_add_epi32(a##_h, b##_h) + // wide add + #define dct_wadd(out, a, b) \ + __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \ + __m128i out##_h = _mm_add_epi32(a##_h, b##_h) -// wide sub -#define dct_wsub(out, a, b) \ - __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \ - __m128i out##_h = _mm_sub_epi32(a##_h, b##_h) + // wide sub + #define dct_wsub(out, a, b) \ + __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \ + __m128i out##_h = _mm_sub_epi32(a##_h, b##_h) -// butterfly a/b, add bias, then shift by "s" and pack -#define dct_bfly32o(out0, out1, a, b, bias, s) \ - { \ - __m128i abiased_l = _mm_add_epi32(a##_l, bias); \ - __m128i abiased_h = _mm_add_epi32(a##_h, bias); \ - dct_wadd(sum, abiased, b); \ - dct_wsub(dif, abiased, b); \ - out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \ - out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \ - } + // butterfly a/b, add bias, then shift by "s" and pack + #define dct_bfly32o(out0, out1, a,b,bias,s) \ + { \ + __m128i abiased_l = _mm_add_epi32(a##_l, bias); \ + __m128i abiased_h = _mm_add_epi32(a##_h, bias); \ + dct_wadd(sum, abiased, b); \ + dct_wsub(dif, abiased, b); \ + out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \ + out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \ + } -// 8-bit interleave step (for transposes) -#define dct_interleave8(a, b) \ - tmp = a; \ - a = _mm_unpacklo_epi8(a, b); \ - b = _mm_unpackhi_epi8(tmp, b) + // 8-bit interleave step (for transposes) + #define dct_interleave8(a, b) \ + tmp = a; \ + a = _mm_unpacklo_epi8(a, b); \ + b = _mm_unpackhi_epi8(tmp, b) -// 16-bit interleave step (for transposes) -#define dct_interleave16(a, b) \ - tmp = a; \ - a = _mm_unpacklo_epi16(a, b); \ - b = _mm_unpackhi_epi16(tmp, b) + // 16-bit interleave step (for transposes) + #define dct_interleave16(a, b) \ + tmp = a; \ + a = _mm_unpacklo_epi16(a, b); \ + b = _mm_unpackhi_epi16(tmp, b) -#define dct_pass(bias, shift) \ - { \ - /* even part */ \ - dct_rot(t2e, t3e, row2, row6, rot0_0, rot0_1); \ - __m128i sum04 = _mm_add_epi16(row0, row4); \ - __m128i dif04 = _mm_sub_epi16(row0, row4); \ - dct_widen(t0e, sum04); \ - dct_widen(t1e, dif04); \ - dct_wadd(x0, t0e, t3e); \ - dct_wsub(x3, t0e, t3e); \ - dct_wadd(x1, t1e, t2e); \ - dct_wsub(x2, t1e, t2e); \ - /* odd part */ \ - dct_rot(y0o, y2o, row7, row3, rot2_0, rot2_1); \ - dct_rot(y1o, y3o, row5, row1, rot3_0, rot3_1); \ - __m128i sum17 = _mm_add_epi16(row1, row7); \ - __m128i sum35 = _mm_add_epi16(row3, row5); \ - dct_rot(y4o, y5o, sum17, sum35, rot1_0, rot1_1); \ - dct_wadd(x4, y0o, y4o); \ - dct_wadd(x5, y1o, y5o); \ - dct_wadd(x6, y2o, y5o); \ - dct_wadd(x7, y3o, y4o); \ - dct_bfly32o(row0, row7, x0, x7, bias, shift); \ - dct_bfly32o(row1, row6, x1, x6, bias, shift); \ - dct_bfly32o(row2, row5, x2, x5, bias, shift); \ - dct_bfly32o(row3, row4, x3, x4, bias, shift); \ - } + #define dct_pass(bias,shift) \ + { \ + /* even part */ \ + dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \ + __m128i sum04 = _mm_add_epi16(row0, row4); \ + __m128i dif04 = _mm_sub_epi16(row0, row4); \ + dct_widen(t0e, sum04); \ + dct_widen(t1e, dif04); \ + dct_wadd(x0, t0e, t3e); \ + dct_wsub(x3, t0e, t3e); \ + dct_wadd(x1, t1e, t2e); \ + dct_wsub(x2, t1e, t2e); \ + /* odd part */ \ + dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \ + dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \ + __m128i sum17 = _mm_add_epi16(row1, row7); \ + __m128i sum35 = _mm_add_epi16(row3, row5); \ + dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \ + dct_wadd(x4, y0o, y4o); \ + dct_wadd(x5, y1o, y5o); \ + dct_wadd(x6, y2o, y5o); \ + dct_wadd(x7, y3o, y4o); \ + dct_bfly32o(row0,row7, x0,x7,bias,shift); \ + dct_bfly32o(row1,row6, x1,x6,bias,shift); \ + dct_bfly32o(row2,row5, x2,x5,bias,shift); \ + dct_bfly32o(row3,row4, x3,x4,bias,shift); \ + } - __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f)); - __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f(0.765366865f), stbi__f2f(0.5411961f)); - __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f)); - __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f)); - __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f(0.298631336f), stbi__f2f(-1.961570560f)); - __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f(3.072711026f)); - __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f(2.053119869f), stbi__f2f(-0.390180644f)); - __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f(1.501321110f)); + __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f)); + __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f)); + __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f)); + __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f)); + __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f)); + __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f)); + __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f)); + __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f)); - // rounding biases in column/row passes, see stbi__idct_block for explanation. - __m128i bias_0 = _mm_set1_epi32(512); - __m128i bias_1 = _mm_set1_epi32(65536 + (128 << 17)); + // rounding biases in column/row passes, see stbi__idct_block for explanation. + __m128i bias_0 = _mm_set1_epi32(512); + __m128i bias_1 = _mm_set1_epi32(65536 + (128<<17)); - // load - row0 = _mm_load_si128((const __m128i *)(data + 0 * 8)); - row1 = _mm_load_si128((const __m128i *)(data + 1 * 8)); - row2 = _mm_load_si128((const __m128i *)(data + 2 * 8)); - row3 = _mm_load_si128((const __m128i *)(data + 3 * 8)); - row4 = _mm_load_si128((const __m128i *)(data + 4 * 8)); - row5 = _mm_load_si128((const __m128i *)(data + 5 * 8)); - row6 = _mm_load_si128((const __m128i *)(data + 6 * 8)); - row7 = _mm_load_si128((const __m128i *)(data + 7 * 8)); + // load + row0 = _mm_load_si128((const __m128i *) (data + 0*8)); + row1 = _mm_load_si128((const __m128i *) (data + 1*8)); + row2 = _mm_load_si128((const __m128i *) (data + 2*8)); + row3 = _mm_load_si128((const __m128i *) (data + 3*8)); + row4 = _mm_load_si128((const __m128i *) (data + 4*8)); + row5 = _mm_load_si128((const __m128i *) (data + 5*8)); + row6 = _mm_load_si128((const __m128i *) (data + 6*8)); + row7 = _mm_load_si128((const __m128i *) (data + 7*8)); - // column pass - dct_pass(bias_0, 10); + // column pass + dct_pass(bias_0, 10); - { - // 16bit 8x8 transpose pass 1 - dct_interleave16(row0, row4); - dct_interleave16(row1, row5); - dct_interleave16(row2, row6); - dct_interleave16(row3, row7); + { + // 16bit 8x8 transpose pass 1 + dct_interleave16(row0, row4); + dct_interleave16(row1, row5); + dct_interleave16(row2, row6); + dct_interleave16(row3, row7); - // transpose pass 2 - dct_interleave16(row0, row2); - dct_interleave16(row1, row3); - dct_interleave16(row4, row6); - dct_interleave16(row5, row7); + // transpose pass 2 + dct_interleave16(row0, row2); + dct_interleave16(row1, row3); + dct_interleave16(row4, row6); + dct_interleave16(row5, row7); - // transpose pass 3 - dct_interleave16(row0, row1); - dct_interleave16(row2, row3); - dct_interleave16(row4, row5); - dct_interleave16(row6, row7); - } + // transpose pass 3 + dct_interleave16(row0, row1); + dct_interleave16(row2, row3); + dct_interleave16(row4, row5); + dct_interleave16(row6, row7); + } - // row pass - dct_pass(bias_1, 17); + // row pass + dct_pass(bias_1, 17); - { - // pack - __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7 - __m128i p1 = _mm_packus_epi16(row2, row3); - __m128i p2 = _mm_packus_epi16(row4, row5); - __m128i p3 = _mm_packus_epi16(row6, row7); + { + // pack + __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7 + __m128i p1 = _mm_packus_epi16(row2, row3); + __m128i p2 = _mm_packus_epi16(row4, row5); + __m128i p3 = _mm_packus_epi16(row6, row7); - // 8bit 8x8 transpose pass 1 - dct_interleave8(p0, p2); // a0e0a1e1... - dct_interleave8(p1, p3); // c0g0c1g1... + // 8bit 8x8 transpose pass 1 + dct_interleave8(p0, p2); // a0e0a1e1... + dct_interleave8(p1, p3); // c0g0c1g1... - // transpose pass 2 - dct_interleave8(p0, p1); // a0c0e0g0... - dct_interleave8(p2, p3); // b0d0f0h0... + // transpose pass 2 + dct_interleave8(p0, p1); // a0c0e0g0... + dct_interleave8(p2, p3); // b0d0f0h0... - // transpose pass 3 - dct_interleave8(p0, p2); // a0b0c0d0... - dct_interleave8(p1, p3); // a4b4c4d4... + // transpose pass 3 + dct_interleave8(p0, p2); // a0b0c0d0... + dct_interleave8(p1, p3); // a4b4c4d4... - // store - _mm_storel_epi64((__m128i *)out, p0); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p0, 0x4e)); - out += out_stride; - _mm_storel_epi64((__m128i *)out, p2); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p2, 0x4e)); - out += out_stride; - _mm_storel_epi64((__m128i *)out, p1); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p1, 0x4e)); - out += out_stride; - _mm_storel_epi64((__m128i *)out, p3); - out += out_stride; - _mm_storel_epi64((__m128i *)out, _mm_shuffle_epi32(p3, 0x4e)); - } + // store + _mm_storel_epi64((__m128i *) out, p0); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p2); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p1); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride; + _mm_storel_epi64((__m128i *) out, p3); out += out_stride; + _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e)); + } #undef dct_const #undef dct_rot @@ -2763,235 +2708,198 @@ static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { // NEON integer IDCT. should produce bit-identical // results to the generic C version. -static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { - int16x8_t row0, row1, row2, row3, row4, row5, row6, row7; +static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64]) +{ + int16x8_t row0, row1, row2, row3, row4, row5, row6, row7; - int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f)); - int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f)); - int16x4_t rot0_2 = vdup_n_s16(stbi__f2f(0.765366865f)); - int16x4_t rot1_0 = vdup_n_s16(stbi__f2f(1.175875602f)); - int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f)); - int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f)); - int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f)); - int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f)); - int16x4_t rot3_0 = vdup_n_s16(stbi__f2f(0.298631336f)); - int16x4_t rot3_1 = vdup_n_s16(stbi__f2f(2.053119869f)); - int16x4_t rot3_2 = vdup_n_s16(stbi__f2f(3.072711026f)); - int16x4_t rot3_3 = vdup_n_s16(stbi__f2f(1.501321110f)); + int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f)); + int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f)); + int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f)); + int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f)); + int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f)); + int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f)); + int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f)); + int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f)); + int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f)); + int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f)); + int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f)); + int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f)); -#define dct_long_mul(out, inq, coeff) \ - int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \ - int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff) +#define dct_long_mul(out, inq, coeff) \ + int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \ + int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff) -#define dct_long_mac(out, acc, inq, coeff) \ - int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \ - int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff) +#define dct_long_mac(out, acc, inq, coeff) \ + int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \ + int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff) -#define dct_widen(out, inq) \ - int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \ - int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12) +#define dct_widen(out, inq) \ + int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \ + int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12) // wide add -#define dct_wadd(out, a, b) \ - int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \ - int32x4_t out##_h = vaddq_s32(a##_h, b##_h) +#define dct_wadd(out, a, b) \ + int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \ + int32x4_t out##_h = vaddq_s32(a##_h, b##_h) // wide sub -#define dct_wsub(out, a, b) \ - int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \ - int32x4_t out##_h = vsubq_s32(a##_h, b##_h) +#define dct_wsub(out, a, b) \ + int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \ + int32x4_t out##_h = vsubq_s32(a##_h, b##_h) // butterfly a/b, then shift using "shiftop" by "s" and pack -#define dct_bfly32o(out0, out1, a, b, shiftop, s) \ - { \ - dct_wadd(sum, a, b); \ - dct_wsub(dif, a, b); \ - out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \ - out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \ - } +#define dct_bfly32o(out0,out1, a,b,shiftop,s) \ + { \ + dct_wadd(sum, a, b); \ + dct_wsub(dif, a, b); \ + out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \ + out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \ + } -#define dct_pass(shiftop, shift) \ - { \ - /* even part */ \ - int16x8_t sum26 = vaddq_s16(row2, row6); \ - dct_long_mul(p1e, sum26, rot0_0); \ - dct_long_mac(t2e, p1e, row6, rot0_1); \ - dct_long_mac(t3e, p1e, row2, rot0_2); \ - int16x8_t sum04 = vaddq_s16(row0, row4); \ - int16x8_t dif04 = vsubq_s16(row0, row4); \ - dct_widen(t0e, sum04); \ - dct_widen(t1e, dif04); \ - dct_wadd(x0, t0e, t3e); \ - dct_wsub(x3, t0e, t3e); \ - dct_wadd(x1, t1e, t2e); \ - dct_wsub(x2, t1e, t2e); \ - /* odd part */ \ - int16x8_t sum15 = vaddq_s16(row1, row5); \ - int16x8_t sum17 = vaddq_s16(row1, row7); \ - int16x8_t sum35 = vaddq_s16(row3, row5); \ - int16x8_t sum37 = vaddq_s16(row3, row7); \ - int16x8_t sumodd = vaddq_s16(sum17, sum35); \ - dct_long_mul(p5o, sumodd, rot1_0); \ - dct_long_mac(p1o, p5o, sum17, rot1_1); \ - dct_long_mac(p2o, p5o, sum35, rot1_2); \ - dct_long_mul(p3o, sum37, rot2_0); \ - dct_long_mul(p4o, sum15, rot2_1); \ - dct_wadd(sump13o, p1o, p3o); \ - dct_wadd(sump24o, p2o, p4o); \ - dct_wadd(sump23o, p2o, p3o); \ - dct_wadd(sump14o, p1o, p4o); \ - dct_long_mac(x4, sump13o, row7, rot3_0); \ - dct_long_mac(x5, sump24o, row5, rot3_1); \ - dct_long_mac(x6, sump23o, row3, rot3_2); \ - dct_long_mac(x7, sump14o, row1, rot3_3); \ - dct_bfly32o(row0, row7, x0, x7, shiftop, shift); \ - dct_bfly32o(row1, row6, x1, x6, shiftop, shift); \ - dct_bfly32o(row2, row5, x2, x5, shiftop, shift); \ - dct_bfly32o(row3, row4, x3, x4, shiftop, shift); \ - } +#define dct_pass(shiftop, shift) \ + { \ + /* even part */ \ + int16x8_t sum26 = vaddq_s16(row2, row6); \ + dct_long_mul(p1e, sum26, rot0_0); \ + dct_long_mac(t2e, p1e, row6, rot0_1); \ + dct_long_mac(t3e, p1e, row2, rot0_2); \ + int16x8_t sum04 = vaddq_s16(row0, row4); \ + int16x8_t dif04 = vsubq_s16(row0, row4); \ + dct_widen(t0e, sum04); \ + dct_widen(t1e, dif04); \ + dct_wadd(x0, t0e, t3e); \ + dct_wsub(x3, t0e, t3e); \ + dct_wadd(x1, t1e, t2e); \ + dct_wsub(x2, t1e, t2e); \ + /* odd part */ \ + int16x8_t sum15 = vaddq_s16(row1, row5); \ + int16x8_t sum17 = vaddq_s16(row1, row7); \ + int16x8_t sum35 = vaddq_s16(row3, row5); \ + int16x8_t sum37 = vaddq_s16(row3, row7); \ + int16x8_t sumodd = vaddq_s16(sum17, sum35); \ + dct_long_mul(p5o, sumodd, rot1_0); \ + dct_long_mac(p1o, p5o, sum17, rot1_1); \ + dct_long_mac(p2o, p5o, sum35, rot1_2); \ + dct_long_mul(p3o, sum37, rot2_0); \ + dct_long_mul(p4o, sum15, rot2_1); \ + dct_wadd(sump13o, p1o, p3o); \ + dct_wadd(sump24o, p2o, p4o); \ + dct_wadd(sump23o, p2o, p3o); \ + dct_wadd(sump14o, p1o, p4o); \ + dct_long_mac(x4, sump13o, row7, rot3_0); \ + dct_long_mac(x5, sump24o, row5, rot3_1); \ + dct_long_mac(x6, sump23o, row3, rot3_2); \ + dct_long_mac(x7, sump14o, row1, rot3_3); \ + dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \ + dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \ + dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \ + dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \ + } - // load - row0 = vld1q_s16(data + 0 * 8); - row1 = vld1q_s16(data + 1 * 8); - row2 = vld1q_s16(data + 2 * 8); - row3 = vld1q_s16(data + 3 * 8); - row4 = vld1q_s16(data + 4 * 8); - row5 = vld1q_s16(data + 5 * 8); - row6 = vld1q_s16(data + 6 * 8); - row7 = vld1q_s16(data + 7 * 8); + // load + row0 = vld1q_s16(data + 0*8); + row1 = vld1q_s16(data + 1*8); + row2 = vld1q_s16(data + 2*8); + row3 = vld1q_s16(data + 3*8); + row4 = vld1q_s16(data + 4*8); + row5 = vld1q_s16(data + 5*8); + row6 = vld1q_s16(data + 6*8); + row7 = vld1q_s16(data + 7*8); - // add DC bias - row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0)); + // add DC bias + row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0)); - // column pass - dct_pass(vrshrn_n_s32, 10); + // column pass + dct_pass(vrshrn_n_s32, 10); - // 16bit 8x8 transpose - { + // 16bit 8x8 transpose + { // these three map to a single VTRN.16, VTRN.32, and VSWP, respectively. // whether compilers actually get this is another story, sadly. -#define dct_trn16(x, y) \ - { \ - int16x8x2_t t = vtrnq_s16(x, y); \ - x = t.val[0]; \ - y = t.val[1]; \ - } -#define dct_trn32(x, y) \ - { \ - int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); \ - x = vreinterpretq_s16_s32(t.val[0]); \ - y = vreinterpretq_s16_s32(t.val[1]); \ - } -#define dct_trn64(x, y) \ - { \ - int16x8_t x0 = x; \ - int16x8_t y0 = y; \ - x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); \ - y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); \ - } +#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; } +#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); } +#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); } - // pass 1 - dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6 - dct_trn16(row2, row3); - dct_trn16(row4, row5); - dct_trn16(row6, row7); + // pass 1 + dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6 + dct_trn16(row2, row3); + dct_trn16(row4, row5); + dct_trn16(row6, row7); - // pass 2 - dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4 - dct_trn32(row1, row3); - dct_trn32(row4, row6); - dct_trn32(row5, row7); + // pass 2 + dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4 + dct_trn32(row1, row3); + dct_trn32(row4, row6); + dct_trn32(row5, row7); - // pass 3 - dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0 - dct_trn64(row1, row5); - dct_trn64(row2, row6); - dct_trn64(row3, row7); + // pass 3 + dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0 + dct_trn64(row1, row5); + dct_trn64(row2, row6); + dct_trn64(row3, row7); #undef dct_trn16 #undef dct_trn32 #undef dct_trn64 - } + } - // row pass - // vrshrn_n_s32 only supports shifts up to 16, we need - // 17. so do a non-rounding shift of 16 first then follow - // up with a rounding shift by 1. - dct_pass(vshrn_n_s32, 16); + // row pass + // vrshrn_n_s32 only supports shifts up to 16, we need + // 17. so do a non-rounding shift of 16 first then follow + // up with a rounding shift by 1. + dct_pass(vshrn_n_s32, 16); - { - // pack and round - uint8x8_t p0 = vqrshrun_n_s16(row0, 1); - uint8x8_t p1 = vqrshrun_n_s16(row1, 1); - uint8x8_t p2 = vqrshrun_n_s16(row2, 1); - uint8x8_t p3 = vqrshrun_n_s16(row3, 1); - uint8x8_t p4 = vqrshrun_n_s16(row4, 1); - uint8x8_t p5 = vqrshrun_n_s16(row5, 1); - uint8x8_t p6 = vqrshrun_n_s16(row6, 1); - uint8x8_t p7 = vqrshrun_n_s16(row7, 1); + { + // pack and round + uint8x8_t p0 = vqrshrun_n_s16(row0, 1); + uint8x8_t p1 = vqrshrun_n_s16(row1, 1); + uint8x8_t p2 = vqrshrun_n_s16(row2, 1); + uint8x8_t p3 = vqrshrun_n_s16(row3, 1); + uint8x8_t p4 = vqrshrun_n_s16(row4, 1); + uint8x8_t p5 = vqrshrun_n_s16(row5, 1); + uint8x8_t p6 = vqrshrun_n_s16(row6, 1); + uint8x8_t p7 = vqrshrun_n_s16(row7, 1); - // again, these can translate into one instruction, but often don't. -#define dct_trn8_8(x, y) \ - { \ - uint8x8x2_t t = vtrn_u8(x, y); \ - x = t.val[0]; \ - y = t.val[1]; \ - } -#define dct_trn8_16(x, y) \ - { \ - uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); \ - x = vreinterpret_u8_u16(t.val[0]); \ - y = vreinterpret_u8_u16(t.val[1]); \ - } -#define dct_trn8_32(x, y) \ - { \ - uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); \ - x = vreinterpret_u8_u32(t.val[0]); \ - y = vreinterpret_u8_u32(t.val[1]); \ - } + // again, these can translate into one instruction, but often don't. +#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; } +#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); } +#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); } - // sadly can't use interleaved stores here since we only write - // 8 bytes to each scan line! + // sadly can't use interleaved stores here since we only write + // 8 bytes to each scan line! - // 8x8 8-bit transpose pass 1 - dct_trn8_8(p0, p1); - dct_trn8_8(p2, p3); - dct_trn8_8(p4, p5); - dct_trn8_8(p6, p7); + // 8x8 8-bit transpose pass 1 + dct_trn8_8(p0, p1); + dct_trn8_8(p2, p3); + dct_trn8_8(p4, p5); + dct_trn8_8(p6, p7); - // pass 2 - dct_trn8_16(p0, p2); - dct_trn8_16(p1, p3); - dct_trn8_16(p4, p6); - dct_trn8_16(p5, p7); + // pass 2 + dct_trn8_16(p0, p2); + dct_trn8_16(p1, p3); + dct_trn8_16(p4, p6); + dct_trn8_16(p5, p7); - // pass 3 - dct_trn8_32(p0, p4); - dct_trn8_32(p1, p5); - dct_trn8_32(p2, p6); - dct_trn8_32(p3, p7); + // pass 3 + dct_trn8_32(p0, p4); + dct_trn8_32(p1, p5); + dct_trn8_32(p2, p6); + dct_trn8_32(p3, p7); - // store - vst1_u8(out, p0); - out += out_stride; - vst1_u8(out, p1); - out += out_stride; - vst1_u8(out, p2); - out += out_stride; - vst1_u8(out, p3); - out += out_stride; - vst1_u8(out, p4); - out += out_stride; - vst1_u8(out, p5); - out += out_stride; - vst1_u8(out, p6); - out += out_stride; - vst1_u8(out, p7); + // store + vst1_u8(out, p0); out += out_stride; + vst1_u8(out, p1); out += out_stride; + vst1_u8(out, p2); out += out_stride; + vst1_u8(out, p3); out += out_stride; + vst1_u8(out, p4); out += out_stride; + vst1_u8(out, p5); out += out_stride; + vst1_u8(out, p6); out += out_stride; + vst1_u8(out, p7); #undef dct_trn8_8 #undef dct_trn8_16 #undef dct_trn8_32 - } + } #undef dct_long_mul #undef dct_long_mac @@ -3004,1267 +2912,1169 @@ static void stbi__idct_simd(stbi_uc * out, int out_stride, short data[64]) { #endif // STBI_NEON -#define STBI__MARKER_none 0xff +#define STBI__MARKER_none 0xff // if there's a pending marker from the entropy stream, return that // otherwise, fetch from the stream and get a marker. if there's no // marker, return 0xff, which is never a valid marker value -static stbi_uc stbi__get_marker(stbi__jpeg * j) { - stbi_uc x; - if (j->marker != STBI__MARKER_none) { - x = j->marker; - j->marker = STBI__MARKER_none; - return x; - } - x = stbi__get8(j->s); - if (x != 0xff) - return STBI__MARKER_none; - while (x == 0xff) - x = stbi__get8(j->s); // consume repeated 0xff fill bytes - return x; +static stbi_uc stbi__get_marker(stbi__jpeg *j) +{ + stbi_uc x; + if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; } + x = stbi__get8(j->s); + if (x != 0xff) return STBI__MARKER_none; + while (x == 0xff) + x = stbi__get8(j->s); // consume repeated 0xff fill bytes + return x; } // in each scan, we'll have scan_n components, and the order // of the components is specified by order[] -#define STBI__RESTART(x) ((x) >= 0xd0 && (x) <= 0xd7) +#define STBI__RESTART(x) ((x) >= 0xd0 && (x) <= 0xd7) // after a restart interval, stbi__jpeg_reset the entropy decoder and // the dc prediction -static void stbi__jpeg_reset(stbi__jpeg * j) { - j->code_bits = 0; - j->code_buffer = 0; - j->nomore = 0; - j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0; - j->marker = STBI__MARKER_none; - j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff; - j->eob_run = 0; - // no more than 1<<31 MCUs if no restart_interal? that's plenty safe, - // since we don't even allow 1<<30 pixels +static void stbi__jpeg_reset(stbi__jpeg *j) +{ + j->code_bits = 0; + j->code_buffer = 0; + j->nomore = 0; + j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0; + j->marker = STBI__MARKER_none; + j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff; + j->eob_run = 0; + // no more than 1<<31 MCUs if no restart_interal? that's plenty safe, + // since we don't even allow 1<<30 pixels } -static int stbi__parse_entropy_coded_data(stbi__jpeg * z) { - stbi__jpeg_reset(z); - if (!z->progressive) { - if (z->scan_n == 1) { - int i, j; - STBI_SIMD_ALIGN(short, data[64]); - int n = z->order[0]; - // non-interleaved data, we just need to process one block at a time, - // in trivial scanline order - // number of blocks to do just depends on how many actual "pixels" this - // component has, independent of interleaved MCU blocking and such - int w = (z->img_comp[n].x + 7) >> 3; - int h = (z->img_comp[n].y + 7) >> 3; - for (j = 0; j < h; ++j) { - for (i = 0; i < w; ++i) { - int ha = z->img_comp[n].ha; - if (!stbi__jpeg_decode_block(z, data, z->huff_dc + z->img_comp[n].hd, z->huff_ac + ha, z->fast_ac[ha], n, - z->dequant[z->img_comp[n].tq])) - return 0; - z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * j * 8 + i * 8, z->img_comp[n].w2, data); - // every data block is an MCU, so countdown the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - // if it's NOT a restart, then just bail, so we get corrupt data - // rather than no data - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } +static int stbi__parse_entropy_coded_data(stbi__jpeg *z) +{ + stbi__jpeg_reset(z); + if (!z->progressive) { + if (z->scan_n == 1) { + int i,j; + STBI_SIMD_ALIGN(short, data[64]); + int n = z->order[0]; + // non-interleaved data, we just need to process one block at a time, + // in trivial scanline order + // number of blocks to do just depends on how many actual "pixels" this + // component has, independent of interleaved MCU blocking and such + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + int ha = z->img_comp[n].ha; + if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0; + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data); + // every data block is an MCU, so countdown the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + // if it's NOT a restart, then just bail, so we get corrupt data + // rather than no data + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } } - return 1; - } else { // interleaved - int i, j, k, x, y; - STBI_SIMD_ALIGN(short, data[64]); - for (j = 0; j < z->img_mcu_y; ++j) { - for (i = 0; i < z->img_mcu_x; ++i) { - // scan an interleaved mcu... process scan_n components in order - for (k = 0; k < z->scan_n; ++k) { - int n = z->order[k]; - // scan out an mcu's worth of this component; that's just determined - // by the basic H and V specified for the component - for (y = 0; y < z->img_comp[n].v; ++y) { - for (x = 0; x < z->img_comp[n].h; ++x) { - int x2 = (i * z->img_comp[n].h + x) * 8; - int y2 = (j * z->img_comp[n].v + y) * 8; - int ha = z->img_comp[n].ha; - if (!stbi__jpeg_decode_block(z, data, z->huff_dc + z->img_comp[n].hd, z->huff_ac + ha, - z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) - return 0; - z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * y2 + x2, z->img_comp[n].w2, - data); - } - } - } - // after all interleaved components, that's an interleaved MCU, - // so now count down the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } - } - return 1; - } - } else { - if (z->scan_n == 1) { - int i, j; - int n = z->order[0]; - // non-interleaved data, we just need to process one block at a time, - // in trivial scanline order - // number of blocks to do just depends on how many actual "pixels" this - // component has, independent of interleaved MCU blocking and such - int w = (z->img_comp[n].x + 7) >> 3; - int h = (z->img_comp[n].y + 7) >> 3; - for (j = 0; j < h; ++j) { - for (i = 0; i < w; ++i) { - short * data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); - if (z->spec_start == 0) { - if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) - return 0; - } else { + } + return 1; + } else { // interleaved + int i,j,k,x,y; + STBI_SIMD_ALIGN(short, data[64]); + for (j=0; j < z->img_mcu_y; ++j) { + for (i=0; i < z->img_mcu_x; ++i) { + // scan an interleaved mcu... process scan_n components in order + for (k=0; k < z->scan_n; ++k) { + int n = z->order[k]; + // scan out an mcu's worth of this component; that's just determined + // by the basic H and V specified for the component + for (y=0; y < z->img_comp[n].v; ++y) { + for (x=0; x < z->img_comp[n].h; ++x) { + int x2 = (i*z->img_comp[n].h + x)*8; + int y2 = (j*z->img_comp[n].v + y)*8; int ha = z->img_comp[n].ha; - if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha])) - return 0; - } - // every data block is an MCU, so countdown the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } + if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0; + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data); + } + } + } + // after all interleaved components, that's an interleaved MCU, + // so now count down the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } } - return 1; - } else { // interleaved - int i, j, k, x, y; - for (j = 0; j < z->img_mcu_y; ++j) { - for (i = 0; i < z->img_mcu_x; ++i) { - // scan an interleaved mcu... process scan_n components in order - for (k = 0; k < z->scan_n; ++k) { - int n = z->order[k]; - // scan out an mcu's worth of this component; that's just determined - // by the basic H and V specified for the component - for (y = 0; y < z->img_comp[n].v; ++y) { - for (x = 0; x < z->img_comp[n].h; ++x) { - int x2 = (i * z->img_comp[n].h + x); - int y2 = (j * z->img_comp[n].v + y); - short * data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w); - if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) - return 0; - } - } - } - // after all interleaved components, that's an interleaved MCU, - // so now count down the restart interval - if (--z->todo <= 0) { - if (z->code_bits < 24) - stbi__grow_buffer_unsafe(z); - if (!STBI__RESTART(z->marker)) - return 1; - stbi__jpeg_reset(z); - } - } + } + return 1; + } + } else { + if (z->scan_n == 1) { + int i,j; + int n = z->order[0]; + // non-interleaved data, we just need to process one block at a time, + // in trivial scanline order + // number of blocks to do just depends on how many actual "pixels" this + // component has, independent of interleaved MCU blocking and such + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); + if (z->spec_start == 0) { + if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) + return 0; + } else { + int ha = z->img_comp[n].ha; + if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha])) + return 0; + } + // every data block is an MCU, so countdown the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } } - return 1; - } - } + } + return 1; + } else { // interleaved + int i,j,k,x,y; + for (j=0; j < z->img_mcu_y; ++j) { + for (i=0; i < z->img_mcu_x; ++i) { + // scan an interleaved mcu... process scan_n components in order + for (k=0; k < z->scan_n; ++k) { + int n = z->order[k]; + // scan out an mcu's worth of this component; that's just determined + // by the basic H and V specified for the component + for (y=0; y < z->img_comp[n].v; ++y) { + for (x=0; x < z->img_comp[n].h; ++x) { + int x2 = (i*z->img_comp[n].h + x); + int y2 = (j*z->img_comp[n].v + y); + short *data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w); + if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n)) + return 0; + } + } + } + // after all interleaved components, that's an interleaved MCU, + // so now count down the restart interval + if (--z->todo <= 0) { + if (z->code_bits < 24) stbi__grow_buffer_unsafe(z); + if (!STBI__RESTART(z->marker)) return 1; + stbi__jpeg_reset(z); + } + } + } + return 1; + } + } } -static void stbi__jpeg_dequantize(short * data, stbi__uint16 * dequant) { - int i; - for (i = 0; i < 64; ++i) - data[i] *= dequant[i]; +static void stbi__jpeg_dequantize(short *data, stbi__uint16 *dequant) +{ + int i; + for (i=0; i < 64; ++i) + data[i] *= dequant[i]; } -static void stbi__jpeg_finish(stbi__jpeg * z) { - if (z->progressive) { - // dequantize and idct the data - int i, j, n; - for (n = 0; n < z->s->img_n; ++n) { - int w = (z->img_comp[n].x + 7) >> 3; - int h = (z->img_comp[n].y + 7) >> 3; - for (j = 0; j < h; ++j) { - for (i = 0; i < w; ++i) { - short * data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); - stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]); - z->idct_block_kernel(z->img_comp[n].data + z->img_comp[n].w2 * j * 8 + i * 8, z->img_comp[n].w2, data); - } +static void stbi__jpeg_finish(stbi__jpeg *z) +{ + if (z->progressive) { + // dequantize and idct the data + int i,j,n; + for (n=0; n < z->s->img_n; ++n) { + int w = (z->img_comp[n].x+7) >> 3; + int h = (z->img_comp[n].y+7) >> 3; + for (j=0; j < h; ++j) { + for (i=0; i < w; ++i) { + short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w); + stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]); + z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data); } - } - } + } + } + } } -static int stbi__process_marker(stbi__jpeg * z, int m) { - int L; - switch (m) { - case STBI__MARKER_none: // no marker found - return stbi__err("expected marker", "Corrupt JPEG"); +static int stbi__process_marker(stbi__jpeg *z, int m) +{ + int L; + switch (m) { + case STBI__MARKER_none: // no marker found + return stbi__err("expected marker","Corrupt JPEG"); - case 0xDD: // DRI - specify restart interval - if (stbi__get16be(z->s) != 4) - return stbi__err("bad DRI len", "Corrupt JPEG"); - z->restart_interval = stbi__get16be(z->s); - return 1; + case 0xDD: // DRI - specify restart interval + if (stbi__get16be(z->s) != 4) return stbi__err("bad DRI len","Corrupt JPEG"); + z->restart_interval = stbi__get16be(z->s); + return 1; - case 0xDB: // DQT - define quantization table - L = stbi__get16be(z->s) - 2; - while (L > 0) { + case 0xDB: // DQT - define quantization table + L = stbi__get16be(z->s)-2; + while (L > 0) { int q = stbi__get8(z->s); int p = q >> 4, sixteen = (p != 0); - int t = q & 15, i; - if (p != 0 && p != 1) - return stbi__err("bad DQT type", "Corrupt JPEG"); - if (t > 3) - return stbi__err("bad DQT table", "Corrupt JPEG"); + int t = q & 15,i; + if (p != 0 && p != 1) return stbi__err("bad DQT type","Corrupt JPEG"); + if (t > 3) return stbi__err("bad DQT table","Corrupt JPEG"); - for (i = 0; i < 64; ++i) - z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s)); + for (i=0; i < 64; ++i) + z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s)); L -= (sixteen ? 