From 37bef8943312d91183ff06d8f1214082a17344a5 Mon Sep 17 00:00:00 2001 From: jaime-m-p <167997752+jaime-m-p@users.noreply.github.com> Date: Tue, 18 Jun 2024 18:40:52 +0200 Subject: [PATCH 1/4] tokenizer : BPE fixes (#7530) * Random test: add_bos_token, add_eos_token * Random test: add BPE models for testing * Custom regex split fails with codepoint 0 * Fix falcon punctuation regex * Refactor llm_tokenizer_bpe: move code to constructor * Move 'add_special_bos/eos' logic to llm_tokenizer_bpe * Move tokenizer flags to vocab structure. * Default values for special_add_bos/eos * Build vocab.special_tokens_cache using vocab token types * Generalize 'jina-v2' per token attributes * Fix unicode whitespaces (deepseek-coder, deepseek-llm) * Skip missing byte tokens (falcon) * Better unicode data generation * Replace char32_t with uint32_t --- llama.cpp | 309 +++--- scripts/gen-unicode-data.py | 180 ++-- tests/test-tokenizer-random.py | 170 +++- unicode-data.cpp | 1652 ++++++++++++++++---------------- unicode.cpp | 29 +- 5 files changed, 1285 insertions(+), 1055 deletions(-) diff --git a/llama.cpp b/llama.cpp index e06c851ad5b6c..8818c69280178 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2310,16 +2310,17 @@ struct llama_vocab { id special_cls_id = -1; id special_mask_id = -1; - int special_add_bos = -1; // -1 unknown, 1 add, 0 don't add. - int special_add_eos = -1; // -1 unknown, 1 add, 0 don't add. - id linefeed_id = 13; id special_prefix_id = -1; id special_suffix_id = -1; id special_middle_id = -1; id special_eot_id = -1; // TODO: move above after "eos_id", and here add "file separator" token - bool add_space_prefix = true; + // tokenizer flags + bool tokenizer_add_space_prefix = true; + bool tokenizer_add_bos = false; + bool tokenizer_add_eos = false; + bool tokenizer_ignore_merges = false; int find_bpe_rank(const std::string & token_left, const std::string & token_right) const { GGML_ASSERT(token_left.find(' ') == std::string::npos); @@ -4770,7 +4771,7 @@ static void llm_load_vocab( const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str()); if (add_space_prefix_keyidx != -1) { - vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); + vocab.tokenizer_add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); } // The default value of add_space_prefix is true. } else if (tokenizer_model == "bert") { vocab.type = LLAMA_VOCAB_TYPE_WPM; @@ -4783,13 +4784,13 @@ static void llm_load_vocab( vocab.special_pad_id = 0; vocab.special_cls_id = 101; vocab.special_mask_id = 103; - vocab.add_space_prefix = false; + vocab.tokenizer_add_space_prefix = false; } else if (tokenizer_model == "gpt2") { vocab.type = LLAMA_VOCAB_TYPE_BPE; const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str()); if (add_space_prefix_keyidx != -1) { - vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); + vocab.tokenizer_add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); } // read bpe merges and populate bpe ranks @@ -4847,6 +4848,8 @@ static void llm_load_vocab( tokenizer_pre == "llama-v3" || tokenizer_pre == "llama-bpe") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3; + vocab.tokenizer_ignore_merges = true; + vocab.tokenizer_add_bos = true; } else if ( tokenizer_pre == "deepseek-llm") { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM; @@ -4897,6 +4900,14 @@ static void llm_load_vocab( } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } + } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + vocab.tokenizer_add_bos = true; + vocab.tokenizer_add_eos = false; + } else if (vocab.type == LLAMA_VOCAB_TYPE_WPM) { + vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; + vocab.tokenizer_add_bos = true; + vocab.tokenizer_add_eos = false; } else { vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT; } @@ -5041,10 +5052,10 @@ static void llm_load_vocab( bool temp = true; if (ml.get_key(LLM_KV_TOKENIZER_ADD_BOS, temp, false)) { - vocab.special_add_bos = int(temp); + vocab.tokenizer_add_bos = temp; } if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) { - vocab.special_add_eos = int(temp); + vocab.tokenizer_add_eos = temp; } } @@ -5144,7 +5155,7 @@ static void llm_load_vocab( ); // set attributes by model/tokenizer name - if (_contains_any(tokenizer_pre, {"jina-v2-es", "jina-v2-de"})) { + if (_contains_any(tokenizer_pre, {"jina-v2-de", "jina-v2-es", "jina-v2-code"})) { _set_token_attr("", LLAMA_TOKEN_ATTR_LSTRIP, true); } else if (_contains_any(model_name, {"phi-3", "phi3"})) { for (auto id : vocab.cache_special_tokens) { @@ -13158,112 +13169,142 @@ struct llm_bigram_bpe { }; struct llm_tokenizer_bpe { - llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) {} - - void tokenize(const std::string & text, std::vector & output) { - int final_prev_index = -1; - bool ignore_merges = false; - - std::vector word_collection; - switch (vocab.type) { - case LLAMA_VOCAB_TYPE_BPE: - switch (vocab.type_pre) { - case LLAMA_VOCAB_PRE_TYPE_LLAMA3: - ignore_merges = true; - word_collection = unicode_regex_split(text, { - // original regex from tokenizer.json - //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - - // adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989 - "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_DBRX: - case LLAMA_VOCAB_PRE_TYPE_SMAUG: - word_collection = unicode_regex_split(text, { - // same as llama3 - "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM: - word_collection = unicode_regex_split(text, { - "[\r\n]", - "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+", - "\\s?[!-/:-~！-／：-～‘-‟　-。]+", - "\\s+$", - "[一-龥ࠀ-一가-퟿]+", - "\\p{N}+", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER: - word_collection = unicode_regex_split(text, { - "[\r\n]", - "\\s?\\p{L}+", - "\\s?\\p{P}+", - "[一-龥ࠀ-一가-퟿]+", - "\\p{N}", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_FALCON: - word_collection = unicode_regex_split(text, { - "[\\p{P}\\$\\+<=>\\^~\\|]+", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - "[0-9][0-9][0-9]", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_MPT: - // TODO: MPT pre-tokenization regexes are unknown - // the following are close, but not exact. run the following: - // ./bin/test-tokenizer-0 ../models/ggml-vocab-mpt.gguf - GGML_ASSERT("MPT pre-tokenization regexes are unknown - fixes needed"); - word_collection = unicode_regex_split(text, { - "\\s?\\p{L}+", - "\\s?\\p{P}+", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_STARCODER: - case LLAMA_VOCAB_PRE_TYPE_REFACT: - case LLAMA_VOCAB_PRE_TYPE_COMMAND_R: - word_collection = unicode_regex_split(text, { - "\\p{N}", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_GPT2: - case LLAMA_VOCAB_PRE_TYPE_OLMO: - word_collection = unicode_regex_split(text, { - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_STABLELM2: - case LLAMA_VOCAB_PRE_TYPE_QWEN2: - word_collection = unicode_regex_split(text, { - // original regex from tokenizer.json - // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" - "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", - }); - break; - case LLAMA_VOCAB_PRE_TYPE_PORO: - word_collection = unicode_regex_split(text, { - " ?[^(\\s|.,!?…。，、।۔،)]+", - }); - break; - default: - // default regex for BPE tokenization pre-processing - word_collection = unicode_regex_split(text, { - "[\\p{P}\\$\\+<=>\\^~\\|]+", - "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", - "\\p{N}+", - "[0-9][0-9][0-9]", - }); - break; - } + llm_tokenizer_bpe(const llama_vocab & vocab): vocab(vocab) { + GGML_ASSERT(vocab.type == LLAMA_VOCAB_TYPE_BPE); + switch (vocab.type_pre) { + case LLAMA_VOCAB_PRE_TYPE_LLAMA3: + regex_exprs = { + // original regex from tokenizer.json + //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + + // adapted: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2080233989 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DBRX: + case LLAMA_VOCAB_PRE_TYPE_SMAUG: + regex_exprs = { + // same as llama3 + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM: + regex_exprs = { + "[\r\n]", + "\\s?[A-Za-zµÀ-ÖØ-öø-ƺƼ-ƿDŽ-ʓʕ-ʯͰ-ͳͶͷͻ-ͽͿΆΈ-ΊΌΎ-ΡΣ-ϵϷ-ҁҊ-ԯԱ-ՖႠ-ჅᎠ-Ᏽᏸ-ᏽᲐ-ᲺᲽ-Ჿᴀ-ᴫᵫ-ᵷᵹ-ᶚḀ-ἕἘ-Ἕἠ-ὅὈ-Ὅὐ-ὗὙὛὝὟ-ώᾀ-ᾴᾶ-ᾼιῂ-ῄῆ-ῌῐ-ΐῖ-Ίῠ-Ῥῲ-ῴῶ-ῼℂℇℊ-ℓℕℙ-ℝℤΩℨK-ℭℯ-ℴℹℼ-ℿⅅ-ⅉⅎↃↄⰀ-ⱻⱾ-ⳤⳫ-ⳮⳲⳳꙀ-ꙭꚀ-ꚛꜢ-ꝯꝱ-ꞇꞋ-ꞎꭰ-ꮿff-stﬓ-ﬗＡ-Ｚａ-ｚ𐐀-𐑏𐒰-𐓓𐓘-𐓻𐲀-𐲲𐳀-𐳲𑢠-𑣟𞤀-𞥃]+", + "\\s?[!-/:-~！-／：-～‘-‟　-。]+", + "\\s+$", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER: + regex_exprs = { + "[\r\n]", + "\\s?\\p{L}+", + "\\s?\\p{P}+", + "[一-龥ࠀ-一가-퟿]+", + "\\p{N}", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_FALCON: + regex_exprs = { + "[\\p{P}\\$\\+<=>\\^~\\|`]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "[0-9][0-9][0-9]", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_MPT: + // TODO: MPT pre-tokenization regexes are unknown + // the following are close, but not exact. run the following: + // ./bin/test-tokenizer-0 ../models/ggml-vocab-mpt.gguf + GGML_ASSERT("MPT pre-tokenization regexes are unknown - fixes needed"); + regex_exprs = { + "\\s?\\p{L}+", + "\\s?\\p{P}+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_STARCODER: + case LLAMA_VOCAB_PRE_TYPE_REFACT: + case LLAMA_VOCAB_PRE_TYPE_COMMAND_R: + regex_exprs = { + "\\p{N}", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_GPT2: + case LLAMA_VOCAB_PRE_TYPE_OLMO: + regex_exprs = { + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_STABLELM2: + case LLAMA_VOCAB_PRE_TYPE_QWEN2: + regex_exprs = { + // original regex from tokenizer.json + // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_PORO: + regex_exprs = { + " ?[^(\\s|.,!?…。，、।۔،)]+", + }; break; default: - GGML_ASSERT(false); + // default regex for BPE tokenization pre-processing + regex_exprs = { + "[\\p{P}\\$\\+<=>\\^~\\|]+", + "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", + "\\p{N}+", + "[0-9][0-9][0-9]", + }; break; } + } + + void append(const llama_vocab::id token_id, std::vector & output) const { + output.push_back(token_id); + } + + bool append_bos(std::vector & output) const { + if (vocab.tokenizer_add_bos) { + GGML_ASSERT(vocab.special_bos_id != -1); + output.push_back(vocab.special_bos_id); + return true; + } + return false; + } + + bool append_eos(std::vector & output) const { + if (vocab.tokenizer_add_eos) { + GGML_ASSERT(vocab.special_eos_id != -1); + output.push_back(vocab.special_eos_id); + return true; + } + return false; + } + + void check_double_bos_eos(const std::vector & output) const { + if (vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) { + LLAMA_LOG_WARN( + "%s: Added a BOS token to the prompt as specified by the model but the prompt " + "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. " + "Are you sure this is what you want?\n", __FUNCTION__); + } + if (vocab.tokenizer_add_eos && output.size() >= 2 && *(output.end()-2) == vocab.special_eos_id) { + LLAMA_LOG_WARN( + "%s: Added a EOS token to the prompt as specified by the model but the prompt " + "also ends with a EOS token. So now the final prompt ends with 2 EOS tokens. " + "Are you sure this is what you want?\n", __FUNCTION__); + } + } + + void tokenize(const std::string & text, std::vector & output) { + int final_prev_index = -1; + + const auto word_collection = unicode_regex_split(text, regex_exprs); symbols_final.clear(); @@ -13274,7 +13315,7 @@ struct llm_tokenizer_bpe { int index = 0; size_t offset = 0; - if (ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) { + if (vocab.tokenizer_ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) { symbols.emplace_back(llm_symbol{-1, -1, word.c_str(), word.size()}); offset = word.size(); } @@ -13355,10 +13396,9 @@ struct llm_tokenizer_bpe { for (auto j = str.begin(); j != str.end(); ++j) { std::string byte_str(1, *j); auto token_multibyte = vocab.token_to_id.find(byte_str); - if (token_multibyte == vocab.token_to_id.end()) { - throw std::runtime_error("ERROR: byte not found in vocab"); + if (token_multibyte != vocab.token_to_id.end()) { + output.push_back(token_multibyte->second); } - output.push_back((*token_multibyte).second); } } else { output.push_back((*token).second); @@ -13397,6 +13437,8 @@ struct llm_tokenizer_bpe { const llama_vocab & vocab; + std::vector regex_exprs; + std::vector symbols; std::vector symbols_final; @@ -13677,7 +13719,7 @@ static std::vector llama_tokenize_internal(const llama_vocab & bool is_prev_special = false; - if (add_special && vocab.special_add_bos != 0) { + if (add_special && vocab.tokenizer_add_bos) { GGML_ASSERT(vocab.special_bos_id != -1); output.push_back(vocab.special_bos_id); is_prev_special = true; @@ -13687,7 +13729,7 @@ static std::vector llama_tokenize_internal(const llama_vocab & if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) { auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); - if (vocab.add_space_prefix) { + if (vocab.tokenizer_add_space_prefix) { if (!output.size() || is_prev_special) { // prefix with space if first token raw_text = " " + raw_text; } @@ -13705,23 +13747,24 @@ static std::vector llama_tokenize_internal(const llama_vocab & } } - if (add_special && vocab.special_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) { + if (add_special && vocab.tokenizer_add_bos && output.size() >= 2 && output[1] == vocab.special_bos_id) { LLAMA_LOG_WARN( "%s: Added a BOS token to the prompt as specified by the model but the prompt " "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. " "Are you sure this is what you want?\n", __FUNCTION__); } - if (add_special && vocab.special_add_eos == 1) { + if (add_special && vocab.tokenizer_add_eos) { GGML_ASSERT(vocab.special_eos_id != -1); output.push_back(vocab.special_eos_id); } } break; case LLAMA_VOCAB_TYPE_BPE: { - if (add_special && vocab.special_add_bos != 0) { - GGML_ASSERT(vocab.special_bos_id != -1); - output.push_back(vocab.special_bos_id); + llm_tokenizer_bpe tokenizer(vocab); + + if (add_special) { + tokenizer.append_bos(output); } for (const auto & fragment : fragment_buffer) { @@ -13731,23 +13774,15 @@ static std::vector llama_tokenize_internal(const llama_vocab & #ifdef PRETOKENIZERDEBUG LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); #endif - llm_tokenizer_bpe tokenizer(vocab); tokenizer.tokenize(raw_text, output); } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN) - output.push_back(fragment.token); + tokenizer.append(fragment.token, output); } } - if (add_special && vocab.special_add_bos != 0 && output.size() >= 2 && output[1] == vocab.special_bos_id) { - LLAMA_LOG_WARN( - "%s: Added a BOS token to the prompt as specified by the model but the prompt " - "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. " - "Are you sure this is what you want?\n", __FUNCTION__); - } - - if (add_special && vocab.special_add_eos == 1) { - GGML_ASSERT(vocab.special_add_eos != -1); - output.push_back(vocab.special_eos_id); + if (add_special) { + tokenizer.append_eos(output); + tokenizer.check_double_bos_eos(output); } } break; case LLAMA_VOCAB_TYPE_WPM: @@ -18320,11 +18355,11 @@ llama_token llama_token_nl(const struct llama_model * model) { } int32_t llama_add_bos_token(const struct llama_model * model) { - return model->vocab.special_add_bos; + return model->vocab.tokenizer_add_bos; } int32_t llama_add_eos_token(const struct llama_model * model) { - return model->vocab.special_add_eos; + return model->vocab.tokenizer_add_eos; } llama_token llama_token_prefix(const struct llama_model * model) { diff --git a/scripts/gen-unicode-data.py b/scripts/gen-unicode-data.py index 744873c2a4986..890e4d7c24eeb 100644 --- a/scripts/gen-unicode-data.py +++ b/scripts/gen-unicode-data.py @@ -1,83 +1,143 @@ -import regex -import ctypes +import array import unicodedata - - -class CoodepointFlags (ctypes.Structure): - _fields_ = [ # see definition in unicode.h - ("is_undefined", ctypes.c_uint16, 1), - ("is_number", ctypes.c_uint16, 1), # regex: \p{N} - ("is_letter", ctypes.c_uint16, 1), # regex: \p{L} - ("is_separator", ctypes.c_uint16, 1), # regex: \p{Z} - ("is_accent_mark", ctypes.c_uint16, 1), # regex: \p{M} - ("is_punctuation", ctypes.c_uint16, 1), # regex: \p{P} - ("is_symbol", ctypes.c_uint16, 1), # regex: \p{S} - ("is_control", ctypes.c_uint16, 1), # regex: \p{C} - ] - - -assert (ctypes.sizeof(CoodepointFlags) == 2) +import requests MAX_CODEPOINTS = 0x110000 -regex_number = regex.compile(r'\p{N}') -regex_letter = regex.compile(r'\p{L}') -regex_separator = regex.compile(r'\p{Z}') -regex_accent_mark = regex.compile(r'\p{M}') -regex_punctuation = regex.compile(r'\p{P}') -regex_symbol = regex.compile(r'\p{S}') -regex_control = regex.compile(r'\p{C}') -regex_whitespace = regex.compile(r'\s') - -codepoint_flags = (CoodepointFlags * MAX_CODEPOINTS)() +UNICODE_DATA_URL = "https://www.unicode.org/Public/UCD/latest/ucd/UnicodeData.txt" + + +# see https://www.unicode.org/L2/L1999/UnicodeData.html +def unicode_data_iter(): + res = requests.get(UNICODE_DATA_URL) + res.raise_for_status() + data = res.content.decode() + + prev = [] + + for line in data.splitlines(): + # ej: 0000;;Cc;0;BN;;;;;N;NULL;;;; + line = line.split(";") + + cpt = int(line[0], base=16) + assert cpt < MAX_CODEPOINTS + + cpt_lower = int(line[-2] or "0", base=16) + assert cpt_lower < MAX_CODEPOINTS + + cpt_upper = int(line[-3] or "0", base=16) + assert cpt_upper < MAX_CODEPOINTS + + categ = line[2].strip() + assert len(categ) == 2 + + bidir = line[4].strip() + assert len(categ) == 2 + + name = line[1] + if name.endswith(", First>"): + prev = (cpt, cpt_lower, cpt_upper, categ, bidir) + continue + if name.endswith(", Last>"): + assert prev[1:] == (0, 0, categ, bidir) + for c in range(prev[0], cpt): + yield (c, cpt_lower, cpt_upper, categ, bidir) + + yield (cpt, cpt_lower, cpt_upper, categ, bidir) + + +# see definition in unicode.h +CODEPOINT_FLAG_UNDEFINED = 0x0001 # +CODEPOINT_FLAG_NUMBER = 0x0002 # \p{N} +CODEPOINT_FLAG_LETTER = 0x0004 # \p{L} +CODEPOINT_FLAG_SEPARATOR = 0x0008 # \p{Z} +CODEPOINT_FLAG_MARK = 0x0010 # \p{M} +CODEPOINT_FLAG_PUNCTUATION = 0x0020 # \p{P} +CODEPOINT_FLAG_SYMBOL = 0x0040 # \p{S} +CODEPOINT_FLAG_CONTROL = 0x0080 # \p{C} + +UNICODE_CATEGORY_TO_FLAG = { + "Cn": CODEPOINT_FLAG_UNDEFINED, # Undefined + "Cc": CODEPOINT_FLAG_CONTROL, # Control + "Cf": CODEPOINT_FLAG_CONTROL, # Format + "Co": CODEPOINT_FLAG_CONTROL, # Private Use + "Cs": CODEPOINT_FLAG_CONTROL, # Surrrogate + "Ll": CODEPOINT_FLAG_LETTER, # Lowercase Letter + "Lm": CODEPOINT_FLAG_LETTER, # Modifier Letter + "Lo": CODEPOINT_FLAG_LETTER, # Other Letter + "Lt": CODEPOINT_FLAG_LETTER, # Titlecase Letter + "Lu": CODEPOINT_FLAG_LETTER, # Uppercase Letter + "L&": CODEPOINT_FLAG_LETTER, # Cased Letter + "Mc": CODEPOINT_FLAG_MARK, # Spacing Mark + "Me": CODEPOINT_FLAG_MARK, # Enclosing Mark + "Mn": CODEPOINT_FLAG_MARK, # Nonspacing Mark + "Nd": CODEPOINT_FLAG_NUMBER, # Decimal Number + "Nl": CODEPOINT_FLAG_NUMBER, # Letter Number + "No": CODEPOINT_FLAG_NUMBER, # Other Number + "Pc": CODEPOINT_FLAG_PUNCTUATION, # Connector Punctuation + "Pd": CODEPOINT_FLAG_PUNCTUATION, # Dash Punctuation + "Pe": CODEPOINT_FLAG_PUNCTUATION, # Close Punctuation + "Pf": CODEPOINT_FLAG_PUNCTUATION, # Final Punctuation + "Pi": CODEPOINT_FLAG_PUNCTUATION, # Initial Punctuation + "Po": CODEPOINT_FLAG_PUNCTUATION, # Other Punctuation + "Ps": CODEPOINT_FLAG_PUNCTUATION, # Open Punctuation + "Sc": CODEPOINT_FLAG_SYMBOL, # Currency Symbol + "Sk": CODEPOINT_FLAG_SYMBOL, # Modifier Symbol + "Sm": CODEPOINT_FLAG_SYMBOL, # Math Symbol + "So": CODEPOINT_FLAG_SYMBOL, # Other Symbol + "Zl": CODEPOINT_FLAG_SEPARATOR, # Line Separator + "Zp": CODEPOINT_FLAG_SEPARATOR, # Paragraph Separator + "Zs": CODEPOINT_FLAG_SEPARATOR, # Space Separator +} + + +codepoint_flags = array.array('H', [CODEPOINT_FLAG_UNDEFINED]) * MAX_CODEPOINTS table_whitespace = [] table_lowercase = [] table_uppercase = [] table_nfd = [] -for codepoint in range(MAX_CODEPOINTS): +for (cpt, cpt_lower, cpt_upper, categ, bidir) in unicode_data_iter(): # convert codepoint to unicode character - char = chr(codepoint) - - # regex categories - flags = codepoint_flags[codepoint] - flags.is_number = bool(regex_number.match(char)) - flags.is_letter = bool(regex_letter.match(char)) - flags.is_separator = bool(regex_separator.match(char)) - flags.is_accent_mark = bool(regex_accent_mark.match(char)) - flags.is_punctuation = bool(regex_punctuation.match(char)) - flags.is_symbol = bool(regex_symbol.match(char)) - flags.is_control = bool(regex_control.match(char)) - flags.is_undefined = bytes(flags)[0] == 0 - assert (not flags.is_undefined) - - # whitespaces - if bool(regex_whitespace.match(char)): - table_whitespace.append(codepoint) + char = chr(cpt) + + # codepoint category flags + codepoint_flags[cpt] = UNICODE_CATEGORY_TO_FLAG[categ] # lowercase conversion - lower = ord(char.lower()[0]) - if codepoint != lower: - table_lowercase.append((codepoint, lower)) + if cpt_lower: + table_lowercase.append((cpt, cpt_lower)) # uppercase conversion - upper = ord(char.upper()[0]) - if codepoint != upper: - table_uppercase.append((codepoint, upper)) + if cpt_upper: + table_uppercase.append((cpt, cpt_upper)) # NFD normalization norm = ord(unicodedata.normalize('NFD', char)[0]) - if codepoint != norm: - table_nfd.append((codepoint, norm)) + if cpt != norm: + table_nfd.append((cpt, norm)) + + +# whitespaces, see "" https://www.unicode.org/Public/UCD/latest/ucd/PropList.txt +table_whitespace.extend(range(0x0009, 0x000D + 1)) +table_whitespace.extend(range(0x2000, 0x200A + 1)) +table_whitespace.extend([0x0020, 0x0085, 0x00A0, 0x1680, 0x2028, 0x2029, 0x202F, 0x205F, 0x3000]) + + +# sort by codepoint +table_whitespace.sort() +table_lowercase.sort() +table_uppercase.sort() +table_nfd.sort() # group ranges with same flags ranges_flags = [(0, codepoint_flags[0])] # start, flags for codepoint, flags in enumerate(codepoint_flags): - if bytes(flags) != bytes(ranges_flags[-1][1]): + if flags != ranges_flags[-1][1]: ranges_flags.append((codepoint, flags)) -ranges_flags.append((MAX_CODEPOINTS, CoodepointFlags())) +ranges_flags.append((MAX_CODEPOINTS, 0x0000)) # group ranges with same nfd @@ -90,8 +150,8 @@ class CoodepointFlags (ctypes.Structure): ranges_nfd[-1] = (start, codepoint, norm) -# Generate 'unicode-data.cpp' - +# Generate 'unicode-data.cpp': +# python ./scripts//gen-unicode-data.py > unicode-data.cpp def out(line=""): print(line, end='\n') # noqa @@ -110,12 +170,12 @@ def out(line=""): out("const std::vector> unicode_ranges_flags = { // start, flags // last=next_start-1") for codepoint, flags in ranges_flags: - flags = int.from_bytes(bytes(flags), "little") out("{0x%06X, 0x%04X}," % (codepoint, flags)) out("};\n") out("const std::unordered_set unicode_set_whitespace = {") -out(", ".join("0x%06X" % cpt for cpt in table_whitespace)) +for codepoint in table_whitespace: + out("0x%06X," % codepoint) out("};\n") out("const std::unordered_map unicode_map_lowercase = {") diff --git a/tests/test-tokenizer-random.py b/tests/test-tokenizer-random.py index 52f589511e470..a07c52fb3fc60 100644 --- a/tests/test-tokenizer-random.py +++ b/tests/test-tokenizer-random.py @@ -11,13 +11,15 @@ import argparse import subprocess import random +import unicodedata from typing import Callable, Iterator import cffi from transformers import AutoTokenizer -logger = logging.getLogger("test-tokenizer-random-bpe") + +logger = logging.getLogger("test-tokenizer-random") class LibLlama: @@ -155,9 +157,14 @@ def generator_custom_text_edge_cases() -> Iterator[str]: 'Cửa Việt', # llama-3, ignore_merges = true 'a', # Phi-3 fail '<|endoftext|>', # Phi-3 fail - 'a\na', # TODO: Bert fail - 'a b', # rstrip phi-3 - 'a b', # lstrip jina-v2 + 'a\na', # bert fail + '"`', # falcon + ' \u2e4e', # falcon + 'a\xa0\xa0\x00b', # jina-v2-es + 'one ', # jina-v2-es lstrip=true + 'a b', # rstrip phi-3 + 'a b', # lstrip jina-v2 + '\xa0aC', # deepseek ] @@ -189,17 +196,23 @@ def generator_random_added_tokens(tokenizer, iterations=100) -> Iterator[str]: for m in range(iterations): rand.seed(m) words = rand.choices(all_tokens, k=500) - if words[0] == tokenizer.bos_token: # skip spam warning of double BOS + if words and words[0] == tokenizer.bos_token: # skip spam warning of double BOS while len(words) > 1 and words[1] == tokenizer.bos_token: # leave one starting BOS words.pop(0) if tokenizer.add_bos_token: # drop all starting BOS words.pop(0) + if words and words[-1] == tokenizer.eos_token: # skip spam warning of double EOS + while len(words) > 1 and words[-2] == tokenizer.eos_token: # leave one trailing EOS + words.pop(-1) + if tokenizer.add_bos_token: # drop all trailing EOS + words.pop(-1) yield "".join(words) def generator_random_chars(iterations=100) -> Iterator[str]: """Brute force random text with simple characters""" + NUM_WORDS = 400 WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5) CHARS = list(sorted(set(""" ABCDEFGHIJKLMNOPQRSTUVWXYZ @@ -213,12 +226,50 @@ def generator_random_chars(iterations=100) -> Iterator[str]: for m in range(iterations): rand.seed(m) text = [] - num_words = rand.randint(300, 400) - for i in range(num_words): + for _ in range(NUM_WORDS): k = rand.randint(1, 7) word = rand.choices(CHARS, k=k) - space = rand.choice(WHITESPACES) - text.append("".join(word) + space) + word.append(rand.choice(WHITESPACES)) + text.append("".join(word)) + yield "".join(text) + + +def generator_unicodes() -> Iterator[str]: + """Iterate unicode characters""" + + MAX_CODEPOINTS = 0x30000 # 0x110000 + + def _valid(cpt): + if cpt >= 0x30000: # unassigned and supplement­ary + return False + if 0x00D800 <= cpt <= 0x00F8FF: # Surrogates + return False + if unicodedata.category(chr(cpt)) == "Cn": + return False + return True + + characters = [chr(cpt) for cpt in range(1, MAX_CODEPOINTS) if _valid(cpt)] + + yield from characters + + +def generator_random_unicodes(iterations=100) -> Iterator[str]: + """Brute force random text with unicode characters""" + + NUM_WORDS = 200 + WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5) + + characters = list(generator_unicodes()) + + rand = random.Random() + for m in range(iterations): + rand.seed(m) + text = [] + for _ in range(NUM_WORDS): + k = rand.randint(1, 7) + word = rand.choices(characters, k=k) + word.append(rand.choice(WHITESPACES)) + text.append("".join(word)) yield "".join(text) @@ -256,25 +307,7 @@ def generator_random_vocab_words(vocab: list[str], iterations=100) -> Iterator[s yield "".join(text) -def generator_random_bytes(iterations=100) -> Iterator[str]: - """Brute force random bytes""" - - WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5) - - rand = random.Random() - for m in range(iterations): - rand.seed(m) - text = [] - num_words = rand.randint(300, 400) - for i in range(num_words): - k = rand.randint(1, 8) - word = [chr(r) for r in rand.randbytes(k) if r] - word.append(rand.choice(WHITESPACES)) - text.append("".join(word)) - yield "".join(text) - - -def test_compare_tokenizer(func_tokenize1: Callable, func_tokenize2: Callable, generator: Iterator[str]): +def compare_tokenizers(func_tokenize1: Callable, func_tokenize2: Callable, generator: Iterator[str]): def find_first_mismatch(ids1: list[int], ids2: list[int]): for i, (a, b) in enumerate(zip(ids1, ids2)): @@ -284,20 +317,34 @@ def find_first_mismatch(ids1: list[int], ids2: list[int]): return -1 return min(len(ids1), len(ids2)) - t0 = time.perf_counter() + t_tokenizer1 = 0 + t_tokenizer2 = 0 + t_start = time.perf_counter() + num_errors = 10 + logger.info("%s: %s" % (generator.__name__, "ini")) for text in generator: + # print(repr(text), hex(ord(text[0])), text.encode()) + t0 = time.perf_counter() ids1 = func_tokenize1(text) + t1 = time.perf_counter() ids2 = func_tokenize2(text) + t2 = time.perf_counter() + t_tokenizer1 += t1 - t0 + t_tokenizer2 += t2 - t1 if ids1 != ids2: i = find_first_mismatch(ids1, ids2) ids1 = list(ids1)[max(0, i - 2) : i + 5 + 1] ids2 = list(ids2)[max(0, i - 2) : i + 5 + 1] - logger.info(" TokenIDs: " + str(ids1)) - logger.info(" Expected: " + str(ids2)) - raise Exception() - t1 = time.perf_counter() - logger.info("%s: end, time: %.3f secs" % (generator.__name__, t1 - t0)) + logger.error(" TokenIDs: " + str(ids1)) + logger.error(" Expected: " + str(ids2)) + # raise Exception() + num_errors += 1 + if num_errors > 10: + break + + t_total = time.perf_counter() - t_start + logger.info("%s: end, tok1: %.3f tok2: %.3f total: %.3f" % (generator.__name__, t_tokenizer1, t_tokenizer2, t_total)) def main(argv: list[str] = None): @@ -307,7 +354,8 @@ def main(argv: list[str] = None): parser.add_argument("--verbose", action="store_true", help="increase output verbosity") args = parser.parse_args(argv) - logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO) + logging.basicConfig(level = logging.DEBUG if args.verbose else logging.INFO) + logger.info(f"VOCABFILE: '{args.vocab_file}'") model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096)) tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer) @@ -321,18 +369,22 @@ def func_tokenize2(text: str): ids = func_tokenize2("a") assert 1 <= len(ids) <= 3 add_bos_token = len(ids) > 1 and tokenizer.bos_token_id == ids[0] + add_eos_token = len(ids) > 1 and tokenizer.eos_token_id == ids[-1] tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", add_bos_token) + tokenizer.add_eos_token = getattr(tokenizer, "add_eos_token", add_eos_token) vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True))) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text()) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases()) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_vocab_words(vocab)) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_added_lr_strip(tokenizer)) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_added_tokens(tokenizer, 10_000)) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_chars(10_000)) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000)) - test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 5_000)) - # test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_bytes(10_000)) # FAIL + + compare_tokenizers(func_tokenize1, func_tokenize2, generator_custom_text()) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases()) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_unicodes()) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_vocab_words(vocab)) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_added_lr_strip(tokenizer)) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_added_tokens(tokenizer, 10_000)) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_chars(10_000)) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_unicodes(10_000)) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000)) + compare_tokenizers(func_tokenize1, func_tokenize2, generator_random_vocab_words(vocab, 5_000)) model.free() @@ -340,20 +392,40 @@ def func_tokenize2(text: str): if __name__ == "__main__": # main() + logging.basicConfig( + level = logging.DEBUG, + format = "%(asctime)s.%(msecs)03d %(name)s %(levelname)s %(message)s", + datefmt = "%Y-%m-%d %H:%M:%S", + filename = logger.name + ".log", + filemode = "a" + ) + path_tokenizers = "./models/tokenizers/" path_vocab_format = "./models/ggml-vocab-%s.gguf" # import os # tokenizers = os.listdir(path_tokenizers) tokenizers = [ - "llama-spm", # SPM - "phi-3", # SPM - "jina-v2-en", # WPM - "bert-bge", # WPM + # "llama-spm", # SPM + # "phi-3", # SPM + # "bert-bge", # WPM + # "jina-v2-en", # WPM + "gpt-2", # BPE + "llama-bpe", # BPE + "falcon", # BPE + "starcoder", # BPE + "jina-v2-es", # BPE + "jina-v2-de", # BPE + "jina-v2-code", # BPE + "smaug-bpe", # BPE + "phi-2", # BPE + "deepseek-coder", # BPE + "deepseek-llm", # BPE ] for tokenizer in tokenizers: - print("\n" + "=" * 50 + "\n" + tokenizer + "\n") # noqa + logger.info("=" * 50) + logger.info(f"TOKENIZER: '{tokenizer}'") vocab_file = path_vocab_format % tokenizer dir_tokenizer = path_tokenizers + "/" + tokenizer main([vocab_file, dir_tokenizer, "--verbose"]) diff --git a/unicode-data.cpp b/unicode-data.cpp index d7c1c898d8828..4a939898b367f 100644 --- a/unicode-data.cpp +++ b/unicode-data.cpp @@ -68,36 +68,36 @@ const std::vector> unicode_ranges_flags = { // st {0x000370, 0x0004}, {0x000375, 0x0040}, {0x000376, 0x0004}, -{0x000378, 0x0080}, +{0x000378, 0x0001}, {0x00037A, 0x0004}, {0x00037E, 0x0020}, {0x00037F, 0x0004}, -{0x000380, 0x0080}, +{0x000380, 0x0001}, {0x000384, 0x0040}, {0x000386, 0x0004}, {0x000387, 0x0020}, {0x000388, 0x0004}, -{0x00038B, 0x0080}, +{0x00038B, 0x0001}, {0x00038C, 0x0004}, -{0x00038D, 0x0080}, +{0x00038D, 0x0001}, {0x00038E, 0x0004}, -{0x0003A2, 0x0080}, +{0x0003A2, 0x0001}, {0x0003A3, 0x0004}, {0x0003F6, 0x0040}, {0x0003F7, 0x0004}, {0x000482, 0x0040}, {0x000483, 0x0010}, {0x00048A, 0x0004}, -{0x000530, 0x0080}, +{0x000530, 0x0001}, {0x000531, 0x0004}, -{0x000557, 0x0080}, +{0x000557, 0x0001}, {0x000559, 0x0004}, {0x00055A, 0x0020}, {0x000560, 0x0004}, {0x000589, 0x0020}, -{0x00058B, 0x0080}, +{0x00058B, 0x0001}, {0x00058D, 0x0040}, -{0x000590, 0x0080}, +{0x000590, 0x0001}, {0x000591, 0x0010}, {0x0005BE, 0x0020}, {0x0005BF, 0x0010}, @@ -107,12 +107,13 @@ const std::vector> unicode_ranges_flags = { // st {0x0005C4, 0x0010}, {0x0005C6, 0x0020}, {0x0005C7, 0x0010}, -{0x0005C8, 0x0080}, +{0x0005C8, 0x0001}, {0x0005D0, 0x0004}, -{0x0005EB, 0x0080}, +{0x0005EB, 0x0001}, {0x0005EF, 0x0004}, {0x0005F3, 0x0020}, -{0x0005F5, 0x0080}, +{0x0005F5, 0x0001}, +{0x000600, 0x0080}, {0x000606, 0x0040}, {0x000609, 0x0020}, {0x00060B, 0x0040}, @@ -145,16 +146,17 @@ const std::vector> unicode_ranges_flags = { // st {0x0006FD, 0x0040}, {0x0006FF, 0x0004}, {0x000700, 0x0020}, -{0x00070E, 0x0080}, +{0x00070E, 0x0001}, +{0x00070F, 0x0080}, {0x000710, 0x0004}, {0x000711, 0x0010}, {0x000712, 0x0004}, {0x000730, 0x0010}, -{0x00074B, 0x0080}, +{0x00074B, 0x0001}, {0x00074D, 0x0004}, {0x0007A6, 0x0010}, {0x0007B1, 0x0004}, -{0x0007B2, 0x0080}, +{0x0007B2, 0x0001}, {0x0007C0, 0x0002}, {0x0007CA, 0x0004}, {0x0007EB, 0x0010}, @@ -162,7 +164,7 @@ const std::vector> unicode_ranges_flags = { // st {0x0007F6, 0x0040}, {0x0007F7, 0x0020}, {0x0007FA, 0x0004}, -{0x0007FB, 0x0080}, +{0x0007FB, 0x0001}, {0x0007FD, 0x0010}, {0x0007FE, 0x0040}, {0x000800, 0x0004}, @@ -173,20 +175,22 @@ const std::vector> unicode_ranges_flags = { // st {0x000825, 0x0010}, {0x000828, 0x0004}, {0x000829, 0x0010}, -{0x00082E, 0x0080}, +{0x00082E, 0x0001}, {0x000830, 0x0020}, -{0x00083F, 0x0080}, +{0x00083F, 0x0001}, {0x000840, 0x0004}, {0x000859, 0x0010}, -{0x00085C, 0x0080}, +{0x00085C, 0x0001}, {0x00085E, 0x0020}, -{0x00085F, 0x0080}, +{0x00085F, 0x0001}, {0x000860, 0x0004}, -{0x00086B, 0x0080}, +{0x00086B, 0x0001}, {0x000870, 0x0004}, {0x000888, 0x0040}, {0x000889, 0x0004}, -{0x00088F, 0x0080}, +{0x00088F, 0x0001}, +{0x000890, 0x0080}, +{0x000892, 0x0001}, {0x000898, 0x0010}, {0x0008A0, 0x0004}, {0x0008CA, 0x0010}, @@ -205,35 +209,35 @@ const std::vector> unicode_ranges_flags = { // st {0x000970, 0x0020}, {0x000971, 0x0004}, {0x000981, 0x0010}, -{0x000984, 0x0080}, +{0x000984, 0x0001}, {0x000985, 0x0004}, -{0x00098D, 0x0080}, +{0x00098D, 0x0001}, {0x00098F, 0x0004}, -{0x000991, 0x0080}, +{0x000991, 0x0001}, {0x000993, 0x0004}, -{0x0009A9, 0x0080}, +{0x0009A9, 0x0001}, {0x0009AA, 0x0004}, -{0x0009B1, 0x0080}, +{0x0009B1, 0x0001}, {0x0009B2, 0x0004}, -{0x0009B3, 0x0080}, +{0x0009B3, 0x0001}, {0x0009B6, 0x0004}, -{0x0009BA, 0x0080}, +{0x0009BA, 0x0001}, {0x0009BC, 0x0010}, {0x0009BD, 0x0004}, {0x0009BE, 0x0010}, -{0x0009C5, 0x0080}, +{0x0009C5, 0x0001}, {0x0009C7, 0x0010}, -{0x0009C9, 0x0080}, +{0x0009C9, 0x0001}, {0x0009CB, 0x0010}, {0x0009CE, 0x0004}, -{0x0009CF, 0x0080}, +{0x0009CF, 0x0001}, {0x0009D7, 0x0010}, -{0x0009D8, 0x0080}, +{0x0009D8, 0x0001}, {0x0009DC, 0x0004}, -{0x0009DE, 0x0080}, +{0x0009DE, 0x0001}, {0x0009DF, 0x0004}, {0x0009E2, 0x0010}, -{0x0009E4, 0x0080}, +{0x0009E4, 0x0001}, {0x0009E6, 0x0002}, {0x0009F0, 0x0004}, {0x0009F2, 0x0040}, @@ -242,173 +246,173 @@ const std::vector> unicode_ranges_flags = { // st {0x0009FC, 0x0004}, {0x0009FD, 0x0020}, {0x0009FE, 0x0010}, -{0x0009FF, 0x0080}, +{0x0009FF, 0x0001}, {0x000A01, 0x0010}, -{0x000A04, 0x0080}, +{0x000A04, 0x0001}, {0x000A05, 0x0004}, -{0x000A0B, 0x0080}, +{0x000A0B, 0x0001}, {0x000A0F, 0x0004}, -{0x000A11, 0x0080}, +{0x000A11, 0x0001}, {0x000A13, 0x0004}, -{0x000A29, 0x0080}, +{0x000A29, 0x0001}, {0x000A2A, 0x0004}, -{0x000A31, 0x0080}, +{0x000A31, 0x0001}, {0x000A32, 0x0004}, -{0x000A34, 0x0080}, +{0x000A34, 0x0001}, {0x000A35, 0x0004}, -{0x000A37, 0x0080}, +{0x000A37, 0x0001}, {0x000A38, 0x0004}, -{0x000A3A, 0x0080}, +{0x000A3A, 0x0001}, {0x000A3C, 0x0010}, -{0x000A3D, 0x0080}, +{0x000A3D, 0x0001}, {0x000A3E, 0x0010}, -{0x000A43, 0x0080}, +{0x000A43, 0x0001}, {0x000A47, 0x0010}, -{0x000A49, 0x0080}, +{0x000A49, 0x0001}, {0x000A4B, 0x0010}, -{0x000A4E, 0x0080}, +{0x000A4E, 0x0001}, {0x000A51, 0x0010}, -{0x000A52, 0x0080}, +{0x000A52, 0x0001}, {0x000A59, 0x0004}, -{0x000A5D, 0x0080}, +{0x000A5D, 0x0001}, {0x000A5E, 0x0004}, -{0x000A5F, 0x0080}, +{0x000A5F, 0x0001}, {0x000A66, 0x0002}, {0x000A70, 0x0010}, {0x000A72, 0x0004}, {0x000A75, 0x0010}, {0x000A76, 0x0020}, -{0x000A77, 0x0080}, +{0x000A77, 0x0001}, {0x000A81, 0x0010}, -{0x000A84, 0x0080}, +{0x000A84, 0x0001}, {0x000A85, 0x0004}, -{0x000A8E, 0x0080}, +{0x000A8E, 0x0001}, {0x000A8F, 0x0004}, -{0x000A92, 0x0080}, +{0x000A92, 0x0001}, {0x000A93, 0x0004}, -{0x000AA9, 0x0080}, +{0x000AA9, 0x0001}, {0x000AAA, 0x0004}, -{0x000AB1, 0x0080}, +{0x000AB1, 0x0001}, {0x000AB2, 0x0004}, -{0x000AB4, 0x0080}, +{0x000AB4, 0x0001}, {0x000AB5, 0x0004}, -{0x000ABA, 0x0080}, +{0x000ABA, 0x0001}, {0x000ABC, 0x0010}, {0x000ABD, 0x0004}, {0x000ABE, 0x0010}, -{0x000AC6, 0x0080}, +{0x000AC6, 0x0001}, {0x000AC7, 0x0010}, -{0x000ACA, 0x0080}, +{0x000ACA, 0x0001}, {0x000ACB, 0x0010}, -{0x000ACE, 0x0080}, +{0x000ACE, 0x0001}, {0x000AD0, 0x0004}, -{0x000AD1, 0x0080}, +{0x000AD1, 0x0001}, {0x000AE0, 0x0004}, {0x000AE2, 0x0010}, -{0x000AE4, 0x0080}, +{0x000AE4, 0x0001}, {0x000AE6, 0x0002}, {0x000AF0, 0x0020}, {0x000AF1, 0x0040}, -{0x000AF2, 0x0080}, +{0x000AF2, 0x0001}, {0x000AF9, 0x0004}, {0x000AFA, 0x0010}, -{0x000B00, 0x0080}, +{0x000B00, 0x0001}, {0x000B01, 0x0010}, -{0x000B04, 0x0080}, +{0x000B04, 0x0001}, {0x000B05, 0x0004}, -{0x000B0D, 0x0080}, +{0x000B0D, 0x0001}, {0x000B0F, 0x0004}, -{0x000B11, 0x0080}, +{0x000B11, 0x0001}, {0x000B13, 0x0004}, -{0x000B29, 0x0080}, +{0x000B29, 0x0001}, {0x000B2A, 0x0004}, -{0x000B31, 0x0080}, +{0x000B31, 0x0001}, {0x000B32, 0x0004}, -{0x000B34, 0x0080}, +{0x000B34, 0x0001}, {0x000B35, 0x0004}, -{0x000B3A, 0x0080}, +{0x000B3A, 0x0001}, {0x000B3C, 0x0010}, {0x000B3D, 0x0004}, {0x000B3E, 0x0010}, -{0x000B45, 0x0080}, +{0x000B45, 0x0001}, {0x000B47, 0x0010}, -{0x000B49, 0x0080}, +{0x000B49, 0x0001}, {0x000B4B, 0x0010}, -{0x000B4E, 0x0080}, +{0x000B4E, 0x0001}, {0x000B55, 0x0010}, -{0x000B58, 0x0080}, +{0x000B58, 0x0001}, {0x000B5C, 0x0004}, -{0x000B5E, 0x0080}, +{0x000B5E, 0x0001}, {0x000B5F, 0x0004}, {0x000B62, 0x0010}, -{0x000B64, 0x0080}, +{0x000B64, 0x0001}, {0x000B66, 0x0002}, {0x000B70, 0x0040}, {0x000B71, 0x0004}, {0x000B72, 0x0002}, -{0x000B78, 0x0080}, +{0x000B78, 0x0001}, {0x000B82, 0x0010}, {0x000B83, 0x0004}, -{0x000B84, 0x0080}, +{0x000B84, 0x0001}, {0x000B85, 0x0004}, -{0x000B8B, 0x0080}, +{0x000B8B, 0x0001}, {0x000B8E, 0x0004}, -{0x000B91, 0x0080}, +{0x000B91, 0x0001}, {0x000B92, 0x0004}, -{0x000B96, 0x0080}, +{0x000B96, 0x0001}, {0x000B99, 0x0004}, -{0x000B9B, 0x0080}, +{0x000B9B, 0x0001}, {0x000B9C, 0x0004}, -{0x000B9D, 0x0080}, +{0x000B9D, 0x0001}, {0x000B9E, 0x0004}, -{0x000BA0, 0x0080}, +{0x000BA0, 0x0001}, {0x000BA3, 0x0004}, -{0x000BA5, 0x0080}, +{0x000BA5, 0x0001}, {0x000BA8, 0x0004}, -{0x000BAB, 0x0080}, +{0x000BAB, 0x0001}, {0x000BAE, 0x0004}, -{0x000BBA, 0x0080}, +{0x000BBA, 0x0001}, {0x000BBE, 0x0010}, -{0x000BC3, 0x0080}, +{0x000BC3, 0x0001}, {0x000BC6, 0x0010}, -{0x000BC9, 0x0080}, +{0x000BC9, 0x0001}, {0x000BCA, 0x0010}, -{0x000BCE, 0x0080}, +{0x000BCE, 0x0001}, {0x000BD0, 0x0004}, -{0x000BD1, 0x0080}, +{0x000BD1, 0x0001}, {0x000BD7, 0x0010}, -{0x000BD8, 0x0080}, +{0x000BD8, 0x0001}, {0x000BE6, 0x0002}, {0x000BF3, 0x0040}, -{0x000BFB, 0x0080}, +{0x000BFB, 0x0001}, {0x000C00, 0x0010}, {0x000C05, 0x0004}, -{0x000C0D, 0x0080}, +{0x000C0D, 0x0001}, {0x000C0E, 0x0004}, -{0x000C11, 0x0080}, +{0x000C11, 0x0001}, {0x000C12, 0x0004}, -{0x000C29, 0x0080}, +{0x000C29, 0x0001}, {0x000C2A, 0x0004}, -{0x000C3A, 0x0080}, +{0x000C3A, 0x0001}, {0x000C3C, 0x0010}, {0x000C3D, 0x0004}, {0x000C3E, 0x0010}, -{0x000C45, 0x0080}, +{0x000C45, 0x0001}, {0x000C46, 0x0010}, -{0x000C49, 0x0080}, +{0x000C49, 0x0001}, {0x000C4A, 0x0010}, -{0x000C4E, 0x0080}, +{0x000C4E, 0x0001}, {0x000C55, 0x0010}, -{0x000C57, 0x0080}, +{0x000C57, 0x0001}, {0x000C58, 0x0004}, -{0x000C5B, 0x0080}, +{0x000C5B, 0x0001}, {0x000C5D, 0x0004}, -{0x000C5E, 0x0080}, +{0x000C5E, 0x0001}, {0x000C60, 0x0004}, {0x000C62, 0x0010}, -{0x000C64, 0x0080}, +{0x000C64, 0x0001}, {0x000C66, 0x0002}, -{0x000C70, 0x0080}, +{0x000C70, 0x0001}, {0x000C77, 0x0020}, {0x000C78, 0x0002}, {0x000C7F, 0x0040}, @@ -416,124 +420,124 @@ const std::vector> unicode_ranges_flags = { // st {0x000C81, 0x0010}, {0x000C84, 0x0020}, {0x000C85, 0x0004}, -{0x000C8D, 0x0080}, +{0x000C8D, 0x0001}, {0x000C8E, 0x0004}, -{0x000C91, 0x0080}, +{0x000C91, 0x0001}, {0x000C92, 0x0004}, -{0x000CA9, 0x0080}, +{0x000CA9, 0x0001}, {0x000CAA, 0x0004}, -{0x000CB4, 0x0080}, +{0x000CB4, 0x0001}, {0x000CB5, 0x0004}, -{0x000CBA, 0x0080}, +{0x000CBA, 0x0001}, {0x000CBC, 0x0010}, {0x000CBD, 0x0004}, {0x000CBE, 0x0010}, -{0x000CC5, 0x0080}, +{0x000CC5, 0x0001}, {0x000CC6, 0x0010}, -{0x000CC9, 0x0080}, +{0x000CC9, 0x0001}, {0x000CCA, 0x0010}, -{0x000CCE, 0x0080}, +{0x000CCE, 0x0001}, {0x000CD5, 0x0010}, -{0x000CD7, 0x0080}, +{0x000CD7, 0x0001}, {0x000CDD, 0x0004}, -{0x000CDF, 0x0080}, +{0x000CDF, 0x0001}, {0x000CE0, 0x0004}, {0x000CE2, 0x0010}, -{0x000CE4, 0x0080}, +{0x000CE4, 0x0001}, {0x000CE6, 0x0002}, -{0x000CF0, 0x0080}, +{0x000CF0, 0x0001}, {0x000CF1, 0x0004}, {0x000CF3, 0x0010}, -{0x000CF4, 0x0080}, +{0x000CF4, 0x0001}, {0x000D00, 0x0010}, {0x000D04, 0x0004}, -{0x000D0D, 0x0080}, +{0x000D0D, 0x0001}, {0x000D0E, 0x0004}, -{0x000D11, 0x0080}, +{0x000D11, 0x0001}, {0x000D12, 0x0004}, {0x000D3B, 0x0010}, {0x000D3D, 0x0004}, {0x000D3E, 0x0010}, -{0x000D45, 0x0080}, +{0x000D45, 0x0001}, {0x000D46, 0x0010}, -{0x000D49, 0x0080}, +{0x000D49, 0x0001}, {0x000D4A, 0x0010}, {0x000D4E, 0x0004}, {0x000D4F, 0x0040}, -{0x000D50, 0x0080}, +{0x000D50, 0x0001}, {0x000D54, 0x0004}, {0x000D57, 0x0010}, {0x000D58, 0x0002}, {0x000D5F, 0x0004}, {0x000D62, 0x0010}, -{0x000D64, 0x0080}, +{0x000D64, 0x0001}, {0x000D66, 0x0002}, {0x000D79, 0x0040}, {0x000D7A, 0x0004}, -{0x000D80, 0x0080}, +{0x000D80, 0x0001}, {0x000D81, 0x0010}, -{0x000D84, 0x0080}, +{0x000D84, 0x0001}, {0x000D85, 0x0004}, -{0x000D97, 0x0080}, +{0x000D97, 0x0001}, {0x000D9A, 0x0004}, -{0x000DB2, 0x0080}, +{0x000DB2, 0x0001}, {0x000DB3, 0x0004}, -{0x000DBC, 0x0080}, +{0x000DBC, 0x0001}, {0x000DBD, 0x0004}, -{0x000DBE, 0x0080}, +{0x000DBE, 0x0001}, {0x000DC0, 0x0004}, -{0x000DC7, 0x0080}, +{0x000DC7, 0x0001}, {0x000DCA, 0x0010}, -{0x000DCB, 0x0080}, +{0x000DCB, 0x0001}, {0x000DCF, 0x0010}, -{0x000DD5, 0x0080}, +{0x000DD5, 0x0001}, {0x000DD6, 0x0010}, -{0x000DD7, 0x0080}, +{0x000DD7, 0x0001}, {0x000DD8, 0x0010}, -{0x000DE0, 0x0080}, +{0x000DE0, 0x0001}, {0x000DE6, 0x0002}, -{0x000DF0, 0x0080}, +{0x000DF0, 0x0001}, {0x000DF2, 0x0010}, {0x000DF4, 0x0020}, -{0x000DF5, 0x0080}, +{0x000DF5, 0x0001}, {0x000E01, 0x0004}, {0x000E31, 0x0010}, {0x000E32, 0x0004}, {0x000E34, 0x0010}, -{0x000E3B, 0x0080}, +{0x000E3B, 0x0001}, {0x000E3F, 0x0040}, {0x000E40, 0x0004}, {0x000E47, 0x0010}, {0x000E4F, 0x0020}, {0x000E50, 0x0002}, {0x000E5A, 0x0020}, -{0x000E5C, 0x0080}, +{0x000E5C, 0x0001}, {0x000E81, 0x0004}, -{0x000E83, 0x0080}, +{0x000E83, 0x0001}, {0x000E84, 0x0004}, -{0x000E85, 0x0080}, +{0x000E85, 0x0001}, {0x000E86, 0x0004}, -{0x000E8B, 0x0080}, +{0x000E8B, 0x0001}, {0x000E8C, 0x0004}, -{0x000EA4, 0x0080}, +{0x000EA4, 0x0001}, {0x000EA5, 0x0004}, -{0x000EA6, 0x0080}, +{0x000EA6, 0x0001}, {0x000EA7, 0x0004}, {0x000EB1, 0x0010}, {0x000EB2, 0x0004}, {0x000EB4, 0x0010}, {0x000EBD, 0x0004}, -{0x000EBE, 0x0080}, +{0x000EBE, 0x0001}, {0x000EC0, 0x0004}, -{0x000EC5, 0x0080}, +{0x000EC5, 0x0001}, {0x000EC6, 0x0004}, -{0x000EC7, 0x0080}, +{0x000EC7, 0x0001}, {0x000EC8, 0x0010}, -{0x000ECF, 0x0080}, +{0x000ECF, 0x0001}, {0x000ED0, 0x0002}, -{0x000EDA, 0x0080}, +{0x000EDA, 0x0001}, {0x000EDC, 0x0004}, -{0x000EE0, 0x0080}, +{0x000EE0, 0x0001}, {0x000F00, 0x0004}, {0x000F01, 0x0040}, {0x000F04, 0x0020}, @@ -552,26 +556,26 @@ const std::vector> unicode_ranges_flags = { // st {0x000F3A, 0x0020}, {0x000F3E, 0x0010}, {0x000F40, 0x0004}, -{0x000F48, 0x0080}, +{0x000F48, 0x0001}, {0x000F49, 0x0004}, -{0x000F6D, 0x0080}, +{0x000F6D, 0x0001}, {0x000F71, 0x0010}, {0x000F85, 0x0020}, {0x000F86, 0x0010}, {0x000F88, 0x0004}, {0x000F8D, 0x0010}, -{0x000F98, 0x0080}, +{0x000F98, 0x0001}, {0x000F99, 0x0010}, -{0x000FBD, 0x0080}, +{0x000FBD, 0x0001}, {0x000FBE, 0x0040}, {0x000FC6, 0x0010}, {0x000FC7, 0x0040}, -{0x000FCD, 0x0080}, +{0x000FCD, 0x0001}, {0x000FCE, 