From d27b3ca1758dfb1718e333d497ef4b68ad109bc2 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 20 Jun 2025 11:19:15 +0300 Subject: [PATCH 1/8] ggml : fix repack work size for mul_mat_id (#14292) ggml-ci --- ggml/src/ggml-cpu/repack.cpp | 38 ++++++++++++++++++++++++------------ 1 file changed, 26 insertions(+), 12 deletions(-) diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 5c6715d5c..290719290 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -1163,13 +1163,24 @@ template op) { case GGML_OP_MUL_MAT: - size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); - return true; + { + size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); + return true; + } case GGML_OP_MUL_MAT_ID: - size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); - size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. - size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2]; - return true; + { + size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); + size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. + + const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert + const int64_t ne12 = op->src[1]->ne[2]; // n_tokens + + const size_t sizeof_mmid_row_mapping = sizeof(int64_t); + + size += sizeof_mmid_row_mapping*ne02*(ne12 + 1); + + return true; + } default: // GGML_ABORT("fatal error"); break; @@ -1305,14 +1316,17 @@ template wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) + - n_as * ne12 * sizeof(mmid_row_mapping))); + GGML_ASSERT(params->wsize >= + (GGML_PAD(nbw3, sizeof(int64_t)) + + n_as*(ne12 + 1)*sizeof(mmid_row_mapping)) + ); - auto * wdata = (char *) params->wdata; - auto * wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t)); - auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + auto * wdata = (char *)params->wdata; + auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t)); - struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] + // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t) + auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] // src1: float32 => param type for (int64_t i12 = 0; i12 < ne12; ++i12) { From e28c1b93fd7d3f8faf9551d962e8a65fe2122e38 Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Fri, 20 Jun 2025 04:57:36 -0700 Subject: [PATCH 2/8] cuda : synchronize graph capture and cublas handle destruction (#14288) Workarounds an issue that may cause CUDA graph capture to fail when a cuBLAS handle is destroyed in a different thread --- ggml/src/ggml-cuda/common.cuh | 18 ++------------ ggml/src/ggml-cuda/ggml-cuda.cu | 44 ++++++++++++++++++++++++++++++--- 2 files changed, 43 insertions(+), 19 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index c14a12f54..364efcaec 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -19,10 +19,10 @@ #endif #include "ggml-common.h" -#include #include #include #include +#include #include #include @@ -767,21 +767,7 @@ struct ggml_backend_cuda_context { name(GGML_CUDA_NAME + std::to_string(device)) { } - ~ggml_backend_cuda_context() { - if (copy_event != nullptr) { - CUDA_CHECK(cudaEventDestroy(copy_event)); - } - for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) { - for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) { - if (streams[i][j] != nullptr) { - CUDA_CHECK(cudaStreamDestroy(streams[i][j])); - } - } - if (cublas_handles[i] != nullptr) { - CUBLAS_CHECK(cublasDestroy(cublas_handles[i])); - } - } - } + ~ggml_backend_cuda_context(); cudaStream_t stream(int device, int stream) { if (streams[device][stream] == nullptr) { diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 80fe05073..530f541f9 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -48,6 +48,7 @@ #include #include #include +#include #include #include #include @@ -55,9 +56,8 @@ #include #include #include -#include -#include #include +#include #include #include #include @@ -515,6 +515,33 @@ std::unique_ptr ggml_backend_cuda_context::new_pool_for_device(i return std::unique_ptr(new ggml_cuda_pool_leg(device)); } +// destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error +// this lock is used to ensure that no cuBLAS handle is destroyed while a graph is being captured + +static std::mutex ggml_cuda_lock; +static std::condition_variable ggml_cuda_lock_cv; +static std::atomic ggml_cuda_lock_counter; + +ggml_backend_cuda_context::~ggml_backend_cuda_context() { + std::unique_lock lock(ggml_cuda_lock); + ggml_cuda_lock_cv.wait(lock, []{ return ggml_cuda_lock_counter.load(std::memory_order_relaxed) == 0; }); + + if (copy_event != nullptr) { + CUDA_CHECK(cudaEventDestroy(copy_event)); + } + for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) { + for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) { + if (streams[i][j] != nullptr) { + CUDA_CHECK(cudaStreamDestroy(streams[i][j])); + } + } + if (cublas_handles[i] != nullptr) { + CUBLAS_CHECK(cublasDestroy(cublas_handles[i])); + } + } +} + + // cuda buffer struct ggml_backend_cuda_buffer_context { @@ -2689,6 +2716,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph)); graph_evaluated_or_captured = true; // CUDA graph has been captured + + std::lock_guard lock(ggml_cuda_lock); + if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) { + ggml_cuda_lock_cv.notify_all(); + } } else { graph_evaluated_or_captured = true; // ggml graph has been directly evaluated } @@ -2764,7 +2796,13 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, } } - if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture + if (use_cuda_graph && cuda_graph_update_required) { + // Start CUDA graph capture + { + std::lock_guard lock(ggml_cuda_lock); + ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed); + } + CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed)); } From 88fc854b4bd2e3caf10e705e6afcbbca136f0a3c Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Fri, 20 Jun 2025 14:04:09 +0200 Subject: [PATCH 3/8] llama : improve sep token handling (#14272) --- ci/run.sh | 2 +- common/arg.cpp | 7 +++ common/common.h | 1 + convert_hf_to_gguf.py | 14 ------ examples/embedding/embedding.cpp | 34 +++++++++++-- gguf-py/gguf/constants.py | 1 + gguf-py/gguf/gguf_writer.py | 3 ++ gguf-py/gguf/vocab.py | 83 ++++++++++++++++++++++++++++++-- include/llama.h | 1 + src/llama-arch.cpp | 1 + src/llama-arch.h | 1 + src/llama-model-saver.cpp | 1 + src/llama-vocab.cpp | 24 +++++++-- src/llama-vocab.h | 1 + tools/server/utils.hpp | 16 ++++-- 15 files changed, 161 insertions(+), 29 deletions(-) diff --git a/ci/run.sh b/ci/run.sh index 940055705..e1b777c30 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -779,7 +779,7 @@ function gg_run_rerank_tiny { model_f16="${path_models}/ggml-model-f16.gguf" # for this model, the SEP token is "" - (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?hi\nwhat is panda?it's a bear\nwhat is panda?The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log + (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log # sample output # rerank score 0: 0.029 diff --git a/common/arg.cpp b/common/arg.cpp index 3dfaa71ef..c4ad85c47 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -2706,6 +2706,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.embd_sep = value; } ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + add_opt(common_arg( + {"--cls-separator"}, "STRING", + "separator of classification sequences (default \\t) for example \"<#seq#>\"", + [](common_params & params, const std::string & value) { + params.cls_sep = value; + } + ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); add_opt(common_arg( {"--host"}, "HOST", string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()), diff --git a/common/common.h b/common/common.h index 5710c4e97..e08a59eae 100644 --- a/common/common.h +++ b/common/common.h @@ -358,6 +358,7 @@ struct common_params { int32_t embd_normalize = 2; // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm) std::string embd_out = ""; // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix std::string embd_sep = "\n"; // separator of embeddings + std::string cls_sep = "\t"; // separator of classification sequences // server params int32_t port = 8080; // server listens on this network port diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 2e08db345..2fe76589e 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -2145,7 +2145,6 @@ class Llama4Model(LlamaModel): def set_vocab(self): self._set_vocab_gpt2() - self.gguf_writer.add_add_bos_token(True) def set_gguf_parameters(self): super().set_gguf_parameters() @@ -3918,9 +3917,6 @@ class BertModel(TextModel): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) - self.gguf_writer.add_add_bos_token(True) - self.gguf_writer.add_add_eos_token(True) - @ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification") class DistilBertModel(BertModel): @@ -3962,8 +3958,6 @@ class RobertaModel(BertModel): bpe_tok_path = self.dir_model / "tokenizer.json" if bpe_tok_path.exists(): self._set_vocab_gpt2() - self.gguf_writer.add_add_bos_token(True) - self.gguf_writer.add_add_eos_token(True) # we need this to validate the size of the token_type embeddings # though currently we are passing all zeros to the token_type embeddings @@ -4848,8 +4842,6 @@ class JinaBertV2Model(BertModel): self.gguf_writer.add_token_type_count(2) else: raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel') - self.gguf_writer.add_add_bos_token(True) - self.gguf_writer.add_add_eos_token(True) @ModelBase.register("OpenELMForCausalLM") @@ -5451,9 +5443,6 @@ class T5Model(TextModel): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) - self.gguf_writer.add_add_bos_token(False) - self.gguf_writer.add_add_eos_token(True) - def set_gguf_parameters(self): if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: logger.warning("Couldn't find context length in config.json, assuming default value of 512") @@ -5591,9 +5580,6 @@ class T5EncoderModel(TextModel): special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) - self.gguf_writer.add_add_bos_token(False) - self.gguf_writer.add_add_eos_token(True) - def set_gguf_parameters(self): if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None: logger.warning("Couldn't find context length in config.json, assuming default value of 512") diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 681929d27..0ec2999a0 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -133,10 +133,36 @@ int main(int argc, char ** argv) { // max batch size const uint64_t n_batch = params.n_batch; + // get added sep and eos token, if any + const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : ""; + const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : ""; + // tokenize the prompts and trim std::vector> inputs; for (const auto & prompt : prompts) { - auto inp = common_tokenize(ctx, prompt, true, true); + std::vector inp; + + // split classification pairs and insert expected separator tokens + if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) { + std::vector pairs = split_lines(prompt, params.cls_sep); + std::string final_prompt; + + for (size_t i = 0; i < pairs.size(); i++) { + final_prompt += pairs[i]; + if (i != pairs.size() - 1) { + if (!added_eos_token.empty()) { + final_prompt += added_eos_token; + } + if (!added_sep_token.empty()) { + final_prompt += added_sep_token; + } + } + } + + inp = common_tokenize(ctx, final_prompt, true, true); + } else { + inp = common_tokenize(ctx, prompt, true, true); + } if (inp.size() > n_batch) { LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n", __func__, (long long int) inp.size(), (long long int) n_batch); @@ -145,11 +171,11 @@ int main(int argc, char ** argv) { inputs.push_back(inp); } - // check if the last token is SEP + // check if the last token is SEP/EOS // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true' for (auto & inp : inputs) { - if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) { - LOG_WRN("%s: last token in the prompt is not SEP\n", __func__); + if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) { + LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__); LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__); } } diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 834a1d5e1..0429b0aaf 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -198,6 +198,7 @@ class Keys: MASK_ID = "tokenizer.ggml.mask_token_id" ADD_BOS = "tokenizer.ggml.add_bos_token" ADD_EOS = "tokenizer.ggml.add_eos_token" + ADD_SEP = "tokenizer.ggml.add_sep_token" ADD_PREFIX = "tokenizer.ggml.add_space_prefix" REMOVE_EXTRA_WS = "tokenizer.ggml.remove_extra_whitespaces" PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap" diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 54ca0c33f..b9b63d052 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -891,6 +891,9 @@ class GGUFWriter: def add_add_eos_token(self, value: bool) -> None: self.add_bool(Keys.Tokenizer.ADD_EOS, value) + def add_add_sep_token(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.ADD_SEP, value) + def add_add_space_prefix(self, value: bool) -> None: self.add_bool(Keys.Tokenizer.ADD_PREFIX, value) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index 44d066ee7..6c4d3a422 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -119,6 +119,7 @@ class SpecialVocab: logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping') def _try_load_from_tokenizer_json(self, path: Path) -> bool: + tokenizer = None tokenizer_file = path / 'tokenizer.json' if tokenizer_file.is_file(): with open(tokenizer_file, encoding = 'utf-8') as f: @@ -152,11 +153,87 @@ class SpecialVocab: added_tokens = tokenizer.get('added_tokens', {}) else: added_tokens = {} + tokenizer_config = None tokenizer_config_file = path / 'tokenizer_config.json' - if not tokenizer_config_file.is_file(): + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, encoding = 'utf-8') as f: + tokenizer_config = json.load(f) + if tokenizer: + special_bos = (tokenizer_config or {}).get('bos_token') + special_cls = (tokenizer_config or {}).get('cls_token') + special_eos = (tokenizer_config or {}).get('eos_token') + special_sep = (tokenizer_config or {}).get('sep_token') + if not special_bos and special_cls and tokenizer_config: + tokenizer_config['bos_token'] = special_bos = special_cls + if not special_eos and special_sep and tokenizer_config: + tokenizer_config['eos_token'] = special_eos = special_sep + post_processor = tokenizer.get('post_processor', {}) + for processor in post_processor.get('processors', [post_processor]): + if processor.get('type') == 'RobertaProcessing': + self.add_special_token['bos'] = True + self.add_special_token['eos'] = True + self.add_special_token['sep'] = True + if not special_cls and tokenizer_config: + special_cls = processor.get('cls', [special_bos])[0] + tokenizer_config['cls_token'] = special_cls + if not special_sep and tokenizer_config: + special_sep = processor.get('sep', [special_eos])[0] + tokenizer_config['sep_token'] = special_sep + continue + # Crude parsing of TemplateProcessing to determine if BOS/SEP/EOS should be added + # Only works with simple templates, **will** get it wrong on unusual sequences + if processor.get('type') == 'TemplateProcessing': + tmpl_single = processor.get('single', []) + tmpl_pair = processor.get('pair', []) + special_first = None + special_last = None + if len(tmpl_single) > 1: + if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'): + if not tokenizer_config: + special_bos = special_first + self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False + if special_first not in (special_bos, special_cls): + logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing') + if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'): + if not tokenizer_config: + special_eos = special_last + self.add_special_token['eos'] = True if special_last == special_eos else False + if special_last != special_eos: + logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing') + if tmpl_pair: + seq_start = 1 if tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0 + seq_stop = -1 if tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None + if seq_start == 0 or seq_stop is None: + logger.warning('TemplateProcessing leading/trailing special tokens do not match TemplateProcessing') + if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]: + tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id') + tmpl_b = tmpl_pair[-1].get('Sequence', {}).get('id') + if tmpl_a != 'A' or tmpl_b != 'B': + logger.warning(f'Unknown sequence {tmpl_a}...{tmpl_b} in TemplateProcessing') + # A [sep] [eos] B + if tmpl_a == 'A' and tmpl_b == 'B' and (tmpl_pair := tmpl_pair[1:-1]): + add_sep = False + if special_entry := tmpl_pair[0].get('SpecialToken', {}).get('id'): + if special_entry in (special_sep, special_eos) and not special_last: + add_sep = True + if special_entry not in (special_sep, special_eos): + logger.warning(f'Unknown separator token {special_entry!r} in TemplateProcessing') + else: + logger.warning(f'Unknown middle sequence {tmpl_pair[0]!r} in TemplateProcessing') + if len(tmpl_pair) == 2: + if special_entry := tmpl_pair[1].get('SpecialToken', {}).get('id'): + if special_entry in (special_sep, special_eos): + add_sep = True + if special_entry not in (special_sep, special_eos): + logger.warning(f'Unknown second separator token {special_entry!r} in TemplateProcessing') + else: + logger.warning(f'Unknown second middle sequence {tmpl_pair[1]!r} in TemplateProcessing') + self.add_special_token['sep'] = add_sep + if add_sep and not special_sep and tokenizer_config: + tokenizer_config['sep_token'] = special_eos + continue + if not tokenizer_config: return True - with open(tokenizer_config_file, encoding = 'utf-8') as f: - tokenizer_config = json.load(f) chat_template_alt = None chat_template_file = path / 'chat_template.json' if chat_template_file.is_file(): diff --git a/include/llama.h b/include/llama.h index 635508b10..3475d5965 100644 --- a/include/llama.h +++ b/include/llama.h @@ -1044,6 +1044,7 @@ extern "C" { LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab); LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab); + LLAMA_API bool llama_vocab_get_add_sep(const struct llama_vocab * vocab); LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab); LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab); diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 0bc60565d..8dadef204 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -198,6 +198,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_TOKENIZER_MASK_ID, "tokenizer.ggml.mask_token_id" }, { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" }, { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" }, + { LLM_KV_TOKENIZER_ADD_SEP, "tokenizer.ggml.add_sep_token" }, { LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" }, { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, "tokenizer.ggml.remove_extra_whitespaces" }, { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap" }, diff --git a/src/llama-arch.h b/src/llama-arch.h index 51b242c66..5b0230c15 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -194,6 +194,7 @@ enum llm_kv { LLM_KV_TOKENIZER_MASK_ID, LLM_KV_TOKENIZER_ADD_BOS, LLM_KV_TOKENIZER_ADD_EOS, + LLM_KV_TOKENIZER_ADD_SEP, LLM_KV_TOKENIZER_ADD_PREFIX, LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp index a70b98923..563823dc3 100644 --- a/src/llama-model-saver.cpp +++ b/src/llama-model-saver.cpp @@ -228,6 +228,7 @@ void llama_model_saver::add_kv_from_model() { // add_kv(LLM_KV_TOKENIZER_MASK_ID, ???); add_kv(LLM_KV_TOKENIZER_ADD_BOS, vocab.get_add_bos()); add_kv(LLM_KV_TOKENIZER_ADD_EOS, vocab.get_add_eos()); + add_kv(LLM_KV_TOKENIZER_ADD_SEP, vocab.get_add_sep()); add_kv(LLM_KV_TOKENIZER_ADD_PREFIX, vocab.get_add_space_prefix()); add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, vocab.get_remove_extra_whitespaces()); add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, vocab.get_precompiled_charsmap()); diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index d90f1d6b1..4ab120d9b 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -1269,6 +1269,7 @@ struct llama_vocab::impl { bool add_space_prefix = false; bool add_bos = false; bool add_eos = false; + bool add_sep = false; bool ignore_merges = false; bool clean_spaces = false; // clean_up_tokenization_spaces bool remove_extra_whitespaces = false; @@ -1421,6 +1422,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { special_sep_id = 102; special_pad_id = 0; special_mask_id = 103; + + add_sep = true; } else if (tokenizer_model == "gpt2") { type = LLAMA_VOCAB_TYPE_BPE; @@ -1550,12 +1553,15 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "jina-es" || tokenizer_pre == "jina-de" || tokenizer_pre == "gigachat" || - tokenizer_pre == "jina-v1-en" || tokenizer_pre == "jina-v2-es" || - tokenizer_pre == "jina-v2-de" || + tokenizer_pre == "jina-v2-de") { + pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2; + } else if ( + tokenizer_pre == "jina-v1-en" || tokenizer_pre == "jina-v2-code" || tokenizer_pre == "roberta-bpe") { pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2; + add_sep = true; } else if ( tokenizer_pre == "refact") { pre_type = LLAMA_VOCAB_PRE_TYPE_REFACT; @@ -1665,6 +1671,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { clean_spaces = true; add_bos = true; add_eos = false; + add_sep = true; } else if (type == LLAMA_VOCAB_TYPE_UGM) { pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT; add_bos = false; @@ -1801,7 +1808,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { } } - // Handle add_bos and add_eos + // Handle add_bos, add_eos and add_sep { bool temp = true; @@ -1811,6 +1818,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) { add_eos = temp; } + if (ml.get_key(LLM_KV_TOKENIZER_ADD_SEP, temp, false)) { + add_sep = temp; + } } // auto-detect special tokens by text @@ -3000,6 +3010,10 @@ bool llama_vocab::get_add_eos() const { return pimpl->add_eos; } +bool llama_vocab::get_add_sep() const { + return pimpl->add_sep; +} + bool llama_vocab::get_ignore_merges() const { return pimpl->ignore_merges; } @@ -3191,6 +3205,10 @@ bool llama_vocab_get_add_eos(const struct llama_vocab * vocab) { return vocab->get_add_eos(); } +bool llama_vocab_get_add_sep(const struct llama_vocab * vocab) { + return vocab->get_add_sep(); +} + llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab) { return vocab->token_fim_pre(); } diff --git a/src/llama-vocab.h b/src/llama-vocab.h index daa6cf308..40e4d1c05 100644 --- a/src/llama-vocab.h +++ b/src/llama-vocab.h @@ -74,6 +74,7 @@ struct llama_vocab { bool get_add_space_prefix () const; bool get_add_bos () const; bool get_add_eos () const; + bool get_add_sep () const; bool get_ignore_merges () const; bool get_clean_spaces () const; bool get_remove_extra_whitespaces () const; diff --git a/tools/server/utils.hpp b/tools/server/utils.hpp index f3e0392a4..f8fab2c86 100644 --- a/tools/server/utils.hpp +++ b/tools/server/utils.hpp @@ -271,12 +271,20 @@ static llama_tokens format_rerank(const struct llama_vocab * vocab, const llama_ } result.reserve(doc.size() + query.size() + 4); - result.push_back(llama_vocab_bos(vocab)); + if (llama_vocab_get_add_bos(vocab)) { + result.push_back(llama_vocab_bos(vocab)); + } result.insert(result.end(), query.begin(), query.end()); - result.push_back(eos_token); - result.push_back(llama_vocab_sep(vocab)); + if (llama_vocab_get_add_eos(vocab)) { + result.push_back(eos_token); + } + if (llama_vocab_get_add_sep(vocab)) { + result.push_back(llama_vocab_sep(vocab)); + } result.insert(result.end(), doc.begin(), doc.end()); - result.push_back(eos_token); + if (llama_vocab_get_add_eos(vocab)) { + result.push_back(eos_token); + } return result; } From 6369be07359d03723f38c0a4a014ff0f698a0738 Mon Sep 17 00:00:00 2001 From: Christian Kastner Date: Fri, 20 Jun 2025 12:17:32 +0000 Subject: [PATCH 4/8] Implement GGML_CPU_ALL_VARIANTS for PowerPC (#14286) * Add PowerPC feature detection and scoring * ggml-cpu: Implement GGML_CPU_ALL_VARIANTS for PowerPC * ggml-cpu: Delay some initializations until function is called When using GGML_BACKEND_DL=ON, these initializations might use instructions that are not supported by the current CPU. --------- Co-authored-by: Diego Devesa --- ggml/src/CMakeLists.txt | 17 ++++ ggml/src/ggml-cpu/CMakeLists.txt | 21 +++++ ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp | 82 ++++++++++++++++++++ ggml/src/ggml-cpu/repack.cpp | 29 +++---- 4 files changed, 135 insertions(+), 14 deletions(-) create mode 100644 ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 0c453741b..9cb2c228d 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -286,6 +286,10 @@ function(ggml_add_cpu_backend_variant tag_name) foreach (feat ${ARGN}) set(GGML_INTERNAL_${feat} ON) endforeach() + elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC") + foreach (feat ${ARGN}) + set(GGML_INTERNAL_${feat} ON) + endforeach() endif() ggml_add_cpu_backend_variant_impl(${tag_name}) @@ -337,6 +341,19 @@ if (GGML_CPU_ALL_VARIANTS) else() message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}") endif() + elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC") + if (CMAKE_SYSTEM_NAME MATCHES "Linux") + ggml_add_cpu_backend_variant(power0) + ggml_add_cpu_backend_variant(power7_1 POWER7) + ggml_add_cpu_backend_variant(power7_2 POWER7 VSX) + ggml_add_cpu_backend_variant(power8_1 POWER8) + ggml_add_cpu_backend_variant(power8_2 POWER8 VSX) + ggml_add_cpu_backend_variant(power9 POWER9 VSX) + ggml_add_cpu_backend_variant(power10 POWER10 VSX) + ggml_add_cpu_backend_variant(power11 POWER11 VSX) + else() + message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}") + endif() else() message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}") endif() diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index 52cae778c..71b1d67b8 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -388,6 +388,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name) else() list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64) endif() + elseif(GGML_CPU_ALL_VARIANTS) + # Begin with the lowest baseline + set(ARCH_DEFINITIONS "") + + # When a feature is selected, bump the MCPU to the first + # version that supported it + foreach(PVER RANGE 7 11) + if(DEFINED GGML_INTERNAL_POWER${PVER}) + set(POWERPC_MCPU "power${PVER}") + list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER}) + endif() + endforeach() + if (GGML_INTERNAL_VSX) + list(APPEND ARCH_DEFINITIONS GGML_USE_VSX) + list(APPEND ARCH_FLAGS -mvsx) + endif() + + if (DEFINED POWERPC_MCPU) + list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU}) + endif() + ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS}) else() if (GGML_CPU_POWERPC_CPUTYPE) list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE}) diff --git a/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp b/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp new file mode 100644 index 000000000..fedd64302 --- /dev/null +++ b/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp @@ -0,0 +1,82 @@ +# include "ggml-backend-impl.h" + +#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__) + +#if defined(__linux__) +#include +#endif + +#include + +struct powerpc_features { + std::string platform = ""; + int power_version = -1; + + bool has_vsx = false; + + powerpc_features() { +#if defined(__linux__) + unsigned long auxval = getauxval(AT_PLATFORM); + if (auxval) { + platform = std::string(reinterpret_cast(auxval)); + // TBD: Do systems exist that return this in uppercase? + if (platform.substr(0, 5) == "power") { + // Extractt a numeric suffix, if one exists + int vpos = -1; + for (int i = platform.length() - 1; i >= 0; i--) { + if (std::isdigit(platform[i])) { + vpos = i; + } else { + break; + } + } + if (vpos > -1) { + power_version = std::stoi(platform.substr(vpos)); + } + } + } +#endif + if (power_version >= 9) { + has_vsx = true; + } + } +}; + +static int ggml_backend_cpu_powerpc_score() { + int score = 1; + powerpc_features pf; + +// Platform scores +#if defined(GGML_USE_POWER7) + if (pf.power_version < 7) { return 0; } + score += 1<<1; +#endif +#if defined(GGML_USE_POWER8) + if (pf.power_version < 8) { return 0; } + score += 1<<2; +#endif +#if defined(GGML_USE_POWER9) + if (pf.power_version < 9) { return 0; } + score += 1<<3; +#endif +#if defined(GGML_USE_POWER10) + if (pf.power_version < 10) { return 0; } + score += 1<<4; +#endif +#if defined(GGML_USE_POWER11) + if (pf.power_version < 11) { return 0; } + score += 1<<5; +#endif + +// Feature scores +#if defined(GGML_USE_VSX) + if (!pf.has_vsx) { return 0; } + score += 1<<6; +#endif + + return score; +} + +GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score) + +#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__) diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 290719290..692c53e01 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -1411,44 +1411,45 @@ template q4_0_4x4_q8_0; -static const tensor_traits q4_0_4x8_q8_0; -static const tensor_traits q4_0_8x8_q8_0; -static const tensor_traits q4_K_8x8_q8_K; - -// instance for IQ4 -static const tensor_traits iq4_nl_4x4_q8_0; - } // namespace ggml::cpu::repack static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) { + + // instance for Q4 + static const ggml::cpu::repack::tensor_traits q4_0_4x4_q8_0; + static const ggml::cpu::repack::tensor_traits q4_0_4x8_q8_0; + static const ggml::cpu::repack::tensor_traits q4_0_8x8_q8_0; + static const ggml::cpu::repack::tensor_traits q4_K_8x8_q8_K; + + // instance for IQ4 + static const ggml::cpu::repack::tensor_traits iq4_nl_4x4_q8_0; + if (cur->type == GGML_TYPE_Q4_0) { if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) { if (cur->ne[1] % 8 == 0) { - return &ggml::cpu::repack::q4_0_8x8_q8_0; + return &q4_0_8x8_q8_0; } } if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { if (cur->ne[1] % 4 == 0) { - return &ggml::cpu::repack::q4_0_4x8_q8_0; + return &q4_0_4x8_q8_0; } } if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { if (cur->ne[1] % 4 == 0) { - return &ggml::cpu::repack::q4_0_4x4_q8_0; + return &q4_0_4x4_q8_0; } } } else if (cur->type == GGML_TYPE_Q4_K) { if (ggml_cpu_has_avx2()) { if (cur->ne[1] % 8 == 0) { - return &ggml::cpu::repack::q4_K_8x8_q8_K; + return &q4_K_8x8_q8_K; } } } else if (cur->type == GGML_TYPE_IQ4_NL) { if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { if (cur->ne[1] % 4 == 0) { - return &ggml::cpu::repack::iq4_nl_4x4_q8_0; + return &iq4_nl_4x4_q8_0; } } } From 8308f98c7fb778e54bf75538f5234d8bd20915e9 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Nicol=C3=B2=20Scipione?= Date: Fri, 20 Jun 2025 15:07:21 +0200 Subject: [PATCH 5/8] sycl: add usage of enqueue_functions extension (#14244) * Add header and namespace to use enqueue_functions extension * Convert submit and parallel_for to use new extension in convert.cpp * Convert submit and parallel_for to use extension in ggml-sycl.cpp * Convert submit and parallel_for to use extension in gla.cpp * Convert submit and parallel_for in mmq.cpp * Convert submit and parallel_for in mmvq.cpp * Convert submit and parallel_for in remaining files * Convert all simple parallel_for to nd_launch from enqueue_functions extension * Wrapping extension in general function Create a general function that enable the enqueue_functions extension if it is enable in the compiler, otherwise call the general SYCL function to launch kernels. --------- Signed-off-by: nscipione --- ggml/src/ggml-sycl/binbcast.cpp | 11 +- ggml/src/ggml-sycl/concat.cpp | 69 +++--- ggml/src/ggml-sycl/conv.cpp | 14 +- ggml/src/ggml-sycl/convert.cpp | 265 +++++++++------------- ggml/src/ggml-sycl/cpy.cpp | 166 ++++++++------ ggml/src/ggml-sycl/dmmv.cpp | 116 ++++------ ggml/src/ggml-sycl/dpct/helper.hpp | 32 ++- ggml/src/ggml-sycl/element_wise.cpp | 258 +++++++++------------ ggml/src/ggml-sycl/getrows.cpp | 15 +- ggml/src/ggml-sycl/ggml-sycl.cpp | 93 ++++---- ggml/src/ggml-sycl/gla.cpp | 4 +- ggml/src/ggml-sycl/im2col.cpp | 2 +- ggml/src/ggml-sycl/mmq.cpp | 140 +++++------- ggml/src/ggml-sycl/mmvq.cpp | 333 +++++++++++----------------- ggml/src/ggml-sycl/norm.cpp | 129 +++++------ ggml/src/ggml-sycl/rope.cpp | 44 ++-- ggml/src/ggml-sycl/softmax.cpp | 6 +- ggml/src/ggml-sycl/tsembd.cpp | 11 +- ggml/src/ggml-sycl/wkv.cpp | 28 +-- 19 files changed, 750 insertions(+), 986 deletions(-) diff --git a/ggml/src/ggml-sycl/binbcast.cpp b/ggml/src/ggml-sycl/binbcast.cpp index 0a3883ae1..741630dba 100644 --- a/ggml/src/ggml-sycl/binbcast.cpp +++ b/ggml/src/ggml-sycl/binbcast.cpp @@ -225,9 +225,9 @@ struct bin_bcast_sycl { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * - sycl::range<3>(1, 1, block_size), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * sycl::range<3>(1, 1, block_size), sycl::range<3>(1, 1, block_size)), [=](sycl::nd_item<3> item_ct1) { k_bin_bcast_unravel( @@ -246,9 +246,8 @@ struct bin_bcast_sycl { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_bin_bcast(src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02, s03, s11, s12, s13, diff --git a/ggml/src/ggml-sycl/concat.cpp b/ggml/src/ggml-sycl/concat.cpp index 7aa91c861..3501484a1 100644 --- a/ggml/src/ggml-sycl/concat.cpp +++ b/ggml/src/ggml-sycl/concat.cpp @@ -89,33 +89,24 @@ static void concat_f32_sycl(const float *x, const float *y, float *dst, sycl::range<3> gridDim(ne2, ne1, num_blocks); switch (dim) { case 0: - stream->parallel_for( - sycl::nd_range<3>(gridDim * - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1); - }); - break; + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1); }); + break; case 1: - stream->parallel_for( - sycl::nd_range<3>(gridDim * - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1); - }); - break; + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1); }); + break; // dim >=2 will be dispatched to the default path default: - stream->parallel_for( - sycl::nd_range<3>(gridDim * - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1); - }); - break; + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1); }); + break; } } @@ -129,33 +120,29 @@ static void concat_f32_sycl_non_cont( int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2, uint64_t nb3, int32_t dim) { sycl::range<3> gridDim(ne3, ne2, ne1); - stream->parallel_for( - sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - int64_t i3 = item_ct1.get_group(0); - int64_t i2 = item_ct1.get_group(1); - int64_t i1 = item_ct1.get_group(2); + sycl_parallel_for(stream, sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + int64_t i3 = item_ct1.get_group(0); + int64_t i2 = item_ct1.get_group(1); + int64_t i1 = item_ct1.get_group(2); - int64_t o[4] = {0, 0, 0, 0}; - o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03)); + int64_t o[4] = { 0, 0, 0, 0 }; + o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03)); - const float *x; + const float * x; - for (int i0 = item_ct1.get_local_id(2); i0 < ne0; - i0 += item_ct1.get_local_range(2)) { + for (int i0 = item_ct1.get_local_id(2); i0 < ne0; i0 += item_ct1.get_local_range(2)) { if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { - x = (const float *)(src0 + (i3)*nb03 + (i2)*nb02 + (i1)*nb01 + - (i0)*nb00); + x = (const float *) (src0 + (i3) *nb03 + (i2) *nb02 + (i1) *nb01 + (i0) *nb00); } else { - x = (const float *)(src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + - (i1 - o[1]) * nb11 + (i0 - o[0]) * nb10); + x = (const float *) (src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + (i1 - o[1]) * nb11 + + (i0 - o[0]) * nb10); } float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0); *y = *x; - } - }); + } + }); } void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { diff --git a/ggml/src/ggml-sycl/conv.cpp b/ggml/src/ggml-sycl/conv.cpp index 475bd34a2..c2f991e8d 100644 --- a/ggml/src/ggml-sycl/conv.cpp +++ b/ggml/src/ggml-sycl/conv.cpp @@ -59,16 +59,10 @@ static void conv_transpose_1d_f32_f32_sycl( const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE; const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE); const sycl::range<3> block_nums(1, 1, num_blocks); - stream->parallel_for( - sycl::nd_range<3>( - block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - conv_transpose_1d_kernel( - s0, output_size, - src0_ne0, src0_ne1, src0_ne2, - src1_ne0, dst_ne0, - src0, src1, dst, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + conv_transpose_1d_kernel(s0, output_size, src0_ne0, src0_ne1, src0_ne2, src1_ne0, dst_ne0, src0, src1, dst, + item_ct1); + }); } void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index 96d2583b1..0ef567122 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -33,14 +33,11 @@ static void dequantize_block_sycl(const void *__restrict__ vx, { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>( - sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block(vx, y, k, item_ct1); - }); + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block(vx, y, k, item_ct1); }); } } @@ -53,24 +50,18 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q2_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q2_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); }); } #endif @@ -85,24 +76,18 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q3_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q3_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); }); } #endif } @@ -116,12 +101,9 @@ static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_0(vx, y, nb32, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_0(vx, y, nb32, item_ct1); }); } } @@ -135,13 +117,12 @@ static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int int constexpr WARP_K = WARP_SIZE * QK4_0; const int n_warp = (k + WARP_K - 1) / WARP_K; GGML_ASSERT(k % 2 == 0); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) * - sycl::range<3>(1, 1, WARP_SIZE), - sycl::range<3>(1, 1, WARP_SIZE)), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{ - dequantize_block_q4_0_reorder(vx, y, k, item_ct1); - }); - + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) * sycl::range<3>(1, 1, WARP_SIZE), + sycl::range<3>(1, 1, WARP_SIZE)), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_block_q4_0_reorder(vx, y, k, item_ct1); + }); } template @@ -153,12 +134,9 @@ static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_1(vx, y, nb32, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_1(vx, y, nb32, item_ct1); }); } } @@ -171,14 +149,13 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor scale_local_acc(sycl::range<1>(12), cgh); - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1); - }); + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1); + }); }); } } @@ -191,13 +168,13 @@ static void dequantize_row_q4_K_sycl_reorder(const void * vx, dst_t * y, const i dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->submit([&](sycl::handler & cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor scale_local_acc(sycl::range<1>(12), cgh); - cgh.parallel_for(sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)), - [=](sycl::nd_item<1> item_ct1) { - dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb); - }); + sycl_parallel_for<1>(cgh, sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)), + [=](sycl::nd_item<1> item_ct1) { + dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb); + }); }); } @@ -210,24 +187,18 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q5_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q5_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); }); } #endif @@ -242,24 +213,18 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 64), - sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q6_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); }); } #else { dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_q6_K(vx, y, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); }); } #endif @@ -271,9 +236,9 @@ static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const i dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), - [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); }); } template @@ -284,15 +249,10 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq1_s( - vx, y, item_ct1, iq1s_grid_gpu - ); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_s(vx, y, item_ct1, iq1s_grid_gpu); }); }); } } @@ -305,15 +265,10 @@ static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq1_m( - vx, y, item_ct1, iq1s_grid_gpu - ); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_m(vx, y, item_ct1, iq1s_grid_gpu); }); }); } } @@ -326,15 +281,12 @@ static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int64_t dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq2_xxs( - vx, y, item_ct1, iq2xxs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq2_xxs(vx, y, item_ct1, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs); + }); }); } } @@ -347,15 +299,12 @@ static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int64_t k dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq2_xs( - vx, y, item_ct1, iq2xs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq2_xs(vx, y, item_ct1, iq2xs_grid, ksigns_iq2xs, kmask_iq2xs); + }); }); } } @@ -368,13 +317,10 @@ static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq2_s(vx, y, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq2_s(vx, y, item_ct1); }); }); } } @@ -388,15 +334,12 @@ static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int64_t dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq3_xxs( - vx, y, item_ct1, iq3xxs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_iq3_xxs(vx, y, item_ct1, iq3xxs_grid, ksigns_iq2xs, kmask_iq2xs); + }); }); } } @@ -409,14 +352,10 @@ static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int64_t k, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq3_s( - vx, y, item_ct1, kmask_iq2xs, iq3s_grid); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq3_s(vx, y, item_ct1, kmask_iq2xs, iq3s_grid); }); }); } } @@ -432,14 +371,11 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int64_t k dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq4_xs(vx, y, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, + sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_xs(vx, y, item_ct1); }); }); } #endif @@ -453,14 +389,11 @@ static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * - sycl::range<3>(1, 1, 32), - sycl::range<3>(1, 1, 32)), - [=](sycl::nd_item<3> item_ct1) { - dequantize_block_iq4_nl(vx, y, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for( + cgh, + sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_nl(vx, y, item_ct1); }); }); } } diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp index bec137140..1ffd7f122 100644 --- a/ggml/src/ggml-sycl/cpy.cpp +++ b/ggml/src/ggml-sycl/cpy.cpp @@ -413,7 +413,8 @@ static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -431,7 +432,8 @@ static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -449,7 +451,8 @@ static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -465,11 +468,11 @@ static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK8_0 == 0); const int num_blocks = ne / QK8_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -477,11 +480,11 @@ static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_q_f32(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -490,11 +493,11 @@ static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK4_0 == 0); const int num_blocks = ne / QK4_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -502,8 +505,9 @@ static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -516,11 +520,11 @@ static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK4_1 == 0); const int num_blocks = ne / QK4_1; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -528,8 +532,9 @@ static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -542,11 +547,11 @@ static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK5_0 == 0); const int num_blocks = ne / QK5_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -554,8 +559,9 @@ static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -568,11 +574,11 @@ static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, c const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK5_1 == 0); const int num_blocks = ne / QK5_1; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, - ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -580,8 +586,9 @@ static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, c const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ne; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { cpy_q_f32, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); @@ -594,11 +601,11 @@ static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne, const int nb12, const int nb13, queue_ptr stream) { GGML_ASSERT(ne % QK4_NL == 0); const int num_blocks = ne / QK4_NL; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, - ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, @@ -609,7 +616,8 @@ static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, co { dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -628,7 +636,8 @@ static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, co // dpct::has_capability_or_fail(stream->get_device(), // {sycl::aspect::fp16}); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -647,7 +656,8 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co // dpct::has_capability_or_fail(stream->get_device(), // {sycl::aspect::fp16}); - stream->parallel_for( + sycl_parallel_for( + stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { @@ -662,11 +672,13 @@ static void ggml_cpy_q8_0_q8_0(const char * cx, char * cdst, const int ne, const const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -675,11 +687,13 @@ static void ggml_cpy_q5_0_q5_0(const char * cx, char * cdst, const int ne, const const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -689,11 +703,13 @@ static void ggml_cpy_q5_1_q5_1(const char * cx, char * cdst, const int ne, const const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -702,10 +718,13 @@ static void ggml_cpy_q4_0_q4_0(const char * cx, char * cdst, const int ne, const const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } @@ -715,10 +734,13 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const const int nb12, const int nb13, queue_ptr stream) { const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); } void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try { diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp index 4f2760110..70579c0c3 100644 --- a/ggml/src/ggml-sycl/dmmv.cpp +++ b/ggml/src/ggml-sycl/dmmv.cpp @@ -208,12 +208,10 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, - nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -877,12 +875,11 @@ static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloa dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec_reorder( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec_reorder(vx, y, dst, ncols, + nrows, item_ct1); + }); } } @@ -900,12 +897,10 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -921,12 +916,10 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -942,12 +935,10 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -963,12 +954,10 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -984,12 +973,10 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec(vx, y, dst, ncols, nrows, item_ct1); + }); } } @@ -1002,11 +989,10 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); + }); } static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, @@ -1018,11 +1004,10 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); + }); } static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, @@ -1034,11 +1019,10 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); + }); } static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, @@ -1047,11 +1031,10 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, dpct::queue_ptr stream) { GGML_ASSERT(ncols % QK_K == 0); const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); + }); } static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, @@ -1063,11 +1046,10 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, const int block_num_y = (nrows + ny - 1) / ny; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { - dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { + dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); + }); } void ggml_sycl_op_dequantize_mul_mat_vec( diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index d538965b0..27c727860 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -13,10 +13,10 @@ #ifndef GGML_SYCL_DPCT_HELPER_HPP #define GGML_SYCL_DPCT_HELPER_HPP +#include #include #include #include -#include #ifdef GGML_SYCL_USE_INTEL_ONEMKL #include @@ -118,6 +118,36 @@ inline auto get_onemath_backend(sycl::queue& queue) #endif } +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + namespace syclex = sycl::ext::oneapi::experimental; +#endif + +template +__dpct_inline__ void sycl_parallel_for(sycl::handler & cgh, sycl::nd_range nd_range, Func && func) { +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + syclex::nd_launch(cgh, nd_range, func); +#else + cgh.parallel_for(nd_range, func); +#endif +} + +template +__dpct_inline__ void sycl_parallel_for(sycl::queue * q, sycl::nd_range nd_range, Func && func) { +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + syclex::nd_launch(*q, nd_range, func); +#else + q->parallel_for(nd_range, func); +#endif +} + +template __dpct_inline__ void sycl_launch(sycl::queue * stream, Func && func) { +#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS + syclex::submit(*stream, func); +#else + stream->submit(func); +#endif +} + namespace dpct { typedef sycl::queue *queue_ptr; diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cpp index 5b7c4f0b4..c56924ce8 100644 --- a/ggml/src/ggml-sycl/element_wise.cpp +++ b/ggml/src/ggml-sycl/element_wise.cpp @@ -329,60 +329,51 @@ static void acc_f32_sycl(const float *x, const float *y, float *dst, const int ne12, const int nb1, const int nb2, const int offset, queue_ptr stream) { int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset, - item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset, item_ct1); + }); } template static void gelu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - gelu(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { gelu(x, dst, k, item_ct1); }); } template static void silu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - silu(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { silu(x, dst, k, item_ct1); }); } template static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) { // hard code for now const int num_blocks = ceil_div(k, 256); - stream->parallel_for( - sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { - sgn(x, dst, k, item_ct1); - }); + sycl_parallel_for( + stream, sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), + [=](sycl::nd_item<3> item_ct1) { sgn(x, dst, k, item_ct1); }); } template static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) { // hard code for now const int num_blocks = ceil_div(k, 256); - stream->parallel_for( - sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { - abs_op(x, dst, k, item_ct1); - }); + sycl_parallel_for( + stream, + sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), + [=](sycl::nd_item<3> item_ct1) { abs_op(x, dst, k, item_ct1); }); } @@ -390,23 +381,20 @@ template static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) { // hard code for now const int num_blocks = ceil_div(k, 256); - stream->parallel_for( - sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { - elu_op(x, dst, k, item_ct1); - }); + sycl_parallel_for( + stream, + sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), + [=](sycl::nd_item<3> item_ct1) { elu_op(x, dst, k, item_ct1); }); } template static void gelu_quick_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - gelu_quick(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { gelu_quick(x, dst, k, item_ct1); }); } @@ -414,169 +402,133 @@ template static void gelu_erf_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - gelu_erf(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { gelu_erf(x, dst, k, item_ct1); }); } template static void tanh_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - tanh(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { tanh(x, dst, k, item_ct1); }); } template static void relu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - relu(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { relu(x, dst, k, item_ct1); }); } template static void hardsigmoid_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - hardsigmoid(x, dst, k, item_ct1); - }); + [=](sycl::nd_item<3> item_ct1) { hardsigmoid(x, dst, k, item_ct1); }); } template static void hardswish_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - hardswish(x, dst, k, item_ct1); - }); + [=](sycl::nd_item<3> item_ct1) { hardswish(x, dst, k, item_ct1); }); } template static void exp_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - exp(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { exp(x, dst, k, item_ct1); }); } template static void log_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - log(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { log(x, dst, k, item_ct1); }); } template static void neg_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - neg(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { neg(x, dst, k, item_ct1); }); } template static void step_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - step(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { step(x, dst, k, item_ct1); }); } template static void sigmoid_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIGMOID_BLOCK_SIZE - 1) / SYCL_SIGMOID_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE), + sycl_parallel_for( + stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sigmoid(x, dst, k, item_ct1); - }); + [=](sycl::nd_item<3> item_ct1) { sigmoid(x, dst, k, item_ct1); }); } template static void sqrt_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SQRT_BLOCK_SIZE - 1) / SYCL_SQRT_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sqrt(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { sqrt(x, dst, k, item_ct1); }); } template static void sin_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sin(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { sin(x, dst, k, item_ct1); }); } template static void cos_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cos(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { cos(x, dst, k, item_ct1); }); } template @@ -584,26 +536,20 @@ static void leaky_relu_sycl(const T *x, T *dst, const int k, const float negative_slope, queue_ptr stream) { const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - leaky_relu(x, dst, k, negative_slope, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { leaky_relu(x, dst, k, negative_slope, item_ct1); }); } template static void sqr_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - sqr(x, dst, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { sqr(x, dst, k, item_ct1); }); } template @@ -614,9 +560,8 @@ static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01, int dst_size = ne10 * ne11 * ne12 * ne13; int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE; sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE); - stream->parallel_for( - sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), - [=](sycl::nd_item<1> item_ct1) { + sycl_parallel_for<1>( + stream, sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) { upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1); }); } @@ -627,12 +572,10 @@ static void pad_sycl(const T *x, T *dst, const int ne00, const int ne1, const int ne2, queue_ptr stream) { int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE; sycl::range<3> gridDim(ne2, ne1, num_blocks); - stream->parallel_for( - sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); }); } template @@ -640,13 +583,10 @@ static void clamp_sycl(const T *x, T *dst, const float min, const float max, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - clamp(x, dst, min, max, k, item_ct1); - }); + sycl_parallel_for(stream, + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { clamp(x, dst, min, max, k, item_ct1); }); } inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-sycl/getrows.cpp b/ggml/src/ggml-sycl/getrows.cpp