129 : 65); - } - return L == 0; + } + return L==0; - case 0xC4: // DHT - define huffman table - L = stbi__get16be(z->s) - 2; - while (L > 0) { - stbi_uc * v; - int sizes[16], i, n = 0; + case 0xC4: // DHT - define huffman table + L = stbi__get16be(z->s)-2; + while (L > 0) { + stbi_uc *v; + int sizes[16],i,n=0; int q = stbi__get8(z->s); int tc = q >> 4; int th = q & 15; - if (tc > 1 || th > 3) - return stbi__err("bad DHT header", "Corrupt JPEG"); - for (i = 0; i < 16; ++i) { - sizes[i] = stbi__get8(z->s); - n += sizes[i]; + if (tc > 1 || th > 3) return stbi__err("bad DHT header","Corrupt JPEG"); + for (i=0; i < 16; ++i) { + sizes[i] = stbi__get8(z->s); + n += sizes[i]; } - if (n > 256) - return stbi__err("bad DHT header", "Corrupt JPEG"); // Loop over i < n would write past end of values! + if(n > 256) return stbi__err("bad DHT header","Corrupt JPEG"); // Loop over i < n would write past end of values! L -= 17; if (tc == 0) { - if (!stbi__build_huffman(z->huff_dc + th, sizes)) - return 0; - v = z->huff_dc[th].values; + if (!stbi__build_huffman(z->huff_dc+th, sizes)) return 0; + v = z->huff_dc[th].values; } else { - if (!stbi__build_huffman(z->huff_ac + th, sizes)) - return 0; - v = z->huff_ac[th].values; + if (!stbi__build_huffman(z->huff_ac+th, sizes)) return 0; + v = z->huff_ac[th].values; } - for (i = 0; i < n; ++i) - v[i] = stbi__get8(z->s); + for (i=0; i < n; ++i) + v[i] = stbi__get8(z->s); if (tc != 0) - stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th); + stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th); L -= n; - } - return L == 0; - } + } + return L==0; + } - // check for comment block or APP blocks - if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) { - L = stbi__get16be(z->s); - if (L < 2) { - if (m == 0xFE) - return stbi__err("bad COM len", "Corrupt JPEG"); - else - return stbi__err("bad APP len", "Corrupt JPEG"); - } - L -= 2; + // check for comment block or APP blocks + if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) { + L = stbi__get16be(z->s); + if (L < 2) { + if (m == 0xFE) + return stbi__err("bad COM len","Corrupt JPEG"); + else + return stbi__err("bad APP len","Corrupt JPEG"); + } + L -= 2; - if (m == 0xE0 && L >= 5) { // JFIF APP0 segment - static const unsigned char tag[5] = {'J', 'F', 'I', 'F', '\0'}; - int ok = 1; - int i; - for (i = 0; i < 5; ++i) - if (stbi__get8(z->s) != tag[i]) - ok = 0; - L -= 5; - if (ok) - z->jfif = 1; - } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment - static const unsigned char tag[6] = {'A', 'd', 'o', 'b', 'e', '\0'}; - int ok = 1; - int i; - for (i = 0; i < 6; ++i) - if (stbi__get8(z->s) != tag[i]) - ok = 0; + if (m == 0xE0 && L >= 5) { // JFIF APP0 segment + static const unsigned char tag[5] = {'J','F','I','F','\0'}; + int ok = 1; + int i; + for (i=0; i < 5; ++i) + if (stbi__get8(z->s) != tag[i]) + ok = 0; + L -= 5; + if (ok) + z->jfif = 1; + } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment + static const unsigned char tag[6] = {'A','d','o','b','e','\0'}; + int ok = 1; + int i; + for (i=0; i < 6; ++i) + if (stbi__get8(z->s) != tag[i]) + ok = 0; + L -= 6; + if (ok) { + stbi__get8(z->s); // version + stbi__get16be(z->s); // flags0 + stbi__get16be(z->s); // flags1 + z->app14_color_transform = stbi__get8(z->s); // color transform L -= 6; - if (ok) { - stbi__get8(z->s); // version - stbi__get16be(z->s); // flags0 - stbi__get16be(z->s); // flags1 - z->app14_color_transform = stbi__get8(z->s); // color transform - L -= 6; - } - } + } + } - stbi__skip(z->s, L); - return 1; - } + stbi__skip(z->s, L); + return 1; + } - return stbi__err("unknown marker", "Corrupt JPEG"); + return stbi__err("unknown marker","Corrupt JPEG"); } // after we see SOS -static int stbi__process_scan_header(stbi__jpeg * z) { - int i; - int Ls = stbi__get16be(z->s); - z->scan_n = stbi__get8(z->s); - if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int)z->s->img_n) - return stbi__err("bad SOS component count", "Corrupt JPEG"); - if (Ls != 6 + 2 * z->scan_n) - return stbi__err("bad SOS len", "Corrupt JPEG"); - for (i = 0; i < z->scan_n; ++i) { - int id = stbi__get8(z->s), which; - int q = stbi__get8(z->s); - for (which = 0; which < z->s->img_n; ++which) - if (z->img_comp[which].id == id) - break; - if (which == z->s->img_n) - return 0; // no match - z->img_comp[which].hd = q >> 4; - if (z->img_comp[which].hd > 3) - return stbi__err("bad DC huff", "Corrupt JPEG"); - z->img_comp[which].ha = q & 15; - if (z->img_comp[which].ha > 3) - return stbi__err("bad AC huff", "Corrupt JPEG"); - z->order[i] = which; - } +static int stbi__process_scan_header(stbi__jpeg *z) +{ + int i; + int Ls = stbi__get16be(z->s); + z->scan_n = stbi__get8(z->s); + if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int) z->s->img_n) return stbi__err("bad SOS component count","Corrupt JPEG"); + if (Ls != 6+2*z->scan_n) return stbi__err("bad SOS len","Corrupt JPEG"); + for (i=0; i < z->scan_n; ++i) { + int id = stbi__get8(z->s), which; + int q = stbi__get8(z->s); + for (which = 0; which < z->s->img_n; ++which) + if (z->img_comp[which].id == id) + break; + if (which == z->s->img_n) return 0; // no match + z->img_comp[which].hd = q >> 4; if (z->img_comp[which].hd > 3) return stbi__err("bad DC huff","Corrupt JPEG"); + z->img_comp[which].ha = q & 15; if (z->img_comp[which].ha > 3) return stbi__err("bad AC huff","Corrupt JPEG"); + z->order[i] = which; + } - { - int aa; - z->spec_start = stbi__get8(z->s); - z->spec_end = stbi__get8(z->s); // should be 63, but might be 0 - aa = stbi__get8(z->s); - z->succ_high = (aa >> 4); - z->succ_low = (aa & 15); - if (z->progressive) { - if (z->spec_start > 63 || z->spec_end > 63 || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13) - return stbi__err("bad SOS", "Corrupt JPEG"); - } else { - if (z->spec_start != 0) - return stbi__err("bad SOS", "Corrupt JPEG"); - if (z->succ_high != 0 || z->succ_low != 0) - return stbi__err("bad SOS", "Corrupt JPEG"); - z->spec_end = 63; - } - } + { + int aa; + z->spec_start = stbi__get8(z->s); + z->spec_end = stbi__get8(z->s); // should be 63, but might be 0 + aa = stbi__get8(z->s); + z->succ_high = (aa >> 4); + z->succ_low = (aa & 15); + if (z->progressive) { + if (z->spec_start > 63 || z->spec_end > 63 || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13) + return stbi__err("bad SOS", "Corrupt JPEG"); + } else { + if (z->spec_start != 0) return stbi__err("bad SOS","Corrupt JPEG"); + if (z->succ_high != 0 || z->succ_low != 0) return stbi__err("bad SOS","Corrupt JPEG"); + z->spec_end = 63; + } + } - return 1; + return 1; } -static int stbi__free_jpeg_components(stbi__jpeg * z, int ncomp, int why) { - int i; - for (i = 0; i < ncomp; ++i) { - if (z->img_comp[i].raw_data) { - STBI_FREE(z->img_comp[i].raw_data); - z->img_comp[i].raw_data = NULL; - z->img_comp[i].data = NULL; - } - if (z->img_comp[i].raw_coeff) { - STBI_FREE(z->img_comp[i].raw_coeff); - z->img_comp[i].raw_coeff = 0; - z->img_comp[i].coeff = 0; - } - if (z->img_comp[i].linebuf) { - STBI_FREE(z->img_comp[i].linebuf); - z->img_comp[i].linebuf = NULL; - } - } - return why; +static int stbi__free_jpeg_components(stbi__jpeg *z, int ncomp, int why) +{ + int i; + for (i=0; i < ncomp; ++i) { + if (z->img_comp[i].raw_data) { + STBI_FREE(z->img_comp[i].raw_data); + z->img_comp[i].raw_data = NULL; + z->img_comp[i].data = NULL; + } + if (z->img_comp[i].raw_coeff) { + STBI_FREE(z->img_comp[i].raw_coeff); + z->img_comp[i].raw_coeff = 0; + z->img_comp[i].coeff = 0; + } + if (z->img_comp[i].linebuf) { + STBI_FREE(z->img_comp[i].linebuf); + z->img_comp[i].linebuf = NULL; + } + } + return why; } -static int stbi__process_frame_header(stbi__jpeg * z, int scan) { - stbi__context * s = z->s; - int Lf, p, i, q, h_max = 1, v_max = 1, c; - Lf = stbi__get16be(s); - if (Lf < 11) - return stbi__err("bad SOF len", "Corrupt JPEG"); // JPEG - p = stbi__get8(s); - if (p != 8) - return stbi__err("only 8-bit", "JPEG format not supported: 8-bit only"); // JPEG baseline - s->img_y = stbi__get16be(s); - if (s->img_y == 0) - return stbi__err("no header height", - "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG - s->img_x = stbi__get16be(s); - if (s->img_x == 0) - return stbi__err("0 width", "Corrupt JPEG"); // JPEG requires - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - c = stbi__get8(s); - if (c != 3 && c != 1 && c != 4) - return stbi__err("bad component count", "Corrupt JPEG"); - s->img_n = c; - for (i = 0; i < c; ++i) { - z->img_comp[i].data = NULL; - z->img_comp[i].linebuf = NULL; - } +static int stbi__process_frame_header(stbi__jpeg *z, int scan) +{ + stbi__context *s = z->s; + int Lf,p,i,q, h_max=1,v_max=1,c; + Lf = stbi__get16be(s); if (Lf < 11) return stbi__err("bad SOF len","Corrupt JPEG"); // JPEG + p = stbi__get8(s); if (p != 8) return stbi__err("only 8-bit","JPEG format not supported: 8-bit only"); // JPEG baseline + s->img_y = stbi__get16be(s); if (s->img_y == 0) return stbi__err("no header height", "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG + s->img_x = stbi__get16be(s); if (s->img_x == 0) return stbi__err("0 width","Corrupt JPEG"); // JPEG requires + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + c = stbi__get8(s); + if (c != 3 && c != 1 && c != 4) return stbi__err("bad component count","Corrupt JPEG"); + s->img_n = c; + for (i=0; i < c; ++i) { + z->img_comp[i].data = NULL; + z->img_comp[i].linebuf = NULL; + } - if (Lf != 8 + 3 * s->img_n) - return stbi__err("bad SOF len", "Corrupt JPEG"); + if (Lf != 8+3*s->img_n) return stbi__err("bad SOF len","Corrupt JPEG"); - z->rgb = 0; - for (i = 0; i < s->img_n; ++i) { - static const unsigned char rgb[3] = {'R', 'G', 'B'}; - z->img_comp[i].id = stbi__get8(s); - if (s->img_n == 3 && z->img_comp[i].id == rgb[i]) - ++z->rgb; - q = stbi__get8(s); - z->img_comp[i].h = (q >> 4); - if (!z->img_comp[i].h || z->img_comp[i].h > 4) - return stbi__err("bad H", "Corrupt JPEG"); - z->img_comp[i].v = q & 15; - if (!z->img_comp[i].v || z->img_comp[i].v > 4) - return stbi__err("bad V", "Corrupt JPEG"); - z->img_comp[i].tq = stbi__get8(s); - if (z->img_comp[i].tq > 3) - return stbi__err("bad TQ", "Corrupt JPEG"); - } + z->rgb = 0; + for (i=0; i < s->img_n; ++i) { + static const unsigned char rgb[3] = { 'R', 'G', 'B' }; + z->img_comp[i].id = stbi__get8(s); + if (s->img_n == 3 && z->img_comp[i].id == rgb[i]) + ++z->rgb; + q = stbi__get8(s); + z->img_comp[i].h = (q >> 4); if (!z->img_comp[i].h || z->img_comp[i].h > 4) return stbi__err("bad H","Corrupt JPEG"); + z->img_comp[i].v = q & 15; if (!z->img_comp[i].v || z->img_comp[i].v > 4) return stbi__err("bad V","Corrupt JPEG"); + z->img_comp[i].tq = stbi__get8(s); if (z->img_comp[i].tq > 3) return stbi__err("bad TQ","Corrupt JPEG"); + } - if (scan != STBI__SCAN_load) - return 1; + if (scan != STBI__SCAN_load) return 1; - if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) - return stbi__err("too large", "Image too large to decode"); + if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) return stbi__err("too large", "Image too large to decode"); - for (i = 0; i < s->img_n; ++i) { - if (z->img_comp[i].h > h_max) - h_max = z->img_comp[i].h; - if (z->img_comp[i].v > v_max) - v_max = z->img_comp[i].v; - } + for (i=0; i < s->img_n; ++i) { + if (z->img_comp[i].h > h_max) h_max = z->img_comp[i].h; + if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v; + } - // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios - // and I've never seen a non-corrupted JPEG file actually use them - for (i = 0; i < s->img_n; ++i) { - if (h_max % z->img_comp[i].h != 0) - return stbi__err("bad H", "Corrupt JPEG"); - if (v_max % z->img_comp[i].v != 0) - return stbi__err("bad V", "Corrupt JPEG"); - } + // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios + // and I've never seen a non-corrupted JPEG file actually use them + for (i=0; i < s->img_n; ++i) { + if (h_max % z->img_comp[i].h != 0) return stbi__err("bad H","Corrupt JPEG"); + if (v_max % z->img_comp[i].v != 0) return stbi__err("bad V","Corrupt JPEG"); + } - // compute interleaved mcu info - z->img_h_max = h_max; - z->img_v_max = v_max; - z->img_mcu_w = h_max * 8; - z->img_mcu_h = v_max * 8; - // these sizes can't be more than 17 bits - z->img_mcu_x = (s->img_x + z->img_mcu_w - 1) / z->img_mcu_w; - z->img_mcu_y = (s->img_y + z->img_mcu_h - 1) / z->img_mcu_h; + // compute interleaved mcu info + z->img_h_max = h_max; + z->img_v_max = v_max; + z->img_mcu_w = h_max * 8; + z->img_mcu_h = v_max * 8; + // these sizes can't be more than 17 bits + z->img_mcu_x = (s->img_x + z->img_mcu_w-1) / z->img_mcu_w; + z->img_mcu_y = (s->img_y + z->img_mcu_h-1) / z->img_mcu_h; - for (i = 0; i < s->img_n; ++i) { - // number of effective pixels (e.g. for non-interleaved MCU) - z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max - 1) / h_max; - z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max - 1) / v_max; - // to simplify generation, we'll allocate enough memory to decode - // the bogus oversized data from using interleaved MCUs and their - // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't - // discard the extra data until colorspace conversion - // - // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier) - // so these muls can't overflow with 32-bit ints (which we require) - z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8; - z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8; - z->img_comp[i].coeff = 0; - z->img_comp[i].raw_coeff = 0; - z->img_comp[i].linebuf = NULL; - z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15); - if (z->img_comp[i].raw_data == NULL) - return stbi__free_jpeg_components(z, i + 1, stbi__err("outofmem", "Out of memory")); - // align blocks for idct using mmx/sse - z->img_comp[i].data = (stbi_uc *)(((size_t)z->img_comp[i].raw_data + 15) & ~15); - if (z->progressive) { - // w2, h2 are multiples of 8 (see above) - z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8; - z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8; - z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15); - if (z->img_comp[i].raw_coeff == NULL) - return stbi__free_jpeg_components(z, i + 1, stbi__err("outofmem", "Out of memory")); - z->img_comp[i].coeff = (short *)(((size_t)z->img_comp[i].raw_coeff + 15) & ~15); - } - } + for (i=0; i < s->img_n; ++i) { + // number of effective pixels (e.g. for non-interleaved MCU) + z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max-1) / h_max; + z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max-1) / v_max; + // to simplify generation, we'll allocate enough memory to decode + // the bogus oversized data from using interleaved MCUs and their + // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't + // discard the extra data until colorspace conversion + // + // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier) + // so these muls can't overflow with 32-bit ints (which we require) + z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8; + z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8; + z->img_comp[i].coeff = 0; + z->img_comp[i].raw_coeff = 0; + z->img_comp[i].linebuf = NULL; + z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15); + if (z->img_comp[i].raw_data == NULL) + return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory")); + // align blocks for idct using mmx/sse + z->img_comp[i].data = (stbi_uc*) (((size_t) z->img_comp[i].raw_data + 15) & ~15); + if (z->progressive) { + // w2, h2 are multiples of 8 (see above) + z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8; + z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8; + z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15); + if (z->img_comp[i].raw_coeff == NULL) + return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory")); + z->img_comp[i].coeff = (short*) (((size_t) z->img_comp[i].raw_coeff + 15) & ~15); + } + } - return 1; + return 1; } // use comparisons since in some cases we handle more than one case (e.g. SOF) -#define stbi__DNL(x) ((x) == 0xdc) -#define stbi__SOI(x) ((x) == 0xd8) -#define stbi__EOI(x) ((x) == 0xd9) -#define stbi__SOF(x) ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2) -#define stbi__SOS(x) ((x) == 0xda) +#define stbi__DNL(x) ((x) == 0xdc) +#define stbi__SOI(x) ((x) == 0xd8) +#define stbi__EOI(x) ((x) == 0xd9) +#define stbi__SOF(x) ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2) +#define stbi__SOS(x) ((x) == 0xda) -#define stbi__SOF_progressive(x) ((x) == 0xc2) +#define stbi__SOF_progressive(x) ((x) == 0xc2) -static int stbi__decode_jpeg_header(stbi__jpeg * z, int scan) { - int m; - z->jfif = 0; - z->app14_color_transform = -1; // valid values are 0,1,2 - z->marker = STBI__MARKER_none; // initialize cached marker to empty - m = stbi__get_marker(z); - if (!stbi__SOI(m)) - return stbi__err("no SOI", "Corrupt JPEG"); - if (scan == STBI__SCAN_type) - return 1; - m = stbi__get_marker(z); - while (!stbi__SOF(m)) { - if (!stbi__process_marker(z, m)) - return 0; - m = stbi__get_marker(z); - while (m == STBI__MARKER_none) { - // some files have extra padding after their blocks, so ok, we'll scan - if (stbi__at_eof(z->s)) - return stbi__err("no SOF", "Corrupt JPEG"); - m = stbi__get_marker(z); - } - } - z->progressive = stbi__SOF_progressive(m); - if (!stbi__process_frame_header(z, scan)) - return 0; - return 1; +static int stbi__decode_jpeg_header(stbi__jpeg *z, int scan) +{ + int m; + z->jfif = 0; + z->app14_color_transform = -1; // valid values are 0,1,2 + z->marker = STBI__MARKER_none; // initialize cached marker to empty + m = stbi__get_marker(z); + if (!stbi__SOI(m)) return stbi__err("no SOI","Corrupt JPEG"); + if (scan == STBI__SCAN_type) return 1; + m = stbi__get_marker(z); + while (!stbi__SOF(m)) { + if (!stbi__process_marker(z,m)) return 0; + m = stbi__get_marker(z); + while (m == STBI__MARKER_none) { + // some files have extra padding after their blocks, so ok, we'll scan + if (stbi__at_eof(z->s)) return stbi__err("no SOF", "Corrupt JPEG"); + m = stbi__get_marker(z); + } + } + z->progressive = stbi__SOF_progressive(m); + if (!stbi__process_frame_header(z, scan)) return 0; + return 1; } -static int stbi__skip_jpeg_junk_at_end(stbi__jpeg * j) { - // some JPEGs have junk at end, skip over it but if we find what looks - // like a valid marker, resume there - while (!stbi__at_eof(j->s)) { - int x = stbi__get8(j->s); - while (x == 255) { // might be a marker - if (stbi__at_eof(j->s)) - return STBI__MARKER_none; - x = stbi__get8(j->s); - if (x != 0x00 && x != 0xff) { - // not a stuffed zero or lead-in to another marker, looks - // like an actual marker, return it - return x; - } - // stuffed zero has x=0 now which ends the loop, meaning we go - // back to regular scan loop. - // repeated 0xff keeps trying to read the next byte of the marker. - } - } - return STBI__MARKER_none; +static stbi_uc stbi__skip_jpeg_junk_at_end(stbi__jpeg *j) +{ + // some JPEGs have junk at end, skip over it but if we find what looks + // like a valid marker, resume there + while (!stbi__at_eof(j->s)) { + stbi_uc x = stbi__get8(j->s); + while (x == 0xff) { // might be a marker + if (stbi__at_eof(j->s)) return STBI__MARKER_none; + x = stbi__get8(j->s); + if (x != 0x00 && x != 0xff) { + // not a stuffed zero or lead-in to another marker, looks + // like an actual marker, return it + return x; + } + // stuffed zero has x=0 now which ends the loop, meaning we go + // back to regular scan loop. + // repeated 0xff keeps trying to read the next byte of the marker. + } + } + return STBI__MARKER_none; } // decode image to YCbCr format -static int stbi__decode_jpeg_image(stbi__jpeg * j) { - int m; - for (m = 0; m < 4; m++) { - j->img_comp[m].raw_data = NULL; - j->img_comp[m].raw_coeff = NULL; - } - j->restart_interval = 0; - if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) - return 0; - m = stbi__get_marker(j); - while (!stbi__EOI(m)) { - if (stbi__SOS(m)) { - if (!stbi__process_scan_header(j)) - return 0; - if (!stbi__parse_entropy_coded_data(j)) - return 0; - if (j->marker == STBI__MARKER_none) { - j->marker = stbi__skip_jpeg_junk_at_end(j); - // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0 - } +static int stbi__decode_jpeg_image(stbi__jpeg *j) +{ + int m; + for (m = 0; m < 4; m++) { + j->img_comp[m].raw_data = NULL; + j->img_comp[m].raw_coeff = NULL; + } + j->restart_interval = 0; + if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) return 0; + m = stbi__get_marker(j); + while (!stbi__EOI(m)) { + if (stbi__SOS(m)) { + if (!stbi__process_scan_header(j)) return 0; + if (!stbi__parse_entropy_coded_data(j)) return 0; + if (j->marker == STBI__MARKER_none ) { + j->marker = stbi__skip_jpeg_junk_at_end(j); + // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0 + } + m = stbi__get_marker(j); + if (STBI__RESTART(m)) m = stbi__get_marker(j); - if (STBI__RESTART(m)) - m = stbi__get_marker(j); - } else if (stbi__DNL(m)) { - int Ld = stbi__get16be(j->s); - stbi__uint32 NL = stbi__get16be(j->s); - if (Ld != 4) - return stbi__err("bad DNL len", "Corrupt JPEG"); - if (NL != j->s->img_y) - return stbi__err("bad DNL height", "Corrupt JPEG"); - m = stbi__get_marker(j); - } else { - if (!stbi__process_marker(j, m)) - return 1; - m = stbi__get_marker(j); - } - } - if (j->progressive) - stbi__jpeg_finish(j); - return 1; + } else if (stbi__DNL(m)) { + int Ld = stbi__get16be(j->s); + stbi__uint32 NL = stbi__get16be(j->s); + if (Ld != 4) return stbi__err("bad DNL len", "Corrupt JPEG"); + if (NL != j->s->img_y) return stbi__err("bad DNL height", "Corrupt JPEG"); + m = stbi__get_marker(j); + } else { + if (!stbi__process_marker(j, m)) return 1; + m = stbi__get_marker(j); + } + } + if (j->progressive) + stbi__jpeg_finish(j); + return 1; } // static jfif-centered resampling (across block boundaries) -typedef stbi_uc * (*resample_row_func)(stbi_uc * out, stbi_uc * in0, stbi_uc * in1, int w, int hs); +typedef stbi_uc *(*resample_row_func)(stbi_uc *out, stbi_uc *in0, stbi_uc *in1, + int w, int hs); -#define stbi__div4(x) ((stbi_uc)((x) >> 2)) +#define stbi__div4(x) ((stbi_uc) ((x) >> 2)) -static stbi_uc * resample_row_1(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - STBI_NOTUSED(out); - STBI_NOTUSED(in_far); - STBI_NOTUSED(w); - STBI_NOTUSED(hs); - return in_near; +static stbi_uc *resample_row_1(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + STBI_NOTUSED(out); + STBI_NOTUSED(in_far); + STBI_NOTUSED(w); + STBI_NOTUSED(hs); + return in_near; } -static stbi_uc * stbi__resample_row_v_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate two samples vertically for every one in input - int i; - STBI_NOTUSED(hs); - for (i = 0; i < w; ++i) - out[i] = stbi__div4(3 * in_near[i] + in_far[i] + 2); - return out; +static stbi_uc* stbi__resample_row_v_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate two samples vertically for every one in input + int i; + STBI_NOTUSED(hs); + for (i=0; i < w; ++i) + out[i] = stbi__div4(3*in_near[i] + in_far[i] + 2); + return out; } -static stbi_uc * stbi__resample_row_h_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate two samples horizontally for every one in input - int i; - stbi_uc * input = in_near; +static stbi_uc* stbi__resample_row_h_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate two samples horizontally for every one in input + int i; + stbi_uc *input = in_near; - if (w == 1) { - // if only one sample, can't do any interpolation - out[0] = out[1] = input[0]; - return out; - } + if (w == 1) { + // if only one sample, can't do any interpolation + out[0] = out[1] = input[0]; + return out; + } - out[0] = input[0]; - out[1] = stbi__div4(input[0] * 3 + input[1] + 2); - for (i = 1; i < w - 1; ++i) { - int n = 3 * input[i] + 2; - out[i * 2 + 0] = stbi__div4(n + input[i - 1]); - out[i * 2 + 1] = stbi__div4(n + input[i + 1]); - } - out[i * 2 + 0] = stbi__div4(input[w - 2] * 3 + input[w - 1] + 2); - out[i * 2 + 1] = input[w - 1]; + out[0] = input[0]; + out[1] = stbi__div4(input[0]*3 + input[1] + 2); + for (i=1; i < w-1; ++i) { + int n = 3*input[i]+2; + out[i*2+0] = stbi__div4(n+input[i-1]); + out[i*2+1] = stbi__div4(n+input[i+1]); + } + out[i*2+0] = stbi__div4(input[w-2]*3 + input[w-1] + 2); + out[i*2+1] = input[w-1]; - STBI_NOTUSED(in_far); - STBI_NOTUSED(hs); + STBI_NOTUSED(in_far); + STBI_NOTUSED(hs); - return out; + return out; } -#define stbi__div16(x) ((stbi_uc)((x) >> 4)) +#define stbi__div16(x) ((stbi_uc) ((x) >> 4)) -static stbi_uc * stbi__resample_row_hv_2(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate 2x2 samples for every one in input - int i, t0, t1; - if (w == 1) { - out[0] = out[1] = stbi__div4(3 * in_near[0] + in_far[0] + 2); - return out; - } +static stbi_uc *stbi__resample_row_hv_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate 2x2 samples for every one in input + int i,t0,t1; + if (w == 1) { + out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2); + return out; + } - t1 = 3 * in_near[0] + in_far[0]; - out[0] = stbi__div4(t1 + 2); - for (i = 1; i < w; ++i) { - t0 = t1; - t1 = 3 * in_near[i] + in_far[i]; - out[i * 2 - 1] = stbi__div16(3 * t0 + t1 + 8); - out[i * 2] = stbi__div16(3 * t1 + t0 + 8); - } - out[w * 2 - 1] = stbi__div4(t1 + 2); + t1 = 3*in_near[0] + in_far[0]; + out[0] = stbi__div4(t1+2); + for (i=1; i < w; ++i) { + t0 = t1; + t1 = 3*in_near[i]+in_far[i]; + out[i*2-1] = stbi__div16(3*t0 + t1 + 8); + out[i*2 ] = stbi__div16(3*t1 + t0 + 8); + } + out[w*2-1] = stbi__div4(t1+2); - STBI_NOTUSED(hs); + STBI_NOTUSED(hs); - return out; + return out; } #if defined(STBI_SSE2) || defined(STBI_NEON) -static stbi_uc * stbi__resample_row_hv_2_simd(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // need to generate 2x2 samples for every one in input - int i = 0, t0, t1; +static stbi_uc *stbi__resample_row_hv_2_simd(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // need to generate 2x2 samples for every one in input + int i=0,t0,t1; - if (w == 1) { - out[0] = out[1] = stbi__div4(3 * in_near[0] + in_far[0] + 2); - return out; - } + if (w == 1) { + out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2); + return out; + } - t1 = 3 * in_near[0] + in_far[0]; - // process groups of 8 pixels for as long as we can. - // note we can't handle the last pixel in a row in this loop - // because we need to handle the filter boundary conditions. - for (; i < ((w - 1) & ~7); i += 8) { + t1 = 3*in_near[0] + in_far[0]; + // process groups of 8 pixels for as long as we can. + // note we can't handle the last pixel in a row in this loop + // because we need to handle the filter boundary conditions. + for (; i < ((w-1) & ~7); i += 8) { #if defined(STBI_SSE2) - // load and perform the vertical filtering pass - // this uses 3*x + y = 4*x + (y - x) - __m128i zero = _mm_setzero_si128(); - __m128i farb = _mm_loadl_epi64((__m128i *)(in_far + i)); - __m128i nearb = _mm_loadl_epi64((__m128i *)(in_near + i)); - __m128i farw = _mm_unpacklo_epi8(farb, zero); - __m128i nearw = _mm_unpacklo_epi8(nearb, zero); - __m128i diff = _mm_sub_epi16(farw, nearw); - __m128i nears = _mm_slli_epi16(nearw, 2); - __m128i curr = _mm_add_epi16(nears, diff); // current row + // load and perform the vertical filtering pass + // this uses 3*x + y = 4*x + (y - x) + __m128i zero = _mm_setzero_si128(); + __m128i farb = _mm_loadl_epi64((__m128i *) (in_far + i)); + __m128i nearb = _mm_loadl_epi64((__m128i *) (in_near + i)); + __m128i farw = _mm_unpacklo_epi8(farb, zero); + __m128i nearw = _mm_unpacklo_epi8(nearb, zero); + __m128i diff = _mm_sub_epi16(farw, nearw); + __m128i nears = _mm_slli_epi16(nearw, 2); + __m128i curr = _mm_add_epi16(nears, diff); // current row - // horizontal filter works the same based on shifted vers of current - // row. "prev" is current row shifted right by 1 pixel; we need to - // insert the previous pixel value (from t1). - // "next" is current row shifted left by 1 pixel, with first pixel - // of next block of 8 pixels added in. - __m128i prv0 = _mm_slli_si128(curr, 2); - __m128i nxt0 = _mm_srli_si128(curr, 2); - __m128i prev = _mm_insert_epi16(prv0, t1, 0); - __m128i next = _mm_insert_epi16(nxt0, 3 * in_near[i + 8] + in_far[i + 8], 7); + // horizontal filter works the same based on shifted vers of current + // row. "prev" is current row shifted right by 1 pixel; we need to + // insert the previous pixel value (from t1). + // "next" is current row shifted left by 1 pixel, with first pixel + // of next block of 8 pixels added in. + __m128i prv0 = _mm_slli_si128(curr, 2); + __m128i nxt0 = _mm_srli_si128(curr, 2); + __m128i prev = _mm_insert_epi16(prv0, t1, 0); + __m128i next = _mm_insert_epi16(nxt0, 3*in_near[i+8] + in_far[i+8], 7); - // horizontal filter, polyphase implementation since it's convenient: - // even pixels = 3*cur + prev = cur*4 + (prev - cur) - // odd pixels = 3*cur + next = cur*4 + (next - cur) - // note the shared term. - __m128i bias = _mm_set1_epi16(8); - __m128i curs = _mm_slli_epi16(curr, 2); - __m128i prvd = _mm_sub_epi16(prev, curr); - __m128i nxtd = _mm_sub_epi16(next, curr); - __m128i curb = _mm_add_epi16(curs, bias); - __m128i even = _mm_add_epi16(prvd, curb); - __m128i odd = _mm_add_epi16(nxtd, curb); + // horizontal filter, polyphase implementation since it's convenient: + // even pixels = 3*cur + prev = cur*4 + (prev - cur) + // odd pixels = 3*cur + next = cur*4 + (next - cur) + // note the shared term. + __m128i bias = _mm_set1_epi16(8); + __m128i curs = _mm_slli_epi16(curr, 2); + __m128i prvd = _mm_sub_epi16(prev, curr); + __m128i nxtd = _mm_sub_epi16(next, curr); + __m128i curb = _mm_add_epi16(curs, bias); + __m128i even = _mm_add_epi16(prvd, curb); + __m128i odd = _mm_add_epi16(nxtd, curb); - // interleave even and odd pixels, then undo scaling. - __m128i int0 = _mm_unpacklo_epi16(even, odd); - __m128i int1 = _mm_unpackhi_epi16(even, odd); - __m128i de0 = _mm_srli_epi16(int0, 4); - __m128i de1 = _mm_srli_epi16(int1, 4); + // interleave even and odd pixels, then undo scaling. + __m128i int0 = _mm_unpacklo_epi16(even, odd); + __m128i int1 = _mm_unpackhi_epi16(even, odd); + __m128i de0 = _mm_srli_epi16(int0, 4); + __m128i de1 = _mm_srli_epi16(int1, 4); - // pack and write output - __m128i outv = _mm_packus_epi16(de0, de1); - _mm_storeu_si128((__m128i *)(out + i * 2), outv); + // pack and write output + __m128i outv = _mm_packus_epi16(de0, de1); + _mm_storeu_si128((__m128i *) (out + i*2), outv); #elif defined(STBI_NEON) - // load and perform the vertical filtering pass - // this uses 3*x + y = 4*x + (y - x) - uint8x8_t farb = vld1_u8(in_far + i); - uint8x8_t nearb = vld1_u8(in_near + i); - int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(farb, nearb)); - int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2)); - int16x8_t curr = vaddq_s16(nears, diff); // current row + // load and perform the vertical filtering pass + // this uses 3*x + y = 4*x + (y - x) + uint8x8_t farb = vld1_u8(in_far + i); + uint8x8_t nearb = vld1_u8(in_near + i); + int16x8_t diff = vreinterpretq_s16_u16(vsubl_u8(farb, nearb)); + int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2)); + int16x8_t curr = vaddq_s16(nears, diff); // current row - // horizontal filter works the same based on shifted vers of current - // row. "prev" is current row shifted right by 1 pixel; we need to - // insert the previous pixel value (from t1). - // "next" is current row shifted left by 1 pixel, with first pixel - // of next block of 8 pixels added in. - int16x8_t prv0 = vextq_s16(curr, curr, 7); - int16x8_t nxt0 = vextq_s16(curr, curr, 1); - int16x8_t prev = vsetq_lane_s16(t1, prv0, 0); - int16x8_t next = vsetq_lane_s16(3 * in_near[i + 8] + in_far[i + 8], nxt0, 7); + // horizontal filter works the same based on shifted vers of current + // row. "prev" is current row shifted right by 1 pixel; we need to + // insert the previous pixel value (from t1). + // "next" is current row shifted left by 1 pixel, with first pixel + // of next block of 8 pixels added in. + int16x8_t prv0 = vextq_s16(curr, curr, 7); + int16x8_t nxt0 = vextq_s16(curr, curr, 1); + int16x8_t prev = vsetq_lane_s16(t1, prv0, 0); + int16x8_t next = vsetq_lane_s16(3*in_near[i+8] + in_far[i+8], nxt0, 7); - // horizontal filter, polyphase implementation since it's convenient: - // even pixels = 3*cur + prev = cur*4 + (prev - cur) - // odd pixels = 3*cur + next = cur*4 + (next - cur) - // note the shared term. - int16x8_t curs = vshlq_n_s16(curr, 2); - int16x8_t prvd = vsubq_s16(prev, curr); - int16x8_t nxtd = vsubq_s16(next, curr); - int16x8_t even = vaddq_s16(curs, prvd); - int16x8_t odd = vaddq_s16(curs, nxtd); + // horizontal filter, polyphase implementation since it's convenient: + // even pixels = 3*cur + prev = cur*4 + (prev - cur) + // odd pixels = 3*cur + next = cur*4 + (next - cur) + // note the shared term. + int16x8_t curs = vshlq_n_s16(curr, 2); + int16x8_t prvd = vsubq_s16(prev, curr); + int16x8_t nxtd = vsubq_s16(next, curr); + int16x8_t even = vaddq_s16(curs, prvd); + int16x8_t odd = vaddq_s16(curs, nxtd); - // undo scaling and round, then store with even/odd phases interleaved - uint8x8x2_t o; - o.val[0] = vqrshrun_n_s16(even, 4); - o.val[1] = vqrshrun_n_s16(odd, 4); - vst2_u8(out + i * 2, o); + // undo scaling and round, then store with even/odd phases interleaved + uint8x8x2_t o; + o.val[0] = vqrshrun_n_s16(even, 4); + o.val[1] = vqrshrun_n_s16(odd, 4); + vst2_u8(out + i*2, o); #endif - // "previous" value for next iter - t1 = 3 * in_near[i + 7] + in_far[i + 7]; - } + // "previous" value for next iter + t1 = 3*in_near[i+7] + in_far[i+7]; + } - t0 = t1; - t1 = 3 * in_near[i] + in_far[i]; - out[i * 2] = stbi__div16(3 * t1 + t0 + 8); + t0 = t1; + t1 = 3*in_near[i] + in_far[i]; + out[i*2] = stbi__div16(3*t1 + t0 + 8); - for (++i; i < w; ++i) { - t0 = t1; - t1 = 3 * in_near[i] + in_far[i]; - out[i * 2 - 1] = stbi__div16(3 * t0 + t1 + 8); - out[i * 2] = stbi__div16(3 * t1 + t0 + 8); - } - out[w * 2 - 1] = stbi__div4(t1 + 2); + for (++i; i < w; ++i) { + t0 = t1; + t1 = 3*in_near[i]+in_far[i]; + out[i*2-1] = stbi__div16(3*t0 + t1 + 8); + out[i*2 ] = stbi__div16(3*t1 + t0 + 8); + } + out[w*2-1] = stbi__div4(t1+2); - STBI_NOTUSED(hs); + STBI_NOTUSED(hs); - return out; + return out; } #endif -static stbi_uc * stbi__resample_row_generic(stbi_uc * out, stbi_uc * in_near, stbi_uc * in_far, int w, int hs) { - // resample with nearest-neighbor - int i, j; - STBI_NOTUSED(in_far); - for (i = 0; i < w; ++i) - for (j = 0; j < hs; ++j) - out[i * hs + j] = in_near[i]; - return out; +static stbi_uc *stbi__resample_row_generic(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs) +{ + // resample with nearest-neighbor + int i,j; + STBI_NOTUSED(in_far); + for (i=0; i < w; ++i) + for (j=0; j < hs; ++j) + out[i*hs+j] = in_near[i]; + return out; } // this is a reduced-precision calculation of YCbCr-to-RGB introduced // to make sure the code produces the same results in both SIMD and scalar -#define stbi__float2fixed(x) (((int)((x)*4096.0f + 0.5f)) << 8) -static void stbi__YCbCr_to_RGB_row(stbi_uc * out, const stbi_uc * y, const stbi_uc * pcb, const stbi_uc * pcr, int count, - int step) { - int i; - for (i = 0; i < count; ++i) { - int y_fixed = (y[i] << 20) + (1 << 19); // rounding - int r, g, b; - int cr = pcr[i] - 128; - int cb = pcb[i] - 128; - r = y_fixed + cr * stbi__float2fixed(1.40200f); - g = y_fixed + (cr * -stbi__float2fixed(0.71414f)) + ((cb * -stbi__float2fixed(0.34414f)) & 0xffff0000); - b = y_fixed + cb * stbi__float2fixed(1.77200f); - r >>= 20; - g >>= 20; - b >>= 20; - if ((unsigned)r > 255) { - if (r < 0) - r = 0; - else - r = 255; - } - if ((unsigned)g > 255) { - if (g < 0) - g = 0; - else - g = 255; - } - if ((unsigned)b > 255) { - if (b < 0) - b = 0; - else - b = 255; - } - out[0] = (stbi_uc)r; - out[1] = (stbi_uc)g; - out[2] = (stbi_uc)b; - out[3] = 255; - out += step; - } +#define stbi__float2fixed(x) (((int) ((x) * 4096.0f + 0.5f)) << 8) +static void stbi__YCbCr_to_RGB_row(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step) +{ + int i; + for (i=0; i < count; ++i) { + int y_fixed = (y[i] << 20) + (1<<19); // rounding + int r,g,b; + int cr = pcr[i] - 128; + int cb = pcb[i] - 128; + r = y_fixed + cr* stbi__float2fixed(1.40200f); + g = y_fixed + (cr*-stbi__float2fixed(0.71414f)) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000); + b = y_fixed + cb* stbi__float2fixed(1.77200f); + r >>= 20; + g >>= 20; + b >>= 20; + if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; } + if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; } + if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; } + out[0] = (stbi_uc)r; + out[1] = (stbi_uc)g; + out[2] = (stbi_uc)b; + out[3] = 255; + out += step; + } } #if defined(STBI_SSE2) || defined(STBI_NEON) -static void stbi__YCbCr_to_RGB_simd(stbi_uc * out, stbi_uc const * y, stbi_uc const * pcb, stbi_uc const * pcr, int count, - int step) { - int i = 0; +static void stbi__YCbCr_to_RGB_simd(stbi_uc *out, stbi_uc const *y, stbi_uc const *pcb, stbi_uc const *pcr, int count, int step) +{ + int i = 0; #ifdef STBI_SSE2 - // step == 3 is pretty ugly on the final interleave, and i'm not convinced - // it's useful in practice (you wouldn't use it for textures, for example). - // so just accelerate step == 4 case. - if (step == 4) { - // this is a fairly straightforward implementation and not super-optimized. - __m128i signflip = _mm_set1_epi8(-0x80); - __m128i cr_const0 = _mm_set1_epi16((short)(1.40200f * 4096.0f + 0.5f)); - __m128i cr_const1 = _mm_set1_epi16(-(short)(0.71414f * 4096.0f + 0.5f)); - __m128i cb_const0 = _mm_set1_epi16(-(short)(0.34414f * 4096.0f + 0.5f)); - __m128i cb_const1 = _mm_set1_epi16((short)(1.77200f * 4096.0f + 0.5f)); - __m128i y_bias = _mm_set1_epi8((char)(unsigned char)128); - __m128i xw = _mm_set1_epi16(255); // alpha channel + // step == 3 is pretty ugly on the final interleave, and i'm not convinced + // it's useful in practice (you wouldn't use it for textures, for example). + // so just accelerate step == 4 case. + if (step == 4) { + // this is a fairly straightforward implementation and not super-optimized. + __m128i signflip = _mm_set1_epi8(-0x80); + __m128i cr_const0 = _mm_set1_epi16( (short) ( 1.40200f*4096.0f+0.5f)); + __m128i cr_const1 = _mm_set1_epi16( - (short) ( 0.71414f*4096.0f+0.5f)); + __m128i cb_const0 = _mm_set1_epi16( - (short) ( 0.34414f*4096.0f+0.5f)); + __m128i cb_const1 = _mm_set1_epi16( (short) ( 1.77200f*4096.0f+0.5f)); + __m128i y_bias = _mm_set1_epi8((char) (unsigned char) 128); + __m128i xw = _mm_set1_epi16(255); // alpha channel - for (; i + 7 < count; i += 8) { - // load - __m128i y_bytes = _mm_loadl_epi64((__m128i *)(y + i)); - __m128i cr_bytes = _mm_loadl_epi64((__m128i *)(pcr + i)); - __m128i cb_bytes = _mm_loadl_epi64((__m128i *)(pcb + i)); - __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128 - __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128 + for (; i+7 < count; i += 8) { + // load + __m128i y_bytes = _mm_loadl_epi64((__m128i *) (y+i)); + __m128i cr_bytes = _mm_loadl_epi64((__m128i *) (pcr+i)); + __m128i cb_bytes = _mm_loadl_epi64((__m128i *) (pcb+i)); + __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128 + __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128 - // unpack to short (and left-shift cr, cb by 8) - __m128i yw = _mm_unpacklo_epi8(y_bias, y_bytes); - __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased); - __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased); + // unpack to short (and left-shift cr, cb by 8) + __m128i yw = _mm_unpacklo_epi8(y_bias, y_bytes); + __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased); + __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased); - // color transform - __m128i yws = _mm_srli_epi16(yw, 4); - __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw); - __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw); - __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1); - __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1); - __m128i rws = _mm_add_epi16(cr0, yws); - __m128i gwt = _mm_add_epi16(cb0, yws); - __m128i bws = _mm_add_epi16(yws, cb1); - __m128i gws = _mm_add_epi16(gwt, cr1); + // color transform + __m128i yws = _mm_srli_epi16(yw, 4); + __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw); + __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw); + __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1); + __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1); + __m128i rws = _mm_add_epi16(cr0, yws); + __m128i gwt = _mm_add_epi16(cb0, yws); + __m128i bws = _mm_add_epi16(yws, cb1); + __m128i gws = _mm_add_epi16(gwt, cr1); - // descale - __m128i rw = _mm_srai_epi16(rws, 4); - __m128i bw = _mm_srai_epi16(bws, 4); - __m128i gw = _mm_srai_epi16(gws, 4); + // descale + __m128i rw = _mm_srai_epi16(rws, 4); + __m128i bw = _mm_srai_epi16(bws, 4); + __m128i gw = _mm_srai_epi16(gws, 4); - // back to byte, set up for transpose - __m128i brb = _mm_packus_epi16(rw, bw); - __m128i gxb = _mm_packus_epi16(gw, xw); + // back to byte, set up for transpose + __m128i brb = _mm_packus_epi16(rw, bw); + __m128i gxb = _mm_packus_epi16(gw, xw); - // transpose to interleave channels - __m128i t0 = _mm_unpacklo_epi8(brb, gxb); - __m128i t1 = _mm_unpackhi_epi8(brb, gxb); - __m128i o0 = _mm_unpacklo_epi16(t0, t1); - __m128i o1 = _mm_unpackhi_epi16(t0, t1); + // transpose to interleave channels + __m128i t0 = _mm_unpacklo_epi8(brb, gxb); + __m128i t1 = _mm_unpackhi_epi8(brb, gxb); + __m128i o0 = _mm_unpacklo_epi16(t0, t1); + __m128i o1 = _mm_unpackhi_epi16(t0, t1); - // store - _mm_storeu_si128((__m128i *)(out + 0), o0); - _mm_storeu_si128((__m128i *)(out + 16), o1); - out += 32; - } - } + // store + _mm_storeu_si128((__m128i *) (out + 0), o0); + _mm_storeu_si128((__m128i *) (out + 16), o1); + out += 32; + } + } #endif #ifdef STBI_NEON - // in this version, step=3 support would be easy to add. but is there demand? - if (step == 4) { - // this is a fairly straightforward implementation and not super-optimized. - uint8x8_t signflip = vdup_n_u8(0x80); - int16x8_t cr_const0 = vdupq_n_s16((short)(1.40200f * 4096.0f + 0.5f)); - int16x8_t cr_const1 = vdupq_n_s16(-(short)(0.71414f * 4096.0f + 0.5f)); - int16x8_t cb_const0 = vdupq_n_s16(-(short)(0.34414f * 4096.0f + 0.5f)); - int16x8_t cb_const1 = vdupq_n_s16((short)(1.77200f * 4096.0f + 0.5f)); + // in this version, step=3 support would be easy to add. but is there demand? + if (step == 4) { + // this is a fairly straightforward implementation and not super-optimized. + uint8x8_t signflip = vdup_n_u8(0x80); + int16x8_t cr_const0 = vdupq_n_s16( (short) ( 1.40200f*4096.0f+0.5f)); + int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f)); + int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f)); + int16x8_t cb_const1 = vdupq_n_s16( (short) ( 1.77200f*4096.0f+0.5f)); - for (; i + 7 < count; i += 8) { - // load - uint8x8_t y_bytes = vld1_u8(y + i); - uint8x8_t cr_bytes = vld1_u8(pcr + i); - uint8x8_t cb_bytes = vld1_u8(pcb + i); - int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip)); - int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip)); + for (; i+7 < count; i += 8) { + // load + uint8x8_t y_bytes = vld1_u8(y + i); + uint8x8_t cr_bytes = vld1_u8(pcr + i); + uint8x8_t cb_bytes = vld1_u8(pcb + i); + int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip)); + int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip)); - // expand to s16 - int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4)); - int16x8_t crw = vshll_n_s8(cr_biased, 7); - int16x8_t cbw = vshll_n_s8(cb_biased, 7); + // expand to s16 + int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4)); + int16x8_t crw = vshll_n_s8(cr_biased, 7); + int16x8_t cbw = vshll_n_s8(cb_biased, 7); - // color transform - int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0); - int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0); - int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1); - int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1); - int16x8_t rws = vaddq_s16(yws, cr0); - int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1); - int16x8_t bws = vaddq_s16(yws, cb1); + // color transform + int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0); + int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0); + int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1); + int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1); + int16x8_t rws = vaddq_s16(yws, cr0); + int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1); + int16x8_t bws = vaddq_s16(yws, cb1); - // undo scaling, round, convert to byte - uint8x8x4_t o; - o.val[0] = vqrshrun_n_s16(rws, 4); - o.val[1] = vqrshrun_n_s16(gws, 4); - o.val[2] = vqrshrun_n_s16(bws, 4); - o.val[3] = vdup_n_u8(255); + // undo scaling, round, convert to byte + uint8x8x4_t o; + o.val[0] = vqrshrun_n_s16(rws, 4); + o.val[1] = vqrshrun_n_s16(gws, 4); + o.val[2] = vqrshrun_n_s16(bws, 4); + o.val[3] = vdup_n_u8(255); - // store, interleaving r/g/b/a - vst4_u8(out, o); - out += 8 * 4; - } - } + // store, interleaving r/g/b/a + vst4_u8(out, o); + out += 8*4; + } + } #endif - for (; i < count; ++i) { - int y_fixed = (y[i] << 20) + (1 << 19); // rounding - int r, g, b; - int cr = pcr[i] - 128; - int cb = pcb[i] - 128; - r = y_fixed + cr * stbi__float2fixed(1.40200f); - g = y_fixed + cr * -stbi__float2fixed(0.71414f) + ((cb * -stbi__float2fixed(0.34414f)) & 0xffff0000); - b = y_fixed + cb * stbi__float2fixed(1.77200f); - r >>= 20; - g >>= 20; - b >>= 20; - if ((unsigned)r > 255) { - if (r < 0) - r = 0; - else - r = 255; - } - if ((unsigned)g > 255) { - if (g < 0) - g = 0; - else - g = 255; - } - if ((unsigned)b > 255) { - if (b < 0) - b = 0; - else - b = 255; - } - out[0] = (stbi_uc)r; - out[1] = (stbi_uc)g; - out[2] = (stbi_uc)b; - out[3] = 255; - out += step; - } + for (; i < count; ++i) { + int y_fixed = (y[i] << 20) + (1<<19); // rounding + int r,g,b; + int cr = pcr[i] - 128; + int cb = pcb[i] - 128; + r = y_fixed + cr* stbi__float2fixed(1.40200f); + g = y_fixed + cr*-stbi__float2fixed(0.71414f) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000); + b = y_fixed + cb* stbi__float2fixed(1.77200f); + r >>= 20; + g >>= 20; + b >>= 20; + if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; } + if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; } + if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; } + out[0] = (stbi_uc)r; + out[1] = (stbi_uc)g; + out[2] = (stbi_uc)b; + out[3] = 255; + out += step; + } } #endif // set up the kernels -static void stbi__setup_jpeg(stbi__jpeg * j) { - j->idct_block_kernel = stbi__idct_block; - j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row; - j->resample_row_hv_2_kernel = stbi__resample_row_hv_2; +static void stbi__setup_jpeg(stbi__jpeg *j) +{ + j->idct_block_kernel = stbi__idct_block; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2; #ifdef STBI_SSE2 - if (stbi__sse2_available()) { - j->idct_block_kernel = stbi__idct_simd; - j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; - j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; - } + if (stbi__sse2_available()) { + j->idct_block_kernel = stbi__idct_simd; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; + } #endif #ifdef STBI_NEON - j->idct_block_kernel = stbi__idct_simd; - j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; - j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; + j->idct_block_kernel = stbi__idct_simd; + j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd; + j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd; #endif } // clean up the temporary component buffers -static void stbi__cleanup_jpeg(stbi__jpeg * j) { stbi__free_jpeg_components(j, j->s->img_n, 0); } +static void stbi__cleanup_jpeg(stbi__jpeg *j) +{ + stbi__free_jpeg_components(j, j->s->img_n, 0); +} -typedef struct { - resample_row_func resample; - stbi_uc *line0, *line1; - int hs, vs; // expansion factor in each axis - int w_lores; // horizontal pixels pre-expansion - int ystep; // how far through vertical expansion we are - int ypos; // which pre-expansion row we're on +typedef struct +{ + resample_row_func resample; + stbi_uc *line0,*line1; + int hs,vs; // expansion factor in each axis + int w_lores; // horizontal pixels pre-expansion + int ystep; // how far through vertical expansion we are + int ypos; // which pre-expansion row we're on } stbi__resample; // fast 0..255 * 0..255 => 0..255 rounded multiplication -static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y) { - unsigned int t = x * y + 128; - return (stbi_uc)((t + (t >> 8)) >> 8); +static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y) +{ + unsigned int t = x*y + 128; + return (stbi_uc) ((t + (t >>8)) >> 8); } -static stbi_uc * load_jpeg_image(stbi__jpeg * z, int * out_x, int * out_y, int * comp, int req_comp) { - int n, decode_n, is_rgb; - z->s->img_n = 0; // make stbi__cleanup_jpeg safe +static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp, int req_comp) +{ + int n, decode_n, is_rgb; + z->s->img_n = 0; // make stbi__cleanup_jpeg safe - // validate req_comp - if (req_comp < 0 || req_comp > 4) - return stbi__errpuc("bad req_comp", "Internal error"); + // validate req_comp + if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error"); - // load a jpeg image from whichever source, but leave in YCbCr format - if (!stbi__decode_jpeg_image(z)) { - stbi__cleanup_jpeg(z); - return NULL; - } + // load a jpeg image from whichever source, but leave in YCbCr format + if (!stbi__decode_jpeg_image(z)) { stbi__cleanup_jpeg(z); return NULL; } - // determine actual number of components to generate - n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1; + // determine actual number of components to generate + n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1; - is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif)); + is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif)); - if (z->s->img_n == 3 && n < 3 && !is_rgb) - decode_n = 1; - else - decode_n = z->s->img_n; + if (z->s->img_n == 3 && n < 3 && !is_rgb) + decode_n = 1; + else + decode_n = z->s->img_n; - // nothing to do if no components requested; check this now to avoid - // accessing uninitialized coutput[0] later - if (decode_n <= 0) { - stbi__cleanup_jpeg(z); - return NULL; - } + // nothing to do if no components requested; check this now to avoid + // accessing uninitialized coutput[0] later + if (decode_n <= 0) { stbi__cleanup_jpeg(z); return NULL; } - // resample and color-convert - { - int k; - unsigned int i, j; - stbi_uc * output; - stbi_uc * coutput[4] = {NULL, NULL, NULL, NULL}; + // resample and color-convert + { + int k; + unsigned int i,j; + stbi_uc *output; + stbi_uc *coutput[4] = { NULL, NULL, NULL, NULL }; - stbi__resample res_comp[4]; + stbi__resample res_comp[4]; - for (k = 0; k < decode_n; ++k) { - stbi__resample * r = &res_comp[k]; + for (k=0; k < decode_n; ++k) { + stbi__resample *r = &res_comp[k]; - // allocate line buffer big enough for upsampling off the edges - // with upsample factor of 4 - z->img_comp[k].linebuf = (stbi_uc *)stbi__malloc(z->s->img_x + 3); - if (!z->img_comp[k].linebuf) { - stbi__cleanup_jpeg(z); - return stbi__errpuc("outofmem", "Out of memory"); + // allocate line buffer big enough for upsampling off the edges + // with upsample factor of 4 + z->img_comp[k].linebuf = (stbi_uc *) stbi__malloc(z->s->img_x + 3); + if (!z->img_comp[k].linebuf) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); } + + r->hs = z->img_h_max / z->img_comp[k].h; + r->vs = z->img_v_max / z->img_comp[k].v; + r->ystep = r->vs >> 1; + r->w_lores = (z->s->img_x + r->hs-1) / r->hs; + r->ypos = 0; + r->line0 = r->line1 = z->img_comp[k].data; + + if (r->hs == 1 && r->vs == 1) r->resample = resample_row_1; + else if (r->hs == 1 && r->vs == 2) r->resample = stbi__resample_row_v_2; + else if (r->hs == 2 && r->vs == 1) r->resample = stbi__resample_row_h_2; + else if (r->hs == 2 && r->vs == 2) r->resample = z->resample_row_hv_2_kernel; + else r->resample = stbi__resample_row_generic; + } + + // can't error after this so, this is safe + output = (stbi_uc *) stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1); + if (!output) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); } + + // now go ahead and resample + for (j=0; j < z->s->img_y; ++j) { + stbi_uc *out = output + n * z->s->img_x * j; + for (k=0; k < decode_n; ++k) { + stbi__resample *r = &res_comp[k]; + int y_bot = r->ystep >= (r->vs >> 1); + coutput[k] = r->resample(z->img_comp[k].linebuf, + y_bot ? r->line1 : r->line0, + y_bot ? r->line0 : r->line1, + r->w_lores, r->hs); + if (++r->ystep >= r->vs) { + r->ystep = 0; + r->line0 = r->line1; + if (++r->ypos < z->img_comp[k].y) + r->line1 += z->img_comp[k].w2; } - - r->hs = z->img_h_max / z->img_comp[k].h; - r->vs = z->img_v_max / z->img_comp[k].v; - r->ystep = r->vs >> 1; - r->w_lores = (z->s->img_x + r->hs - 1) / r->hs; - r->ypos = 0; - r->line0 = r->line1 = z->img_comp[k].data; - - if (r->hs == 1 && r->vs == 1) - r->resample = resample_row_1; - else if (r->hs == 1 && r->vs == 2) - r->resample = stbi__resample_row_v_2; - else if (r->hs == 2 && r->vs == 1) - r->resample = stbi__resample_row_h_2; - else if (r->hs == 2 && r->vs == 2) - r->resample = z->resample_row_hv_2_kernel; - else - r->resample = stbi__resample_row_generic; - } - - // can't error after this so, this is safe - output = (stbi_uc *)stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1); - if (!output) { - stbi__cleanup_jpeg(z); - return stbi__errpuc("outofmem", "Out of memory"); - } - - // now go ahead and resample - for (j = 0; j < z->s->img_y; ++j) { - stbi_uc * out = output + n * z->s->img_x * j; - for (k = 0; k < decode_n; ++k) { - stbi__resample * r = &res_comp[k]; - int y_bot = r->ystep >= (r->vs >> 1); - coutput[k] = r->resample(z->img_comp[k].linebuf, y_bot ? r->line1 : r->line0, y_bot ? r->line0 : r->line1, - r->w_lores, r->hs); - if (++r->ystep >= r->vs) { - r->ystep = 0; - r->line0 = r->line1; - if (++r->ypos < z->img_comp[k].y) - r->line1 += z->img_comp[k].w2; - } - } - if (n >= 3) { - stbi_uc * y = coutput[0]; - if (z->s->img_n == 3) { - if (is_rgb) { - for (i = 0; i < z->s->img_x; ++i) { - out[0] = y[i]; - out[1] = coutput[1][i]; - out[2] = coutput[2][i]; - out[3] = 255; - out += n; - } - } else { - z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); - } - } else if (z->s->img_n == 4) { - if (z->app14_color_transform == 0) { // CMYK - for (i = 0; i < z->s->img_x; ++i) { - stbi_uc m = coutput[3][i]; - out[0] = stbi__blinn_8x8(coutput[0][i], m); - out[1] = stbi__blinn_8x8(coutput[1][i], m); - out[2] = stbi__blinn_8x8(coutput[2][i], m); - out[3] = 255; - out += n; - } - } else if (z->app14_color_transform == 2) { // YCCK - z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); - for (i = 0; i < z->s->img_x; ++i) { - stbi_uc m = coutput[3][i]; - out[0] = stbi__blinn_8x8(255 - out[0], m); - out[1] = stbi__blinn_8x8(255 - out[1], m); - out[2] = stbi__blinn_8x8(255 - out[2], m); - out += n; - } - } else { // YCbCr + alpha? Ignore the fourth channel for now - z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); - } - } else - for (i = 0; i < z->s->img_x; ++i) { - out[0] = out[1] = out[2] = y[i]; - out[3] = 255; // not used if n==3 - out += n; - } + } + if (n >= 3) { + stbi_uc *y = coutput[0]; + if (z->s->img_n == 3) { + if (is_rgb) { + for (i=0; i < z->s->img_x; ++i) { + out[0] = y[i]; + out[1] = coutput[1][i]; + out[2] = coutput[2][i]; + out[3] = 255; + out += n; + } + } else { + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + } + } else if (z->s->img_n == 4) { + if (z->app14_color_transform == 0) { // CMYK + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + out[0] = stbi__blinn_8x8(coutput[0][i], m); + out[1] = stbi__blinn_8x8(coutput[1][i], m); + out[2] = stbi__blinn_8x8(coutput[2][i], m); + out[3] = 255; + out += n; + } + } else if (z->app14_color_transform == 2) { // YCCK + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + out[0] = stbi__blinn_8x8(255 - out[0], m); + out[1] = stbi__blinn_8x8(255 - out[1], m); + out[2] = stbi__blinn_8x8(255 - out[2], m); + out += n; + } + } else { // YCbCr + alpha? Ignore the fourth channel for now + z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n); + } + } else + for (i=0; i < z->s->img_x; ++i) { + out[0] = out[1] = out[2] = y[i]; + out[3] = 255; // not used if n==3 + out += n; + } + } else { + if (is_rgb) { + if (n == 1) + for (i=0; i < z->s->img_x; ++i) + *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); + else { + for (i=0; i < z->s->img_x; ++i, out += 2) { + out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); + out[1] = 255; + } + } + } else if (z->s->img_n == 4 && z->app14_color_transform == 0) { + for (i=0; i < z->s->img_x; ++i) { + stbi_uc m = coutput[3][i]; + stbi_uc r = stbi__blinn_8x8(coutput[0][i], m); + stbi_uc g = stbi__blinn_8x8(coutput[1][i], m); + stbi_uc b = stbi__blinn_8x8(coutput[2][i], m); + out[0] = stbi__compute_y(r, g, b); + out[1] = 255; + out += n; + } + } else if (z->s->img_n == 4 && z->app14_color_transform == 2) { + for (i=0; i < z->s->img_x; ++i) { + out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]); + out[1] = 255; + out += n; + } } else { - if (is_rgb) { - if (n == 1) - for (i = 0; i < z->s->img_x; ++i) - *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); - else { - for (i = 0; i < z->s->img_x; ++i, out += 2) { - out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]); - out[1] = 255; - } - } - } else if (z->s->img_n == 4 && z->app14_color_transform == 0) { - for (i = 0; i < z->s->img_x; ++i) { - stbi_uc m = coutput[3][i]; - stbi_uc r = stbi__blinn_8x8(coutput[0][i], m); - stbi_uc g = stbi__blinn_8x8(coutput[1][i], m); - stbi_uc b = stbi__blinn_8x8(coutput[2][i], m); - out[0] = stbi__compute_y(r, g, b); - out[1] = 255; - out += n; - } - } else if (z->s->img_n == 4 && z->app14_color_transform == 2) { - for (i = 0; i < z->s->img_x; ++i) { - out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]); - out[1] = 255; - out += n; - } - } else { - stbi_uc * y = coutput[0]; - if (n == 1) - for (i = 0; i < z->s->img_x; ++i) - out[i] = y[i]; - else - for (i = 0; i < z->s->img_x; ++i) { - *out++ = y[i]; - *out++ = 255; - } - } + stbi_uc *y = coutput[0]; + if (n == 1) + for (i=0; i < z->s->img_x; ++i) out[i] = y[i]; + else + for (i=0; i < z->s->img_x; ++i) { *out++ = y[i]; *out++ = 255; } } - } - stbi__cleanup_jpeg(z); - *out_x = z->s->img_x; - *out_y = z->s->img_y; - if (comp) - *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output - return output; - } + } + } + stbi__cleanup_jpeg(z); + *out_x = z->s->img_x; + *out_y = z->s->img_y; + if (comp) *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output + return output; + } } -static void * stbi__jpeg_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - unsigned char * result; - stbi__jpeg * j = (stbi__jpeg *)stbi__malloc(sizeof(stbi__jpeg)); - if (!j) - return stbi__errpuc("outofmem", "Out of memory"); - memset(j, 0, sizeof(stbi__jpeg)); - STBI_NOTUSED(ri); - j->s = s; - stbi__setup_jpeg(j); - result = load_jpeg_image(j, x, y, comp, req_comp); - STBI_FREE(j); - return result; +static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + unsigned char* result; + stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg)); + if (!j) return stbi__errpuc("outofmem", "Out of memory"); + memset(j, 0, sizeof(stbi__jpeg)); + STBI_NOTUSED(ri); + j->s = s; + stbi__setup_jpeg(j); + result = load_jpeg_image(j, x,y,comp,req_comp); + STBI_FREE(j); + return result; } -static int stbi__jpeg_test(stbi__context * s) { - int r; - stbi__jpeg * j = (stbi__jpeg *)stbi__malloc(sizeof(stbi__jpeg)); - if (!j) - return stbi__err("outofmem", "Out of memory"); - memset(j, 0, sizeof(stbi__jpeg)); - j->s = s; - stbi__setup_jpeg(j); - r = stbi__decode_jpeg_header(j, STBI__SCAN_type); - stbi__rewind(s); - STBI_FREE(j); - return r; +static int stbi__jpeg_test(stbi__context *s) +{ + int r; + stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg)); + if (!j) return stbi__err("outofmem", "Out of memory"); + memset(j, 0, sizeof(stbi__jpeg)); + j->s = s; + stbi__setup_jpeg(j); + r = stbi__decode_jpeg_header(j, STBI__SCAN_type); + stbi__rewind(s); + STBI_FREE(j); + return r; } -static int stbi__jpeg_info_raw(stbi__jpeg * j, int * x, int * y, int * comp) { - if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) { - stbi__rewind(j->s); - return 0; - } - if (x) - *x = j->s->img_x; - if (y) - *y = j->s->img_y; - if (comp) - *comp = j->s->img_n >= 3 ? 3 : 1; - return 1; +static int stbi__jpeg_info_raw(stbi__jpeg *j, int *x, int *y, int *comp) +{ + if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) { + stbi__rewind( j->s ); + return 0; + } + if (x) *x = j->s->img_x; + if (y) *y = j->s->img_y; + if (comp) *comp = j->s->img_n >= 3 ? 3 : 1; + return 1; } -static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp) { - int result; - stbi__jpeg * j = (stbi__jpeg *)(stbi__malloc(sizeof(stbi__jpeg))); - if (!j) - return stbi__err("outofmem", "Out of memory"); - memset(j, 0, sizeof(stbi__jpeg)); - j->s = s; - result = stbi__jpeg_info_raw(j, x, y, comp); - STBI_FREE(j); - return result; +static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp) +{ + int result; + stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg))); + if (!j) return stbi__err("outofmem", "Out of memory"); + memset(j, 0, sizeof(stbi__jpeg)); + j->s = s; + result = stbi__jpeg_info_raw(j, x, y, comp); + STBI_FREE(j); + return result; } #endif @@ -4278,81 +4088,84 @@ static int stbi__jpeg_info(stbi__context * s, int * x, int * y, int * comp) { #ifndef STBI_NO_ZLIB // fast-way is faster to check than jpeg huffman, but slow way is slower -#define STBI__ZFAST_BITS 9 // accelerate all cases in default tables -#define STBI__ZFAST_MASK ((1 << STBI__ZFAST_BITS) - 1) +#define STBI__ZFAST_BITS 9 // accelerate all cases in default tables +#define STBI__ZFAST_MASK ((1 << STBI__ZFAST_BITS) - 1) #define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet // zlib-style huffman encoding // (jpegs packs from left, zlib from right, so can't share code) -typedef struct { - stbi__uint16 fast[1 << STBI__ZFAST_BITS]; - stbi__uint16 firstcode[16]; - int maxcode[17]; - stbi__uint16 firstsymbol[16]; - stbi_uc size[STBI__ZNSYMS]; - stbi__uint16 value[STBI__ZNSYMS]; +typedef struct +{ + stbi__uint16 fast[1 << STBI__ZFAST_BITS]; + stbi__uint16 firstcode[16]; + int maxcode[17]; + stbi__uint16 firstsymbol[16]; + stbi_uc size[STBI__ZNSYMS]; + stbi__uint16 value[STBI__ZNSYMS]; } stbi__zhuffman; -stbi_inline static int stbi__bitreverse16(int n) { - n = ((n & 0xAAAA) >> 1) | ((n & 0x5555) << 1); - n = ((n & 0xCCCC) >> 2) | ((n & 0x3333) << 2); - n = ((n & 0xF0F0) >> 4) | ((n & 0x0F0F) << 4); - n = ((n & 0xFF00) >> 8) | ((n & 0x00FF) << 8); - return n; +stbi_inline static int stbi__bitreverse16(int n) +{ + n = ((n & 0xAAAA) >> 1) | ((n & 0x5555) << 1); + n = ((n & 0xCCCC) >> 2) | ((n & 0x3333) << 2); + n = ((n & 0xF0F0) >> 4) | ((n & 0x0F0F) << 4); + n = ((n & 0xFF00) >> 8) | ((n & 0x00FF) << 8); + return n; } -stbi_inline static int stbi__bit_reverse(int v, int bits) { - STBI_ASSERT(bits <= 16); - // to bit reverse n bits, reverse 16 and shift - // e.g. 11 bits, bit reverse and shift away 5 - return stbi__bitreverse16(v) >> (16 - bits); +stbi_inline static int stbi__bit_reverse(int v, int bits) +{ + STBI_ASSERT(bits <= 16); + // to bit reverse n bits, reverse 16 and shift + // e.g. 11 bits, bit reverse and shift away 5 + return stbi__bitreverse16(v) >> (16-bits); } -static int stbi__zbuild_huffman(stbi__zhuffman * z, const stbi_uc * sizelist, int num) { - int i, k = 0; - int code, next_code[16], sizes[17]; +static int stbi__zbuild_huffman(stbi__zhuffman *z, const stbi_uc *sizelist, int num) +{ + int i,k=0; + int code, next_code[16], sizes[17]; - // DEFLATE spec for generating codes - memset(sizes, 0, sizeof(sizes)); - memset(z->fast, 0, sizeof(z->fast)); - for (i = 0; i < num; ++i) - ++sizes[sizelist[i]]; - sizes[0] = 0; - for (i = 1; i < 16; ++i) - if (sizes[i] > (1 << i)) - return stbi__err("bad sizes", "Corrupt PNG"); - code = 0; - for (i = 1; i < 16; ++i) { - next_code[i] = code; - z->firstcode[i] = (stbi__uint16)code; - z->firstsymbol[i] = (stbi__uint16)k; - code = (code + sizes[i]); - if (sizes[i]) - if (code - 1 >= (1 << i)) - return stbi__err("bad codelengths", "Corrupt PNG"); - z->maxcode[i] = code << (16 - i); // preshift for inner loop - code <<= 1; - k += sizes[i]; - } - z->maxcode[16] = 0x10000; // sentinel - for (i = 0; i < num; ++i) { - int s = sizelist[i]; - if (s) { - int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s]; - stbi__uint16 fastv = (stbi__uint16)((s << 9) | i); - z->size[c] = (stbi_uc)s; - z->value[c] = (stbi__uint16)i; - if (s <= STBI__ZFAST_BITS) { - int j = stbi__bit_reverse(next_code[s], s); - while (j < (1 << STBI__ZFAST_BITS)) { - z->fast[j] = fastv; - j += (1 << s); - } + // DEFLATE spec for generating codes + memset(sizes, 0, sizeof(sizes)); + memset(z->fast, 0, sizeof(z->fast)); + for (i=0; i < num; ++i) + ++sizes[sizelist[i]]; + sizes[0] = 0; + for (i=1; i < 16; ++i) + if (sizes[i] > (1 << i)) + return stbi__err("bad sizes", "Corrupt PNG"); + code = 0; + for (i=1; i < 16; ++i) { + next_code[i] = code; + z->firstcode[i] = (stbi__uint16) code; + z->firstsymbol[i] = (stbi__uint16) k; + code = (code + sizes[i]); + if (sizes[i]) + if (code-1 >= (1 << i)) return stbi__err("bad codelengths","Corrupt PNG"); + z->maxcode[i] = code << (16-i); // preshift for inner loop + code <<= 1; + k += sizes[i]; + } + z->maxcode[16] = 0x10000; // sentinel + for (i=0; i < num; ++i) { + int s = sizelist[i]; + if (s) { + int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s]; + stbi__uint16 fastv = (stbi__uint16) ((s << 9) | i); + z->size [c] = (stbi_uc ) s; + z->value[c] = (stbi__uint16) i; + if (s <= STBI__ZFAST_BITS) { + int j = stbi__bit_reverse(next_code[s],s); + while (j < (1 << STBI__ZFAST_BITS)) { + z->fast[j] = fastv; + j += (1 << s); } - ++next_code[s]; - } - } - return 1; + } + ++next_code[s]; + } + } + return 1; } // zlib-from-memory implementation for PNG reading @@ -4361,298 +4174,297 @@ static int stbi__zbuild_huffman(stbi__zhuffman * z, const stbi_uc * sizelist, in // we require PNG read all the IDATs and combine them into a single // memory buffer -typedef struct { - stbi_uc *zbuffer, *zbuffer_end; - int num_bits; - stbi__uint32 code_buffer; +typedef struct +{ + stbi_uc *zbuffer, *zbuffer_end; + int num_bits; + int hit_zeof_once; + stbi__uint32 code_buffer; - char * zout; - char * zout_start; - char * zout_end; - int z_expandable; + char *zout; + char *zout_start; + char *zout_end; + int z_expandable; - stbi__zhuffman z_length, z_distance; + stbi__zhuffman z_length, z_distance; } stbi__zbuf; -stbi_inline static int stbi__zeof(stbi__zbuf * z) { return (z->zbuffer >= z->zbuffer_end); } - -stbi_inline static stbi_uc stbi__zget8(stbi__zbuf * z) { return stbi__zeof(z) ? 0 : *z->zbuffer++; } - -static void stbi__fill_bits(stbi__zbuf * z) { - do { - if (z->code_buffer >= (1U << z->num_bits)) { - z->zbuffer = z->zbuffer_end; /* treat this as EOF so we fail. */ - return; - } - z->code_buffer |= (unsigned int)stbi__zget8(z) << z->num_bits; - z->num_bits += 8; - } while (z->num_bits <= 24); -} - -stbi_inline static unsigned int stbi__zreceive(stbi__zbuf * z, int n) { - unsigned int k; - if (z->num_bits < n) - stbi__fill_bits(z); - k = z->code_buffer & ((1 << n) - 1); - z->code_buffer >>= n; - z->num_bits -= n; - return k; -} - -static int stbi__zhuffman_decode_slowpath(stbi__zbuf * a, stbi__zhuffman * z) { - int b, s, k; - // not resolved by fast table, so compute it the slow way - // use jpeg approach, which requires MSbits at top - k = stbi__bit_reverse(a->code_buffer, 16); - for (s = STBI__ZFAST_BITS + 1;; ++s) - if (k < z->maxcode[s]) - break; - if (s >= 16) - return -1; // invalid code! - // code size is s, so: - b = (k >> (16 - s)) - z->firstcode[s] + z->firstsymbol[s]; - if (b >= STBI__ZNSYMS) - return -1; // some data was corrupt somewhere! - if (z->size[b] != s) - return -1; // was originally an assert, but report failure instead. - a->code_buffer >>= s; - a->num_bits -= s; - return z->value[b]; -} - -stbi_inline static int stbi__zhuffman_decode(stbi__zbuf * a, stbi__zhuffman * z) { - int b, s; - if (a->num_bits < 16) { - if (stbi__zeof(a)) { - return -1; /* report error for unexpected end of data. */ - } - stbi__fill_bits(a); - } - b = z->fast[a->code_buffer & STBI__ZFAST_MASK]; - if (b) { - s = b >> 9; - a->code_buffer >>= s; - a->num_bits -= s; - return b & 511; - } - return stbi__zhuffman_decode_slowpath(a, z); -} - -static int stbi__zexpand(stbi__zbuf * z, char * zout, int n) // need to make room for n bytes +stbi_inline static int stbi__zeof(stbi__zbuf *z) { - char * q; - unsigned int cur, limit, old_limit; - z->zout = zout; - if (!z->z_expandable) - return stbi__err("output buffer limit", "Corrupt PNG"); - cur = (unsigned int)(z->zout - z->zout_start); - limit = old_limit = (unsigned)(z->zout_end - z->zout_start); - if (UINT_MAX - cur < (unsigned)n) - return stbi__err("outofmem", "Out of memory"); - while (cur + n > limit) { - if (limit > UINT_MAX / 2) - return stbi__err("outofmem", "Out of memory"); - limit *= 2; - } - q = (char *)STBI_REALLOC_SIZED(z->zout_start, old_limit, limit); - STBI_NOTUSED(old_limit); - if (q == NULL) - return stbi__err("outofmem", "Out of memory"); - z->zout_start = q; - z->zout = q + cur; - z->zout_end = q + limit; - return 1; + return (z->zbuffer >= z->zbuffer_end); } -static const int stbi__zlength_base[31] = {3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31, - 35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258, 0, 0}; - -static const int stbi__zlength_extra[31] = {0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, - 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0, 0, 0}; - -static const int stbi__zdist_base[32] = {1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, - 49, 65, 97, 129, 193, 257, 385, 513, 769, 1025, 1537, - 2049, 3073, 4097, 6145, 8193, 12289, 16385, 24577, 0, 0}; - -static const int stbi__zdist_extra[32] = {0, 0, 0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, - 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13}; - -static int stbi__parse_huffman_block(stbi__zbuf * a) { - char * zout = a->zout; - for (;;) { - int z = stbi__zhuffman_decode(a, &a->z_length); - if (z < 256) { - if (z < 0) - return stbi__err("bad huffman code", "Corrupt PNG"); // error in huffman codes - if (zout >= a->zout_end) { - if (!stbi__zexpand(a, zout, 1)) - return 0; - zout = a->zout; - } - *zout++ = (char)z; - } else { - stbi_uc * p; - int len, dist; - if (z == 256) { - a->zout = zout; - return 1; - } - if (z >= 286) - return stbi__err("bad huffman code", - "Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data - z -= 257; - len = stbi__zlength_base[z]; - if (stbi__zlength_extra[z]) - len += stbi__zreceive(a, stbi__zlength_extra[z]); - z = stbi__zhuffman_decode(a, &a->z_distance); - if (z < 0 || z >= 30) - return stbi__err("bad huffman code", - "Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data - dist = stbi__zdist_base[z]; - if (stbi__zdist_extra[z]) - dist += stbi__zreceive(a, stbi__zdist_extra[z]); - if (zout - a->zout_start < dist) - return stbi__err("bad dist", "Corrupt PNG"); - if (zout + len > a->zout_end) { - if (!stbi__zexpand(a, zout, len)) - return 0; - zout = a->zout; - } - p = (stbi_uc *)(zout - dist); - if (dist == 1) { // run of one byte; common in images. - stbi_uc v = *p; - if (len) { - do - *zout++ = v; - while (--len); - } - } else { - if (len) { - do - *zout++ = *p++; - while (--len); - } - } - } - } +stbi_inline static stbi_uc stbi__zget8(stbi__zbuf *z) +{ + return stbi__zeof(z) ? 