0x0040}, {0x000FD0, 0x0020}, {0x000FD5, 0x0040}, {0x000FD9, 0x0020}, -{0x000FDB, 0x0080}, +{0x000FDB, 0x0001}, {0x001000, 0x0004}, {0x00102B, 0x0010}, {0x00103F, 0x0004}, @@ -595,56 +599,56 @@ const std::vector> unicode_ranges_flags = { // st {0x00109A, 0x0010}, {0x00109E, 0x0040}, {0x0010A0, 0x0004}, -{0x0010C6, 0x0080}, +{0x0010C6, 0x0001}, {0x0010C7, 0x0004}, -{0x0010C8, 0x0080}, +{0x0010C8, 0x0001}, {0x0010CD, 0x0004}, -{0x0010CE, 0x0080}, +{0x0010CE, 0x0001}, {0x0010D0, 0x0004}, {0x0010FB, 0x0020}, {0x0010FC, 0x0004}, -{0x001249, 0x0080}, +{0x001249, 0x0001}, {0x00124A, 0x0004}, -{0x00124E, 0x0080}, +{0x00124E, 0x0001}, {0x001250, 0x0004}, -{0x001257, 0x0080}, +{0x001257, 0x0001}, {0x001258, 0x0004}, -{0x001259, 0x0080}, +{0x001259, 0x0001}, {0x00125A, 0x0004}, -{0x00125E, 0x0080}, +{0x00125E, 0x0001}, {0x001260, 0x0004}, -{0x001289, 0x0080}, +{0x001289, 0x0001}, {0x00128A, 0x0004}, -{0x00128E, 0x0080}, +{0x00128E, 0x0001}, {0x001290, 0x0004}, -{0x0012B1, 0x0080}, +{0x0012B1, 0x0001}, {0x0012B2, 0x0004}, -{0x0012B6, 0x0080}, +{0x0012B6, 0x0001}, {0x0012B8, 0x0004}, -{0x0012BF, 0x0080}, +{0x0012BF, 0x0001}, {0x0012C0, 0x0004}, -{0x0012C1, 0x0080}, +{0x0012C1, 0x0001}, {0x0012C2, 0x0004}, -{0x0012C6, 0x0080}, +{0x0012C6, 0x0001}, {0x0012C8, 0x0004}, -{0x0012D7, 0x0080}, +{0x0012D7, 0x0001}, {0x0012D8, 0x0004}, -{0x001311, 0x0080}, +{0x001311, 0x0001}, {0x001312, 0x0004}, -{0x001316, 0x0080}, +{0x001316, 0x0001}, {0x001318, 0x0004}, -{0x00135B, 0x0080}, +{0x00135B, 0x0001}, {0x00135D, 0x0010}, {0x001360, 0x0020}, {0x001369, 0x0002}, -{0x00137D, 0x0080}, +{0x00137D, 0x0001}, {0x001380, 0x0004}, {0x001390, 0x0040}, -{0x00139A, 0x0080}, +{0x00139A, 0x0001}, {0x0013A0, 0x0004}, -{0x0013F6, 0x0080}, +{0x0013F6, 0x0001}, {0x0013F8, 0x0004}, -{0x0013FE, 0x0080}, +{0x0013FE, 0x0001}, {0x001400, 0x0020}, {0x001401, 0x0004}, {0x00166D, 0x0040}, @@ -653,28 +657,28 @@ const std::vector> unicode_ranges_flags = { // st {0x001680, 0x0008}, {0x001681, 0x0004}, {0x00169B, 0x0020}, -{0x00169D, 0x0080}, +{0x00169D, 0x0001}, {0x0016A0, 0x0004}, {0x0016EB, 0x0020}, {0x0016EE, 0x0002}, {0x0016F1, 0x0004}, -{0x0016F9, 0x0080}, +{0x0016F9, 0x0001}, {0x001700, 0x0004}, {0x001712, 0x0010}, -{0x001716, 0x0080}, +{0x001716, 0x0001}, {0x00171F, 0x0004}, {0x001732, 0x0010}, {0x001735, 0x0020}, -{0x001737, 0x0080}, +{0x001737, 0x0001}, {0x001740, 0x0004}, {0x001752, 0x0010}, -{0x001754, 0x0080}, +{0x001754, 0x0001}, {0x001760, 0x0004}, -{0x00176D, 0x0080}, +{0x00176D, 0x0001}, {0x00176E, 0x0004}, -{0x001771, 0x0080}, +{0x001771, 0x0001}, {0x001772, 0x0010}, -{0x001774, 0x0080}, +{0x001774, 0x0001}, {0x001780, 0x0004}, {0x0017B4, 0x0010}, {0x0017D4, 0x0020}, @@ -683,80 +687,80 @@ const std::vector> unicode_ranges_flags = { // st {0x0017DB, 0x0040}, {0x0017DC, 0x0004}, {0x0017DD, 0x0010}, -{0x0017DE, 0x0080}, +{0x0017DE, 0x0001}, {0x0017E0, 0x0002}, -{0x0017EA, 0x0080}, +{0x0017EA, 0x0001}, {0x0017F0, 0x0002}, -{0x0017FA, 0x0080}, +{0x0017FA, 0x0001}, {0x001800, 0x0020}, {0x00180B, 0x0010}, {0x00180E, 0x0080}, {0x00180F, 0x0010}, {0x001810, 0x0002}, -{0x00181A, 0x0080}, +{0x00181A, 0x0001}, {0x001820, 0x0004}, -{0x001879, 0x0080}, +{0x001879, 0x0001}, {0x001880, 0x0004}, {0x001885, 0x0010}, {0x001887, 0x0004}, {0x0018A9, 0x0010}, {0x0018AA, 0x0004}, -{0x0018AB, 0x0080}, +{0x0018AB, 0x0001}, {0x0018B0, 0x0004}, -{0x0018F6, 0x0080}, +{0x0018F6, 0x0001}, {0x001900, 0x0004}, -{0x00191F, 0x0080}, +{0x00191F, 0x0001}, {0x001920, 0x0010}, -{0x00192C, 0x0080}, +{0x00192C, 0x0001}, {0x001930, 0x0010}, -{0x00193C, 0x0080}, +{0x00193C, 0x0001}, {0x001940, 0x0040}, -{0x001941, 0x0080}, +{0x001941, 0x0001}, {0x001944, 0x0020}, {0x001946, 0x0002}, {0x001950, 0x0004}, -{0x00196E, 0x0080}, +{0x00196E, 0x0001}, {0x001970, 0x0004}, -{0x001975, 0x0080}, +{0x001975, 0x0001}, {0x001980, 0x0004}, -{0x0019AC, 0x0080}, +{0x0019AC, 0x0001}, {0x0019B0, 0x0004}, -{0x0019CA, 0x0080}, +{0x0019CA, 0x0001}, {0x0019D0, 0x0002}, -{0x0019DB, 0x0080}, +{0x0019DB, 0x0001}, {0x0019DE, 0x0040}, {0x001A00, 0x0004}, {0x001A17, 0x0010}, -{0x001A1C, 0x0080}, +{0x001A1C, 0x0001}, {0x001A1E, 0x0020}, {0x001A20, 0x0004}, {0x001A55, 0x0010}, -{0x001A5F, 0x0080}, +{0x001A5F, 0x0001}, {0x001A60, 0x0010}, -{0x001A7D, 0x0080}, +{0x001A7D, 0x0001}, {0x001A7F, 0x0010}, {0x001A80, 0x0002}, -{0x001A8A, 0x0080}, +{0x001A8A, 0x0001}, {0x001A90, 0x0002}, -{0x001A9A, 0x0080}, +{0x001A9A, 0x0001}, {0x001AA0, 0x0020}, {0x001AA7, 0x0004}, {0x001AA8, 0x0020}, -{0x001AAE, 0x0080}, +{0x001AAE, 0x0001}, {0x001AB0, 0x0010}, -{0x001ACF, 0x0080}, +{0x001ACF, 0x0001}, {0x001B00, 0x0010}, {0x001B05, 0x0004}, {0x001B34, 0x0010}, {0x001B45, 0x0004}, -{0x001B4D, 0x0080}, +{0x001B4D, 0x0001}, {0x001B50, 0x0002}, {0x001B5A, 0x0020}, {0x001B61, 0x0040}, {0x001B6B, 0x0010}, {0x001B74, 0x0040}, {0x001B7D, 0x0020}, -{0x001B7F, 0x0080}, +{0x001B7F, 0x0001}, {0x001B80, 0x0010}, {0x001B83, 0x0004}, {0x001BA1, 0x0010}, @@ -764,25 +768,25 @@ const std::vector> unicode_ranges_flags = { // st {0x001BB0, 0x0002}, {0x001BBA, 0x0004}, {0x001BE6, 0x0010}, -{0x001BF4, 0x0080}, +{0x001BF4, 0x0001}, {0x001BFC, 0x0020}, {0x001C00, 0x0004}, {0x001C24, 0x0010}, -{0x001C38, 0x0080}, +{0x001C38, 0x0001}, {0x001C3B, 0x0020}, {0x001C40, 0x0002}, -{0x001C4A, 0x0080}, +{0x001C4A, 0x0001}, {0x001C4D, 0x0004}, {0x001C50, 0x0002}, {0x001C5A, 0x0004}, {0x001C7E, 0x0020}, {0x001C80, 0x0004}, -{0x001C89, 0x0080}, +{0x001C89, 0x0001}, {0x001C90, 0x0004}, -{0x001CBB, 0x0080}, +{0x001CBB, 0x0001}, {0x001CBD, 0x0004}, {0x001CC0, 0x0020}, -{0x001CC8, 0x0080}, +{0x001CC8, 0x0001}, {0x001CD0, 0x0010}, {0x001CD3, 0x0020}, {0x001CD4, 0x0010}, @@ -793,50 +797,50 @@ const std::vector> unicode_ranges_flags = { // st {0x001CF5, 0x0004}, {0x001CF7, 0x0010}, {0x001CFA, 0x0004}, -{0x001CFB, 0x0080}, +{0x001CFB, 0x0001}, {0x001D00, 0x0004}, {0x001DC0, 0x0010}, {0x001E00, 0x0004}, -{0x001F16, 0x0080}, +{0x001F16, 0x0001}, {0x001F18, 0x0004}, -{0x001F1E, 0x0080}, +{0x001F1E, 0x0001}, {0x001F20, 0x0004}, -{0x001F46, 0x0080}, +{0x001F46, 0x0001}, {0x001F48, 0x0004}, -{0x001F4E, 0x0080}, +{0x001F4E, 0x0001}, {0x001F50, 0x0004}, -{0x001F58, 0x0080}, +{0x001F58, 0x0001}, {0x001F59, 0x0004}, -{0x001F5A, 0x0080}, +{0x001F5A, 0x0001}, {0x001F5B, 0x0004}, -{0x001F5C, 0x0080}, +{0x001F5C, 0x0001}, {0x001F5D, 0x0004}, -{0x001F5E, 0x0080}, +{0x001F5E, 0x0001}, {0x001F5F, 0x0004}, -{0x001F7E, 0x0080}, +{0x001F7E, 0x0001}, {0x001F80, 0x0004}, -{0x001FB5, 0x0080}, +{0x001FB5, 0x0001}, {0x001FB6, 0x0004}, {0x001FBD, 0x0040}, {0x001FBE, 0x0004}, {0x001FBF, 0x0040}, {0x001FC2, 0x0004}, -{0x001FC5, 0x0080}, +{0x001FC5, 0x0001}, {0x001FC6, 0x0004}, {0x001FCD, 0x0040}, {0x001FD0, 0x0004}, -{0x001FD4, 0x0080}, +{0x001FD4, 0x0001}, {0x001FD6, 0x0004}, -{0x001FDC, 0x0080}, +{0x001FDC, 0x0001}, {0x001FDD, 0x0040}, {0x001FE0, 0x0004}, {0x001FED, 0x0040}, -{0x001FF0, 0x0080}, +{0x001FF0, 0x0001}, {0x001FF2, 0x0004}, -{0x001FF5, 0x0080}, +{0x001FF5, 0x0001}, {0x001FF6, 0x0004}, {0x001FFD, 0x0040}, -{0x001FFF, 0x0080}, +{0x001FFF, 0x0001}, {0x002000, 0x0008}, {0x00200B, 0x0080}, {0x002010, 0x0020}, @@ -850,9 +854,11 @@ const std::vector> unicode_ranges_flags = { // st {0x002053, 0x0020}, {0x00205F, 0x0008}, {0x002060, 0x0080}, +{0x002065, 0x0001}, +{0x002066, 0x0080}, {0x002070, 0x0002}, {0x002071, 0x0004}, -{0x002072, 0x0080}, +{0x002072, 0x0001}, {0x002074, 0x0002}, {0x00207A, 0x0040}, {0x00207D, 0x0020}, @@ -860,13 +866,13 @@ const std::vector> unicode_ranges_flags = { // st {0x002080, 0x0002}, {0x00208A, 0x0040}, {0x00208D, 0x0020}, -{0x00208F, 0x0080}, +{0x00208F, 0x0001}, {0x002090, 0x0004}, -{0x00209D, 0x0080}, +{0x00209D, 0x0001}, {0x0020A0, 0x0040}, -{0x0020C1, 0x0080}, +{0x0020C1, 0x0001}, {0x0020D0, 0x0010}, -{0x0020F1, 0x0080}, +{0x0020F1, 0x0001}, {0x002100, 0x0040}, {0x002102, 0x0004}, {0x002103, 0x0040}, @@ -898,15 +904,15 @@ const std::vector> unicode_ranges_flags = { // st {0x002183, 0x0004}, {0x002185, 0x0002}, {0x00218A, 0x0040}, -{0x00218C, 0x0080}, +{0x00218C, 0x0001}, {0x002190, 0x0040}, {0x002308, 0x0020}, {0x00230C, 0x0040}, {0x002329, 0x0020}, {0x00232B, 0x0040}, -{0x002427, 0x0080}, +{0x002427, 0x0001}, {0x002440, 0x0040}, -{0x00244B, 0x0080}, +{0x00244B, 0x0001}, {0x002460, 0x0002}, {0x00249C, 0x0040}, {0x0024EA, 0x0002}, @@ -924,62 +930,62 @@ const std::vector> unicode_ranges_flags = { // st {0x0029DC, 0x0040}, {0x0029FC, 0x0020}, {0x0029FE, 0x0040}, -{0x002B74, 0x0080}, +{0x002B74, 0x0001}, {0x002B76, 0x0040}, -{0x002B96, 0x0080}, +{0x002B96, 0x0001}, {0x002B97, 0x0040}, {0x002C00, 0x0004}, {0x002CE5, 0x0040}, {0x002CEB, 0x0004}, {0x002CEF, 0x0010}, {0x002CF2, 0x0004}, -{0x002CF4, 0x0080}, +{0x002CF4, 0x0001}, {0x002CF9, 0x0020}, {0x002CFD, 0x0002}, {0x002CFE, 0x0020}, {0x002D00, 0x0004}, -{0x002D26, 0x0080}, +{0x002D26, 0x0001}, {0x002D27, 0x0004}, -{0x002D28, 0x0080}, +{0x002D28, 0x0001}, {0x002D2D, 0x0004}, -{0x002D2E, 0x0080}, +{0x002D2E, 0x0001}, {0x002D30, 0x0004}, -{0x002D68, 0x0080}, +{0x002D68, 0x0001}, {0x002D6F, 0x0004}, {0x002D70, 0x0020}, -{0x002D71, 0x0080}, +{0x002D71, 0x0001}, {0x002D7F, 0x0010}, {0x002D80, 0x0004}, -{0x002D97, 0x0080}, +{0x002D97, 0x0001}, {0x002DA0, 0x0004}, -{0x002DA7, 0x0080}, +{0x002DA7, 0x0001}, {0x002DA8, 0x0004}, -{0x002DAF, 0x0080}, +{0x002DAF, 0x0001}, {0x002DB0, 0x0004}, -{0x002DB7, 0x0080}, +{0x002DB7, 0x0001}, {0x002DB8, 0x0004}, -{0x002DBF, 0x0080}, +{0x002DBF, 0x0001}, {0x002DC0, 0x0004}, -{0x002DC7, 0x0080}, +{0x002DC7, 0x0001}, {0x002DC8, 0x0004}, -{0x002DCF, 0x0080}, +{0x002DCF, 0x0001}, {0x002DD0, 0x0004}, -{0x002DD7, 0x0080}, +{0x002DD7, 0x0001}, {0x002DD8, 0x0004}, -{0x002DDF, 0x0080}, +{0x002DDF, 0x0001}, {0x002DE0, 0x0010}, {0x002E00, 0x0020}, {0x002E2F, 0x0004}, {0x002E30, 0x0020}, {0x002E50, 0x0040}, {0x002E52, 0x0020}, -{0x002E5E, 0x0080}, +{0x002E5E, 0x0001}, {0x002E80, 0x0040}, -{0x002E9A, 0x0080}, +{0x002E9A, 0x0001}, {0x002E9B, 0x0040}, -{0x002EF4, 0x0080}, +{0x002EF4, 0x0001}, {0x002F00, 0x0040}, -{0x002FD6, 0x0080}, +{0x002FD6, 0x0001}, {0x002FF0, 0x0040}, {0x003000, 0x0008}, {0x003001, 0x0020}, @@ -999,9 +1005,9 @@ const std::vector> unicode_ranges_flags = { // st {0x00303B, 0x0004}, {0x00303D, 0x0020}, {0x00303E, 0x0040}, -{0x003040, 0x0080}, +{0x003040, 0x0001}, {0x003041, 0x0004}, -{0x003097, 0x0080}, +{0x003097, 0x0001}, {0x003099, 0x0010}, {0x00309B, 0x0040}, {0x00309D, 0x0004}, @@ -1009,21 +1015,21 @@ const std::vector> unicode_ranges_flags = { // st {0x0030A1, 0x0004}, {0x0030FB, 0x0020}, {0x0030FC, 0x0004}, -{0x003100, 0x0080}, +{0x003100, 0x0001}, {0x003105, 0x0004}, -{0x003130, 0x0080}, +{0x003130, 0x0001}, {0x003131, 0x0004}, -{0x00318F, 0x0080}, +{0x00318F, 0x0001}, {0x003190, 0x0040}, {0x003192, 0x0002}, {0x003196, 0x0040}, {0x0031A0, 0x0004}, {0x0031C0, 0x0040}, -{0x0031E4, 0x0080}, +{0x0031E4, 0x0001}, {0x0031EF, 0x0040}, {0x0031F0, 0x0004}, {0x003200, 0x0040}, -{0x00321F, 0x0080}, +{0x00321F, 0x0001}, {0x003220, 0x0002}, {0x00322A, 0x0040}, {0x003248, 0x0002}, @@ -1037,9 +1043,9 @@ const std::vector> unicode_ranges_flags = { // st {0x003400, 0x0004}, {0x004DC0, 0x0040}, {0x004E00, 0x0004}, -{0x00A48D, 0x0080}, +{0x00A48D, 0x0001}, {0x00A490, 0x0040}, -{0x00A4C7, 0x0080}, +{0x00A4C7, 0x0001}, {0x00A4D0, 0x0004}, {0x00A4FE, 0x0020}, {0x00A500, 0x0004}, @@ -1047,7 +1053,7 @@ const std::vector> unicode_ranges_flags = { // st {0x00A610, 0x0004}, {0x00A620, 0x0002}, {0x00A62A, 0x0004}, -{0x00A62C, 0x0080}, +{0x00A62C, 0x0001}, {0x00A640, 0x0004}, {0x00A66F, 0x0010}, {0x00A673, 0x0020}, @@ -1059,20 +1065,20 @@ const std::vector> unicode_ranges_flags = { // st {0x00A6E6, 0x0002}, {0x00A6F0, 0x0010}, {0x00A6F2, 0x0020}, -{0x00A6F8, 0x0080}, +{0x00A6F8, 0x0001}, {0x00A700, 0x0040}, {0x00A717, 0x0004}, {0x00A720, 0x0040}, {0x00A722, 0x0004}, {0x00A789, 0x0040}, {0x00A78B, 0x0004}, -{0x00A7CB, 0x0080}, +{0x00A7CB, 0x0001}, {0x00A7D0, 0x0004}, -{0x00A7D2, 0x0080}, +{0x00A7D2, 0x0001}, {0x00A7D3, 0x0004}, -{0x00A7D4, 0x0080}, +{0x00A7D4, 0x0001}, {0x00A7D5, 0x0004}, -{0x00A7DA, 0x0080}, +{0x00A7DA, 0x0001}, {0x00A7F2, 0x0004}, {0x00A802, 0x0010}, {0x00A803, 0x0004}, @@ -1083,20 +1089,20 @@ const std::vector> unicode_ranges_flags = { // st {0x00A823, 0x0010}, {0x00A828, 0x0040}, {0x00A82C, 0x0010}, -{0x00A82D, 0x0080}, +{0x00A82D, 0x0001}, {0x00A830, 0x0002}, {0x00A836, 0x0040}, -{0x00A83A, 0x0080}, +{0x00A83A, 0x0001}, {0x00A840, 0x0004}, {0x00A874, 0x0020}, -{0x00A878, 0x0080}, +{0x00A878, 0x0001}, {0x00A880, 0x0010}, {0x00A882, 0x0004}, {0x00A8B4, 0x0010}, -{0x00A8C6, 0x0080}, +{0x00A8C6, 0x0001}, {0x00A8CE, 0x0020}, {0x00A8D0, 0x0002}, -{0x00A8DA, 0x0080}, +{0x00A8DA, 0x0001}, {0x00A8E0, 0x0010}, {0x00A8F2, 0x0004}, {0x00A8F8, 0x0020}, @@ -1110,35 +1116,35 @@ const std::vector> unicode_ranges_flags = { // st {0x00A92E, 0x0020}, {0x00A930, 0x0004}, {0x00A947, 0x0010}, -{0x00A954, 0x0080}, +{0x00A954, 0x0001}, {0x00A95F, 0x0020}, {0x00A960, 0x0004}, -{0x00A97D, 0x0080}, +{0x00A97D, 0x0001}, {0x00A980, 0x0010}, {0x00A984, 0x0004}, {0x00A9B3, 0x0010}, {0x00A9C1, 0x0020}, -{0x00A9CE, 0x0080}, +{0x00A9CE, 0x0001}, {0x00A9CF, 0x0004}, {0x00A9D0, 0x0002}, -{0x00A9DA, 0x0080}, +{0x00A9DA, 0x0001}, {0x00A9DE, 0x0020}, {0x00A9E0, 0x0004}, {0x00A9E5, 0x0010}, {0x00A9E6, 0x0004}, {0x00A9F0, 0x0002}, {0x00A9FA, 0x0004}, -{0x00A9FF, 0x0080}, +{0x00A9FF, 0x0001}, {0x00AA00, 0x0004}, {0x00AA29, 0x0010}, -{0x00AA37, 0x0080}, +{0x00AA37, 0x0001}, {0x00AA40, 0x0004}, {0x00AA43, 0x0010}, {0x00AA44, 0x0004}, {0x00AA4C, 0x0010}, -{0x00AA4E, 0x0080}, +{0x00AA4E, 0x0001}, {0x00AA50, 0x0002}, -{0x00AA5A, 0x0080}, +{0x00AA5A, 0x0001}, {0x00AA5C, 0x0020}, {0x00AA60, 0x0004}, {0x00AA77, 0x0040}, @@ -1155,7 +1161,7 @@ const std::vector> unicode_ranges_flags = { // st {0x00AAC0, 0x0004}, {0x00AAC1, 0x0010}, {0x00AAC2, 0x0004}, -{0x00AAC3, 0x0080}, +{0x00AAC3, 0x0001}, {0x00AADB, 0x0004}, {0x00AADE, 0x0020}, {0x00AAE0, 0x0004}, @@ -1163,90 +1169,93 @@ const std::vector> unicode_ranges_flags = { // st {0x00AAF0, 0x0020}, {0x00AAF2, 0x0004}, {0x00AAF5, 0x0010}, -{0x00AAF7, 0x0080}, +{0x00AAF7, 0x0001}, {0x00AB01, 0x0004}, -{0x00AB07, 0x0080}, +{0x00AB07, 0x0001}, {0x00AB09, 0x0004}, -{0x00AB0F, 0x0080}, +{0x00AB0F, 0x0001}, {0x00AB11, 0x0004}, -{0x00AB17, 0x0080}, +{0x00AB17, 0x0001}, {0x00AB20, 0x0004}, -{0x00AB27, 0x0080}, +{0x00AB27, 0x0001}, {0x00AB28, 0x0004}, -{0x00AB2F, 0x0080}, +{0x00AB2F, 0x0001}, {0x00AB30, 0x0004}, {0x00AB5B, 0x0040}, {0x00AB5C, 0x0004}, {0x00AB6A, 0x0040}, -{0x00AB6C, 0x0080}, +{0x00AB6C, 0x0001}, {0x00AB70, 0x0004}, {0x00ABE3, 0x0010}, {0x00ABEB, 0x0020}, {0x00ABEC, 0x0010}, -{0x00ABEE, 0x0080}, +{0x00ABEE, 0x0001}, {0x00ABF0, 0x0002}, -{0x00ABFA, 0x0080}, +{0x00ABFA, 0x0001}, {0x00AC00, 0x0004}, -{0x00D7A4, 0x0080}, +{0x00D7A4, 0x0001}, {0x00D7B0, 0x0004}, -{0x00D7C7, 0x0080}, +{0x00D7C7, 0x0001}, {0x00D7CB, 0x0004}, -{0x00D7FC, 0x0080}, +{0x00D7FC, 0x0001}, +{0x00D800, 0x0080}, {0x00F900, 0x0004}, -{0x00FA6E, 0x0080}, +{0x00FA6E, 0x0001}, {0x00FA70, 0x0004}, -{0x00FADA, 0x0080}, +{0x00FADA, 0x0001}, {0x00FB00, 0x0004}, -{0x00FB07, 0x0080}, +{0x00FB07, 0x0001}, {0x00FB13, 0x0004}, -{0x00FB18, 0x0080}, +{0x00FB18, 0x0001}, {0x00FB1D, 0x0004}, {0x00FB1E, 0x0010}, {0x00FB1F, 0x0004}, {0x00FB29, 0x0040}, {0x00FB2A, 0x0004}, -{0x00FB37, 0x0080}, +{0x00FB37, 0x0001}, {0x00FB38, 0x0004}, -{0x00FB3D, 0x0080}, +{0x00FB3D, 0x0001}, {0x00FB3E, 0x0004}, -{0x00FB3F, 0x0080}, +{0x00FB3F, 0x0001}, {0x00FB40, 0x0004}, -{0x00FB42, 0x0080}, +{0x00FB42, 0x0001}, {0x00FB43, 0x0004}, -{0x00FB45, 0x0080}, +{0x00FB45, 0x0001}, {0x00FB46, 0x0004}, {0x00FBB2, 0x0040}, -{0x00FBC3, 0x0080}, +{0x00FBC3, 0x0001}, {0x00FBD3, 0x0004}, {0x00FD3E, 0x0020}, {0x00FD40, 0x0040}, {0x00FD50, 0x0004}, -{0x00FD90, 0x0080}, +{0x00FD90, 0x0001}, {0x00FD92, 0x0004}, -{0x00FDC8, 0x0080}, +{0x00FDC8, 0x0001}, {0x00FDCF, 0x0040}, -{0x00FDD0, 0x0080}, +{0x00FDD0, 0x0001}, {0x00FDF0, 0x0004}, {0x00FDFC, 0x0040}, {0x00FE00, 0x0010}, {0x00FE10, 0x0020}, -{0x00FE1A, 0x0080}, +{0x00FE1A, 0x0001}, {0x00FE20, 0x0010}, {0x00FE30, 0x0020}, -{0x00FE53, 0x0080}, +{0x00FE53, 0x0001}, {0x00FE54, 0x0020}, {0x00FE62, 0x0040}, {0x00FE63, 0x0020}, {0x00FE64, 0x0040}, -{0x00FE67, 0x0080}, +{0x00FE67, 0x0001}, {0x00FE68, 0x0020}, {0x00FE69, 0x0040}, {0x00FE6A, 0x0020}, -{0x00FE6C, 0x0080}, +{0x00FE6C, 0x0001}, {0x00FE70, 0x0004}, -{0x00FE75, 0x0080}, +{0x00FE75, 0x0001}, {0x00FE76, 0x0004}, -{0x00FEFD, 0x0080}, +{0x00FEFD, 0x0001}, +{0x00FEFF, 0x0080}, +{0x00FF00, 0x0001}, {0x00FF01, 0x0020}, {0x00FF04, 0x0040}, {0x00FF05, 0x0020}, @@ -1268,260 +1277,261 @@ const std::vector> unicode_ranges_flags = { // st {0x00FF5E, 0x0040}, {0x00FF5F, 0x0020}, {0x00FF66, 0x0004}, -{0x00FFBF, 0x0080}, +{0x00FFBF, 0x0001}, {0x00FFC2, 0x0004}, -{0x00FFC8, 0x0080}, +{0x00FFC8, 0x0001}, {0x00FFCA, 0x0004}, -{0x00FFD0, 0x0080}, +{0x00FFD0, 0x0001}, {0x00FFD2, 0x0004}, -{0x00FFD8, 0x0080}, +{0x00FFD8, 0x0001}, {0x00FFDA, 0x0004}, -{0x00FFDD, 0x0080}, +{0x00FFDD, 0x0001}, {0x00FFE0, 0x0040}, -{0x00FFE7, 0x0080}, +{0x00FFE7, 0x0001}, {0x00FFE8, 0x0040}, -{0x00FFEF, 0x0080}, +{0x00FFEF, 0x0001}, +{0x00FFF9, 0x0080}, {0x00FFFC, 0x0040}, -{0x00FFFE, 0x0080}, +{0x00FFFE, 0x0001}, {0x010000, 0x0004}, -{0x01000C, 0x0080}, +{0x01000C, 0x0001}, {0x01000D, 0x0004}, -{0x010027, 0x0080}, +{0x010027, 0x0001}, {0x010028, 0x0004}, -{0x01003B, 0x0080}, +{0x01003B, 0x0001}, {0x01003C, 0x0004}, -{0x01003E, 0x0080}, +{0x01003E, 0x0001}, {0x01003F, 0x0004}, -{0x01004E, 0x0080}, +{0x01004E, 0x0001}, {0x010050, 0x0004}, -{0x01005E, 0x0080}, +{0x01005E, 0x0001}, {0x010080, 0x0004}, -{0x0100FB, 0x0080}, +{0x0100FB, 0x0001}, {0x010100, 0x0020}, -{0x010103, 0x0080}, +{0x010103, 0x0001}, {0x010107, 0x0002}, -{0x010134, 0x0080}, +{0x010134, 0x0001}, {0x010137, 0x0040}, {0x010140, 0x0002}, {0x010179, 0x0040}, {0x01018A, 0x0002}, {0x01018C, 0x0040}, -{0x01018F, 0x0080}, +{0x01018F, 0x0001}, {0x010190, 0x0040}, -{0x01019D, 0x0080}, +{0x01019D, 0x0001}, {0x0101A0, 0x0040}, -{0x0101A1, 0x0080}, +{0x0101A1, 0x0001}, {0x0101D0, 0x0040}, {0x0101FD, 0x0010}, -{0x0101FE, 0x0080}, +{0x0101FE, 0x0001}, {0x010280, 0x0004}, -{0x01029D, 0x0080}, +{0x01029D, 0x0001}, {0x0102A0, 0x0004}, -{0x0102D1, 0x0080}, +{0x0102D1, 0x0001}, {0x0102E0, 0x0010}, {0x0102E1, 0x0002}, -{0x0102FC, 0x0080}, +{0x0102FC, 0x0001}, {0x010300, 0x0004}, {0x010320, 0x0002}, -{0x010324, 0x0080}, +{0x010324, 0x0001}, {0x01032D, 0x0004}, {0x010341, 0x0002}, {0x010342, 0x0004}, {0x01034A, 0x0002}, -{0x01034B, 0x0080}, +{0x01034B, 0x0001}, {0x010350, 0x0004}, {0x010376, 0x0010}, -{0x01037B, 0x0080}, +{0x01037B, 0x0001}, {0x010380, 0x0004}, -{0x01039E, 0x0080}, +{0x01039E, 0x0001}, {0x01039F, 0x0020}, {0x0103A0, 0x0004}, -{0x0103C4, 0x0080}, +{0x0103C4, 0x0001}, {0x0103C8, 0x0004}, {0x0103D0, 0x0020}, {0x0103D1, 0x0002}, -{0x0103D6, 0x0080}, +{0x0103D6, 0x0001}, {0x010400, 0x0004}, -{0x01049E, 0x0080}, +{0x01049E, 0x0001}, {0x0104A0, 0x0002}, -{0x0104AA, 0x0080}, +{0x0104AA, 0x0001}, {0x0104B0, 0x0004}, -{0x0104D4, 0x0080}, +{0x0104D4, 0x0001}, {0x0104D8, 0x0004}, -{0x0104FC, 0x0080}, +{0x0104FC, 0x0001}, {0x010500, 0x0004}, -{0x010528, 0x0080}, +{0x010528, 0x0001}, {0x010530, 0x0004}, -{0x010564, 0x0080}, +{0x010564, 0x0001}, {0x01056F, 0x0020}, {0x010570, 0x0004}, -{0x01057B, 0x0080}, +{0x01057B, 0x0001}, {0x01057C, 0x0004}, -{0x01058B, 0x0080}, +{0x01058B, 0x0001}, {0x01058C, 0x0004}, -{0x010593, 0x0080}, +{0x010593, 0x0001}, {0x010594, 0x0004}, -{0x010596, 0x0080}, +{0x010596, 0x0001}, {0x010597, 0x0004}, -{0x0105A2, 0x0080}, +{0x0105A2, 0x0001}, {0x0105A3, 0x0004}, -{0x0105B2, 0x0080}, +{0x0105B2, 0x0001}, {0x0105B3, 0x0004}, -{0x0105BA, 0x0080}, +{0x0105BA, 0x0001}, {0x0105BB, 0x0004}, -{0x0105BD, 0x0080}, +{0x0105BD, 0x0001}, {0x010600, 0x0004}, -{0x010737, 0x0080}, +{0x010737, 0x0001}, {0x010740, 0x0004}, -{0x010756, 0x0080}, +{0x010756, 0x0001}, {0x010760, 0x0004}, -{0x010768, 0x0080}, +{0x010768, 0x0001}, {0x010780, 0x0004}, -{0x010786, 0x0080}, +{0x010786, 0x0001}, {0x010787, 0x0004}, -{0x0107B1, 0x0080}, +{0x0107B1, 0x0001}, {0x0107B2, 0x0004}, -{0x0107BB, 0x0080}, +{0x0107BB, 0x0001}, {0x010800, 0x0004}, -{0x010806, 0x0080}, +{0x010806, 0x0001}, {0x010808, 0x0004}, -{0x010809, 0x0080}, +{0x010809, 0x0001}, {0x01080A, 0x0004}, -{0x010836, 0x0080}, +{0x010836, 0x0001}, {0x010837, 0x0004}, -{0x010839, 0x0080}, +{0x010839, 0x0001}, {0x01083C, 0x0004}, -{0x01083D, 0x0080}, +{0x01083D, 0x0001}, {0x01083F, 0x0004}, -{0x010856, 0x0080}, +{0x010856, 0x0001}, {0x010857, 0x0020}, {0x010858, 0x0002}, {0x010860, 0x0004}, {0x010877, 0x0040}, {0x010879, 0x0002}, {0x010880, 0x0004}, -{0x01089F, 0x0080}, +{0x01089F, 0x0001}, {0x0108A7, 0x0002}, -{0x0108B0, 0x0080}, +{0x0108B0, 0x0001}, {0x0108E0, 0x0004}, -{0x0108F3, 0x0080}, +{0x0108F3, 0x0001}, {0x0108F4, 0x0004}, -{0x0108F6, 0x0080}, +{0x0108F6, 0x0001}, {0x0108FB, 0x0002}, {0x010900, 0x0004}, {0x010916, 0x0002}, -{0x01091C, 0x0080}, +{0x01091C, 0x0001}, {0x01091F, 0x0020}, {0x010920, 0x0004}, -{0x01093A, 0x0080}, +{0x01093A, 0x0001}, {0x01093F, 0x0020}, -{0x010940, 0x0080}, +{0x010940, 0x0001}, {0x010980, 0x0004}, -{0x0109B8, 0x0080}, +{0x0109B8, 0x0001}, {0x0109BC, 0x0002}, {0x0109BE, 0x0004}, {0x0109C0, 0x0002}, -{0x0109D0, 0x0080}, +{0x0109D0, 0x0001}, {0x0109D2, 0x0002}, {0x010A00, 0x0004}, {0x010A01, 0x0010}, -{0x010A04, 0x0080}, +{0x010A04, 0x0001}, {0x010A05, 0x0010}, -{0x010A07, 0x0080}, +{0x010A07, 0x0001}, {0x010A0C, 0x0010}, {0x010A10, 0x0004}, -{0x010A14, 0x0080}, +{0x010A14, 0x0001}, {0x010A15, 0x0004}, -{0x010A18, 0x0080}, +{0x010A18, 0x0001}, {0x010A19, 0x0004}, -{0x010A36, 0x0080}, +{0x010A36, 0x0001}, {0x010A38, 0x0010}, -{0x010A3B, 0x0080}, +{0x010A3B, 0x0001}, {0x010A3F, 0x0010}, {0x010A40, 0x0002}, -{0x010A49, 0x0080}, +{0x010A49, 0x0001}, {0x010A50, 0x0020}, -{0x010A59, 0x0080}, +{0x010A59, 0x0001}, {0x010A60, 0x0004}, {0x010A7D, 0x0002}, {0x010A7F, 0x0020}, {0x010A80, 0x0004}, {0x010A9D, 0x0002}, -{0x010AA0, 0x0080}, +{0x010AA0, 0x0001}, {0x010AC0, 0x0004}, {0x010AC8, 0x0040}, {0x010AC9, 0x0004}, {0x010AE5, 0x0010}, -{0x010AE7, 0x0080}, +{0x010AE7, 0x0001}, {0x010AEB, 0x0002}, {0x010AF0, 0x0020}, -{0x010AF7, 0x0080}, +{0x010AF7, 0x0001}, {0x010B00, 0x0004}, -{0x010B36, 0x0080}, +{0x010B36, 0x0001}, {0x010B39, 0x0020}, {0x010B40, 0x0004}, -{0x010B56, 0x0080}, +{0x010B56, 0x0001}, {0x010B58, 0x0002}, {0x010B60, 0x0004}, -{0x010B73, 0x0080}, +{0x010B73, 0x0001}, {0x010B78, 0x0002}, {0x010B80, 0x0004}, -{0x010B92, 0x0080}, +{0x010B92, 0x0001}, {0x010B99, 0x0020}, -{0x010B9D, 0x0080}, +{0x010B9D, 0x0001}, {0x010BA9, 0x0002}, -{0x010BB0, 0x0080}, +{0x010BB0, 0x0001}, {0x010C00, 0x0004}, -{0x010C49, 0x0080}, +{0x010C49, 0x0001}, {0x010C80, 0x0004}, -{0x010CB3, 0x0080}, +{0x010CB3, 0x0001}, {0x010CC0, 0x0004}, -{0x010CF3, 0x0080}, +{0x010CF3, 0x0001}, {0x010CFA, 0x0002}, {0x010D00, 0x0004}, {0x010D24, 0x0010}, -{0x010D28, 0x0080}, +{0x010D28, 0x0001}, {0x010D30, 0x0002}, -{0x010D3A, 0x0080}, +{0x010D3A, 0x0001}, {0x010E60, 0x0002}, -{0x010E7F, 0x0080}, +{0x010E7F, 0x0001}, {0x010E80, 0x0004}, -{0x010EAA, 0x0080}, +{0x010EAA, 0x0001}, {0x010EAB, 0x0010}, {0x010EAD, 0x0020}, -{0x010EAE, 0x0080}, +{0x010EAE, 0x0001}, {0x010EB0, 0x0004}, -{0x010EB2, 0x0080}, +{0x010EB2, 0x0001}, {0x010EFD, 0x0010}, {0x010F00, 0x0004}, {0x010F1D, 0x0002}, {0x010F27, 0x0004}, -{0x010F28, 0x0080}, +{0x010F28, 0x0001}, {0x010F30, 0x0004}, {0x010F46, 0x0010}, {0x010F51, 0x0002}, {0x010F55, 0x0020}, -{0x010F5A, 0x0080}, +{0x010F5A, 0x0001}, {0x010F70, 0x0004}, {0x010F82, 0x0010}, {0x010F86, 0x0020}, -{0x010F8A, 0x0080}, +{0x010F8A, 0x0001}, {0x010FB0, 0x0004}, {0x010FC5, 0x0002}, -{0x010FCC, 0x0080}, +{0x010FCC, 0x0001}, {0x010FE0, 0x0004}, -{0x010FF7, 0x0080}, +{0x010FF7, 0x0001}, {0x011000, 0x0010}, {0x011003, 0x0004}, {0x011038, 0x0010}, {0x011047, 0x0020}, -{0x01104E, 0x0080}, +{0x01104E, 0x0001}, {0x011052, 0x0002}, {0x011070, 0x0010}, {0x011071, 0x0004}, {0x011073, 0x0010}, {0x011075, 0x0004}, -{0x011076, 0x0080}, +{0x011076, 0x0001}, {0x01107F, 0x0010}, {0x011083, 0x0004}, {0x0110B0, 0x0010}, @@ -1529,26 +1539,28 @@ const std::vector> unicode_ranges_flags = { // st {0x0110BD, 0x0080}, {0x0110BE, 0x0020}, {0x0110C2, 0x0010}, -{0x0110C3, 0x0080}, +{0x0110C3, 0x0001}, +{0x0110CD, 0x0080}, +{0x0110CE, 0x0001}, {0x0110D0, 0x0004}, -{0x0110E9, 0x0080}, +{0x0110E9, 0x0001}, {0x0110F0, 0x0002}, -{0x0110FA, 0x0080}, +{0x0110FA, 0x0001}, {0x011100, 0x0010}, {0x011103, 0x0004}, {0x011127, 0x0010}, -{0x011135, 0x0080}, +{0x011135, 0x0001}, {0x011136, 0x0002}, {0x011140, 0x0020}, {0x011144, 0x0004}, {0x011145, 0x0010}, {0x011147, 0x0004}, -{0x011148, 0x0080}, +{0x011148, 0x0001}, {0x011150, 0x0004}, {0x011173, 0x0010}, {0x011174, 0x0020}, {0x011176, 0x0004}, -{0x011177, 0x0080}, +{0x011177, 0x0001}, {0x011180, 0x0010}, {0x011183, 0x0004}, {0x0111B3, 0x0010}, @@ -1562,159 +1574,159 @@ const std::vector> unicode_ranges_flags = { // st {0x0111DB, 0x0020}, {0x0111DC, 0x0004}, {0x0111DD, 0x0020}, -{0x0111E0, 0x0080}, +{0x0111E0, 0x0001}, {0x0111E1, 0x0002}, -{0x0111F5, 0x0080}, +{0x0111F5, 0x0001}, {0x011200, 0x0004}, -{0x011212, 0x0080}, +{0x011212, 0x0001}, {0x011213, 0x0004}, {0x01122C, 0x0010}, {0x011238, 0x0020}, {0x01123E, 0x0010}, {0x01123F, 0x0004}, {0x011241, 0x0010}, -{0x011242, 0x0080}, +{0x011242, 0x0001}, {0x011280, 0x0004}, -{0x011287, 0x0080}, +{0x011287, 0x0001}, {0x011288, 0x0004}, -{0x011289, 0x0080}, +{0x011289, 0x0001}, {0x01128A, 0x0004}, -{0x01128E, 0x0080}, +{0x01128E, 0x0001}, {0x01128F, 0x0004}, -{0x01129E, 0x0080}, +{0x01129E, 0x0001}, {0x01129F, 0x0004}, {0x0112A9, 0x0020}, -{0x0112AA, 0x0080}, +{0x0112AA, 0x0001}, {0x0112B0, 0x0004}, {0x0112DF, 0x0010}, -{0x0112EB, 0x0080}, +{0x0112EB, 0x0001}, {0x0112F0, 0x0002}, -{0x0112FA, 0x0080}, +{0x0112FA, 0x0001}, {0x011300, 0x0010}, -{0x011304, 0x0080}, +{0x011304, 0x0001}, {0x011305, 0x0004}, -{0x01130D, 0x0080}, +{0x01130D, 0x0001}, {0x01130F, 0x0004}, -{0x011311, 0x0080}, +{0x011311, 0x0001}, {0x011313, 0x0004}, -{0x011329, 0x0080}, +{0x011329, 0x0001}, {0x01132A, 0x0004}, -{0x011331, 0x0080}, +{0x011331, 0x0001}, {0x011332, 0x0004}, -{0x011334, 0x0080}, +{0x011334, 0x0001}, {0x011335, 0x0004}, -{0x01133A, 0x0080}, +{0x01133A, 0x0001}, {0x01133B, 0x0010}, {0x01133D, 0x0004}, {0x01133E, 0x0010}, -{0x011345, 0x0080}, +{0x011345, 0x0001}, {0x011347, 0x0010}, -{0x011349, 0x0080}, +{0x011349, 0x0001}, {0x01134B, 0x0010}, -{0x01134E, 0x0080}, +{0x01134E, 0x0001}, {0x011350, 0x0004}, -{0x011351, 0x0080}, +{0x011351, 0x0001}, {0x011357, 0x0010}, -{0x011358, 0x0080}, +{0x011358, 0x0001}, {0x01135D, 0x0004}, {0x011362, 0x0010}, -{0x011364, 0x0080}, +{0x011364, 0x0001}, {0x011366, 0x0010}, -{0x01136D, 0x0080}, +{0x01136D, 0x0001}, {0x011370, 0x0010}, -{0x011375, 0x0080}, +{0x011375, 0x0001}, {0x011400, 0x0004}, {0x011435, 0x0010}, {0x011447, 0x0004}, {0x01144B, 0x0020}, {0x011450, 0x0002}, {0x01145A, 0x0020}, -{0x01145C, 0x0080}, +{0x01145C, 0x0001}, {0x01145D, 0x0020}, {0x01145E, 0x0010}, {0x01145F, 0x0004}, -{0x011462, 0x0080}, +{0x011462, 0x0001}, {0x011480, 0x0004}, {0x0114B0, 0x0010}, {0x0114C4, 0x0004}, {0x0114C6, 0x0020}, {0x0114C7, 0x0004}, -{0x0114C8, 0x0080}, +{0x0114C8, 0x0001}, {0x0114D0, 0x0002}, -{0x0114DA, 0x0080}, +{0x0114DA, 0x0001}, {0x011580, 0x0004}, {0x0115AF, 0x0010}, -{0x0115B6, 0x0080}, +{0x0115B6, 0x0001}, {0x0115B8, 0x0010}, {0x0115C1, 0x0020}, {0x0115D8, 0x0004}, {0x0115DC, 0x0010}, -{0x0115DE, 0x0080}, +{0x0115DE, 0x0001}, {0x011600, 0x0004}, {0x011630, 0x0010}, {0x011641, 0x0020}, {0x011644, 0x0004}, -{0x011645, 0x0080}, +{0x011645, 0x0001}, {0x011650, 0x0002}, -{0x01165A, 0x0080}, +{0x01165A, 0x0001}, {0x011660, 0x0020}, -{0x01166D, 0x0080}, +{0x01166D, 0x0001}, {0x011680, 0x0004}, {0x0116AB, 0x0010}, {0x0116B8, 0x0004}, {0x0116B9, 0x0020}, -{0x0116BA, 0x0080}, +{0x0116BA, 0x0001}, {0x0116C0, 0x0002}, -{0x0116CA, 0x0080}, +{0x0116CA, 0x0001}, {0x011700, 0x0004}, -{0x01171B, 0x0080}, +{0x01171B, 0x0001}, {0x01171D, 0x0010}, -{0x01172C, 0x0080}, +{0x01172C, 0x0001}, {0x011730, 0x0002}, {0x01173C, 0x0020}, {0x01173F, 0x0040}, {0x011740, 0x0004}, -{0x011747, 0x0080}, +{0x011747, 0x0001}, {0x011800, 0x0004}, {0x01182C, 0x0010}, {0x01183B, 0x0020}, -{0x01183C, 0x0080}, +{0x01183C, 0x0001}, {0x0118A0, 0x0004}, {0x0118E0, 0x0002}, -{0x0118F3, 0x0080}, +{0x0118F3, 0x0001}, {0x0118FF, 0x0004}, -{0x011907, 0x0080}, +{0x011907, 0x0001}, {0x011909, 0x0004}, -{0x01190A, 0x0080}, +{0x01190A, 0x0001}, {0x01190C, 0x0004}, -{0x011914, 0x0080}, +{0x011914, 0x0001}, {0x011915, 0x0004}, -{0x011917, 0x0080}, +{0x011917, 0x0001}, {0x011918, 0x0004}, {0x011930, 0x0010}, -{0x011936, 0x0080}, +{0x011936, 0x0001}, {0x011937, 0x0010}, -{0x011939, 0x0080}, +{0x011939, 0x0001}, {0x01193B, 0x0010}, {0x01193F, 0x0004}, {0x011940, 0x0010}, {0x011941, 0x0004}, {0x011942, 0x0010}, {0x011944, 0x0020}, -{0x011947, 0x0080}, +{0x011947, 0x0001}, {0x011950, 0x0002}, -{0x01195A, 0x0080}, +{0x01195A, 0x0001}, {0x0119A0, 0x0004}, -{0x0119A8, 0x0080}, +{0x0119A8, 0x0001}, {0x0119AA, 0x0004}, {0x0119D1, 0x0010}, -{0x0119D8, 0x0080}, +{0x0119D8, 0x0001}, {0x0119DA, 0x0010}, {0x0119E1, 0x0004}, {0x0119E2, 0x0020}, {0x0119E3, 0x0004}, {0x0119E4, 0x0010}, -{0x0119E5, 0x0080}, +{0x0119E5, 0x0001}, {0x011A00, 0x0004}, {0x011A01, 0x0010}, {0x011A0B, 0x0004}, @@ -1723,7 +1735,7 @@ const std::vector> unicode_ranges_flags = { // st {0x011A3B, 0x0010}, {0x011A3F, 0x0020}, {0x011A47, 0x0010}, -{0x011A48, 0x0080}, +{0x011A48, 0x0001}, {0x011A50, 0x0004}, {0x011A51, 0x0010}, {0x011A5C, 0x0004}, @@ -1731,117 +1743,117 @@ const std::vector> unicode_ranges_flags = { // st {0x011A9A, 0x0020}, {0x011A9D, 0x0004}, {0x011A9E, 0x0020}, -{0x011AA3, 0x0080}, +{0x011AA3, 0x0001}, {0x011AB0, 0x0004}, -{0x011AF9, 0x0080}, +{0x011AF9, 0x0001}, {0x011B00, 0x0020}, -{0x011B0A, 0x0080}, +{0x011B0A, 0x0001}, {0x011C00, 0x0004}, -{0x011C09, 0x0080}, +{0x011C09, 0x0001}, {0x011C0A, 0x0004}, {0x011C2F, 0x0010}, -{0x011C37, 0x0080}, +{0x011C37, 0x0001}, {0x011C38, 0x0010}, {0x011C40, 0x0004}, {0x011C41, 0x0020}, -{0x011C46, 0x0080}, +{0x011C46, 0x0001}, {0x011C50, 0x0002}, -{0x011C6D, 0x0080}, +{0x011C6D, 0x0001}, {0x011C70, 0x0020}, {0x011C72, 0x0004}, -{0x011C90, 0x0080}, +{0x011C90, 0x0001}, {0x011C92, 0x0010}, -{0x011CA8, 0x0080}, +{0x011CA8, 0x0001}, {0x011CA9, 0x0010}, -{0x011CB7, 0x0080}, +{0x011CB7, 0x0001}, {0x011D00, 0x0004}, -{0x011D07, 0x0080}, +{0x011D07, 0x0001}, {0x011D08, 0x0004}, -{0x011D0A, 0x0080}, +{0x011D0A, 0x0001}, {0x011D0B, 0x0004}, {0x011D31, 0x0010}, -{0x011D37, 0x0080}, +{0x011D37, 0x0001}, {0x011D3A, 0x0010}, -{0x011D3B, 0x0080}, +{0x011D3B, 0x0001}, {0x011D3C, 0x0010}, -{0x011D3E, 0x0080}, +{0x011D3E, 0x0001}, {0x011D3F, 0x0010}, {0x011D46, 0x0004}, {0x011D47, 0x0010}, -{0x011D48, 0x0080}, +{0x011D48, 0x0001}, {0x011D50, 0x0002}, -{0x011D5A, 0x0080}, +{0x011D5A, 0x0001}, {0x011D60, 0x0004}, -{0x011D66, 0x0080}, +{0x011D66, 0x0001}, {0x011D67, 0x0004}, -{0x011D69, 0x0080}, +{0x011D69, 0x0001}, {0x011D6A, 0x0004}, {0x011D8A, 0x0010}, -{0x011D8F, 0x0080}, +{0x011D8F, 0x0001}, {0x011D90, 0x0010}, -{0x011D92, 0x0080}, +{0x011D92, 0x0001}, {0x011D93, 0x0010}, {0x011D98, 0x0004}, -{0x011D99, 0x0080}, +{0x011D99, 0x0001}, {0x011DA0, 0x0002}, -{0x011DAA, 0x0080}, +{0x011DAA, 0x0001}, {0x011EE0, 0x0004}, {0x011EF3, 0x0010}, {0x011EF7, 0x0020}, -{0x011EF9, 0x0080}, +{0x011EF9, 0x0001}, {0x011F00, 0x0010}, {0x011F02, 0x0004}, {0x011F03, 0x0010}, {0x011F04, 0x0004}, -{0x011F11, 0x0080}, +{0x011F11, 0x0001}, {0x011F12, 0x0004}, {0x011F34, 0x0010}, -{0x011F3B, 0x0080}, +{0x011F3B, 0x0001}, {0x011F3E, 0x0010}, {0x011F43, 0x0020}, {0x011F50, 0x0002}, -{0x011F5A, 0x0080}, +{0x011F5A, 0x0001}, {0x011FB0, 0x0004}, -{0x011FB1, 0x0080}, +{0x011FB1, 0x0001}, {0x011FC0, 0x0002}, {0x011FD5, 0x0040}, -{0x011FF2, 0x0080}, +{0x011FF2, 0x0001}, {0x011FFF, 0x0020}, {0x012000, 0x0004}, -{0x01239A, 0x0080}, +{0x01239A, 0x0001}, {0x012400, 0x0002}, -{0x01246F, 0x0080}, +{0x01246F, 0x0001}, {0x012470, 0x0020}, -{0x012475, 0x0080}, +{0x012475, 0x0001}, {0x012480, 0x0004}, -{0x012544, 0x0080}, +{0x012544, 0x0001}, {0x012F90, 0x0004}, {0x012FF1, 0x0020}, -{0x012FF3, 0x0080}, +{0x012FF3, 0x0001}, {0x013000, 0x0004}, {0x013430, 0x0080}, {0x013440, 0x0010}, {0x013441, 0x0004}, {0x013447, 0x0010}, -{0x013456, 0x0080}, +{0x013456, 0x0001}, {0x014400, 0x0004}, -{0x014647, 0x0080}, +{0x014647, 0x0001}, {0x016800, 0x0004}, -{0x016A39, 0x0080}, +{0x016A39, 0x0001}, {0x016A40, 0x0004}, -{0x016A5F, 0x0080}, +{0x016A5F, 0x0001}, {0x016A60, 0x0002}, -{0x016A6A, 0x0080}, +{0x016A6A, 0x0001}, {0x016A6E, 0x0020}, {0x016A70, 0x0004}, -{0x016ABF, 0x0080}, +{0x016ABF, 0x0001}, {0x016AC0, 0x0002}, -{0x016ACA, 0x0080}, +{0x016ACA, 0x0001}, {0x016AD0, 0x0004}, -{0x016AEE, 0x0080}, +{0x016AEE, 0x0001}, {0x016AF0, 0x0010}, {0x016AF5, 0x0020}, -{0x016AF6, 0x0080}, +{0x016AF6, 0x0001}, {0x016B00, 0x0004}, {0x016B30, 0x0010}, {0x016B37, 0x0020}, @@ -1849,81 +1861,82 @@ const std::vector> unicode_ranges_flags = { // st {0x016B40, 0x0004}, {0x016B44, 0x0020}, {0x016B45, 0x0040}, -{0x016B46, 0x0080}, +{0x016B46, 0x0001}, {0x016B50, 0x0002}, -{0x016B5A, 0x0080}, +{0x016B5A, 0x0001}, {0x016B5B, 0x0002}, -{0x016B62, 0x0080}, +{0x016B62, 0x0001}, {0x016B63, 0x0004}, -{0x016B78, 0x0080}, +{0x016B78, 0x0001}, {0x016B7D, 0x0004}, -{0x016B90, 0x0080}, +{0x016B90, 0x0001}, {0x016E40, 0x0004}, {0x016E80, 0x0002}, {0x016E97, 0x0020}, -{0x016E9B, 0x0080}, +{0x016E9B, 0x0001}, {0x016F00, 0x0004}, -{0x016F4B, 0x0080}, +{0x016F4B, 0x0001}, {0x016F4F, 0x0010}, {0x016F50, 0x0004}, {0x016F51, 0x0010}, -{0x016F88, 0x0080}, +{0x016F88, 0x0001}, {0x016F8F, 0x0010}, {0x016F93, 0x0004}, -{0x016FA0, 0x0080}, +{0x016FA0, 0x0001}, {0x016FE0, 0x0004}, {0x016FE2, 0x0020}, {0x016FE3, 0x0004}, {0x016FE4, 0x0010}, -{0x016FE5, 0x0080}, +{0x016FE5, 0x0001}, {0x016FF0, 0x0010}, -{0x016FF2, 0x0080}, +{0x016FF2, 0x0001}, {0x017000, 0x0004}, -{0x0187F8, 0x0080}, +{0x0187F8, 0x0001}, {0x018800, 0x0004}, -{0x018CD6, 0x0080}, +{0x018CD6, 0x0001}, {0x018D00, 0x0004}, -{0x018D09, 0x0080}, +{0x018D09, 0x0001}, {0x01AFF0, 0x0004}, -{0x01AFF4, 0x0080}, +{0x01AFF4, 0x0001}, {0x01AFF5, 0x0004}, -{0x01AFFC, 0x0080}, +{0x01AFFC, 0x0001}, {0x01AFFD, 0x0004}, -{0x01AFFF, 0x0080}, +{0x01AFFF, 0x0001}, {0x01B000, 0x0004}, -{0x01B123, 0x0080}, +{0x01B123, 0x0001}, {0x01B132, 0x0004}, -{0x01B133, 0x0080}, +{0x01B133, 0x0001}, {0x01B150, 0x0004}, -{0x01B153, 0x0080}, +{0x01B153, 0x0001}, {0x01B155, 0x0004}, -{0x01B156, 0x0080}, +{0x01B156, 0x0001}, {0x01B164, 0x0004}, -{0x01B168, 0x0080}, +{0x01B168, 0x0001}, {0x01B170, 0x0004}, -{0x01B2FC, 0x0080}, +{0x01B2FC, 0x0001}, {0x01BC00, 0x0004}, -{0x01BC6B, 0x0080}, +{0x01BC6B, 0x0001}, {0x01BC70, 0x0004}, -{0x01BC7D, 0x0080}, +{0x01BC7D, 0x0001}, {0x01BC80, 0x0004}, -{0x01BC89, 0x0080}, +{0x01BC89, 0x0001}, {0x01BC90, 0x0004}, -{0x01BC9A, 0x0080}, +{0x01BC9A, 0x0001}, {0x01BC9C, 0x0040}, {0x01BC9D, 0x0010}, {0x01BC9F, 0x0020}, {0x01BCA0, 0x0080}, +{0x01BCA4, 0x0001}, {0x01CF00, 0x0010}, -{0x01CF2E, 0x0080}, +{0x01CF2E, 0x0001}, {0x01CF30, 0x0010}, -{0x01CF47, 0x0080}, +{0x01CF47, 0x0001}, {0x01CF50, 0x0040}, -{0x01CFC4, 0x0080}, +{0x01CFC4, 0x0001}, {0x01D000, 0x0040}, -{0x01D0F6, 0x0080}, +{0x01D0F6, 0x0001}, {0x01D100, 0x0040}, -{0x01D127, 0x0080}, +{0x01D127, 0x0001}, {0x01D129, 0x0040}, {0x01D165, 0x0010}, {0x01D16A, 0x0040}, @@ -1935,57 +1948,57 @@ const std::vector> unicode_ranges_flags = { // st {0x01D18C, 0x0040}, {0x01D1AA, 0x0010}, {0x01D1AE, 0x0040}, -{0x01D1EB, 0x0080}, +{0x01D1EB, 0x0001}, {0x01D200, 0x0040}, {0x01D242, 0x0010}, {0x01D245, 0x0040}, -{0x01D246, 0x0080}, +{0x01D246, 0x0001}, {0x01D2C0, 0x0002}, -{0x01D2D4, 0x0080}, +{0x01D2D4, 0x0001}, {0x01D2E0, 0x0002}, -{0x01D2F4, 0x0080}, +{0x01D2F4, 0x0001}, {0x01D300, 0x0040}, -{0x01D357, 0x0080}, +{0x01D357, 0x0001}, {0x01D360, 0x0002}, -{0x01D379, 0x0080}, +{0x01D379, 0x0001}, {0x01D400, 0x0004}, -{0x01D455, 0x0080}, +{0x01D455, 0x0001}, {0x01D456, 0x0004}, -{0x01D49D, 0x0080}, +{0x01D49D, 0x0001}, {0x01D49E, 0x0004}, -{0x01D4A0, 0x0080}, +{0x01D4A0, 0x0001}, {0x01D4A2, 0x0004}, -{0x01D4A3, 0x0080}, +{0x01D4A3, 0x0001}, {0x01D4A5, 0x0004}, -{0x01D4A7, 0x0080}, +{0x01D4A7, 0x0001}, {0x01D4A9, 0x0004}, -{0x01D4AD, 0x0080}, +{0x01D4AD, 0x0001}, {0x01D4AE, 0x0004}, -{0x01D4BA, 0x0080}, +{0x01D4BA, 0x0001}, {0x01D4BB, 0x0004}, -{0x01D4BC, 0x0080}, +{0x01D4BC, 0x0001}, {0x01D4BD, 0x0004}, -{0x01D4C4, 0x0080}, +{0x01D4C4, 0x0001}, {0x01D4C5, 0x0004}, -{0x01D506, 0x0080}, +{0x01D506, 0x0001}, {0x01D507, 0x0004}, -{0x01D50B, 0x0080}, +{0x01D50B, 0x0001}, {0x01D50D, 0x0004}, -{0x01D515, 0x0080}, +{0x01D515, 0x0001}, {0x01D516, 0x0004}, -{0x01D51D, 0x0080}, +{0x01D51D, 0x0001}, {0x01D51E, 0x0004}, -{0x01D53A, 0x0080}, +{0x01D53A, 0x0001}, {0x01D53B, 0x0004}, -{0x01D53F, 0x0080}, +{0x01D53F, 0x0001}, {0x01D540, 0x0004}, -{0x01D545, 0x0080}, +{0x01D545, 0x0001}, {0x01D546, 0x0004}, -{0x01D547, 0x0080}, +{0x01D547, 0x0001}, {0x01D54A, 0x0004}, -{0x01D551, 0x0080}, +{0x01D551, 0x0001}, {0x01D552, 0x0004}, -{0x01D6A6, 0x0080}, +{0x01D6A6, 0x0001}, {0x01D6A8, 0x0004}, {0x01D6C1, 0x0040}, {0x01D6C2, 0x0004}, @@ -2007,7 +2020,7 @@ const std::vector> unicode_ranges_flags = { // st {0x01D7AA, 0x0004}, {0x01D7C3, 0x0040}, {0x01D7C4, 0x0004}, -{0x01D7CC, 0x0080}, +{0x01D7CC, 0x0001}, {0x01D7CE, 0x0002}, {0x01D800, 0x0040}, {0x01DA00, 0x0010}, @@ -2019,251 +2032,283 @@ const std::vector> unicode_ranges_flags = { // st {0x01DA84, 0x0010}, {0x01DA85, 0x0040}, {0x01DA87, 0x0020}, -{0x01DA8C, 0x0080}, +{0x01DA8C, 0x0001}, {0x01DA9B, 0x0010}, -{0x01DAA0, 0x0080}, +{0x01DAA0, 0x0001}, {0x01DAA1, 0x0010}, -{0x01DAB0, 0x0080}, +{0x01DAB0, 0x0001}, {0x01DF00, 0x0004}, -{0x01DF1F, 0x0080}, +{0x01DF1F, 0x0001}, {0x01DF25, 0x0004}, -{0x01DF2B, 0x0080}, +{0x01DF2B, 0x0001}, {0x01E000, 0x0010}, -{0x01E007, 0x0080}, +{0x01E007, 0x0001}, {0x01E008, 0x0010}, -{0x01E019, 0x0080}, +{0x01E019, 0x0001}, {0x01E01B, 0x0010}, -{0x01E022, 0x0080}, +{0x01E022, 0x0001}, {0x01E023, 0x0010}, -{0x01E025, 0x0080}, +{0x01E025, 0x0001}, {0x01E026, 0x0010}, -{0x01E02B, 0x0080}, +{0x01E02B, 0x0001}, {0x01E030, 0x0004}, -{0x01E06E, 0x0080}, +{0x01E06E, 0x0001}, {0x01E08F, 0x0010}, -{0x01E090, 0x0080}, +{0x01E090, 0x0001}, {0x01E100, 0x0004}, -{0x01E12D, 0x0080}, +{0x01E12D, 0x0001}, {0x01E130, 0x0010}, {0x01E137, 0x0004}, -{0x01E13E, 0x0080}, +{0x01E13E, 0x0001}, {0x01E140, 0x0002}, -{0x01E14A, 0x0080}, +{0x01E14A, 0x0001}, {0x01E14E, 0x0004}, {0x01E14F, 0x0040}, -{0x01E150, 0x0080}, +{0x01E150, 0x0001}, {0x01E290, 0x0004}, {0x01E2AE, 0x0010}, -{0x01E2AF, 0x0080}, +{0x01E2AF, 0x0001}, {0x01E2C0, 0x0004}, {0x01E2EC, 0x0010}, {0x01E2F0, 0x0002}, -{0x01E2FA, 0x0080}, +{0x01E2FA, 0x0001}, {0x01E2FF, 0x0040}, -{0x01E300, 0x0080}, +{0x01E300, 0x0001}, {0x01E4D0, 0x0004}, {0x01E4EC, 0x0010}, {0x01E4F0, 0x0002}, -{0x01E4FA, 0x0080}, +{0x01E4FA, 0x0001}, {0x01E7E0, 0x0004}, -{0x01E7E7, 0x0080}, +{0x01E7E7, 0x0001}, {0x01E7E8, 0x0004}, -{0x01E7EC, 0x0080}, +{0x01E7EC, 0x0001}, {0x01E7ED, 0x0004}, -{0x01E7EF, 0x0080}, +{0x01E7EF, 0x0001}, {0x01E7F0, 0x0004}, -{0x01E7FF, 0x0080}, +{0x01E7FF, 0x0001}, {0x01E800, 0x0004}, -{0x01E8C5, 0x0080}, +{0x01E8C5, 0x0001}, {0x01E8C7, 0x0002}, {0x01E8D0, 0x0010}, -{0x01E8D7, 0x0080}, +{0x01E8D7, 0x0001}, {0x01E900, 0x0004}, {0x01E944, 0x0010}, {0x01E94B, 0x0004}, -{0x01E94C, 0x0080}, +{0x01E94C, 0x0001}, {0x01E950, 0x0002}, -{0x01E95A, 0x0080}, +{0x01E95A, 0x0001}, {0x01E95E, 0x0020}, -{0x01E960, 0x0080}, +{0x01E960, 0x0001}, {0x01EC71, 0x0002}, {0x01ECAC, 0x0040}, {0x01ECAD, 0x0002}, {0x01ECB0, 0x0040}, {0x01ECB1, 0x0002}, -{0x01ECB5, 0x0080}, +{0x01ECB5, 0x0001}, {0x01ED01, 0x0002}, {0x01ED2E, 0x0040}, {0x01ED2F, 0x0002}, -{0x01ED3E, 0x0080}, +{0x01ED3E, 0x0001}, {0x01EE00, 0x0004}, -{0x01EE04, 0x0080}, +{0x01EE04, 0x0001}, {0x01EE05, 0x0004}, -{0x01EE20, 0x0080}, +{0x01EE20, 0x0001}, {0x01EE21, 0x0004}, -{0x01EE23, 0x0080}, +{0x01EE23, 0x0001}, {0x01EE24, 0x0004}, -{0x01EE25, 0x0080}, +{0x01EE25, 0x0001}, {0x01EE27, 0x0004}, -{0x01EE28, 0x0080}, +{0x01EE28, 0x0001}, {0x01EE29, 0x0004}, -{0x01EE33, 0x0080}, +{0x01EE33, 0x0001}, {0x01EE34, 0x0004}, -{0x01EE38, 0x0080}, +{0x01EE38, 0x0001}, {0x01EE39, 0x0004}, -{0x01EE3A, 0x0080}, +{0x01EE3A, 0x0001}, {0x01EE3B, 0x0004}, -{0x01EE3C, 0x0080}, +{0x01EE3C, 0x0001}, {0x01EE42, 0x0004}, -{0x01EE43, 0x0080}, +{0x01EE43, 0x0001}, {0x01EE47, 0x0004}, -{0x01EE48, 0x0080}, +{0x01EE48, 0x0001}, {0x01EE49, 0x0004}, -{0x01EE4A, 0x0080}, +{0x01EE4A, 0x0001}, {0x01EE4B, 0x0004}, -{0x01EE4C, 0x0080}, +{0x01EE4C, 0x0001}, {0x01EE4D, 0x0004}, -{0x01EE50, 0x0080}, +{0x01EE50, 0x0001}, {0x01EE51, 0x0004}, -{0x01EE53, 0x0080}, +{0x01EE53, 0x0001}, {0x01EE54, 0x0004}, -{0x01EE55, 0x0080}, +{0x01EE55, 0x0001}, {0x01EE57, 0x0004}, -{0x01EE58, 0x0080}, +{0x01EE58, 0x0001}, {0x01EE59, 0x0004}, -{0x01EE5A, 0x0080}, +{0x01EE5A, 0x0001}, {0x01EE5B, 0x0004}, -{0x01EE5C, 0x0080}, +{0x01EE5C, 0x0001}, {0x01EE5D, 0x0004}, -{0x01EE5E, 0x0080}, +{0x01EE5E, 0x0001}, {0x01EE5F, 0x0004}, -{0x01EE60, 0x0080}, +{0x01EE60, 0x0001}, {0x01EE61, 0x0004}, -{0x01EE63, 0x0080}, +{0x01EE63, 0x0001}, {0x01EE64, 0x0004}, -{0x01EE65, 0x0080}, +{0x01EE65, 0x0001}, {0x01EE67, 0x0004}, -{0x01EE6B, 0x0080}, +{0x01EE6B, 0x0001}, {0x01EE6C, 0x0004}, -{0x01EE73, 0x0080}, +{0x01EE73, 0x0001}, {0x01EE74, 0x0004}, -{0x01EE78, 0x0080}, +{0x01EE78, 0x0001}, {0x01EE79, 0x0004}, -{0x01EE7D, 0x0080}, +{0x01EE7D, 0x0001}, {0x01EE7E, 0x0004}, -{0x01EE7F, 0x0080}, +{0x01EE7F, 0x0001}, {0x01EE80, 0x0004}, -{0x01EE8A, 0x0080}, +{0x01EE8A, 0x0001}, {0x01EE8B, 0x0004}, -{0x01EE9C, 0x0080}, +{0x01EE9C, 0x0001}, {0x01EEA1, 0x0004}, -{0x01EEA4, 0x0080}, +{0x01EEA4, 0x0001}, {0x01EEA5, 0x0004}, -{0x01EEAA, 0x0080}, +{0x01EEAA, 0x0001}, {0x01EEAB, 0x0004}, -{0x01EEBC, 0x0080}, +{0x01EEBC, 0x0001}, {0x01EEF0, 0x0040}, -{0x01EEF2, 0x0080}, +{0x01EEF2, 0x0001}, {0x01F000, 0x0040}, -{0x01F02C, 0x0080}, +{0x01F02C, 0x0001}, {0x01F030, 0x0040}, -{0x01F094, 0x0080}, +{0x01F094, 0x0001}, {0x01F0A0, 0x0040}, -{0x01F0AF, 0x0080}, +{0x01F0AF, 0x0001}, {0x01F0B1, 0x0040}, -{0x01F0C0, 0x0080}, +{0x01F0C0, 0x0001}, {0x01F0C1, 0x0040}, -{0x01F0D0, 0x0080}, +{0x01F0D0, 0x0001}, {0x01F0D1, 0x0040}, -{0x01F0F6, 0x0080}, +{0x01F0F6, 0x0001}, {0x01F100, 0x0002}, {0x01F10D, 0x0040}, -{0x01F1AE, 0x0080}, +{0x01F1AE, 0x0001}, {0x01F1E6, 0x0040}, -{0x01F203, 0x0080}, +{0x01F203, 0x0001}, {0x01F210, 0x0040}, -{0x01F23C, 0x0080}, +{0x01F23C, 0x0001}, {0x01F240, 0x0040}, -{0x01F249, 0x0080}, +{0x01F249, 0x0001}, {0x01F250, 0x0040}, -{0x01F252, 0x0080}, +{0x01F252, 0x0001}, {0x01F260, 0x0040}, -{0x01F266, 0x0080}, +{0x01F266, 0x0001}, {0x01F300, 0x0040}, -{0x01F6D8, 0x0080}, +{0x01F6D8, 0x0001}, {0x01F6DC, 0x0040}, -{0x01F6ED, 0x0080}, +{0x01F6ED, 0x0001}, {0x01F6F0, 0x0040}, -{0x01F6FD, 0x0080}, +{0x01F6FD, 0x0001}, {0x01F700, 0x0040}, -{0x01F777, 0x0080}, +{0x01F777, 0x0001}, {0x01F77B, 0x0040}, -{0x01F7DA, 0x0080}, +{0x01F7DA, 0x0001}, {0x01F7E0, 0x0040}, -{0x01F7EC, 0x0080}, +{0x01F7EC, 0x0001}, {0x01F7F0, 0x0040}, -{0x01F7F1, 0x0080}, +{0x01F7F1, 0x0001}, {0x01F800, 0x0040}, -{0x01F80C, 0x0080}, +{0x01F80C, 0x0001}, {0x01F810, 0x0040}, -{0x01F848, 0x0080}, +{0x01F848, 0x0001}, {0x01F850, 0x0040}, -{0x01F85A, 0x0080}, +{0x01F85A, 0x0001}, {0x01F860, 0x0040}, -{0x01F888, 0x0080}, +{0x01F888, 0x0001}, {0x01F890, 0x0040}, -{0x01F8AE, 0x0080}, +{0x01F8AE, 0x0001}, {0x01F8B0, 0x0040}, -{0x01F8B2, 0x0080}, +{0x01F8B2, 0x0001}, {0x01F900, 0x0040}, -{0x01FA54, 0x0080}, +{0x01FA54, 0x0001}, {0x01FA60, 0x0040}, -{0x01FA6E, 0x0080}, +{0x01FA6E, 0x0001}, {0x01FA70, 0x0040}, -{0x01FA7D, 0x0080}, +{0x01FA7D, 0x0001}, {0x01FA80, 0x0040}, -{0x01FA89, 0x0080}, +{0x01FA89, 0x0001}, {0x01FA90, 0x0040}, -{0x01FABE, 0x0080}, +{0x01FABE, 0x0001}, {0x01FABF, 0x0040}, -{0x01FAC6, 0x0080}, +{0x01FAC6, 0x0001}, {0x01FACE, 0x0040}, -{0x01FADC, 0x0080}, +{0x01FADC, 0x0001}, {0x01FAE0, 0x0040}, -{0x01FAE9, 0x0080}, +{0x01FAE9, 0x0001}, {0x01FAF0, 0x0040}, -{0x01FAF9, 0x0080}, +{0x01FAF9, 0x0001}, {0x01FB00, 0x0040}, -{0x01FB93, 0x0080}, +{0x01FB93, 0x0001}, {0x01FB94, 0x0040}, -{0x01FBCB, 0x0080}, +{0x01FBCB, 0x0001}, {0x01FBF0, 0x0002}, -{0x01FBFA, 0x0080}, +{0x01FBFA, 0x0001}, {0x020000, 0x0004}, -{0x02A6E0, 0x0080}, +{0x02A6E0, 0x0001}, {0x02A700, 0x0004}, -{0x02B73A, 0x0080}, +{0x02B73A, 0x0001}, {0x02B740, 0x0004}, -{0x02B81E, 0x0080}, +{0x02B81E, 0x0001}, {0x02B820, 0x0004}, -{0x02CEA2, 0x0080}, +{0x02CEA2, 0x0001}, {0x02CEB0, 0x0004}, -{0x02EBE1, 0x0080}, +{0x02EBE1, 0x0001}, {0x02EBF0, 0x0004}, -{0x02EE5E, 0x0080}, +{0x02EE5E, 0x0001}, {0x02F800, 0x0004}, -{0x02FA1E, 0x0080}, +{0x02FA1E, 0x0001}, {0x030000, 0x0004}, -{0x03134B, 0x0080}, +{0x03134B, 0x0001}, {0x031350, 0x0004}, -{0x0323B0, 0x0080}, +{0x0323B0, 0x0001}, +{0x0E0001, 0x0080}, +{0x0E0002, 0x0001}, +{0x0E0020, 0x0080}, +{0x0E0080, 0x0001}, {0x0E0100, 0x0010}, -{0x0E01F0, 0x0080}, +{0x0E01F0, 0x0001}, +{0x0F0000, 0x0080}, +{0x0FFFFE, 0x0001}, +{0x100000, 0x0080}, +{0x10FFFE, 0x0001}, {0x110000, 0x0000}, }; const std::unordered_set unicode_set_whitespace = { -0x000009, 0x00000A, 0x00000B, 0x00000C, 0x00000D, 0x000020, 0x000085, 0x0000A0, 0x001680, 0x002000, 0x002001, 0x002002, 0x002003, 0x002004, 0x002005, 0x002006, 0x002007, 0x002008, 0x002009, 0x00200A, 0x002028, 0x002029, 0x00202F, 0x00205F, 0x003000 +0x000009, +0x00000A, +0x00000B, +0x00000C, +0x00000D, +0x000020, +0x000085, +0x0000A0, +0x001680, +0x002000, +0x002001, +0x002002, +0x002003, +0x002004, +0x002005, +0x002006, +0x002007, +0x002008, +0x002009, +0x00200A, +0x002028, +0x002029, +0x00202F, +0x00205F, +0x003000, }; const std::unordered_map unicode_map_lowercase = { @@ -3222,6 +3267,7 @@ const std::unordered_map unicode_map_lowercase = { {0x002C2C, 0x002C5C}, {0x002C2D, 0x002C5D}, {0x002C2E, 0x002C5E}, +{0x002C2F, 0x002C5F}, {0x002C60, 0x002C61}, {0x002C62, 0x00026B}, {0x002C63, 0x001D7D}, @@ -3402,12 +3448,16 @@ const std::unordered_map unicode_map_lowercase = { {0x00A7BA, 0x00A7BB}, {0x00A7BC, 0x00A7BD}, {0x00A7BE, 0x00A7BF}, +{0x00A7C0, 0x00A7C1}, {0x00A7C2, 0x00A7C3}, {0x00A7C4, 0x00A794}, {0x00A7C5, 0x000282}, {0x00A7C6, 0x001D8E}, {0x00A7C7, 0x00A7C8}, {0x00A7C9, 0x00A7CA}, +{0x00A7D0, 0x00A7D1}, +{0x00A7D6, 0x00A7D7}, +{0x00A7D8, 0x00A7D9}, {0x00A7F5, 0x00A7F6}, {0x00FF21, 0x00FF41}, {0x00FF22, 0x00FF42}, @@ -3511,6 +3561,41 @@ const std::unordered_map unicode_map_lowercase = { {0x0104D1, 0x0104F9}, {0x0104D2, 0x0104FA}, {0x0104D3, 0x0104FB}, +{0x010570, 0x010597}, +{0x010571, 0x010598}, +{0x010572, 0x010599}, +{0x010573, 0x01059A}, +{0x010574, 0x01059B}, +{0x010575, 0x01059C}, +{0x010576, 0x01059D}, +{0x010577, 0x01059E}, +{0x010578, 0x01059F}, +{0x010579, 0x0105A0}, +{0x01057A, 0x0105A1}, +{0x01057C, 0x0105A3}, +{0x01057D, 0x0105A4}, +{0x01057E, 0x0105A5}, +{0x01057F, 0x0105A6}, +{0x010580, 0x0105A7}, +{0x010581, 0x0105A8}, +{0x010582, 0x0105A9}, +{0x010583, 0x0105AA}, +{0x010584, 0x0105AB}, +{0x010585, 0x0105AC}, +{0x010586, 0x0105AD}, +{0x010587, 0x0105AE}, +{0x010588, 0x0105AF}, +{0x010589, 0x0105B0}, +{0x01058A, 0x0105B1}, +{0x01058C, 0x0105B3}, +{0x01058D, 0x0105B4}, +{0x01058E, 0x0105B5}, +{0x01058F, 0x0105B6}, +{0x010590, 0x0105B7}, +{0x010591, 0x0105B8}, +{0x010592, 0x0105B9}, +{0x010594, 0x0105BB}, +{0x010595, 0x0105BC}, {0x010C80, 0x010CC0}, {0x010C81, 0x010CC1}, {0x010C82, 0x010CC2}, @@ -3690,7 +3775,6 @@ const std::unordered_map unicode_map_uppercase = { {0x000079, 0x000059}, {0x00007A, 0x00005A}, {0x0000B5, 0x00039C}, -{0x0000DF, 0x000053}, {0x0000E0, 0x0000C0}, {0x0000E1, 0x0000C1}, {0x0000E2, 0x0000C2}, @@ -3758,7 +3842,6 @@ const std::unordered_map unicode_map_uppercase = { {0x000144, 0x000143}, {0x000146, 0x000145}, {0x000148, 0x000147}, -{0x000149, 0x0002BC}, {0x00014B, 0x00014A}, {0x00014D, 0x00014C}, {0x00014F, 0x00014E}, @@ -3831,7 +3914,6 @@ const std::unordered_map unicode_map_uppercase = { {0x0001EB, 0x0001EA}, {0x0001ED, 0x0001EC}, {0x0001EF, 0x0001EE}, -{0x0001F0, 0x00004A}, {0x0001F2, 0x0001F1}, {0x0001F3, 0x0001F1}, {0x0001F5, 0x0001F4}, @@ -3917,12 +3999,10 @@ const std::unordered_map unicode_map_uppercase = { {0x00037B, 0x0003FD}, {0x00037C, 0x0003FE}, {0x00037D, 0x0003FF}, -{0x000390, 0x000399}, {0x0003AC, 0x000386}, {0x0003AD, 0x000388}, {0x0003AE, 0x000389}, {0x0003AF, 0x00038A}, -{0x0003B0, 0x0003A5}, {0x0003B1, 0x000391}, {0x0003B2, 0x000392}, {0x0003B3, 0x000393}, @@ -4163,7 +4243,6 @@ const std::unordered_map unicode_map_uppercase = { {0x000584, 0x000554}, {0x000585, 0x000555}, {0x000586, 0x000556}, -{0x000587, 0x000535}, {0x0010D0, 0x001C90}, {0x0010D1, 0x001C91}, {0x0010D2, 0x001C92}, @@ -4303,11 +4382,6 @@ const std::unordered_map unicode_map_uppercase = { {0x001E91, 0x001E90}, {0x001E93, 0x001E92}, {0x001E95, 0x001E94}, -{0x001E96, 0x000048}, -{0x001E97, 0x000054}, -{0x001E98, 0x000057}, -{0x001E99, 0x000059}, -{0x001E9A, 0x000041}, {0x001E9B, 0x001E60}, {0x001EA1, 0x001EA0}, {0x001EA3, 0x001EA2}, @@ -4393,13 +4467,9 @@ const std::unordered_map unicode_map_uppercase = { {0x001F43, 0x001F4B}, {0x001F44, 0x001F4C}, {0x001F45, 0x001F4D}, -{0x001F50, 0x0003A5}, {0x001F51, 0x001F59}, -{0x001F52, 0x0003A5}, {0x001F53, 0x001F5B}, -{0x001F54, 0x0003A5}, {0x001F55, 0x001F5D}, -{0x001F56, 0x0003A5}, {0x001F57, 0x001F5F}, {0x001F60, 0x001F68}, {0x001F61, 0x001F69}, @@ -4423,89 +4493,41 @@ const std::unordered_map unicode_map_uppercase = { {0x001F7B, 0x001FEB}, {0x001F7C, 0x001FFA}, {0x001F7D, 0x001FFB}, -{0x001F80, 0x001F08}, -{0x001F81, 0x001F09}, -{0x001F82, 0x001F0A}, -{0x001F83, 0x001F0B}, -{0x001F84, 0x001F0C}, -{0x001F85, 0x001F0D}, -{0x001F86, 0x001F0E}, -{0x001F87, 0x001F0F}, -{0x001F88, 0x001F08}, -{0x001F89, 0x001F09}, -{0x001F8A, 0x001F0A}, -{0x001F8B, 0x001F0B}, -{0x001F8C, 0x001F0C}, -{0x001F8D, 0x001F0D}, -{0x001F8E, 0x001F0E}, -{0x001F8F, 0x001F0F}, -{0x001F90, 0x001F28}, -{0x001F91, 0x001F29}, -{0x001F92, 0x001F2A}, -{0x001F93, 0x001F2B}, -{0x001F94, 0x001F2C}, -{0x001F95, 0x001F2D}, -{0x001F96, 0x001F2E}, -{0x001F97, 0x001F2F}, -{0x001F98, 0x001F28}, -{0x001F99, 0x001F29}, -{0x001F9A, 0x001F2A}, -{0x001F9B, 0x001F2B}, -{0x001F9C, 0x001F2C}, -{0x001F9D, 0x001F2D}, -{0x001F9E, 0x001F2E}, -{0x001F9F, 0x001F2F}, -{0x001FA0, 0x001F68}, -{0x001FA1, 0x001F69}, -{0x001FA2, 0x001F6A}, -{0x001FA3, 0x001F6B}, -{0x001FA4, 0x001F6C}, -{0x001FA5, 0x001F6D}, -{0x001FA6, 0x001F6E}, -{0x001FA7, 0x001F6F}, -{0x001FA8, 0x001F68}, -{0x001FA9, 0x001F69}, -{0x001FAA, 0x001F6A}, -{0x001FAB, 0x001F6B}, -{0x001FAC, 0x001F6C}, -{0x001FAD, 0x001F6D}, -{0x001FAE, 0x001F6E}, -{0x001FAF, 0x001F6F}, +{0x001F80, 0x001F88}, +{0x001F81, 0x001F89}, +{0x001F82, 0x001F8A}, +{0x001F83, 0x001F8B}, +{0x001F84, 0x001F8C}, +{0x001F85, 0x001F8D}, +{0x001F86, 0x001F8E}, +{0x001F87, 0x001F8F}, +{0x001F90, 0x001F98}, +{0x001F91, 0x001F99}, +{0x001F92, 0x001F9A}, +{0x001F93, 0x001F9B}, +{0x001F94, 0x001F9C}, +{0x001F95, 0x001F9D}, +{0x001F96, 0x001F9E}, +{0x001F97, 0x001F9F}, +{0x001FA0, 0x001FA8}, +{0x001FA1, 0x001FA9}, +{0x001FA2, 0x001FAA}, +{0x001FA3, 0x001FAB}, +{0x001FA4, 0x001FAC}, +{0x001FA5, 0x001FAD}, +{0x001FA6, 0x001FAE}, +{0x001FA7, 0x001FAF}, {0x001FB0, 0x001FB8}, {0x001FB1, 0x001FB9}, -{0x001FB2, 0x001FBA}, -{0x001FB3, 0x000391}, -{0x001FB4, 0x000386}, -{0x001FB6, 0x000391}, -{0x001FB7, 0x000391}, -{0x001FBC, 0x000391}, +{0x001FB3, 0x001FBC}, {0x001FBE, 0x000399}, -{0x001FC2, 0x001FCA}, -{0x001FC3, 0x000397}, -{0x001FC4, 0x000389}, -{0x001FC6, 0x000397}, -{0x001FC7, 0x000397}, -{0x001FCC, 0x000397}, +{0x001FC3, 0x001FCC}, {0x001FD0, 0x001FD8}, {0x001FD1, 0x001FD9}, -{0x001FD2, 0x000399}, -{0x001FD3, 0x000399}, -{0x001FD6, 0x000399}, -{0x001FD7, 0x000399}, {0x001FE0, 0x001FE8}, {0x001FE1, 0x001FE9}, -{0x001FE2, 0x0003A5}, -{0x001FE3, 0x0003A5}, -{0x001FE4, 0x0003A1}, {0x001FE5, 0x001FEC}, -{0x001FE6, 0x0003A5}, -{0x001FE7, 0x0003A5}, -{0x001FF2, 0x001FFA}, -{0x001FF3, 0x0003A9}, -{0x001FF4, 0x00038F}, -{0x001FF6, 0x0003A9}, -{0x001FF7, 0x0003A9}, -{0x001FFC, 0x0003A9}, +{0x001FF3, 0x001FFC}, {0x00214E, 0x002132}, {0x002170, 0x002160}, {0x002171, 0x002161}, @@ -4597,6 +4619,7 @@ const std::unordered_map unicode_map_uppercase = { {0x002C5C, 0x002C2C}, {0x002C5D, 0x002C2D}, {0x002C5E, 0x002C2E}, +{0x002C5F, 0x002C2F}, {0x002C61, 0x002C60}, {0x002C65, 0x00023A}, {0x002C66, 0x00023E}, @@ -4800,9 +4823,13 @@ const std::unordered_map unicode_map_uppercase = { {0x00A7BB, 0x00A7BA}, {0x00A7BD, 0x00A7BC}, {0x00A7BF, 0x00A7BE}, +{0x00A7C1, 0x00A7C0}, {0x00A7C3, 0x00A7C2}, {0x00A7C8, 0x00A7C7}, {0x00A7CA, 0x00A7C9}, +{0x00A7D1, 0x00A7D0}, +{0x00A7D7, 0x00A7D6}, +{0x00A7D9, 0x00A7D8}, {0x00A7F6, 0x00A7F5}, {0x00AB53, 0x00A7B3}, {0x00AB70, 0x0013A0}, @@ -4885,18 +4912,6 @@ const std::unordered_map unicode_map_uppercase = { {0x00ABBD, 0x0013ED}, {0x00ABBE, 0x0013EE}, {0x00ABBF, 0x0013EF}, -{0x00FB00, 0x000046}, -{0x00FB01, 0x000046}, -{0x00FB02, 0x000046}, -{0x00FB03, 0x000046}, -{0x00FB04, 0x000046}, -{0x00FB05, 0x000053}, -{0x00FB06, 0x000053}, -{0x00FB13, 0x000544}, -{0x00FB14, 0x000544}, -{0x00FB15, 0x000544}, -{0x00FB16, 0x00054E}, -{0x00FB17, 0x000544}, {0x00FF41, 0x00FF21}, {0x00FF42, 0x00FF22}, {0x00FF43, 0x00FF23}, @@ -4999,6 +5014,41 @@ const std::unordered_map unicode_map_uppercase = { {0x0104F9, 0x0104D1}, {0x0104FA, 0x0104D2}, {0x0104FB, 0x0104D3}, +{0x010597, 0x010570}, +{0x010598, 0x010571}, +{0x010599, 0x010572}, +{0x01059A, 0x010573}, +{0x01059B, 0x010574}, +{0x01059C, 0x010575}, +{0x01059D, 0x010576}, +{0x01059E, 0x010577}, +{0x01059F, 0x010578}, +{0x0105A0, 0x010579}, +{0x0105A1, 0x01057A}, +{0x0105A3, 0x01057C}, +{0x0105A4, 0x01057D}, +{0x0105A5, 0x01057E}, +{0x0105A6, 0x01057F}, +{0x0105A7, 0x010580}, +{0x0105A8, 0x010581}, +{0x0105A9, 0x010582}, +{0x0105AA, 0x010583}, +{0x0105AB, 0x010584}, +{0x0105AC, 0x010585}, +{0x0105AD, 0x010586}, +{0x0105AE, 0x010587}, +{0x0105AF, 0x010588}, +{0x0105B0, 0x010589}, +{0x0105B1, 0x01058A}, +{0x0105B3, 0x01058C}, +{0x0105B4, 0x01058D}, +{0x0105B5, 0x01058E}, +{0x0105B6, 0x01058F}, +{0x0105B7, 0x010590}, +{0x0105B8, 0x010591}, +{0x0105B9, 0x010592}, +{0x0105BB, 0x010594}, +{0x0105BC, 0x010595}, {0x010CC0, 0x010C80}, {0x010CC1, 0x010C81}, {0x010CC2, 0x010C82}, diff --git a/unicode.cpp b/unicode.cpp index 2f8d73832d107..913c34b9b7bd6 100644 --- a/unicode.cpp +++ b/unicode.cpp @@ -226,8 +226,9 @@ static std::vector unicode_regex_split_custom_gpt2(const std::string & t assert(offset_end <= cpts.size()); start = offset_end; + static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF; auto _get_cpt = [&] (const size_t pos) -> uint32_t { - return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0; + return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE; }; auto _get_flags = [&] (const size_t pos) -> codepoint_flags { @@ -309,7 +310,7 @@ static std::vector unicode_regex_split_custom_gpt2(const std::string & t } // regex: \s+(?!\S) - if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) { + if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) { pos += num_whitespaces - 1; _add_token(pos); continue; @@ -344,8 +345,9 @@ static std::vector unicode_regex_split_custom_llama3(const std::string & assert(offset_end <= cpts.size()); start = offset_end; + static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF; auto _get_cpt = [&] (const size_t pos) -> uint32_t { - return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0; + return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE; }; auto _get_flags = [&] (const size_t pos) -> codepoint_flags { @@ -450,7 +452,7 @@ static std::vector unicode_regex_split_custom_llama3(const std::string & } // regex: \s+(?!\S) - if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) { + if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != OUT_OF_RANGE) { pos += num_whitespaces - 1; _add_token(pos); continue; @@ -679,10 +681,14 @@ std::vector unicode_regex_split(const std::string & text, const std continue; } - const int cpt_flag = unicode_cpt_flags(cpts[i]).category_flag(); + const auto flags = unicode_cpt_flags(cpts[i]); - if (k_ucat_cpt.find(cpt_flag) != k_ucat_cpt.end()) { - text_collapsed[i] = k_ucat_cpt.at(cpt_flag); + if (flags.is_whitespace) { + //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does. + //text_collapsed[i] = (char) 0x85; // as whitespace fallback + text_collapsed[i] = (char) 0x0B; // as whitespace fallback + } else if (k_ucat_cpt.find(flags.category_flag()) != k_ucat_cpt.end()) { + text_collapsed[i] = k_ucat_cpt.at(flags.category_flag()); } else { text_collapsed[i] = (char) 0xD0; // fallback } @@ -766,9 +772,16 @@ std::vector unicode_regex_split(const std::string & text, const std bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets); } else { // no unicode category used, we can use std::wregex directly - const std::wstring wtext = unicode_wstring_from_utf8(text); const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr); + // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback + std::wstring wtext(cpts.begin(), cpts.end()); + for (size_t i = 0; i < wtext.size(); ++i) { + if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) { + wtext[i] = 0x0B; + } + } + //printf("text: %s\n", text.c_str()); //printf("regex_expr: %s\n", regex_expr.c_str()); bpe_offsets = unicode_regex_split_stl(wtext, wregex_expr, bpe_offsets); From 623494a478134432fd2d7ee40135770a3340674f Mon Sep 17 00:00:00 2001 From: "Meng, Hengyu" Date: Wed, 19 Jun 2024 09:11:51 +0800 Subject: [PATCH 2/4] [SYCL] refactor (#6408) * seperate lower precision GEMM from the main files * fix workgroup size hardcode --- ggml-sycl.cpp | 9132 +++++--------------------------------- ggml-sycl/backend.hpp | 5 + ggml-sycl/convert.cpp | 544 +++ ggml-sycl/convert.hpp | 27 + ggml-sycl/dequantize.hpp | 690 +++ ggml-sycl/dmmv.cpp | 1022 +++++ ggml-sycl/dmmv.hpp | 27 + ggml-sycl/mmq.cpp | 3031 +++++++++++++ ggml-sycl/mmq.hpp | 33 + ggml-sycl/mmvq.cpp | 1024 +++++ ggml-sycl/mmvq.hpp | 27 + ggml-sycl/presets.hpp | 2 - ggml-sycl/vecdotq.hpp | 1161 +++++ 13 files changed, 8664 insertions(+), 8061 deletions(-) create mode 100644 ggml-sycl/convert.cpp create mode 100644 ggml-sycl/convert.hpp create mode 100644 ggml-sycl/dequantize.hpp create mode 100644 ggml-sycl/dmmv.cpp create mode 100644 ggml-sycl/dmmv.hpp create mode 100644 ggml-sycl/mmq.cpp create mode 100644 ggml-sycl/mmq.hpp create mode 100644 ggml-sycl/mmvq.cpp create mode 100644 ggml-sycl/mmvq.hpp create mode 100644 ggml-sycl/vecdotq.hpp diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index 6bd42b9609882..485f06ad331f8 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -38,67 +38,17 @@ #include "ggml-sycl/backend.hpp" -/* -Following definition copied from DPCT head files, which are used by ggml-sycl.cpp -*/ -// COPY from DPCT head files -#include -#include -#include - -#if defined(__linux__) -#include -#elif defined(_WIN64) -#ifndef NOMINMAX -#define NOMINMAX -#endif -#include -#else -#error "Only support Windows and Linux." -#endif - -#if defined(__linux__) -#include -#include -#endif -#if defined(_WIN64) -#ifndef NOMINMAX -#define NOMINMAX -#endif -#include -#endif - -#define DPCT_COMPATIBILITY_TEMP (900) - -#if defined(_MSC_VER) -#define __dpct_align__(n) __declspec(align(n)) -#define __dpct_inline__ __forceinline -#else -#define __dpct_align__(n) __attribute__((aligned(n))) -#define __dpct_inline__ __inline__ __attribute__((always_inline)) -#endif - -#if defined(_MSC_VER) -#define __dpct_noinline__ __declspec(noinline) -#else -#define __dpct_noinline__ __attribute__((noinline)) -#endif - bool ggml_sycl_loaded(void); void ggml_sycl_free_data(struct ggml_tensor * tensor); -void ggml_sycl_assign_buffers(struct ggml_tensor * tensor); -void ggml_sycl_assign_buffers_no_scratch(struct ggml_tensor * tensor); -void ggml_sycl_assign_buffers_force_inplace(struct ggml_tensor * tensor); -void ggml_sycl_assign_buffers_no_alloc(struct ggml_tensor * tensor); void ggml_sycl_copy_to_device(struct ggml_tensor * tensor); void ggml_sycl_set_main_device(int main_device); void ggml_sycl_set_mul_mat_q(bool mul_mat_q); -void ggml_sycl_set_scratch_size(size_t scratch_size); -void ggml_sycl_free_scratch(void); void ggml_sycl_get_device_description(int device, char * description, size_t description_size); bool ggml_backend_is_sycl(ggml_backend_t backend); int ggml_backend_sycl_get_device(ggml_backend_t backend); static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer); +static inline int get_sycl_env(const char *env_name, int default_val); +static inline int get_work_group_size(const sycl::device& device); void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst, const void *ptr_src, size_t size) { @@ -108,45 +58,6 @@ void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst, free(host_buf); } -static __dpct_inline__ int get_int_from_int8(const int8_t *x8, const int &i32) { - const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment - - int x32 = 0; - x32 |= x16[0] << 0; - x32 |= x16[1] << 16; - - return x32; -} - -static __dpct_inline__ int get_int_from_uint8(const uint8_t *x8, - const int &i32) { - const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment - - int x32 = 0; - x32 |= x16[0] << 0; - x32 |= x16[1] << 16; - - return x32; -} - -static __dpct_inline__ int get_int_from_int8_aligned(const int8_t *x8, - const int &i32) { - return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment -} - -static __dpct_inline__ int get_int_from_uint8_aligned(const uint8_t *x8, - const int &i32) { - return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment -} - -template -using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y, - int k, queue_ptr stream); -typedef to_t_sycl_t to_fp32_sycl_t; -typedef to_t_sycl_t to_fp16_sycl_t; - -typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v); -typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v); typedef void (*cpy_kernel_t)(const char * cx, char * cdst); typedef void (*ggml_sycl_func_t)(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst); typedef void (*ggml_sycl_op_mul_mat_t)( @@ -162,22 +73,6 @@ typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const gg const float *src1_dd, float *dst_dd, const queue_ptr &main_stream); -typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs); -typedef void (*allocate_tiles_sycl_t)(int **x_ql, sycl::half2 **x_dm, - int **x_qh, int **x_sc); -typedef void (*load_tiles_sycl_t)(const void *__restrict__ vx, - int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, - int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, - const int &i_max, const int &k, - const int &blocks_per_row); -typedef float (*vec_dot_q_mul_mat_sycl_t)( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ms, - const int &i, const int &j, const int &k); - static __dpct_inline__ float warp_reduce_sum(float x, const sycl::nd_item<3> &item_ct1) { #pragma unroll @@ -664,8029 +559,1392 @@ static void rms_norm_f32(const float * x, float * dst, const int ncols, const fl } } -static __dpct_inline__ void dequantize_q4_0(const void *vx, const int ib, - const int iqs, dfloat2 &v) { - const block_q4_0 * x = (const block_q4_0 *) vx; +static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded, + const sycl::nd_item<3> &item_ct1) { + const int ix = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); - const dfloat d = x[ib].d; + if (ix >= kx_padded) { + return; + } - const int vui = x[ib].qs[iqs]; + const int iy = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); - v.x() = vui & 0xF; - v.y() = vui >> 4; + const int i_padded = iy*kx_padded + ix; -#ifdef GGML_SYCL_F16 - // v = v - {8.0f, 8.0f}; - // v = v * {d, d}; - v.s0() = (v.s0() - 8.0f) * d; - v.s1() = (v.s1() - 8.0f) * d; + block_q8_1 * y = (block_q8_1 *) vy; -#else - v.x() = (v.x() - 8.0f) * d; - v.y() = (v.y() - 8.0f) * d; -#endif // GGML_SYCL_F16 -} + const int ib = i_padded / QK8_1; // block index + const int iqs = i_padded % QK8_1; // quant index -static __dpct_inline__ void dequantize_q4_1(const void *vx, const int ib, - const int iqs, dfloat2 &v) { - const block_q4_1 * x = (const block_q4_1 *) vx; + const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; + float amax = sycl::fabs((float)xi); + float sum = xi; - const dfloat d = x[ib].dm[0]; - const dfloat m = x[ib].dm[1]; +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + amax = sycl::fmax(amax, dpct::permute_sub_group_by_xor( + item_ct1.get_sub_group(), amax, mask)); + sum += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), sum, mask); + } - const int vui = x[ib].qs[iqs]; + const float d = amax / 127; + const int8_t q = amax == 0.0f ? 0 : sycl::round(xi / d); - v.x() = vui & 0xF; - v.y() = vui >> 4; + y[ib].qs[iqs] = q; -#ifdef GGML_SYCL_F16 - // v = v * {d, d}; - // v = v + {m, m}; - v.s0() = (v.s0() * d) + m; - v.s1() = (v.s1() * d) + m; + if (iqs > 0) { + return; + } -#else - v.x() = (v.x() * d) + m; - v.y() = (v.y() * d) + m; -#endif // GGML_SYCL_F16 + reinterpret_cast(y[ib].ds.x()) = d; + reinterpret_cast(y[ib].ds.y()) = sum; } -static __dpct_inline__ void dequantize_q5_0(const void *vx, const int ib, - const int iqs, dfloat2 &v) { - const block_q5_0 * x = (const block_q5_0 *) vx; +template +static void k_get_rows( + const void * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12, + const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { - const dfloat d = x[ib].d; + const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2)) * + 2; + const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) / + ne12; + const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) % + ne12; - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); + if (i00 >= ne00) { + return; + } - const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; - const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; - v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0); - v.y() = ((x[ib].qs[iqs] >> 4) | xh_1); + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03; -#ifdef GGML_SYCL_F16 - // v = v - {16.0f, 16.0f}; - // v = v * {d, d}; - v.s0() = (v.s0() - 16.0f) * d; - v.s1() = (v.s1() - 16.0f) * d; + const int ib = i00/qk; // block index + const int iqs = (i00%qk)/qr; // quant index + const int iybs = i00 - i00%qk; // dst block start index + const int y_offset = qr == 1 ? 1 : qk/2; -#else - v.x() = (v.x() - 16.0f) * d; - v.y() = (v.y() - 16.0f) * d; -#endif // GGML_SYCL_F16 + // dequantize + dfloat2 v; + dequantize_kernel(src0_row, ib, iqs, v); + + dst_row[iybs + iqs + 0] = v.x(); + dst_row[iybs + iqs + y_offset] = v.y(); } -static __dpct_inline__ void dequantize_q5_1(const void *vx, const int ib, - const int iqs, dfloat2 &v) { - const block_q5_1 * x = (const block_q5_1 *) vx; +template +static void k_get_rows_float( + const src0_t * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12, + const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { - const dfloat d = x[ib].dm[0]; - const dfloat m = x[ib].dm[1]; + const int i00 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2); + const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) / + ne12; + const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) % + ne12; - uint32_t qh; - memcpy(&qh, x[ib].qh, sizeof(qh)); + if (i00 >= ne00) { + return; + } - const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; - const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; - v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0); - v.y() = ((x[ib].qs[iqs] >> 4) | xh_1); + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03); -#ifdef GGML_SYCL_F16 - // v = v * {d, d}; - // v = v + {m, m}; - v.s0() = (v.s0() * d) + m; - v.s1() = (v.s1() * d) + m; -#else - v.x() = (v.x() * d) + m; - v.y() = (v.y() * d) + m; -#endif // GGML_SYCL_F16 + dst_row[i00] = src0_row[i00]; } -static __dpct_inline__ void dequantize_q8_0(const void *vx, const int ib, - const int iqs, dfloat2 &v) { - const block_q8_0 * x = (const block_q8_0 *) vx; - - const dfloat d = x[ib].d; - - v.x() = x[ib].qs[iqs + 0]; - v.y() = x[ib].qs[iqs + 1]; +static void mul_mat_p021_f16_f32( + const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y, + const sycl::nd_item<3> &item_ct1) { -#ifdef GGML_SYCL_F16 - // v = v * {d, d}; - v.s0() *= d; - v.s1() *= d; -#else - v.x() *= d; - v.y() *= d; -#endif // GGML_SYCL_F16 -} + const sycl::half *x = (const sycl::half *)vx; -template -static void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32, - const sycl::nd_item<3> &item_ct1) { + const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) + + item_ct1.get_local_id(0); + const int channel_x = channel / (nchannels_y / nchannels_x); - const int i = item_ct1.get_group(2); + const int nrows_y = ncols_x; + const int nrows_dst = nrows_x; + const int row_dst = row_x; - // assume 32 threads - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; - const int ir = tid%8; - const int ib = 8*i + ir; - if (ib >= nb32) { - return; - } + float tmp = 0.0f; - dst_t * y = yy + 256*i + 32*ir + 4*il; + for (int col_x0 = 0; col_x0 < ncols_x; + col_x0 += item_ct1.get_local_range(2)) { + const int col_x = col_x0 + item_ct1.get_local_id(2); - const block_q4_0 * x = (const block_q4_0 *)vx + ib; - const float d = sycl::vec(x->d) - .convert()[0]; - const float dm = -8*d; + if (col_x >= ncols_x) { + break; + } - const uint8_t * q = x->qs + 4*il; + // x is transposed and permuted + const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x; + const float xi = + sycl::vec(x[ix]) + .convert()[0]; - for (int l = 0; l < 4; ++l) { - y[l+ 0] = d * (q[l] & 0xF) + dm; - y[l+16] = d * (q[l] >> 4) + dm; - } -} + const int row_y = col_x; -template -static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32, - const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_group(2); + // y is not transposed but permuted + const int iy = channel*nrows_y + row_y; - // assume 32 threads - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; - const int ir = tid%8; - const int ib = 8*i + ir; - if (ib >= nb32) { - return; + tmp += xi * y[iy]; } - dst_t * y = yy + 256*i + 32*ir + 4*il; - - const block_q4_1 * x = (const block_q4_1 *)vx + ib; - const sycl::float2 d = - x->dm.convert(); + // dst is not transposed and not permuted + const int idst = channel*nrows_dst + row_dst; - const uint8_t * q = x->qs + 4*il; + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } - for (int l = 0; l < 4; ++l) { - y[l + 0] = d.x() * (q[l] & 0xF) + d.y(); - y[l + 16] = d.x() * (q[l] >> 4) + d.y(); + if (item_ct1.get_local_id(2) == 0) { + dst[idst] = tmp; } } +static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous + const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x, + const int row_stride_x, const int channel_stride_x, const int channel_x_divisor, + const sycl::nd_item<3> &item_ct1) { -//================================== k-quants - -template -static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { + const sycl::half *x = (const sycl::half *)vx; - const int i = item_ct1.get_group(2); - const block_q2_K * x = (const block_q2_K *) vx; + const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) + + item_ct1.get_local_id(0); + const int channel_x = channel / channel_x_divisor; - const int tid = item_ct1.get_local_id(2); - const int n = tid/32; - const int l = tid - 32*n; - const int is = 8*n + l/16; + const int nrows_y = ncols_x; + const int nrows_dst = nrows_x; + const int row_dst = row_x; - const uint8_t q = x[i].qs[32*n + l]; - dst_t * y = yy + i*QK_K + 128*n; + const int idst = channel*nrows_dst + row_dst; - float dall = x[i].dm[0]; - float dmin = x[i].dm[1]; - y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); - y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); - y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); - y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); -} + float tmp = 0.0f; -template -static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { + for (int col_x0 = 0; col_x0 < ncols_x; + col_x0 += item_ct1.get_local_range(2)) { + const int col_x = col_x0 + item_ct1.get_local_id(2); - const int i = item_ct1.get_group(2); - const block_q3_K * x = (const block_q3_K *) vx; + if (col_x >= ncols_x) { + break; + } - const int r = item_ct1.get_local_id(2) / 4; - const int tid = r/2; - const int is0 = r%2; - const int l0 = 16 * is0 + 4 * (item_ct1.get_local_id(2) % 4); - const int n = tid / 4; - const int j = tid - 4*n; + const int row_y = col_x; - uint8_t m = 1 << (4*n + j); - int is = 8*n + 2*j + is0; - int shift = 2*j; + const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x; + const int iy = channel*nrows_y + row_y; - int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) : - is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) : - is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) : - (x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4); - float d_all = x[i].d; - float dl = d_all * (us - 32); + const float xi = + sycl::vec(x[ix]) + .convert()[0]; - dst_t * y = yy + i*QK_K + 128*n + 32*j; - const uint8_t * q = x[i].qs + 32*n; - const uint8_t * hm = x[i].hmask; + tmp += xi * y[iy]; + } - for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); -} + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } -static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { - if (j < 4) { - d = q[j] & 63; m = q[j + 4] & 63; - } else { - d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); - m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); + if (item_ct1.get_local_id(2) == 0) { + dst[idst] = tmp; } } -template -static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { - const block_q4_K * x = (const block_q4_K *) vx; - - const int i = item_ct1.get_group(2); +static void cpy_1_f32_f32(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + float * dsti = (float *) cdsti; - // assume 32 threads - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; - const int ir = tid%8; - const int is = 2*il; - const int n = 4; + *dsti = *xi; +} - dst_t * y = yy + i*QK_K + 64*il + n*ir; +static void cpy_1_f32_f16(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + sycl::half *dsti = (sycl::half *)cdsti; - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; + *dsti = sycl::vec(*xi) + .convert()[0]; +} - const uint8_t * q = x[i].qs + 32*il + n*ir; +static void cpy_1_f16_f16(const char * cxi, char * cdsti) { + const sycl::half *xi = (const sycl::half *)cxi; + sycl::half *dsti = (sycl::half *)cdsti; - uint8_t sc, m; - get_scale_min_k4(is + 0, x[i].scales, sc, m); - const float d1 = dall * sc; const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[i].scales, sc, m); - const float d2 = dall * sc; const float m2 = dmin * m; - for (int l = 0; l < n; ++l) { - y[l + 0] = d1 * (q[l] & 0xF) - m1; - y[l +32] = d2 * (q[l] >> 4) - m2; - } + *dsti = *xi; } -template -static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { - const block_q5_K * x = (const block_q5_K *) vx; +static void cpy_1_f16_f32(const char * cxi, char * cdsti) { + const sycl::half *xi = (const sycl::half *)cxi; + float * dsti = (float *) cdsti; - const int i = item_ct1.get_group(2); + *dsti = *xi; +} - // assume 64 threads - this is very slightly better than the one below - const int tid = item_ct1.get_local_id(2); - const int il = tid/16; // il is in 0...3 - const int ir = tid%16; // ir is in 0...15 - const int is = 2*il; // is is in 0...6 +static void cpy_1_i16_i16(const char * cxi, char * cdsti) { + const int16_t *xi = (const int16_t *)cxi; + int16_t *dsti = (int16_t *)cdsti; - dst_t * y = yy + i*QK_K + 64*il + 2*ir; - - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; - - const uint8_t * ql = x[i].qs + 32*il + 2*ir; - const uint8_t * qh = x[i].qh + 2*ir; - - uint8_t sc, m; - get_scale_min_k4(is + 0, x[i].scales, sc, m); - const float d1 = dall * sc; const float m1 = dmin * m; - get_scale_min_k4(is + 1, x[i].scales, sc, m); - const float d2 = dall * sc; const float m2 = dmin * m; - - uint8_t hm = 1 << (2*il); - y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1; - y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1; - hm <<= 1; - y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; - y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; + *dsti = *xi; } -template -static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1) { - const block_q6_K * x = (const block_q6_K *) vx; +static void cpy_1_i32_i32(const char * cxi, char * cdsti) { + const int32_t *xi = (const int32_t *)cxi; + int32_t *dsti = (int32_t *)cdsti; - const int i = item_ct1.get_group(2); + *dsti = *xi; +} - // assume 64 threads - this is very slightly better than the one below - const int tid = item_ct1.get_local_id(2); - const int ip = tid/32; // ip is 0 or 1 - const int il = tid - 32*ip; // 0...32 - const int is = 8*ip + il/16; +template +static void cpy_f32_f16(const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); - dst_t * y = yy + i*QK_K + 128*ip + il; + if (i >= ne) { + return; + } - const float d = x[i].d; + // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor + // then combine those indices with the corresponding byte offsets to get the total offsets + const int i03 = i/(ne00 * ne01 * ne02); + const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); + const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; + const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; + const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - const uint8_t * ql = x[i].ql + 64*ip + il; - const uint8_t qh = x[i].qh[32*ip + il]; - const int8_t * sc = x[i].scales + is; + const int i13 = i/(ne10 * ne11 * ne12); + const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); + const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; + const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; + const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; - y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32); - y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32); - y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32); - y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); + cpy_1(cx + x_offset, cdst + dst_offset); } -template -static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1, - const uint64_t *iq2xxs_grid_ptr, - const uint8_t *ksigns_iq2xs_ptr, - const uint8_t *kmask_iq2xs_ptr) { +static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q8_0 * dsti = (block_q8_0 *) cdsti; - const int i = item_ct1.get_group(2); - const block_iq2_xxs * x = (const block_iq2_xxs *) vx; + float amax = 0.0f; // absolute max - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint16_t * q2 = x[i].qs + 4*ib; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid_ptr + aux8[il]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; - const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127]; - for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f); -} - -template -static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1, - const uint64_t *iq2xs_grid, - const uint8_t *ksigns_iq2xs, - const uint8_t *kmask_iq2xs) { - - const int i = item_ct1.get_group(2); - const block_iq2_xs * x = (const block_iq2_xs *) vx; + for (int j = 0; j < QK8_0; j++) { + const float v = xi[j]; + amax = sycl::fmax(amax, sycl::fabs((float)v)); + } - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint16_t * q2 = x[i].qs + 4*ib; - const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511)); - const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; - const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; - for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); -} + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; -template -__dpct_inline__ static void -dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy, - const sycl::nd_item<3> &item_ct1) { + dsti->d = d; - const int i = item_ct1.get_group(2); - const block_iq2_s * x = (const block_iq2_s *) vx; + for (int j = 0; j < QK8_0; ++j) { + const float x0 = xi[j]*id; - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300))); - const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; - const uint8_t signs = x[i].qs[QK_K/8+4*ib+il]; -#pragma unroll - for (int j = 0; j < 8; ++j) { - y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); + dsti->qs[j] = sycl::round((float)x0); } } -template -static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy, - const sycl::nd_item<3> &item_ct1, - const uint32_t *iq3xxs_grid, - const uint8_t *ksigns_iq2xs, - const uint8_t *kmask_iq2xs) { - - const int i = item_ct1.get_group(2); - const block_iq3_xxs * x = (const block_iq3_xxs *) vx; - - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint8_t * q3 = x[i].qs + 8*ib; - const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib; - const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]); - const uint32_t aux32 = gas[0] | (gas[1] << 16); - const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f; - const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; - for (int j = 0; j < 4; ++j) { - y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); - y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); - } -} - -template -__dpct_inline__ static void -dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy, - const sycl::nd_item<3> &item_ct1, - const uint8_t *kmask_iq2xs, const uint32_t *iq3s_grid) { +static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_0 * dsti = (block_q4_0 *) cdsti; - const int i = item_ct1.get_group(2); - const block_iq3_s * x = (const block_iq3_s *) vx; + float amax = 0.0f; + float vmax = 0.0f; - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint8_t * qs = x[i].qs + 8*ib; - const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256))); - const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256))); - const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)); - const uint8_t signs = x[i].signs[4*ib + il]; -#pragma unroll - for (int j = 0; j < 4; ++j) { - y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); - y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); + for (int j = 0; j < QK4_0; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float)v)) { + amax = sycl::fabs((float)v); + vmax = v; + } } -} - -template -__dpct_inline__ static void -dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy, - const sycl::nd_item<3> &item_ct1, - const uint32_t *iq1s_grid_gpu) { - const int i = item_ct1.get_group(2); - const block_iq1_s * x = (const block_iq1_s *) vx; + const float d = vmax / -8; + const float id = d ? 1.0f/d : 0.0f; - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA; - const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1); - uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32; - grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)]; - grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f; - grid32[0] &= 0x0f0f0f0f; -#pragma unroll - for (int j = 0; j < 8; ++j) { - y[j] = d * (q[j] + delta); - } -} + dsti->d = d; -template -__dpct_inline__ static void -dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy, - const sycl::nd_item<3> &item_ct1, - const uint32_t *iq1s_grid_gpu) { + for (int j = 0; j < QK4_0/2; ++j) { + const float x0 = xi[0 + j]*id; + const float x1 = xi[QK4_0/2 + j]*id; - const int i = item_ct1.get_group(2); - const block_iq1_m * x = (const block_iq1_m *) vx; + const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 8.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 8.5f)); - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 8*il; - const uint16_t * sc = (const uint16_t *)x[i].scales; - iq1m_scale_t scale; - scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); - const int ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4); - const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1); - const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA; - uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32; - grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)]; - grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f; - grid32[0] &= 0x0f0f0f0f; -#pragma unroll - for (int j = 0; j < 8; ++j) { - y[j] = d * (q[j] + delta); + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; } } -template -__dpct_inline__ static void -dequantize_block_iq4_nl(const void *__restrict__ vx, dst_t *__restrict__ yy, - const sycl::nd_item<3> &item_ct1) { +static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_1 * dsti = (block_q4_1 *) cdsti; - const int i = item_ct1.get_group(2); - const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL); + float vmin = FLT_MAX; + float vmax = -FLT_MAX; - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 4*il; - const uint8_t * q4 = x[ib].qs + 4*il; - const float d = (float)x[ib].d; -#pragma unroll - for (int j = 0; j < 4; ++j) { - y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; - y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; + for (int j = 0; j < QK4_1; ++j) { + const float v = xi[j]; + + if (v < vmin) vmin = v; + if (v > vmax) vmax = v; } -} + const float d = (vmax - vmin) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + dsti->dm.x() = d; + dsti->dm.y() = vmin; -template -__dpct_inline__ static void -dequantize_block_iq4_xs(const void *__restrict__ vx, dst_t *__restrict__ yy, - const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_group(2); - const block_iq4_xs * x = (const block_iq4_xs *)vx; + for (int j = 0; j < QK4_1/2; ++j) { + const float x0 = (xi[0 + j] - vmin)*id; + const float x1 = (xi[QK4_1/2 + j] - vmin)*id; - const int tid = item_ct1.get_local_id(2); - const int il = tid/8; // 0...3 - const int ib = tid%8; // 0...7 - dst_t * y = yy + i*QK_K + 32*ib + 4*il; - const uint8_t * q4 = x[i].qs + 16*ib + 4*il; - const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32); -#pragma unroll - for (int j = 0; j < 4; ++j) { - y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; - y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; + const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 0.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 0.