index 03f8dd907..9c76ffeb9 100644 --- a/ggml/src/ggml-sycl/getrows.cpp +++ b/ggml/src/ggml-sycl/getrows.cpp @@ -118,12 +118,10 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr GGML_ASSERT(ne00 % 2 == 0); - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - k_get_rows( - src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, - s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + k_get_rows(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, s3, nb01, nb02, nb03, s10, s11, s12, + item_ct1); + }); GGML_UNUSED(dst); GGML_UNUSED(ctx); @@ -156,9 +154,8 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); }); diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index 4b7610362..f25a96a62 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -1887,13 +1887,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols, const size_t shared_mem = ncols_pad * sizeof(int); if (order == GGML_SORT_ORDER_ASC) { - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor dpct_local_acc_ct1( sycl::range<1>(shared_mem), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_argsort_f32_i32( x, dst, ncols, ncols_pad, item_ct1, dpct_local_acc_ct1.get_multi_ptr() @@ -1901,13 +1900,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols, }); }); } else if (order == GGML_SORT_ORDER_DESC) { - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor dpct_local_acc_ct1( sycl::range<1>(shared_mem), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_argsort_f32_i32( x, dst, ncols, ncols_pad, item_ct1, dpct_local_acc_ct1.get_multi_ptr() @@ -1925,50 +1923,47 @@ static void argmax_f32_i32_sycl(const float *x, int *dst, const int ncols, const sycl::range<3> block_nums(1, nrows, 1); const size_t shared_mem = 256 * sizeof(float); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_data( sycl::range<1>(shared_mem/sizeof(float)), cgh); sycl::local_accessor shared_indices( sycl::range<1>(shared_mem/sizeof(float)), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - const int tid = item_ct1.get_local_id(2); - const int row = item_ct1.get_global_id(1); + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + const int tid = item_ct1.get_local_id(2); + const int row = item_ct1.get_global_id(1); - float max_val = -INFINITY; - int max_idx = -1; + float max_val = -INFINITY; + int max_idx = -1; - for (int col = tid; col < ncols; col += 256) { - float val = x[row * ncols + col]; - if (val > max_val) { - max_val = val; - max_idx = col; + for (int col = tid; col < ncols; col += 256) { + float val = x[row * ncols + col]; + if (val > max_val) { + max_val = val; + max_idx = col; + } + } + + shared_data[tid] = max_val; + shared_indices[tid] = max_idx; + item_ct1.barrier(sycl::access::fence_space::local_space); + + for (int stride = 256 / 2; stride > 0; stride >>= 1) { + if (tid < stride) { + float val1 = shared_data[tid]; + float val2 = shared_data[tid + stride]; + if (val2 > val1) { + shared_data[tid] = val2; + shared_indices[tid] = shared_indices[tid + stride]; } } - - shared_data[tid] = max_val; - shared_indices[tid] = max_idx; item_ct1.barrier(sycl::access::fence_space::local_space); + } - for (int stride = 256/2; stride > 0; stride >>= 1) { - if (tid < stride) { - float val1 = shared_data[tid]; - float val2 = shared_data[tid + stride]; - if (val2 > val1) { - shared_data[tid] = val2; - shared_indices[tid] = shared_indices[tid + stride]; - } - } - item_ct1.barrier(sycl::access::fence_space::local_space); - } - - - if (tid == 0) { - dst[row] = shared_indices[0]; - } - }); + if (tid == 0) { + dst[row] = shared_indices[0]; + } + }); }); } static void diag_mask_inf_f32_sycl(const float *x, float *dst, @@ -2952,7 +2947,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons void ** ptrs_dst_get = ptrs_dst.get(); size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half); size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half); - cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_compute_batched_ptrs(src0_f16, src1_f16, dst_ddf, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02, nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1); }); @@ -3456,7 +3451,7 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, { sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u)); sycl::range<3> grid_dims(1, n_ids, ids->ne[1]); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor src1_row_acc(cgh); char *__restrict src1_contiguous_get = @@ -3468,9 +3463,8 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, size_t ids_nb_ct6 = ids->nb[1]; size_t ids_nb_ct7 = ids->nb[0]; - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_copy_src1_to_contiguous( src1_original, src1_contiguous_get, dev_cur_src1_row_get, @@ -3501,15 +3495,14 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, { sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u)); sycl::range<3> grid_dims(1, 1, num_src1_rows); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { const char *__restrict dst_contiguous_get = dst_contiguous.get(); const mmid_row_mapping *__restrict dev_row_mapping_get = dev_row_mapping.get(); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { k_copy_dst_from_contiguous(dst_original, dst_contiguous_get, dev_row_mapping_get, diff --git a/ggml/src/ggml-sycl/gla.cpp b/ggml/src/ggml-sycl/gla.cpp index 879184fdd..b40cbf1f1 100644 --- a/ggml/src/ggml-sycl/gla.cpp +++ b/ggml/src/ggml-sycl/gla.cpp @@ -11,13 +11,13 @@ static void gated_linear_attn_f32_kernel(const dpct::queue_ptr stream, u_int B, const u_int n_seq_tokens = T / B; sycl::range<1> block_dims((C / H)); sycl::range<1> grid_dims((B * H)); - stream->submit([&](sycl::handler & cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { /* local memory accessors*/ auto _k = sycl::local_accessor(sycl::range<1>(head_size), cgh); auto _r = sycl::local_accessor(sycl::range<1>(head_size), cgh); auto _td = sycl::local_accessor(sycl::range<1>(head_size), cgh); - cgh.parallel_for(sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) { + sycl_parallel_for<1>(cgh, sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) { u_int tid = item.get_local_id(0); u_int bid = item.get_group(0); diff --git a/ggml/src/ggml-sycl/im2col.cpp b/ggml/src/ggml-sycl/im2col.cpp index aa19c2527..52737cc74 100644 --- a/ggml/src/ggml-sycl/im2col.cpp +++ b/ggml/src/ggml-sycl/im2col.cpp @@ -70,7 +70,7 @@ static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t I const int64_t CHW = IC * KH * KW; - stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) { im2col_kernel(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1, p0, p1, d0, d1, item_ct1); }); diff --git a/ggml/src/ggml-sycl/mmq.cpp b/ggml/src/ggml-sycl/mmq.cpp index ffb272aa2..c72fcd38e 100644 --- a/ggml/src/ggml-sycl/mmq.cpp +++ b/ggml/src/ggml-sycl/mmq.cpp @@ -1818,7 +1818,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q4_0_acc_ct1( @@ -1829,9 +1829,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -1853,7 +1852,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q4_0_acc_ct1( @@ -1864,9 +1863,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -1933,7 +1931,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_1_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_1_acc_ct1( @@ -1944,9 +1942,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -1968,7 +1965,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q4_1_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_1_acc_ct1( @@ -1979,9 +1976,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2048,7 +2044,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_0_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q5_0_acc_ct1( @@ -2059,9 +2055,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2083,7 +2078,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_0_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q5_0_acc_ct1( @@ -2094,9 +2089,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2163,7 +2157,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_1_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_1_acc_ct1( @@ -2174,9 +2168,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2198,7 +2191,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_1_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_1_acc_ct1( @@ -2209,9 +2202,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_1( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2278,7 +2270,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q8_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q8_0_acc_ct1( @@ -2289,9 +2281,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q8_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2313,7 +2304,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_qs_q8_0_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_d_q8_0_acc_ct1( @@ -2324,9 +2315,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q8_0( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2393,7 +2383,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q2_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q2_K_acc_ct1( @@ -2406,9 +2396,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q2_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2431,7 +2420,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q2_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q2_K_acc_ct1( @@ -2444,9 +2433,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q2_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2516,7 +2504,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q3_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q3_K_acc_ct1( @@ -2531,9 +2519,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q3_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2557,7 +2544,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q3_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q3_K_acc_ct1( @@ -2572,9 +2559,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q3_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2644,7 +2630,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q4_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_K_acc_ct1( @@ -2657,9 +2643,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2682,7 +2667,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q4_K_acc_ct1( sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q4_K_acc_ct1( @@ -2695,9 +2680,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q4_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2765,7 +2749,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_K_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_K_acc_ct1( @@ -2778,9 +2762,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2803,7 +2786,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_q5_K_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_q5_K_acc_ct1( @@ -2816,9 +2799,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q5_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2886,7 +2868,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_acc_ct1( @@ -2899,9 +2881,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q6_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, @@ -2924,7 +2905,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, dpct::has_capability_or_fail(stream->get_device(), {sycl::aspect::fp16}); - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor tile_x_ql_acc_ct1( sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); sycl::local_accessor tile_x_dm_acc_ct1( @@ -2937,9 +2918,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, sycl::local_accessor tile_y_ds_acc_ct1( sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { mul_mat_q6_K( vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, item_ct1, diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 5b7f06407..c21929d51 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -544,12 +544,12 @@ static void reorder_mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, (block_num_y * WARP_SIZE)); const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE); - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size), - [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, - nd_item); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size), + [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, + nd_item); + }); }); } @@ -561,12 +561,12 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float * const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -580,17 +580,12 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -604,17 +599,12 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -628,17 +618,12 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -652,17 +637,12 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -676,17 +656,12 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -700,17 +675,12 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -724,17 +694,12 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -750,12 +715,12 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy, const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE); const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE); - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size), - [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_reorder>(vx, vy, dst, ncols, - nrows, nd_item); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size), + [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, + nd_item); + }); }); } @@ -769,17 +734,12 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -794,12 +754,12 @@ static void reorder_mul_mat_vec_q6_k_q8_1_sycl(const void * vx, const void * vy, const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE); const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE); - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size), - [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, - nd_item); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size), + [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, + nd_item); + }); }); } static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, @@ -811,17 +771,12 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -836,14 +791,12 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq2_xxs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq2_xxs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } @@ -857,14 +810,12 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler & cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq2_xs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq2_xs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } @@ -878,15 +829,12 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq2_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq2_s_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -900,15 +848,12 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq3_xxs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq3_xxs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } @@ -922,15 +867,12 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq3_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq3_s_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -944,15 +886,12 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq1_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq1_s_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -966,14 +905,12 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq1_m_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq1_m_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -987,15 +924,12 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq4_nl_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq4_nl_q8_1(vx, vy, dst, ncols, nrows, + item_ct1); + }); }); } } @@ -1009,15 +943,12 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { - - stream->submit([&](sycl::handler &cgh) { - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q_iq4_xs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_iq4_xs_q8_1(vx, vy, dst, ncols, + nrows, item_ct1); + }); }); } } diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp index 4ec141684..79d846b41 100644 --- a/ggml/src/ggml-sycl/norm.cpp +++ b/ggml/src/ggml-sycl/norm.cpp @@ -254,14 +254,13 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i GGML_ASSERT(ncols % WARP_SIZE == 0); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE); - }); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -272,16 +271,15 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1( sycl::range<1>(work_group_size / WARP_SIZE), cgh); - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + get_pointer(s_sum_acc_ct1), work_group_size); + }); + }); } } @@ -290,18 +288,14 @@ static void group_norm_f32_sycl(const float* x, float* dst, const int ne_elements, queue_ptr stream, int device) { if (group_size < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { const float eps_ct4 = eps; - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - group_norm_f32( - x, dst, group_size, ne_elements, eps_ct4, item_ct1, - nullptr, WARP_SIZE); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, nullptr, + WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -313,22 +307,18 @@ static void group_norm_f32_sycl(const float* x, float* dst, info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), cgh); const float eps_ct4 = eps; - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - group_norm_f32(x, dst, group_size, ne_elements, - eps_ct4, item_ct1, - get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, + get_pointer(s_sum_acc_ct1), work_group_size); + }); + }); } } @@ -340,14 +330,13 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const const sycl::range<3> global_dims(nsamples, nchannels, nrows); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE); - }); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -358,16 +347,15 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), cgh); - cgh.parallel_for( - sycl::nd_range<3>(global_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, + get_pointer(s_sum_acc_ct1), work_group_size); + }); + }); } } @@ -378,16 +366,12 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols, // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); if (ncols < 1024) { const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler& cgh) { - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - l2_norm_f32(x, dst, ncols, eps, item_ct1, - nullptr, WARP_SIZE); - }); - }); + sycl_launch(stream, [&](sycl::handler & cgh) { + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + l2_norm_f32(x, dst, ncols, eps, item_ct1, nullptr, WARP_SIZE); + }); + }); } else { const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; @@ -398,18 +382,15 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols, the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), cgh); - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - l2_norm_f32(x, dst, ncols, eps, item_ct1, - get_pointer(s_sum_acc_ct1), work_group_size); - }); - }); + sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + l2_norm_f32(x, dst, ncols, eps, item_ct1, get_pointer(s_sum_acc_ct1), + work_group_size); + }); + }); } } diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp index 44473e1e5..e44c6b6ef 100644 --- a/ggml/src/ggml-sycl/rope.cpp +++ b/ggml/src/ggml-sycl/rope.cpp @@ -235,20 +235,22 @@ static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, c the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } else { /* DPCT1049:41: The work-group size passed to the SYCL kernel may exceed the limit. To get the device limit, query info::device::max_work_group_size. Adjust the work-group size if needed. */ - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_norm(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } } @@ -267,15 +269,17 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); if (freq_factors == nullptr) { - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } else { - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { - rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, - theta_scale, freq_factors, item_ct1); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, + attn_factor, corr_dims, theta_scale, freq_factors, item_ct1); + }); } } @@ -298,12 +302,12 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1, } // launch kernel if (freq_factors == nullptr) { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_multi(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); } else { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_multi(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); @@ -333,12 +337,12 @@ static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1, } // launch kernel if (freq_factors == nullptr) { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_vision(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); } else { - stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) { rope_vision(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale, freq_factors, sections, item_ct1); }); diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp index 52fcf4b3d..7b60c292e 100644 --- a/ggml/src/ggml-sycl/softmax.cpp +++ b/ggml/src/ggml-sycl/softmax.cpp @@ -127,11 +127,11 @@ static void soft_max_f32_submitter(const float * x, const T * mask, float * dst, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims, const size_t n_local_scratch, queue_ptr stream) { - stream->submit([&](sycl::handler &cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor local_buf_acc(n_local_scratch, cgh); - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), + sycl_parallel_for( + cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { soft_max_f32(x, mask, dst, ncols_par, nrows_y, scale, max_bias, m0, diff --git a/ggml/src/ggml-sycl/tsembd.cpp b/ggml/src/ggml-sycl/tsembd.cpp index f6ca626ea..721c8fa6f 100644 --- a/ggml/src/ggml-sycl/tsembd.cpp +++ b/ggml/src/ggml-sycl/tsembd.cpp @@ -45,14 +45,9 @@ static void timestep_embedding_f32_sycl( int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE; sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE); sycl::range<3> gridDim(1, ne00, num_blocks); - stream->parallel_for( - sycl::nd_range<3>( - gridDim * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - timestep_embedding_f32( - x, dst, nb1, dim, max_period, item_ct1 - ); - }); + sycl_parallel_for(stream, sycl::nd_range<3>(gridDim * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + timestep_embedding_f32(x, dst, nb1, dim, max_period, item_ct1); + }); } void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-sycl/wkv.cpp b/ggml/src/ggml-sycl/wkv.cpp index c10e2f764..3ed5bbf35 100644 --- a/ggml/src/ggml-sycl/wkv.cpp +++ b/ggml/src/ggml-sycl/wkv.cpp @@ -207,12 +207,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { // Submit kernel if (C / H == WKV_BLOCK_SIZE) { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv6_f32_kernel( B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() @@ -220,12 +219,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { }); }); } else { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv6_f32_kernel( B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() @@ -264,12 +262,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { // Submit kernel if (C / H == WKV_BLOCK_SIZE) { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv7_f32_kernel( B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() @@ -277,12 +274,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { }); }); } else { - stream->submit([&](sycl::handler& cgh) { + sycl_launch(stream, [&](sycl::handler & cgh) { sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { + sycl_parallel_for( + cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { rwkv_wkv7_f32_kernel( B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d, item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() From dd6e6d0b6a4bbe3ebfc931d1eb14db2f2b1d70af Mon Sep 17 00:00:00 2001 From: Ruikai Peng Date: Fri, 20 Jun 2025 22:13:06 +0800 Subject: [PATCH 6/8] vocab : prevent tokenizer overflow (#14301) * vocab : prevent stack overflow in tokenize * vocab : return error instead of aborting on oversized token count * vocab : INT32_MIN from llama_tokenize on overflow --- common/common.cpp | 3 +++ include/llama.h | 1 + src/llama-vocab.cpp | 5 +++++ 3 files changed, 9 insertions(+) diff --git a/common/common.cpp b/common/common.cpp index c2c94e7ae..e4e71ad13 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1290,6 +1290,9 @@ std::vector common_tokenize( int n_tokens = text.length() + 2 * add_special; std::vector result(n_tokens); n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special); + if (n_tokens == std::numeric_limits::min()) { + throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit"); + } if (n_tokens < 0) { result.resize(-n_tokens); int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special); diff --git a/include/llama.h b/include/llama.h index 3475d5965..b04720bee 100644 --- a/include/llama.h +++ b/include/llama.h @@ -1088,6 +1088,7 @@ extern "C" { /// @param tokens The tokens pointer must be large enough to hold the resulting tokens. /// @return Returns the number of tokens on success, no more than n_tokens_max /// @return Returns a negative number on failure - the number of tokens that would have been returned + /// @return Returns INT32_MIN on overflow (e.g., tokenization result size exceeds int32_t limit) /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so. /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated /// as plaintext. Does not insert a leading space. diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 4ab120d9b..4aaf4c825 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -3074,6 +3074,11 @@ int32_t llama_vocab::tokenize( bool add_special, bool parse_special) const { auto res = tokenize(std::string(text, text_len), add_special, parse_special); + if (res.size() >= static_cast(std::numeric_limits::max())) { + LLAMA_LOG_ERROR("%s: tokenization result size %zu exceeds int32_t limit\n", __func__, res.size()); + return std::numeric_limits::min(); + } + if (n_tokens_max < (int) res.size()) { // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); return -((int) res.size()); From 22015b2092e291022ea3cfedc6aeb8f2643807da Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Fri, 20 Jun 2025 16:37:44 +0200 Subject: [PATCH 7/8] lint : remove trailing whitepace (#14304) --- src/llama-vocab.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 4aaf4c825..5c9eb8756 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -3078,7 +3078,7 @@ int32_t llama_vocab::tokenize( LLAMA_LOG_ERROR("%s: tokenization result size %zu exceeds int32_t limit\n", __func__, res.size()); return std::numeric_limits::min(); } - + if (n_tokens_max < (int) res.size()) { // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__); return -((int) res.size()); From c959f462a0b4d42eaf930ffd72df0e435c97d5d5 Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Fri, 20 Jun 2025 22:48:24 +0800 Subject: [PATCH 8/8] CUDA: add conv_2d_transpose (#14287) * CUDA: add conv_2d_transpose * remove direct include of cuda_fp16 * Review: add brackets for readability, remove ggml_set_param and add asserts --- ggml/src/ggml-cuda/conv2d-transpose.cu | 91 +++++++++++++++++++++++++ ggml/src/ggml-cuda/conv2d-transpose.cuh | 4 ++ ggml/src/ggml-cuda/ggml-cuda.cu | 5 ++ tests/test-backend-ops.cpp | 34 +++++++++ 4 files changed, 134 insertions(+) create mode 100644 ggml/src/ggml-cuda/conv2d-transpose.cu create mode 100644 ggml/src/ggml-cuda/conv2d-transpose.cuh diff --git a/ggml/src/ggml-cuda/conv2d-transpose.cu b/ggml/src/ggml-cuda/conv2d-transpose.cu new file mode 100644 index 000000000..03224e404 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-transpose.cu @@ -0,0 +1,91 @@ +#include + +#include "conv2d-transpose.cuh" +#include "ggml.h" + +__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel, + float * __restrict__ output, const int in_w, const int in_h, const int out_w, + const int out_h, const int kernel_w, const int kernel_h, const int stride, + const int c_in, const int c_out, const int batches) { + const int global_idx = blockIdx.x * blockDim.x + threadIdx.x; + + const int total_elements = out_w * out_h * c_out * batches; + + if (global_idx >= total_elements) { + return; + } + + const int out_x_idx = global_idx % out_w; + const int out_y_idx = (global_idx / out_w) % out_h; + const int c_idx = (global_idx / (out_w * out_h)) % c_out; + const int n_idx = global_idx / (out_w * out_h * c_out); + + float accumulator = 0; + // For each output idx, find the inputs that contribute to it by checking stride alignment and bounds + + for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) { + for (int kh = 0; kh < kernel_h; ++kh) { + int in_y = out_y_idx - kh; + if (in_y < 0 || in_y % stride) continue; + in_y /= stride; + if (in_y >= in_h) continue; + + for (int kw = 0; kw < kernel_w; ++kw) { + int in_x = out_x_idx - kw; + if (in_x < 0 || in_x % stride) continue; + in_x /= stride; + if (in_x >= in_w) continue; + + const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x; + const int kernel_idx = + (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw; + + float input_val = input[input_idx]; + half kern_val = kernel[kernel_idx]; + + accumulator += input_val * (float) kern_val; + } + } + } + + output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + (out_w) *out_y_idx + out_x_idx] = accumulator; +} + +//input is (W, H, C_in, N), Kernel is (W, H, C_out, C_in) +void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * kernel = dst->src[0]; + const ggml_tensor * input = dst->src[1]; + + GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); + + const float * input_data = (const float *) input->data; + float * output_data = (float *) dst->data; + const half * kernel_data = (const half *) kernel->data; + + const int input_w = input->ne[0]; + const int input_h = input->ne[1]; + const int output_w = dst->ne[0]; + const int output_h = dst->ne[1]; + const int channels_in = input->ne[2]; + const int channels_out = kernel->ne[2]; + const int kernel_w = kernel->ne[0]; + const int kernel_h = kernel->ne[1]; + const int stride = dst->op_params[0]; + const int batches = input->ne[3]; + + GGML_ASSERT(channels_in == kernel->ne[3]); + GGML_ASSERT(stride > 0); + + cudaStream_t st = ctx.stream(); + + GGML_ASSERT(ggml_is_contiguous(input)); + GGML_ASSERT(ggml_is_contiguous(kernel)); + GGML_ASSERT(ggml_is_contiguous(dst)); + + const int total = (output_w * output_h * channels_out * batches); + const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE; + + conv2d_transpose_kernel<<>>( + input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride, + channels_in, channels_out, batches); +} diff --git a/ggml/src/ggml-cuda/conv2d-transpose.cuh b/ggml/src/ggml-cuda/conv2d-transpose.cuh new file mode 100644 index 000000000..c9430b248 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-transpose.cuh @@ -0,0 +1,4 @@ +#include "common.cuh" + +#define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256 +void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 530f541f9..5bab92e34 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -12,6 +12,7 @@ #include "ggml-cuda/concat.cuh" #include "ggml-cuda/conv-transpose-1d.cuh" #include "ggml-cuda/conv2d-dw.cuh" +#include "ggml-cuda/conv2d-transpose.cuh" #include "ggml-cuda/convert.cuh" #include "ggml-cuda/count-equal.cuh" #include "ggml-cuda/cpy.cuh" @@ -2341,6 +2342,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_CONV_2D_DW: ggml_cuda_op_conv2d_dw(ctx, dst); break; + case GGML_OP_CONV_TRANSPOSE_2D: + ggml_cuda_conv_2d_transpose_p0(ctx, dst); + break; case GGML_OP_CONV_TRANSPOSE_1D: ggml_cuda_op_conv_transpose_1d(ctx,dst); break; @@ -3252,6 +3256,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g } case GGML_OP_IM2COL: case GGML_OP_CONV_2D_DW: + case GGML_OP_CONV_TRANSPOSE_2D: case GGML_OP_POOL_2D: case GGML_OP_SUM: case GGML_OP_SUM_ROWS: diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 509a4b35f..772bee346 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -2725,6 +2725,35 @@ struct test_conv_transpose_1d : public test_case { } }; +// GGML_OP_CONV_TRANSPOSE_2D +struct test_conv_transpose_2d : public test_case { + const std::array ne_input; + const std::array ne_kernel; + const int stride; + + std::string vars() override { + return VARS_TO_STR3(ne_input, ne_kernel, stride); + } + + test_conv_transpose_2d(std::array ne_input = {10, 10, 3, 1}, // [input_width, input_height, input_channels, 1] + std::array ne_kernel = {3, 3, 3, 1}, // [kernel_width, kernel_height, input_channels, 1] + int stride = 1) + : ne_input(ne_input), ne_kernel(ne_kernel), stride(stride){} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data()); + ggml_set_name(input, "input"); + + ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne_kernel.data()); + ggml_set_name(kernel, "kernel"); + + ggml_tensor * out = ggml_conv_transpose_2d_p0(ctx, kernel, input, stride); + ggml_set_name(out, "out"); + + return out; + } +}; + // GGML_OP_IM2COL struct test_im2col : public test_case { const ggml_type type_input; @@ -4050,6 +4079,9 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1)); test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1)); + test_cases.emplace_back(new test_conv_transpose_2d({3, 2, 3, 1}, {2, 2, 1, 3}, 1)); + test_cases.emplace_back(new test_conv_transpose_2d({10, 10, 9, 1}, {3, 3, 1, 9}, 2)); + test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 500, 1, 1})); test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 5000, 1, 1})); @@ -4618,6 +4650,8 @@ static std::vector> make_test_cases_perf() { test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, false)); test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, true)); + test_cases.emplace_back(new test_conv_transpose_2d({256, 256, 256, 1}, {3, 3, 16, 256}, 1)); + return test_cases; }