0 : *z->zbuffer++; } -static int stbi__compute_huffman_codes(stbi__zbuf * a) { - static const stbi_uc length_dezigzag[19] = {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}; - stbi__zhuffman z_codelength; - stbi_uc lencodes[286 + 32 + 137]; // padding for maximum single op - stbi_uc codelength_sizes[19]; - int i, n; +static void stbi__fill_bits(stbi__zbuf *z) +{ + do { + if (z->code_buffer >= (1U << z->num_bits)) { + z->zbuffer = z->zbuffer_end; /* treat this as EOF so we fail. */ + return; + } + z->code_buffer |= (unsigned int) stbi__zget8(z) << z->num_bits; + z->num_bits += 8; + } while (z->num_bits <= 24); +} - int hlit = stbi__zreceive(a, 5) + 257; - int hdist = stbi__zreceive(a, 5) + 1; - int hclen = stbi__zreceive(a, 4) + 4; - int ntot = hlit + hdist; +stbi_inline static unsigned int stbi__zreceive(stbi__zbuf *z, int n) +{ + unsigned int k; + if (z->num_bits < n) stbi__fill_bits(z); + k = z->code_buffer & ((1 << n) - 1); + z->code_buffer >>= n; + z->num_bits -= n; + return k; +} - memset(codelength_sizes, 0, sizeof(codelength_sizes)); - for (i = 0; i < hclen; ++i) { - int s = stbi__zreceive(a, 3); - codelength_sizes[length_dezigzag[i]] = (stbi_uc)s; - } - if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) - return 0; +static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z) +{ + int b,s,k; + // not resolved by fast table, so compute it the slow way + // use jpeg approach, which requires MSbits at top + k = stbi__bit_reverse(a->code_buffer, 16); + for (s=STBI__ZFAST_BITS+1; ; ++s) + if (k < z->maxcode[s]) + break; + if (s >= 16) return -1; // invalid code! + // code size is s, so: + b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s]; + if (b >= STBI__ZNSYMS) return -1; // some data was corrupt somewhere! + if (z->size[b] != s) return -1; // was originally an assert, but report failure instead. + a->code_buffer >>= s; + a->num_bits -= s; + return z->value[b]; +} - n = 0; - while (n < ntot) { - int c = stbi__zhuffman_decode(a, &z_codelength); - if (c < 0 || c >= 19) +stbi_inline static int stbi__zhuffman_decode(stbi__zbuf *a, stbi__zhuffman *z) +{ + int b,s; + if (a->num_bits < 16) { + if (stbi__zeof(a)) { + if (!a->hit_zeof_once) { + // This is the first time we hit eof, insert 16 extra padding btis + // to allow us to keep going; if we actually consume any of them + // though, that is invalid data. This is caught later. + a->hit_zeof_once = 1; + a->num_bits += 16; // add 16 implicit zero bits + } else { + // We already inserted our extra 16 padding bits and are again + // out, this stream is actually prematurely terminated. + return -1; + } + } else { + stbi__fill_bits(a); + } + } + b = z->fast[a->code_buffer & STBI__ZFAST_MASK]; + if (b) { + s = b >> 9; + a->code_buffer >>= s; + a->num_bits -= s; + return b & 511; + } + return stbi__zhuffman_decode_slowpath(a, z); +} + +static int stbi__zexpand(stbi__zbuf *z, char *zout, int n) // need to make room for n bytes +{ + char *q; + unsigned int cur, limit, old_limit; + z->zout = zout; + if (!z->z_expandable) return stbi__err("output buffer limit","Corrupt PNG"); + cur = (unsigned int) (z->zout - z->zout_start); + limit = old_limit = (unsigned) (z->zout_end - z->zout_start); + if (UINT_MAX - cur < (unsigned) n) return stbi__err("outofmem", "Out of memory"); + while (cur + n > limit) { + if(limit > UINT_MAX / 2) return stbi__err("outofmem", "Out of memory"); + limit *= 2; + } + q = (char *) STBI_REALLOC_SIZED(z->zout_start, old_limit, limit); + STBI_NOTUSED(old_limit); + if (q == NULL) return stbi__err("outofmem", "Out of memory"); + z->zout_start = q; + z->zout = q + cur; + z->zout_end = q + limit; + return 1; +} + +static const int stbi__zlength_base[31] = { + 3,4,5,6,7,8,9,10,11,13, + 15,17,19,23,27,31,35,43,51,59, + 67,83,99,115,131,163,195,227,258,0,0 }; + +static const int stbi__zlength_extra[31]= +{ 0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0,0,0 }; + +static const int stbi__zdist_base[32] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193, +257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577,0,0}; + +static const int stbi__zdist_extra[32] = +{ 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13}; + +static int stbi__parse_huffman_block(stbi__zbuf *a) +{ + char *zout = a->zout; + for(;;) { + int z = stbi__zhuffman_decode(a, &a->z_length); + if (z < 256) { + if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); // error in huffman codes + if (zout >= a->zout_end) { + if (!stbi__zexpand(a, zout, 1)) return 0; + zout = a->zout; + } + *zout++ = (char) z; + } else { + stbi_uc *p; + int len,dist; + if (z == 256) { + a->zout = zout; + if (a->hit_zeof_once && a->num_bits < 16) { + // The first time we hit zeof, we inserted 16 extra zero bits into our bit + // buffer so the decoder can just do its speculative decoding. But if we + // actually consumed any of those bits (which is the case when num_bits < 16), + // the stream actually read past the end so it is malformed. + return stbi__err("unexpected end","Corrupt PNG"); + } + return 1; + } + if (z >= 286) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data + z -= 257; + len = stbi__zlength_base[z]; + if (stbi__zlength_extra[z]) len += stbi__zreceive(a, stbi__zlength_extra[z]); + z = stbi__zhuffman_decode(a, &a->z_distance); + if (z < 0 || z >= 30) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data + dist = stbi__zdist_base[z]; + if (stbi__zdist_extra[z]) dist += stbi__zreceive(a, stbi__zdist_extra[z]); + if (zout - a->zout_start < dist) return stbi__err("bad dist","Corrupt PNG"); + if (len > a->zout_end - zout) { + if (!stbi__zexpand(a, zout, len)) return 0; + zout = a->zout; + } + p = (stbi_uc *) (zout - dist); + if (dist == 1) { // run of one byte; common in images. + stbi_uc v = *p; + if (len) { do *zout++ = v; while (--len); } + } else { + if (len) { do *zout++ = *p++; while (--len); } + } + } + } +} + +static int stbi__compute_huffman_codes(stbi__zbuf *a) +{ + static const stbi_uc length_dezigzag[19] = { 16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15 }; + stbi__zhuffman z_codelength; + stbi_uc lencodes[286+32+137];//padding for maximum single op + stbi_uc codelength_sizes[19]; + int i,n; + + int hlit = stbi__zreceive(a,5) + 257; + int hdist = stbi__zreceive(a,5) + 1; + int hclen = stbi__zreceive(a,4) + 4; + int ntot = hlit + hdist; + + memset(codelength_sizes, 0, sizeof(codelength_sizes)); + for (i=0; i < hclen; ++i) { + int s = stbi__zreceive(a,3); + codelength_sizes[length_dezigzag[i]] = (stbi_uc) s; + } + if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) return 0; + + n = 0; + while (n < ntot) { + int c = stbi__zhuffman_decode(a, &z_codelength); + if (c < 0 || c >= 19) return stbi__err("bad codelengths", "Corrupt PNG"); + if (c < 16) + lencodes[n++] = (stbi_uc) c; + else { + stbi_uc fill = 0; + if (c == 16) { + c = stbi__zreceive(a,2)+3; + if (n == 0) return stbi__err("bad codelengths", "Corrupt PNG"); + fill = lencodes[n-1]; + } else if (c == 17) { + c = stbi__zreceive(a,3)+3; + } else if (c == 18) { + c = stbi__zreceive(a,7)+11; + } else { return stbi__err("bad codelengths", "Corrupt PNG"); - if (c < 16) - lencodes[n++] = (stbi_uc)c; - else { - stbi_uc fill = 0; - if (c == 16) { - c = stbi__zreceive(a, 2) + 3; - if (n == 0) - return stbi__err("bad codelengths", "Corrupt PNG"); - fill = lencodes[n - 1]; - } else if (c == 17) { - c = stbi__zreceive(a, 3) + 3; - } else if (c == 18) { - c = stbi__zreceive(a, 7) + 11; - } else { - return stbi__err("bad codelengths", "Corrupt PNG"); - } - if (ntot - n < c) - return stbi__err("bad codelengths", "Corrupt PNG"); - memset(lencodes + n, fill, c); - n += c; - } - } - if (n != ntot) - return stbi__err("bad codelengths", "Corrupt PNG"); - if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) - return 0; - if (!stbi__zbuild_huffman(&a->z_distance, lencodes + hlit, hdist)) - return 0; - return 1; + } + if (ntot - n < c) return stbi__err("bad codelengths", "Corrupt PNG"); + memset(lencodes+n, fill, c); + n += c; + } + } + if (n != ntot) return stbi__err("bad codelengths","Corrupt PNG"); + if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) return 0; + if (!stbi__zbuild_huffman(&a->z_distance, lencodes+hlit, hdist)) return 0; + return 1; } -static int stbi__parse_uncompressed_block(stbi__zbuf * a) { - stbi_uc header[4]; - int len, nlen, k; - if (a->num_bits & 7) - stbi__zreceive(a, a->num_bits & 7); // discard - // drain the bit-packed data into header - k = 0; - while (a->num_bits > 0) { - header[k++] = (stbi_uc)(a->code_buffer & 255); // suppress MSVC run-time check - a->code_buffer >>= 8; - a->num_bits -= 8; - } - if (a->num_bits < 0) - return stbi__err("zlib corrupt", "Corrupt PNG"); - // now fill header the normal way - while (k < 4) - header[k++] = stbi__zget8(a); - len = header[1] * 256 + header[0]; - nlen = header[3] * 256 + header[2]; - if (nlen != (len ^ 0xffff)) - return stbi__err("zlib corrupt", "Corrupt PNG"); - if (a->zbuffer + len > a->zbuffer_end) - return stbi__err("read past buffer", "Corrupt PNG"); - if (a->zout + len > a->zout_end) - if (!stbi__zexpand(a, a->zout, len)) - return 0; - memcpy(a->zout, a->zbuffer, len); - a->zbuffer += len; - a->zout += len; - return 1; +static int stbi__parse_uncompressed_block(stbi__zbuf *a) +{ + stbi_uc header[4]; + int len,nlen,k; + if (a->num_bits & 7) + stbi__zreceive(a, a->num_bits & 7); // discard + // drain the bit-packed data into header + k = 0; + while (a->num_bits > 0) { + header[k++] = (stbi_uc) (a->code_buffer & 255); // suppress MSVC run-time check + a->code_buffer >>= 8; + a->num_bits -= 8; + } + if (a->num_bits < 0) return stbi__err("zlib corrupt","Corrupt PNG"); + // now fill header the normal way + while (k < 4) + header[k++] = stbi__zget8(a); + len = header[1] * 256 + header[0]; + nlen = header[3] * 256 + header[2]; + if (nlen != (len ^ 0xffff)) return stbi__err("zlib corrupt","Corrupt PNG"); + if (a->zbuffer + len > a->zbuffer_end) return stbi__err("read past buffer","Corrupt PNG"); + if (a->zout + len > a->zout_end) + if (!stbi__zexpand(a, a->zout, len)) return 0; + memcpy(a->zout, a->zbuffer, len); + a->zbuffer += len; + a->zout += len; + return 1; } -static int stbi__parse_zlib_header(stbi__zbuf * a) { - int cmf = stbi__zget8(a); - int cm = cmf & 15; - /* int cinfo = cmf >> 4; */ - int flg = stbi__zget8(a); - if (stbi__zeof(a)) - return stbi__err("bad zlib header", "Corrupt PNG"); // zlib spec - if ((cmf * 256 + flg) % 31 != 0) - return stbi__err("bad zlib header", "Corrupt PNG"); // zlib spec - if (flg & 32) - return stbi__err("no preset dict", "Corrupt PNG"); // preset dictionary not allowed in png - if (cm != 8) - return stbi__err("bad compression", "Corrupt PNG"); // DEFLATE required for png - // window = 1 << (8 + cinfo)... but who cares, we fully buffer output - return 1; +static int stbi__parse_zlib_header(stbi__zbuf *a) +{ + int cmf = stbi__zget8(a); + int cm = cmf & 15; + /* int cinfo = cmf >> 4; */ + int flg = stbi__zget8(a); + if (stbi__zeof(a)) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec + if ((cmf*256+flg) % 31 != 0) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec + if (flg & 32) return stbi__err("no preset dict","Corrupt PNG"); // preset dictionary not allowed in png + if (cm != 8) return stbi__err("bad compression","Corrupt PNG"); // DEFLATE required for png + // window = 1 << (8 + cinfo)... but who cares, we fully buffer output + return 1; } -static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] = { - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, - 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, - 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, - 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, - 9, 9, 9, 9, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 8, 8}; -static const stbi_uc stbi__zdefault_distance[32] = {5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, - 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5}; +static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] = +{ + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, + 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, + 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7, 7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,8 +}; +static const stbi_uc stbi__zdefault_distance[32] = +{ + 5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5 +}; /* Init algorithm: { @@ -4666,122 +4478,118 @@ Init algorithm: } */ -static int stbi__parse_zlib(stbi__zbuf * a, int parse_header) { - int final, type; - if (parse_header) - if (!stbi__parse_zlib_header(a)) - return 0; - a->num_bits = 0; - a->code_buffer = 0; - do { - final = stbi__zreceive(a, 1); - type = stbi__zreceive(a, 2); - if (type == 0) { - if (!stbi__parse_uncompressed_block(a)) - return 0; - } else if (type == 3) { - return 0; - } else { - if (type == 1) { - // use fixed code lengths - if (!stbi__zbuild_huffman(&a->z_length, stbi__zdefault_length, STBI__ZNSYMS)) - return 0; - if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance, 32)) - return 0; - } else { - if (!stbi__compute_huffman_codes(a)) - return 0; - } - if (!stbi__parse_huffman_block(a)) - return 0; - } - } while (!final); - return 1; +static int stbi__parse_zlib(stbi__zbuf *a, int parse_header) +{ + int final, type; + if (parse_header) + if (!stbi__parse_zlib_header(a)) return 0; + a->num_bits = 0; + a->code_buffer = 0; + a->hit_zeof_once = 0; + do { + final = stbi__zreceive(a,1); + type = stbi__zreceive(a,2); + if (type == 0) { + if (!stbi__parse_uncompressed_block(a)) return 0; + } else if (type == 3) { + return 0; + } else { + if (type == 1) { + // use fixed code lengths + if (!stbi__zbuild_huffman(&a->z_length , stbi__zdefault_length , STBI__ZNSYMS)) return 0; + if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance, 32)) return 0; + } else { + if (!stbi__compute_huffman_codes(a)) return 0; + } + if (!stbi__parse_huffman_block(a)) return 0; + } + } while (!final); + return 1; } -static int stbi__do_zlib(stbi__zbuf * a, char * obuf, int olen, int exp, int parse_header) { - a->zout_start = obuf; - a->zout = obuf; - a->zout_end = obuf + olen; - a->z_expandable = exp; +static int stbi__do_zlib(stbi__zbuf *a, char *obuf, int olen, int exp, int parse_header) +{ + a->zout_start = obuf; + a->zout = obuf; + a->zout_end = obuf + olen; + a->z_expandable = exp; - return stbi__parse_zlib(a, parse_header); + return stbi__parse_zlib(a, parse_header); } -STBIDEF char * stbi_zlib_decode_malloc_guesssize(const char * buffer, int len, int initial_size, int * outlen) { - stbi__zbuf a; - char * p = (char *)stbi__malloc(initial_size); - if (p == NULL) - return NULL; - a.zbuffer = (stbi_uc *)buffer; - a.zbuffer_end = (stbi_uc *)buffer + len; - if (stbi__do_zlib(&a, p, initial_size, 1, 1)) { - if (outlen) - *outlen = (int)(a.zout - a.zout_start); - return a.zout_start; - } else { - STBI_FREE(a.zout_start); - return NULL; - } +STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(initial_size); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer + len; + if (stbi__do_zlib(&a, p, initial_size, 1, 1)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } } -STBIDEF char * stbi_zlib_decode_malloc(char const * buffer, int len, int * outlen) { - return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen); +STBIDEF char *stbi_zlib_decode_malloc(char const *buffer, int len, int *outlen) +{ + return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen); } -STBIDEF char * stbi_zlib_decode_malloc_guesssize_headerflag(const char * buffer, int len, int initial_size, int * outlen, - int parse_header) { - stbi__zbuf a; - char * p = (char *)stbi__malloc(initial_size); - if (p == NULL) - return NULL; - a.zbuffer = (stbi_uc *)buffer; - a.zbuffer_end = (stbi_uc *)buffer + len; - if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) { - if (outlen) - *outlen = (int)(a.zout - a.zout_start); - return a.zout_start; - } else { - STBI_FREE(a.zout_start); - return NULL; - } +STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(initial_size); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer + len; + if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } } -STBIDEF int stbi_zlib_decode_buffer(char * obuffer, int olen, char const * ibuffer, int ilen) { - stbi__zbuf a; - a.zbuffer = (stbi_uc *)ibuffer; - a.zbuffer_end = (stbi_uc *)ibuffer + ilen; - if (stbi__do_zlib(&a, obuffer, olen, 0, 1)) - return (int)(a.zout - a.zout_start); - else - return -1; +STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, char const *ibuffer, int ilen) +{ + stbi__zbuf a; + a.zbuffer = (stbi_uc *) ibuffer; + a.zbuffer_end = (stbi_uc *) ibuffer + ilen; + if (stbi__do_zlib(&a, obuffer, olen, 0, 1)) + return (int) (a.zout - a.zout_start); + else + return -1; } -STBIDEF char * stbi_zlib_decode_noheader_malloc(char const * buffer, int len, int * outlen) { - stbi__zbuf a; - char * p = (char *)stbi__malloc(16384); - if (p == NULL) - return NULL; - a.zbuffer = (stbi_uc *)buffer; - a.zbuffer_end = (stbi_uc *)buffer + len; - if (stbi__do_zlib(&a, p, 16384, 1, 0)) { - if (outlen) - *outlen = (int)(a.zout - a.zout_start); - return a.zout_start; - } else { - STBI_FREE(a.zout_start); - return NULL; - } +STBIDEF char *stbi_zlib_decode_noheader_malloc(char const *buffer, int len, int *outlen) +{ + stbi__zbuf a; + char *p = (char *) stbi__malloc(16384); + if (p == NULL) return NULL; + a.zbuffer = (stbi_uc *) buffer; + a.zbuffer_end = (stbi_uc *) buffer+len; + if (stbi__do_zlib(&a, p, 16384, 1, 0)) { + if (outlen) *outlen = (int) (a.zout - a.zout_start); + return a.zout_start; + } else { + STBI_FREE(a.zout_start); + return NULL; + } } -STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const char * ibuffer, int ilen) { - stbi__zbuf a; - a.zbuffer = (stbi_uc *)ibuffer; - a.zbuffer_end = (stbi_uc *)ibuffer + ilen; - if (stbi__do_zlib(&a, obuffer, olen, 0, 0)) - return (int)(a.zout - a.zout_start); - else - return -1; +STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen) +{ + stbi__zbuf a; + a.zbuffer = (stbi_uc *) ibuffer; + a.zbuffer_end = (stbi_uc *) ibuffer + ilen; + if (stbi__do_zlib(&a, obuffer, olen, 0, 0)) + return (int) (a.zout - a.zout_start); + else + return -1; } #endif @@ -4796,1303 +4604,1131 @@ STBIDEF int stbi_zlib_decode_noheader_buffer(char * obuffer, int olen, const cha // - uses stb_zlib, a PD zlib implementation with fast huffman decoding #ifndef STBI_NO_PNG -typedef struct { - stbi__uint32 length; - stbi__uint32 type; +typedef struct +{ + stbi__uint32 length; + stbi__uint32 type; } stbi__pngchunk; -static stbi__pngchunk stbi__get_chunk_header(stbi__context * s) { - stbi__pngchunk c; - c.length = stbi__get32be(s); - c.type = stbi__get32be(s); - return c; +static stbi__pngchunk stbi__get_chunk_header(stbi__context *s) +{ + stbi__pngchunk c; + c.length = stbi__get32be(s); + c.type = stbi__get32be(s); + return c; } -static int stbi__check_png_header(stbi__context * s) { - static const stbi_uc png_sig[8] = {137, 80, 78, 71, 13, 10, 26, 10}; - int i; - for (i = 0; i < 8; ++i) - if (stbi__get8(s) != png_sig[i]) - return stbi__err("bad png sig", "Not a PNG"); - return 1; +static int stbi__check_png_header(stbi__context *s) +{ + static const stbi_uc png_sig[8] = { 137,80,78,71,13,10,26,10 }; + int i; + for (i=0; i < 8; ++i) + if (stbi__get8(s) != png_sig[i]) return stbi__err("bad png sig","Not a PNG"); + return 1; } -typedef struct { - stbi__context * s; - stbi_uc *idata, *expanded, *out; - int depth; +typedef struct +{ + stbi__context *s; + stbi_uc *idata, *expanded, *out; + int depth; } stbi__png; + enum { - STBI__F_none = 0, - STBI__F_sub = 1, - STBI__F_up = 2, - STBI__F_avg = 3, - STBI__F_paeth = 4, - // synthetic filters used for first scanline to avoid needing a dummy row of 0s - STBI__F_avg_first, - STBI__F_paeth_first + STBI__F_none=0, + STBI__F_sub=1, + STBI__F_up=2, + STBI__F_avg=3, + STBI__F_paeth=4, + // synthetic filter used for first scanline to avoid needing a dummy row of 0s + STBI__F_avg_first }; -static stbi_uc first_row_filter[5] = {STBI__F_none, STBI__F_sub, STBI__F_none, STBI__F_avg_first, STBI__F_paeth_first}; +static stbi_uc first_row_filter[5] = +{ + STBI__F_none, + STBI__F_sub, + STBI__F_none, + STBI__F_avg_first, + STBI__F_sub // Paeth with b=c=0 turns out to be equivalent to sub +}; -static int stbi__paeth(int a, int b, int c) { - int p = a + b - c; - int pa = abs(p - a); - int pb = abs(p - b); - int pc = abs(p - c); - if (pa <= pb && pa <= pc) - return a; - if (pb <= pc) - return b; - return c; +static int stbi__paeth(int a, int b, int c) +{ + // This formulation looks very different from the reference in the PNG spec, but is + // actually equivalent and has favorable data dependencies and admits straightforward + // generation of branch-free code, which helps performance significantly. + int thresh = c*3 - (a + b); + int lo = a < b ? a : b; + int hi = a < b ? b : a; + int t0 = (hi <= thresh) ? lo : c; + int t1 = (thresh <= lo) ? hi : t0; + return t1; } -static const stbi_uc stbi__depth_scale_table[9] = {0, 0xff, 0x55, 0, 0x11, 0, 0, 0, 0x01}; +static const stbi_uc stbi__depth_scale_table[9] = { 0, 0xff, 0x55, 0, 0x11, 0,0,0, 0x01 }; + +// adds an extra all-255 alpha channel +// dest == src is legal +// img_n must be 1 or 3 +static void stbi__create_png_alpha_expand8(stbi_uc *dest, stbi_uc *src, stbi__uint32 x, int img_n) +{ + int i; + // must process data backwards since we allow dest==src + if (img_n == 1) { + for (i=x-1; i >= 0; --i) { + dest[i*2+1] = 255; + dest[i*2+0] = src[i]; + } + } else { + STBI_ASSERT(img_n == 3); + for (i=x-1; i >= 0; --i) { + dest[i*4+3] = 255; + dest[i*4+2] = src[i*3+2]; + dest[i*4+1] = src[i*3+1]; + dest[i*4+0] = src[i*3+0]; + } + } +} // create the png data from post-deflated data -static int stbi__create_png_image_raw(stbi__png * a, stbi_uc * raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, - stbi__uint32 y, int depth, int color) { - int bytes = (depth == 16 ? 2 : 1); - stbi__context * s = a->s; - stbi__uint32 i, j, stride = x * out_n * bytes; - stbi__uint32 img_len, img_width_bytes; - int k; - int img_n = s->img_n; // copy it into a local for later +static int stbi__create_png_image_raw(stbi__png *a, stbi_uc *raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, stbi__uint32 y, int depth, int color) +{ + int bytes = (depth == 16 ? 2 : 1); + stbi__context *s = a->s; + stbi__uint32 i,j,stride = x*out_n*bytes; + stbi__uint32 img_len, img_width_bytes; + stbi_uc *filter_buf; + int all_ok = 1; + int k; + int img_n = s->img_n; // copy it into a local for later - int output_bytes = out_n * bytes; - int filter_bytes = img_n * bytes; - int width = x; + int output_bytes = out_n*bytes; + int filter_bytes = img_n*bytes; + int width = x; - STBI_ASSERT(out_n == s->img_n || out_n == s->img_n + 1); - a->out = (stbi_uc *)stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into - if (!a->out) - return stbi__err("outofmem", "Out of memory"); + STBI_ASSERT(out_n == s->img_n || out_n == s->img_n+1); + a->out = (stbi_uc *) stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into + if (!a->out) return stbi__err("outofmem", "Out of memory"); - if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) - return stbi__err("too large", "Corrupt PNG"); - img_width_bytes = (((img_n * x * depth) + 7) >> 3); - img_len = (img_width_bytes + 1) * y; + // note: error exits here don't need to clean up a->out individually, + // stbi__do_png always does on error. + if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) return stbi__err("too large", "Corrupt PNG"); + img_width_bytes = (((img_n * x * depth) + 7) >> 3); + if (!stbi__mad2sizes_valid(img_width_bytes, y, img_width_bytes)) return stbi__err("too large", "Corrupt PNG"); + img_len = (img_width_bytes + 1) * y; - // we used to check for exact match between raw_len and img_len on non-interlaced PNGs, - // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros), - // so just check for raw_len < img_len always. - if (raw_len < img_len) - return stbi__err("not enough pixels", "Corrupt PNG"); + // we used to check for exact match between raw_len and img_len on non-interlaced PNGs, + // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros), + // so just check for raw_len < img_len always. + if (raw_len < img_len) return stbi__err("not enough pixels","Corrupt PNG"); - for (j = 0; j < y; ++j) { - stbi_uc * cur = a->out + stride * j; - stbi_uc * prior; - int filter = *raw++; + // Allocate two scan lines worth of filter workspace buffer. + filter_buf = (stbi_uc *) stbi__malloc_mad2(img_width_bytes, 2, 0); + if (!filter_buf) return stbi__err("outofmem", "Out of memory"); - if (filter > 4) - return stbi__err("invalid filter", "Corrupt PNG"); + // Filtering for low-bit-depth images + if (depth < 8) { + filter_bytes = 1; + width = img_width_bytes; + } - if (depth < 8) { - if (img_width_bytes > x) - return stbi__err("invalid width", "Corrupt PNG"); - cur += x * out_n - img_width_bytes; // store output to the rightmost img_len bytes, so we can decode in place - filter_bytes = 1; - width = img_width_bytes; - } - prior = cur - stride; // bugfix: need to compute this after 'cur +=' computation above + for (j=0; j < y; ++j) { + // cur/prior filter buffers alternate + stbi_uc *cur = filter_buf + (j & 1)*img_width_bytes; + stbi_uc *prior = filter_buf + (~j & 1)*img_width_bytes; + stbi_uc *dest = a->out + stride*j; + int nk = width * filter_bytes; + int filter = *raw++; - // if first row, use special filter that doesn't sample previous row - if (j == 0) - filter = first_row_filter[filter]; + // check filter type + if (filter > 4) { + all_ok = stbi__err("invalid filter","Corrupt PNG"); + break; + } - // handle first byte explicitly - for (k = 0; k < filter_bytes; ++k) { - switch (filter) { - case STBI__F_none: - cur[k] = raw[k]; - break; - case STBI__F_sub: - cur[k] = raw[k]; - break; - case STBI__F_up: - cur[k] = STBI__BYTECAST(raw[k] + prior[k]); - break; - case STBI__F_avg: - cur[k] = STBI__BYTECAST(raw[k] + (prior[k] >> 1)); - break; - case STBI__F_paeth: - cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(0, prior[k], 0)); - break; - case STBI__F_avg_first: - cur[k] = raw[k]; - break; - case STBI__F_paeth_first: - cur[k] = raw[k]; - break; + // if first row, use special filter that doesn't sample previous row + if (j == 0) filter = first_row_filter[filter]; + + // perform actual filtering + switch (filter) { + case STBI__F_none: + memcpy(cur, raw, nk); + break; + case STBI__F_sub: + memcpy(cur, raw, filter_bytes); + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + cur[k-filter_bytes]); + break; + case STBI__F_up: + for (k = 0; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + prior[k]); + break; + case STBI__F_avg: + for (k = 0; k < filter_bytes; ++k) + cur[k] = STBI__BYTECAST(raw[k] + (prior[k]>>1)); + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k-filter_bytes])>>1)); + break; + case STBI__F_paeth: + for (k = 0; k < filter_bytes; ++k) + cur[k] = STBI__BYTECAST(raw[k] + prior[k]); // prior[k] == stbi__paeth(0,prior[k],0) + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes], prior[k], prior[k-filter_bytes])); + break; + case STBI__F_avg_first: + memcpy(cur, raw, filter_bytes); + for (k = filter_bytes; k < nk; ++k) + cur[k] = STBI__BYTECAST(raw[k] + (cur[k-filter_bytes] >> 1)); + break; + } + + raw += nk; + + // expand decoded bits in cur to dest, also adding an extra alpha channel if desired + if (depth < 8) { + stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range + stbi_uc *in = cur; + stbi_uc *out = dest; + stbi_uc inb = 0; + stbi__uint32 nsmp = x*img_n; + + // expand bits to bytes first + if (depth == 4) { + for (i=0; i < nsmp; ++i) { + if ((i & 1) == 0) inb = *in++; + *out++ = scale * (inb >> 4); + inb <<= 4; } - } - - if (depth == 8) { - if (img_n != out_n) - cur[img_n] = 255; // first pixel - raw += img_n; - cur += out_n; - prior += out_n; - } else if (depth == 16) { - if (img_n != out_n) { - cur[filter_bytes] = 255; // first pixel top byte - cur[filter_bytes + 1] = 255; // first pixel bottom byte + } else if (depth == 2) { + for (i=0; i < nsmp; ++i) { + if ((i & 3) == 0) inb = *in++; + *out++ = scale * (inb >> 6); + inb <<= 2; } - raw += filter_bytes; - cur += output_bytes; - prior += output_bytes; - } else { - raw += 1; - cur += 1; - prior += 1; - } - - // this is a little gross, so that we don't switch per-pixel or per-component - if (depth < 8 || img_n == out_n) { - int nk = (width - 1) * filter_bytes; -#define STBI__CASE(f) \ - case f: \ - for (k = 0; k < nk; ++k) - switch (filter) { - // "none" filter turns into a memcpy here; make that explicit. - case STBI__F_none: - memcpy(cur, raw, nk); - break; - STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k - filter_bytes]); } - break; - STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } - break; - STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k - filter_bytes]) >> 1)); } - break; - STBI__CASE(STBI__F_paeth) { - cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - filter_bytes], prior[k], prior[k - filter_bytes])); - } - break; - STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k - filter_bytes] >> 1)); } - break; - STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - filter_bytes], 0, 0)); } - break; + } else { + STBI_ASSERT(depth == 1); + for (i=0; i < nsmp; ++i) { + if ((i & 7) == 0) inb = *in++; + *out++ = scale * (inb >> 7); + inb <<= 1; } -#undef STBI__CASE - raw += nk; - } else { - STBI_ASSERT(img_n + 1 == out_n); -#define STBI__CASE(f) \ - case f: \ - for (i = x - 1; i >= 1; --i, cur[filter_bytes] = 255, raw += filter_bytes, cur += output_bytes, prior += output_bytes) \ - for (k = 0; k < filter_bytes; ++k) - switch (filter) { - STBI__CASE(STBI__F_none) { cur[k] = raw[k]; } - break; - STBI__CASE(STBI__F_sub) { cur[k] = STBI__BYTECAST(raw[k] + cur[k - output_bytes]); } - break; - STBI__CASE(STBI__F_up) { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } - break; - STBI__CASE(STBI__F_avg) { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k - output_bytes]) >> 1)); } - break; - STBI__CASE(STBI__F_paeth) { - cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - output_bytes], prior[k], prior[k - output_bytes])); - } - break; - STBI__CASE(STBI__F_avg_first) { cur[k] = STBI__BYTECAST(raw[k] + (cur[k - output_bytes] >> 1)); } - break; - STBI__CASE(STBI__F_paeth_first) { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k - output_bytes], 0, 0)); } - break; + } + + // insert alpha=255 values if desired + if (img_n != out_n) + stbi__create_png_alpha_expand8(dest, dest, x, img_n); + } else if (depth == 8) { + if (img_n == out_n) + memcpy(dest, cur, x*img_n); + else + stbi__create_png_alpha_expand8(dest, cur, x, img_n); + } else if (depth == 16) { + // convert the image data from big-endian to platform-native + stbi__uint16 *dest16 = (stbi__uint16*)dest; + stbi__uint32 nsmp = x*img_n; + + if (img_n == out_n) { + for (i = 0; i < nsmp; ++i, ++dest16, cur += 2) + *dest16 = (cur[0] << 8) | cur[1]; + } else { + STBI_ASSERT(img_n+1 == out_n); + if (img_n == 1) { + for (i = 0; i < x; ++i, dest16 += 2, cur += 2) { + dest16[0] = (cur[0] << 8) | cur[1]; + dest16[1] = 0xffff; + } + } else { + STBI_ASSERT(img_n == 3); + for (i = 0; i < x; ++i, dest16 += 4, cur += 6) { + dest16[0] = (cur[0] << 8) | cur[1]; + dest16[1] = (cur[2] << 8) | cur[3]; + dest16[2] = (cur[4] << 8) | cur[5]; + dest16[3] = 0xffff; + } } -#undef STBI__CASE + } + } + } - // the loop above sets the high byte of the pixels' alpha, but for - // 16 bit png files we also need the low byte set. we'll do that here. - if (depth == 16) { - cur = a->out + stride * j; // start at the beginning of the row again - for (i = 0; i < x; ++i, cur += output_bytes) { - cur[filter_bytes + 1] = 255; - } - } - } - } + STBI_FREE(filter_buf); + if (!all_ok) return 0; - // we make a separate pass to expand bits to pixels; for performance, - // this could run two scanlines behind the above code, so it won't - // intefere with filtering but will still be in the cache. - if (depth < 8) { - for (j = 0; j < y; ++j) { - stbi_uc * cur = a->out + stride * j; - stbi_uc * in = a->out + stride * j + x * out_n - img_width_bytes; - // unpack 1/2/4-bit into a 8-bit buffer. allows us to keep the common 8-bit path optimal at minimal cost for - // 1/2/4-bit png guarante byte alignment, if width is not multiple of 8/4/2 we'll decode dummy trailing data that - // will be skipped in the later loop - stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range - - // note that the final byte might overshoot and write more data than desired. - // we can allocate enough data that this never writes out of memory, but it - // could also overwrite the next scanline. can it overwrite non-empty data - // on the next scanline? yes, consider 1-pixel-wide scanlines with 1-bit-per-pixel. - // so we need to explicitly clamp the final ones - - if (depth == 4) { - for (k = x * img_n; k >= 2; k -= 2, ++in) { - *cur++ = scale * ((*in >> 4)); - *cur++ = scale * ((*in) & 0x0f); - } - if (k > 0) - *cur++ = scale * ((*in >> 4)); - } else if (depth == 2) { - for (k = x * img_n; k >= 4; k -= 4, ++in) { - *cur++ = scale * ((*in >> 6)); - *cur++ = scale * ((*in >> 4) & 0x03); - *cur++ = scale * ((*in >> 2) & 0x03); - *cur++ = scale * ((*in) & 0x03); - } - if (k > 0) - *cur++ = scale * ((*in >> 6)); - if (k > 1) - *cur++ = scale * ((*in >> 4) & 0x03); - if (k > 2) - *cur++ = scale * ((*in >> 2) & 0x03); - } else if (depth == 1) { - for (k = x * img_n; k >= 8; k -= 8, ++in) { - *cur++ = scale * ((*in >> 7)); - *cur++ = scale * ((*in >> 6) & 0x01); - *cur++ = scale * ((*in >> 5) & 0x01); - *cur++ = scale * ((*in >> 4) & 0x01); - *cur++ = scale * ((*in >> 3) & 0x01); - *cur++ = scale * ((*in >> 2) & 0x01); - *cur++ = scale * ((*in >> 1) & 0x01); - *cur++ = scale * ((*in) & 0x01); - } - if (k > 0) - *cur++ = scale * ((*in >> 7)); - if (k > 1) - *cur++ = scale * ((*in >> 6) & 0x01); - if (k > 2) - *cur++ = scale * ((*in >> 5) & 0x01); - if (k > 3) - *cur++ = scale * ((*in >> 4) & 0x01); - if (k > 4) - *cur++ = scale * ((*in >> 3) & 0x01); - if (k > 5) - *cur++ = scale * ((*in >> 2) & 0x01); - if (k > 6) - *cur++ = scale * ((*in >> 1) & 0x01); - } - if (img_n != out_n) { - int q; - // insert alpha = 255 - cur = a->out + stride * j; - if (img_n == 1) { - for (q = x - 1; q >= 0; --q) { - cur[q * 2 + 1] = 255; - cur[q * 2 + 0] = cur[q]; - } - } else { - STBI_ASSERT(img_n == 3); - for (q = x - 1; q >= 0; --q) { - cur[q * 4 + 3] = 255; - cur[q * 4 + 2] = cur[q * 3 + 2]; - cur[q * 4 + 1] = cur[q * 3 + 1]; - cur[q * 4 + 0] = cur[q * 3 + 0]; - } - } - } - } - } else if (depth == 16) { - // force the image data from big-endian to platform-native. - // this is done in a separate pass due to the decoding relying - // on the data being untouched, but could probably be done - // per-line during decode if care is taken. - stbi_uc * cur = a->out; - stbi__uint16 * cur16 = (stbi__uint16 *)cur; - - for (i = 0; i < x * y * out_n; ++i, cur16++, cur += 2) { - *cur16 = (cur[0] << 8) | cur[1]; - } - } - - return 1; + return 1; } -static int stbi__create_png_image(stbi__png * a, stbi_uc * image_data, stbi__uint32 image_data_len, int out_n, int depth, - int color, int interlaced) { - int bytes = (depth == 16 ? 2 : 1); - int out_bytes = out_n * bytes; - stbi_uc * final; - int p; - if (!interlaced) - return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color); +static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint32 image_data_len, int out_n, int depth, int color, int interlaced) +{ + int bytes = (depth == 16 ? 2 : 1); + int out_bytes = out_n * bytes; + stbi_uc *final; + int p; + if (!interlaced) + return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color); - // de-interlacing - final = (stbi_uc *)stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0); - if (!final) - return stbi__err("outofmem", "Out of memory"); - for (p = 0; p < 7; ++p) { - int xorig[] = {0, 4, 0, 2, 0, 1, 0}; - int yorig[] = {0, 0, 4, 0, 2, 0, 1}; - int xspc[] = {8, 8, 4, 4, 2, 2, 1}; - int yspc[] = {8, 8, 8, 4, 4, 2, 2}; - int i, j, x, y; - // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1 - x = (a->s->img_x - xorig[p] + xspc[p] - 1) / xspc[p]; - y = (a->s->img_y - yorig[p] + yspc[p] - 1) / yspc[p]; - if (x && y) { - stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y; - if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) { - STBI_FREE(final); - return 0; + // de-interlacing + final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0); + if (!final) return stbi__err("outofmem", "Out of memory"); + for (p=0; p < 7; ++p) { + int xorig[] = { 0,4,0,2,0,1,0 }; + int yorig[] = { 0,0,4,0,2,0,1 }; + int xspc[] = { 8,8,4,4,2,2,1 }; + int yspc[] = { 8,8,8,4,4,2,2 }; + int i,j,x,y; + // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1 + x = (a->s->img_x - xorig[p] + xspc[p]-1) / xspc[p]; + y = (a->s->img_y - yorig[p] + yspc[p]-1) / yspc[p]; + if (x && y) { + stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y; + if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) { + STBI_FREE(final); + return 0; + } + for (j=0; j < y; ++j) { + for (i=0; i < x; ++i) { + int out_y = j*yspc[p]+yorig[p]; + int out_x = i*xspc[p]+xorig[p]; + memcpy(final + out_y*a->s->img_x*out_bytes + out_x*out_bytes, + a->out + (j*x+i)*out_bytes, out_bytes); } - for (j = 0; j < y; ++j) { - for (i = 0; i < x; ++i) { - int out_y = j * yspc[p] + yorig[p]; - int out_x = i * xspc[p] + xorig[p]; - memcpy(final + out_y * a->s->img_x * out_bytes + out_x * out_bytes, a->out + (j * x + i) * out_bytes, - out_bytes); - } - } - STBI_FREE(a->out); - image_data += img_len; - image_data_len -= img_len; - } - } - a->out = final; + } + STBI_FREE(a->out); + image_data += img_len; + image_data_len -= img_len; + } + } + a->out = final; - return 1; + return 1; } -static int stbi__compute_transparency(stbi__png * z, stbi_uc tc[3], int out_n) { - stbi__context * s = z->s; - stbi__uint32 i, pixel_count = s->img_x * s->img_y; - stbi_uc * p = z->out; +static int stbi__compute_transparency(stbi__png *z, stbi_uc tc[3], int out_n) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi_uc *p = z->out; - // compute color-based transparency, assuming we've - // already got 255 as the alpha value in the output - STBI_ASSERT(out_n == 2 || out_n == 4); + // compute color-based transparency, assuming we've + // already got 255 as the alpha value in the output + STBI_ASSERT(out_n == 2 || out_n == 4); - if (out_n == 2) { - for (i = 0; i < pixel_count; ++i) { - p[1] = (p[0] == tc[0] ? 0 : 255); - p += 2; - } - } else { - for (i = 0; i < pixel_count; ++i) { - if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) - p[3] = 0; - p += 4; - } - } - return 1; + if (out_n == 2) { + for (i=0; i < pixel_count; ++i) { + p[1] = (p[0] == tc[0] ? 0 : 255); + p += 2; + } + } else { + for (i=0; i < pixel_count; ++i) { + if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) + p[3] = 0; + p += 4; + } + } + return 1; } -static int stbi__compute_transparency16(stbi__png * z, stbi__uint16 tc[3], int out_n) { - stbi__context * s = z->s; - stbi__uint32 i, pixel_count = s->img_x * s->img_y; - stbi__uint16 * p = (stbi__uint16 *)z->out; +static int stbi__compute_transparency16(stbi__png *z, stbi__uint16 tc[3], int out_n) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi__uint16 *p = (stbi__uint16*) z->out; - // compute color-based transparency, assuming we've - // already got 65535 as the alpha value in the output - STBI_ASSERT(out_n == 2 || out_n == 4); + // compute color-based transparency, assuming we've + // already got 65535 as the alpha value in the output + STBI_ASSERT(out_n == 2 || out_n == 4); - if (out_n == 2) { - for (i = 0; i < pixel_count; ++i) { - p[1] = (p[0] == tc[0] ? 0 : 65535); - p += 2; - } - } else { - for (i = 0; i < pixel_count; ++i) { - if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) - p[3] = 0; - p += 4; - } - } - return 1; + if (out_n == 2) { + for (i = 0; i < pixel_count; ++i) { + p[1] = (p[0] == tc[0] ? 