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; } } +template +static void cpy_f32_q(const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { + const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2)) * + qk; + if (i >= ne) { + return; + } -/* -DPCT1110:4: The total declared local variable size in device function -dequantize_mul_mat_vec_q2_k exceeds 128 bytes and may cause high register -pressure. Consult with your hardware vendor to find the total register size -available and adjust the code, or use smaller sub-group size to avoid high -register pressure. -*/ -static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, - const float *__restrict__ yy, - float *__restrict__ dst, - const int ncols, int nrows, - const sycl::nd_item<3> &item_ct1) { - - static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); - - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - if (row > nrows) return; + const int i03 = i/(ne00 * ne01 * ne02); + const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); + const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; + const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; + const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; + const int i13 = i/(ne10 * ne11 * ne12); + const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); + const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; + const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; + const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13; - const block_q2_K * x = (const block_q2_K *)vx + ib0; + cpy_blck(cx + x_offset, cdst + dst_offset); +} - float tmp = 0; // partial sum for thread in warp +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); + return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); +} - const int tid = - item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...15 - const int ix = - item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 +struct rope_corr_dims { + float v[4]; +}; - const int step = 16/K_QUANTS_PER_ITERATION; +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; - const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const int in = tid - step*im; // 0...15 or 0...7 + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale); + } + *cos_theta = sycl::cos(theta) * mscale; + *sin_theta = sycl::sin(theta) * mscale; +} - const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2 - const int q_offset = 32*im + l0; - const int s_offset = 8*im; - const int y_offset = 128*im + l0; +// rope == RoPE == rotary positional embedding +template +static void rope( + const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, + float ext_factor, float attn_factor, rope_corr_dims corr_dims +, + const sycl::nd_item<3> &item_ct1) { + const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); - uint32_t aux[4]; - const uint8_t * d = (const uint8_t *)aux; - const uint8_t * m = (const uint8_t *)(aux + 2); + if (col >= ncols) { + return; + } - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + const int i = row*ncols + col; + const int i2 = row/p_delta_rows; - const float * y = yy + i * QK_K + y_offset; - const uint8_t * q = x[i].qs + q_offset; + const int p = has_pos ? pos[i2] : 0; + const float theta_base = p * dpct::pow(freq_base, -float(col) / ncols); - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); - const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset); - aux[0] = a[0] & 0x0f0f0f0f; - aux[1] = a[1] & 0x0f0f0f0f; - aux[2] = (a[0] >> 4) & 0x0f0f0f0f; - aux[3] = (a[1] >> 4) & 0x0f0f0f0f; + const float x0 = x[i + 0]; + const float x1 = x[i + 1]; - float sum1 = 0, sum2 = 0; - for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { - sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3) - + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3) - + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3) - + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3) - + y[l+16] * d[1] * ((q[l+16] >> 0) & 3) - + y[l+48] * d[3] * ((q[l+16] >> 2) & 3) - + y[l+80] * d[5] * ((q[l+16] >> 4) & 3) - +y[l+112] * d[7] * ((q[l+16] >> 6) & 3); - sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6] - + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7]; + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + 1] = x0*sin_theta + x1*cos_theta; +} - } - tmp += dall * sum1 - dmin * sum2; +template +static void rope_neox( + const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, + const float * freq_factors, const sycl::nd_item<3> &item_ct1) { + const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + if (col >= ncols) { + return; } - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + const int ib = col / n_dims; + const int ic = col % n_dims; - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} + if (ib > 0) { + const int i = row*ncols + ib*n_dims + ic; -/* -DPCT1110:5: The total declared local variable size in device function -dequantize_mul_mat_vec_q3_k exceeds 128 bytes and may cause high register -pressure. Consult with your hardware vendor to find the total register size -available and adjust the code, or use smaller sub-group size to avoid high -register pressure. -*/ -static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, - const float *__restrict__ yy, - float *__restrict__ dst, - const int ncols, int nrows, - const sycl::nd_item<3> &item_ct1) { + dst[i + 0] = x[i + 0]; + dst[i + 1] = x[i + 1]; - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - if (row > nrows) return; + return; + } - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; + const int i = row*ncols + ib*n_dims + ic/2; + const int i2 = row/p_delta_rows; - const block_q3_K * x = (const block_q3_K *)vx + ib0; + float cur_rot = inv_ndims * ic - ib; - float tmp = 0; // partial sum for thread in warp + const int p = has_pos ? pos[i2] : 0; + const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f; - const uint16_t kmask1 = 0x0303; - const uint16_t kmask2 = 0x0f0f; + const float theta_base = + p * freq_scale * dpct::pow(theta_scale, col / 2.0f)/freq_factor; - const int tid = - item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const int ix = - item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); - const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop - const int step = 16/K_QUANTS_PER_ITERATION; - const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const int in = tid - step*im; // 0....15 or 0...7 + const float x0 = x[i + 0]; + const float x1 = x[i + n_dims/2]; - const uint8_t m = 1 << (4*im); + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; +} - const int l0 = n*in; // 0...15 or 0...14 in steps of 2 - const int q_offset = 32*im + l0; - const int y_offset = 128*im + l0; +static void k_sum_rows_f32(const float * x, float * dst, const int ncols, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(1); + const int col = item_ct1.get_local_id(2); - uint16_t utmp[4]; - const int8_t * s = (const int8_t *)utmp; + float sum = 0.0f; + for (int i = col; i < ncols; i += item_ct1.get_local_range(2)) { + sum += x[row * ncols + i]; + } - const uint16_t s_shift = 4*im; + sum = warp_reduce_sum(sum, item_ct1); - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + y_offset; - const uint8_t * q = x[i].qs + q_offset; - const uint8_t * h = x[i].hmask + l0; + if (col == 0) { + dst[row] = sum; + } +} - const uint16_t * a = (const uint16_t *)x[i].scales; - utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4); - utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4); - utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4); - utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4); - const float d = x[i].d; +template +static inline void ggml_sycl_swap(T & a, T & b) { + T tmp = a; + a = b; + b = tmp; +} - float sum = 0; - for (int l = 0; l < n; ++l) { - sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4)) - + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4)) - + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4)) - + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4)); - sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4)) - + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4)) - + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4)) - + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4)); - } - tmp += d * sum; +template +__dpct_inline__ static void +k_argsort_f32_i32(const float *x, int *dst, const int ncols, int ncols_pad, + const sycl::nd_item<3> &item_ct1, uint8_t *dpct_local) { + // bitonic sort + int col = item_ct1.get_local_id(2); + int row = item_ct1.get_group(1); + if (col >= ncols_pad) { + return; } - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + const float * x_row = x + row * ncols; + auto dst_row = (int *)dpct_local; + + // initialize indices + dst_row[col] = col; + + item_ct1.barrier(sycl::access::fence_space::local_space); + + for (int k = 2; k <= ncols_pad; k *= 2) { + for (int j = k / 2; j > 0; j /= 2) { + int ixj = col ^ j; + if (ixj > col) { + if ((col & k) == 0) { + if (dst_row[col] >= ncols || + (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ? + x_row[dst_row[col]] > x_row[dst_row[ixj]] : + x_row[dst_row[col]] < x_row[dst_row[ixj]])) + ) { + ggml_sycl_swap(dst_row[col], dst_row[ixj]); + } + } else { + if (dst_row[ixj] >= ncols || + (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ? + x_row[dst_row[col]] < x_row[dst_row[ixj]] : + x_row[dst_row[col]] > x_row[dst_row[ixj]])) + ) { + ggml_sycl_swap(dst_row[col], dst_row[ixj]); + } + } + } + /* + DPCT1118:1: SYCL group functions and algorithms must be encountered + in converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + } } - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; + // copy the result to dst without the padding + if (col < ncols) { + dst[row * ncols + col] = dst_row[col]; } } -/* -DPCT1110:6: The total declared local variable size in device function -dequantize_mul_mat_vec_q4_k exceeds 128 bytes and may cause high register -pressure. Consult with your hardware vendor to find the total register size -available and adjust the code, or use smaller sub-group size to avoid high -register pressure. -*/ -static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, - const float *__restrict__ yy, - float *__restrict__ dst, - const int ncols, int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + +static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past, + const sycl::nd_item<3> &item_ct1) { + const int col = item_ct1.get_local_range(1) * item_ct1.get_group(1) + item_ct1.get_local_id(1); - if (row > nrows) return; - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q4_K * x = (const block_q4_K *)vx + ib0; - - const uint16_t kmask1 = 0x3f3f; - const uint16_t kmask2 = 0x0f0f; - const uint16_t kmask3 = 0xc0c0; - - const int tid = - item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const int ix = - item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); - const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4 + if (col >= ncols) { + return; + } - const int il = tid/step; // 0...3 - const int ir = tid - step*il; // 0...7 or 0...3 - const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4 + const int i = row*ncols + col; + //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i]; + //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU + dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX; +} - const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 - const int in = il%2; - const int l0 = n*(2*ir + in); - const int q_offset = 32*im + l0; - const int y_offset = 64*im + l0; +template +static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par, + const int nrows_y, const float scale, const float max_bias, const float m0, + const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) { + const int ncols = ncols_template == 0 ? ncols_par : ncols_template; - uint16_t aux[4]; - const uint8_t * sc = (const uint8_t *)aux; + const int tid = item_ct1.get_local_id(2); + const int rowx = item_ct1.get_group(2); + const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension -#if K_QUANTS_PER_ITERATION == 2 - uint32_t q32[4]; - const uint8_t * q4 = (const uint8_t *)q32; -#else - uint16_t q16[4]; - const uint8_t * q4 = (const uint8_t *)q16; -#endif + const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template; - float tmp = 0; // partial sum for thread in warp + const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + float slope = 1.0f; - const float * y1 = yy + i*QK_K + y_offset; - const float * y2 = y1 + 128; + // ALiBi + if (max_bias > 0.0f) { + const uint32_t h = rowx/nrows_y; // head index - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; + const float base = h < n_head_log2 ? m0 : m1; + const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; - const uint16_t * a = (const uint16_t *)x[i].scales; - aux[0] = a[im+0] & kmask1; - aux[1] = a[im+2] & kmask1; - aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); - aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + slope = sycl::pow(base, float(exp)); + } -#if K_QUANTS_PER_ITERATION == 2 - const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset); - const uint32_t * q2 = q1 + 16; + float * vals = vals_smem ? buf + WARP_SIZE : dst + rowx*ncols; + float max_val = -INFINITY; - q32[0] = q1[0] & 0x0f0f0f0f; - q32[1] = q1[0] & 0xf0f0f0f0; - q32[2] = q2[0] & 0x0f0f0f0f; - q32[3] = q2[0] & 0xf0f0f0f0; + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; - sycl::float4 s = {0.f, 0.f, 0.f, 0.f}; - float smin = 0; - for (int l = 0; l < 4; ++l) { - s.x() += y1[l] * q4[l + 0]; s.y() += y1[l + 32] * q4[l + 4]; - s.z() += y2[l] * q4[l + 8]; s.w() += y2[l + 32] * q4[l + 12]; - smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; - } - tmp += dall * (s.x() * sc[0] + s.y() * sc[1] * 1.f / 16.f + - s.z() * sc[4] + s.w() * sc[5] * 1.f / 16.f) - - dmin * smin; -#else - const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset); - const uint16_t * q2 = q1 + 32; - - q16[0] = q1[0] & 0x0f0f; - q16[1] = q1[0] & 0xf0f0; - q16[2] = q2[0] & 0x0f0f; - q16[3] = q2[0] & 0xf0f0; - - float4 s = {0.f, 0.f, 0.f, 0.f}; - float smin = 0; - for (int l = 0; l < 2; ++l) { - s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2]; - s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6]; - smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + if (ncols_template == 0 && col >= ncols) { + break; } - tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin; -#endif - } + const int ix = rowx*ncols + col; + const int iy = rowy*ncols + col; - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } + const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f); - if (tid == 0) { - dst[row] = tmp; + vals[col] = val; + max_val = sycl::max(max_val, val); } -} -/* -DPCT1110:7: The total declared local variable size in device function -dequantize_mul_mat_vec_q5_k exceeds 128 bytes and may cause high register -pressure. Consult with your hardware vendor to find the total register size -available and adjust the code, or use smaller sub-group size to avoid high -register pressure. -*/ -static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, - const float *__restrict__ yy, - float *__restrict__ dst, - const int ncols, - const sycl::nd_item<3> &item_ct1) { - - const int row = item_ct1.get_group(2); - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q5_K * x = (const block_q5_K *)vx + ib0; - - float tmp = 0; // partial sum for thread in warp - - const uint16_t kmask1 = 0x3f3f; - const uint16_t kmask2 = 0x0f0f; - const uint16_t kmask3 = 0xc0c0; - - const int tid = item_ct1.get_local_id(2) / 2; // 0...15 - const int ix = item_ct1.get_local_id(2) % 2; - - const int il = tid/4; // 0...3 - const int ir = tid - 4*il;// 0...3 - const int n = 2; - - const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 - const int in = il%2; - - const int l0 = n*(2*ir + in); - const int q_offset = 32*im + l0; - const int y_offset = 64*im + l0; - - const uint8_t hm1 = 1 << (2*im); - const uint8_t hm2 = hm1 << 4; - - uint16_t aux[4]; - const uint8_t * sc = (const uint8_t *)aux; - - uint16_t q16[8]; - const uint8_t * q4 = (const uint8_t *)q16; - - for (int i = ix; i < num_blocks_per_row; i += 2) { - - const uint8_t * ql1 = x[i].qs + q_offset; - const uint8_t * qh = x[i].qh + l0; - const float * y1 = yy + i*QK_K + y_offset; - const float * y2 = y1 + 128; - - const float dall = x[i].dm[0]; - const float dmin = x[i].dm[1]; - - const uint16_t * a = (const uint16_t *)x[i].scales; - aux[0] = a[im+0] & kmask1; - aux[1] = a[im+2] & kmask1; - aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); - aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); - - sycl::float4 sum = {0.f, 0.f, 0.f, 0.f}; - float smin = 0; - const uint16_t * q1 = (const uint16_t *)ql1; - const uint16_t * q2 = q1 + 32; - q16[0] = q1[0] & 0x0f0f; - q16[1] = q1[8] & 0x0f0f; - q16[2] = (q1[0] >> 4) & 0x0f0f; - q16[3] = (q1[8] >> 4) & 0x0f0f; - q16[4] = q2[0] & 0x0f0f; - q16[5] = q2[8] & 0x0f0f; - q16[6] = (q2[0] >> 4) & 0x0f0f; - q16[7] = (q2[8] >> 4) & 0x0f0f; - for (int l = 0; l < n; ++l) { - sum.x() += - y1[l + 0] * (q4[l + 0] + (qh[l + 0] & (hm1 << 0) ? 16 : 0)) + - y1[l + 16] * (q4[l + 2] + (qh[l + 16] & (hm1 << 0) ? 16 : 0)); - sum.y() += - y1[l + 32] * (q4[l + 4] + (qh[l + 0] & (hm1 << 1) ? 16 : 0)) + - y1[l + 48] * (q4[l + 6] + (qh[l + 16] & (hm1 << 1) ? 16 : 0)); - sum.z() += - y2[l + 0] * (q4[l + 8] + (qh[l + 0] & (hm2 << 0) ? 16 : 0)) + - y2[l + 16] * (q4[l + 10] + (qh[l + 16] & (hm2 << 0) ? 16 : 0)); - sum.w() += - y2[l + 32] * (q4[l + 12] + (qh[l + 0] & (hm2 << 1) ? 16 : 0)) + - y2[l + 48] * (q4[l + 14] + (qh[l + 16] & (hm2 << 1) ? 16 : 0)); - smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3] - + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7]; + // find the max value in the block + max_val = warp_reduce_max(max_val, item_ct1); + if (block_size > WARP_SIZE) { + if (warp_id == 0) { + buf[lane_id] = -INFINITY; } - tmp += dall * (sum.x() * sc[0] + sum.y() * sc[1] + sum.z() * sc[4] + - sum.w() * sc[5]) - - dmin * smin; - } + item_ct1.barrier(sycl::access::fence_space::local_space); - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } + if (lane_id == 0) { + buf[warp_id] = max_val; + } + item_ct1.barrier(sycl::access::fence_space::local_space); - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; + max_val = buf[lane_id]; + max_val = warp_reduce_max(max_val, item_ct1); } -} - -static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows, - const sycl::nd_item<3> &item_ct1) { - static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); - - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - if (row > nrows) return; - - const int num_blocks_per_row = ncols / QK_K; - const int ib0 = row*num_blocks_per_row; - - const block_q6_K * x = (const block_q6_K *)vx + ib0; + float tmp = 0.f; - const int tid = - item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 - const int ix = - item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0, 1 +#pragma unroll + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + if (ncols_template == 0 && col >= ncols) { + break; + } - const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + const float val = sycl::native::exp(vals[col] - max_val); + tmp += val; + vals[col] = val; + } - const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... - const int in = tid - step*im; // 0...15 or 0...7 + // find the sum of exps in the block + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + item_ct1.barrier(sycl::access::fence_space::local_space); + if (warp_id == 0) { + buf[lane_id] = 0.f; + } + item_ct1.barrier(sycl::access::fence_space::local_space); -#if K_QUANTS_PER_ITERATION == 1 - const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 - const int is = 0; -#else - const int l0 = 4 * in; // 0, 4, 8, ..., 28 - const int is = in / 4; -#endif - const int ql_offset = 64*im + l0; - const int qh_offset = 32*im + l0; - const int s_offset = 8*im + is; - const int y_offset = 128*im + l0; - - float tmp = 0; // partial sum for thread in warp - - for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { - - const float * y = yy + i * QK_K + y_offset; - const uint8_t * ql = x[i].ql + ql_offset; - const uint8_t * qh = x[i].qh + qh_offset; - const int8_t * s = x[i].scales + s_offset; - - const float d = x[i].d; - -#if K_QUANTS_PER_ITERATION == 1 - float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32) - + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32) - + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32) - + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32) - + y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32) - + y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32) - + y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32) - +y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32); - tmp += sum; -#else - float sum = 0; - for (int l = 0; l < 4; ++l) { - sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32) - + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32) - + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32) - + y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32); + if (lane_id == 0) { + buf[warp_id] = tmp; } - tmp += sum; -#endif + item_ct1.barrier(sycl::access::fence_space::local_space); + tmp = buf[lane_id]; + tmp = warp_reduce_sum(tmp, item_ct1); } - // sum up partial sums and write back result + const float inv_sum = 1.f / tmp; + #pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } + for (int col0 = 0; col0 < ncols; col0 += block_size) { + const int col = col0 + tid; + + if (ncols_template == 0 && col >= ncols) { + return; + } - if (tid == 0) { - dst[row] = tmp; + const int idst = rowx*ncols + col; + dst[idst] = vals[col] * inv_sum; } } -static void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const sycl::half *x = (const sycl::half *)vx; - - // automatic half -> float type cast if dfloat == float - v.x() = x[ib + iqs + 0]; - v.y() = x[ib + iqs + 1]; -} +static void scale_f32(const float * x, float * dst, const float scale, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); -static void convert_f32(const void * vx, const int ib, const int iqs, dfloat2 & v){ - const float * x = (const float *) vx; + if (i >= k) { + return; + } - // automatic half -> float type cast if dfloat == float - v.x() = x[ib + iqs + 0]; - v.y() = x[ib + iqs + 1]; + dst[i] = scale * x[i]; } -static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded, - const sycl::nd_item<3> &item_ct1) { - const int ix = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); +static void clamp_f32(const float * x, float * dst, const float min, const float max, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); - if (ix >= kx_padded) { + if (i >= k) { return; } - const int iy = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); + dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); +} - const int i_padded = iy*kx_padded + ix; +template +static void im2col_kernel(const float *x, T *dst, int offset_delta, + int IW, int IH, int OW, int KW, int KH, + int pelements, int CHW, int s0, int s1, int p0, + int p1, int d0, int d1, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_id(2) + + item_ct1.get_group(2) * item_ct1.get_local_range(2); + if (i >= pelements) { + return; + } - block_q8_1 * y = (block_q8_1 *) vy; + const int ksize = OW * (KH > 1 ? KW : 1); + const int kx = i / ksize; + const int kd = kx * ksize; + const int ky = (i - kd) / OW; + const int ix = i % OW; - const int ib = i_padded / QK8_1; // block index - const int iqs = i_padded % QK8_1; // quant index + const int64_t iiw = ix * s0 + kx * d0 - p0; + const int64_t iih = item_ct1.get_group(1) * s1 + ky * d1 - p1; - const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; - float amax = sycl::fabs((float)xi); - float sum = xi; + const int64_t offset_dst = + (item_ct1.get_group(1) * OW + ix) * CHW + + (item_ct1.get_group(0) * (KW * KH) + ky * KW + kx); -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - amax = sycl::fmax(amax, dpct::permute_sub_group_by_xor( - item_ct1.get_sub_group(), amax, mask)); - sum += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), sum, mask); - } - - const float d = amax / 127; - const int8_t q = amax == 0.0f ? 0 : sycl::round(xi / d); - - y[ib].qs[iqs] = q; - - if (iqs > 0) { - return; + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = + sycl::vec(0.0f) + .convert()[0]; + } else { + const int64_t offset_src = item_ct1.get_group(0) * offset_delta; + dst[offset_dst] = + sycl::vec(x[offset_src + iih * IW + iiw]) + .convert()[0]; } - - reinterpret_cast(y[ib].ds.x()) = d; - reinterpret_cast(y[ib].ds.y()) = sum; } -template -static void k_get_rows( - const void * src0, const int32_t * src1, dst_t * dst, - int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ - /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ - /*size_t s0,*/ size_t s1, size_t s2, size_t s3, - /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, - size_t s10, size_t s11, size_t s12, - const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { - - const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) + - item_ct1.get_local_id(2)) * - 2; - const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) / - ne12; - const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) % - ne12; - - if (i00 >= ne00) { - return; - } - - const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; +template +static void pool2d_nchw_kernel( + const int ih, const int iw, const int oh, const int ow, + const int kh, const int kw, const int sh, const int sw, + const int ph, const int pw, const int parallel_elements, + const Ti* src, To* dst, const enum ggml_op_pool op, + const sycl::nd_item<3> &item_ct1) { + int idx = item_ct1.get_local_id(2) + + item_ct1.get_group(2) * item_ct1.get_local_range(2); + if (idx >= parallel_elements) { + return; + } - dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; - const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03; + const int I_HW = ih * iw; + const int O_HW = oh * ow; + const int nc = idx / O_HW; + const int cur_oh = idx % O_HW / ow; + const int cur_ow = idx % O_HW % ow; + const Ti* i_ptr = src + nc * I_HW; + To* o_ptr = dst + nc * O_HW; + const int start_h = cur_oh * sh - ph; + const int bh = sycl::max(0, start_h); + const int eh = sycl::min(ih, start_h + kh); + const int start_w = cur_ow * sw - pw; + const int bw = sycl::max(0, start_w); + const int ew = sycl::min(iw, start_w + kw); - const int ib = i00/qk; // block index - const int iqs = (i00%qk)/qr; // quant index - const int iybs = i00 - i00%qk; // dst block start index - const int y_offset = qr == 1 ? 1 : qk/2; + To res = 0; - // dequantize - dfloat2 v; - dequantize_kernel(src0_row, ib, iqs, v); + switch (op) { + case GGML_OP_POOL_AVG: res = 0; break; + case GGML_OP_POOL_MAX: res = -FLT_MAX; break; + } - dst_row[iybs + iqs + 0] = v.x(); - dst_row[iybs + iqs + y_offset] = v.y(); + for (int i = bh; i < eh; i += 1) { + for (int j = bw; j < ew; j += 1) { +#if DPCT_COMPATIBILITY_TEMP >= 350 + /* + DPCT1098:106: The '*' expression is used instead of the __ldg + call. These two expressions do not provide the exact same + functionality. Check the generated code for potential precision + and/or performance issues. + */ + Ti cur = *(i_ptr + i * iw + j); +#else + Ti cur = i_ptr[i * iw + j]; +#endif + switch (op) { + case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break; + case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break; + } + } + } + o_ptr[cur_oh * ow + cur_ow] = res; } -template -static void k_get_rows_float( - const src0_t * src0, const int32_t * src1, dst_t * dst, - int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ - /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ - /*size_t s0,*/ size_t s1, size_t s2, size_t s3, - /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, - size_t s10, size_t s11, size_t s12, - const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { - - const int i00 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + - item_ct1.get_local_id(2); - const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) / - ne12; - const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) % - ne12; - - if (i00 >= ne00) { - return; - } - - const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; +template +static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const void *src0_dd, + const int32_t *src1_dd, float *dst_dd, + queue_ptr stream) { - dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; - const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03); + GGML_TENSOR_BINARY_OP_LOCALS - dst_row[i00] = src0_row[i00]; -} + const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); + const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE); + const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); -template -static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, - const sycl::nd_item<3> &item_ct1) { - const int i = 2 * (item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2)); + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); - if (i >= k) { - return; - } + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); - const int ib = i/qk; // block index - const int iqs = (i%qk)/qr; // quant index - const int iybs = i - i%qk; // y block start index - const int y_offset = qr == 1 ? 1 : qk/2; + GGML_ASSERT(ne00 % 2 == 0); - // dequantize - dfloat2 v; - dequantize_kernel(vx, ib, iqs, v); + 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); + }); - y[iybs + iqs + 0] = v.x(); - y[iybs + iqs + y_offset] = v.y(); + (void) dst; } -template -static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, - const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (i >= k) { - return; - } +template +static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const src0_t *src0_dd, const int32_t *src1_dd, + float *dst_dd, queue_ptr stream) { - const src_t * x = (src_t *) vx; + GGML_TENSOR_BINARY_OP_LOCALS - y[i] = x[i]; -} + const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); + const int block_num_x = (ne00 + SYCL_GET_ROWS_BLOCK_SIZE - 1) / SYCL_GET_ROWS_BLOCK_SIZE; + const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); -// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called -// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); -#define VDR_Q4_0_Q8_1_MMVQ 2 -#define VDR_Q4_0_Q8_1_MMQ 4 + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); -template -static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u, - const float &d4, - const sycl::half2 &ds8) { - int sumi = 0; -#pragma unroll - for (int i = 0; i < vdr; ++i) { - const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; - const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); - // SIMD dot product of quantized values - sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi); - sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); + stream->parallel_for( + 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); + }); } - const sycl::float2 ds8f = - ds8.convert(); - - // second part effectively subtracts 8 from each quant value - return d4 * (sumi * ds8f.x() - (8 * vdr / QI4_0) * ds8f.y()); + (void) dst; } -#define VDR_Q4_1_Q8_1_MMVQ 2 -#define VDR_Q4_1_Q8_1_MMQ 4 - -template -static __dpct_inline__ float vec_dot_q4_1_q8_1_impl(const int *v, const int *u, - const sycl::half2 &dm4, - const sycl::half2 &ds8) { - - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; - const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; - - // SIMD dot product of quantized values - sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi); - sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); - } - -#ifdef GGML_SYCL_F16 - const sycl::float2 tmp = - (dm4 * ds8).convert(); - const float d4d8 = tmp.x(); - const float m4s8 = tmp.y(); -#else - const sycl::float2 dm4f = - dm4.convert(); - const sycl::float2 ds8f = - ds8.convert(); - const float d4d8 = dm4f.x() * ds8f.x(); - const float m4s8 = dm4f.y() * ds8f.y(); -#endif // GGML_SYCL_F16 - - // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it - return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); -} +template +struct bin_bcast_sycl { + template + void operator()(ggml_backend_sycl_context & ctx, + const struct ggml_tensor *src0, + const struct ggml_tensor *src1, struct ggml_tensor *dst, + const src0_t *src0_dd, const src1_t *src1_dd, dst_t *dst_dd, + queue_ptr stream) { -#define VDR_Q5_0_Q8_1_MMVQ 2 -#define VDR_Q5_0_Q8_1_MMQ 4 + GGML_TENSOR_BINARY_OP_LOCALS -template -static __dpct_inline__ float -vec_dot_q5_0_q8_1_impl(const int *vl, const int *vh, const int *u, - const float &d5, const sycl::half2 &ds8) { - int sumi = 0; + int nr0 = ne10/ne0; + int nr1 = ne11/ne1; + int nr2 = ne12/ne2; + int nr3 = ne13/ne3; -#pragma unroll - for (int i = 0; i < vdr; ++i) { - int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits - vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 - vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 - vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 - vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = dpct::dp4a(vi0, u[2 * i + 0], - sumi); // SIMD dot product of quantized values + int nr[4] = { nr0, nr1, nr2, nr3 }; - int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits - vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 - vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 - vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 - vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = dpct::dp4a(vi1, u[2 * i + 1], - sumi); // SIMD dot product of quantized values - } + // collapse dimensions until first broadcast dimension + int64_t cne0[] = {ne0, ne1, ne2, ne3}; + int64_t cne1[] = {ne10, ne11, ne12, ne13}; + size_t cnb0[] = {nb0, nb1, nb2, nb3}; + size_t cnb1[] = {nb10, nb11, nb12, nb13}; + auto collapse = [](int64_t cne[]) { + cne[0] *= cne[1]; + cne[1] = cne[2]; + cne[2] = cne[3]; + cne[3] = 1; + }; - const sycl::float2 ds8f = - ds8.convert(); + auto collapse_nb = [](size_t cnb[], int64_t cne[]) { + cnb[1] *= cne[1]; + cnb[2] *= cne[2]; + cnb[3] *= cne[3]; + }; - // second part effectively subtracts 16 from each quant value - return d5 * (sumi * ds8f.x() - (16 * vdr / QI5_0) * ds8f.y()); -} - -#define VDR_Q5_1_Q8_1_MMVQ 2 -#define VDR_Q5_1_Q8_1_MMQ 4 - -template -static __dpct_inline__ float -vec_dot_q5_1_q8_1_impl(const int *vl, const int *vh, const int *u, - const sycl::half2 &dm5, const sycl::half2 &ds8) { - - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits - vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 - vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 - vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 - vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 - sumi = dpct::dp4a(vi0, u[2 * i + 0], - sumi); // SIMD dot product of quantized values - - int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits - vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 - vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 - vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 - vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 - sumi = dpct::dp4a(vi1, u[2 * i + 1], - sumi); // SIMD dot product of quantized values - } - -#ifdef GGML_SYCL_F16 - const sycl::float2 tmp = - (dm5 * ds8).convert(); - const float d5d8 = tmp.x(); - const float m5s8 = tmp.y(); - - -#else - const sycl::float2 dm5f = - dm5.convert(); - const sycl::float2 ds8f = - ds8.convert(); - const float d5d8 = dm5f.x() * ds8f.x(); - const float m5s8 = dm5f.y() * ds8f.y(); -#endif // GGML_SYCL_F16 - - // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it - return sumi*d5d8 + m5s8 / (QI5_1 / vdr); -} - -#define VDR_Q8_0_Q8_1_MMVQ 2 -#define VDR_Q8_0_Q8_1_MMQ 8 - -template -static __dpct_inline__ float vec_dot_q8_0_q8_1_impl(const int *v, const int *u, - const float &d8_0, - const float &d8_1) { - - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - // SIMD dot product of quantized values - sumi = dpct::dp4a(v[i], u[i], sumi); - } - - return d8_0*d8_1 * sumi; -} - -template -static __dpct_inline__ float vec_dot_q8_1_q8_1_impl(const int *v, const int *u, - const sycl::half2 &dm8, - const sycl::half2 &ds8) { - - int sumi = 0; - -#pragma unroll - for (int i = 0; i < vdr; ++i) { - // SIMD dot product of quantized values - sumi = dpct::dp4a(v[i], u[i], sumi); - } - -#ifdef GGML_SYCL_F16 - const sycl::float2 tmp = - (dm8 * ds8).convert(); - const float d8d8 = tmp.x(); - const float m8s8 = tmp.y(); -#else - const sycl::float2 dm8f = - dm8.convert(); - const sycl::float2 ds8f = - ds8.convert(); - const float d8d8 = dm8f.x() * ds8f.x(); - const float m8s8 = dm8f.y() * ds8f.y(); -#endif // GGML_SYCL_F16 - - // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it - return sumi*d8d8 + m8s8 / (QI8_1 / vdr); -} - -#define VDR_Q2_K_Q8_1_MMVQ 1 -#define VDR_Q2_K_Q8_1_MMQ 2 - -// contiguous v/x values -static __dpct_inline__ float vec_dot_q2_K_q8_1_impl_mmvq( - const int &v, const int *__restrict__ u, const uint8_t *__restrict__ scales, - const sycl::half2 &dm2, const float *__restrict__ d8) { - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR2_K; ++i) { - const int sc = scales[2*i]; - - const int vi = (v >> (2*i)) & 0x03030303; - - sumf_d += - d8[i] * (dpct::dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product - - // fill int with 4x m - int m = sc >> 4; - m |= m << 8; - m |= m << 16; - sumf_m += d8[i] * - dpct::dp4a( - m, u[i], - 0); // multiply constant q2_K part with sum of q8_1 values - } - - const sycl::float2 dm2f = - dm2.convert(); - - return dm2f.x() * sumf_d - dm2f.y() * sumf_m; -} - -// contiguous u/y values -static __dpct_inline__ float -vec_dot_q2_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, - const uint8_t *__restrict__ scales, - const sycl::half2 &dm2, const float &d8) { - - int sumi_d = 0; - int sumi_m = 0; - -#pragma unroll - for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) { - int sumi_d_sc = 0; - - const int sc = scales[i0 / (QI8_1/2)]; - - // fill int with 4x m - int m = sc >> 4; - m |= m << 8; - m |= m << 16; - -#pragma unroll - for (int i = i0; i < i0 + QI8_1/2; ++i) { - sumi_d_sc = dpct::dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product - sumi_m = dpct::dp4a(m, u[i], - sumi_m); // multiply sum of q8_1 values with m - } - - sumi_d += sumi_d_sc * (sc & 0xF); - } - - const sycl::float2 dm2f = - dm2.convert(); - - return d8 * (dm2f.x() * sumi_d - dm2f.y() * sumi_m); -} - -#define VDR_Q3_K_Q8_1_MMVQ 1 -#define VDR_Q3_K_Q8_1_MMQ 2 - -// contiguous v/x values -static __dpct_inline__ float vec_dot_q3_K_q8_1_impl_mmvq( - const int &vl, const int &vh, const int *__restrict__ u, - const uint8_t *__restrict__ scales, const int &scale_offset, - const float &d3, const float *__restrict__ d8) { - - float sumf = 0.0f; - -#pragma unroll - for (int i = 0; i < QR3_K; ++i) { - const int isc = scale_offset + 2*i; - - const int isc_low = isc % (QK_K/32); - const int sc_shift_low = 4 * (isc / (QK_K/32)); - const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF; - - const int isc_high = isc % (QK_K/64); - const int sc_shift_high = 2 * (isc / (QK_K/64)); - const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4; - - const int sc = (sc_low | sc_high) - 32; - - const int vil = (vl >> (2*i)) & 0x03030303; - - const int vih = ((vh >> i) << 2) & 0x04040404; - - const int vi = - dpct::vectorized_binary(vil, vih, dpct::sub_sat()); - - sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product - } - - return d3 * sumf; -} - -// contiguous u/y values -static __dpct_inline__ float -vec_dot_q3_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, - const int8_t *__restrict__ scales, const float &d3, - const float &d8) { - - int sumi = 0; - -#pragma unroll - for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) { - int sumi_sc = 0; - - for (int i = i0; i < i0 + QI8_1/2; ++i) { - sumi_sc = dpct::dp4a(v[i], u[i], sumi_sc); // SIMD dot product - } - - sumi += sumi_sc * scales[i0 / (QI8_1/2)]; - } - - return d3*d8 * sumi; -} - -#define VDR_Q4_K_Q8_1_MMVQ 2 -#define VDR_Q4_K_Q8_1_MMQ 8 - -// contiguous v/x values -static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_vmmq( - const int *__restrict__ v, const int *__restrict__ u, - const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, - const sycl::half2 &dm4, const float *__restrict__ d8) { - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR4_K; ++i) { - const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F; - const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F; - - const int dot1 = - dpct::dp4a(v1i, u[2 * i + 1], - dpct::dp4a(v0i, u[2 * i + 0], 0)); // SIMD dot product - const int dot2 = - dpct::dp4a(0x01010101, u[2 * i + 1], - dpct::dp4a(0x01010101, u[2 * i + 0], 0)); // sum of u - - sumf_d += d8[i] * (dot1 * sc[i]); - sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values - } - - const sycl::float2 dm4f = - dm4.convert(); - - return dm4f.x() * sumf_d - dm4f.y() * sumf_m; -} - -// contiguous u/y values -static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_mmq( - const int *__restrict__ v, const int *__restrict__ u, - const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, - const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) { - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) { - int sumi_d = 0; - -#pragma unroll - for (int j = 0; j < QI8_1; ++j) { - sumi_d = dpct::dp4a((v[j] >> (4 * i)) & 0x0F0F0F0F, - u[i * QI8_1 + j], sumi_d); // SIMD dot product - } - - const sycl::float2 ds8f = - ds8[i].convert(); - - sumf_d += ds8f.x() * (sc[i] * sumi_d); - sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val - } - - const sycl::float2 dm4f = - dm4.convert(); - - return dm4f.x() * sumf_d - dm4f.y() * sumf_m; -} - -#define VDR_Q5_K_Q8_1_MMVQ 2 -#define VDR_Q5_K_Q8_1_MMQ 8 - -// contiguous v/x values -static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_vmmq( - const int *__restrict__ vl, const int *__restrict__ vh, - const int *__restrict__ u, const uint8_t *__restrict__ sc, - const uint8_t *__restrict__ m, const sycl::half2 &dm5, - const float *__restrict__ d8) { - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR5_K; ++i) { - const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F; - const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F; - - const int vh0i = ((vh[0] >> i) << 4) & 0x10101010; - const int vh1i = ((vh[1] >> i) << 4) & 0x10101010; - - const int v0i = vl0i | vh0i; - const int v1i = vl1i | vh1i; - - const int dot1 = - dpct::dp4a(v0i, u[2 * i + 0], - dpct::dp4a(v1i, u[2 * i + 1], 0)); // SIMD dot product - const int dot2 = - dpct::dp4a(0x01010101, u[2 * i + 0], - dpct::dp4a(0x01010101, u[2 * i + 1], 0)); // sum of u - - sumf_d += d8[i] * (dot1 * sc[i]); - sumf_m += d8[i] * (dot2 * m[i]); - - } - - const sycl::float2 dm5f = - dm5.convert(); - - return dm5f.x() * sumf_d - dm5f.y() * sumf_m; -} - -// contiguous u/y values -static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_mmq( - const int *__restrict__ v, const int *__restrict__ u, - const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, - const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) { - - float sumf_d = 0.0f; - float sumf_m = 0.0f; - -#pragma unroll - for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) { - int sumi_d = 0; - -#pragma unroll - for (int j = 0; j < QI8_1; ++j) { - sumi_d = dpct::dp4a(v[i * QI8_1 + j], u[i * QI8_1 + j], - sumi_d); // SIMD dot product - } - - const sycl::float2 ds8f = - ds8[i].convert(); - - sumf_d += ds8f.x() * (sc[i] * sumi_d); - sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val - } - - const sycl::float2 dm4f = - dm4.convert(); - - return dm4f.x() * sumf_d - dm4f.y() * sumf_m; -} - -#define VDR_Q6_K_Q8_1_MMVQ 1 -#define VDR_Q6_K_Q8_1_MMQ 8 - -// contiguous v/x values -static __dpct_inline__ float -vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh, - const int *__restrict__ u, - const int8_t *__restrict__ scales, const float &d, - const float *__restrict__ d8) { - - float sumf = 0.0f; - -#pragma unroll - for (int i = 0; i < QR6_K; ++i) { - const int sc = scales[4*i]; - - const int vil = (vl >> (4*i)) & 0x0F0F0F0F; - - const int vih = ((vh >> (4*i)) << 4) & 0x30303030; - - const int vi = dpct::vectorized_binary( - (vil | vih), 0x20202020, dpct::sub_sat()); // vi = (vil | vih) - 32 - - sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product - } - - return d*sumf; -} - -// contiguous u/y values -static __dpct_inline__ float -vec_dot_q6_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, - const int8_t *__restrict__ sc, const float &d6, - const float *__restrict__ d8) { - - float sumf_d = 0.0f; - -#pragma unroll - for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) { - sycl::int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale - -#pragma unroll - for (int i = i0; i < i0 + 2; ++i) { - sumi_d.x() = dpct::dp4a(v[2 * i + 0], u[2 * i + 0], - sumi_d.x()); // SIMD dot product - sumi_d.x() = dpct::dp4a(v[2 * i + 1], u[2 * i + 1], - sumi_d.x()); // SIMD dot product - - sumi_d.y() = dpct::dp4a(v[2 * i + 4], u[2 * i + 4], - sumi_d.y()); // SIMD dot product - sumi_d.y() = dpct::dp4a(v[2 * i + 5], u[2 * i + 5], - sumi_d.y()); // SIMD dot product - } - - sumf_d += d8[i0 / 4] * - (sc[i0 / 2 + 0] * sumi_d.x() + sc[i0 / 2 + 1] * sumi_d.y()); - } - - return d6 * sumf_d; -} - -static __dpct_inline__ float -vec_dot_q4_0_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq; - - int v[VDR_Q4_0_Q8_1_MMVQ]; - int u[2*VDR_Q4_0_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); - } - - return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); -} - -template -static __dpct_inline__ void -allocate_tiles_q4_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_qs_q4_0, float *tile_x_d_q4_0) { - (void)x_qh; (void)x_sc; - - *x_ql = tile_x_qs_q4_0; - *x_dm = (sycl::half2 *)tile_x_d_q4_0; -} - -template -static __dpct_inline__ void -load_tiles_q4_0(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; (void)x_sc; - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI4_0; - const int kqsx = k % QI4_0; - - const block_q4_0 * bx0 = (const block_q4_0 *) vx; - - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); - // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI4_0; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) { - int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d; - } -} - -static __dpct_inline__ float vec_dot_q4_0_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - const float * x_dmf = (const float *) x_dm; - - int u[2*VDR_Q4_0_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE]; - } - - return vec_dot_q4_0_q8_1_impl - (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0], - y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __dpct_inline__ float -vec_dot_q4_1_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq; - - int v[VDR_Q4_1_Q8_1_MMVQ]; - int u[2*VDR_Q4_1_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1); - } - - return vec_dot_q4_1_q8_1_impl(v, u, bq4_1->dm, bq8_1->ds); -} - -template -static __dpct_inline__ void -allocate_tiles_q4_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_qs_q4_1, sycl::half2 *tile_x_dm_q4_1) { - (void)x_qh; (void)x_sc; - - *x_ql = tile_x_qs_q4_1; - *x_dm = tile_x_dm_q4_1; -} - -template -static __dpct_inline__ void -load_tiles_q4_1(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI4_1; - const int kqsx = k % QI4_1; - - const block_q4_1 * bx0 = (const block_q4_1 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI4_1; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) { - int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm; - } -} - -static __dpct_inline__ float vec_dot_q4_1_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - - int u[2*VDR_Q4_1_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE]; - } - - return vec_dot_q4_1_q8_1_impl - (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1], - y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __dpct_inline__ float -vec_dot_q5_0_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq; - - int vl[VDR_Q5_0_Q8_1_MMVQ]; - int vh[VDR_Q5_0_Q8_1_MMVQ]; - int u[2*VDR_Q5_0_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i); - vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0); - } - - return vec_dot_q5_0_q8_1_impl(vl, vh, u, bq5_0->d, bq8_1->ds); -} - -template -static __dpct_inline__ void -allocate_tiles_q5_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_ql_q5_0, float *tile_x_d_q5_0) { - (void)x_qh; (void)x_sc; - - *x_ql = tile_x_ql_q5_0; - *x_dm = (sycl::half2 *)tile_x_d_q5_0; -} - -template -static __dpct_inline__ void -load_tiles_q5_0(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI5_0; - const int kqsx = k % QI5_0; - - const block_q5_0 * bx0 = (const block_q5_0 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx; - - const int ql = get_int_from_uint8(bxi->qs, kqsx); - const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0)); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - qs0 = dpct::vectorized_binary( - qs0, 0x10101010, dpct::sub_sat()); // subtract 16 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - qs1 = dpct::vectorized_binary( - qs1, 0x10101010, dpct::sub_sat()); // subtract 16 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI5_0; - const int kbxd = k % blocks_per_tile_x_row; - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) { - int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d; - } -} - -static __dpct_inline__ float vec_dot_q5_0_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0; - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - int u[2*VDR_Q5_0_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE]; - } - - return vec_dot_q8_0_q8_1_impl - (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __dpct_inline__ float -vec_dot_q5_1_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq; - - int vl[VDR_Q5_1_Q8_1_MMVQ]; - int vh[VDR_Q5_1_Q8_1_MMVQ]; - int u[2*VDR_Q5_1_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) { - vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i); - vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i)); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1); - } - - return vec_dot_q5_1_q8_1_impl(vl, vh, u, bq5_1->dm, bq8_1->ds); -} - -template -static __dpct_inline__ void -allocate_tiles_q5_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_ql_q5_1, sycl::half2 *tile_x_dm_q5_1) { - (void)x_qh; (void)x_sc; - - *x_ql = tile_x_ql_q5_1; - *x_dm = tile_x_dm_q5_1; -} - -template -static __dpct_inline__ void -load_tiles_q5_1(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI5_1; - const int kqsx = k % QI5_1; - - const block_q5_1 * bx0 = (const block_q5_1 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx; - - const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); - const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1)); - - int qs0 = (ql >> 0) & 0x0F0F0F0F; - qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 - qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 - qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 - qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; - - int qs1 = (ql >> 4) & 0x0F0F0F0F; - qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 - qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 - qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 - qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 - - x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI5_1; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) { - int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm; - } -} - -static __dpct_inline__ float vec_dot_q5_1_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; (void)x_sc; - - const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1; - - int u[2*VDR_Q5_1_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) { - u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; - u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE]; - } - - return vec_dot_q8_1_q8_1_impl - (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); -} - -static __dpct_inline__ float -vec_dot_q8_0_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq; - - int v[VDR_Q8_0_Q8_1_MMVQ]; - int u[VDR_Q8_0_Q8_1_MMVQ]; - -#pragma unroll - for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_int8(bq8_0->qs, iqs + i); - u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - } - - return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, - bq8_1->ds[0]); -} - -template -static __dpct_inline__ void -allocate_tiles_q8_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_qs_q8_0, float *tile_x_d_q8_0) { - (void)x_qh; (void)x_sc; - - *x_ql = tile_x_qs_q8_0; - *x_dm = (sycl::half2 *)tile_x_d_q8_0; -} - -template -static __dpct_inline__ void -load_tiles_q8_0(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; (void)x_sc; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI8_0; - const int kqsx = k % QI8_0; - float * x_dmf = (float *) x_dm; - - const block_q8_0 * bx0 = (const block_q8_0 *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI8_0; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) { - int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d; - } -} - -static __dpct_inline__ float vec_dot_q8_0_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; (void)x_sc; - - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - return vec_dot_q8_0_q8_1_impl - (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0], - y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]); -} - -static __dpct_inline__ float -vec_dot_q2_K_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q2_K * bq2_K = (const block_q2_K *) vbq; - - const int bq8_offset = QR2_K * (iqs / QI8_1); - const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); - - const uint8_t * scales = bq2_K->scales + scale_offset; - - const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs); - int u[QR2_K]; - float d8[QR2_K]; - -#pragma unroll - for (int i = 0; i < QR2_K; ++ i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); - d8[i] = bq8_1[bq8_offset + i].ds[0]; - } - - return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8); -} - -template -static __dpct_inline__ void -allocate_tiles_q2_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_ql_q2_K, sycl::half2 *tile_x_dm_q2_K, - int *tile_x_sc_q2_K) { - (void)x_qh; - - *x_ql = tile_x_ql_q2_K; - *x_dm = tile_x_dm_q2_K; - *x_sc = tile_x_sc_q2_K; -} - -template -static __dpct_inline__ void -load_tiles_q2_K(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI2_K; - const int kqsx = k % QI2_K; - - const block_q2_K * bx0 = (const block_q2_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI2_K; - const int kbxd = k % blocks_per_tile_x_row; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) { - int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { - int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4); - - x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4)); - } -} - -static __dpct_inline__ float vec_dot_q2_K_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; - - const int kbx = k / QI2_K; - const int ky = (k % QI2_K) * QR2_K; - const float * y_df = (const float *) y_ds; - - int v[QR2_K*VDR_Q2_K_Q8_1_MMQ]; - - const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2); - const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2)); - -#pragma unroll - for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) { - v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303; - } - - const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4; - - const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE; - return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]); -} - -static __dpct_inline__ float -vec_dot_q3_K_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q3_K * bq3_K = (const block_q3_K *) vbq; - - const int bq8_offset = QR3_K * (iqs / (QI3_K/2)); - const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); - - const float d = bq3_K->d; - - const int vl = get_int_from_uint8(bq3_K->qs, iqs); - - // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted - const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; - - int u[QR3_K]; - float d8[QR3_K]; - -#pragma unroll - for (int i = 0; i < QR3_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); - d8[i] = bq8_1[bq8_offset + i].ds[0]; - } - - return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8); -} - -template -static __dpct_inline__ void -allocate_tiles_q3_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_ql_q3_K, sycl::half2 *tile_x_dm_q3_K, - int *tile_x_qh_q3_K, int *tile_x_sc_q3_K) { - - *x_ql = tile_x_ql_q3_K; - *x_dm = tile_x_dm_q3_K; - *x_qh = tile_x_qh_q3_K; - *x_sc = tile_x_sc_q3_K; -} - -template -static __dpct_inline__ void -load_tiles_q3_K(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI3_K; - const int kqsx = k % QI3_K; - - const block_q3_K * bx0 = (const block_q3_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI3_K; - const int kbxd = k % blocks_per_tile_x_row; - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) { - int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) { - int i = i0 + i_offset * 2 + k / (WARP_SIZE/2); - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2); - - // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted - x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2)); - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { - int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4); - - const int ksc = k % (QI3_K/4); - - const int ksc_low = ksc % (QI3_K/8); - const int shift_low = 4 * (ksc / (QI3_K/8)); - const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; - - const int ksc_high = QI3_K/8; - const int shift_high = 2 * ksc; - const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; - - const int sc = dpct::vectorized_binary( - sc_low | sc_high, 0x20202020, dpct::sub_sat()); - - x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc; - } -} - -static __dpct_inline__ float vec_dot_q3_K_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - - const int kbx = k / QI3_K; - const int ky = (k % QI3_K) * QR3_K; - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4; - - int v[QR3_K*VDR_Q3_K_Q8_1_MMQ]; - -#pragma unroll - for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) { - const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2); - const int shift = 2 * ((ky % 32) / 8); - const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303; - - const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8); - const int vlh = (vh << 2) & 0x04040404; - - v[l] = dpct::vectorized_binary(vll, vlh, dpct::sub_sat()); - } - - const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE; - return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]); -} - -static __dpct_inline__ float -vec_dot_q4_K_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q4_K * bq4_K = (const block_q4_K *) vbq; - - int v[2]; - int u[2*QR4_K]; - float d8[QR4_K]; - - // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6 - const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2)); - - // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12 - // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44 - // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76 - // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108 - - const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); - v[0] = q4[0]; - v[1] = q4[4]; - - const uint16_t * scales = (const uint16_t *)bq4_K->scales; - uint16_t aux[2]; - const int j = bq8_offset/2; - if (j < 2) { - aux[0] = scales[j+0] & 0x3f3f; - aux[1] = scales[j+2] & 0x3f3f; - } else { - aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); - aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); - } - const uint8_t * sc = (const uint8_t *)aux; - const uint8_t * m = sc + 2; - - for (int i = 0; i < QR4_K; ++i) { - const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; - d8[i] = bq8i->ds[0]; - - const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); - u[2*i+0] = q8[0]; - u[2*i+1] = q8[4]; - } - - return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8); -} - -template -static __dpct_inline__ void -allocate_tiles_q4_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_ql_q4_K, sycl::half2 *tile_x_dm_q4_K, - int *tile_x_sc_q4_K) { - (void)x_qh; - - *x_ql = tile_x_ql_q4_K; - *x_dm = tile_x_dm_q4_K; - *x_sc = tile_x_sc_q4_K; -} - -template -static __dpct_inline__ void -load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI4_K; // == 0 if QK_K == 256 - const int kqsx = k % QI4_K; // == k if QK_K == 256 - - const block_q4_K * bx0 = (const block_q4_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx; - - x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256 - const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) { - int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { - int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8); - - const int * scales = (const int *) bxi->scales; - - const int ksc = k % (WARP_SIZE/8); - - // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 - int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits - scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits - - x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; - } -} - -static __dpct_inline__ float vec_dot_q4_K_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; - - const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8); - - const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE; - return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8, - x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]); -} - -static __dpct_inline__ float -vec_dot_q5_K_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q5_K * bq5_K = (const block_q5_K *) vbq; - - int vl[2]; - int vh[2]; - int u[2*QR5_K]; - float d8[QR5_K]; - - const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2)); - const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); - const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4)); - - vl[0] = ql[0]; - vl[1] = ql[4]; - - vh[0] = qh[0] >> bq8_offset; - vh[1] = qh[4] >> bq8_offset; - - const uint16_t * scales = (const uint16_t *)bq5_K->scales; - uint16_t aux[2]; - const int j = bq8_offset/2; - if (j < 2) { - aux[0] = scales[j+0] & 0x3f3f; - aux[1] = scales[j+2] & 0x3f3f; - } else { - aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); - aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); - } - const uint8_t * sc = (const uint8_t *)aux; - const uint8_t * m = sc + 2; - -#pragma unroll - for (int i = 0; i < QR5_K; ++i) { - const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; - d8[i] = bq8i->ds[0]; - - const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); - u[2*i+0] = q8[0]; - u[2*i+1] = q8[4]; - } - - return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8); -} - -template -static __dpct_inline__ void -allocate_tiles_q5_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_ql_q5_K, sycl::half2 *tile_x_dm_q5_K, - int *tile_x_sc_q5_K) { - (void)x_qh; - - *x_ql = tile_x_ql_q5_K; - *x_dm = tile_x_dm_q5_K; - *x_sc = tile_x_sc_q5_K; -} - -template -static __dpct_inline__ void -load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI5_K; // == 0 if QK_K == 256 - const int kqsx = k % QI5_K; // == k if QK_K == 256 - - const block_q5_K * bx0 = (const block_q5_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx; - const int ky = QR5_K*kqsx; - - const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4)); - const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010; - const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010; - - const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0; - const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4); - - x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0; - x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1; - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256 - const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) { - int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { - int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8); - - const int * scales = (const int *) bxi->scales; - - const int ksc = k % (WARP_SIZE/8); - - // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 - int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits - scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits - - x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; - } -} - -static __dpct_inline__ float vec_dot_q5_K_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; - - const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8); - - const int index_x = i * (QR5_K*WARP_SIZE + 1) + QR5_K*k; - const int index_y = j * WARP_SIZE + (QR5_K*k) % WARP_SIZE; - return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8, - x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]); -} - -static __dpct_inline__ float -vec_dot_q6_K_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_q6_K * bq6_K = (const block_q6_K *) vbq; - - const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4); - const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8); - const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4)); - - const int vl = get_int_from_uint8(bq6_K->ql, iqs); - const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; - - const int8_t * scales = bq6_K->scales + scale_offset; - - int u[QR6_K]; - float d8[QR6_K]; - -#pragma unroll - for (int i = 0; i < QR6_K; ++i) { - u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); - d8[i] = bq8_1[bq8_offset + 2 * i].ds[0]; - } - - return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); -} - -template -static __dpct_inline__ void -allocate_tiles_q6_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, - int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_sc) { - (void)x_qh; - - *x_ql = tile_x_ql; - *x_dm = tile_x_dm; - *x_sc = tile_x_sc; -} - -template -static __dpct_inline__ void -load_tiles_q6_K(const void *__restrict__ vx, int *__restrict__ x_ql, - sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, - int *__restrict__ x_sc, const int &i_offset, const int &i_max, - const int &k, const int &blocks_per_row) { - (void)x_qh; - - GGML_SYCL_ASSUME(i_offset >= 0); - GGML_SYCL_ASSUME(i_offset < nwarps); - GGML_SYCL_ASSUME(k >= 0); - GGML_SYCL_ASSUME(k < WARP_SIZE); - - const int kbx = k / QI6_K; // == 0 if QK_K == 256 - const int kqsx = k % QI6_K; // == k if QK_K == 256 - - const block_q6_K * bx0 = (const block_q6_K *) vx; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { - int i = i0 + i_offset; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx; - const int ky = QR6_K*kqsx; - - const int ql = get_int_from_uint8(bxi->ql, kqsx); - const int ql0 = (ql >> 0) & 0x0F0F0F0F; - const int ql1 = (ql >> 4) & 0x0F0F0F0F; - - const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4)); - const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030; - const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030; - - const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0; - const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2); - - x_ql[i * (2 * WARP_SIZE + 1) + kq0] = - dpct::vectorized_binary(ql0 | qh0, 0x20202020, - dpct::sub_sat()); - x_ql[i * (2 * WARP_SIZE + 1) + kq1] = - dpct::vectorized_binary(ql1 | qh1, 0x20202020, - dpct::sub_sat()); - } - - const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256 - const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 - float * x_dmf = (float *) x_dm; - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) { - int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd; - - x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d; - } - -#pragma unroll - for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { - int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; - - if (need_check) { - i = sycl::min(i, i_max); - } - - const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4; - - x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8)); - } -} - -static __dpct_inline__ float vec_dot_q6_K_q8_1_mul_mat( - const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, - const int *__restrict__ x_qh, const int *__restrict__ x_sc, - const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, - const int &i, const int &j, const int &k) { - (void)x_qh; - - const float * x_dmf = (const float *) x_dm; - const float * y_df = (const float *) y_ds; - - const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]); - - const int index_x = i * (QR6_K*WARP_SIZE + 1) + QR6_K*k; - const int index_y = j * WARP_SIZE + (QR6_K*k) % WARP_SIZE; - return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]); -} - - -static __dpct_inline__ float -vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs, - const uint64_t *iq2xxs_grid, const uint8_t *ksigns_iq2xs, - const uint8_t *kmask_iq2xs) { - const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; - -#if QR2_XXS == 8 - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = q2[2] | (q2[3] << 16); - int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); - const uint8_t signs = ksigns_iq2xs[aux32 & 127]; - for (int j = 0; j < 8; ++j) { - sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); - } - q8 += 8; - aux32 >>= 7; - } - const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.25f; - return d * sumi; -#else - // iqs is 0...15 - const int ib32 = iqs/2; - const int il = iqs%2; - const uint16_t * q2 = bq2->qs + 4*ib32; - const uint8_t * aux8 = (const uint8_t *)q2; - const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); - const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); - const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * bq8_1[ib32].ds[0] * 0.25f; - const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; - const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; - const int8_t * q8 = bq8_1[ib32].qs + 16*il; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 8; ++j) { - sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); - sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); - } - return d * (sumi1 + sumi2); -#endif -} - -static __dpct_inline__ float -vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs, - const uint64_t *iq2xs_grid, const uint64_t *ksigns64) { -#if DPCT_COMPATIBILITY_TEMP >= \ - MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; - - const int ib32 = iqs; - const uint16_t * q2 = bq2->qs + 4*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; - int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = dpct::vectorized_binary( - grid[0] ^ signs[0], signs[0], std::minus<>()); - const int grid_h = dpct::vectorized_binary( - grid[1] ^ signs[1], signs[1], std::minus<>()); - sumi1 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; - } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); - const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); - const int grid_l = dpct::vectorized_binary( - grid[0] ^ signs[0], signs[0], std::minus<>()); - const int grid_h = dpct::vectorized_binary( - grid[1] ^ signs[1], signs[1], std::minus<>()); - sumi2 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; - } - const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -#else - assert(false); - return 0.f; -#endif -} - -static __dpct_inline__ float -vec_dot_iq2_s_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - const block_iq2_s * bq2 = (const block_iq2_s *) vbq; - - const int ib32 = iqs; - const int8_t * q8 = bq8_1[ib32].qs; - const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32; - const uint8_t ls1 = bq2->scales[ib32] & 0xf; - const uint8_t ls2 = bq2->scales[ib32] >> 4; - int sumi1 = 0; - for (int l = 0; l < 2; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = dpct::vectorized_binary( - ((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201, - std::equal_to<>()); - const uint32_t signs1 = dpct::vectorized_binary( - ((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201, - std::equal_to<>()); - const int grid_l = dpct::vectorized_binary( - grid[0] ^ signs0, signs0, std::minus<>()); - const int grid_h = dpct::vectorized_binary( - grid[1] ^ signs1, signs1, std::minus<>()); - sumi1 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi1); - sumi1 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi1); - q8 += 8; - } - int sumi2 = 0; - for (int l = 2; l < 4; ++l) { - const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); - const uint32_t signs0 = dpct::vectorized_binary( - ((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201, - std::equal_to<>()); - const uint32_t signs1 = dpct::vectorized_binary( - ((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201, - std::equal_to<>()); - const int grid_l = dpct::vectorized_binary( - grid[0] ^ signs0, signs0, std::minus<>()); - const int grid_h = dpct::vectorized_binary( - grid[1] ^ signs1, signs1, std::minus<>()); - sumi2 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi2); - sumi2 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi2); - q8 += 8; - } - const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f; - return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); -} - -static __dpct_inline__ float -vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs, - const uint32_t *iq3xxs_grid, const uint64_t *ksigns64) { -#if DPCT_COMPATIBILITY_TEMP >= \ - MIN_CC_DP4A // lowest compute capability for integer intrinsics - const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq; - - const int ib32 = iqs; - const uint8_t * q3 = bq2->qs + 8*ib32; - const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - uint32_t aux32 = gas[0] | (gas[1] << 16); - int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0]; - const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1]; - const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127)); - const int grid_l = dpct::vectorized_binary( - grid1[0] ^ signs[0], signs[0], std::minus<>()); - const int grid_h = dpct::vectorized_binary( - grid2[0] ^ signs[1], signs[1], std::minus<>()); - sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi); - sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi); - q8 += 8; - aux32 >>= 7; - } - const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.5f; - return d * sumi; -#else - assert(false); - return 0.f; -#endif -} - -static __dpct_inline__ float -vec_dot_iq3_s_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs, - const uint32_t *iq3s_grid) { - const block_iq3_s * bq2 = (const block_iq3_s *) vbq; - - const int ib32 = iqs; - const uint8_t * qs = bq2->qs + 8*ib32; - const int8_t * q8 = bq8_1[ib32].qs; - int sumi = 0; - for (int l = 0; l < 4; ++l) { - const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); - const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); - uint32_t signs0 = dpct::vectorized_binary( - ((bq2->signs[4 * ib32 + l] & 0xf) * 0x01010101) & 0x08040201, - 0x08040201, std::equal_to<>()); - uint32_t signs1 = dpct::vectorized_binary( - ((bq2->signs[4 * ib32 + l] >> 4) * 0x01010101) & 0x08040201, - 0x08040201, std::equal_to<>()); - const int grid_l = dpct::vectorized_binary( - grid1[0] ^ signs0, signs0, std::minus<>()); - const int grid_h = dpct::vectorized_binary( - grid2[0] ^ signs1, signs1, std::minus<>()); - sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi); - sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi); - q8 += 8; - } - const float d = - (float)bq2->d * - (1 + 2 * ((bq2->scales[ib32 / 2] >> 4 * (ib32 % 2)) & 0xf)) * - bq8_1[ib32].ds[0]; - return d * sumi; -} - -static __dpct_inline__ float -vec_dot_iq1_s_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs, - const uint32_t *iq1s_grid_gpu) { - const block_iq1_s * bq1 = (const block_iq1_s *) vbq; - - const int ib32 = iqs; - int sumi = 0; - const int * q8 = (const int *)bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); - int grid0 = grid[0] & 0x0f0f0f0f; - int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; - sumi = dpct::dp4a(q8[2 * l + 1], grid1, - dpct::dp4a(q8[2 * l + 0], grid0, sumi)); - } - - const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA; - const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1); - const float d = d1q * bq8_1[ib32].ds[0]; - const float m = d1q * bq8_1[ib32].ds[1]; - return d * sumi + m * delta; -} - -static __dpct_inline__ float -vec_dot_iq1_m_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - const block_iq1_m * bq1 = (const block_iq1_m *) vbq; - - const int ib32 = iqs; - int sumi[2] = {0, 0}; - float sumf[2] = {0.f, 0.f}; - - const int * q8 = (const int *)bq8_1[ib32].qs; - for (int l = 0; l < 4; ++l) { - const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8))); - int grid0 = grid[0] & 0x0f0f0f0f; - int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; - sumi[l / 2] = dpct::dp4a(q8[2 * l + 1], grid1, - dpct::dp4a(q8[2 * l + 0], grid0, sumi[l / 2])); - const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA; - const int sumy = dpct::dp4a(q8[2 * l + 1], 0x01010101, - dpct::dp4a(q8[2 * l + 0], 0x01010101, 0)); - sumf[l/2] += delta*sumy; - } - - iq1m_scale_t scale; - const uint16_t * sc = (const uint16_t *)bq1->scales; - scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); - const float d = (float)scale.f16 * bq8_1[ib32].ds[0]; - return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1)); -} - -static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4, - const uint8_t *values, - int &val1, int &val2) { - - uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32; - aux32 = q4 & 0x0f0f0f0f; - uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8); - uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8); - val1 = v1 | (v2 << 16); - aux32 = (q4 >> 4) & 0x0f0f0f0f; - v1 = values[q8[0]] | (values[q8[1]] << 8); - v2 = values[q8[2]] | (values[q8[3]] << 8); - val2 = v1 | (v2 << 16); -} - - -static __dpct_inline__ float -vec_dot_iq4_nl_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_iq4_nl * bq = (const block_iq4_nl *) vbq; - - const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs; - const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs; - - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - - int v1, v2; - int sumi1 = 0, sumi2 = 0; - for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) { - const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16); - get_int_from_table_16(aux, values, v1, v2); - sumi1 = dpct::dp4a(v1, q8[l + 0], sumi1); - sumi2 = dpct::dp4a(v2, q8[l + 4], sumi2); - } - - const float d = (float)bq->d * bq8_1->ds[0]; - return d * (sumi1 + sumi2); -} - - -static __dpct_inline__ float -vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq, - const block_q8_1 *__restrict__ bq8_1, const int &iqs) { - - const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq; - const uint8_t * values = (const uint8_t *)kvalues_iq4nl; - - // iqs is 0...7 - const int ib32 = iqs; - const int32_t * q8 = (const int *)bq8_1[ib32].qs; - const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32; - const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); - const float d = (float)bq4->d * (ls - 32) * bq8_1[ib32].ds[0]; - int v1, v2; - int sumi1 = 0, sumi2 = 0; - for (int j = 0; j < 4; ++j) { - get_int_from_table_16(q4[j], values, v1, v2); - sumi1 = dpct::dp4a(v1, q8[j + 0], sumi1); - sumi2 = dpct::dp4a(v2, q8[j + 4], sumi2); - } - return d * (sumi1 + sumi2); -} - -template -/* -DPCT1110:8: The total declared local variable size in device function mul_mat_q -exceeds 128 bytes and may cause high register pressure. Consult with your -hardware vendor to find the total register size available and adjust the code, -or use smaller sub-group size to avoid high register pressure. -*/ -static __dpct_inline__ void -mul_mat_q(const void *__restrict__ vx, const void *__restrict__ vy, - float *__restrict__ dst, const int ncols_x, const int nrows_x, - const int ncols_y, const int nrows_y, const int nrows_dst, - int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_qh, - int *tile_x_sc, const sycl::nd_item<3> &item_ct1, int *tile_y_qs, - sycl::half2 *tile_y_ds) { - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - const int blocks_per_row_x = ncols_x / qk; - const int blocks_per_col_y = nrows_y / QK8_1; - const int blocks_per_warp = WARP_SIZE / qi; - - const int & ncols_dst = ncols_y; - - const int row_dst_0 = item_ct1.get_group(2) * mmq_y; - const int & row_x_0 = row_dst_0; - - const int col_dst_0 = item_ct1.get_group(1) * mmq_x; - const int & col_y_0 = col_dst_0; - - float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}}; - - for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) { - - load_tiles(x + row_x_0 * blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, - tile_x_qh, tile_x_sc, item_ct1.get_local_id(1), - nrows_x - row_x_0 - 1, item_ct1.get_local_id(2), - blocks_per_row_x); - -#pragma unroll - for (int ir = 0; ir < qr; ++ir) { - const int kqs = ir * WARP_SIZE + item_ct1.get_local_id(2); - const int kbxd = kqs / QI8_1; - -#pragma unroll - for (int i = 0; i < mmq_x; i += nwarps) { - const int col_y_eff = dpct::min( - (unsigned int)(col_y_0 + item_ct1.get_local_id(1) + i), - ncols_y - 1); // to prevent out-of-bounds memory accesses - - const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd]; - - const int index_y = (item_ct1.get_local_id(1) + i) * WARP_SIZE + - kqs % WARP_SIZE; - tile_y_qs[index_y] = get_int_from_int8_aligned( - by0->qs, item_ct1.get_local_id(2) % QI8_1); - } - -#pragma unroll - for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) { - const int ids = - (ids0 + item_ct1.get_local_id(1) * QI8_1 + - item_ct1.get_local_id(2) / (WARP_SIZE / QI8_1)) % - mmq_x; - const int kby = item_ct1.get_local_id(2) % (WARP_SIZE / QI8_1); - const int col_y_eff = sycl::min(col_y_0 + ids, ncols_y - 1); - - // if the sum is not needed it's faster to transform the scale to f32 ahead of time - const sycl::half2 *dsi_src = - &y[col_y_eff * blocks_per_col_y + ib0 * (qk / QK8_1) + - ir * (WARP_SIZE / QI8_1) + kby] - .ds; - sycl::half2 *dsi_dst = - &tile_y_ds[ids * (WARP_SIZE / QI8_1) + kby]; - if (need_sum) { - *dsi_dst = *dsi_src; - } else { - float * dfi_dst = (float *) dsi_dst; - *dfi_dst = (*dsi_src)[0]; - } - } - - /* - DPCT1118:9: SYCL group functions and algorithms must be encountered - in converged control flow. You may need to adjust the code. - */ - /* - DPCT1065:56: Consider replacing sycl::nd_item::barrier() with - sycl::nd_item::barrier(sycl::access::fence_space::local_space) for - better performance if there is no access to global memory. - */ - item_ct1.barrier(); - -// #pragma unroll // unrolling this loop causes too much register pressure - for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) { -#pragma unroll - for (int j = 0; j < mmq_x; j += nwarps) { -#pragma unroll - for (int i = 0; i < mmq_y; i += WARP_SIZE) { - sum[i / WARP_SIZE][j / nwarps] += vec_dot( - tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, - tile_y_qs, tile_y_ds, item_ct1.get_local_id(2) + i, - item_ct1.get_local_id(1) + j, k); - } - } - } - - /* - DPCT1118:10: SYCL group functions and algorithms must be encountered - in converged control flow. You may need to adjust the code. - */ - /* - DPCT1065:57: Consider replacing sycl::nd_item::barrier() with - sycl::nd_item::barrier(sycl::access::fence_space::local_space) for - better performance if there is no access to global memory. - */ - item_ct1.barrier(); - } - } - -#pragma unroll - for (int j = 0; j < mmq_x; j += nwarps) { - const int col_dst = col_dst_0 + j + item_ct1.get_local_id(1); - - if (col_dst >= ncols_dst) { - return; - } - -#pragma unroll - for (int i = 0; i < mmq_y; i += WARP_SIZE) { - const int row_dst = row_dst_0 + item_ct1.get_local_id(2) + i; - - if (row_dst >= nrows_dst) { - continue; - } - - dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps]; - } - } -} - -#define MMQ_X_Q4_0_RDNA2 64 -#define MMQ_Y_Q4_0_RDNA2 128 -#define NWARPS_Q4_0_RDNA2 8 -#define MMQ_X_Q4_0_RDNA1 64 -#define MMQ_Y_Q4_0_RDNA1 64 -#define NWARPS_Q4_0_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q4_0_AMPERE 4 -#define MMQ_Y_Q4_0_AMPERE 32 -#define NWARPS_Q4_0_AMPERE 4 -#else -#define MMQ_X_Q4_0_AMPERE 64 -#define MMQ_Y_Q4_0_AMPERE 128 -#define NWARPS_Q4_0_AMPERE 4 -#endif -#define MMQ_X_Q4_0_PASCAL 64 -#define MMQ_Y_Q4_0_PASCAL 64 -#define NWARPS_Q4_0_PASCAL 8 - -template static void - mul_mat_q4_0( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_0, float *tile_x_d_q4_0, - int *tile_y_qs, sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - - const int mmq_x = MMQ_X_Q4_0_AMPERE; - const int mmq_y = MMQ_Y_Q4_0_AMPERE; - const int nwarps = NWARPS_Q4_0_AMPERE; - allocate_tiles_q4_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_qs_q4_0, tile_x_d_q4_0); - mul_mat_q, VDR_Q4_0_Q8_1_MMQ, - vec_dot_q4_0_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q4_1_RDNA2 64 -#define MMQ_Y_Q4_1_RDNA2 128 -#define NWARPS_Q4_1_RDNA2 8 -#define MMQ_X_Q4_1_RDNA1 64 -#define MMQ_Y_Q4_1_RDNA1 64 -#define NWARPS_Q4_1_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q4_1_AMPERE 4 -#define MMQ_Y_Q4_1_AMPERE 32 -#define NWARPS_Q4_1_AMPERE 4 -#else -#define MMQ_X_Q4_1_AMPERE 64 -#define MMQ_Y_Q4_1_AMPERE 128 -#define NWARPS_Q4_1_AMPERE 4 -#endif -#define MMQ_X_Q4_1_PASCAL 64 -#define MMQ_Y_Q4_1_PASCAL 64 -#define NWARPS_Q4_1_PASCAL 8 - -template static void - mul_mat_q4_1( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_1, - sycl::half2 *tile_x_dm_q4_1, int *tile_y_qs, sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q4_1_AMPERE; - const int mmq_y = MMQ_Y_Q4_1_AMPERE; - const int nwarps = NWARPS_Q4_1_AMPERE; - allocate_tiles_q4_1(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_qs_q4_1, tile_x_dm_q4_1); - mul_mat_q, VDR_Q4_1_Q8_1_MMQ, - vec_dot_q4_1_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q5_0_RDNA2 64 -#define MMQ_Y_Q5_0_RDNA2 128 -#define NWARPS_Q5_0_RDNA2 8 -#define MMQ_X_Q5_0_RDNA1 64 -#define MMQ_Y_Q5_0_RDNA1 64 -#define NWARPS_Q5_0_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q5_0_AMPERE 4 -#define MMQ_Y_Q5_0_AMPERE 32 -#define NWARPS_Q5_0_AMPERE 4 -#else -#define MMQ_X_Q5_0_AMPERE 128 -#define MMQ_Y_Q5_0_AMPERE 64 -#define NWARPS_Q5_0_AMPERE 4 -#endif -#define MMQ_X_Q5_0_PASCAL 64 -#define MMQ_Y_Q5_0_PASCAL 64 -#define NWARPS_Q5_0_PASCAL 8 - -template static void - mul_mat_q5_0( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_0, float *tile_x_d_q5_0, - int *tile_y_qs, sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q5_0_AMPERE; - const int mmq_y = MMQ_Y_Q5_0_AMPERE; - const int nwarps = NWARPS_Q5_0_AMPERE; - allocate_tiles_q5_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_ql_q5_0, tile_x_d_q5_0); - mul_mat_q, VDR_Q5_0_Q8_1_MMQ, - vec_dot_q5_0_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q5_1_RDNA2 64 -#define MMQ_Y_Q5_1_RDNA2 128 -#define NWARPS_Q5_1_RDNA2 8 -#define MMQ_X_Q5_1_RDNA1 64 -#define MMQ_Y_Q5_1_RDNA1 64 -#define NWARPS_Q5_1_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q5_1_AMPERE 4 -#define MMQ_Y_Q5_1_AMPERE 32 -#define NWARPS_Q5_1_AMPERE 4 -#else -#define MMQ_X_Q5_1_AMPERE 128 -#define MMQ_Y_Q5_1_AMPERE 64 -#define NWARPS_Q5_1_AMPERE 4 -#endif -#define MMQ_X_Q5_1_PASCAL 64 -#define MMQ_Y_Q5_1_PASCAL 64 -#define NWARPS_Q5_1_PASCAL 8 - -template static void -mul_mat_q5_1( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_1, - sycl::half2 *tile_x_dm_q5_1, int *tile_y_qs, sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q5_1_AMPERE; - const int mmq_y = MMQ_Y_Q5_1_AMPERE; - const int nwarps = NWARPS_Q5_1_AMPERE; - allocate_tiles_q5_1(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_ql_q5_1, tile_x_dm_q5_1); - mul_mat_q, VDR_Q5_1_Q8_1_MMQ, - vec_dot_q5_1_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q8_0_RDNA2 64 -#define MMQ_Y_Q8_0_RDNA2 128 -#define NWARPS_Q8_0_RDNA2 8 -#define MMQ_X_Q8_0_RDNA1 64 -#define MMQ_Y_Q8_0_RDNA1 64 -#define NWARPS_Q8_0_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q8_0_AMPERE 4 -#define MMQ_Y_Q8_0_AMPERE 32 -#define NWARPS_Q8_0_AMPERE 4 -#else -#define MMQ_X_Q8_0_AMPERE 128 -#define MMQ_Y_Q8_0_AMPERE 64 -#define NWARPS_Q8_0_AMPERE 4 -#endif -#define MMQ_X_Q8_0_PASCAL 64 -#define MMQ_Y_Q8_0_PASCAL 64 -#define NWARPS_Q8_0_PASCAL 8 - -template static void - mul_mat_q8_0( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q8_0, float *tile_x_d_q8_0, - int *tile_y_qs, sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q8_0_AMPERE; - const int mmq_y = MMQ_Y_Q8_0_AMPERE; - const int nwarps = NWARPS_Q8_0_AMPERE; - allocate_tiles_q8_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_qs_q8_0, tile_x_d_q8_0); - mul_mat_q, VDR_Q8_0_Q8_1_MMQ, - vec_dot_q8_0_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q2_K_RDNA2 64 -#define MMQ_Y_Q2_K_RDNA2 128 -#define NWARPS_Q2_K_RDNA2 8 -#define MMQ_X_Q2_K_RDNA1 128 -#define MMQ_Y_Q2_K_RDNA1 32 -#define NWARPS_Q2_K_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q2_K_AMPERE 4 -#define MMQ_Y_Q2_K_AMPERE 32 -#define NWARPS_Q2_K_AMPERE 4 -#else -#define MMQ_X_Q2_K_AMPERE 64 -#define MMQ_Y_Q2_K_AMPERE 128 -#define NWARPS_Q2_K_AMPERE 4 -#endif -#define MMQ_X_Q2_K_PASCAL 64 -#define MMQ_Y_Q2_K_PASCAL 64 -#define NWARPS_Q2_K_PASCAL 8 - -template static void -mul_mat_q2_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q2_K, - sycl::half2 *tile_x_dm_q2_K, int *tile_x_sc_q2_K, int *tile_y_qs, - sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q2_K_AMPERE; - const int mmq_y = MMQ_Y_Q2_K_AMPERE; - const int nwarps = NWARPS_Q2_K_AMPERE; - allocate_tiles_q2_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_ql_q2_K, tile_x_dm_q2_K, tile_x_sc_q2_K); - mul_mat_q, VDR_Q2_K_Q8_1_MMQ, - vec_dot_q2_K_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q3_K_RDNA2 128 -#define MMQ_Y_Q3_K_RDNA2 64 -#define NWARPS_Q3_K_RDNA2 8 -#define MMQ_X_Q3_K_RDNA1 32 -#define MMQ_Y_Q3_K_RDNA1 128 -#define NWARPS_Q3_K_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q3_K_AMPERE 4 -#define MMQ_Y_Q3_K_AMPERE 32 -#define NWARPS_Q3_K_AMPERE 4 -#else -#define MMQ_X_Q3_K_AMPERE 128 -#define MMQ_Y_Q3_K_AMPERE 128 -#define NWARPS_Q3_K_AMPERE 4 -#endif -#define MMQ_X_Q3_K_PASCAL 64 -#define MMQ_Y_Q3_K_PASCAL 64 -#define NWARPS_Q3_K_PASCAL 8 - -template static void -mul_mat_q3_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q3_K, - sycl::half2 *tile_x_dm_q3_K, int *tile_x_qh_q3_K, int *tile_x_sc_q3_K, - int *tile_y_qs, sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q3_K_AMPERE; - const int mmq_y = MMQ_Y_Q3_K_AMPERE; - const int nwarps = NWARPS_Q3_K_AMPERE; - allocate_tiles_q3_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_ql_q3_K, tile_x_dm_q3_K, tile_x_qh_q3_K, - tile_x_sc_q3_K); - mul_mat_q, VDR_Q3_K_Q8_1_MMQ, - vec_dot_q3_K_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q4_K_RDNA2 64 -#define MMQ_Y_Q4_K_RDNA2 128 -#define NWARPS_Q4_K_RDNA2 8 -#define MMQ_X_Q4_K_RDNA1 32 -#define MMQ_Y_Q4_K_RDNA1 64 -#define NWARPS_Q4_K_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q4_K_AMPERE 4 -#define MMQ_Y_Q4_K_AMPERE 32 -#define NWARPS_Q4_K_AMPERE 4 -#else -#define MMQ_X_Q4_K_AMPERE 64 -#define MMQ_Y_Q4_K_AMPERE 128 -#define NWARPS_Q4_K_AMPERE 4 -#endif -#define MMQ_X_Q4_K_PASCAL 64 -#define MMQ_Y_Q4_K_PASCAL 64 -#define NWARPS_Q4_K_PASCAL 8 - -template static void - mul_mat_q4_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q4_K, - sycl::half2 *tile_x_dm_q4_K, int *tile_x_sc_q4_K, int *tile_y_qs, - sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q4_K_AMPERE; - const int mmq_y = MMQ_Y_Q4_K_AMPERE; - const int nwarps = NWARPS_Q4_K_AMPERE; - allocate_tiles_q4_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_ql_q4_K, tile_x_dm_q4_K, tile_x_sc_q4_K); - mul_mat_q, VDR_Q4_K_Q8_1_MMQ, - vec_dot_q4_K_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q5_K_RDNA2 64 -#define MMQ_Y_Q5_K_RDNA2 128 -#define NWARPS_Q5_K_RDNA2 8 -#define MMQ_X_Q5_K_RDNA1 32 -#define MMQ_Y_Q5_K_RDNA1 64 -#define NWARPS_Q5_K_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q5_K_AMPERE 4 -#define MMQ_Y_Q5_K_AMPERE 32 -#define NWARPS_Q5_K_AMPERE 4 -#else -#define MMQ_X_Q5_K_AMPERE 64 -#define MMQ_Y_Q5_K_AMPERE 128 -#define NWARPS_Q5_K_AMPERE 4 -#endif -#define MMQ_X_Q5_K_PASCAL 64 -#define MMQ_Y_Q5_K_PASCAL 64 -#define NWARPS_Q5_K_PASCAL 8 - -template static void -mul_mat_q5_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_K, - sycl::half2 *tile_x_dm_q5_K, int *tile_x_sc_q5_K, int *tile_y_qs, - sycl::half2 *tile_y_ds) { - int * tile_x_ql = nullptr; - sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q5_K_AMPERE; - const int mmq_y = MMQ_Y_Q5_K_AMPERE; - const int nwarps = NWARPS_Q5_K_AMPERE; - allocate_tiles_q5_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_ql_q5_K, tile_x_dm_q5_K, tile_x_sc_q5_K); - mul_mat_q, VDR_Q5_K_Q8_1_MMQ, - vec_dot_q5_K_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -#define MMQ_X_Q6_K_RDNA2 64 -#define MMQ_Y_Q6_K_RDNA2 128 -#define NWARPS_Q6_K_RDNA2 8 -#define MMQ_X_Q6_K_RDNA1 32 -#define MMQ_Y_Q6_K_RDNA1 64 -#define NWARPS_Q6_K_RDNA1 8 -#if defined(SYCL_USE_XMX) -#define MMQ_X_Q6_K_AMPERE 4 -#define MMQ_Y_Q6_K_AMPERE 32 -#define NWARPS_Q6_K_AMPERE 4 -#else -#define MMQ_X_Q6_K_AMPERE 64 -#define MMQ_Y_Q6_K_AMPERE 64 -#define NWARPS_Q6_K_AMPERE 4 -#endif -#define MMQ_X_Q6_K_PASCAL 64 -#define MMQ_Y_Q6_K_PASCAL 64 -#define NWARPS_Q6_K_PASCAL 8 - -template static void - mul_mat_q6_K( - const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, - const sycl::nd_item<3> &item_ct1, int *tile_x_ql, sycl::half2 *tile_x_dm, - int *tile_x_sc, int *tile_y_qs, sycl::half2 *tile_y_ds) { - // int * tile_x_ql = nullptr; - // sycl::half2 *tile_x_dm = nullptr; - int * tile_x_qh = nullptr; - // int * tile_x_sc = nullptr; - -//sycl_todo: change according to hardware - const int mmq_x = MMQ_X_Q6_K_AMPERE; - const int mmq_y = MMQ_Y_Q6_K_AMPERE; - const int nwarps = NWARPS_Q6_K_AMPERE; - allocate_tiles_q6_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, - tile_x_ql, tile_x_dm, tile_x_sc); - mul_mat_q, VDR_Q6_K_Q8_1_MMQ, - vec_dot_q6_K_q8_1_mul_mat>( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, - tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); -} - -template -static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - - const int qi_vdr = (qi / vdr); // N_threads processing 1 qk block - - // partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / qi_vdr; i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row * blocks_per_row + i; // x block index - - const int iby = i * (qk / QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) - - i * qi_vdr); // x block quant index when casting the quants to int - - tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq2_xxs_q8_1(&x[ibx], &y[iby], iqs, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq2_xs_q8_1(&x[ibx], &y[iby], iqs, iq2xs_grid, ksigns64); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq2_s_q8_1(&x[ibx], &y[iby], iqs); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq3_xxs_q8_1(&x[ibx], &y[iby], iqs, iq3xxs_grid, ksigns64); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq3_s_q8_1(&x[ibx], &y[iby], iqs, iq3s_grid); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq1_s_q8_1(&x[ibx], &y[iby], iqs, iq1s_grid_gpu); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq1_m_q8_1(&x[ibx], &y[iby], iqs); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - -template -static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq4_nl_q8_1(&x[ibx], &y[iby], iqs); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - - -template -static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, - const void *__restrict__ vy, - float *__restrict__ dst, const int ncols, - const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * WARP_SIZE / qi; - -// partial sum for each thread - float tmp = 0.0f; - - const block_q_t * x = (const block_q_t *) vx; - const block_q8_1 * y = (const block_q8_1 *) vy; - - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index - - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx - - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int - - tmp += vec_dot_iq4_xs_q8_1(&x[ibx], &y[iby], iqs); - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[row] = tmp; - } -} - - -template -static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows, - const sycl::nd_item<3> &item_ct1) { - // qk = quantized weights per x block - // qr = number of quantized weights per data value in x block - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); - - if (row >= nrows) { - return; - } - - const int tid = item_ct1.get_local_id(2); - - const int iter_stride = 2*GGML_SYCL_DMMV_X; - const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter - const int y_offset = qr == 1 ? 