0 : 65535); + p += 2; + } + } else { + for (i = 0; i < pixel_count; ++i) { + if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2]) + p[3] = 0; + p += 4; + } + } + return 1; } -static int stbi__expand_png_palette(stbi__png * a, stbi_uc * palette, int len, int pal_img_n) { - stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y; - stbi_uc *p, *temp_out, *orig = a->out; +static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int pal_img_n) +{ + stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y; + stbi_uc *p, *temp_out, *orig = a->out; - p = (stbi_uc *)stbi__malloc_mad2(pixel_count, pal_img_n, 0); - if (p == NULL) - return stbi__err("outofmem", "Out of memory"); + p = (stbi_uc *) stbi__malloc_mad2(pixel_count, pal_img_n, 0); + if (p == NULL) return stbi__err("outofmem", "Out of memory"); - // between here and free(out) below, exitting would leak - temp_out = p; + // between here and free(out) below, exitting would leak + temp_out = p; - if (pal_img_n == 3) { - for (i = 0; i < pixel_count; ++i) { - int n = orig[i] * 4; - p[0] = palette[n]; - p[1] = palette[n + 1]; - p[2] = palette[n + 2]; - p += 3; - } - } else { - for (i = 0; i < pixel_count; ++i) { - int n = orig[i] * 4; - p[0] = palette[n]; - p[1] = palette[n + 1]; - p[2] = palette[n + 2]; - p[3] = palette[n + 3]; - p += 4; - } - } - STBI_FREE(a->out); - a->out = temp_out; + if (pal_img_n == 3) { + for (i=0; i < pixel_count; ++i) { + int n = orig[i]*4; + p[0] = palette[n ]; + p[1] = palette[n+1]; + p[2] = palette[n+2]; + p += 3; + } + } else { + for (i=0; i < pixel_count; ++i) { + int n = orig[i]*4; + p[0] = palette[n ]; + p[1] = palette[n+1]; + p[2] = palette[n+2]; + p[3] = palette[n+3]; + p += 4; + } + } + STBI_FREE(a->out); + a->out = temp_out; - STBI_NOTUSED(len); + STBI_NOTUSED(len); - return 1; + return 1; } static int stbi__unpremultiply_on_load_global = 0; static int stbi__de_iphone_flag_global = 0; -STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply) { - stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply; +STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply) +{ + stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply; } -STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert) { - stbi__de_iphone_flag_global = flag_true_if_should_convert; +STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert) +{ + stbi__de_iphone_flag_global = flag_true_if_should_convert; } #ifndef STBI_THREAD_LOCAL -#define stbi__unpremultiply_on_load stbi__unpremultiply_on_load_global -#define stbi__de_iphone_flag stbi__de_iphone_flag_global +#define stbi__unpremultiply_on_load stbi__unpremultiply_on_load_global +#define stbi__de_iphone_flag stbi__de_iphone_flag_global #else static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set; static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set; -STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply) { - stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply; - stbi__unpremultiply_on_load_set = 1; +STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply) +{ + stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply; + stbi__unpremultiply_on_load_set = 1; } -STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert) { - stbi__de_iphone_flag_local = flag_true_if_should_convert; - stbi__de_iphone_flag_set = 1; +STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert) +{ + stbi__de_iphone_flag_local = flag_true_if_should_convert; + stbi__de_iphone_flag_set = 1; } -#define stbi__unpremultiply_on_load \ - (stbi__unpremultiply_on_load_set ? stbi__unpremultiply_on_load_local : stbi__unpremultiply_on_load_global) -#define stbi__de_iphone_flag (stbi__de_iphone_flag_set ? stbi__de_iphone_flag_local : stbi__de_iphone_flag_global) +#define stbi__unpremultiply_on_load (stbi__unpremultiply_on_load_set \ + ? stbi__unpremultiply_on_load_local \ + : stbi__unpremultiply_on_load_global) +#define stbi__de_iphone_flag (stbi__de_iphone_flag_set \ + ? stbi__de_iphone_flag_local \ + : stbi__de_iphone_flag_global) #endif // STBI_THREAD_LOCAL -static void stbi__de_iphone(stbi__png * z) { - stbi__context * s = z->s; - stbi__uint32 i, pixel_count = s->img_x * s->img_y; - stbi_uc * p = z->out; +static void stbi__de_iphone(stbi__png *z) +{ + stbi__context *s = z->s; + stbi__uint32 i, pixel_count = s->img_x * s->img_y; + stbi_uc *p = z->out; - if (s->img_out_n == 3) { // convert bgr to rgb - for (i = 0; i < pixel_count; ++i) { + if (s->img_out_n == 3) { // convert bgr to rgb + for (i=0; i < pixel_count; ++i) { + stbi_uc t = p[0]; + p[0] = p[2]; + p[2] = t; + p += 3; + } + } else { + STBI_ASSERT(s->img_out_n == 4); + if (stbi__unpremultiply_on_load) { + // convert bgr to rgb and unpremultiply + for (i=0; i < pixel_count; ++i) { + stbi_uc a = p[3]; + stbi_uc t = p[0]; + if (a) { + stbi_uc half = a / 2; + p[0] = (p[2] * 255 + half) / a; + p[1] = (p[1] * 255 + half) / a; + p[2] = ( t * 255 + half) / a; + } else { + p[0] = p[2]; + p[2] = t; + } + p += 4; + } + } else { + // convert bgr to rgb + for (i=0; i < pixel_count; ++i) { stbi_uc t = p[0]; p[0] = p[2]; p[2] = t; - p += 3; - } - } else { - STBI_ASSERT(s->img_out_n == 4); - if (stbi__unpremultiply_on_load) { - // convert bgr to rgb and unpremultiply - for (i = 0; i < pixel_count; ++i) { - stbi_uc a = p[3]; - stbi_uc t = p[0]; - if (a) { - stbi_uc half = a / 2; - p[0] = (p[2] * 255 + half) / a; - p[1] = (p[1] * 255 + half) / a; - p[2] = (t * 255 + half) / a; - } else { - p[0] = p[2]; - p[2] = t; - } - p += 4; - } - } else { - // convert bgr to rgb - for (i = 0; i < pixel_count; ++i) { - stbi_uc t = p[0]; - p[0] = p[2]; - p[2] = t; - p += 4; - } - } - } + p += 4; + } + } + } } -#define STBI__PNG_TYPE(a, b, c, d) (((unsigned)(a) << 24) + ((unsigned)(b) << 16) + ((unsigned)(c) << 8) + (unsigned)(d)) +#define STBI__PNG_TYPE(a,b,c,d) (((unsigned) (a) << 24) + ((unsigned) (b) << 16) + ((unsigned) (c) << 8) + (unsigned) (d)) -static int stbi__parse_png_file(stbi__png * z, int scan, int req_comp) { - stbi_uc palette[1024], pal_img_n = 0; - stbi_uc has_trans = 0, tc[3] = {0}; - stbi__uint16 tc16[3]; - stbi__uint32 ioff = 0, idata_limit = 0, i, pal_len = 0; - int first = 1, k, interlace = 0, color = 0, is_iphone = 0; - stbi__context * s = z->s; +static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp) +{ + stbi_uc palette[1024], pal_img_n=0; + stbi_uc has_trans=0, tc[3]={0}; + stbi__uint16 tc16[3]; + stbi__uint32 ioff=0, idata_limit=0, i, pal_len=0; + int first=1,k,interlace=0, color=0, is_iphone=0; + stbi__context *s = z->s; - z->expanded = NULL; - z->idata = NULL; - z->out = NULL; + z->expanded = NULL; + z->idata = NULL; + z->out = NULL; - if (!stbi__check_png_header(s)) - return 0; + if (!stbi__check_png_header(s)) return 0; - if (scan == STBI__SCAN_type) - return 1; + if (scan == STBI__SCAN_type) return 1; - for (;;) { - stbi__pngchunk c = stbi__get_chunk_header(s); - switch (c.type) { - case STBI__PNG_TYPE('C', 'g', 'B', 'I'): + for (;;) { + stbi__pngchunk c = stbi__get_chunk_header(s); + switch (c.type) { + case STBI__PNG_TYPE('C','g','B','I'): is_iphone = 1; stbi__skip(s, c.length); break; - case STBI__PNG_TYPE('I', 'H', 'D', 'R'): { - int comp, filter; - if (!first) - return stbi__err("multiple IHDR", "Corrupt PNG"); + case STBI__PNG_TYPE('I','H','D','R'): { + int comp,filter; + if (!first) return stbi__err("multiple IHDR","Corrupt PNG"); first = 0; - if (c.length != 13) - return stbi__err("bad IHDR len", "Corrupt PNG"); + if (c.length != 13) return stbi__err("bad IHDR len","Corrupt PNG"); s->img_x = stbi__get32be(s); s->img_y = stbi__get32be(s); - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - z->depth = stbi__get8(s); - if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16) - return stbi__err("1/2/4/8/16-bit only", "PNG not supported: 1/2/4/8/16-bit only"); - color = stbi__get8(s); - if (color > 6) - return stbi__err("bad ctype", "Corrupt PNG"); - if (color == 3 && z->depth == 16) - return stbi__err("bad ctype", "Corrupt PNG"); - if (color == 3) - pal_img_n = 3; - else if (color & 1) - return stbi__err("bad ctype", "Corrupt PNG"); - comp = stbi__get8(s); - if (comp) - return stbi__err("bad comp method", "Corrupt PNG"); - filter = stbi__get8(s); - if (filter) - return stbi__err("bad filter method", "Corrupt PNG"); - interlace = stbi__get8(s); - if (interlace > 1) - return stbi__err("bad interlace method", "Corrupt PNG"); - if (!s->img_x || !s->img_y) - return stbi__err("0-pixel image", "Corrupt PNG"); + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + z->depth = stbi__get8(s); if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16) return stbi__err("1/2/4/8/16-bit only","PNG not supported: 1/2/4/8/16-bit only"); + color = stbi__get8(s); if (color > 6) return stbi__err("bad ctype","Corrupt PNG"); + if (color == 3 && z->depth == 16) return stbi__err("bad ctype","Corrupt PNG"); + if (color == 3) pal_img_n = 3; else if (color & 1) return stbi__err("bad ctype","Corrupt PNG"); + comp = stbi__get8(s); if (comp) return stbi__err("bad comp method","Corrupt PNG"); + filter= stbi__get8(s); if (filter) return stbi__err("bad filter method","Corrupt PNG"); + interlace = stbi__get8(s); if (interlace>1) return stbi__err("bad interlace method","Corrupt PNG"); + if (!s->img_x || !s->img_y) return stbi__err("0-pixel image","Corrupt PNG"); if (!pal_img_n) { - s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0); - if ((1 << 30) / s->img_x / s->img_n < s->img_y) - return stbi__err("too large", "Image too large to decode"); + s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0); + if ((1 << 30) / s->img_x / s->img_n < s->img_y) return stbi__err("too large", "Image too large to decode"); } else { - // if paletted, then pal_n is our final components, and - // img_n is # components to decompress/filter. - s->img_n = 1; - if ((1 << 30) / s->img_x / 4 < s->img_y) - return stbi__err("too large", "Corrupt PNG"); + // if paletted, then pal_n is our final components, and + // img_n is # components to decompress/filter. + s->img_n = 1; + if ((1 << 30) / s->img_x / 4 < s->img_y) return stbi__err("too large","Corrupt PNG"); } // even with SCAN_header, have to scan to see if we have a tRNS break; - } + } - case STBI__PNG_TYPE('P', 'L', 'T', 'E'): { - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (c.length > 256 * 3) - return stbi__err("invalid PLTE", "Corrupt PNG"); + case STBI__PNG_TYPE('P','L','T','E'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (c.length > 256*3) return stbi__err("invalid PLTE","Corrupt PNG"); pal_len = c.length / 3; - if (pal_len * 3 != c.length) - return stbi__err("invalid PLTE", "Corrupt PNG"); - for (i = 0; i < pal_len; ++i) { - palette[i * 4 + 0] = stbi__get8(s); - palette[i * 4 + 1] = stbi__get8(s); - palette[i * 4 + 2] = stbi__get8(s); - palette[i * 4 + 3] = 255; + if (pal_len * 3 != c.length) return stbi__err("invalid PLTE","Corrupt PNG"); + for (i=0; i < pal_len; ++i) { + palette[i*4+0] = stbi__get8(s); + palette[i*4+1] = stbi__get8(s); + palette[i*4+2] = stbi__get8(s); + palette[i*4+3] = 255; } break; - } + } - case STBI__PNG_TYPE('t', 'R', 'N', 'S'): { - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (z->idata) - return stbi__err("tRNS after IDAT", "Corrupt PNG"); + case STBI__PNG_TYPE('t','R','N','S'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (z->idata) return stbi__err("tRNS after IDAT","Corrupt PNG"); if (pal_img_n) { - if (scan == STBI__SCAN_header) { - s->img_n = 4; - return 1; - } - if (pal_len == 0) - return stbi__err("tRNS before PLTE", "Corrupt PNG"); - if (c.length > pal_len) - return stbi__err("bad tRNS len", "Corrupt PNG"); - pal_img_n = 4; - for (i = 0; i < c.length; ++i) - palette[i * 4 + 3] = stbi__get8(s); + if (scan == STBI__SCAN_header) { s->img_n = 4; return 1; } + if (pal_len == 0) return stbi__err("tRNS before PLTE","Corrupt PNG"); + if (c.length > pal_len) return stbi__err("bad tRNS len","Corrupt PNG"); + pal_img_n = 4; + for (i=0; i < c.length; ++i) + palette[i*4+3] = stbi__get8(s); } else { - if (!(s->img_n & 1)) - return stbi__err("tRNS with alpha", "Corrupt PNG"); - if (c.length != (stbi__uint32)s->img_n * 2) - return stbi__err("bad tRNS len", "Corrupt PNG"); - has_trans = 1; - // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now. - if (scan == STBI__SCAN_header) { - ++s->img_n; - return 1; - } - if (z->depth == 16) { - for (k = 0; k < s->img_n; ++k) - tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is - } else { - for (k = 0; k < s->img_n; ++k) - tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * - stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger - } + if (!(s->img_n & 1)) return stbi__err("tRNS with alpha","Corrupt PNG"); + if (c.length != (stbi__uint32) s->img_n*2) return stbi__err("bad tRNS len","Corrupt PNG"); + has_trans = 1; + // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now. + if (scan == STBI__SCAN_header) { ++s->img_n; return 1; } + if (z->depth == 16) { + for (k = 0; k < s->img_n && k < 3; ++k) // extra loop test to suppress false GCC warning + tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is + } else { + for (k = 0; k < s->img_n && k < 3; ++k) + tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger + } } break; - } + } - case STBI__PNG_TYPE('I', 'D', 'A', 'T'): { - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (pal_img_n && !pal_len) - return stbi__err("no PLTE", "Corrupt PNG"); + case STBI__PNG_TYPE('I','D','A','T'): { + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (pal_img_n && !pal_len) return stbi__err("no PLTE","Corrupt PNG"); if (scan == STBI__SCAN_header) { - // header scan definitely stops at first IDAT - if (pal_img_n) - s->img_n = pal_img_n; - return 1; + // header scan definitely stops at first IDAT + if (pal_img_n) + s->img_n = pal_img_n; + return 1; } - if (c.length > (1u << 30)) - return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes"); - if ((int)(ioff + c.length) < (int)ioff) - return 0; + if (c.length > (1u << 30)) return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes"); + if ((int)(ioff + c.length) < (int)ioff) return 0; if (ioff + c.length > idata_limit) { - stbi__uint32 idata_limit_old = idata_limit; - stbi_uc * p; - if (idata_limit == 0) - idata_limit = c.length > 4096 ? c.length : 4096; - while (ioff + c.length > idata_limit) - idata_limit *= 2; - STBI_NOTUSED(idata_limit_old); - p = (stbi_uc *)STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); - if (p == NULL) - return stbi__err("outofmem", "Out of memory"); - z->idata = p; + stbi__uint32 idata_limit_old = idata_limit; + stbi_uc *p; + if (idata_limit == 0) idata_limit = c.length > 4096 ? c.length : 4096; + while (ioff + c.length > idata_limit) + idata_limit *= 2; + STBI_NOTUSED(idata_limit_old); + p = (stbi_uc *) STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); if (p == NULL) return stbi__err("outofmem", "Out of memory"); + z->idata = p; } - if (!stbi__getn(s, z->idata + ioff, c.length)) - return stbi__err("outofdata", "Corrupt PNG"); + if (!stbi__getn(s, z->idata+ioff,c.length)) return stbi__err("outofdata","Corrupt PNG"); ioff += c.length; break; - } + } - case STBI__PNG_TYPE('I', 'E', 'N', 'D'): { + case STBI__PNG_TYPE('I','E','N','D'): { stbi__uint32 raw_len, bpl; - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); - if (scan != STBI__SCAN_load) - return 1; - if (z->idata == NULL) - return stbi__err("no IDAT", "Corrupt PNG"); + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); + if (scan != STBI__SCAN_load) return 1; + if (z->idata == NULL) return stbi__err("no IDAT","Corrupt PNG"); // initial guess for decoded data size to avoid unnecessary reallocs bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */; - z->expanded = (stbi_uc *)stbi_zlib_decode_malloc_guesssize_headerflag((char *)z->idata, ioff, raw_len, - (int *)&raw_len, !is_iphone); - if (z->expanded == NULL) - return 0; // zlib should set error - STBI_FREE(z->idata); - z->idata = NULL; - if ((req_comp == s->img_n + 1 && req_comp != 3 && !pal_img_n) || has_trans) - s->img_out_n = s->img_n + 1; + z->expanded = (stbi_uc *) stbi_zlib_decode_malloc_guesssize_headerflag((char *) z->idata, ioff, raw_len, (int *) &raw_len, !is_iphone); + if (z->expanded == NULL) return 0; // zlib should set error + STBI_FREE(z->idata); z->idata = NULL; + if ((req_comp == s->img_n+1 && req_comp != 3 && !pal_img_n) || has_trans) + s->img_out_n = s->img_n+1; else - s->img_out_n = s->img_n; - if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) - return 0; + s->img_out_n = s->img_n; + if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) return 0; if (has_trans) { - if (z->depth == 16) { - if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) - return 0; - } else { - if (!stbi__compute_transparency(z, tc, s->img_out_n)) - return 0; - } + if (z->depth == 16) { + if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) return 0; + } else { + if (!stbi__compute_transparency(z, tc, s->img_out_n)) return 0; + } } if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2) - stbi__de_iphone(z); + stbi__de_iphone(z); if (pal_img_n) { - // pal_img_n == 3 or 4 - s->img_n = pal_img_n; // record the actual colors we had - s->img_out_n = pal_img_n; - if (req_comp >= 3) - s->img_out_n = req_comp; - if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n)) - return 0; + // pal_img_n == 3 or 4 + s->img_n = pal_img_n; // record the actual colors we had + s->img_out_n = pal_img_n; + if (req_comp >= 3) s->img_out_n = req_comp; + if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n)) + return 0; } else if (has_trans) { - // non-paletted image with tRNS -> source image has (constant) alpha - ++s->img_n; + // non-paletted image with tRNS -> source image has (constant) alpha + ++s->img_n; } - STBI_FREE(z->expanded); - z->expanded = NULL; + STBI_FREE(z->expanded); z->expanded = NULL; // end of PNG chunk, read and skip CRC stbi__get32be(s); return 1; - } + } - default: + default: // if critical, fail - if (first) - return stbi__err("first not IHDR", "Corrupt PNG"); + if (first) return stbi__err("first not IHDR", "Corrupt PNG"); if ((c.type & (1 << 29)) == 0) { -#ifndef STBI_NO_FAILURE_STRINGS - // not threadsafe - static char invalid_chunk[] = "XXXX PNG chunk not known"; - invalid_chunk[0] = STBI__BYTECAST(c.type >> 24); - invalid_chunk[1] = STBI__BYTECAST(c.type >> 16); - invalid_chunk[2] = STBI__BYTECAST(c.type >> 8); - invalid_chunk[3] = STBI__BYTECAST(c.type >> 0); -#endif - return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type"); + #ifndef STBI_NO_FAILURE_STRINGS + // not threadsafe + static char invalid_chunk[] = "XXXX PNG chunk not known"; + invalid_chunk[0] = STBI__BYTECAST(c.type >> 24); + invalid_chunk[1] = STBI__BYTECAST(c.type >> 16); + invalid_chunk[2] = STBI__BYTECAST(c.type >> 8); + invalid_chunk[3] = STBI__BYTECAST(c.type >> 0); + #endif + return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type"); } stbi__skip(s, c.length); break; - } - // end of PNG chunk, read and skip CRC - stbi__get32be(s); - } + } + // end of PNG chunk, read and skip CRC + stbi__get32be(s); + } } -static void * stbi__do_png(stbi__png * p, int * x, int * y, int * n, int req_comp, stbi__result_info * ri) { - void * result = NULL; - if (req_comp < 0 || req_comp > 4) - return stbi__errpuc("bad req_comp", "Internal error"); - if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) { - if (p->depth <= 8) - ri->bits_per_channel = 8; - else if (p->depth == 16) - ri->bits_per_channel = 16; - else - return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth"); - result = p->out; - p->out = NULL; - if (req_comp && req_comp != p->s->img_out_n) { - if (ri->bits_per_channel == 8) - result = stbi__convert_format((unsigned char *)result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); - else - result = stbi__convert_format16((stbi__uint16 *)result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); - p->s->img_out_n = req_comp; - if (result == NULL) - return result; - } - *x = p->s->img_x; - *y = p->s->img_y; - if (n) - *n = p->s->img_n; - } - STBI_FREE(p->out); - p->out = NULL; - STBI_FREE(p->expanded); - p->expanded = NULL; - STBI_FREE(p->idata); - p->idata = NULL; +static void *stbi__do_png(stbi__png *p, int *x, int *y, int *n, int req_comp, stbi__result_info *ri) +{ + void *result=NULL; + if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error"); + if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) { + if (p->depth <= 8) + ri->bits_per_channel = 8; + else if (p->depth == 16) + ri->bits_per_channel = 16; + else + return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth"); + result = p->out; + p->out = NULL; + if (req_comp && req_comp != p->s->img_out_n) { + if (ri->bits_per_channel == 8) + result = stbi__convert_format((unsigned char *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); + else + result = stbi__convert_format16((stbi__uint16 *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y); + p->s->img_out_n = req_comp; + if (result == NULL) return result; + } + *x = p->s->img_x; + *y = p->s->img_y; + if (n) *n = p->s->img_n; + } + STBI_FREE(p->out); p->out = NULL; + STBI_FREE(p->expanded); p->expanded = NULL; + STBI_FREE(p->idata); p->idata = NULL; - return result; + return result; } -static void * stbi__png_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi__png p; - p.s = s; - return stbi__do_png(&p, x, y, comp, req_comp, ri); +static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi__png p; + p.s = s; + return stbi__do_png(&p, x,y,comp,req_comp, ri); } -static int stbi__png_test(stbi__context * s) { - int r; - r = stbi__check_png_header(s); - stbi__rewind(s); - return r; +static int stbi__png_test(stbi__context *s) +{ + int r; + r = stbi__check_png_header(s); + stbi__rewind(s); + return r; } -static int stbi__png_info_raw(stbi__png * p, int * x, int * y, int * comp) { - if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) { - stbi__rewind(p->s); - return 0; - } - if (x) - *x = p->s->img_x; - if (y) - *y = p->s->img_y; - if (comp) - *comp = p->s->img_n; - return 1; +static int stbi__png_info_raw(stbi__png *p, int *x, int *y, int *comp) +{ + if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) { + stbi__rewind( p->s ); + return 0; + } + if (x) *x = p->s->img_x; + if (y) *y = p->s->img_y; + if (comp) *comp = p->s->img_n; + return 1; } -static int stbi__png_info(stbi__context * s, int * x, int * y, int * comp) { - stbi__png p; - p.s = s; - return stbi__png_info_raw(&p, x, y, comp); +static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp) +{ + stbi__png p; + p.s = s; + return stbi__png_info_raw(&p, x, y, comp); } -static int stbi__png_is16(stbi__context * s) { - stbi__png p; - p.s = s; - if (!stbi__png_info_raw(&p, NULL, NULL, NULL)) - return 0; - if (p.depth != 16) { - stbi__rewind(p.s); - return 0; - } - return 1; +static int stbi__png_is16(stbi__context *s) +{ + stbi__png p; + p.s = s; + if (!stbi__png_info_raw(&p, NULL, NULL, NULL)) + return 0; + if (p.depth != 16) { + stbi__rewind(p.s); + return 0; + } + return 1; } #endif // Microsoft/Windows BMP image #ifndef STBI_NO_BMP -static int stbi__bmp_test_raw(stbi__context * s) { - int r; - int sz; - if (stbi__get8(s) != 'B') - return 0; - if (stbi__get8(s) != 'M') - return 0; - stbi__get32le(s); // discard filesize - stbi__get16le(s); // discard reserved - stbi__get16le(s); // discard reserved - stbi__get32le(s); // discard data offset - sz = stbi__get32le(s); - r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124); - return r; +static int stbi__bmp_test_raw(stbi__context *s) +{ + int r; + int sz; + if (stbi__get8(s) != 'B') return 0; + if (stbi__get8(s) != 'M') return 0; + stbi__get32le(s); // discard filesize + stbi__get16le(s); // discard reserved + stbi__get16le(s); // discard reserved + stbi__get32le(s); // discard data offset + sz = stbi__get32le(s); + r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124); + return r; } -static int stbi__bmp_test(stbi__context * s) { - int r = stbi__bmp_test_raw(s); - stbi__rewind(s); - return r; +static int stbi__bmp_test(stbi__context *s) +{ + int r = stbi__bmp_test_raw(s); + stbi__rewind(s); + return r; } + // returns 0..31 for the highest set bit -static int stbi__high_bit(unsigned int z) { - int n = 0; - if (z == 0) - return -1; - if (z >= 0x10000) { - n += 16; - z >>= 16; - } - if (z >= 0x00100) { - n += 8; - z >>= 8; - } - if (z >= 0x00010) { - n += 4; - z >>= 4; - } - if (z >= 0x00004) { - n += 2; - z >>= 2; - } - if (z >= 0x00002) { - n += 1; /* >>= 1;*/ - } - return n; +static int stbi__high_bit(unsigned int z) +{ + int n=0; + if (z == 0) return -1; + if (z >= 0x10000) { n += 16; z >>= 16; } + if (z >= 0x00100) { n += 8; z >>= 8; } + if (z >= 0x00010) { n += 4; z >>= 4; } + if (z >= 0x00004) { n += 2; z >>= 2; } + if (z >= 0x00002) { n += 1;/* >>= 1;*/ } + return n; } -static int stbi__bitcount(unsigned int a) { - a = (a & 0x55555555) + ((a >> 1) & 0x55555555); // max 2 - a = (a & 0x33333333) + ((a >> 2) & 0x33333333); // max 4 - a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits - a = (a + (a >> 8)); // max 16 per 8 bits - a = (a + (a >> 16)); // max 32 per 8 bits - return a & 0xff; +static int stbi__bitcount(unsigned int a) +{ + a = (a & 0x55555555) + ((a >> 1) & 0x55555555); // max 2 + a = (a & 0x33333333) + ((a >> 2) & 0x33333333); // max 4 + a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits + a = (a + (a >> 8)); // max 16 per 8 bits + a = (a + (a >> 16)); // max 32 per 8 bits + return a & 0xff; } // extract an arbitrarily-aligned N-bit value (N=bits) // from v, and then make it 8-bits long and fractionally // extend it to full full range. -static int stbi__shiftsigned(unsigned int v, int shift, int bits) { - static unsigned int mul_table[9] = { - 0, - 0xff /*0b11111111*/, - 0x55 /*0b01010101*/, - 0x49 /*0b01001001*/, - 0x11 /*0b00010001*/, - 0x21 /*0b00100001*/, - 0x41 /*0b01000001*/, - 0x81 /*0b10000001*/, - 0x01 /*0b00000001*/, - }; - static unsigned int shift_table[9] = { - 0, 0, 0, 1, 0, 2, 4, 6, 0, - }; - if (shift < 0) - v <<= -shift; - else - v >>= shift; - STBI_ASSERT(v < 256); - v >>= (8 - bits); - STBI_ASSERT(bits >= 0 && bits <= 8); - return (int)((unsigned)v * mul_table[bits]) >> shift_table[bits]; +static int stbi__shiftsigned(unsigned int v, int shift, int bits) +{ + static unsigned int mul_table[9] = { + 0, + 0xff/*0b11111111*/, 0x55/*0b01010101*/, 0x49/*0b01001001*/, 0x11/*0b00010001*/, + 0x21/*0b00100001*/, 0x41/*0b01000001*/, 0x81/*0b10000001*/, 0x01/*0b00000001*/, + }; + static unsigned int shift_table[9] = { + 0, 0,0,1,0,2,4,6,0, + }; + if (shift < 0) + v <<= -shift; + else + v >>= shift; + STBI_ASSERT(v < 256); + v >>= (8-bits); + STBI_ASSERT(bits >= 0 && bits <= 8); + return (int) ((unsigned) v * mul_table[bits]) >> shift_table[bits]; } -typedef struct { - int bpp, offset, hsz; - unsigned int mr, mg, mb, ma, all_a; - int extra_read; +typedef struct +{ + int bpp, offset, hsz; + unsigned int mr,mg,mb,ma, all_a; + int extra_read; } stbi__bmp_data; -static int stbi__bmp_set_mask_defaults(stbi__bmp_data * info, int compress) { - // BI_BITFIELDS specifies masks explicitly, don't override - if (compress == 3) - return 1; +static int stbi__bmp_set_mask_defaults(stbi__bmp_data *info, int compress) +{ + // BI_BITFIELDS specifies masks explicitly, don't override + if (compress == 3) + return 1; - if (compress == 0) { - if (info->bpp == 16) { - info->mr = 31u << 10; - info->mg = 31u << 5; - info->mb = 31u << 0; - } else if (info->bpp == 32) { - info->mr = 0xffu << 16; - info->mg = 0xffu << 8; - info->mb = 0xffu << 0; - info->ma = 0xffu << 24; - info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0 - } else { - // otherwise, use defaults, which is all-0 - info->mr = info->mg = info->mb = info->ma = 0; - } - return 1; - } - return 0; // error + if (compress == 0) { + if (info->bpp == 16) { + info->mr = 31u << 10; + info->mg = 31u << 5; + info->mb = 31u << 0; + } else if (info->bpp == 32) { + info->mr = 0xffu << 16; + info->mg = 0xffu << 8; + info->mb = 0xffu << 0; + info->ma = 0xffu << 24; + info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0 + } else { + // otherwise, use defaults, which is all-0 + info->mr = info->mg = info->mb = info->ma = 0; + } + return 1; + } + return 0; // error } -static void * stbi__bmp_parse_header(stbi__context * s, stbi__bmp_data * info) { - int hsz; - if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') - return stbi__errpuc("not BMP", "Corrupt BMP"); - stbi__get32le(s); // discard filesize - stbi__get16le(s); // discard reserved - stbi__get16le(s); // discard reserved - info->offset = stbi__get32le(s); - info->hsz = hsz = stbi__get32le(s); - info->mr = info->mg = info->mb = info->ma = 0; - info->extra_read = 14; +static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info) +{ + int hsz; + if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') return stbi__errpuc("not BMP", "Corrupt BMP"); + stbi__get32le(s); // discard filesize + stbi__get16le(s); // discard reserved + stbi__get16le(s); // discard reserved + info->offset = stbi__get32le(s); + info->hsz = hsz = stbi__get32le(s); + info->mr = info->mg = info->mb = info->ma = 0; + info->extra_read = 14; - if (info->offset < 0) - return stbi__errpuc("bad BMP", "bad BMP"); + if (info->offset < 0) return stbi__errpuc("bad BMP", "bad BMP"); - if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) - return stbi__errpuc("unknown BMP", "BMP type not supported: unknown"); - if (hsz == 12) { - s->img_x = stbi__get16le(s); - s->img_y = stbi__get16le(s); - } else { - s->img_x = stbi__get32le(s); - s->img_y = stbi__get32le(s); - } - if (stbi__get16le(s) != 1) - return stbi__errpuc("bad BMP", "bad BMP"); - info->bpp = stbi__get16le(s); - if (hsz != 12) { - int compress = stbi__get32le(s); - if (compress == 1 || compress == 2) - return stbi__errpuc("BMP RLE", "BMP type not supported: RLE"); - if (compress >= 4) - return stbi__errpuc("BMP JPEG/PNG", - "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes - if (compress == 3 && info->bpp != 16 && info->bpp != 32) - return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel - stbi__get32le(s); // discard sizeof - stbi__get32le(s); // discard hres - stbi__get32le(s); // discard vres - stbi__get32le(s); // discard colorsused - stbi__get32le(s); // discard max important - if (hsz == 40 || hsz == 56) { - if (hsz == 56) { - stbi__get32le(s); - stbi__get32le(s); - stbi__get32le(s); - stbi__get32le(s); - } - if (info->bpp == 16 || info->bpp == 32) { - if (compress == 0) { - stbi__bmp_set_mask_defaults(info, compress); - } else if (compress == 3) { - info->mr = stbi__get32le(s); - info->mg = stbi__get32le(s); - info->mb = stbi__get32le(s); - info->extra_read += 12; - // not documented, but generated by photoshop and handled by mspaint - if (info->mr == info->mg && info->mg == info->mb) { - // ?!?!? - return stbi__errpuc("bad BMP", "bad BMP"); - } - } else - return stbi__errpuc("bad BMP", "bad BMP"); - } - } else { - // V4/V5 header - int i; - if (hsz != 108 && hsz != 124) - return stbi__errpuc("bad BMP", "bad BMP"); - info->mr = stbi__get32le(s); - info->mg = stbi__get32le(s); - info->mb = stbi__get32le(s); - info->ma = stbi__get32le(s); - if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs - stbi__bmp_set_mask_defaults(info, compress); - stbi__get32le(s); // discard color space - for (i = 0; i < 12; ++i) - stbi__get32le(s); // discard color space parameters - if (hsz == 124) { - stbi__get32le(s); // discard rendering intent - stbi__get32le(s); // discard offset of profile data - stbi__get32le(s); // discard size of profile data - stbi__get32le(s); // discard reserved - } - } - } - return (void *)1; + if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) return stbi__errpuc("unknown BMP", "BMP type not supported: unknown"); + if (hsz == 12) { + s->img_x = stbi__get16le(s); + s->img_y = stbi__get16le(s); + } else { + s->img_x = stbi__get32le(s); + s->img_y = stbi__get32le(s); + } + if (stbi__get16le(s) != 1) return stbi__errpuc("bad BMP", "bad BMP"); + info->bpp = stbi__get16le(s); + if (hsz != 12) { + int compress = stbi__get32le(s); + if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE"); + if (compress >= 4) return stbi__errpuc("BMP JPEG/PNG", "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes + if (compress == 3 && info->bpp != 16 && info->bpp != 32) return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel + stbi__get32le(s); // discard sizeof + stbi__get32le(s); // discard hres + stbi__get32le(s); // discard vres + stbi__get32le(s); // discard colorsused + stbi__get32le(s); // discard max important + if (hsz == 40 || hsz == 56) { + if (hsz == 56) { + stbi__get32le(s); + stbi__get32le(s); + stbi__get32le(s); + stbi__get32le(s); + } + if (info->bpp == 16 || info->bpp == 32) { + if (compress == 0) { + stbi__bmp_set_mask_defaults(info, compress); + } else if (compress == 3) { + info->mr = stbi__get32le(s); + info->mg = stbi__get32le(s); + info->mb = stbi__get32le(s); + info->extra_read += 12; + // not documented, but generated by photoshop and handled by mspaint + if (info->mr == info->mg && info->mg == info->mb) { + // ?!?!? + return stbi__errpuc("bad BMP", "bad BMP"); + } + } else + return stbi__errpuc("bad BMP", "bad BMP"); + } + } else { + // V4/V5 header + int i; + if (hsz != 108 && hsz != 124) + return stbi__errpuc("bad BMP", "bad BMP"); + info->mr = stbi__get32le(s); + info->mg = stbi__get32le(s); + info->mb = stbi__get32le(s); + info->ma = stbi__get32le(s); + if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs + stbi__bmp_set_mask_defaults(info, compress); + stbi__get32le(s); // discard color space + for (i=0; i < 12; ++i) + stbi__get32le(s); // discard color space parameters + if (hsz == 124) { + stbi__get32le(s); // discard rendering intent + stbi__get32le(s); // discard offset of profile data + stbi__get32le(s); // discard size of profile data + stbi__get32le(s); // discard reserved + } + } + } + return (void *) 1; } -static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * out; - unsigned int mr = 0, mg = 0, mb = 0, ma = 0, all_a; - stbi_uc pal[256][4]; - int psize = 0, i, j, width; - int flip_vertically, pad, target; - stbi__bmp_data info; - STBI_NOTUSED(ri); - info.all_a = 255; - if (stbi__bmp_parse_header(s, &info) == NULL) - return NULL; // error code already set +static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *out; + unsigned int mr=0,mg=0,mb=0,ma=0, all_a; + stbi_uc pal[256][4]; + int psize=0,i,j,width; + int flip_vertically, pad, target; + stbi__bmp_data info; + STBI_NOTUSED(ri); - flip_vertically = ((int)s->img_y) > 0; - s->img_y = abs((int)s->img_y); + info.all_a = 255; + if (stbi__bmp_parse_header(s, &info) == NULL) + return NULL; // error code already set - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + flip_vertically = ((int) s->img_y) > 0; + s->img_y = abs((int) s->img_y); - mr = info.mr; - mg = info.mg; - mb = info.mb; - ma = info.ma; - all_a = info.all_a; + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - if (info.hsz == 12) { - if (info.bpp < 24) - psize = (info.offset - info.extra_read - 24) / 3; - } else { - if (info.bpp < 16) - psize = (info.offset - info.extra_read - info.hsz) >> 2; - } - if (psize == 0) { - // accept some number of extra bytes after the header, but if the offset points either to before - // the header ends or implies a large amount of extra data, reject the file as malformed - int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original); - int header_limit = 1024; // max we actually read is below 256 bytes currently. - int extra_data_limit = 256 * 4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size. - if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) { - return stbi__errpuc("bad header", "Corrupt BMP"); - } - // we established that bytes_read_so_far is positive and sensible. - // the first half of this test rejects offsets that are either too small positives, or - // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn - // ensures the number computed in the second half of the test can't overflow. - if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) { - return stbi__errpuc("bad offset", "Corrupt BMP"); - } else { - stbi__skip(s, info.offset - bytes_read_so_far); - } - } + mr = info.mr; + mg = info.mg; + mb = info.mb; + ma = info.ma; + all_a = info.all_a; - if (info.bpp == 24 && ma == 0xff000000) - s->img_n = 3; - else - s->img_n = ma ? 4 : 3; - if (req_comp && req_comp >= 3) // we can directly decode 3 or 4 - target = req_comp; - else - target = s->img_n; // if they want monochrome, we'll post-convert + if (info.hsz == 12) { + if (info.bpp < 24) + psize = (info.offset - info.extra_read - 24) / 3; + } else { + if (info.bpp < 16) + psize = (info.offset - info.extra_read - info.hsz) >> 2; + } + if (psize == 0) { + // accept some number of extra bytes after the header, but if the offset points either to before + // the header ends or implies a large amount of extra data, reject the file as malformed + int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original); + int header_limit = 1024; // max we actually read is below 256 bytes currently. + int extra_data_limit = 256*4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size. + if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) { + return stbi__errpuc("bad header", "Corrupt BMP"); + } + // we established that bytes_read_so_far is positive and sensible. + // the first half of this test rejects offsets that are either too small positives, or + // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn + // ensures the number computed in the second half of the test can't overflow. + if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) { + return stbi__errpuc("bad offset", "Corrupt BMP"); + } else { + stbi__skip(s, info.offset - bytes_read_so_far); + } + } - // sanity-check size - if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0)) - return stbi__errpuc("too large", "Corrupt BMP"); + if (info.bpp == 24 && ma == 0xff000000) + s->img_n = 3; + else + s->img_n = ma ? 4 : 3; + if (req_comp && req_comp >= 3) // we can directly decode 3 or 4 + target = req_comp; + else + target = s->img_n; // if they want monochrome, we'll post-convert - out = (stbi_uc *)stbi__malloc_mad3(target, s->img_x, s->img_y, 0); - if (!out) - return stbi__errpuc("outofmem", "Out of memory"); - if (info.bpp < 16) { - int z = 0; - if (psize == 0 || psize > 256) { - STBI_FREE(out); - return stbi__errpuc("invalid", "Corrupt BMP"); - } - for (i = 0; i < psize; ++i) { - pal[i][2] = stbi__get8(s); - pal[i][1] = stbi__get8(s); - pal[i][0] = stbi__get8(s); - if (info.hsz != 12) - stbi__get8(s); - pal[i][3] = 255; - } - stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4)); - if (info.bpp == 1) - width = (s->img_x + 7) >> 3; - else if (info.bpp == 4) - width = (s->img_x + 1) >> 1; - else if (info.bpp == 8) - width = s->img_x; - else { - STBI_FREE(out); - return stbi__errpuc("bad bpp", "Corrupt BMP"); - } - pad = (-width) & 3; - if (info.bpp == 1) { - for (j = 0; j < (int)s->img_y; ++j) { - int bit_offset = 7, v = stbi__get8(s); - for (i = 0; i < (int)s->img_x; ++i) { - int color = (v >> bit_offset) & 0x1; - out[z++] = pal[color][0]; - out[z++] = pal[color][1]; - out[z++] = pal[color][2]; - if (target == 4) - out[z++] = 255; - if (i + 1 == (int)s->img_x) - break; - if ((--bit_offset) < 0) { - bit_offset = 7; - v = stbi__get8(s); - } - } - stbi__skip(s, pad); - } - } else { - for (j = 0; j < (int)s->img_y; ++j) { - for (i = 0; i < (int)s->img_x; i += 2) { - int v = stbi__get8(s), v2 = 0; - if (info.bpp == 4) { - v2 = v & 15; - v >>= 4; - } - out[z++] = pal[v][0]; - out[z++] = pal[v][1]; - out[z++] = pal[v][2]; - if (target == 4) - out[z++] = 255; - if (i + 1 == (int)s->img_x) - break; - v = (info.bpp == 8) ? stbi__get8(s) : v2; - out[z++] = pal[v][0]; - out[z++] = pal[v][1]; - out[z++] = pal[v][2]; - if (target == 4) - out[z++] = 255; - } - stbi__skip(s, pad); - } - } - } else { - int rshift = 0, gshift = 0, bshift = 0, ashift = 0, rcount = 0, gcount = 0, bcount = 0, acount = 0; - int z = 0; - int easy = 0; - stbi__skip(s, info.offset - info.extra_read - info.hsz); - if (info.bpp == 24) - width = 3 * s->img_x; - else if (info.bpp == 16) - width = 2 * s->img_x; - else /* bpp = 32 and pad = 0 */ - width = 0; - pad = (-width) & 3; - if (info.bpp == 24) { - easy = 1; - } else if (info.bpp == 32) { - if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000) - easy = 2; - } - if (!easy) { - if (!mr || !mg || !mb) { - STBI_FREE(out); - return stbi__errpuc("bad masks", "Corrupt BMP"); - } - // right shift amt to put high bit in position #7 - rshift = stbi__high_bit(mr) - 7; - rcount = stbi__bitcount(mr); - gshift = stbi__high_bit(mg) - 7; - gcount = stbi__bitcount(mg); - bshift = stbi__high_bit(mb) - 7; - bcount = stbi__bitcount(mb); - ashift = stbi__high_bit(ma) - 7; - acount = stbi__bitcount(ma); - if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { - STBI_FREE(out); - return stbi__errpuc("bad masks", "Corrupt BMP"); - } - } - for (j = 0; j < (int)s->img_y; ++j) { - if (easy) { - for (i = 0; i < (int)s->img_x; ++i) { - unsigned char a; - out[z + 2] = stbi__get8(s); - out[z + 1] = stbi__get8(s); - out[z + 0] = stbi__get8(s); - z += 3; - a = (easy == 2 ? stbi__get8(s) : 255); - all_a |= a; - if (target == 4) - out[z++] = a; - } - } else { - int bpp = info.bpp; - for (i = 0; i < (int)s->img_x; ++i) { - stbi__uint32 v = (bpp == 16 ? (stbi__uint32)stbi__get16le(s) : stbi__get32le(s)); - unsigned int a; - out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount)); - out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount)); - out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount)); - a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255); - all_a |= a; - if (target == 4) - out[z++] = STBI__BYTECAST(a); - } + // sanity-check size + if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0)) + return stbi__errpuc("too large", "Corrupt BMP"); + + out = (stbi_uc *) stbi__malloc_mad3(target, s->img_x, s->img_y, 0); + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + if (info.bpp < 16) { + int z=0; + if (psize == 0 || psize > 256) { STBI_FREE(out); return stbi__errpuc("invalid", "Corrupt BMP"); } + for (i=0; i < psize; ++i) { + pal[i][2] = stbi__get8(s); + pal[i][1] = stbi__get8(s); + pal[i][0] = stbi__get8(s); + if (info.hsz != 12) stbi__get8(s); + pal[i][3] = 255; + } + stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4)); + if (info.bpp == 1) width = (s->img_x + 7) >> 3; + else if (info.bpp == 4) width = (s->img_x + 1) >> 1; + else if (info.bpp == 8) width = s->img_x; + else { STBI_FREE(out); return stbi__errpuc("bad bpp", "Corrupt BMP"); } + pad = (-width)&3; + if (info.bpp == 1) { + for (j=0; j < (int) s->img_y; ++j) { + int bit_offset = 7, v = stbi__get8(s); + for (i=0; i < (int) s->img_x; ++i) { + int color = (v>>bit_offset)&0x1; + out[z++] = pal[color][0]; + out[z++] = pal[color][1]; + out[z++] = pal[color][2]; + if (target == 4) out[z++] = 255; + if (i+1 == (int) s->img_x) break; + if((--bit_offset) < 0) { + bit_offset = 7; + v = stbi__get8(s); + } } stbi__skip(s, pad); - } - } - - // if alpha channel is all 0s, replace with all 255s - if (target == 4 && all_a == 0) - for (i = 4 * s->img_x * s->img_y - 1; i >= 0; i -= 4) - out[i] = 255; - - if (flip_vertically) { - stbi_uc t; - for (j = 0; j < (int)s->img_y >> 1; ++j) { - stbi_uc * p1 = out + j * s->img_x * target; - stbi_uc * p2 = out + (s->img_y - 1 - j) * s->img_x * target; - for (i = 0; i < (int)s->img_x * target; ++i) { - t = p1[i]; - p1[i] = p2[i]; - p2[i] = t; + } + } else { + for (j=0; j < (int) s->img_y; ++j) { + for (i=0; i < (int) s->img_x; i += 2) { + int v=stbi__get8(s),v2=0; + if (info.bpp == 4) { + v2 = v & 15; + v >>= 4; + } + out[z++] = pal[v][0]; + out[z++] = pal[v][1]; + out[z++] = pal[v][2]; + if (target == 4) out[z++] = 255; + if (i+1 == (int) s->img_x) break; + v = (info.bpp == 8) ? stbi__get8(s) : v2; + out[z++] = pal[v][0]; + out[z++] = pal[v][1]; + out[z++] = pal[v][2]; + if (target == 4) out[z++] = 255; } - } - } + stbi__skip(s, pad); + } + } + } else { + int rshift=0,gshift=0,bshift=0,ashift=0,rcount=0,gcount=0,bcount=0,acount=0; + int z = 0; + int easy=0; + stbi__skip(s, info.offset - info.extra_read - info.hsz); + if (info.bpp == 24) width = 3 * s->img_x; + else if (info.bpp == 16) width = 2*s->img_x; + else /* bpp = 32 and pad = 0 */ width=0; + pad = (-width) & 3; + if (info.bpp == 24) { + easy = 1; + } else if (info.bpp == 32) { + if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000) + easy = 2; + } + if (!easy) { + if (!mr || !mg || !mb) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); } + // right shift amt to put high bit in position #7 + rshift = stbi__high_bit(mr)-7; rcount = stbi__bitcount(mr); + gshift = stbi__high_bit(mg)-7; gcount = stbi__bitcount(mg); + bshift = stbi__high_bit(mb)-7; bcount = stbi__bitcount(mb); + ashift = stbi__high_bit(ma)-7; acount = stbi__bitcount(ma); + if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); } + } + for (j=0; j < (int) s->img_y; ++j) { + if (easy) { + for (i=0; i < (int) s->img_x; ++i) { + unsigned char a; + out[z+2] = stbi__get8(s); + out[z+1] = stbi__get8(s); + out[z+0] = stbi__get8(s); + z += 3; + a = (easy == 2 ? stbi__get8(s) : 255); + all_a |= a; + if (target == 4) out[z++] = a; + } + } else { + int bpp = info.bpp; + for (i=0; i < (int) s->img_x; ++i) { + stbi__uint32 v = (bpp == 16 ? (stbi__uint32) stbi__get16le(s) : stbi__get32le(s)); + unsigned int a; + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount)); + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount)); + out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount)); + a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255); + all_a |= a; + if (target == 4) out[z++] = STBI__BYTECAST(a); + } + } + stbi__skip(s, pad); + } + } - if (req_comp && req_comp != target) { - out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y); - if (out == NULL) - return out; // stbi__convert_format frees input on failure - } + // if alpha channel is all 0s, replace with all 255s + if (target == 4 && all_a == 0) + for (i=4*s->img_x*s->img_y-1; i >= 0; i -= 4) + out[i] = 255; - *x = s->img_x; - *y = s->img_y; - if (comp) - *comp = s->img_n; - return out; + if (flip_vertically) { + stbi_uc t; + for (j=0; j < (int) s->img_y>>1; ++j) { + stbi_uc *p1 = out + j *s->img_x*target; + stbi_uc *p2 = out + (s->img_y-1-j)*s->img_x*target; + for (i=0; i < (int) s->img_x*target; ++i) { + t = p1[i]; p1[i] = p2[i]; p2[i] = t; + } + } + } + + if (req_comp && req_comp != target) { + out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y); + if (out == NULL) return out; // stbi__convert_format frees input on failure + } + + *x = s->img_x; + *y = s->img_y; + if (comp) *comp = s->img_n; + return out; } #endif @@ -6100,74 +5736,68 @@ static void * stbi__bmp_load(stbi__context * s, int * x, int * y, int * comp, in // by Jonathan Dummer #ifndef STBI_NO_TGA // returns STBI_rgb or whatever, 0 on error -static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int * is_rgb16) { - // only RGB or RGBA (incl. 16bit) or grey allowed - if (is_rgb16) - *is_rgb16 = 0; - switch (bits_per_pixel) { - case 8: - return STBI_grey; - case 16: - if (is_grey) - return STBI_grey_alpha; - // fallthrough - case 15: - if (is_rgb16) - *is_rgb16 = 1; - return STBI_rgb; - case 24: // fallthrough - case 32: - return bits_per_pixel / 8; - default: - return 0; - } +static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int* is_rgb16) +{ + // only RGB or RGBA (incl. 16bit) or grey allowed + if (is_rgb16) *is_rgb16 = 0; + switch(bits_per_pixel) { + case 8: return STBI_grey; + case 16: if(is_grey) return STBI_grey_alpha; + // fallthrough + case 15: if(is_rgb16) *is_rgb16 = 1; + return STBI_rgb; + case 24: // fallthrough + case 32: return bits_per_pixel/8; + default: return 0; + } } -static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp) { +static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp) +{ int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp; int sz, tga_colormap_type; - stbi__get8(s); // discard Offset + stbi__get8(s); // discard Offset tga_colormap_type = stbi__get8(s); // colormap type - if (tga_colormap_type > 1) { + if( tga_colormap_type > 1 ) { stbi__rewind(s); - return 0; // only RGB or indexed allowed + return 0; // only RGB or indexed allowed } tga_image_type = stbi__get8(s); // image type - if (tga_colormap_type == 1) { // colormapped (paletted) image + if ( tga_colormap_type == 1 ) { // colormapped (paletted) image if (tga_image_type != 1 && tga_image_type != 9) { stbi__rewind(s); return 0; } - stbi__skip(s, 4); // skip index of first colormap entry and number of entries - sz = stbi__get8(s); // check bits per palette color entry - if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32)) { + stbi__skip(s,4); // skip index of first colormap entry and number of entries + sz = stbi__get8(s); // check bits per palette color entry + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) { stbi__rewind(s); return 0; } - stbi__skip(s, 4); // skip image x and y origin + stbi__skip(s,4); // skip image x and y origin tga_colormap_bpp = sz; } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE - if ((tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11)) { + if ( (tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11) ) { stbi__rewind(s); return 0; // only RGB or grey allowed, +/- RLE } - stbi__skip(s, 9); // skip colormap specification and image x/y origin + stbi__skip(s,9); // skip colormap specification and image x/y origin tga_colormap_bpp = 0; } tga_w = stbi__get16le(s); - if (tga_w < 1) { + if( tga_w < 1 ) { stbi__rewind(s); - return 0; // test width + return 0; // test width } tga_h = stbi__get16le(s); - if (tga_h < 1) { + if( tga_h < 1 ) { stbi__rewind(s); - return 0; // test height + return 0; // test height } tga_bits_per_pixel = stbi__get8(s); // bits per pixel - stbi__get8(s); // ignore alpha bits + stbi__get8(s); // ignore alpha bits if (tga_colormap_bpp != 0) { - if ((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) { + if((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) { // when using a colormap, tga_bits_per_pixel is the size of the indexes // I don't think anything but 8 or 16bit indexes makes sense stbi__rewind(s); @@ -6177,268 +5807,270 @@ static int stbi__tga_info(stbi__context * s, int * x, int * y, int * comp) { } else { tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL); } - if (!tga_comp) { - stbi__rewind(s); - return 0; + if(!tga_comp) { + stbi__rewind(s); + return 0; } - if (x) - *x = tga_w; - if (y) - *y = tga_h; - if (comp) - *comp = tga_comp; - return 1; // seems to have passed everything + if (x) *x = tga_w; + if (y) *y = tga_h; + if (comp) *comp = tga_comp; + return 1; // seems to have passed everything } -static int stbi__tga_test(stbi__context * s) { - int res = 0; - int sz, tga_color_type; - stbi__get8(s); // discard Offset - tga_color_type = stbi__get8(s); // color type - if (tga_color_type > 1) - goto errorEnd; // only RGB or indexed allowed - sz = stbi__get8(s); // image type - if (tga_color_type == 1) { // colormapped (paletted) image - if (sz != 1 && sz != 9) - goto errorEnd; // colortype 1 demands image type 1 or 9 - stbi__skip(s, 4); // skip index of first colormap entry and number of entries - sz = stbi__get8(s); // check bits per palette color entry - if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32)) - goto errorEnd; - stbi__skip(s, 4); // skip image x and y origin - } else { // "normal" image w/o colormap - if ((sz != 2) && (sz != 3) && (sz != 10) && (sz != 11)) - goto errorEnd; // only RGB or grey allowed, +/- RLE - stbi__skip(s, 9); // skip colormap specification and image x/y origin - } - if (stbi__get16le(s) < 1) - goto errorEnd; // test width - if (stbi__get16le(s) < 1) - goto errorEnd; // test height - sz = stbi__get8(s); // bits per pixel - if ((tga_color_type == 1) && (sz != 8) && (sz != 16)) - goto errorEnd; // for colormapped images, bpp is size of an index - if ((sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32)) - goto errorEnd; +static int stbi__tga_test(stbi__context *s) +{ + int res = 0; + int sz, tga_color_type; + stbi__get8(s); // discard Offset + tga_color_type = stbi__get8(s); // color type + if ( tga_color_type > 1 ) goto errorEnd; // only RGB or indexed allowed + sz = stbi__get8(s); // image type + if ( tga_color_type == 1 ) { // colormapped (paletted) image + if (sz != 1 && sz != 9) goto errorEnd; // colortype 1 demands image type 1 or 9 + stbi__skip(s,4); // skip index of first colormap entry and number of entries + sz = stbi__get8(s); // check bits per palette color entry + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd; + stbi__skip(s,4); // skip image x and y origin + } else { // "normal" image w/o colormap + if ( (sz != 2) && (sz != 3) && (sz != 10) && (sz != 11) ) goto errorEnd; // only RGB or grey allowed, +/- RLE + stbi__skip(s,9); // skip colormap specification and image x/y origin + } + if ( stbi__get16le(s) < 1 ) goto errorEnd; // test width + if ( stbi__get16le(s) < 1 ) goto errorEnd; // test height + sz = stbi__get8(s); // bits per pixel + if ( (tga_color_type == 1) && (sz != 8) && (sz != 16) ) goto errorEnd; // for colormapped images, bpp is size of an index + if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd; - res = 1; // if we got this far, everything's good and we can return 1 instead of 0 + res = 1; // if we got this far, everything's good and we can return 1 instead of 0 errorEnd: - stbi__rewind(s); - return res; + stbi__rewind(s); + return res; } // read 16bit value and convert to 24bit RGB -static void stbi__tga_read_rgb16(stbi__context * s, stbi_uc * out) { - stbi__uint16 px = (stbi__uint16)stbi__get16le(s); - stbi__uint16 fiveBitMask = 31; - // we have 3 channels with 5bits each - int r = (px >> 10) & fiveBitMask; - int g = (px >> 5) & fiveBitMask; - int b = px & fiveBitMask; - // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later - out[0] = (stbi_uc)((r * 255) / 31); - out[1] = (stbi_uc)((g * 255) / 31); - out[2] = (stbi_uc)((b * 255) / 31); +static void stbi__tga_read_rgb16(stbi__context *s, stbi_uc* out) +{ + stbi__uint16 px = (stbi__uint16)stbi__get16le(s); + stbi__uint16 fiveBitMask = 31; + // we have 3 channels with 5bits each + int r = (px >> 10) & fiveBitMask; + int g = (px >> 5) & fiveBitMask; + int b = px & fiveBitMask; + // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later + out[0] = (stbi_uc)((r * 255)/31); + out[1] = (stbi_uc)((g * 255)/31); + out[2] = (stbi_uc)((b * 255)/31); - // some people claim that the most significant bit might be used for alpha - // (possibly if an alpha-bit is set in the "image descriptor byte") - // but that only made 16bit test images completely translucent.. - // so let's treat all 15 and 16bit TGAs as RGB with no alpha. + // some people claim that the most significant bit might be used for alpha + // (possibly if an alpha-bit is set in the "image descriptor byte") + // but that only made 16bit test images completely translucent.. + // so let's treat all 15 and 16bit TGAs as RGB with no alpha. } -static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - // read in the TGA header stuff - int tga_offset = stbi__get8(s); - int tga_indexed = stbi__get8(s); - int tga_image_type = stbi__get8(s); - int tga_is_RLE = 0; - int tga_palette_start = stbi__get16le(s); - int tga_palette_len = stbi__get16le(s); - int tga_palette_bits = stbi__get8(s); - int tga_x_origin = stbi__get16le(s); - int tga_y_origin = stbi__get16le(s); - int tga_width = stbi__get16le(s); - int tga_height = stbi__get16le(s); - int tga_bits_per_pixel = stbi__get8(s); - int tga_comp, tga_rgb16 = 0; - int tga_inverted = stbi__get8(s); - // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?) - // image data - unsigned char * tga_data; - unsigned char * tga_palette = NULL; - int i, j; - unsigned char raw_data[4] = {0}; - int RLE_count = 0; - int RLE_repeating = 0; - int read_next_pixel = 1; - STBI_NOTUSED(ri); - STBI_NOTUSED(tga_x_origin); // @TODO - STBI_NOTUSED(tga_y_origin); // @TODO +static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + // read in the TGA header stuff + int tga_offset = stbi__get8(s); + int tga_indexed = stbi__get8(s); + int tga_image_type = stbi__get8(s); + int tga_is_RLE = 0; + int tga_palette_start = stbi__get16le(s); + int tga_palette_len = stbi__get16le(s); + int tga_palette_bits = stbi__get8(s); + int tga_x_origin = stbi__get16le(s); + int tga_y_origin = stbi__get16le(s); + int tga_width = stbi__get16le(s); + int tga_height = stbi__get16le(s); + int tga_bits_per_pixel = stbi__get8(s); + int tga_comp, tga_rgb16=0; + int tga_inverted = stbi__get8(s); + // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?) + // image data + unsigned char *tga_data; + unsigned char *tga_palette = NULL; + int i, j; + unsigned char raw_data[4] = {0}; + int RLE_count = 0; + int RLE_repeating = 0; + int read_next_pixel = 1; + STBI_NOTUSED(ri); + STBI_NOTUSED(tga_x_origin); // @TODO + STBI_NOTUSED(tga_y_origin); // @TODO - if (tga_height > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (tga_width > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (tga_height > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (tga_width > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - // do a tiny bit of precessing - if (tga_image_type >= 8) { - tga_image_type -= 8; - tga_is_RLE = 1; - } - tga_inverted = 1 - ((tga_inverted >> 5) & 1); + // do a tiny bit of precessing + if ( tga_image_type >= 8 ) + { + tga_image_type -= 8; + tga_is_RLE = 1; + } + tga_inverted = 1 - ((tga_inverted >> 5) & 1); - // If I'm paletted, then I'll use the number of bits from the palette - if (tga_indexed) - tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16); - else - tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16); + // If I'm paletted, then I'll use the number of bits from the palette + if ( tga_indexed ) tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16); + else tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16); - if (!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency - return stbi__errpuc("bad format", "Can't find out TGA pixelformat"); + if(!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency + return stbi__errpuc("bad format", "Can't find out TGA pixelformat"); - // tga info - *x = tga_width; - *y = tga_height; - if (comp) - *comp = tga_comp; + // tga info + *x = tga_width; + *y = tga_height; + if (comp) *comp = tga_comp; - if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0)) - return stbi__errpuc("too large", "Corrupt TGA"); + if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0)) + return stbi__errpuc("too large", "Corrupt TGA"); - tga_data = (unsigned char *)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0); - if (!tga_data) - return stbi__errpuc("outofmem", "Out of memory"); + tga_data = (unsigned char*)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0); + if (!tga_data) return stbi__errpuc("outofmem", "Out of memory"); - // skip to the data's starting position (offset usually = 0) - stbi__skip(s, tga_offset); + // skip to the data's starting position (offset usually = 0) + stbi__skip(s, tga_offset ); - if (!tga_indexed && !tga_is_RLE && !tga_rgb16) { - for (i = 0; i < tga_height; ++i) { - int row = tga_inverted ? tga_height - i - 1 : i; - stbi_uc * tga_row = tga_data + row * tga_width * tga_comp; - stbi__getn(s, tga_row, tga_width * tga_comp); - } - } else { - // do I need to load a palette? - if (tga_indexed) { - if (tga_palette_len == 0) { /* you have to have at least one entry! */ - STBI_FREE(tga_data); - return stbi__errpuc("bad palette", "Corrupt TGA"); + if ( !tga_indexed && !tga_is_RLE && !tga_rgb16 ) { + for (i=0; i < tga_height; ++i) { + int row = tga_inverted ? tga_height -i - 1 : i; + stbi_uc *tga_row = tga_data + row*tga_width*tga_comp; + stbi__getn(s, tga_row, tga_width * tga_comp); + } + } else { + // do I need to load a palette? + if ( tga_indexed) + { + if (tga_palette_len == 0) { /* you have to have at least one entry! */ + STBI_FREE(tga_data); + return stbi__errpuc("bad palette", "Corrupt TGA"); + } + + // any data to skip? (offset usually = 0) + stbi__skip(s, tga_palette_start ); + // load the palette + tga_palette = (unsigned char*)stbi__malloc_mad2(tga_palette_len, tga_comp, 0); + if (!tga_palette) { + STBI_FREE(tga_data); + return stbi__errpuc("outofmem", "Out of memory"); + } + if (tga_rgb16) { + stbi_uc *pal_entry = tga_palette; + STBI_ASSERT(tga_comp == STBI_rgb); + for (i=0; i < tga_palette_len; ++i) { + stbi__tga_read_rgb16(s, pal_entry); + pal_entry += tga_comp; } - - // any data to skip? (offset usually = 0) - stbi__skip(s, tga_palette_start); - // load the palette - tga_palette = (unsigned char *)stbi__malloc_mad2(tga_palette_len, tga_comp, 0); - if (!tga_palette) { - STBI_FREE(tga_data); - return stbi__errpuc("outofmem", "Out of memory"); + } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) { + STBI_FREE(tga_data); + STBI_FREE(tga_palette); + return stbi__errpuc("bad palette", "Corrupt TGA"); + } + } + // load the data + for (i=0; i < tga_width * tga_height; ++i) + { + // if I'm in RLE mode, do I need to get a RLE stbi__pngchunk? + if ( tga_is_RLE ) + { + if ( RLE_count == 0 ) + { + // yep, get the next byte as a RLE command + int RLE_cmd = stbi__get8(s); + RLE_count = 1 + (RLE_cmd & 127); + RLE_repeating = RLE_cmd >> 7; + read_next_pixel = 1; + } else if ( !RLE_repeating ) + { + read_next_pixel = 1; } - if (tga_rgb16) { - stbi_uc * pal_entry = tga_palette; - STBI_ASSERT(tga_comp == STBI_rgb); - for (i = 0; i < tga_palette_len; ++i) { - stbi__tga_read_rgb16(s, pal_entry); - pal_entry += tga_comp; - } - } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) { - STBI_FREE(tga_data); - STBI_FREE(tga_palette); - return stbi__errpuc("bad palette", "Corrupt TGA"); - } - } - // load the data - for (i = 0; i < tga_width * tga_height; ++i) { - // if I'm in RLE mode, do I need to get a RLE stbi__pngchunk? - if (tga_is_RLE) { - if (RLE_count == 0) { - // yep, get the next byte as a RLE command - int RLE_cmd = stbi__get8(s); - RLE_count = 1 + (RLE_cmd & 127); - RLE_repeating = RLE_cmd >> 7; - read_next_pixel = 1; - } else if (!RLE_repeating) { - read_next_pixel = 1; - } + } else + { + read_next_pixel = 1; + } + // OK, if I need to read a pixel, do it now + if ( read_next_pixel ) + { + // load however much data we did have + if ( tga_indexed ) + { + // read in index, then perform the lookup + int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s); + if ( pal_idx >= tga_palette_len ) { + // invalid index + pal_idx = 0; + } + pal_idx *= tga_comp; + for (j = 0; j < tga_comp; ++j) { + raw_data[j] = tga_palette[pal_idx+j]; + } + } else if(tga_rgb16) { + STBI_ASSERT(tga_comp == STBI_rgb); + stbi__tga_read_rgb16(s, raw_data); } else { - read_next_pixel = 1; + // read in the data raw + for (j = 0; j < tga_comp; ++j) { + raw_data[j] = stbi__get8(s); + } } - // OK, if I need to read a pixel, do it now - if (read_next_pixel) { - // load however much data we did have - if (tga_indexed) { - // read in index, then perform the lookup - int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s); - if (pal_idx >= tga_palette_len) { - // invalid index - pal_idx = 0; - } - pal_idx *= tga_comp; - for (j = 0; j < tga_comp; ++j) { - raw_data[j] = tga_palette[pal_idx + j]; - } - } else if (tga_rgb16) { - STBI_ASSERT(tga_comp == STBI_rgb); - stbi__tga_read_rgb16(s, raw_data); - } else { - // read in the data raw - for (j = 0; j < tga_comp; ++j) { - raw_data[j] = stbi__get8(s); - } - } - // clear the reading flag for the next pixel - read_next_pixel = 0; - } // end of reading a pixel + // clear the reading flag for the next pixel + read_next_pixel = 0; + } // end of reading a pixel - // copy data - for (j = 0; j < tga_comp; ++j) - tga_data[i * tga_comp + j] = raw_data[j]; + // copy data + for (j = 0; j < tga_comp; ++j) + tga_data[i*tga_comp+j] = raw_data[j]; - // in case we're in RLE mode, keep counting down - --RLE_count; - } - // do I need to invert the image? - if (tga_inverted) { - for (j = 0; j * 2 < tga_height; ++j) { - int index1 = j * tga_width * tga_comp; - int index2 = (tga_height - 1 - j) * tga_width * tga_comp; - for (i = tga_width * tga_comp; i > 0; --i) { - unsigned char temp = tga_data[index1]; - tga_data[index1] = tga_data[index2]; - tga_data[index2] = temp; - ++index1; - ++index2; - } + // in case we're in RLE mode, keep counting down + --RLE_count; + } + // do I need to invert the image? + if ( tga_inverted ) + { + for (j = 0; j*2 < tga_height; ++j) + { + int index1 = j * tga_width * tga_comp; + int index2 = (tga_height - 1 - j) * tga_width * tga_comp; + for (i = tga_width * tga_comp; i > 0; --i) + { + unsigned char temp = tga_data[index1]; + tga_data[index1] = tga_data[index2]; + tga_data[index2] = temp; + ++index1; + ++index2; } - } - // clear my palette, if I had one - if (tga_palette != NULL) { - STBI_FREE(tga_palette); - } - } + } + } + // clear my palette, if I had one + if ( tga_palette != NULL ) + { + STBI_FREE( tga_palette ); + } + } - // swap RGB - if the source data was RGB16, it already is in the right order - if (tga_comp >= 3 && !tga_rgb16) { - unsigned char * tga_pixel = tga_data; - for (i = 0; i < tga_width * tga_height; ++i) { - unsigned char temp = tga_pixel[0]; - tga_pixel[0] = tga_pixel[2]; - tga_pixel[2] = temp; - tga_pixel += tga_comp; - } - } + // swap RGB - if the source data was RGB16, it already is in the right order + if (tga_comp >= 3 && !tga_rgb16) + { + unsigned char* tga_pixel = tga_data; + for (i=0; i < tga_width * tga_height; ++i) + { + unsigned char temp = tga_pixel[0]; + tga_pixel[0] = tga_pixel[2]; + tga_pixel[2] = temp; + tga_pixel += tga_comp; + } + } - // convert to target component count - if (req_comp && req_comp != tga_comp) - tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height); + // convert to target component count + if (req_comp && req_comp != tga_comp) + tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height); - // the things I do to get rid of an error message, and yet keep - // Microsoft's C compilers happy... [8^( - tga_palette_start = tga_palette_len = tga_palette_bits = tga_x_origin = tga_y_origin = 0; - STBI_NOTUSED(tga_palette_start); - // OK, done - return tga_data; + // the things I do to get rid of an error message, and yet keep + // Microsoft's C compilers happy... [8^( + tga_palette_start = tga_palette_len = tga_palette_bits = + tga_x_origin = tga_y_origin = 0; + STBI_NOTUSED(tga_palette_start); + // OK, done + return tga_data; } #endif @@ -6446,253 +6078,250 @@ static void * stbi__tga_load(stbi__context * s, int * x, int * y, int * comp, in // Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB #ifndef STBI_NO_PSD -static int stbi__psd_test(stbi__context * s) { - int r = (stbi__get32be(s) == 0x38425053); - stbi__rewind(s); - return r; +static int stbi__psd_test(stbi__context *s) +{ + int r = (stbi__get32be(s) == 0x38425053); + stbi__rewind(s); + return r; } -static int stbi__psd_decode_rle(stbi__context * s, stbi_uc * p, int pixelCount) { - int count, nleft, len; +static int stbi__psd_decode_rle(stbi__context *s, stbi_uc *p, int pixelCount) +{ + int count, nleft, len; - count = 0; - while ((nleft = pixelCount - count) > 0) { - len = stbi__get8(s); - if (len == 128) { - // No-op. - } else if (len < 128) { - // Copy next len+1 bytes literally. - len++; - if (len > nleft) - return 0; // corrupt data - count += len; - while (len) { - *p = stbi__get8(s); - p += 4; - len--; - } - } else if (len > 128) { - stbi_uc val; - // Next -len+1 bytes in the dest are replicated from next source byte. - // (Interpret len as a negative 8-bit int.) - len = 257 - len; - if (len > nleft) - return 0; // corrupt data - val = stbi__get8(s); - count += len; - while (len) { - *p = val; - p += 4; - len--; - } - } - } + count = 0; + while ((nleft = pixelCount - count) > 0) { + len = stbi__get8(s); + if (len == 128) { + // No-op. + } else if (len < 128) { + // Copy next len+1 bytes literally. + len++; + if (len > nleft) return 0; // corrupt data + count += len; + while (len) { + *p = stbi__get8(s); + p += 4; + len--; + } + } else if (len > 128) { + stbi_uc val; + // Next -len+1 bytes in the dest are replicated from next source byte. + // (Interpret len as a negative 8-bit int.) + len = 257 - len; + if (len > nleft) return 0; // corrupt data + val = stbi__get8(s); + count += len; + while (len) { + *p = val; + p += 4; + len--; + } + } + } - return 1; + return 1; } -static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri, int bpc) { - int pixelCount; - int channelCount, compression; - int channel, i; - int bitdepth; - int w, h; - stbi_uc * out; - STBI_NOTUSED(ri); +static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc) +{ + int pixelCount; + int channelCount, compression; + int channel, i; + int bitdepth; + int w,h; + stbi_uc *out; + STBI_NOTUSED(ri); - // Check identifier - if (stbi__get32be(s) != 0x38425053) // "8BPS" - return stbi__errpuc("not PSD", "Corrupt PSD image"); + // Check identifier + if (stbi__get32be(s) != 0x38425053) // "8BPS" + return stbi__errpuc("not PSD", "Corrupt PSD image"); - // Check file type version. - if (stbi__get16be(s) != 1) - return stbi__errpuc("wrong version", "Unsupported version of PSD image"); + // Check file type version. + if (stbi__get16be(s) != 1) + return stbi__errpuc("wrong version", "Unsupported version of PSD image"); - // Skip 6 reserved bytes. - stbi__skip(s, 6); + // Skip 6 reserved bytes. + stbi__skip(s, 6 ); - // Read the number of channels (R, G, B, A, etc). - channelCount = stbi__get16be(s); - if (channelCount < 0 || channelCount > 16) - return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image"); + // Read the number of channels (R, G, B, A, etc). + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) + return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image"); - // Read the rows and columns of the image. - h = stbi__get32be(s); - w = stbi__get32be(s); + // Read the rows and columns of the image. + h = stbi__get32be(s); + w = stbi__get32be(s); - if (h > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (w > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (h > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (w > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - // Make sure the depth is 8 bits. - bitdepth = stbi__get16be(s); - if (bitdepth != 8 && bitdepth != 16) - return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit"); + // Make sure the depth is 8 bits. + bitdepth = stbi__get16be(s); + if (bitdepth != 8 && bitdepth != 16) + return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit"); - // Make sure the color mode is RGB. - // Valid options are: - // 0: Bitmap - // 1: Grayscale - // 2: Indexed color - // 3: RGB color - // 4: CMYK color - // 7: Multichannel - // 8: Duotone - // 9: Lab color - if (stbi__get16be(s) != 3) - return stbi__errpuc("wrong color format", "PSD is not in RGB color format"); + // Make sure the color mode is RGB. + // Valid options are: + // 0: Bitmap + // 1: Grayscale + // 2: Indexed color + // 3: RGB color + // 4: CMYK color + // 7: Multichannel + // 8: Duotone + // 9: Lab color + if (stbi__get16be(s) != 3) + return stbi__errpuc("wrong color format", "PSD is not in RGB color format"); - // Skip the Mode Data. (It's the palette for indexed color; other info for other modes.) - stbi__skip(s, stbi__get32be(s)); + // Skip the Mode Data. (It's the palette for indexed color; other info for other modes.) + stbi__skip(s,stbi__get32be(s) ); - // Skip the image resources. (resolution, pen tool paths, etc) - stbi__skip(s, stbi__get32be(s)); + // Skip the image resources. (resolution, pen tool paths, etc) + stbi__skip(s, stbi__get32be(s) ); - // Skip the reserved data. - stbi__skip(s, stbi__get32be(s)); + // Skip the reserved data. + stbi__skip(s, stbi__get32be(s) ); - // Find out if the data is compressed. - // Known values: - // 0: no compression - // 1: RLE compressed - compression = stbi__get16be(s); - if (compression > 1) - return stbi__errpuc("bad compression", "PSD has an unknown compression format"); + // Find out if the data is compressed. + // Known values: + // 0: no compression + // 1: RLE compressed + compression = stbi__get16be(s); + if (compression > 1) + return stbi__errpuc("bad compression", "PSD has an unknown compression format"); - // Check size - if (!stbi__mad3sizes_valid(4, w, h, 0)) - return stbi__errpuc("too large", "Corrupt PSD"); + // Check size + if (!stbi__mad3sizes_valid(4, w, h, 0)) + return stbi__errpuc("too large", "Corrupt PSD"); - // Create the destination image. + // Create the destination image. - if (!compression && bitdepth == 16 && bpc == 16) { - out = (stbi_uc *)stbi__malloc_mad3(8, w, h, 0); - ri->bits_per_channel = 16; - } else - out = (stbi_uc *)stbi__malloc(4 * w * h); + if (!compression && bitdepth == 16 && bpc == 16) { + out = (stbi_uc *) stbi__malloc_mad3(8, w, h, 0); + ri->bits_per_channel = 16; + } else + out = (stbi_uc *) stbi__malloc(4 * w*h); - if (!out) - return stbi__errpuc("outofmem", "Out of memory"); - pixelCount = w * h; + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + pixelCount = w*h; - // Initialize the data to zero. - // memset( out, 0, pixelCount * 4 ); + // Initialize the data to zero. + //memset( out, 0, pixelCount * 4 ); - // Finally, the image data. - if (compression) { - // RLE as used by .PSD and .TIFF - // Loop until you get the number of unpacked bytes you are expecting: - // Read the next source byte into n. - // If n is between 0 and 127 inclusive, copy the next n+1 bytes literally. - // Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times. - // Else if n is 128, noop. - // Endloop + // Finally, the image data. + if (compression) { + // RLE as used by .PSD and .TIFF + // Loop until you get the number of unpacked bytes you are expecting: + // Read the next source byte into n. + // If n is between 0 and 127 inclusive, copy the next n+1 bytes literally. + // Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times. + // Else if n is 128, noop. + // Endloop - // The RLE-compressed data is preceded by a 2-byte data count for each row in the data, - // which we're going to just skip. - stbi__skip(s, h * channelCount * 2); + // The RLE-compressed data is preceded by a 2-byte data count for each row in the data, + // which we're going to just skip. + stbi__skip(s, h * channelCount * 2 ); - // Read the RLE data by channel. - for (channel = 0; channel < 4; channel++) { - stbi_uc * p; + // Read the RLE data by channel. + for (channel = 0; channel < 4; channel++) { + stbi_uc *p; - p = out + channel; - if (channel >= channelCount) { - // Fill this channel with default data. - for (i = 0; i < pixelCount; i++, p += 4) - *p = (channel == 3 ? 