1 : qk/2; - -// partial sum for each thread -#ifdef GGML_SYCL_F16 - sycl::half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics -#else - float tmp = 0.0f; -#endif // GGML_SYCL_F16 - - for (int i = 0; i < ncols; i += iter_stride) { - const int col = i + vals_per_iter*tid; - const int ib = (row*ncols + col)/qk; // x block index - const int iqs = (col%qk)/qr; // x quant index - const int iybs = col - col%qk; // y block start index - -// processing >2 values per i iter is faster for fast GPUs -#pragma unroll - for (int j = 0; j < vals_per_iter; j += 2) { - // process 2 vals per j iter - - // dequantize - // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val - dfloat2 v; - dequantize_kernel(vx, ib, iqs + j/qr, v); - - // matrix multiplication - // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2 -#ifdef GGML_SYCL_F16 - dfloat2 t1{y[iybs + iqs + j / qr + 0], - y[iybs + iqs + j / qr + y_offset]}; - - tmp += v * t1; -#else - tmp += v.x() * y[iybs + iqs + j / qr + 0]; - tmp += v.y() * y[iybs + iqs + j / qr + y_offset]; -#endif // GGML_SYCL_F16 - } - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (tid == 0) { -#ifdef GGML_SYCL_F16 - dst[row] = tmp.x() + tmp.y(); -#else - dst[row] = tmp; -#endif // GGML_SYCL_F16 - } -} - -static void mul_mat_p021_f16_f32( - const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, - const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y, - const sycl::nd_item<3> &item_ct1) { - - const sycl::half *x = (const sycl::half *)vx; - - const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) + - item_ct1.get_local_id(0); - const int channel_x = channel / (nchannels_y / nchannels_x); - - const int nrows_y = ncols_x; - const int nrows_dst = nrows_x; - const int row_dst = row_x; - - float tmp = 0.0f; - - for (int col_x0 = 0; col_x0 < ncols_x; - col_x0 += item_ct1.get_local_range(2)) { - const int col_x = col_x0 + item_ct1.get_local_id(2); - - if (col_x >= ncols_x) { - break; - } - - // x is transposed and permuted - const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x; - const float xi = - sycl::vec(x[ix]) - .convert()[0]; - - const int row_y = col_x; - - - // y is not transposed but permuted - const int iy = channel*nrows_y + row_y; - - tmp += xi * y[iy]; - } - - // dst is not transposed and not permuted - const int idst = channel*nrows_dst + row_dst; - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[idst] = tmp; - } -} - -static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous - const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x, - const int row_stride_x, const int channel_stride_x, const int channel_x_divisor, - const sycl::nd_item<3> &item_ct1) { - - const sycl::half *x = (const sycl::half *)vx; - - const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) + - item_ct1.get_local_id(0); - const int channel_x = channel / channel_x_divisor; - - const int nrows_y = ncols_x; - const int nrows_dst = nrows_x; - const int row_dst = row_x; - - const int idst = channel*nrows_dst + row_dst; - - float tmp = 0.0f; - - for (int col_x0 = 0; col_x0 < ncols_x; - col_x0 += item_ct1.get_local_range(2)) { - const int col_x = col_x0 + item_ct1.get_local_id(2); - - if (col_x >= ncols_x) { - break; - } - - const int row_y = col_x; - - const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x; - const int iy = channel*nrows_y + row_y; - - const float xi = - sycl::vec(x[ix]) - .convert()[0]; - - tmp += xi * y[iy]; - } - - // sum up partial sums and write back result -#pragma unroll - for (int mask = 16; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); - } - - if (item_ct1.get_local_id(2) == 0) { - dst[idst] = tmp; - } -} - -static void cpy_1_f32_f32(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -static void cpy_1_f32_f16(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - sycl::half *dsti = (sycl::half *)cdsti; - - *dsti = sycl::vec(*xi) - .convert()[0]; -} - -static void cpy_1_f16_f16(const char * cxi, char * cdsti) { - const sycl::half *xi = (const sycl::half *)cxi; - sycl::half *dsti = (sycl::half *)cdsti; - - *dsti = *xi; -} - -static void cpy_1_f16_f32(const char * cxi, char * cdsti) { - const sycl::half *xi = (const sycl::half *)cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -static void cpy_1_i16_i16(const char * cxi, char * cdsti) { - const int16_t *xi = (const int16_t *)cxi; - int16_t *dsti = (int16_t *)cdsti; - - *dsti = *xi; -} - -static void cpy_1_i32_i32(const char * cxi, char * cdsti) { - const int32_t *xi = (const int32_t *)cxi; - int32_t *dsti = (int32_t *)cdsti; - - *dsti = *xi; -} - -template -static void cpy_f32_f16(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (i >= ne) { - return; - } - - // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor - // then combine those indices with the corresponding byte offsets to get the total offsets - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; - - cpy_1(cx + x_offset, cdst + dst_offset); -} - -static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q8_0 * dsti = (block_q8_0 *) cdsti; - - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - const float v = xi[j]; - amax = sycl::fmax(amax, sycl::fabs((float)v)); - } - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK8_0; ++j) { - const float x0 = xi[j]*id; - - dsti->qs[j] = sycl::round((float)x0); - } -} - -static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_0 * dsti = (block_q4_0 *) cdsti; - - float amax = 0.0f; - float vmax = 0.0f; - - for (int j = 0; j < QK4_0; ++j) { - const float v = xi[j]; - if (amax < sycl::fabs((float)v)) { - amax = sycl::fabs((float)v); - vmax = v; - } - } - - const float d = vmax / -8; - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK4_0/2; ++j) { - const float x0 = xi[0 + j]*id; - const float x1 = xi[QK4_0/2 + j]*id; - - const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 8.5f)); - const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 8.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_1 * dsti = (block_q4_1 *) cdsti; - - float vmin = FLT_MAX; - float vmax = -FLT_MAX; - - for (int j = 0; j < QK4_1; ++j) { - const float v = xi[j]; - - if (v < vmin) vmin = v; - if (v > vmax) vmax = v; - } - - const float d = (vmax - vmin) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->dm.x() = d; - dsti->dm.y() = vmin; - - for (int j = 0; j < QK4_1/2; ++j) { - const float x0 = (xi[0 + j] - vmin)*id; - const float x1 = (xi[QK4_1/2 + j] - vmin)*id; - - const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 0.5f)); - const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 0.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -template -static void cpy_f32_q(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { - const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2)) * - qk; - - if (i >= ne) { - return; - } - - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13; - - cpy_blck(cx + x_offset, cdst + dst_offset); -} - -static float rope_yarn_ramp(const float low, const float high, const int i0) { - const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); - return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); -} - -struct rope_corr_dims { - float v[4]; -}; - -// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn -// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. -static void rope_yarn( - float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, - float * cos_theta, float * sin_theta -) { - // Get n-d rotational scaling corrected for extrapolation - float theta_interp = freq_scale * theta_extrap; - float theta = theta_interp; - if (ext_factor != 0.0f) { - float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; - theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; - - // Get n-d magnitude scaling corrected for interpolation - mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale); - } - *cos_theta = sycl::cos(theta) * mscale; - *sin_theta = sycl::sin(theta) * mscale; -} - -// rope == RoPE == rotary positional embedding -template -static void rope( - const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, - float ext_factor, float attn_factor, rope_corr_dims corr_dims -, - const sycl::nd_item<3> &item_ct1) { - const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1)); - - if (col >= ncols) { - return; - } - - const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - const int i = row*ncols + col; - const int i2 = row/p_delta_rows; - - const int p = has_pos ? pos[i2] : 0; - const float theta_base = p * dpct::pow(freq_base, -float(col) / ncols); - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + 1]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + 1] = x0*sin_theta + x1*cos_theta; -} - -template -static void rope_neox( - const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, - float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, - const float * freq_factors, const sycl::nd_item<3> &item_ct1) { - const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1)); - - if (col >= ncols) { - return; - } - - const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - const int ib = col / n_dims; - const int ic = col % n_dims; - - if (ib > 0) { - const int i = row*ncols + ib*n_dims + ic; - - dst[i + 0] = x[i + 0]; - dst[i + 1] = x[i + 1]; - - return; - } - - const int i = row*ncols + ib*n_dims + ic/2; - const int i2 = row/p_delta_rows; - - float cur_rot = inv_ndims * ic - ib; - - const int p = has_pos ? pos[i2] : 0; - const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f; - - const float theta_base = - p * freq_scale * dpct::pow(theta_scale, col / 2.0f)/freq_factor; - - float cos_theta, sin_theta; - rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); - - const float x0 = x[i + 0]; - const float x1 = x[i + n_dims/2]; - - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; -} - -static void k_sum_rows_f32(const float * x, float * dst, const int ncols, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(1); - const int col = item_ct1.get_local_id(2); - - float sum = 0.0f; - for (int i = col; i < ncols; i += item_ct1.get_local_range(2)) { - sum += x[row * ncols + i]; - } - - sum = warp_reduce_sum(sum, item_ct1); - - if (col == 0) { - dst[row] = sum; - } -} - - -template -static inline void ggml_sycl_swap(T & a, T & b) { - T tmp = a; - a = b; - b = tmp; -} - -template -__dpct_inline__ static void -k_argsort_f32_i32(const float *x, int *dst, const int ncols, int ncols_pad, - const sycl::nd_item<3> &item_ct1, uint8_t *dpct_local) { - // bitonic sort - int col = item_ct1.get_local_id(2); - int row = item_ct1.get_group(1); - - if (col >= ncols_pad) { - return; - } - - const float * x_row = x + row * ncols; - auto dst_row = (int *)dpct_local; - - // initialize indices - dst_row[col] = col; - - item_ct1.barrier(sycl::access::fence_space::local_space); - - for (int k = 2; k <= ncols_pad; k *= 2) { - for (int j = k / 2; j > 0; j /= 2) { - int ixj = col ^ j; - if (ixj > col) { - if ((col & k) == 0) { - if (dst_row[col] >= ncols || - (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ? - x_row[dst_row[col]] > x_row[dst_row[ixj]] : - x_row[dst_row[col]] < x_row[dst_row[ixj]])) - ) { - ggml_sycl_swap(dst_row[col], dst_row[ixj]); - } - } else { - if (dst_row[ixj] >= ncols || - (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ? - x_row[dst_row[col]] < x_row[dst_row[ixj]] : - x_row[dst_row[col]] > x_row[dst_row[ixj]])) - ) { - ggml_sycl_swap(dst_row[col], dst_row[ixj]); - } - } - } - /* - DPCT1118:1: SYCL group functions and algorithms must be encountered - in converged control flow. You may need to adjust the code. - */ - item_ct1.barrier(sycl::access::fence_space::local_space); - } - } - - // copy the result to dst without the padding - if (col < ncols) { - dst[row * ncols + col] = dst_row[col]; - } -} - - -static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past, - const sycl::nd_item<3> &item_ct1) { - const int col = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (col >= ncols) { - return; - } - - const int i = row*ncols + col; - //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i]; - //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU - dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX; -} - - -template -static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par, - const int nrows_y, const float scale, const float max_bias, const float m0, - const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) { - const int ncols = ncols_template == 0 ? ncols_par : ncols_template; - - const int tid = item_ct1.get_local_id(2); - const int rowx = item_ct1.get_group(2); - const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension - - const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template; - - const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; - const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; - - float slope = 1.0f; - - // ALiBi - if (max_bias > 0.0f) { - const uint32_t h = rowx/nrows_y; // head index - - const float base = h < n_head_log2 ? m0 : m1; - const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; - - slope = sycl::pow(base, float(exp)); - } - - float * vals = vals_smem ? buf + WARP_SIZE : dst + rowx*ncols; - float max_val = -INFINITY; - - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - - if (ncols_template == 0 && col >= ncols) { - break; - } - - const int ix = rowx*ncols + col; - const int iy = rowy*ncols + col; - - const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f); - - vals[col] = val; - max_val = sycl::max(max_val, val); - } - - // find the max value in the block - max_val = warp_reduce_max(max_val, item_ct1); - if (block_size > WARP_SIZE) { - if (warp_id == 0) { - buf[lane_id] = -INFINITY; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - if (lane_id == 0) { - buf[warp_id] = max_val; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - max_val = buf[lane_id]; - max_val = warp_reduce_max(max_val, item_ct1); - } - - float tmp = 0.f; - -#pragma unroll - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - if (ncols_template == 0 && col >= ncols) { - break; - } - - const float val = sycl::native::exp(vals[col] - max_val); - tmp += val; - vals[col] = val; - } - - // find the sum of exps in the block - tmp = warp_reduce_sum(tmp, item_ct1); - if (block_size > WARP_SIZE) { - item_ct1.barrier(sycl::access::fence_space::local_space); - if (warp_id == 0) { - buf[lane_id] = 0.f; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - if (lane_id == 0) { - buf[warp_id] = tmp; - } - item_ct1.barrier(sycl::access::fence_space::local_space); - - tmp = buf[lane_id]; - tmp = warp_reduce_sum(tmp, item_ct1); - } - - const float inv_sum = 1.f / tmp; - -#pragma unroll - for (int col0 = 0; col0 < ncols; col0 += block_size) { - const int col = col0 + tid; - - if (ncols_template == 0 && col >= ncols) { - return; - } - - const int idst = rowx*ncols + col; - dst[idst] = vals[col] * inv_sum; - } -} - -static void scale_f32(const float * x, float * dst, const float scale, const int k, - const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (i >= k) { - return; - } - - dst[i] = scale * x[i]; -} - -static void clamp_f32(const float * x, float * dst, const float min, const float max, const int k, - const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (i >= k) { - return; - } - - dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); -} - -template -static void im2col_kernel(const float *x, T *dst, int offset_delta, - int IW, int IH, int OW, int KW, int KH, - int pelements, int CHW, int s0, int s1, int p0, - int p1, int d0, int d1, - const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_id(2) + - item_ct1.get_group(2) * item_ct1.get_local_range(2); - if (i >= pelements) { - return; - } - - const int ksize = OW * (KH > 1 ? KW : 1); - const int kx = i / ksize; - const int kd = kx * ksize; - const int ky = (i - kd) / OW; - const int ix = i % OW; - - const int64_t iiw = ix * s0 + kx * d0 - p0; - const int64_t iih = item_ct1.get_group(1) * s1 + ky * d1 - p1; - - const int64_t offset_dst = - (item_ct1.get_group(1) * OW + ix) * CHW + - (item_ct1.get_group(0) * (KW * KH) + ky * KW + kx); - - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { - dst[offset_dst] = - sycl::vec(0.0f) - .convert()[0]; - } else { - const int64_t offset_src = item_ct1.get_group(0) * offset_delta; - dst[offset_dst] = - sycl::vec(x[offset_src + iih * IW + iiw]) - .convert()[0]; - } -} - -template -static void pool2d_nchw_kernel( - const int ih, const int iw, const int oh, const int ow, - const int kh, const int kw, const int sh, const int sw, - const int ph, const int pw, const int parallel_elements, - const Ti* src, To* dst, const enum ggml_op_pool op, - const sycl::nd_item<3> &item_ct1) { - int idx = item_ct1.get_local_id(2) + - item_ct1.get_group(2) * item_ct1.get_local_range(2); - if (idx >= parallel_elements) { - return; - } - - const int I_HW = ih * iw; - const int O_HW = oh * ow; - const int nc = idx / O_HW; - const int cur_oh = idx % O_HW / ow; - const int cur_ow = idx % O_HW % ow; - const Ti* i_ptr = src + nc * I_HW; - To* o_ptr = dst + nc * O_HW; - const int start_h = cur_oh * sh - ph; - const int bh = sycl::max(0, start_h); - const int eh = sycl::min(ih, start_h + kh); - const int start_w = cur_ow * sw - pw; - const int bw = sycl::max(0, start_w); - const int ew = sycl::min(iw, start_w + kw); - - To res = 0; - - switch (op) { - case GGML_OP_POOL_AVG: res = 0; break; - case GGML_OP_POOL_MAX: res = -FLT_MAX; break; - } - - for (int i = bh; i < eh; i += 1) { - for (int j = bw; j < ew; j += 1) { -#if DPCT_COMPATIBILITY_TEMP >= 350 - /* - DPCT1098:106: The '*' expression is used instead of the __ldg - call. These two expressions do not provide the exact same - functionality. Check the generated code for potential precision - and/or performance issues. - */ - Ti cur = *(i_ptr + i * iw + j); -#else - Ti cur = i_ptr[i * iw + j]; -#endif - switch (op) { - case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break; - case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break; - } - } - } - o_ptr[cur_oh * ow + cur_ow] = res; -} - -template -static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const void *src0_dd, - const int32_t *src1_dd, float *dst_dd, - queue_ptr stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); - const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE); - const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); - - // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); - - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); - - 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); - }); - - (void) dst; -} - -template -static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const src0_t *src0_dd, const int32_t *src1_dd, - float *dst_dd, queue_ptr stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); - const int block_num_x = (ne00 + SYCL_GET_ROWS_BLOCK_SIZE - 1) / SYCL_GET_ROWS_BLOCK_SIZE; - const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); - - // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); - - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); - - { - 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) { - k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, - s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); - }); - } - - (void) dst; -} - -template -struct bin_bcast_sycl { - template - void operator()(ggml_backend_sycl_context & ctx, - const struct ggml_tensor *src0, - const struct ggml_tensor *src1, struct ggml_tensor *dst, - const src0_t *src0_dd, const src1_t *src1_dd, dst_t *dst_dd, - queue_ptr stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - int nr0 = ne10/ne0; - int nr1 = ne11/ne1; - int nr2 = ne12/ne2; - int nr3 = ne13/ne3; - - int nr[4] = { nr0, nr1, nr2, nr3 }; - - // collapse dimensions until first broadcast dimension - int64_t cne0[] = {ne0, ne1, ne2, ne3}; - int64_t cne1[] = {ne10, ne11, ne12, ne13}; - size_t cnb0[] = {nb0, nb1, nb2, nb3}; - size_t cnb1[] = {nb10, nb11, nb12, nb13}; - auto collapse = [](int64_t cne[]) { - cne[0] *= cne[1]; - cne[1] = cne[2]; - cne[2] = cne[3]; - cne[3] = 1; - }; - - auto collapse_nb = [](size_t cnb[], int64_t cne[]) { - cnb[1] *= cne[1]; - cnb[2] *= cne[2]; - cnb[3] *= cne[3]; - }; - - for (int i = 0; i < 4; i++) { - if (nr[i] != 1) { - break; - } - if (i > 0) { - collapse_nb(cnb0, cne0); - collapse_nb(cnb1, cne1); - collapse(cne0); - collapse(cne1); - } - } - { - int64_t ne0 = cne0[0]; - int64_t ne1 = cne0[1]; - int64_t ne2 = cne0[2]; - int64_t ne3 = cne0[3]; - - int64_t ne10 = cne1[0]; - int64_t ne11 = cne1[1]; - int64_t ne12 = cne1[2]; - int64_t ne13 = cne1[3]; - - size_t nb0 = cnb0[0]; - size_t nb1 = cnb0[1]; - size_t nb2 = cnb0[2]; - size_t nb3 = cnb0[3]; - - size_t nb10 = cnb1[0]; - size_t nb11 = cnb1[1]; - size_t nb12 = cnb1[2]; - size_t nb13 = cnb1[3]; - - size_t s0 = nb0 / sizeof(dst_t); - size_t s1 = nb1 / sizeof(dst_t); - size_t s2 = nb2 / sizeof(dst_t); - size_t s3 = nb3 / sizeof(dst_t); - - size_t s10 = nb10 / sizeof(src1_t); - size_t s11 = nb11 / sizeof(src1_t); - size_t s12 = nb12 / sizeof(src1_t); - size_t s13 = nb13 / sizeof(src1_t); - - GGML_ASSERT(s0 == 1); - GGML_ASSERT(s10 == 1); - - const int block_size = 128; - - int64_t hne0 = std::max(ne0/2LL, 1LL); - - sycl::range<3> block_dims(1, 1, 1); - block_dims[2] = std::min(hne0, block_size); - block_dims[1] = std::min( - ne1, block_size / (unsigned int)block_dims[2]); - block_dims[0] = std::min( - std::min( - ne2 * ne3, block_size / (unsigned int)block_dims[2] / - (unsigned int)block_dims[1]), - 64U); - - sycl::range<3> block_nums( - (ne2 * ne3 + block_dims[0] - 1) / block_dims[0], - (ne1 + block_dims[1] - 1) / block_dims[1], - (hne0 + block_dims[2] - 1) / block_dims[2]); - - if (block_nums[0] > 65535) { - // this is the maximum number of blocks in z direction, fallback to 1D grid kernel - int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size; - { - 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::range<3>(1, 1, block_size)), - [=](sycl::nd_item<3> item_ct1) { - k_bin_bcast_unravel( - src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, - ne10, ne11, ne12, ne13, s1, s2, s3, s11, s12, - s13, item_ct1); - }); - } - } else { - /* - DPCT1049:16: 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. - */ - 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) { - k_bin_bcast(src0_dd, src1_dd, dst_dd, ne0, ne1, - ne2, ne3, ne10, ne11, ne12, ne13, - s1, s2, s3, s11, s12, s13, - item_ct1); - }); - } - } - } -}; - -static void acc_f32_sycl(const float *x, const float *y, float *dst, - const int n_elements, const int ne10, const int ne11, - 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); - }); -} - -static void gelu_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); - }); -} - -static void silu_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); - }); -} - -static void gelu_quick_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); - }); -} - -static void tanh_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); - }); -} - -static void relu_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); - }); -} - -static void hardsigmoid_f32_sycl(const float *x, float *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::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - hardsigmoid_f32(x, dst, k, item_ct1); - }); -} - -static void hardswish_f32_sycl(const float *x, float *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::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - hardswish_f32(x, dst, k, item_ct1); - }); -} - -static void leaky_relu_f32_sycl(const float *x, float *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_f32(x, dst, k, negative_slope, item_ct1); - }); -} - -static void sqr_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); - }); -} - -static void norm_f32_sycl(const float *x, float *dst, const int ncols, - const int nrows, const float eps, - queue_ptr stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - if (ncols < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1( - sycl::range<1>(32), cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - // FIXME: 1024 from cuda - const int work_group_size = GROUP_SIZE; - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:17: 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->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1( - sycl::range<1>(32), cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - -static void group_norm_f32_sycl(const float *x, float *dst, - const int num_groups, const int group_size, - const int ne_elements, queue_ptr stream) { - static const float eps = 1e-6f; - if (group_size < 1024) { - const sycl::range<3> block_dims(1, 1, WARP_SIZE); - stream->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - 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) - [[intel::reqd_sub_group_size(32)]] { - group_norm_f32( - x, dst, group_size, ne_elements, eps_ct4, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = GROUP_SIZE; - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:18: 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->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - 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) - [[intel::reqd_sub_group_size(32)]] { - group_norm_f32(x, dst, group_size, ne_elements, - eps_ct4, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - -static void concat_f32_sycl(const float *x, const float *y, float *dst, - const int ne0, int ne1, int ne2, int ne02, - queue_ptr stream) { - int num_blocks = (ne0 + SYCL_CONCAT_BLOCK_SIZE - 1) / SYCL_CONCAT_BLOCK_SIZE; - sycl::range<3> gridDim(ne2, ne1, num_blocks); - 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(x, y, dst, ne0, ne02, item_ct1); - }); -} - -static void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01, - const int nb02, const int nb03, const int ne10, const int ne11, - const int ne12, const int ne13, const float sf0, const float sf1, - const float sf2, const float sf3, queue_ptr stream) { - 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) { - upscale_f32(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1); - }); -} - -static void pad_f32_sycl(const float *x, float *dst, const int ne00, - const int ne01, const int ne02, const int ne0, - 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_f32(x, dst, ne0, ne00, ne01, ne02, item_ct1); - }); -} - -static void rms_norm_f32_sycl(const float *x, float *dst, const int ncols, - const int nrows, const float eps, - queue_ptr stream) { - GGML_ASSERT(ncols % WARP_SIZE == 0); - // 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) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - rms_norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), WARP_SIZE); - }); - }); - } else { - const int work_group_size = GROUP_SIZE; - const sycl::range<3> block_dims(1, 1, work_group_size); - /* - DPCT1049:19: 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->submit([&](sycl::handler &cgh) { - sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), - cgh); - - cgh.parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, - block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - rms_norm_f32(x, dst, ncols, eps, item_ct1, - s_sum_acc_ct1.get_pointer(), work_group_size); - }); - }); - } -} - -static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx, - const int ky, const int kx_padded, - queue_ptr stream) { - const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE; - const sycl::range<3> num_blocks(1, ky, block_num_x); - const sycl::range<3> block_size(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE); - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(num_blocks * block_size, block_size), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - quantize_q8_1(x, vy, kx, kx_padded, item_ct1); - }); - } -} - -template -static void dequantize_block_sycl(const void *__restrict__ vx, - dst_t *__restrict__ y, const int k, - queue_ptr stream) { - const int num_blocks = (k + 2*SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / (2*SYCL_DEQUANTIZE_BLOCK_SIZE); - { - 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); - }); - } -} - -template -static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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); - }); - } -} - -template -static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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); - }); - } -} - -template -static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb32 = k / 32; - const int nb = (k + 255) / 256; - { - 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); - }); - } -} - -template -static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb32 = k / 32; - const int nb = (k + 255) / 256; - { - 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); - }); - } -} - - -template -static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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_K(vx, y, item_ct1); - }); - } -} - -template -static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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); - }); - } -} - -template -static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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); - }); - } -} - -template -static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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 - ); - }); - }); - } -} - -template -static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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 - ); - }); - }); - } -} - -template -static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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_xxs( - vx, y, item_ct1, iq2xxs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); - }); - } -} - -template -static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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); - }); - }); - } -} - -template -static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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); - }); - }); - } -} - - -template -static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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_xxs( - vx, y, item_ct1, iq3xxs_grid, - ksigns_iq2xs, kmask_iq2xs); - }); - }); - } -} - -template -static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = k / QK_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); - }); - }); - } -} - -template -static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = (k + QK_K - 1) / QK_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); - }); - }); - } -} - - -template -static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int k, - queue_ptr stream) { - const int nb = (k + QK_K - 1) / QK_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); - }); - }); - } -} - - - -template -static void convert_unary_sycl(const void *__restrict__ vx, - dst_t *__restrict__ y, const int k, - queue_ptr stream) { - const int num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE; - { - 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) { - convert_unary(vx, y, k, item_ct1); - }); - } -} - - -static to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) try { - int id; - switch (type) { - case GGML_TYPE_Q4_0: - return dequantize_block_sycl; - case GGML_TYPE_Q4_1: - return dequantize_block_sycl; - case GGML_TYPE_Q5_0: - return dequantize_block_sycl; - case GGML_TYPE_Q5_1: - return dequantize_block_sycl; - case GGML_TYPE_Q8_0: - return dequantize_block_sycl; - case GGML_TYPE_Q2_K: - return dequantize_row_q2_K_sycl; - case GGML_TYPE_Q3_K: - return dequantize_row_q3_K_sycl; - case GGML_TYPE_Q4_K: - return dequantize_row_q4_K_sycl; - case GGML_TYPE_Q5_K: - return dequantize_row_q5_K_sycl; - case GGML_TYPE_Q6_K: - return dequantize_row_q6_K_sycl; - case GGML_TYPE_IQ1_S: - return dequantize_row_iq1_s_sycl; - case GGML_TYPE_IQ1_M: - return dequantize_row_iq1_m_sycl; - case GGML_TYPE_IQ2_XXS: - return dequantize_row_iq2_xxs_sycl; - case GGML_TYPE_IQ2_XS: - return dequantize_row_iq2_xs_sycl; - case GGML_TYPE_IQ2_S: - return dequantize_row_iq2_s_sycl; - case GGML_TYPE_IQ3_XXS: - return dequantize_row_iq3_xxs_sycl; - case GGML_TYPE_IQ3_S: - return dequantize_row_iq3_s_sycl; - case GGML_TYPE_IQ4_XS: - return dequantize_row_iq4_xs_sycl; - case GGML_TYPE_IQ4_NL: - return dequantize_row_iq4_nl_sycl; - case GGML_TYPE_F32: - return convert_unary_sycl; - default: - return nullptr; - } -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -static to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) { - switch (type) { - case GGML_TYPE_Q4_0: - return dequantize_row_q4_0_sycl; - case GGML_TYPE_Q4_1: - return dequantize_row_q4_1_sycl; - case GGML_TYPE_Q5_0: - return dequantize_block_sycl; - case GGML_TYPE_Q5_1: - return dequantize_block_sycl; - case GGML_TYPE_Q8_0: - return dequantize_block_sycl; - case GGML_TYPE_Q2_K: - return dequantize_row_q2_K_sycl; - case GGML_TYPE_Q3_K: - return dequantize_row_q3_K_sycl; - case GGML_TYPE_Q4_K: - return dequantize_row_q4_K_sycl; - case GGML_TYPE_Q5_K: - return dequantize_row_q5_K_sycl; - case GGML_TYPE_Q6_K: - return dequantize_row_q6_K_sycl; - case GGML_TYPE_IQ1_S: - return dequantize_row_iq1_s_sycl; - case GGML_TYPE_IQ1_M: - return dequantize_row_iq1_m_sycl; - case GGML_TYPE_IQ2_XXS: - return dequantize_row_iq2_xxs_sycl; - case GGML_TYPE_IQ2_XS: - return dequantize_row_iq2_xs_sycl; - case GGML_TYPE_IQ2_S: - return dequantize_row_iq2_s_sycl; - case GGML_TYPE_IQ3_XXS: - return dequantize_row_iq3_xxs_sycl; - case GGML_TYPE_IQ3_S: - return dequantize_row_iq3_s_sycl; - case GGML_TYPE_IQ4_XS: - return dequantize_row_iq4_xs_sycl; - case GGML_TYPE_IQ4_NL: - return dequantize_row_iq4_nl_sycl; - case GGML_TYPE_F16: - return convert_unary_sycl; - default: - return nullptr; - } -} - -static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead - const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); - { - 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) [[intel::reqd_sub_group_size(32)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); - } -} - -static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); - { - 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) [[intel::reqd_sub_group_size(32)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); - } -} - -static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); - { - 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) [[intel::reqd_sub_group_size(32)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); - } -} - -static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); - { - 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) [[intel::reqd_sub_group_size(32)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); - } -} - -static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); - { - 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) [[intel::reqd_sub_group_size(32)]] { - dequantize_mul_mat_vec( - vx, y, dst, ncols, nrows, item_ct1); - }); - } -} - -static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 - 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, 32); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - 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, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2 / K_QUANTS_PER_ITERATION; - 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, 32); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - 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, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2 / K_QUANTS_PER_ITERATION; - 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, 32); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - 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, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const sycl::range<3> block_dims(1, 1, 32); - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - 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, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int ny = 2 / K_QUANTS_PER_ITERATION; - 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, 32); - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { - dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); - }); -} - -static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); - { - 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) [[intel::reqd_sub_group_size(32)]] { - dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, - nrows, item_ct1); - }); - } -} - - -static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK4_0 == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK4_1 == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK5_0 == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK5_1 == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK8_0 == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - - -static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_xxs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) { - auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0]; - auto ksigns64_ptr_ct1 = &ksigns64[0]; - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_xs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) { - auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0]; - auto ksigns64_ptr_ct1 = &ksigns64[0]; - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq2_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) { - auto iq3xxs_grid_ptr_ct1 = &iq3xxs_grid[0]; - auto ksigns64_ptr_ct1 = &ksigns64[0]; - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq3_xxs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) { - auto iq3s_grid_ptr_ct1 = &iq3s_grid[0]; - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq3_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) { - auto iq1s_grid_ptr_ct1 = &iq1s_grid_gpu[0]; - auto ksigns64_ptr_ct1 = &ksigns64[0]; - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq1_s_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq1_m_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK4_NL == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq4_nl_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, - queue_ptr stream) { - GGML_ASSERT(ncols % QK_K == 0); - const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; - 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) - [[intel::reqd_sub_group_size(32)]] { - mul_mat_vec_q_iq4_xs_q8_1( - vx, vy, dst, ncols, nrows, item_ct1); - }); - }); - } -} - -static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { - - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q4_0_RDNA2; - mmq_y = MMQ_Y_Q4_0_RDNA2; - nwarps = NWARPS_Q4_0_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q4_0_RDNA1; - mmq_y = MMQ_Y_Q4_0_RDNA1; - nwarps = NWARPS_Q4_0_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q4_0_AMPERE; - mmq_y = MMQ_Y_Q4_0_AMPERE; - nwarps = NWARPS_Q4_0_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q4_0_PASCAL; - mmq_y = MMQ_Y_Q4_0_PASCAL; - nwarps = NWARPS_Q4_0_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:20: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q4_0( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_qs_q4_0_acc_ct1.get_pointer(), - tile_x_d_q4_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } else { - const bool need_check = true; - /* - DPCT1049:21: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q4_0( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_qs_q4_0_acc_ct1.get_pointer(), - tile_x_d_q4_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { - - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q4_1_RDNA2; - mmq_y = MMQ_Y_Q4_1_RDNA2; - nwarps = NWARPS_Q4_1_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q4_1_RDNA1; - mmq_y = MMQ_Y_Q4_1_RDNA1; - nwarps = NWARPS_Q4_1_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q4_1_AMPERE; - mmq_y = MMQ_Y_Q4_1_AMPERE; - nwarps = NWARPS_Q4_1_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q4_1_PASCAL; - mmq_y = MMQ_Y_Q4_1_PASCAL; - nwarps = NWARPS_Q4_1_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:22: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q4_1( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_qs_q4_1_acc_ct1.get_pointer(), - tile_x_dm_q4_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } else { - const bool need_check = true; - /* - DPCT1049:23: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q4_1( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_qs_q4_1_acc_ct1.get_pointer(), - tile_x_dm_q4_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { - - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q5_0_RDNA2; - mmq_y = MMQ_Y_Q5_0_RDNA2; - nwarps = NWARPS_Q5_0_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q5_0_RDNA1; - mmq_y = MMQ_Y_Q5_0_RDNA1; - nwarps = NWARPS_Q5_0_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q5_0_AMPERE; - mmq_y = MMQ_Y_Q5_0_AMPERE; - nwarps = NWARPS_Q5_0_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q5_0_PASCAL; - mmq_y = MMQ_Y_Q5_0_PASCAL; - nwarps = NWARPS_Q5_0_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:24: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q5_0( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q5_0_acc_ct1.get_pointer(), - tile_x_d_q5_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); + for (int i = 0; i < 4; i++) { + if (nr[i] != 1) { + break; + } + if (i > 0) { + collapse_nb(cnb0, cne0); + collapse_nb(cnb1, cne1); + collapse(cne0); + collapse(cne1); + } } - } else { - const bool need_check = true; - /* - DPCT1049:25: 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. - */ { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + int64_t ne0 = cne0[0]; + int64_t ne1 = cne0[1]; + int64_t ne2 = cne0[2]; + int64_t ne3 = cne0[3]; - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q5_0( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q5_0_acc_ct1.get_pointer(), - tile_x_d_q5_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} + int64_t ne10 = cne1[0]; + int64_t ne11 = cne1[1]; + int64_t ne12 = cne1[2]; + int64_t ne13 = cne1[3]; -static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { + size_t nb0 = cnb0[0]; + size_t nb1 = cnb0[1]; + size_t nb2 = cnb0[2]; + size_t nb3 = cnb0[3]; - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q5_1_RDNA2; - mmq_y = MMQ_Y_Q5_1_RDNA2; - nwarps = NWARPS_Q5_1_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q5_1_RDNA1; - mmq_y = MMQ_Y_Q5_1_RDNA1; - nwarps = NWARPS_Q5_1_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q5_1_AMPERE; - mmq_y = MMQ_Y_Q5_1_AMPERE; - nwarps = NWARPS_Q5_1_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q5_1_PASCAL; - mmq_y = MMQ_Y_Q5_1_PASCAL; - nwarps = NWARPS_Q5_1_PASCAL; - } else { - GGML_ASSERT(false); - } + size_t nb10 = cnb1[0]; + size_t nb11 = cnb1[1]; + size_t nb12 = cnb1[2]; + size_t nb13 = cnb1[3]; - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + size_t s0 = nb0 / sizeof(dst_t); + size_t s1 = nb1 / sizeof(dst_t); + size_t s2 = nb2 / sizeof(dst_t); + size_t s3 = nb3 / sizeof(dst_t); - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:26: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + size_t s10 = nb10 / sizeof(src1_t); + size_t s11 = nb11 / sizeof(src1_t); + size_t s12 = nb12 / sizeof(src1_t); + size_t s13 = nb13 / sizeof(src1_t); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q5_1( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q5_1_acc_ct1.get_pointer(), - tile_x_dm_q5_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } else { - const bool need_check = true; - /* - DPCT1049:27: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + GGML_ASSERT(s0 == 1); + GGML_ASSERT(s10 == 1); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q5_1( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q5_1_acc_ct1.get_pointer(), - tile_x_dm_q5_1_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} + const int block_size = 128; -static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { + int64_t hne0 = std::max(ne0/2LL, 1LL); - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q8_0_RDNA2; - mmq_y = MMQ_Y_Q8_0_RDNA2; - nwarps = NWARPS_Q8_0_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q8_0_RDNA1; - mmq_y = MMQ_Y_Q8_0_RDNA1; - nwarps = NWARPS_Q8_0_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q8_0_AMPERE; - mmq_y = MMQ_Y_Q8_0_AMPERE; - nwarps = NWARPS_Q8_0_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q8_0_PASCAL; - mmq_y = MMQ_Y_Q8_0_PASCAL; - nwarps = NWARPS_Q8_0_PASCAL; - } else { - GGML_ASSERT(false); - } + sycl::range<3> block_dims(1, 1, 1); + block_dims[2] = std::min(hne0, block_size); + block_dims[1] = std::min( + ne1, block_size / (unsigned int)block_dims[2]); + block_dims[0] = std::min( + std::min( + ne2 * ne3, block_size / (unsigned int)block_dims[2] / + (unsigned int)block_dims[1]), + 64U); - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + sycl::range<3> block_nums( + (ne2 * ne3 + block_dims[0] - 1) / block_dims[0], + (ne1 + block_dims[1] - 1) / block_dims[1], + (hne0 + block_dims[2] - 1) / block_dims[2]); - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:28: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + if (block_nums[0] > 65535) { + // this is the maximum number of blocks in z direction, fallback to 1D grid kernel + int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q8_0( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_qs_q8_0_acc_ct1.get_pointer(), - tile_x_d_q8_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } else { - const bool need_check = true; - /* - DPCT1049:29: 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. - */ - { - 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::range<3>(1, 1, block_size)), + [=](sycl::nd_item<3> item_ct1) { + k_bin_bcast_unravel( + src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, + ne10, ne11, ne12, ne13, s1, s2, s3, s11, s12, + s13, item_ct1); + }); + } + } else { + /* + DPCT1049:16: 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. + */ + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0), - cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - sycl::local_accessor tile_y_ds_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); - - cgh.parallel_for( + stream->parallel_for( 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, - tile_x_qs_q8_0_acc_ct1.get_pointer(), - tile_x_d_q8_0_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + k_bin_bcast(src0_dd, src1_dd, dst_dd, ne0, ne1, + ne2, ne3, ne10, ne11, ne12, ne13, + s1, s2, s3, s11, s12, s13, + item_ct1); }); - }); + } } } -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} +}; -static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { +static void acc_f32_sycl(const float *x, const float *y, float *dst, + const int n_elements, const int ne10, const int ne11, + 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); + }); +} - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q2_K_RDNA2; - mmq_y = MMQ_Y_Q2_K_RDNA2; - nwarps = NWARPS_Q2_K_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q2_K_RDNA1; - mmq_y = MMQ_Y_Q2_K_RDNA1; - nwarps = NWARPS_Q2_K_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q2_K_AMPERE; - mmq_y = MMQ_Y_Q2_K_AMPERE; - nwarps = NWARPS_Q2_K_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q2_K_PASCAL; - mmq_y = MMQ_Y_Q2_K_PASCAL; - nwarps = NWARPS_Q2_K_PASCAL; - } else { - GGML_ASSERT(false); - } +static void gelu_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); + }); +} - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); +static void silu_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); + }); +} - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:30: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); +static void gelu_quick_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); + }); +} - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K), - cgh); - sycl::local_accessor tile_x_sc_q2_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q2_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q2_K_acc_ct1.get_pointer(), - tile_x_dm_q2_K_acc_ct1.get_pointer(), - tile_x_sc_q2_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } else { - const bool need_check = true; - /* - DPCT1049:31: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); +static void tanh_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); + }); +} - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K), - cgh); - sycl::local_accessor tile_x_sc_q2_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q2_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q2_K_acc_ct1.get_pointer(), - tile_x_dm_q2_K_acc_ct1.get_pointer(), - tile_x_sc_q2_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } +static void relu_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); + }); } -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); + +static void hardsigmoid_f32_sycl(const float *x, float *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::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + hardsigmoid_f32(x, dst, k, item_ct1); + }); } -static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { +static void hardswish_f32_sycl(const float *x, float *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::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + hardswish_f32(x, dst, k, item_ct1); + }); +} - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q3_K_RDNA2; - mmq_y = MMQ_Y_Q3_K_RDNA2; - nwarps = NWARPS_Q3_K_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q3_K_RDNA1; - mmq_y = MMQ_Y_Q3_K_RDNA1; - nwarps = NWARPS_Q3_K_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q3_K_AMPERE; - mmq_y = MMQ_Y_Q3_K_AMPERE; - nwarps = NWARPS_Q3_K_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q3_K_PASCAL; - mmq_y = MMQ_Y_Q3_K_PASCAL; - nwarps = NWARPS_Q3_K_PASCAL; - } else { - GGML_ASSERT(false); - } +static void leaky_relu_f32_sycl(const float *x, float *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_f32(x, dst, k, negative_slope, item_ct1); + }); +} - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); +static void sqr_f32_sycl(const float *x, float *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_f32(x, dst, k, item_ct1); + }); +} - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:32: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); +static void norm_f32_sycl(const float *x, float *dst, const int ncols, + const int nrows, const float eps, + queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1( + sycl::range<1>(32), cgh); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K), - cgh); - sycl::local_accessor tile_x_qh_q3_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh); - sycl::local_accessor tile_x_sc_q3_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q3_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q3_K_acc_ct1.get_pointer(), - tile_x_dm_q3_K_acc_ct1.get_pointer(), - tile_x_qh_q3_K_acc_ct1.get_pointer(), - tile_x_sc_q3_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), WARP_SIZE); }); - }); - } + }); } else { - const bool need_check = true; + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); /* - DPCT1049:33: The work-group size passed to the SYCL kernel may exceed + DPCT1049:17: 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. */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1( + sycl::range<1>(32), cgh); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K), - cgh); - sycl::local_accessor tile_x_qh_q3_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh); - sycl::local_accessor tile_x_sc_q3_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q3_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q3_K_acc_ct1.get_pointer(), - tile_x_dm_q3_K_acc_ct1.get_pointer(), - tile_x_qh_q3_K_acc_ct1.get_pointer(), - tile_x_sc_q3_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); }); - }); - } + }); } } -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { - - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q4_K_RDNA2; - mmq_y = MMQ_Y_Q4_K_RDNA2; - nwarps = NWARPS_Q4_K_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q4_K_RDNA1; - mmq_y = MMQ_Y_Q4_K_RDNA1; - nwarps = NWARPS_Q4_K_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q4_K_AMPERE; - mmq_y = MMQ_Y_Q4_K_AMPERE; - nwarps = NWARPS_Q4_K_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q4_K_PASCAL; - mmq_y = MMQ_Y_Q4_K_PASCAL; - nwarps = NWARPS_Q4_K_PASCAL; - } else { - GGML_ASSERT(false); - } - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); +static void group_norm_f32_sycl(const float *x, float *dst, + const int num_groups, const int group_size, + const int ne_elements, queue_ptr stream) { + static const float eps = 1e-6f; + if (group_size < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + cgh); - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:34: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + const float eps_ct4 = eps; - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K), - cgh); - sycl::local_accessor tile_x_sc_q4_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q4_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q4_K_acc_ct1.get_pointer(), - tile_x_dm_q4_K_acc_ct1.get_pointer(), - tile_x_sc_q4_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + group_norm_f32( + x, dst, group_size, ne_elements, eps_ct4, item_ct1, + s_sum_acc_ct1.get_pointer(), WARP_SIZE); }); - }); - } + }); } else { - const bool need_check = true; + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); /* - DPCT1049:35: The work-group size passed to the SYCL kernel may exceed + DPCT1049:18: 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. */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K), - cgh); - sycl::local_accessor tile_x_sc_q4_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q4_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q4_K_acc_ct1.get_pointer(), - tile_x_dm_q4_K_acc_ct1.get_pointer(), - tile_x_sc_q4_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + 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) + [[intel::reqd_sub_group_size(32)]] { + group_norm_f32(x, dst, group_size, ne_elements, + eps_ct4, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); }); - }); - } + }); } } -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} -static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { +static void concat_f32_sycl(const float *x, const float *y, float *dst, + const int ne0, int ne1, int ne2, int ne02, + queue_ptr stream) { + int num_blocks = (ne0 + SYCL_CONCAT_BLOCK_SIZE - 1) / SYCL_CONCAT_BLOCK_SIZE; + sycl::range<3> gridDim(ne2, ne1, num_blocks); + 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(x, y, dst, ne0, ne02, item_ct1); + }); +} - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q5_K_RDNA2; - mmq_y = MMQ_Y_Q5_K_RDNA2; - nwarps = NWARPS_Q5_K_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q5_K_RDNA1; - mmq_y = MMQ_Y_Q5_K_RDNA1; - nwarps = NWARPS_Q5_K_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q5_K_AMPERE; - mmq_y = MMQ_Y_Q5_K_AMPERE; - nwarps = NWARPS_Q5_K_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q5_K_PASCAL; - mmq_y = MMQ_Y_Q5_K_PASCAL; - nwarps = NWARPS_Q5_K_PASCAL; - } else { - GGML_ASSERT(false); - } +static void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01, + const int nb02, const int nb03, const int ne10, const int ne11, + const int ne12, const int ne13, const float sf0, const float sf1, + const float sf2, const float sf3, queue_ptr stream) { + 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) { + upscale_f32(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1); + }); +} - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); +static void pad_f32_sycl(const float *x, float *dst, const int ne00, + const int ne01, const int ne02, const int ne0, + 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_f32(x, dst, ne0, ne00, ne01, ne02, item_ct1); + }); +} - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:36: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); +static void rms_norm_f32_sycl(const float *x, float *dst, const int ncols, + const int nrows, const float eps, + queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + // 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) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + cgh); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K), - cgh); - sycl::local_accessor tile_x_sc_q5_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q5_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q5_K_acc_ct1.get_pointer(), - tile_x_dm_q5_K_acc_ct1.get_pointer(), - tile_x_sc_q5_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), WARP_SIZE); }); - }); - } + }); } else { - const bool need_check = true; + const int work_group_size = get_work_group_size(stream->get_device()); + const sycl::range<3> block_dims(1, 1, work_group_size); /* - DPCT1049:37: The work-group size passed to the SYCL kernel may exceed + DPCT1049:19: 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. */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + cgh); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K), - cgh); - sycl::local_accessor tile_x_sc_q5_K_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q5_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_q5_K_acc_ct1.get_pointer(), - tile_x_dm_q5_K_acc_ct1.get_pointer(), - tile_x_sc_q5_K_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); }); - }); - } + }); } } -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols_x, - const int nrows_x, const int ncols_y, - const int nrows_y, const int nrows_dst, - queue_ptr stream) try { - - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - const int compute_capability = ggml_sycl_info().devices[id].cc; - - int mmq_x, mmq_y, nwarps; - if (compute_capability >= VER_GEN13) { - mmq_x = MMQ_X_Q6_K_RDNA2; - mmq_y = MMQ_Y_Q6_K_RDNA2; - nwarps = NWARPS_Q6_K_RDNA2; - } else if (compute_capability >= VER_GEN12) { - mmq_x = MMQ_X_Q6_K_RDNA1; - mmq_y = MMQ_Y_Q6_K_RDNA1; - nwarps = NWARPS_Q6_K_RDNA1; - } else if (compute_capability >= VER_GEN9) { - mmq_x = MMQ_X_Q6_K_AMPERE; - mmq_y = MMQ_Y_Q6_K_AMPERE; - nwarps = NWARPS_Q6_K_AMPERE; - } else if (compute_capability >= VER_4VEC) { - mmq_x = MMQ_X_Q6_K_PASCAL; - mmq_y = MMQ_Y_Q6_K_PASCAL; - nwarps = NWARPS_Q6_K_PASCAL; - } else { - GGML_ASSERT(false); - } - - const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; - const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; - const sycl::range<3> block_nums(1, block_num_y, block_num_x); - const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); - - if (nrows_x % mmq_y == 0) { - const bool need_check = false; - /* - DPCT1049:38: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K), - cgh); - sycl::local_accessor tile_x_sc_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q6_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_acc_ct1.get_pointer(), - tile_x_dm_acc_ct1.get_pointer(), - tile_x_sc_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); - }); - } - } else { - const bool need_check = true; - /* - DPCT1049:39: 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. - */ - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); +static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx, + const int ky, const int kx_padded, + queue_ptr stream) { + const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE; + const sycl::range<3> num_blocks(1, ky, block_num_x); + const sycl::range<3> block_size(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); - stream->submit([&](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( - sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K), - cgh); - sycl::local_accessor tile_x_sc_acc_ct1( - sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); - sycl::local_accessor tile_y_qs_acc_ct1( - sycl::range<1>(mmq_x * WARP_SIZE), cgh); - 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) { - mul_mat_q6_K( - vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, - nrows_dst, item_ct1, - tile_x_ql_acc_ct1.get_pointer(), - tile_x_dm_acc_ct1.get_pointer(), - tile_x_sc_acc_ct1.get_pointer(), - tile_y_qs_acc_ct1.get_pointer(), - tile_y_ds_acc_ct1.get_pointer()); - }); + stream->parallel_for( + sycl::nd_range<3>(num_blocks * block_size, block_size), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + quantize_q8_1(x, vy, kx, kx_padded, item_ct1); }); - } } } -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y, float *dst, const int ncols_x, @@ -9187,7 +2445,7 @@ static void soft_max_f32_sycl(const float * x, const float * mask, const int nrows_y, const float scale, const float max_bias, queue_ptr stream) { int nth = WARP_SIZE; - int max_block_size = GROUP_SIZE; + int max_block_size = get_work_group_size(stream->get_device()); while (nth < ncols_x && nth < max_block_size) nth *= 2; if (nth>max_block_size) nth = max_block_size; @@ -9339,7 +2597,7 @@ void ggml_backend_sycl_print_sycl_devices() { } } -int get_sycl_env(const char *env_name, int default_val) { +static inline int get_sycl_env(const char *env_name, int default_val) { char *user_device_string = getenv(env_name); int user_number = default_val; @@ -9353,10 +2611,9 @@ int get_sycl_env(const char *env_name, int default_val) { return user_number; } -int get_work_group_size(int user_device_id) { +static inline int get_work_group_size(const sycl::device& device) { dpct::device_info prop; - dpct::get_device_info(prop, - dpct::dev_mgr::instance().get_device(user_device_id)); + dpct::get_device_info(prop, device); return prop.get_max_work_group_size(); } @@ -10042,76 +3299,6 @@ inline void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, const ggml_te (void) src1_dd; } -inline void ggml_sycl_op_mul_mat_q( - ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, - const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, - float *dst_dd_i, const int64_t row_low, const int64_t row_high, - const int64_t src1_ncols, const int64_t src1_padded_row_size, - const queue_ptr &stream) try { - - const int64_t ne00 = src0->ne[0]; - - const int64_t ne10 = src1->ne[0]; - GGML_ASSERT(ne10 % QK8_1 == 0); - - const int64_t ne0 = dst->ne[0]; - - const int64_t row_diff = row_high - row_low; - - int device_id; - SYCL_CHECK( - CHECK_TRY_ERROR(device_id = get_current_device_id())); - - // the main device has a larger memory buffer to hold the results from all GPUs - // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into - const int64_t nrows_dst = device_id == ctx.device ? ne0 : row_diff; - - switch (src0->type) { - case GGML_TYPE_Q4_0: - ggml_mul_mat_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q4_1: - ggml_mul_mat_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q5_0: - ggml_mul_mat_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q5_1: - ggml_mul_mat_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q8_0: - ggml_mul_mat_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q2_K: - ggml_mul_mat_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q3_K: - ggml_mul_mat_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q4_K: - ggml_mul_mat_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q5_K: - ggml_mul_mat_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - case GGML_TYPE_Q6_K: - ggml_mul_mat_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); - break; - default: - GGML_ASSERT(false); - break; - } - - (void) src1; - (void) dst; - (void) src1_ddf_i; -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - static int64_t get_row_rounding(ggml_type type, const std::array & tensor_split) { int64_t min_compute_capability = INT_MAX; int64_t max_compute_capability = INT_MIN; @@ -10160,179 +3347,6 @@ static int64_t get_row_rounding(ggml_type type, const std::arrayne[0]; - GGML_ASSERT(ne10 % QK8_1 == 0); - - const int64_t ne00 = src0->ne[0]; - const int64_t row_diff = row_high - row_low; - - int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); - - // the main device has a larger memory buffer to hold the results from all GPUs - // nrows_dst == nrows of the matrix that the kernel writes into - const int64_t nrows_dst = id == ctx.device ? ne00 : row_diff; - - switch (src0->type) { - case GGML_TYPE_Q4_0: - mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_1: - mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_0: - mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_1: - mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q8_0: - mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q2_K: - mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q3_K: - mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_K: - mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_K: - mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q6_K: - mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ1_S: - mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ1_M: - mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ2_XXS: - mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ2_XS: - mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ2_S: - mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ3_XXS: - mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ3_S: - mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ4_NL: - mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ4_XS: - mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); - break; - default: - GGML_ASSERT(false); - break; - } - - (void) src1; - (void) dst; - (void) src1_ddf_i; - (void) src1_ncols; - (void) src1_padded_row_size; -} - - -inline void ggml_sycl_op_dequantize_mul_mat_vec( - ggml_backend_sycl_context & ctx, - const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, - const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, - float *dst_dd_i, const int64_t row_low, const int64_t row_high, - const int64_t src1_ncols, const int64_t src1_padded_row_size, - const queue_ptr &stream) { - - const int64_t ne00 = src0->ne[0]; - const int64_t row_diff = row_high - row_low; - - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics -#ifdef GGML_SYCL_F16 - ggml_sycl_pool_alloc src1_dfloat_a(ctx.pool()); - sycl::half *src1_dfloat = nullptr; // dfloat == half - - bool src1_convert_f16 = - src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || - src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 || - src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; - - if (src1_convert_f16) { - src1_dfloat = src1_dfloat_a.alloc(ne00); - const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); - GGML_ASSERT(to_fp16_sycl != nullptr); - to_fp16_sycl(src1_ddf_i, src1_dfloat, ne00, stream); - } -#else - const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion -#endif // GGML_SYCL_F16 - - switch (src0->type) { - case GGML_TYPE_Q4_0: - dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_1: - dequantize_mul_mat_vec_q4_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_0: - dequantize_mul_mat_vec_q5_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_1: - dequantize_mul_mat_vec_q5_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q8_0: - dequantize_mul_mat_vec_q8_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q2_K: - dequantize_mul_mat_vec_q2_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q3_K: - dequantize_mul_mat_vec_q3_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_K: - dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_K: - dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_Q6_K: - dequantize_mul_mat_vec_q6_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); - break; - case GGML_TYPE_F16: - convert_mul_mat_vec_f16_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); - break; - default: - printf("ggml_sycl_op_dequantize_mul_mat_vec unsupported GGML_TYPE %d\n", src0->type); - GGML_ASSERT(false); - break; - } - - (void) src1; - (void) dst; - (void) src1_ddq_i; - (void) src1_ncols; - (void) src1_padded_row_size; -} - inline void ggml_sycl_op_mul_mat_sycl( ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, diff --git a/ggml-sycl/backend.hpp b/ggml-sycl/backend.hpp index 88bae59678bdd..2d37e271f9050 100644 --- a/ggml-sycl/backend.hpp +++ b/ggml-sycl/backend.hpp @@ -14,5 +14,10 @@ #define GGML_SYCL_BACKEND_HPP #include "common.hpp" +#include "convert.hpp" +#include "dequantize.hpp" +#include "dmmv.hpp" +#include "mmq.hpp" +#include "mmvq.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml-sycl/convert.cpp b/ggml-sycl/convert.cpp new file mode 100644 index 0000000000000..ce9de2b42b722 --- /dev/null +++ b/ggml-sycl/convert.cpp @@ -0,0 +1,544 @@ +#include "convert.hpp" +#include "dequantize.hpp" +#include "presets.hpp" + +template +static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = 2 * (item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2)); + + if (i >= k) { + return; + } + + const int ib = i/qk; // block index + const int iqs = (i%qk)/qr; // quant index + const int iybs = i - i%qk; // y block start index + const int y_offset = qr == 1 ? 1 : qk/2; + + // dequantize + dfloat2 v; + dequantize_kernel(vx, ib, iqs, v); + + y[iybs + iqs + 0] = v.x(); + y[iybs + iqs + y_offset] = v.y(); +} + +template +static void dequantize_block_sycl(const void *__restrict__ vx, + dst_t *__restrict__ y, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + 2*SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / (2*SYCL_DEQUANTIZE_BLOCK_SIZE); + { + 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); + }); + } +} + +template +static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + 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); + }); + } +#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); + }); + } + +#endif +} + +template +static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + 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); + }); + } +#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); + }); + } +#endif +} + +template +static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb32 = k / 32; + const int nb = (k + 255) / 256; + { + 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); + }); + } +} + +template +static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb32 = k / 32; + const int nb = (k + 255) / 256; + { + 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); + }); + } +} + + +template +static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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_K(vx, y, item_ct1); + }); + } +} + +template +static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + 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); + }); + } +#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); + }); + } + +#endif +} + +template +static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + 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); + }); + } +#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); + }); + } + +#endif +} + +template +static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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 + ); + }); + }); + } +} + +template +static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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 + ); + }); + }); + } +} + +template +static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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_xxs( + vx, y, item_ct1, iq2xxs_grid, + ksigns_iq2xs, kmask_iq2xs); + }); + }); + } +} + +template +static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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); + }); + }); + } +} + +template +static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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); + }); + }); + } +} + + +template +static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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_xxs( + vx, y, item_ct1, iq3xxs_grid, + ksigns_iq2xs, kmask_iq2xs); + }); + }); + } +} + +template +static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_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); + }); + }); + } +} + +template +static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = (k + QK_K - 1) / QK_K; +#if QK_K == 64 + dequantize_row_iq4_nl_sycl(vx, y, k, stream); +#else + { + 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); + }); + }); + } +#endif +} + +template +static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = (k + QK_K - 1) / QK_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); + }); + }); + } +} + +template +static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + + const src_t * x = (src_t *) vx; + + y[i] = x[i]; +} + +template +static void convert_unary_sycl(const void *__restrict__ vx, + dst_t *__restrict__ y, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE; + { + 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) { + convert_unary(vx, y, k, item_ct1); + }); + } +} + +to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_block_sycl; + case GGML_TYPE_Q4_1: + return dequantize_block_sycl; + case GGML_TYPE_Q5_0: + return dequantize_block_sycl; + case GGML_TYPE_Q5_1: + return dequantize_block_sycl; + case GGML_TYPE_Q8_0: + return dequantize_block_sycl; + case GGML_TYPE_Q2_K: + return dequantize_row_q2_K_sycl; + case GGML_TYPE_Q3_K: + return dequantize_row_q3_K_sycl; + case GGML_TYPE_Q4_K: + return dequantize_row_q4_K_sycl; + case GGML_TYPE_Q5_K: + return dequantize_row_q5_K_sycl; + case GGML_TYPE_Q6_K: + return dequantize_row_q6_K_sycl; + case GGML_TYPE_IQ1_S: + return dequantize_row_iq1_s_sycl; + case GGML_TYPE_IQ1_M: + return dequantize_row_iq1_m_sycl; + case GGML_TYPE_IQ2_XXS: + return dequantize_row_iq2_xxs_sycl; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_sycl; + case GGML_TYPE_IQ2_S: + return dequantize_row_iq2_s_sycl; + case GGML_TYPE_IQ3_XXS: + return dequantize_row_iq3_xxs_sycl; + case GGML_TYPE_IQ3_S: + return dequantize_row_iq3_s_sycl; + case GGML_TYPE_IQ4_XS: + return dequantize_row_iq4_xs_sycl; + case GGML_TYPE_IQ4_NL: + return dequantize_row_iq4_nl_sycl; + case GGML_TYPE_F32: + return convert_unary_sycl; + default: + return nullptr; + } +} + +to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_row_q4_0_sycl; + case GGML_TYPE_Q4_1: + return dequantize_row_q4_1_sycl; + case GGML_TYPE_Q5_0: + return dequantize_block_sycl; + case GGML_TYPE_Q5_1: + return dequantize_block_sycl; + case GGML_TYPE_Q8_0: + return dequantize_block_sycl; + case GGML_TYPE_Q2_K: + return dequantize_row_q2_K_sycl; + case GGML_TYPE_Q3_K: + return dequantize_row_q3_K_sycl; + case GGML_TYPE_Q4_K: + return dequantize_row_q4_K_sycl; + case GGML_TYPE_Q5_K: + return dequantize_row_q5_K_sycl; + case GGML_TYPE_Q6_K: + return dequantize_row_q6_K_sycl; + case GGML_TYPE_IQ1_S: + return dequantize_row_iq1_s_sycl; + case GGML_TYPE_IQ1_M: + return dequantize_row_iq1_m_sycl; + case GGML_TYPE_IQ2_XXS: + return dequantize_row_iq2_xxs_sycl; + case GGML_TYPE_IQ2_XS: + return dequantize_row_iq2_xs_sycl; + case GGML_TYPE_IQ2_S: + return dequantize_row_iq2_s_sycl; + case GGML_TYPE_IQ3_XXS: + return dequantize_row_iq3_xxs_sycl; + case GGML_TYPE_IQ3_S: + return dequantize_row_iq3_s_sycl; + case GGML_TYPE_IQ4_XS: + return dequantize_row_iq4_xs_sycl; + case GGML_TYPE_IQ4_NL: + return dequantize_row_iq4_nl_sycl; + case GGML_TYPE_F16: + return convert_unary_sycl; + default: + return nullptr; + } +} diff --git a/ggml-sycl/convert.hpp b/ggml-sycl/convert.hpp new file mode 100644 index 0000000000000..b1f10d6355535 --- /dev/null +++ b/ggml-sycl/convert.hpp @@ -0,0 +1,27 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_CONVERT_HPP +#define GGML_SYCL_CONVERT_HPP + +#include "common.hpp" + +template +using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y, + int k, dpct::queue_ptr stream); +typedef to_t_sycl_t to_fp32_sycl_t; +typedef to_t_sycl_t to_fp16_sycl_t; + +to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type); +to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type); + +#endif // GGML_SYCL_CONVERT_HPP diff --git a/ggml-sycl/dequantize.hpp b/ggml-sycl/dequantize.hpp new file mode 100644 index 0000000000000..b6080d83a33eb --- /dev/null +++ b/ggml-sycl/dequantize.hpp @@ -0,0 +1,690 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_DEQUANTIZE_HPP +#define GGML_SYCL_DEQUANTIZE_HPP + +#include "common.hpp" + +typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v); + +static __dpct_inline__ void dequantize_q4_0(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q4_0 * x = (const block_q4_0 *) vx; + + const dfloat d = x[ib].d; + + const int vui = x[ib].qs[iqs]; + + v.x() = vui & 0xF; + v.y() = vui >> 4; + +#ifdef GGML_SYCL_F16 + // v = v - {8.0f, 8.0f}; + // v = v * {d, d}; + v.s0() = (v.s0() - 8.0f) * d; + v.s1() = (v.s1() - 8.0f) * d; + +#else + v.x() = (v.x() - 8.0f) * d; + v.y() = (v.y() - 8.0f) * d; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q4_1(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q4_1 * x = (const block_q4_1 *) vx; + + const dfloat d = x[ib].dm[0]; + const dfloat m = x[ib].dm[1]; + + const int vui = x[ib].qs[iqs]; + + v.x() = vui & 0xF; + v.y() = vui >> 4; + +#ifdef GGML_SYCL_F16 + // v = v * {d, d}; + // v = v + {m, m}; + v.s0() = (v.s0() * d) + m; + v.s1() = (v.s1() * d) + m; + +#else + v.x() = (v.x() * d) + m; + v.y() = (v.y() * d) + m; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q5_0(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q5_0 * x = (const block_q5_0 *) vx; + + const dfloat d = x[ib].d; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0); + v.y() = ((x[ib].qs[iqs] >> 4) | xh_1); + +#ifdef GGML_SYCL_F16 + // v = v - {16.0f, 16.0f}; + // v = v * {d, d}; + v.s0() = (v.s0() - 16.0f) * d; + v.s1() = (v.s1() - 16.0f) * d; + +#else + v.x() = (v.x() - 16.0f) * d; + v.y() = (v.y() - 16.0f) * d; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q5_1(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q5_1 * x = (const block_q5_1 *) vx; + + const dfloat d = x[ib].dm[0]; + const dfloat m = x[ib].dm[1]; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0); + v.y() = ((x[ib].qs[iqs] >> 4) | xh_1); + +#ifdef GGML_SYCL_F16 + // v = v * {d, d}; + // v = v + {m, m}; + v.s0() = (v.s0() * d) + m; + v.s1() = (v.s1() * d) + m; +#else + v.x() = (v.x() * d) + m; + v.y() = (v.y() * d) + m; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q8_0(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q8_0 * x = (const block_q8_0 *) vx; + + const dfloat d = x[ib].d; + + v.x() = x[ib].qs[iqs + 0]; + v.y() = x[ib].qs[iqs + 1]; + +#ifdef GGML_SYCL_F16 + // v = v * {d, d}; + v.s0() *= d; + v.s1() *= d; +#else + v.x() *= d; + v.y() *= d; +#endif // GGML_SYCL_F16 +} + +template +static void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + + // assume 32 threads + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; + const int ir = tid%8; + const int ib = 8*i + ir; + if (ib >= nb32) { + return; + } + + dst_t * y = yy + 256*i + 32*ir + 4*il; + + const block_q4_0 * x = (const block_q4_0 *)vx + ib; + const float d = sycl::vec(x->d) + .convert()[0]; + const float dm = -8*d; + + const uint8_t * q = x->qs + 4*il; + + for (int l = 0; l < 4; ++l) { + y[l+ 0] = d * (q[l] & 0xF) + dm; + y[l+16] = d * (q[l] >> 4) + dm; + } +} + +template +static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + + // assume 32 threads + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; + const int ir = tid%8; + const int ib = 8*i + ir; + if (ib >= nb32) { + return; + } + + dst_t * y = yy + 256*i + 32*ir + 4*il; + + const block_q4_1 * x = (const block_q4_1 *)vx + ib; + const sycl::float2 d = + x->dm.convert(); + + const uint8_t * q = x->qs + 4*il; + + for (int l = 0; l < 4; ++l) { + y[l + 0] = d.x() * (q[l] & 0xF) + d.y(); + y[l + 16] = d.x() * (q[l] >> 4) + d.y(); + } +} + + +//================================== k-quants + +template +static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + const block_q2_K * x = (const block_q2_K *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int n = tid/32; + const int l = tid - 32*n; + const int is = 8*n + l/16; + + const uint8_t q = x[i].qs[32*n + l]; + dst_t * y = yy + i*QK_K + 128*n; + + float dall = x[i].dm[0]; + float dmin = x[i].dm[1]; + y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); + y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); + y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); + y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); +#else + const int is = tid/16; // 0 or 1 + const int il = tid%16; // 0...15 + const uint8_t q = x[i].qs[il] >> (2*is); + dst_t * y = yy + i*QK_K + 16*is + il; + + float dall = x[i].dm[0]; + float dmin = x[i].dm[1]; + y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); + y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); +#endif + +} + +template +static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + const block_q3_K * x = (const block_q3_K *) vx; + +#if QK_K == 256 + const int r = item_ct1.get_local_id(2) / 4; + const int tid = r/2; + const int is0 = r%2; + const int l0 = 16 * is0 + 4 * (item_ct1.get_local_id(2) % 4); + const int n = tid / 4; + const int j = tid - 4*n; + + uint8_t m = 1 << (4*n + j); + int is = 8*n + 2*j + is0; + int shift = 2*j; + + int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) : + is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) : + is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) : + (x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4); + float d_all = x[i].d; + float dl = d_all * (us - 32); + + dst_t * y = yy + i*QK_K + 128*n + 32*j; + const uint8_t * q = x[i].qs + 32*n; + const uint8_t * hm = x[i].hmask; + + for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); +#else + const int tid = item_ct1.get_local_id(2); + const int is = tid/16; // 0 or 1 + const int il = tid%16; // 0...15 + const int im = il/8; // 0...1 + const int in = il%8; // 0...7 + + dst_t * y = yy + i*QK_K + 16*is + il; + + const uint8_t q = x[i].qs[il] >> (2*is); + const uint8_t h = x[i].hmask[in] >> (2*is + im); + const float d = (float)x[i].d; + + if (is == 0) { + y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); + y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); + } else { + y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); + y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); + } +#endif + +} + +#if QK_K == 256 +static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { + if (j < 4) { + d = q[j] & 63; m = q[j + 4] & 63; + } else { + d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); + m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); + } +} +#endif + +template +static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + const block_q4_K * x = (const block_q4_K *) vx; + + const int i = item_ct1.get_group(2); + +#if QK_K == 256 + // assume 32 threads + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; + const int ir = tid%8; + const int is = 2*il; + const int n = 4; + + dst_t * y = yy + i*QK_K + 64*il + n*ir; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint8_t * q = x[i].qs + 32*il + n*ir; + + uint8_t sc, m; + get_scale_min_k4(is + 0, x[i].scales, sc, m); + const float d1 = dall * sc; const float m1 = dmin * m; + get_scale_min_k4(is + 1, x[i].scales, sc, m); + const float d2 = dall * sc; const float m2 = dmin * m; + for (int l = 0; l < n; ++l) { + y[l + 0] = d1 * (q[l] & 0xF) - m1; + y[l +32] = d2 * (q[l] >> 4) - m2; + } +#else + const int tid = item_ct1.get_local_id(2); + const uint8_t * q = x[i].qs; + dst_t * y = yy + i*QK_K; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; + y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4); + y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4); +#endif +} + +template +static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + const block_q5_K * x = (const block_q5_K *) vx; + + const int i = item_ct1.get_group(2); + +#if QK_K == 256 + // assume 64 threads - this is very slightly better than the one below + const int tid = item_ct1.get_local_id(2); + const int il = tid/16; // il is in 0...3 + const int ir = tid%16; // ir is in 0...15 + const int is = 2*il; // is is in 0...6 + + dst_t * y = yy + i*QK_K + 64*il + 2*ir; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint8_t * ql = x[i].qs + 32*il + 2*ir; + const uint8_t * qh = x[i].qh + 2*ir; + + uint8_t sc, m; + get_scale_min_k4(is + 0, x[i].scales, sc, m); + const float d1 = dall * sc; const float m1 = dmin * m; + get_scale_min_k4(is + 1, x[i].scales, sc, m); + const float d2 = dall * sc; const float m2 = dmin * m; + + uint8_t hm = 1 << (2*il); + y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1; + y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1; + hm <<= 1; + y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; + y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; +#else + const int tid = item_ct1.get_local_id(2); + const uint8_t q = x[i].qs[tid]; + const int im = tid/8; // 0...3 + const int in = tid%8; // 0...7 + const int is = tid/16; // 0 or 1 + const uint8_t h = x[i].qh[in] >> im; + const float d = x[i].d; + dst_t * y = yy + i*QK_K + tid; + y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16)); + y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16)); +#endif +} + +template +static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + const block_q6_K * x = (const block_q6_K *) vx; + + const int i = item_ct1.get_group(2); +#if QK_K == 256 + + // assume 64 threads - this is very slightly better than the one below + const int tid = item_ct1.get_local_id(2); + const int ip = tid/32; // ip is 0 or 1 + const int il = tid - 32*ip; // 0...32 + const int is = 8*ip + il/16; + + dst_t * y = yy + i*QK_K + 128*ip + il; + + const float d = x[i].d; + + const uint8_t * ql = x[i].ql + 64*ip + il; + const uint8_t qh = x[i].qh[32*ip + il]; + const int8_t * sc = x[i].scales + is; + + y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32); + y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32); + y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32); + y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); +#else + + // assume 32 threads + const int tid = item_ct1.get_local_id(2); + const int ip = tid/16; // 0 or 1 + const int il = tid - 16*ip; // 0...15 + + dst_t * y = yy + i*QK_K + 16*ip + il; + + const float d = x[i].d; + + const uint8_t ql = x[i].ql[16*ip + il]; + const uint8_t qh = x[i].qh[il] >> (2*ip); + const int8_t * sc = x[i].scales; + + y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32); + y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32); +#endif +} + +template +static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint64_t *iq2xxs_grid_ptr, + const uint8_t *ksigns_iq2xs_ptr, + const uint8_t *kmask_iq2xs_ptr) { + + const int i = item_ct1.get_group(2); + const block_iq2_xxs * x = (const block_iq2_xxs *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * q2 = x[i].qs + 4*ib; + const uint8_t * aux8 = (const uint8_t *)q2; + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid_ptr + aux8[il]); + const uint32_t aux32 = q2[2] | (q2[3] << 16); + const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f; + const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + +template +static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint64_t *iq2xs_grid, + const uint8_t *ksigns_iq2xs, + const uint8_t *kmask_iq2xs) { + + const int i = item_ct1.get_group(2); + const block_iq2_xs * x = (const block_iq2_xs *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * q2 = x[i].qs + 4*ib; + const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511)); + const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; + const uint8_t signs = ksigns_iq2xs[q2[il] >> 9]; + for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); +#else + assert(false); +#endif + +} + +template +__dpct_inline__ static void +dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + const block_iq2_s * x = (const block_iq2_s *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300))); + const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f; + const uint8_t signs = x[i].qs[QK_K/8+4*ib+il]; +#pragma unroll + for (int j = 0; j < 8; ++j) + y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); +#else + assert(false); + +#endif + +} + +template +static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint32_t *iq3xxs_grid, + const uint8_t *ksigns_iq2xs, + const uint8_t *kmask_iq2xs) { + + const int i = item_ct1.get_group(2); + const block_iq3_xxs * x = (const block_iq3_xxs *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint8_t * q3 = x[i].qs + 8*ib; + const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib; + const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]); + const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]); + const uint32_t aux32 = gas[0] | (gas[1] << 16); + const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f; + const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127]; + for (int j = 0; j < 4; ++j) { + y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); + } +#else + assert(false); +#endif + +} + +template +__dpct_inline__ static void +dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint8_t *kmask_iq2xs, const uint32_t *iq3s_grid) { + + const int i = item_ct1.get_group(2); + const block_iq3_s * x = (const block_iq3_s *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint8_t * qs = x[i].qs + 8*ib; + const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256))); + const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256))); + const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf)); + const uint8_t signs = x[i].signs[4*ib + il]; +#pragma unroll + for (int j = 0; j < 4; ++j) { + y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); + y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); + } +#else + assert(false); +#endif + +} + +template +__dpct_inline__ static void +dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint32_t *iq1s_grid_gpu) { + + const int i = item_ct1.get_group(2); + const block_iq1_s * x = (const block_iq1_s *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA; + const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1); + uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32; + grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)]; + grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f; + grid32[0] &= 0x0f0f0f0f; +#pragma unroll + for (int j = 0; j < 8; ++j) { + y[j] = d * (q[j] + delta); + } +#else + assert(false); +#endif + +} + +template +__dpct_inline__ static void +dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy, + const sycl::nd_item<3> &item_ct1, + const uint32_t *iq1s_grid_gpu) { + + const int i = item_ct1.get_group(2); + const block_iq1_m * x = (const block_iq1_m *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 8*il; + const uint16_t * sc = (const uint16_t *)x[i].scales; + iq1m_scale_t scale; + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); + const int ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4); + const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1); + const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA; + uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32; + grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)]; + grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f; + grid32[0] &= 0x0f0f0f0f; +#pragma unroll + for (int j = 0; j < 8; ++j) { + y[j] = d * (q[j] + delta); + } +#else + assert(false); +#endif + +} + +template +__dpct_inline__ static void +dequantize_block_iq4_nl(const void *__restrict__ vx, dst_t *__restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL); + + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 4*il; + const uint8_t * q4 = x[ib].qs + 4*il; + const float d = (float)x[ib].d; +#pragma unroll + for (int j = 0; j < 4; ++j) { + y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; + y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; + } + +} + + +template +__dpct_inline__ static void +dequantize_block_iq4_xs(const void *__restrict__ vx, dst_t *__restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_group(2); + const block_iq4_xs * x = (const block_iq4_xs *)vx; + + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; // 0...3 + const int ib = tid%8; // 0...7 + dst_t * y = yy + i*QK_K + 32*ib + 4*il; + const uint8_t * q4 = x[i].qs + 16*ib + 4*il; + const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32); +#pragma unroll + for (int j = 0; j < 4; ++j) { + y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf]; + y[j+16] = d * kvalues_iq4nl[q4[j] >> 4]; + } +} + + +#endif // GGML_SYCL_DEQUANTIZE_HPP diff --git a/ggml-sycl/dmmv.cpp b/ggml-sycl/dmmv.cpp new file mode 100644 index 0000000000000..3a87d3ef8e45c --- /dev/null +++ b/ggml-sycl/dmmv.cpp @@ -0,0 +1,1022 @@ +#include "convert.hpp" +#include "dmmv.hpp" +#include "dequantize.hpp" +#include "presets.hpp" + +static void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const sycl::half *x = (const sycl::half *)vx; + + // automatic half -> float type cast if dfloat == float + v.x() = x[ib + iqs + 0]; + v.y() = x[ib + iqs + 1]; +} + +static void convert_f32(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const float * x = (const float *) vx; + + // automatic half -> float type cast if dfloat == float + v.x() = x[ib + iqs + 0]; + v.y() = x[ib + iqs + 1]; +} + +template +static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1) { + // qk = quantized weights per x block + // qr = number of quantized weights per data value in x block + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int tid = item_ct1.get_local_id(2); + + const int iter_stride = 2*GGML_SYCL_DMMV_X; + const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter + const int y_offset = qr == 1 ? 