255 : 0); + p = out+channel; + if (channel >= channelCount) { + // Fill this channel with default data. + for (i = 0; i < pixelCount; i++, p += 4) + *p = (channel == 3 ? 255 : 0); + } else { + // Read the RLE data. + if (!stbi__psd_decode_rle(s, p, pixelCount)) { + STBI_FREE(out); + return stbi__errpuc("corrupt", "bad RLE data"); + } + } + } + + } else { + // We're at the raw image data. It's each channel in order (Red, Green, Blue, Alpha, ...) + // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image. + + // Read the data by channel. + for (channel = 0; channel < 4; channel++) { + if (channel >= channelCount) { + // Fill this channel with default data. + if (bitdepth == 16 && bpc == 16) { + stbi__uint16 *q = ((stbi__uint16 *) out) + channel; + stbi__uint16 val = channel == 3 ? 65535 : 0; + for (i = 0; i < pixelCount; i++, q += 4) + *q = val; } else { - // Read the RLE data. - if (!stbi__psd_decode_rle(s, p, pixelCount)) { - STBI_FREE(out); - return stbi__errpuc("corrupt", "bad RLE data"); - } + stbi_uc *p = out+channel; + stbi_uc val = channel == 3 ? 255 : 0; + for (i = 0; i < pixelCount; i++, p += 4) + *p = val; } - } - } else { - // We're at the raw image data. It's each channel in order (Red, Green, Blue, Alpha, ...) - // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image. - - // Read the data by channel. - for (channel = 0; channel < 4; channel++) { - if (channel >= channelCount) { - // Fill this channel with default data. - if (bitdepth == 16 && bpc == 16) { - stbi__uint16 * q = ((stbi__uint16 *)out) + channel; - stbi__uint16 val = channel == 3 ? 65535 : 0; - for (i = 0; i < pixelCount; i++, q += 4) - *q = val; - } else { - stbi_uc * p = out + channel; - stbi_uc val = channel == 3 ? 255 : 0; - for (i = 0; i < pixelCount; i++, p += 4) - *p = val; - } + } else { + if (ri->bits_per_channel == 16) { // output bpc + stbi__uint16 *q = ((stbi__uint16 *) out) + channel; + for (i = 0; i < pixelCount; i++, q += 4) + *q = (stbi__uint16) stbi__get16be(s); } else { - if (ri->bits_per_channel == 16) { // output bpc - stbi__uint16 * q = ((stbi__uint16 *)out) + channel; - for (i = 0; i < pixelCount; i++, q += 4) - *q = (stbi__uint16)stbi__get16be(s); - } else { - stbi_uc * p = out + channel; - if (bitdepth == 16) { // input bpc - for (i = 0; i < pixelCount; i++, p += 4) - *p = (stbi_uc)(stbi__get16be(s) >> 8); - } else { - for (i = 0; i < pixelCount; i++, p += 4) - *p = stbi__get8(s); - } - } + stbi_uc *p = out+channel; + if (bitdepth == 16) { // input bpc + for (i = 0; i < pixelCount; i++, p += 4) + *p = (stbi_uc) (stbi__get16be(s) >> 8); + } else { + for (i = 0; i < pixelCount; i++, p += 4) + *p = stbi__get8(s); + } } - } - } + } + } + } - // remove weird white matte from PSD - if (channelCount >= 4) { - if (ri->bits_per_channel == 16) { - for (i = 0; i < w * h; ++i) { - stbi__uint16 * pixel = (stbi__uint16 *)out + 4 * i; - if (pixel[3] != 0 && pixel[3] != 65535) { - float a = pixel[3] / 65535.0f; - float ra = 1.0f / a; - float inv_a = 65535.0f * (1 - ra); - pixel[0] = (stbi__uint16)(pixel[0] * ra + inv_a); - pixel[1] = (stbi__uint16)(pixel[1] * ra + inv_a); - pixel[2] = (stbi__uint16)(pixel[2] * ra + inv_a); - } + // remove weird white matte from PSD + if (channelCount >= 4) { + if (ri->bits_per_channel == 16) { + for (i=0; i < w*h; ++i) { + stbi__uint16 *pixel = (stbi__uint16 *) out + 4*i; + if (pixel[3] != 0 && pixel[3] != 65535) { + float a = pixel[3] / 65535.0f; + float ra = 1.0f / a; + float inv_a = 65535.0f * (1 - ra); + pixel[0] = (stbi__uint16) (pixel[0]*ra + inv_a); + pixel[1] = (stbi__uint16) (pixel[1]*ra + inv_a); + pixel[2] = (stbi__uint16) (pixel[2]*ra + inv_a); } - } else { - for (i = 0; i < w * h; ++i) { - unsigned char * pixel = out + 4 * i; - if (pixel[3] != 0 && pixel[3] != 255) { - float a = pixel[3] / 255.0f; - float ra = 1.0f / a; - float inv_a = 255.0f * (1 - ra); - pixel[0] = (unsigned char)(pixel[0] * ra + inv_a); - pixel[1] = (unsigned char)(pixel[1] * ra + inv_a); - pixel[2] = (unsigned char)(pixel[2] * ra + inv_a); - } + } + } else { + for (i=0; i < w*h; ++i) { + unsigned char *pixel = out + 4*i; + if (pixel[3] != 0 && pixel[3] != 255) { + float a = pixel[3] / 255.0f; + float ra = 1.0f / a; + float inv_a = 255.0f * (1 - ra); + pixel[0] = (unsigned char) (pixel[0]*ra + inv_a); + pixel[1] = (unsigned char) (pixel[1]*ra + inv_a); + pixel[2] = (unsigned char) (pixel[2]*ra + inv_a); } - } - } + } + } + } - // convert to desired output format - if (req_comp && req_comp != 4) { - if (ri->bits_per_channel == 16) - out = (stbi_uc *)stbi__convert_format16((stbi__uint16 *)out, 4, req_comp, w, h); - else - out = stbi__convert_format(out, 4, req_comp, w, h); - if (out == NULL) - return out; // stbi__convert_format frees input on failure - } + // convert to desired output format + if (req_comp && req_comp != 4) { + if (ri->bits_per_channel == 16) + out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, 4, req_comp, w, h); + else + out = stbi__convert_format(out, 4, req_comp, w, h); + if (out == NULL) return out; // stbi__convert_format frees input on failure + } - if (comp) - *comp = 4; - *y = h; - *x = w; + if (comp) *comp = 4; + *y = h; + *x = w; - return out; + return out; } #endif @@ -6704,221 +6333,216 @@ static void * stbi__psd_load(stbi__context * s, int * x, int * y, int * comp, in // See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/ #ifndef STBI_NO_PIC -static int stbi__pic_is4(stbi__context * s, const char * str) { - int i; - for (i = 0; i < 4; ++i) - if (stbi__get8(s) != (stbi_uc)str[i]) - return 0; +static int stbi__pic_is4(stbi__context *s,const char *str) +{ + int i; + for (i=0; i<4; ++i) + if (stbi__get8(s) != (stbi_uc)str[i]) + return 0; - return 1; + return 1; } -static int stbi__pic_test_core(stbi__context * s) { - int i; +static int stbi__pic_test_core(stbi__context *s) +{ + int i; - if (!stbi__pic_is4(s, "\x53\x80\xF6\x34")) - return 0; + if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) + return 0; - for (i = 0; i < 84; ++i) - stbi__get8(s); + for(i=0;i<84;++i) + stbi__get8(s); - if (!stbi__pic_is4(s, "PICT")) - return 0; + if (!stbi__pic_is4(s,"PICT")) + return 0; - return 1; + return 1; } -typedef struct { - stbi_uc size, type, channel; +typedef struct +{ + stbi_uc size,type,channel; } stbi__pic_packet; -static stbi_uc * stbi__readval(stbi__context * s, int channel, stbi_uc * dest) { - int mask = 0x80, i; +static stbi_uc *stbi__readval(stbi__context *s, int channel, stbi_uc *dest) +{ + int mask=0x80, i; - for (i = 0; i < 4; ++i, mask >>= 1) { - if (channel & mask) { - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "PIC file too short"); - dest[i] = stbi__get8(s); - } - } + for (i=0; i<4; ++i, mask>>=1) { + if (channel & mask) { + if (stbi__at_eof(s)) return stbi__errpuc("bad file","PIC file too short"); + dest[i]=stbi__get8(s); + } + } - return dest; + return dest; } -static void stbi__copyval(int channel, stbi_uc * dest, const stbi_uc * src) { - int mask = 0x80, i; +static void stbi__copyval(int channel,stbi_uc *dest,const stbi_uc *src) +{ + int mask=0x80,i; - for (i = 0; i < 4; ++i, mask >>= 1) - if (channel & mask) - dest[i] = src[i]; + for (i=0;i<4; ++i, mask>>=1) + if (channel&mask) + dest[i]=src[i]; } -static stbi_uc * stbi__pic_load_core(stbi__context * s, int width, int height, int * comp, stbi_uc * result) { - int act_comp = 0, num_packets = 0, y, chained; - stbi__pic_packet packets[10]; +static stbi_uc *stbi__pic_load_core(stbi__context *s,int width,int height,int *comp, stbi_uc *result) +{ + int act_comp=0,num_packets=0,y,chained; + stbi__pic_packet packets[10]; - // this will (should...) cater for even some bizarre stuff like having data + // this will (should...) cater for even some bizarre stuff like having data // for the same channel in multiple packets. - do { - stbi__pic_packet * packet; + do { + stbi__pic_packet *packet; - if (num_packets == sizeof(packets) / sizeof(packets[0])) - return stbi__errpuc("bad format", "too many packets"); + if (num_packets==sizeof(packets)/sizeof(packets[0])) + return stbi__errpuc("bad format","too many packets"); - packet = &packets[num_packets++]; + packet = &packets[num_packets++]; - chained = stbi__get8(s); - packet->size = stbi__get8(s); - packet->type = stbi__get8(s); - packet->channel = stbi__get8(s); + chained = stbi__get8(s); + packet->size = stbi__get8(s); + packet->type = stbi__get8(s); + packet->channel = stbi__get8(s); - act_comp |= packet->channel; + act_comp |= packet->channel; - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (reading packets)"); - if (packet->size != 8) - return stbi__errpuc("bad format", "packet isn't 8bpp"); - } while (chained); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (reading packets)"); + if (packet->size != 8) return stbi__errpuc("bad format","packet isn't 8bpp"); + } while (chained); - *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel? + *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel? - for (y = 0; y < height; ++y) { - int packet_idx; + for(y=0; ytype) { + switch (packet->type) { default: - return stbi__errpuc("bad format", "packet has bad compression type"); + return stbi__errpuc("bad format","packet has bad compression type"); - case 0: { // uncompressed - int x; + case 0: {//uncompressed + int x; - for (x = 0; x < width; ++x, dest += 4) - if (!stbi__readval(s, packet->channel, dest)) - return 0; - break; + for(x=0;xchannel,dest)) + return 0; + break; } - case 1: // Pure RLE - { - int left = width, i; + case 1://Pure RLE + { + int left=width, i; - while (left > 0) { - stbi_uc count, value[4]; + while (left>0) { + stbi_uc count,value[4]; - count = stbi__get8(s); - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (pure read count)"); + count=stbi__get8(s); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (pure read count)"); - if (count > left) - count = (stbi_uc)left; + if (count > left) + count = (stbi_uc) left; - if (!stbi__readval(s, packet->channel, value)) + if (!stbi__readval(s,packet->channel,value)) return 0; + + for(i=0; ichannel,dest,value); + left -= count; + } + } + break; + + case 2: {//Mixed RLE + int left=width; + while (left>0) { + int count = stbi__get8(s), i; + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (mixed read count)"); + + if (count >= 128) { // Repeated + stbi_uc value[4]; + + if (count==128) + count = stbi__get16be(s); + else + count -= 127; + if (count > left) + return stbi__errpuc("bad file","scanline overrun"); + + if (!stbi__readval(s,packet->channel,value)) return 0; - for (i = 0; i < count; ++i, dest += 4) - stbi__copyval(packet->channel, dest, value); - left -= count; - } - } break; + for(i=0;ichannel,dest,value); + } else { // Raw + ++count; + if (count>left) return stbi__errpuc("bad file","scanline overrun"); - case 2: { // Mixed RLE - int left = width; - while (left > 0) { - int count = stbi__get8(s), i; - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (mixed read count)"); - - if (count >= 128) { // Repeated - stbi_uc value[4]; - - if (count == 128) - count = stbi__get16be(s); - else - count -= 127; - if (count > left) - return stbi__errpuc("bad file", "scanline overrun"); - - if (!stbi__readval(s, packet->channel, value)) - return 0; - - for (i = 0; i < count; ++i, dest += 4) - stbi__copyval(packet->channel, dest, value); - } else { // Raw - ++count; - if (count > left) - return stbi__errpuc("bad file", "scanline overrun"); - - for (i = 0; i < count; ++i, dest += 4) - if (!stbi__readval(s, packet->channel, dest)) - return 0; - } - left -= count; - } - break; + for(i=0;ichannel,dest)) + return 0; + } + left-=count; + } + break; } - } - } - } + } + } + } - return result; + return result; } -static void * stbi__pic_load(stbi__context * s, int * px, int * py, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * result; - int i, x, y, internal_comp; - STBI_NOTUSED(ri); +static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_comp, stbi__result_info *ri) +{ + stbi_uc *result; + int i, x,y, internal_comp; + STBI_NOTUSED(ri); - if (!comp) - comp = &internal_comp; + if (!comp) comp = &internal_comp; - for (i = 0; i < 92; ++i) - stbi__get8(s); + for (i=0; i<92; ++i) + stbi__get8(s); - x = stbi__get16be(s); - y = stbi__get16be(s); + x = stbi__get16be(s); + y = stbi__get16be(s); - if (y > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (x > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - if (stbi__at_eof(s)) - return stbi__errpuc("bad file", "file too short (pic header)"); - if (!stbi__mad3sizes_valid(x, y, 4, 0)) - return stbi__errpuc("too large", "PIC image too large to decode"); + if (stbi__at_eof(s)) return stbi__errpuc("bad file","file too short (pic header)"); + if (!stbi__mad3sizes_valid(x, y, 4, 0)) return stbi__errpuc("too large", "PIC image too large to decode"); - stbi__get32be(s); // skip `ratio' - stbi__get16be(s); // skip `fields' - stbi__get16be(s); // skip `pad' + stbi__get32be(s); //skip `ratio' + stbi__get16be(s); //skip `fields' + stbi__get16be(s); //skip `pad' - // intermediate buffer is RGBA - result = (stbi_uc *)stbi__malloc_mad3(x, y, 4, 0); - if (!result) - return stbi__errpuc("outofmem", "Out of memory"); - memset(result, 0xff, x * y * 4); + // intermediate buffer is RGBA + result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0); + if (!result) return stbi__errpuc("outofmem", "Out of memory"); + memset(result, 0xff, x*y*4); - if (!stbi__pic_load_core(s, x, y, comp, result)) { - STBI_FREE(result); - result = 0; - } - *px = x; - *py = y; - if (req_comp == 0) - req_comp = *comp; - result = stbi__convert_format(result, 4, req_comp, x, y); + if (!stbi__pic_load_core(s,x,y,comp, result)) { + STBI_FREE(result); + result=0; + } + *px = x; + *py = y; + if (req_comp == 0) req_comp = *comp; + result=stbi__convert_format(result,4,req_comp,x,y); - return result; + return result; } -static int stbi__pic_test(stbi__context * s) { - int r = stbi__pic_test_core(s); - stbi__rewind(s); - return r; +static int stbi__pic_test(stbi__context *s) +{ + int r = stbi__pic_test_core(s); + stbi__rewind(s); + return r; } #endif @@ -6926,968 +6550,931 @@ static int stbi__pic_test(stbi__context * s) { // GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb #ifndef STBI_NO_GIF -typedef struct { - stbi__int16 prefix; - stbi_uc first; - stbi_uc suffix; +typedef struct +{ + stbi__int16 prefix; + stbi_uc first; + stbi_uc suffix; } stbi__gif_lzw; -typedef struct { - int w, h; - stbi_uc * out; // output buffer (always 4 components) - stbi_uc * background; // The current "background" as far as a gif is concerned - stbi_uc * history; - int flags, bgindex, ratio, transparent, eflags; - stbi_uc pal[256][4]; - stbi_uc lpal[256][4]; - stbi__gif_lzw codes[8192]; - stbi_uc * color_table; - int parse, step; - int lflags; - int start_x, start_y; - int max_x, max_y; - int cur_x, cur_y; - int line_size; - int delay; +typedef struct +{ + int w,h; + stbi_uc *out; // output buffer (always 4 components) + stbi_uc *background; // The current "background" as far as a gif is concerned + stbi_uc *history; + int flags, bgindex, ratio, transparent, eflags; + stbi_uc pal[256][4]; + stbi_uc lpal[256][4]; + stbi__gif_lzw codes[8192]; + stbi_uc *color_table; + int parse, step; + int lflags; + int start_x, start_y; + int max_x, max_y; + int cur_x, cur_y; + int line_size; + int delay; } stbi__gif; -static int stbi__gif_test_raw(stbi__context * s) { - int sz; - if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') - return 0; - sz = stbi__get8(s); - if (sz != '9' && sz != '7') - return 0; - if (stbi__get8(s) != 'a') - return 0; - return 1; +static int stbi__gif_test_raw(stbi__context *s) +{ + int sz; + if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') return 0; + sz = stbi__get8(s); + if (sz != '9' && sz != '7') return 0; + if (stbi__get8(s) != 'a') return 0; + return 1; } -static int stbi__gif_test(stbi__context * s) { - int r = stbi__gif_test_raw(s); - stbi__rewind(s); - return r; +static int stbi__gif_test(stbi__context *s) +{ + int r = stbi__gif_test_raw(s); + stbi__rewind(s); + return r; } -static void stbi__gif_parse_colortable(stbi__context * s, stbi_uc pal[256][4], int num_entries, int transp) { - int i; - for (i = 0; i < num_entries; ++i) { - pal[i][2] = stbi__get8(s); - pal[i][1] = stbi__get8(s); - pal[i][0] = stbi__get8(s); - pal[i][3] = transp == i ? 0 : 255; - } +static void stbi__gif_parse_colortable(stbi__context *s, stbi_uc pal[256][4], int num_entries, int transp) +{ + int i; + for (i=0; i < num_entries; ++i) { + pal[i][2] = stbi__get8(s); + pal[i][1] = stbi__get8(s); + pal[i][0] = stbi__get8(s); + pal[i][3] = transp == i ? 0 : 255; + } } -static int stbi__gif_header(stbi__context * s, stbi__gif * g, int * comp, int is_info) { - stbi_uc version; - if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') - return stbi__err("not GIF", "Corrupt GIF"); +static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_info) +{ + stbi_uc version; + if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') + return stbi__err("not GIF", "Corrupt GIF"); - version = stbi__get8(s); - if (version != '7' && version != '9') - return stbi__err("not GIF", "Corrupt GIF"); - if (stbi__get8(s) != 'a') - return stbi__err("not GIF", "Corrupt GIF"); + version = stbi__get8(s); + if (version != '7' && version != '9') return stbi__err("not GIF", "Corrupt GIF"); + if (stbi__get8(s) != 'a') return stbi__err("not GIF", "Corrupt GIF"); - stbi__g_failure_reason = ""; - g->w = stbi__get16le(s); - g->h = stbi__get16le(s); - g->flags = stbi__get8(s); - g->bgindex = stbi__get8(s); - g->ratio = stbi__get8(s); - g->transparent = -1; + stbi__g_failure_reason = ""; + g->w = stbi__get16le(s); + g->h = stbi__get16le(s); + g->flags = stbi__get8(s); + g->bgindex = stbi__get8(s); + g->ratio = stbi__get8(s); + g->transparent = -1; - if (g->w > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); - if (g->h > STBI_MAX_DIMENSIONS) - return stbi__err("too large", "Very large image (corrupt?)"); + if (g->w > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); + if (g->h > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)"); - if (comp != 0) - *comp = 4; // can't actually tell whether it's 3 or 4 until we parse the comments + if (comp != 0) *comp = 4; // can't actually tell whether it's 3 or 4 until we parse the comments - if (is_info) - return 1; + if (is_info) return 1; - if (g->flags & 0x80) - stbi__gif_parse_colortable(s, g->pal, 2 << (g->flags & 7), -1); + if (g->flags & 0x80) + stbi__gif_parse_colortable(s,g->pal, 2 << (g->flags & 7), -1); - return 1; + return 1; } -static int stbi__gif_info_raw(stbi__context * s, int * x, int * y, int * comp) { - stbi__gif * g = (stbi__gif *)stbi__malloc(sizeof(stbi__gif)); - if (!g) - return stbi__err("outofmem", "Out of memory"); - if (!stbi__gif_header(s, g, comp, 1)) { - STBI_FREE(g); - stbi__rewind(s); - return 0; - } - if (x) - *x = g->w; - if (y) - *y = g->h; - STBI_FREE(g); - return 1; +static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp) +{ + stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif)); + if (!g) return stbi__err("outofmem", "Out of memory"); + if (!stbi__gif_header(s, g, comp, 1)) { + STBI_FREE(g); + stbi__rewind( s ); + return 0; + } + if (x) *x = g->w; + if (y) *y = g->h; + STBI_FREE(g); + return 1; } -static void stbi__out_gif_code(stbi__gif * g, stbi__uint16 code) { - stbi_uc *p, *c; - int idx; +static void stbi__out_gif_code(stbi__gif *g, stbi__uint16 code) +{ + stbi_uc *p, *c; + int idx; - // recurse to decode the prefixes, since the linked-list is backwards, - // and working backwards through an interleaved image would be nasty - if (g->codes[code].prefix >= 0) - stbi__out_gif_code(g, g->codes[code].prefix); + // recurse to decode the prefixes, since the linked-list is backwards, + // and working backwards through an interleaved image would be nasty + if (g->codes[code].prefix >= 0) + stbi__out_gif_code(g, g->codes[code].prefix); - if (g->cur_y >= g->max_y) - return; + if (g->cur_y >= g->max_y) return; - idx = g->cur_x + g->cur_y; - p = &g->out[idx]; - g->history[idx / 4] = 1; + idx = g->cur_x + g->cur_y; + p = &g->out[idx]; + g->history[idx / 4] = 1; - c = &g->color_table[g->codes[code].suffix * 4]; - if (c[3] > 128) { // don't render transparent pixels; - p[0] = c[2]; - p[1] = c[1]; - p[2] = c[0]; - p[3] = c[3]; - } - g->cur_x += 4; + c = &g->color_table[g->codes[code].suffix * 4]; + if (c[3] > 128) { // don't render transparent pixels; + p[0] = c[2]; + p[1] = c[1]; + p[2] = c[0]; + p[3] = c[3]; + } + g->cur_x += 4; - if (g->cur_x >= g->max_x) { - g->cur_x = g->start_x; - g->cur_y += g->step; + if (g->cur_x >= g->max_x) { + g->cur_x = g->start_x; + g->cur_y += g->step; - while (g->cur_y >= g->max_y && g->parse > 0) { - g->step = (1 << g->parse) * g->line_size; - g->cur_y = g->start_y + (g->step >> 1); - --g->parse; - } - } + while (g->cur_y >= g->max_y && g->parse > 0) { + g->step = (1 << g->parse) * g->line_size; + g->cur_y = g->start_y + (g->step >> 1); + --g->parse; + } + } } -static stbi_uc * stbi__process_gif_raster(stbi__context * s, stbi__gif * g) { - stbi_uc lzw_cs; - stbi__int32 len, init_code; - stbi__uint32 first; - stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear; - stbi__gif_lzw * p; +static stbi_uc *stbi__process_gif_raster(stbi__context *s, stbi__gif *g) +{ + stbi_uc lzw_cs; + stbi__int32 len, init_code; + stbi__uint32 first; + stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear; + stbi__gif_lzw *p; - lzw_cs = stbi__get8(s); - if (lzw_cs > 12) - return NULL; - clear = 1 << lzw_cs; - first = 1; - codesize = lzw_cs + 1; - codemask = (1 << codesize) - 1; - bits = 0; - valid_bits = 0; - for (init_code = 0; init_code < clear; init_code++) { - g->codes[init_code].prefix = -1; - g->codes[init_code].first = (stbi_uc)init_code; - g->codes[init_code].suffix = (stbi_uc)init_code; - } + lzw_cs = stbi__get8(s); + if (lzw_cs > 12) return NULL; + clear = 1 << lzw_cs; + first = 1; + codesize = lzw_cs + 1; + codemask = (1 << codesize) - 1; + bits = 0; + valid_bits = 0; + for (init_code = 0; init_code < clear; init_code++) { + g->codes[init_code].prefix = -1; + g->codes[init_code].first = (stbi_uc) init_code; + g->codes[init_code].suffix = (stbi_uc) init_code; + } - // support no starting clear code - avail = clear + 2; - oldcode = -1; + // support no starting clear code + avail = clear+2; + oldcode = -1; - len = 0; - for (;;) { - if (valid_bits < codesize) { - if (len == 0) { - len = stbi__get8(s); // start new block - if (len == 0) - return g->out; + len = 0; + for(;;) { + if (valid_bits < codesize) { + if (len == 0) { + len = stbi__get8(s); // start new block + if (len == 0) + return g->out; + } + --len; + bits |= (stbi__int32) stbi__get8(s) << valid_bits; + valid_bits += 8; + } else { + stbi__int32 code = bits & codemask; + bits >>= codesize; + valid_bits -= codesize; + // @OPTIMIZE: is there some way we can accelerate the non-clear path? + if (code == clear) { // clear code + codesize = lzw_cs + 1; + codemask = (1 << codesize) - 1; + avail = clear + 2; + oldcode = -1; + first = 0; + } else if (code == clear + 1) { // end of stream code + stbi__skip(s, len); + while ((len = stbi__get8(s)) > 0) + stbi__skip(s,len); + return g->out; + } else if (code <= avail) { + if (first) { + return stbi__errpuc("no clear code", "Corrupt GIF"); } - --len; - bits |= (stbi__int32)stbi__get8(s) << valid_bits; - valid_bits += 8; - } else { - stbi__int32 code = bits & codemask; - bits >>= codesize; - valid_bits -= codesize; - // @OPTIMIZE: is there some way we can accelerate the non-clear path? - if (code == clear) { // clear code - codesize = lzw_cs + 1; - codemask = (1 << codesize) - 1; - avail = clear + 2; - oldcode = -1; - first = 0; - } else if (code == clear + 1) { // end of stream code - stbi__skip(s, len); - while ((len = stbi__get8(s)) > 0) - stbi__skip(s, len); - return g->out; - } else if (code <= avail) { - if (first) { - return stbi__errpuc("no clear code", "Corrupt GIF"); - } - if (oldcode >= 0) { - p = &g->codes[avail++]; - if (avail > 8192) { - return stbi__errpuc("too many codes", "Corrupt GIF"); - } + if (oldcode >= 0) { + p = &g->codes[avail++]; + if (avail > 8192) { + return stbi__errpuc("too many codes", "Corrupt GIF"); + } - p->prefix = (stbi__int16)oldcode; - p->first = g->codes[oldcode].first; - p->suffix = (code == avail) ? p->first : g->codes[code].first; - } else if (code == avail) - return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + p->prefix = (stbi__int16) oldcode; + p->first = g->codes[oldcode].first; + p->suffix = (code == avail) ? p->first : g->codes[code].first; + } else if (code == avail) + return stbi__errpuc("illegal code in raster", "Corrupt GIF"); - stbi__out_gif_code(g, (stbi__uint16)code); + stbi__out_gif_code(g, (stbi__uint16) code); - if ((avail & codemask) == 0 && avail <= 0x0FFF) { - codesize++; - codemask = (1 << codesize) - 1; - } - - oldcode = code; - } else { - return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + if ((avail & codemask) == 0 && avail <= 0x0FFF) { + codesize++; + codemask = (1 << codesize) - 1; } - } - } + + oldcode = code; + } else { + return stbi__errpuc("illegal code in raster", "Corrupt GIF"); + } + } + } } // this function is designed to support animated gifs, although stb_image doesn't support it // two back is the image from two frames ago, used for a very specific disposal format -static stbi_uc * stbi__gif_load_next(stbi__context * s, stbi__gif * g, int * comp, int req_comp, stbi_uc * two_back) { - int dispose; - int first_frame; - int pi; - int pcount; - STBI_NOTUSED(req_comp); +static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, int req_comp, stbi_uc *two_back) +{ + int dispose; + int first_frame; + int pi; + int pcount; + STBI_NOTUSED(req_comp); - // on first frame, any non-written pixels get the background colour (non-transparent) - first_frame = 0; - if (g->out == 0) { - if (!stbi__gif_header(s, g, comp, 0)) - return 0; // stbi__g_failure_reason set by stbi__gif_header - if (!stbi__mad3sizes_valid(4, g->w, g->h, 0)) - return stbi__errpuc("too large", "GIF image is too large"); - pcount = g->w * g->h; - g->out = (stbi_uc *)stbi__malloc(4 * pcount); - g->background = (stbi_uc *)stbi__malloc(4 * pcount); - g->history = (stbi_uc *)stbi__malloc(pcount); - if (!g->out || !g->background || !g->history) - return stbi__errpuc("outofmem", "Out of memory"); + // on first frame, any non-written pixels get the background colour (non-transparent) + first_frame = 0; + if (g->out == 0) { + if (!stbi__gif_header(s, g, comp,0)) return 0; // stbi__g_failure_reason set by stbi__gif_header + if (!stbi__mad3sizes_valid(4, g->w, g->h, 0)) + return stbi__errpuc("too large", "GIF image is too large"); + pcount = g->w * g->h; + g->out = (stbi_uc *) stbi__malloc(4 * pcount); + g->background = (stbi_uc *) stbi__malloc(4 * pcount); + g->history = (stbi_uc *) stbi__malloc(pcount); + if (!g->out || !g->background || !g->history) + return stbi__errpuc("outofmem", "Out of memory"); - // image is treated as "transparent" at the start - ie, nothing overwrites the current background; - // background colour is only used for pixels that are not rendered first frame, after that "background" - // color refers to the color that was there the previous frame. - memset(g->out, 0x00, 4 * pcount); - memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent) - memset(g->history, 0x00, pcount); // pixels that were affected previous frame - first_frame = 1; - } else { - // second frame - how do we dispose of the previous one? - dispose = (g->eflags & 0x1C) >> 2; - pcount = g->w * g->h; + // image is treated as "transparent" at the start - ie, nothing overwrites the current background; + // background colour is only used for pixels that are not rendered first frame, after that "background" + // color refers to the color that was there the previous frame. + memset(g->out, 0x00, 4 * pcount); + memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent) + memset(g->history, 0x00, pcount); // pixels that were affected previous frame + first_frame = 1; + } else { + // second frame - how do we dispose of the previous one? + dispose = (g->eflags & 0x1C) >> 2; + pcount = g->w * g->h; - if ((dispose == 3) && (two_back == 0)) { - dispose = 2; // if I don't have an image to revert back to, default to the old background - } + if ((dispose == 3) && (two_back == 0)) { + dispose = 2; // if I don't have an image to revert back to, default to the old background + } - if (dispose == 3) { // use previous graphic - for (pi = 0; pi < pcount; ++pi) { - if (g->history[pi]) { - memcpy(&g->out[pi * 4], &two_back[pi * 4], 4); - } + if (dispose == 3) { // use previous graphic + for (pi = 0; pi < pcount; ++pi) { + if (g->history[pi]) { + memcpy( &g->out[pi * 4], &two_back[pi * 4], 4 ); } - } else if (dispose == 2) { - // restore what was changed last frame to background before that frame; - for (pi = 0; pi < pcount; ++pi) { - if (g->history[pi]) { - memcpy(&g->out[pi * 4], &g->background[pi * 4], 4); - } + } + } else if (dispose == 2) { + // restore what was changed last frame to background before that frame; + for (pi = 0; pi < pcount; ++pi) { + if (g->history[pi]) { + memcpy( &g->out[pi * 4], &g->background[pi * 4], 4 ); } - } else { - // This is a non-disposal case eithe way, so just - // leave the pixels as is, and they will become the new background - // 1: do not dispose - // 0: not specified. - } + } + } else { + // This is a non-disposal case eithe way, so just + // leave the pixels as is, and they will become the new background + // 1: do not dispose + // 0: not specified. + } - // background is what out is after the undoing of the previou frame; - memcpy(g->background, g->out, 4 * g->w * g->h); - } + // background is what out is after the undoing of the previou frame; + memcpy( g->background, g->out, 4 * g->w * g->h ); + } - // clear my history; - memset(g->history, 0x00, g->w * g->h); // pixels that were affected previous frame + // clear my history; + memset( g->history, 0x00, g->w * g->h ); // pixels that were affected previous frame - for (;;) { - int tag = stbi__get8(s); - switch (tag) { - case 0x2C: /* Image Descriptor */ - { + for (;;) { + int tag = stbi__get8(s); + switch (tag) { + case 0x2C: /* Image Descriptor */ + { stbi__int32 x, y, w, h; - stbi_uc * o; + stbi_uc *o; x = stbi__get16le(s); y = stbi__get16le(s); w = stbi__get16le(s); h = stbi__get16le(s); if (((x + w) > (g->w)) || ((y + h) > (g->h))) - return stbi__errpuc("bad Image Descriptor", "Corrupt GIF"); + return stbi__errpuc("bad Image Descriptor", "Corrupt GIF"); g->line_size = g->w * 4; g->start_x = x * 4; g->start_y = y * g->line_size; - g->max_x = g->start_x + w * 4; - g->max_y = g->start_y + h * g->line_size; - g->cur_x = g->start_x; - g->cur_y = g->start_y; + g->max_x = g->start_x + w * 4; + g->max_y = g->start_y + h * g->line_size; + g->cur_x = g->start_x; + g->cur_y = g->start_y; // if the width of the specified rectangle is 0, that means // we may not see *any* pixels or the image is malformed; // to make sure this is caught, move the current y down to // max_y (which is what out_gif_code checks). if (w == 0) - g->cur_y = g->max_y; + g->cur_y = g->max_y; g->lflags = stbi__get8(s); if (g->lflags & 0x40) { - g->step = 8 * g->line_size; // first interlaced spacing - g->parse = 3; + g->step = 8 * g->line_size; // first interlaced spacing + g->parse = 3; } else { - g->step = g->line_size; - g->parse = 0; + g->step = g->line_size; + g->parse = 0; } if (g->lflags & 0x80) { - stbi__gif_parse_colortable(s, g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1); - g->color_table = (stbi_uc *)g->lpal; + stbi__gif_parse_colortable(s,g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1); + g->color_table = (stbi_uc *) g->lpal; } else if (g->flags & 0x80) { - g->color_table = (stbi_uc *)g->pal; + g->color_table = (stbi_uc *) g->pal; } else - return stbi__errpuc("missing color table", "Corrupt GIF"); + return stbi__errpuc("missing color table", "Corrupt GIF"); o = stbi__process_gif_raster(s, g); - if (!o) - return NULL; + if (!o) return NULL; // if this was the first frame, pcount = g->w * g->h; if (first_frame && (g->bgindex > 0)) { - // if first frame, any pixel not drawn to gets the background color - for (pi = 0; pi < pcount; ++pi) { - if (g->history[pi] == 0) { - g->pal[g->bgindex][3] = - 255; // just in case it was made transparent, undo that; It will be reset next frame if need be; - memcpy(&g->out[pi * 4], &g->pal[g->bgindex], 4); - } - } + // if first frame, any pixel not drawn to gets the background color + for (pi = 0; pi < pcount; ++pi) { + if (g->history[pi] == 0) { + g->pal[g->bgindex][3] = 255; // just in case it was made transparent, undo that; It will be reset next frame if need be; + memcpy( &g->out[pi * 4], &g->pal[g->bgindex], 4 ); + } + } } return o; - } + } - case 0x21: // Comment Extension. - { + case 0x21: // Comment Extension. + { int len; int ext = stbi__get8(s); if (ext == 0xF9) { // Graphic Control Extension. - len = stbi__get8(s); - if (len == 4) { - g->eflags = stbi__get8(s); - g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths. + len = stbi__get8(s); + if (len == 4) { + g->eflags = stbi__get8(s); + g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths. - // unset old transparent - if (g->transparent >= 0) { - g->pal[g->transparent][3] = 255; - } - if (g->eflags & 0x01) { - g->transparent = stbi__get8(s); - if (g->transparent >= 0) { - g->pal[g->transparent][3] = 0; - } - } else { - // don't need transparent - stbi__skip(s, 1); - g->transparent = -1; - } - } else { - stbi__skip(s, len); - break; - } + // unset old transparent + if (g->transparent >= 0) { + g->pal[g->transparent][3] = 255; + } + if (g->eflags & 0x01) { + g->transparent = stbi__get8(s); + if (g->transparent >= 0) { + g->pal[g->transparent][3] = 0; + } + } else { + // don't need transparent + stbi__skip(s, 1); + g->transparent = -1; + } + } else { + stbi__skip(s, len); + break; + } } while ((len = stbi__get8(s)) != 0) { - stbi__skip(s, len); + stbi__skip(s, len); } break; - } + } - case 0x3B: // gif stream termination code - return (stbi_uc *)s; // using '1' causes warning on some compilers + case 0x3B: // gif stream termination code + return (stbi_uc *) s; // using '1' causes warning on some compilers - default: + default: return stbi__errpuc("unknown code", "Corrupt GIF"); - } - } + } + } } -static void * stbi__load_gif_main_outofmem(stbi__gif * g, stbi_uc * out, int ** delays) { - STBI_FREE(g->out); - STBI_FREE(g->history); - STBI_FREE(g->background); +static void *stbi__load_gif_main_outofmem(stbi__gif *g, stbi_uc *out, int **delays) +{ + STBI_FREE(g->out); + STBI_FREE(g->history); + STBI_FREE(g->background); - if (out) - STBI_FREE(out); - if (delays && *delays) - STBI_FREE(*delays); - return stbi__errpuc("outofmem", "Out of memory"); + if (out) STBI_FREE(out); + if (delays && *delays) STBI_FREE(*delays); + return stbi__errpuc("outofmem", "Out of memory"); } -static void * stbi__load_gif_main(stbi__context * s, int ** delays, int * x, int * y, int * z, int * comp, int req_comp) { - if (stbi__gif_test(s)) { - int layers = 0; - stbi_uc * u = 0; - stbi_uc * out = 0; - stbi_uc * two_back = 0; - stbi__gif g; - int stride; - int out_size = 0; - int delays_size = 0; +static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp) +{ + if (stbi__gif_test(s)) { + int layers = 0; + stbi_uc *u = 0; + stbi_uc *out = 0; + stbi_uc *two_back = 0; + stbi__gif g; + int stride; + int out_size = 0; + int delays_size = 0; - STBI_NOTUSED(out_size); - STBI_NOTUSED(delays_size); + STBI_NOTUSED(out_size); + STBI_NOTUSED(delays_size); - memset(&g, 0, sizeof(g)); - if (delays) { - *delays = 0; - } + memset(&g, 0, sizeof(g)); + if (delays) { + *delays = 0; + } - do { - u = stbi__gif_load_next(s, &g, comp, req_comp, two_back); - if (u == (stbi_uc *)s) - u = 0; // end of animated gif marker + do { + u = stbi__gif_load_next(s, &g, comp, req_comp, two_back); + if (u == (stbi_uc *) s) u = 0; // end of animated gif marker - if (u) { - *x = g.w; - *y = g.h; - ++layers; - stride = g.w * g.h * 4; + if (u) { + *x = g.w; + *y = g.h; + ++layers; + stride = g.w * g.h * 4; - if (out) { - void * tmp = (stbi_uc *)STBI_REALLOC_SIZED(out, out_size, layers * stride); - if (!tmp) - return stbi__load_gif_main_outofmem(&g, out, delays); - else { - out = (stbi_uc *)tmp; - out_size = layers * stride; - } + if (out) { + void *tmp = (stbi_uc*) STBI_REALLOC_SIZED( out, out_size, layers * stride ); + if (!tmp) + return stbi__load_gif_main_outofmem(&g, out, delays); + else { + out = (stbi_uc*) tmp; + out_size = layers * stride; + } - if (delays) { - int * new_delays = (int *)STBI_REALLOC_SIZED(*delays, delays_size, sizeof(int) * layers); - if (!new_delays) - return stbi__load_gif_main_outofmem(&g, out, delays); - *delays = new_delays; - delays_size = layers * sizeof(int); - } - } else { - out = (stbi_uc *)stbi__malloc(layers * stride); - if (!out) - return stbi__load_gif_main_outofmem(&g, out, delays); - out_size = layers * stride; - if (delays) { - *delays = (int *)stbi__malloc(layers * sizeof(int)); - if (!*delays) - return stbi__load_gif_main_outofmem(&g, out, delays); - delays_size = layers * sizeof(int); - } - } - memcpy(out + ((layers - 1) * stride), u, stride); - if (layers >= 2) { - two_back = out - 2 * stride; - } - - if (delays) { - (*delays)[layers - 1U] = g.delay; - } + if (delays) { + int *new_delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers ); + if (!new_delays) + return stbi__load_gif_main_outofmem(&g, out, delays); + *delays = new_delays; + delays_size = layers * sizeof(int); + } + } else { + out = (stbi_uc*)stbi__malloc( layers * stride ); + if (!out) + return stbi__load_gif_main_outofmem(&g, out, delays); + out_size = layers * stride; + if (delays) { + *delays = (int*) stbi__malloc( layers * sizeof(int) ); + if (!*delays) + return stbi__load_gif_main_outofmem(&g, out, delays); + delays_size = layers * sizeof(int); + } + } + memcpy( out + ((layers - 1) * stride), u, stride ); + if (layers >= 2) { + two_back = out - 2 * stride; } - } while (u != 0); - // free temp buffer; - STBI_FREE(g.out); - STBI_FREE(g.history); - STBI_FREE(g.background); + if (delays) { + (*delays)[layers - 1U] = g.delay; + } + } + } while (u != 0); - // do the final conversion after loading everything; - if (req_comp && req_comp != 4) - out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h); + // free temp buffer; + STBI_FREE(g.out); + STBI_FREE(g.history); + STBI_FREE(g.background); - *z = layers; - return out; - } else { - return stbi__errpuc("not GIF", "Image was not as a gif type."); - } + // do the final conversion after loading everything; + if (req_comp && req_comp != 4) + out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h); + + *z = layers; + return out; + } else { + return stbi__errpuc("not GIF", "Image was not as a gif type."); + } } -static void * stbi__gif_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * u = 0; - stbi__gif g; - memset(&g, 0, sizeof(g)); - STBI_NOTUSED(ri); +static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *u = 0; + stbi__gif g; + memset(&g, 0, sizeof(g)); + STBI_NOTUSED(ri); - u = stbi__gif_load_next(s, &g, comp, req_comp, 0); - if (u == (stbi_uc *)s) - u = 0; // end of animated gif marker - if (u) { - *x = g.w; - *y = g.h; + u = stbi__gif_load_next(s, &g, comp, req_comp, 0); + if (u == (stbi_uc *) s) u = 0; // end of animated gif marker + if (u) { + *x = g.w; + *y = g.h; - // moved conversion to after successful load so that the same - // can be done for multiple frames. - if (req_comp && req_comp != 4) - u = stbi__convert_format(u, 4, req_comp, g.w, g.h); - } else if (g.out) { - // if there was an error and we allocated an image buffer, free it! - STBI_FREE(g.out); - } + // moved conversion to after successful load so that the same + // can be done for multiple frames. + if (req_comp && req_comp != 4) + u = stbi__convert_format(u, 4, req_comp, g.w, g.h); + } else if (g.out) { + // if there was an error and we allocated an image buffer, free it! + STBI_FREE(g.out); + } - // free buffers needed for multiple frame loading; - STBI_FREE(g.history); - STBI_FREE(g.background); + // free buffers needed for multiple frame loading; + STBI_FREE(g.history); + STBI_FREE(g.background); - return u; + return u; } -static int stbi__gif_info(stbi__context * s, int * x, int * y, int * comp) { return stbi__gif_info_raw(s, x, y, comp); } +static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp) +{ + return stbi__gif_info_raw(s,x,y,comp); +} #endif // ************************************************************************************************* // Radiance RGBE HDR loader // originally by Nicolas Schulz #ifndef STBI_NO_HDR -static int stbi__hdr_test_core(stbi__context * s, const char * signature) { - int i; - for (i = 0; signature[i]; ++i) - if (stbi__get8(s) != signature[i]) - return 0; - stbi__rewind(s); - return 1; +static int stbi__hdr_test_core(stbi__context *s, const char *signature) +{ + int i; + for (i=0; signature[i]; ++i) + if (stbi__get8(s) != signature[i]) + return 0; + stbi__rewind(s); + return 1; } -static int stbi__hdr_test(stbi__context * s) { - int r = stbi__hdr_test_core(s, "#?RADIANCE\n"); - stbi__rewind(s); - if (!r) { - r = stbi__hdr_test_core(s, "#?RGBE\n"); - stbi__rewind(s); - } - return r; +static int stbi__hdr_test(stbi__context* s) +{ + int r = stbi__hdr_test_core(s, "#?RADIANCE\n"); + stbi__rewind(s); + if(!r) { + r = stbi__hdr_test_core(s, "#?RGBE\n"); + stbi__rewind(s); + } + return r; } -#define STBI__HDR_BUFLEN 1024 -static char * stbi__hdr_gettoken(stbi__context * z, char * buffer) { - int len = 0; - char c = '\0'; +#define STBI__HDR_BUFLEN 1024 +static char *stbi__hdr_gettoken(stbi__context *z, char *buffer) +{ + int len=0; + char c = '\0'; - c = (char)stbi__get8(z); + c = (char) stbi__get8(z); - while (!stbi__at_eof(z) && c != '\n') { - buffer[len++] = c; - if (len == STBI__HDR_BUFLEN - 1) { - // flush to end of line - while (!stbi__at_eof(z) && stbi__get8(z) != '\n') - ; - break; - } - c = (char)stbi__get8(z); - } + while (!stbi__at_eof(z) && c != '\n') { + buffer[len++] = c; + if (len == STBI__HDR_BUFLEN-1) { + // flush to end of line + while (!stbi__at_eof(z) && stbi__get8(z) != '\n') + ; + break; + } + c = (char) stbi__get8(z); + } - buffer[len] = 0; - return buffer; + buffer[len] = 0; + return buffer; } -static void stbi__hdr_convert(float * output, stbi_uc * input, int req_comp) { - if (input[3] != 0) { - float f1; - // Exponent - f1 = (float)ldexp(1.0f, input[3] - (int)(128 + 8)); - if (req_comp <= 2) - output[0] = (input[0] + input[1] + input[2]) * f1 / 3; - else { - output[0] = input[0] * f1; - output[1] = input[1] * f1; - output[2] = input[2] * f1; - } - if (req_comp == 2) - output[1] = 1; - if (req_comp == 4) - output[3] = 1; - } else { - switch (req_comp) { - case 4: - output[3] = 1; /* fallthrough */ - case 3: - output[0] = output[1] = output[2] = 0; - break; - case 2: - output[1] = 1; /* fallthrough */ - case 1: - output[0] = 0; - break; - } - } +static void stbi__hdr_convert(float *output, stbi_uc *input, int req_comp) +{ + if ( input[3] != 0 ) { + float f1; + // Exponent + f1 = (float) ldexp(1.0f, input[3] - (int)(128 + 8)); + if (req_comp <= 2) + output[0] = (input[0] + input[1] + input[2]) * f1 / 3; + else { + output[0] = input[0] * f1; + output[1] = input[1] * f1; + output[2] = input[2] * f1; + } + if (req_comp == 2) output[1] = 1; + if (req_comp == 4) output[3] = 1; + } else { + switch (req_comp) { + case 4: output[3] = 1; /* fallthrough */ + case 3: output[0] = output[1] = output[2] = 0; + break; + case 2: output[1] = 1; /* fallthrough */ + case 1: output[0] = 0; + break; + } + } } -static float * stbi__hdr_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - char buffer[STBI__HDR_BUFLEN]; - char * token; - int valid = 0; - int width, height; - stbi_uc * scanline; - float * hdr_data; - int len; - unsigned char count, value; - int i, j, k, c1, c2, z; - const char * headerToken; - STBI_NOTUSED(ri); +static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + char buffer[STBI__HDR_BUFLEN]; + char *token; + int valid = 0; + int width, height; + stbi_uc *scanline; + float *hdr_data; + int len; + unsigned char count, value; + int i, j, k, c1,c2, z; + const char *headerToken; + STBI_NOTUSED(ri); - // Check identifier - headerToken = stbi__hdr_gettoken(s, buffer); - if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0) - return stbi__errpf("not HDR", "Corrupt HDR image"); + // Check identifier + headerToken = stbi__hdr_gettoken(s,buffer); + if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0) + return stbi__errpf("not HDR", "Corrupt HDR image"); - // Parse header - for (;;) { - token = stbi__hdr_gettoken(s, buffer); - if (token[0] == 0) - break; - if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) - valid = 1; - } + // Parse header + for(;;) { + token = stbi__hdr_gettoken(s,buffer); + if (token[0] == 0) break; + if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1; + } - if (!valid) - return stbi__errpf("unsupported format", "Unsupported HDR format"); + if (!valid) return stbi__errpf("unsupported format", "Unsupported HDR format"); - // Parse width and height - // can't use sscanf() if we're not using stdio! - token = stbi__hdr_gettoken(s, buffer); - if (strncmp(token, "-Y ", 3)) - return stbi__errpf("unsupported data layout", "Unsupported HDR format"); - token += 3; - height = (int)strtol(token, &token, 10); - while (*token == ' ') - ++token; - if (strncmp(token, "+X ", 3)) - return stbi__errpf("unsupported data layout", "Unsupported HDR format"); - token += 3; - width = (int)strtol(token, NULL, 10); + // Parse width and height + // can't use sscanf() if we're not using stdio! + token = stbi__hdr_gettoken(s,buffer); + if (strncmp(token, "-Y ", 3)) return stbi__errpf("unsupported data layout", "Unsupported HDR format"); + token += 3; + height = (int) strtol(token, &token, 10); + while (*token == ' ') ++token; + if (strncmp(token, "+X ", 3)) return stbi__errpf("unsupported data layout", "Unsupported HDR format"); + token += 3; + width = (int) strtol(token, NULL, 10); - if (height > STBI_MAX_DIMENSIONS) - return stbi__errpf("too large", "Very large image (corrupt?)"); - if (width > STBI_MAX_DIMENSIONS) - return stbi__errpf("too large", "Very large image (corrupt?)"); + if (height > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)"); + if (width > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)"); - *x = width; - *y = height; + *x = width; + *y = height; - if (comp) - *comp = 3; - if (req_comp == 0) - req_comp = 3; + if (comp) *comp = 3; + if (req_comp == 0) req_comp = 3; - if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0)) - return stbi__errpf("too large", "HDR image is too large"); + if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0)) + return stbi__errpf("too large", "HDR image is too large"); - // Read data - hdr_data = (float *)stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0); - if (!hdr_data) - return stbi__errpf("outofmem", "Out of memory"); + // Read data + hdr_data = (float *) stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0); + if (!hdr_data) + return stbi__errpf("outofmem", "Out of memory"); - // Load image data - // image data is stored as some number of sca - if (width < 8 || width >= 32768) { - // Read flat data - for (j = 0; j < height; ++j) { - for (i = 0; i < width; ++i) { - stbi_uc rgbe[4]; - main_decode_loop: - stbi__getn(s, rgbe, 4); - stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp); - } - } - } else { - // Read RLE-encoded data - scanline = NULL; + // Load image data + // image data is stored as some number of sca + if ( width < 8 || width >= 32768) { + // Read flat data + for (j=0; j < height; ++j) { + for (i=0; i < width; ++i) { + stbi_uc rgbe[4]; + main_decode_loop: + stbi__getn(s, rgbe, 4); + stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp); + } + } + } else { + // Read RLE-encoded data + scanline = NULL; - for (j = 0; j < height; ++j) { - c1 = stbi__get8(s); - c2 = stbi__get8(s); - len = stbi__get8(s); - if (c1 != 2 || c2 != 2 || (len & 0x80)) { - // not run-length encoded, so we have to actually use THIS data as a decoded - // pixel (note this can't be a valid pixel--one of RGB must be >= 128) - stbi_uc rgbe[4]; - rgbe[0] = (stbi_uc)c1; - rgbe[1] = (stbi_uc)c2; - rgbe[2] = (stbi_uc)len; - rgbe[3] = (stbi_uc)stbi__get8(s); - stbi__hdr_convert(hdr_data, rgbe, req_comp); - i = 1; - j = 0; - STBI_FREE(scanline); - goto main_decode_loop; // yes, this makes no sense - } - len <<= 8; - len |= stbi__get8(s); - if (len != width) { - STBI_FREE(hdr_data); - STBI_FREE(scanline); - return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); - } - if (scanline == NULL) { - scanline = (stbi_uc *)stbi__malloc_mad2(width, 4, 0); - if (!scanline) { - STBI_FREE(hdr_data); - return stbi__errpf("outofmem", "Out of memory"); - } - } - - for (k = 0; k < 4; ++k) { - int nleft; - i = 0; - while ((nleft = width - i) > 0) { - count = stbi__get8(s); - if (count > 128) { - // Run - value = stbi__get8(s); - count -= 128; - if ((count == 0) || (count > nleft)) { - STBI_FREE(hdr_data); - STBI_FREE(scanline); - return stbi__errpf("corrupt", "bad RLE data in HDR"); - } - for (z = 0; z < count; ++z) - scanline[i++ * 4 + k] = value; - } else { - // Dump - if ((count == 0) || (count > nleft)) { - STBI_FREE(hdr_data); - STBI_FREE(scanline); - return stbi__errpf("corrupt", "bad RLE data in HDR"); - } - for (z = 0; z < count; ++z) - scanline[i++ * 4 + k] = stbi__get8(s); - } - } - } - for (i = 0; i < width; ++i) - stbi__hdr_convert(hdr_data + (j * width + i) * req_comp, scanline + i * 4, req_comp); - } - if (scanline) + for (j = 0; j < height; ++j) { + c1 = stbi__get8(s); + c2 = stbi__get8(s); + len = stbi__get8(s); + if (c1 != 2 || c2 != 2 || (len & 0x80)) { + // not run-length encoded, so we have to actually use THIS data as a decoded + // pixel (note this can't be a valid pixel--one of RGB must be >= 128) + stbi_uc rgbe[4]; + rgbe[0] = (stbi_uc) c1; + rgbe[1] = (stbi_uc) c2; + rgbe[2] = (stbi_uc) len; + rgbe[3] = (stbi_uc) stbi__get8(s); + stbi__hdr_convert(hdr_data, rgbe, req_comp); + i = 1; + j = 0; STBI_FREE(scanline); - } + goto main_decode_loop; // yes, this makes no sense + } + len <<= 8; + len |= stbi__get8(s); + if (len != width) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); } + if (scanline == NULL) { + scanline = (stbi_uc *) stbi__malloc_mad2(width, 4, 0); + if (!scanline) { + STBI_FREE(hdr_data); + return stbi__errpf("outofmem", "Out of memory"); + } + } - return hdr_data; + for (k = 0; k < 4; ++k) { + int nleft; + i = 0; + while ((nleft = width - i) > 0) { + count = stbi__get8(s); + if (count > 128) { + // Run + value = stbi__get8(s); + count -= 128; + if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); } + for (z = 0; z < count; ++z) + scanline[i++ * 4 + k] = value; + } else { + // Dump + if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); } + for (z = 0; z < count; ++z) + scanline[i++ * 4 + k] = stbi__get8(s); + } + } + } + for (i=0; i < width; ++i) + stbi__hdr_convert(hdr_data+(j*width + i)*req_comp, scanline + i*4, req_comp); + } + if (scanline) + STBI_FREE(scanline); + } + + return hdr_data; } -static int stbi__hdr_info(stbi__context * s, int * x, int * y, int * comp) { - char buffer[STBI__HDR_BUFLEN]; - char * token; - int valid = 0; - int dummy; +static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp) +{ + char buffer[STBI__HDR_BUFLEN]; + char *token; + int valid = 0; + int dummy; - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; - if (stbi__hdr_test(s) == 0) { - stbi__rewind(s); - return 0; - } + if (stbi__hdr_test(s) == 0) { + stbi__rewind( s ); + return 0; + } - for (;;) { - token = stbi__hdr_gettoken(s, buffer); - if (token[0] == 0) - break; - if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) - valid = 1; - } + for(;;) { + token = stbi__hdr_gettoken(s,buffer); + if (token[0] == 0) break; + if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1; + } - if (!valid) { - stbi__rewind(s); - return 0; - } - token = stbi__hdr_gettoken(s, buffer); - if (strncmp(token, "-Y ", 3)) { - stbi__rewind(s); - return 0; - } - token += 3; - *y = (int)strtol(token, &token, 10); - while (*token == ' ') - ++token; - if (strncmp(token, "+X ", 3)) { - stbi__rewind(s); - return 0; - } - token += 3; - *x = (int)strtol(token, NULL, 10); - *comp = 3; - return 1; + if (!valid) { + stbi__rewind( s ); + return 0; + } + token = stbi__hdr_gettoken(s,buffer); + if (strncmp(token, "-Y ", 3)) { + stbi__rewind( s ); + return 0; + } + token += 3; + *y = (int) strtol(token, &token, 10); + while (*token == ' ') ++token; + if (strncmp(token, "+X ", 3)) { + stbi__rewind( s ); + return 0; + } + token += 3; + *x = (int) strtol(token, NULL, 10); + *comp = 3; + return 1; } #endif // STBI_NO_HDR #ifndef STBI_NO_BMP -static int stbi__bmp_info(stbi__context * s, int * x, int * y, int * comp) { - void * p; - stbi__bmp_data info; +static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp) +{ + void *p; + stbi__bmp_data info; - info.all_a = 255; - p = stbi__bmp_parse_header(s, &info); - if (p == NULL) { - stbi__rewind(s); - return 0; - } - if (x) - *x = s->img_x; - if (y) - *y = s->img_y; - if (comp) { - if (info.bpp == 24 && info.ma == 0xff000000) - *comp = 3; - else - *comp = info.ma ? 4 : 3; - } - return 1; + info.all_a = 255; + p = stbi__bmp_parse_header(s, &info); + if (p == NULL) { + stbi__rewind( s ); + return 0; + } + if (x) *x = s->img_x; + if (y) *y = s->img_y; + if (comp) { + if (info.bpp == 24 && info.ma == 0xff000000) + *comp = 3; + else + *comp = info.ma ? 4 : 3; + } + return 1; } #endif #ifndef STBI_NO_PSD -static int stbi__psd_info(stbi__context * s, int * x, int * y, int * comp) { - int channelCount, dummy, depth; - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; - if (stbi__get32be(s) != 0x38425053) { - stbi__rewind(s); - return 0; - } - if (stbi__get16be(s) != 1) { - stbi__rewind(s); - return 0; - } - stbi__skip(s, 6); - channelCount = stbi__get16be(s); - if (channelCount < 0 || channelCount > 16) { - stbi__rewind(s); - return 0; - } - *y = stbi__get32be(s); - *x = stbi__get32be(s); - depth = stbi__get16be(s); - if (depth != 8 && depth != 16) { - stbi__rewind(s); - return 0; - } - if (stbi__get16be(s) != 3) { - stbi__rewind(s); - return 0; - } - *comp = 4; - return 1; +static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp) +{ + int channelCount, dummy, depth; + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + if (stbi__get32be(s) != 0x38425053) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 1) { + stbi__rewind( s ); + return 0; + } + stbi__skip(s, 6); + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) { + stbi__rewind( s ); + return 0; + } + *y = stbi__get32be(s); + *x = stbi__get32be(s); + depth = stbi__get16be(s); + if (depth != 8 && depth != 16) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 3) { + stbi__rewind( s ); + return 0; + } + *comp = 4; + return 1; } -static int stbi__psd_is16(stbi__context * s) { - int channelCount, depth; - if (stbi__get32be(s) != 0x38425053) { - stbi__rewind(s); - return 0; - } - if (stbi__get16be(s) != 1) { - stbi__rewind(s); - return 0; - } - stbi__skip(s, 6); - channelCount = stbi__get16be(s); - if (channelCount < 0 || channelCount > 16) { - stbi__rewind(s); - return 0; - } - STBI_NOTUSED(stbi__get32be(s)); - STBI_NOTUSED(stbi__get32be(s)); - depth = stbi__get16be(s); - if (depth != 16) { - stbi__rewind(s); - return 0; - } - return 1; +static int stbi__psd_is16(stbi__context *s) +{ + int channelCount, depth; + if (stbi__get32be(s) != 0x38425053) { + stbi__rewind( s ); + return 0; + } + if (stbi__get16be(s) != 1) { + stbi__rewind( s ); + return 0; + } + stbi__skip(s, 6); + channelCount = stbi__get16be(s); + if (channelCount < 0 || channelCount > 16) { + stbi__rewind( s ); + return 0; + } + STBI_NOTUSED(stbi__get32be(s)); + STBI_NOTUSED(stbi__get32be(s)); + depth = stbi__get16be(s); + if (depth != 16) { + stbi__rewind( s ); + return 0; + } + return 1; } #endif #ifndef STBI_NO_PIC -static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp) { - int act_comp = 0, num_packets = 0, chained, dummy; - stbi__pic_packet packets[10]; +static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp) +{ + int act_comp=0,num_packets=0,chained,dummy; + stbi__pic_packet packets[10]; - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; - if (!stbi__pic_is4(s, "\x53\x80\xF6\x34")) { - stbi__rewind(s); - return 0; - } + if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) { + stbi__rewind(s); + return 0; + } - stbi__skip(s, 88); + stbi__skip(s, 88); - *x = stbi__get16be(s); - *y = stbi__get16be(s); - if (stbi__at_eof(s)) { - stbi__rewind(s); - return 0; - } - if ((*x) != 0 && (1 << 28) / (*x) < (*y)) { - stbi__rewind(s); - return 0; - } + *x = stbi__get16be(s); + *y = stbi__get16be(s); + if (stbi__at_eof(s)) { + stbi__rewind( s); + return 0; + } + if ( (*x) != 0 && (1 << 28) / (*x) < (*y)) { + stbi__rewind( s ); + return 0; + } - stbi__skip(s, 8); + stbi__skip(s, 8); - do { - stbi__pic_packet * packet; + do { + stbi__pic_packet *packet; - if (num_packets == sizeof(packets) / sizeof(packets[0])) - return 0; + if (num_packets==sizeof(packets)/sizeof(packets[0])) + return 0; - packet = &packets[num_packets++]; - chained = stbi__get8(s); - packet->size = stbi__get8(s); - packet->type = stbi__get8(s); - packet->channel = stbi__get8(s); - act_comp |= packet->channel; + packet = &packets[num_packets++]; + chained = stbi__get8(s); + packet->size = stbi__get8(s); + packet->type = stbi__get8(s); + packet->channel = stbi__get8(s); + act_comp |= packet->channel; - if (stbi__at_eof(s)) { - stbi__rewind(s); - return 0; - } - if (packet->size != 8) { - stbi__rewind(s); - return 0; - } - } while (chained); + if (stbi__at_eof(s)) { + stbi__rewind( s ); + return 0; + } + if (packet->size != 8) { + stbi__rewind( s ); + return 0; + } + } while (chained); - *comp = (act_comp & 0x10 ? 4 : 3); + *comp = (act_comp & 0x10 ? 4 : 3); - return 1; + return 1; } #endif @@ -7904,271 +7491,272 @@ static int stbi__pic_info(stbi__context * s, int * x, int * y, int * comp) { #ifndef STBI_NO_PNM -static int stbi__pnm_test(stbi__context * s) { - char p, t; - p = (char)stbi__get8(s); - t = (char)stbi__get8(s); - if (p != 'P' || (t != '5' && t != '6')) { - stbi__rewind(s); - return 0; - } - return 1; +static int stbi__pnm_test(stbi__context *s) +{ + char p, t; + p = (char) stbi__get8(s); + t = (char) stbi__get8(s); + if (p != 'P' || (t != '5' && t != '6')) { + stbi__rewind( s ); + return 0; + } + return 1; } -static void * stbi__pnm_load(stbi__context * s, int * x, int * y, int * comp, int req_comp, stbi__result_info * ri) { - stbi_uc * out; - STBI_NOTUSED(ri); +static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri) +{ + stbi_uc *out; + STBI_NOTUSED(ri); - ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n); - if (ri->bits_per_channel == 0) - return 0; + ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n); + if (ri->bits_per_channel == 0) + return 0; - if (s->img_y > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); - if (s->img_x > STBI_MAX_DIMENSIONS) - return stbi__errpuc("too large", "Very large image (corrupt?)"); + if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); + if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)"); - *x = s->img_x; - *y = s->img_y; - if (comp) - *comp = s->img_n; + *x = s->img_x; + *y = s->img_y; + if (comp) *comp = s->img_n; - if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0)) - return stbi__errpuc("too large", "PNM too large"); + if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0)) + return stbi__errpuc("too large", "PNM too large"); - out = (stbi_uc *)stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0); - if (!out) - return stbi__errpuc("outofmem", "Out of memory"); - if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) { - STBI_FREE(out); - return stbi__errpuc("bad PNM", "PNM file truncated"); - } + out = (stbi_uc *) stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0); + if (!out) return stbi__errpuc("outofmem", "Out of memory"); + if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) { + STBI_FREE(out); + return stbi__errpuc("bad PNM", "PNM file truncated"); + } - if (req_comp && req_comp != s->img_n) { - if (ri->bits_per_channel == 16) { - out = (stbi_uc *)stbi__convert_format16((stbi__uint16 *)out, s->img_n, req_comp, s->img_x, s->img_y); - } else { - out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y); - } - if (out == NULL) - return out; // stbi__convert_format frees input on failure - } - return out; + if (req_comp && req_comp != s->img_n) { + if (ri->bits_per_channel == 16) { + out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, s->img_n, req_comp, s->img_x, s->img_y); + } else { + out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y); + } + if (out == NULL) return out; // stbi__convert_format frees input on failure + } + return out; } -static int stbi__pnm_isspace(char c) { return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r'; } - -static void stbi__pnm_skip_whitespace(stbi__context * s, char * c) { - for (;;) { - while (!stbi__at_eof(s) && stbi__pnm_isspace(*c)) - *c = (char)stbi__get8(s); - - if (stbi__at_eof(s) || *c != '#') - break; - - while (!stbi__at_eof(s) && *c != '\n' && *c != '\r') - *c = (char)stbi__get8(s); - } +static int stbi__pnm_isspace(char c) +{ + return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r'; } -static int stbi__pnm_isdigit(char c) { return c >= '0' && c <= '9'; } +static void stbi__pnm_skip_whitespace(stbi__context *s, char *c) +{ + for (;;) { + while (!stbi__at_eof(s) && stbi__pnm_isspace(*c)) + *c = (char) stbi__get8(s); -static int stbi__pnm_getinteger(stbi__context * s, char * c) { - int value = 0; + if (stbi__at_eof(s) || *c != '#') + break; - while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) { - value = value * 10 + (*c - '0'); - *c = (char)stbi__get8(s); - if ((value > 214748364) || (value == 214748364 && *c > '7')) - return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int"); - } - - return value; + while (!stbi__at_eof(s) && *c != '\n' && *c != '\r' ) + *c = (char) stbi__get8(s); + } } -static int stbi__pnm_info(stbi__context * s, int * x, int * y, int * comp) { - int maxv, dummy; - char c, p, t; - - if (!x) - x = &dummy; - if (!y) - y = &dummy; - if (!comp) - comp = &dummy; - - stbi__rewind(s); - - // Get identifier - p = (char)stbi__get8(s); - t = (char)stbi__get8(s); - if (p != 'P' || (t != '5' && t != '6')) { - stbi__rewind(s); - return 0; - } - - *comp = (t == '6') ? 3 : 1; // '5' is 1-component .pgm; '6' is 3-component .ppm - - c = (char)stbi__get8(s); - stbi__pnm_skip_whitespace(s, &c); - - *x = stbi__pnm_getinteger(s, &c); // read width - if (*x == 0) - return stbi__err("invalid width", "PPM image header had zero or overflowing width"); - stbi__pnm_skip_whitespace(s, &c); - - *y = stbi__pnm_getinteger(s, &c); // read height - if (*y == 0) - return stbi__err("invalid width", "PPM image header had zero or overflowing width"); - stbi__pnm_skip_whitespace(s, &c); - - maxv = stbi__pnm_getinteger(s, &c); // read max value - if (maxv > 65535) - return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images"); - else if (maxv > 255) - return 16; - else - return 8; +static int stbi__pnm_isdigit(char c) +{ + return c >= '0' && c <= '9'; } -static int stbi__pnm_is16(stbi__context * s) { - if (stbi__pnm_info(s, NULL, NULL, NULL) == 16) - return 1; - return 0; +static int stbi__pnm_getinteger(stbi__context *s, char *c) +{ + int value = 0; + + while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) { + value = value*10 + (*c - '0'); + *c = (char) stbi__get8(s); + if((value > 214748364) || (value == 214748364 && *c > '7')) + return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int"); + } + + return value; +} + +static int stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp) +{ + int maxv, dummy; + char c, p, t; + + if (!x) x = &dummy; + if (!y) y = &dummy; + if (!comp) comp = &dummy; + + stbi__rewind(s); + + // Get identifier + p = (char) stbi__get8(s); + t = (char) stbi__get8(s); + if (p != 'P' || (t != '5' && t != '6')) { + stbi__rewind(s); + return 0; + } + + *comp = (t == '6') ? 3 : 1; // '5' is 1-component .pgm; '6' is 3-component .ppm + + c = (char) stbi__get8(s); + stbi__pnm_skip_whitespace(s, &c); + + *x = stbi__pnm_getinteger(s, &c); // read width + if(*x == 0) + return stbi__err("invalid width", "PPM image header had zero or overflowing width"); + stbi__pnm_skip_whitespace(s, &c); + + *y = stbi__pnm_getinteger(s, &c); // read height + if (*y == 0) + return stbi__err("invalid width", "PPM image header had zero or overflowing width"); + stbi__pnm_skip_whitespace(s, &c); + + maxv = stbi__pnm_getinteger(s, &c); // read max value + if (maxv > 65535) + return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images"); + else if (maxv > 255) + return 16; + else + return 8; +} + +static int stbi__pnm_is16(stbi__context *s) +{ + if (stbi__pnm_info(s, NULL, NULL, NULL) == 16) + return 1; + return 0; } #endif -static int stbi__info_main(stbi__context * s, int * x, int * y, int * comp) { -#ifndef STBI_NO_JPEG - if (stbi__jpeg_info(s, x, y, comp)) - return 1; -#endif +static int stbi__info_main(stbi__context *s, int *x, int *y, int *comp) +{ + #ifndef STBI_NO_JPEG + if (stbi__jpeg_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PNG - if (stbi__png_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PNG + if (stbi__png_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_GIF - if (stbi__gif_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_GIF + if (stbi__gif_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_BMP - if (stbi__bmp_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_BMP + if (stbi__bmp_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PSD - if (stbi__psd_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PSD + if (stbi__psd_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PIC - if (stbi__pic_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PIC + if (stbi__pic_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_PNM - if (stbi__pnm_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_PNM + if (stbi__pnm_info(s, x, y, comp)) return 1; + #endif -#ifndef STBI_NO_HDR - if (stbi__hdr_info(s, x, y, comp)) - return 1; -#endif + #ifndef STBI_NO_HDR + if (stbi__hdr_info(s, x, y, comp)) return 1; + #endif -// test tga last because it's a crappy test! -#ifndef STBI_NO_TGA - if (stbi__tga_info(s, x, y, comp)) - return 1; -#endif - return stbi__err("unknown image type", "Image not of any known type, or corrupt"); + // test tga last because it's a crappy test! + #ifndef STBI_NO_TGA + if (stbi__tga_info(s, x, y, comp)) + return 1; + #endif + return stbi__err("unknown image type", "Image not of any known type, or corrupt"); } -static int stbi__is_16_main(stbi__context * s) { -#ifndef STBI_NO_PNG - if (stbi__png_is16(s)) - return 1; -#endif +static int stbi__is_16_main(stbi__context *s) +{ + #ifndef STBI_NO_PNG + if (stbi__png_is16(s)) return 1; + #endif -#ifndef STBI_NO_PSD - if (stbi__psd_is16(s)) - return 1; -#endif + #ifndef STBI_NO_PSD + if (stbi__psd_is16(s)) return 1; + #endif -#ifndef STBI_NO_PNM - if (stbi__pnm_is16(s)) - return 1; -#endif - return 0; + #ifndef STBI_NO_PNM + if (stbi__pnm_is16(s)) return 1; + #endif + return 0; } #ifndef STBI_NO_STDIO -STBIDEF int stbi_info(char const * filename, int * x, int * y, int * comp) { - FILE * f = stbi__fopen(filename, "rb"); +STBIDEF int stbi_info(char const *filename, int *x, int *y, int *comp) +{ + FILE *f = stbi__fopen(filename, "rb"); int result; - if (!f) - return stbi__err("can't fopen", "Unable to open file"); + if (!f) return stbi__err("can't fopen", "Unable to open file"); result = stbi_info_from_file(f, x, y, comp); fclose(f); return result; } -STBIDEF int stbi_info_from_file(FILE * f, int * x, int * y, int * comp) { - int r; - stbi__context s; - long pos = ftell(f); - stbi__start_file(&s, f); - r = stbi__info_main(&s, x, y, comp); - fseek(f, pos, SEEK_SET); - return r; +STBIDEF int stbi_info_from_file(FILE *f, int *x, int *y, int *comp) +{ + int r; + stbi__context s; + long pos = ftell(f); + stbi__start_file(&s, f); + r = stbi__info_main(&s,x,y,comp); + fseek(f,pos,SEEK_SET); + return r; } -STBIDEF int stbi_is_16_bit(char const * filename) { - FILE * f = stbi__fopen(filename, "rb"); +STBIDEF int stbi_is_16_bit(char const *filename) +{ + FILE *f = stbi__fopen(filename, "rb"); int result; - if (!f) - return stbi__err("can't fopen", "Unable to open file"); + if (!f) return stbi__err("can't fopen", "Unable to open file"); result = stbi_is_16_bit_from_file(f); fclose(f); return result; } -STBIDEF int stbi_is_16_bit_from_file(FILE * f) { - int r; - stbi__context s; - long pos = ftell(f); - stbi__start_file(&s, f); - r = stbi__is_16_main(&s); - fseek(f, pos, SEEK_SET); - return r; +STBIDEF int stbi_is_16_bit_from_file(FILE *f) +{ + int r; + stbi__context s; + long pos = ftell(f); + stbi__start_file(&s, f); + r = stbi__is_16_main(&s); + fseek(f,pos,SEEK_SET); + return r; } #endif // !STBI_NO_STDIO -STBIDEF int stbi_info_from_memory(stbi_uc const * buffer, int len, int * x, int * y, int * comp) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__info_main(&s, x, y, comp); +STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__info_main(&s,x,y,comp); } -STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const * c, void * user, int * x, int * y, int * comp) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)c, user); - return stbi__info_main(&s, x, y, comp); +STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *c, void *user, int *x, int *y, int *comp) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user); + return stbi__info_main(&s,x,y,comp); } -STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const * buffer, int len) { - stbi__context s; - stbi__start_mem(&s, buffer, len); - return stbi__is_16_main(&s); +STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len) +{ + stbi__context s; + stbi__start_mem(&s,buffer,len); + return stbi__is_16_main(&s); } -STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * c, void * user) { - stbi__context s; - stbi__start_callbacks(&s, (stbi_io_callbacks *)c, user); - return stbi__is_16_main(&s); +STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *c, void *user) +{ + stbi__context s; + stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user); + return stbi__is_16_main(&s); } #endif // STB_IMAGE_IMPLEMENTATION @@ -8279,9 +7867,12 @@ STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * c, void * us 1.30 (2011-06-11) added ability to load files via callbacks to accomidate custom input streams (Ben Wenger) removed deprecated format-specific test/load functions - removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks - anyway error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha) fix inefficiency in - decoding 32-bit BMP (David Woo) 1.29 (2010-08-16) various warning fixes from Aurelien Pocheville 1.28 (2010-08-01) + removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks anyway + error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha) + fix inefficiency in decoding 32-bit BMP (David Woo) + 1.29 (2010-08-16) + various warning fixes from Aurelien Pocheville + 1.28 (2010-08-01) fix bug in GIF palette transparency (SpartanJ) 1.27 (2010-08-01) cast-to-stbi_uc to fix warnings @@ -8353,6 +7944,7 @@ STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const * c, void * us first released version */ + /* ------------------------------------------------------------------------------ This software is available under 2 licenses -- choose whichever you prefer. diff --git a/llama/unicode-data.cpp b/llama/unicode-data.cpp index 7e89db07..71de6795 100644 --- a/llama/unicode-data.cpp +++ b/llama/unicode-data.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/unicode-data.h b/llama/unicode-data.h index 9cd3d5f9..bb31731d 100644 --- a/llama/unicode-data.h +++ b/llama/unicode-data.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/unicode.cpp b/llama/unicode.cpp index b7bdf334..d3d8b05f 100644 --- a/llama/unicode.cpp +++ b/llama/unicode.cpp @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License * diff --git a/llama/unicode.h b/llama/unicode.h index 2ea5e168..34f2bf7a 100644 --- a/llama/unicode.h +++ b/llama/unicode.h @@ -1,5 +1,5 @@ /** - * llama.cpp - commit 1e6f6554aa11fa10160a5fda689e736c3c34169f - do not edit this file + * llama.cpp - commit 8962422b1c6f9b8b15f5aeaea42600bcc2d44177 - do not edit this file * * MIT License *