1 : qk/2; + +// partial sum for each thread +#ifdef GGML_SYCL_F16 + sycl::half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics +#else + float tmp = 0.0f; +#endif // GGML_SYCL_F16 + + for (int i = 0; i < ncols; i += iter_stride) { + const int col = i + vals_per_iter*tid; + const int ib = (row*ncols + col)/qk; // x block index + const int iqs = (col%qk)/qr; // x quant index + const int iybs = col - col%qk; // y block start index + +// processing >2 values per i iter is faster for fast GPUs +#pragma unroll + for (int j = 0; j < vals_per_iter; j += 2) { + // process 2 vals per j iter + + // dequantize + // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val + dfloat2 v; + dequantize_kernel(vx, ib, iqs + j/qr, v); + + // matrix multiplication + // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2 +#ifdef GGML_SYCL_F16 + dfloat2 t1{y[iybs + iqs + j / qr + 0], + y[iybs + iqs + j / qr + y_offset]}; + + tmp += v * t1; +#else + tmp += v.x() * y[iybs + iqs + j / qr + 0]; + tmp += v.y() * y[iybs + iqs + j / qr + y_offset]; +#endif // GGML_SYCL_F16 + } + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (tid == 0) { +#ifdef GGML_SYCL_F16 + dst[row] = tmp.x() + tmp.y(); +#else + dst[row] = tmp; +#endif // GGML_SYCL_F16 + } +} + +static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + 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) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, + nrows, item_ct1); + }); + } +} + +/* +DPCT1110:4: The total declared local variable size in device function +dequantize_mul_mat_vec_q2_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q2_K * x = (const block_q2_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...15 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int step = 16/K_QUANTS_PER_ITERATION; + + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 + + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int s_offset = 8*im; + const int y_offset = 128*im + l0; + + uint32_t aux[4]; + const uint8_t * d = (const uint8_t *)aux; + const uint8_t * m = (const uint8_t *)(aux + 2); + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * q = x[i].qs + q_offset; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset); + aux[0] = a[0] & 0x0f0f0f0f; + aux[1] = a[1] & 0x0f0f0f0f; + aux[2] = (a[0] >> 4) & 0x0f0f0f0f; + aux[3] = (a[1] >> 4) & 0x0f0f0f0f; + + float sum1 = 0, sum2 = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3) + + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3) + + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3) + + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3) + + y[l+16] * d[1] * ((q[l+16] >> 0) & 3) + + y[l+48] * d[3] * ((q[l+16] >> 2) & 3) + + y[l+80] * d[5] * ((q[l+16] >> 4) & 3) + +y[l+112] * d[7] * ((q[l+16] >> 6) & 3); + sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6] + + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7]; + + } + tmp += dall * sum1 - dmin * sum2; + + } +#else + const int tid = item_ct1.get_local_id(2) / + (2 * K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = item_ct1.get_local_id(2) % + (2 * K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int offset = tid * K_QUANTS_PER_ITERATION; + + uint32_t uaux[2]; + const uint8_t * d = (const uint8_t *)uaux; + + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + offset; + const uint8_t * q = x[i].qs + offset; + const uint32_t * s = (const uint32_t *)x[i].scales; + + uaux[0] = s[0] & 0x0f0f0f0f; + uaux[1] = (s[0] >> 4) & 0x0f0f0f0f; + + const sycl::float2 dall = + x[i].dm.convert(); + + float sum1 = 0, sum2 = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + const uint8_t ql = q[l]; + sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3) + + y[l+16] * d[1] * ((ql >> 2) & 3) + + y[l+32] * d[2] * ((ql >> 4) & 3) + + y[l+48] * d[3] * ((ql >> 6) & 3); + sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7]; + } + tmp += dall.x() * sum1 - dall.y() * sum2; + } + +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +/* +DPCT1110:5: The total declared local variable size in device function +dequantize_mul_mat_vec_q3_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q3_K * x = (const block_q3_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + + const uint16_t kmask1 = 0x0303; + const uint16_t kmask2 = 0x0f0f; + + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop + const int step = 16/K_QUANTS_PER_ITERATION; + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0....15 or 0...7 + + const uint8_t m = 1 << (4*im); + + const int l0 = n*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int y_offset = 128*im + l0; + + uint16_t utmp[4]; + const int8_t * s = (const int8_t *)utmp; + + const uint16_t s_shift = 4*im; + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * q = x[i].qs + q_offset; + const uint8_t * h = x[i].hmask + l0; + + const uint16_t * a = (const uint16_t *)x[i].scales; + utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4); + utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4); + utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4); + utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4); + + const float d = x[i].d; + + float sum = 0; + for (int l = 0; l < n; ++l) { + sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4)) + + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4)) + + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4)) + + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4)); + sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4)) + + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4)) + + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4)) + + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4)); + } + tmp += d * sum; + + } +#else + + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14 + const int in = offset/8; // 0 or 1 + const int im = offset%8; // 0...7 + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + offset; + const uint8_t * q = x[i].qs + offset; + const uint8_t * s = x[i].scales; + + const float dall = (float)x[i].d; + + float sum = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + const uint8_t hl = x[i].hmask[im+l] >> in; + const uint8_t ql = q[l]; + sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4)) + + y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4)) + + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4)) + + y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4)); + } + tmp += sum; + } +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +/* +DPCT1110:6: The total declared local variable size in device function +dequantize_mul_mat_vec_q4_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q4_K * x = (const block_q4_K *)vx + ib0; + +#if QK_K == 256 + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; + + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4 + + const int il = tid/step; // 0...3 + const int ir = tid - step*il; // 0...7 or 0...3 + const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4 + + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + +#if K_QUANTS_PER_ITERATION == 2 + uint32_t q32[4]; + const uint8_t * q4 = (const uint8_t *)q32; +#else + uint16_t q16[4]; + const uint8_t * q4 = (const uint8_t *)q16; +#endif + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y1 = yy + i*QK_K + y_offset; + const float * y2 = y1 + 128; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint16_t * a = (const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + +#if K_QUANTS_PER_ITERATION == 2 + const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset); + const uint32_t * q2 = q1 + 16; + + q32[0] = q1[0] & 0x0f0f0f0f; + q32[1] = q1[0] & 0xf0f0f0f0; + q32[2] = q2[0] & 0x0f0f0f0f; + q32[3] = q2[0] & 0xf0f0f0f0; + + sycl::float4 s = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + for (int l = 0; l < 4; ++l) { + s.x() += y1[l] * q4[l + 0]; s.y() += y1[l + 32] * q4[l + 4]; + s.z() += y2[l] * q4[l + 8]; s.w() += y2[l + 32] * q4[l + 12]; + smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + } + tmp += dall * (s.x() * sc[0] + s.y() * sc[1] * 1.f / 16.f + + s.z() * sc[4] + s.w() * sc[5] * 1.f / 16.f) - + dmin * smin; +#else + const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset); + const uint16_t * q2 = q1 + 32; + + q16[0] = q1[0] & 0x0f0f; + q16[1] = q1[0] & 0xf0f0; + q16[2] = q2[0] & 0x0f0f; + q16[3] = q2[0] & 0xf0f0; + + float4 s = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + for (int l = 0; l < 2; ++l) { + s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2]; + s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6]; + smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + } + tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin; +#endif + + } +#else + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); + + const int step = tid * K_QUANTS_PER_ITERATION; + + uint16_t aux16[2]; + const uint8_t * s = (const uint8_t *)aux16; + + float tmp = 0; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + const uint8_t * q = x[i].qs + step; + const float * y = yy + i*QK_K + step; + const uint16_t * a = (const uint16_t *)x[i].scales; + aux16[0] = a[0] & 0x0f0f; + aux16[1] = (a[0] >> 4) & 0x0f0f; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; + float sum = 0.f; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2]) + + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2]) + + y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3]) + + y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]); + } + tmp += sum; + } + +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (tid == 0) { + dst[row] = tmp; + } +} + +/* +DPCT1110:7: The total declared local variable size in device function +dequantize_mul_mat_vec_q5_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, + const sycl::nd_item<3> &item_ct1) { + + const int row = item_ct1.get_group(2); + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q5_K * x = (const block_q5_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; + + const int tid = item_ct1.get_local_id(2) / 2; // 0...15 + const int ix = item_ct1.get_local_id(2) % 2; + + const int il = tid/4; // 0...3 + const int ir = tid - 4*il;// 0...3 + const int n = 2; + + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + const uint8_t hm1 = 1 << (2*im); + const uint8_t hm2 = hm1 << 4; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + + uint16_t q16[8]; + const uint8_t * q4 = (const uint8_t *)q16; + + for (int i = ix; i < num_blocks_per_row; i += 2) { + + const uint8_t * ql1 = x[i].qs + q_offset; + const uint8_t * qh = x[i].qh + l0; + const float * y1 = yy + i*QK_K + y_offset; + const float * y2 = y1 + 128; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint16_t * a = (const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + + sycl::float4 sum = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + const uint16_t * q1 = (const uint16_t *)ql1; + const uint16_t * q2 = q1 + 32; + q16[0] = q1[0] & 0x0f0f; + q16[1] = q1[8] & 0x0f0f; + q16[2] = (q1[0] >> 4) & 0x0f0f; + q16[3] = (q1[8] >> 4) & 0x0f0f; + q16[4] = q2[0] & 0x0f0f; + q16[5] = q2[8] & 0x0f0f; + q16[6] = (q2[0] >> 4) & 0x0f0f; + q16[7] = (q2[8] >> 4) & 0x0f0f; + for (int l = 0; l < n; ++l) { + sum.x() += + y1[l + 0] * (q4[l + 0] + (qh[l + 0] & (hm1 << 0) ? 16 : 0)) + + y1[l + 16] * (q4[l + 2] + (qh[l + 16] & (hm1 << 0) ? 16 : 0)); + sum.y() += + y1[l + 32] * (q4[l + 4] + (qh[l + 0] & (hm1 << 1) ? 16 : 0)) + + y1[l + 48] * (q4[l + 6] + (qh[l + 16] & (hm1 << 1) ? 16 : 0)); + sum.z() += + y2[l + 0] * (q4[l + 8] + (qh[l + 0] & (hm2 << 0) ? 16 : 0)) + + y2[l + 16] * (q4[l + 10] + (qh[l + 16] & (hm2 << 0) ? 16 : 0)); + sum.w() += + y2[l + 32] * (q4[l + 12] + (qh[l + 0] & (hm2 << 1) ? 16 : 0)) + + y2[l + 48] * (q4[l + 14] + (qh[l + 16] & (hm2 << 1) ? 16 : 0)); + smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3] + + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7]; + } + tmp += dall * (sum.x() * sc[0] + sum.y() * sc[1] + sum.z() * sc[4] + + sum.w() * sc[5]) - + dmin * smin; + } + +#else + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); + const int step = tid * K_QUANTS_PER_ITERATION; + const int im = step/8; + const int in = step%8; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + const uint8_t * q = x[i].qs + step; + const int8_t * s = x[i].scales; + const float * y = yy + i*QK_K + step; + const float d = x[i].d; + float sum = 0.f; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + const uint8_t h = x[i].qh[in+j] >> im; + sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16)) + + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16)) + + y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16)) + + y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16)); + } + tmp += sum; + } +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q6_K * x = (const block_q6_K *)vx + ib0; + +#if QK_K == 256 + + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0, 1 + + const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 + +#if K_QUANTS_PER_ITERATION == 1 + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 + const int is = 0; +#else + const int l0 = 4 * in; // 0, 4, 8, ..., 28 + const int is = in / 4; +#endif + const int ql_offset = 64*im + l0; + const int qh_offset = 32*im + l0; + const int s_offset = 8*im + is; + const int y_offset = 128*im + l0; + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * ql = x[i].ql + ql_offset; + const uint8_t * qh = x[i].qh + qh_offset; + const int8_t * s = x[i].scales + s_offset; + + const float d = x[i].d; + +#if K_QUANTS_PER_ITERATION == 1 + float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32) + + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32) + + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32) + + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32) + + y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32) + + y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32) + + y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32) + +y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32); + tmp += sum; +#else + float sum = 0; + for (int l = 0; l < 4; ++l) { + sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32) + + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32) + + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32) + + y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32); + } + tmp += sum; +#endif + + } + +#else + + const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION); // 0...7 + const int ix = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION); // 0...3 + + const int step = tid * K_QUANTS_PER_ITERATION; + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + step; + const uint8_t * ql = x[i].ql + step; + const uint8_t * qh = x[i].qh + step; + const int8_t * s = x[i].scales; + + const float d = x[i+0].d; + + float sum = 0; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32) + + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32) + + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32) + + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32); + } + tmp += sum; + + } + +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (tid == 0) { + dst[row] = tmp; + } +} + + +static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + 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) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + 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) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + 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) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + 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) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + 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) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 + 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, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + 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, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + 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, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + 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, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + 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, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + 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, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const sycl::range<3> block_dims(1, 1, 32); + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + 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, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + 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, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); + }); +} + +void ggml_sycl_op_dequantize_mul_mat_vec( + ggml_backend_sycl_context & ctx, + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream) { + + const int64_t ne00 = src0->ne[0]; + const int64_t row_diff = row_high - row_low; + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics +#ifdef GGML_SYCL_F16 + ggml_sycl_pool_alloc src1_dfloat_a(ctx.pool()); + sycl::half *src1_dfloat = nullptr; // dfloat == half + + bool src1_convert_f16 = + src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || + src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 || + src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; + + if (src1_convert_f16) { + src1_dfloat = src1_dfloat_a.alloc(ne00); + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + GGML_ASSERT(to_fp16_sycl != nullptr); + to_fp16_sycl(src1_ddf_i, src1_dfloat, ne00, stream); + } +#else + const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion +#endif // GGML_SYCL_F16 + + switch (src0->type) { + case GGML_TYPE_Q4_0: + dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_1: + dequantize_mul_mat_vec_q4_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + dequantize_mul_mat_vec_q5_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + dequantize_mul_mat_vec_q5_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + dequantize_mul_mat_vec_q8_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + dequantize_mul_mat_vec_q2_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + dequantize_mul_mat_vec_q3_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + dequantize_mul_mat_vec_q6_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_F16: + convert_mul_mat_vec_f16_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + default: + printf("ggml_sycl_op_dequantize_mul_mat_vec unsupported GGML_TYPE %d\n", src0->type); + GGML_ASSERT(false); + break; + } + + (void) src1; + (void) dst; + (void) src1_ddq_i; + (void) src1_ncols; + (void) src1_padded_row_size; +} diff --git a/ggml-sycl/dmmv.hpp b/ggml-sycl/dmmv.hpp new file mode 100644 index 0000000000000..bd83735641533 --- /dev/null +++ b/ggml-sycl/dmmv.hpp @@ -0,0 +1,27 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_DMMV_HPP +#define GGML_SYCL_DMMV_HPP + +#include "common.hpp" + + +void ggml_sycl_op_dequantize_mul_mat_vec( + ggml_backend_sycl_context & ctx, + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream); + +#endif // GGML_SYCL_DMMV_HPP diff --git a/ggml-sycl/mmq.cpp b/ggml-sycl/mmq.cpp new file mode 100644 index 0000000000000..b514f00404e1f --- /dev/null +++ b/ggml-sycl/mmq.cpp @@ -0,0 +1,3031 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#include "mmq.hpp" +#include "vecdotq.hpp" + +typedef void (*allocate_tiles_sycl_t)( + int** x_ql, + sycl::half2** x_dm, + int** x_qh, + int** x_sc); +typedef void (*load_tiles_sycl_t)( + const void* __restrict__ vx, + int* __restrict__ x_ql, + sycl::half2* __restrict__ x_dm, + int* __restrict__ x_qh, + int* __restrict__ x_sc, + const int& i_offset, + const int& i_max, + const int& k, + const int& blocks_per_row); +typedef float (*vec_dot_q_mul_mat_sycl_t)( + const int* __restrict__ x_ql, + const sycl::half2* __restrict__ x_dm, + const int* __restrict__ x_qh, + const int* __restrict__ x_sc, + const int* __restrict__ y_qs, + const sycl::half2* __restrict__ y_ms, + const int& i, + const int& j, + const int& k); + + +template +static __dpct_inline__ void +allocate_tiles_q4_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_qs_q4_0, float *tile_x_d_q4_0) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_qs_q4_0; + *x_dm = (sycl::half2 *)tile_x_d_q4_0; +} + +template +static __dpct_inline__ void +load_tiles_q4_0(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_0; + const int kqsx = k % QI4_0; + + const block_q4_0 * bx0 = (const block_q4_0 *) vx; + + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); + // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_0; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) { + int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d; + } +} + +static __dpct_inline__ float vec_dot_q4_0_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const float * x_dmf = (const float *) x_dm; + + int u[2*VDR_Q4_0_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE]; + } + + return vec_dot_q4_0_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0], + y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +template +static __dpct_inline__ void +allocate_tiles_q4_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_qs_q4_1, sycl::half2 *tile_x_dm_q4_1) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_qs_q4_1; + *x_dm = tile_x_dm_q4_1; +} + + +template +static __dpct_inline__ void +load_tiles_q4_1(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_1; + const int kqsx = k % QI4_1; + + const block_q4_1 * bx0 = (const block_q4_1 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_1; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) { + int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm; + } +} + +static __dpct_inline__ float vec_dot_q4_1_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + + int u[2*VDR_Q4_1_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE]; + } + + return vec_dot_q4_1_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1], + y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +template +static __dpct_inline__ void +allocate_tiles_q5_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q5_0, float *tile_x_d_q5_0) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_ql_q5_0; + *x_dm = (sycl::half2 *)tile_x_d_q5_0; +} + +template +static __dpct_inline__ void +load_tiles_q5_0(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_0; + const int kqsx = k % QI5_0; + + const block_q5_0 * bx0 = (const block_q5_0 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx; + + const int ql = get_int_from_uint8(bxi->qs, kqsx); + const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0)); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + qs0 = dpct::vectorized_binary( + qs0, 0x10101010, dpct::sub_sat()); // subtract 16 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + qs1 = dpct::vectorized_binary( + qs1, 0x10101010, dpct::sub_sat()); // subtract 16 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_0; + const int kbxd = k % blocks_per_tile_x_row; + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) { + int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d; + } +} + +static __dpct_inline__ float vec_dot_q5_0_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0; + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + int u[2*VDR_Q5_0_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE]; + } + + return vec_dot_q8_0_q8_1_impl + (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +template +static __dpct_inline__ void +allocate_tiles_q5_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q5_1, sycl::half2 *tile_x_dm_q5_1) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_ql_q5_1; + *x_dm = tile_x_dm_q5_1; +} + +template +static __dpct_inline__ void +load_tiles_q5_1(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_1; + const int kqsx = k % QI5_1; + + const block_q5_1 * bx0 = (const block_q5_1 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx; + + const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); + const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1)); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_1; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) { + int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm; + } +} + +static __dpct_inline__ float vec_dot_q5_1_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1; + + int u[2*VDR_Q5_1_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE]; + } + + return vec_dot_q8_1_q8_1_impl + (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +template +static __dpct_inline__ void +allocate_tiles_q8_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_qs_q8_0, float *tile_x_d_q8_0) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_qs_q8_0; + *x_dm = (sycl::half2 *)tile_x_d_q8_0; +} + +template +static __dpct_inline__ void +load_tiles_q8_0(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI8_0; + const int kqsx = k % QI8_0; + float * x_dmf = (float *) x_dm; + + const block_q8_0 * bx0 = (const block_q8_0 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI8_0; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) { + int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d; + } +} + +static __dpct_inline__ float vec_dot_q8_0_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + return vec_dot_q8_0_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0], + y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]); +} + +template +static __dpct_inline__ void +allocate_tiles_q2_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q2_K, sycl::half2 *tile_x_dm_q2_K, + int *tile_x_sc_q2_K) { + (void)x_qh; + + *x_ql = tile_x_ql_q2_K; + *x_dm = tile_x_dm_q2_K; + *x_sc = tile_x_sc_q2_K; +} + +template +static __dpct_inline__ void +load_tiles_q2_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI2_K; + const int kqsx = k % QI2_K; + + const block_q2_K * bx0 = (const block_q2_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI2_K; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) { + int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { + int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4); + + x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4)); + } +} + +#define VDR_Q2_K_Q8_1_MMQ 2 +// contiguous u/y values +static __dpct_inline__ float +vec_dot_q2_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ scales, + const sycl::half2 &dm2, const float &d8) { + + int sumi_d = 0; + int sumi_m = 0; + +#pragma unroll + for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) { + int sumi_d_sc = 0; + + const int sc = scales[i0 / (QI8_1/2)]; + + // fill int with 4x m + int m = sc >> 4; + m |= m << 8; + m |= m << 16; + +#pragma unroll + for (int i = i0; i < i0 + QI8_1/2; ++i) { + sumi_d_sc = dpct::dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product + sumi_m = dpct::dp4a(m, u[i], + sumi_m); // multiply sum of q8_1 values with m + } + + sumi_d += sumi_d_sc * (sc & 0xF); + } + + const sycl::float2 dm2f = + dm2.convert(); + + return d8 * (dm2f.x() * sumi_d - dm2f.y() * sumi_m); +} + +static __dpct_inline__ float vec_dot_q2_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const int kbx = k / QI2_K; + const int ky = (k % QI2_K) * QR2_K; + const float * y_df = (const float *) y_ds; + + int v[QR2_K*VDR_Q2_K_Q8_1_MMQ]; + + const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2); + const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2)); + +#pragma unroll + for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) { + v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303; + } + + const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4; + + const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE; + return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]); +} + +template +static __dpct_inline__ void +allocate_tiles_q3_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q3_K, sycl::half2 *tile_x_dm_q3_K, + int *tile_x_qh_q3_K, int *tile_x_sc_q3_K) { + + *x_ql = tile_x_ql_q3_K; + *x_dm = tile_x_dm_q3_K; + *x_qh = tile_x_qh_q3_K; + *x_sc = tile_x_sc_q3_K; +} + +template +static __dpct_inline__ void +load_tiles_q3_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI3_K; + const int kqsx = k % QI3_K; + + const block_q3_K * bx0 = (const block_q3_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI3_K; + const int kbxd = k % blocks_per_tile_x_row; + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) { + int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) { + int i = i0 + i_offset * 2 + k / (WARP_SIZE/2); + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2); + + // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted + x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2)); + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { + int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4); + + const int ksc = k % (QI3_K/4); + + const int ksc_low = ksc % (QI3_K/8); + const int shift_low = 4 * (ksc / (QI3_K/8)); + const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; + + const int ksc_high = QI3_K/8; + const int shift_high = 2 * ksc; + const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; + + const int sc = dpct::vectorized_binary( + sc_low | sc_high, 0x20202020, dpct::sub_sat()); + + x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc; + } +} + +#define VDR_Q3_K_Q8_1_MMQ 2 +// contiguous u/y values +static __dpct_inline__ float +vec_dot_q3_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, + const int8_t *__restrict__ scales, const float &d3, + const float &d8) { + + int sumi = 0; + +#pragma unroll + for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) { + int sumi_sc = 0; + + for (int i = i0; i < i0 + QI8_1/2; ++i) { + sumi_sc = dpct::dp4a(v[i], u[i], sumi_sc); // SIMD dot product + } + + sumi += sumi_sc * scales[i0 / (QI8_1/2)]; + } + + return d3*d8 * sumi; +} + +static __dpct_inline__ float vec_dot_q3_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + + const int kbx = k / QI3_K; + const int ky = (k % QI3_K) * QR3_K; + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4; + + int v[QR3_K*VDR_Q3_K_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) { + const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2); + const int shift = 2 * ((ky % 32) / 8); + const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303; + + const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8); + const int vlh = (vh << 2) & 0x04040404; + + v[l] = dpct::vectorized_binary(vll, vlh, dpct::sub_sat()); + } + + const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE; + return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]); +} + +template +static __dpct_inline__ void +allocate_tiles_q4_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q4_K, sycl::half2 *tile_x_dm_q4_K, + int *tile_x_sc_q4_K) { + (void)x_qh; + + *x_ql = tile_x_ql_q4_K; + *x_dm = tile_x_dm_q4_K; + *x_sc = tile_x_sc_q4_K; +} + +template +static __dpct_inline__ void +load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_K; // == 0 if QK_K == 256 + const int kqsx = k % QI4_K; // == k if QK_K == 256 + + const block_q4_K * bx0 = (const block_q4_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) { + int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; + +#if QK_K == 256 + x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; +#else + x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]}; +#endif + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = k % (WARP_SIZE/8); + + // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 + int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits + scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits + + x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; + } +} + + +#define VDR_Q4_K_Q8_1_MMQ 8 + +// contiguous u/y values +static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_mmq( + const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, + const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) { + int sumi_d = 0; + +#pragma unroll + for (int j = 0; j < QI8_1; ++j) { + sumi_d = dpct::dp4a((v[j] >> (4 * i)) & 0x0F0F0F0F, + u[i * QI8_1 + j], sumi_d); // SIMD dot product + } + + const sycl::float2 ds8f = + ds8[i].convert(); + + sumf_d += ds8f.x() * (sc[i] * sumi_d); + sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val + } + + const sycl::float2 dm4f = + dm4.convert(); + + return dm4f.x() * sumf_d - dm4f.y() * sumf_m; +} + + +static __dpct_inline__ float vec_dot_q4_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8); + + const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE; + return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8, + x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]); +} + +template +static __dpct_inline__ void +allocate_tiles_q5_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q5_K, sycl::half2 *tile_x_dm_q5_K, + int *tile_x_sc_q5_K) { + (void)x_qh; + + *x_ql = tile_x_ql_q5_K; + *x_dm = tile_x_dm_q5_K; + *x_sc = tile_x_sc_q5_K; +} + +template +static __dpct_inline__ void +load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_K; // == 0 if QK_K == 256 + const int kqsx = k % QI5_K; // == k if QK_K == 256 + + const block_q5_K * bx0 = (const block_q5_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx; + const int ky = QR5_K*kqsx; + + const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4)); + const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010; + const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010; + + const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0; + const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4); + + x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0; + x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) { + int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; + +#if QK_K == 256 + x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; +#endif + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = k % (WARP_SIZE/8); + + // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 + int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits + scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits + + x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; + } +} + +#define VDR_Q5_K_Q8_1_MMQ 8 + +// contiguous u/y values +static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_mmq( + const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, + const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) { + int sumi_d = 0; + +#pragma unroll + for (int j = 0; j < QI8_1; ++j) { + sumi_d = dpct::dp4a(v[i * QI8_1 + j], u[i * QI8_1 + j], + sumi_d); // SIMD dot product + } + + const sycl::float2 ds8f = + ds8[i].convert(); + + sumf_d += ds8f.x() * (sc[i] * sumi_d); + sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val + } + + const sycl::float2 dm4f = + dm4.convert(); + + return dm4f.x() * sumf_d - dm4f.y() * sumf_m; +} + +static __dpct_inline__ float vec_dot_q5_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8); + + const int index_x = i * (QR5_K*WARP_SIZE + 1) + QR5_K*k; + const int index_y = j * WARP_SIZE + (QR5_K*k) % WARP_SIZE; + return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8, + x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]); +} + +template +static __dpct_inline__ void +allocate_tiles_q6_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_sc) { + (void)x_qh; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; + *x_sc = tile_x_sc; +} + +template +static __dpct_inline__ void +load_tiles_q6_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI6_K; // == 0 if QK_K == 256 + const int kqsx = k % QI6_K; // == k if QK_K == 256 + + const block_q6_K * bx0 = (const block_q6_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx; + const int ky = QR6_K*kqsx; + + const int ql = get_int_from_uint8(bxi->ql, kqsx); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4)); + const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030; + const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030; + + const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0; + const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2); + + x_ql[i * (2 * WARP_SIZE + 1) + kq0] = + dpct::vectorized_binary(ql0 | qh0, 0x20202020, + dpct::sub_sat()); + x_ql[i * (2 * WARP_SIZE + 1) + kq1] = + dpct::vectorized_binary(ql1 | qh1, 0x20202020, + dpct::sub_sat()); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) { + int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4; + + x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8)); + } +} + +#define VDR_Q6_K_Q8_1_MMQ 8 + +// contiguous u/y values +static __dpct_inline__ float +vec_dot_q6_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, + const int8_t *__restrict__ sc, const float &d6, + const float *__restrict__ d8) { + + float sumf_d = 0.0f; + +#pragma unroll + for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) { + sycl::int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale + +#pragma unroll + for (int i = i0; i < i0 + 2; ++i) { + sumi_d.x() = dpct::dp4a(v[2 * i + 0], u[2 * i + 0], + sumi_d.x()); // SIMD dot product + sumi_d.x() = dpct::dp4a(v[2 * i + 1], u[2 * i + 1], + sumi_d.x()); // SIMD dot product + + sumi_d.y() = dpct::dp4a(v[2 * i + 4], u[2 * i + 4], + sumi_d.y()); // SIMD dot product + sumi_d.y() = dpct::dp4a(v[2 * i + 5], u[2 * i + 5], + sumi_d.y()); // SIMD dot product + } + + sumf_d += d8[i0 / 4] * + (sc[i0 / 2 + 0] * sumi_d.x() + sc[i0 / 2 + 1] * sumi_d.y()); + } + + return d6 * sumf_d; +} + +static __dpct_inline__ float vec_dot_q6_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]); + + const int index_x = i * (QR6_K*WARP_SIZE + 1) + QR6_K*k; + const int index_y = j * WARP_SIZE + (QR6_K*k) % WARP_SIZE; + return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]); +} + +template +/* +DPCT1110:8: The total declared local variable size in device function mul_mat_q +exceeds 128 bytes and may cause high register pressure. Consult with your +hardware vendor to find the total register size available and adjust the code, +or use smaller sub-group size to avoid high register pressure. +*/ +static __dpct_inline__ void +mul_mat_q(const void *__restrict__ vx, const void *__restrict__ vy, + float *__restrict__ dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, + int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_qh, + int *tile_x_sc, const sycl::nd_item<3> &item_ct1, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + const int blocks_per_row_x = ncols_x / qk; + const int blocks_per_col_y = nrows_y / QK8_1; + const int blocks_per_warp = WARP_SIZE / qi; + + const int & ncols_dst = ncols_y; + + const int row_dst_0 = item_ct1.get_group(2) * mmq_y; + const int & row_x_0 = row_dst_0; + + const int col_dst_0 = item_ct1.get_group(1) * mmq_x; + const int & col_y_0 = col_dst_0; + + float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}}; + + for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) { + + load_tiles(x + row_x_0 * blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, + tile_x_qh, tile_x_sc, item_ct1.get_local_id(1), + nrows_x - row_x_0 - 1, item_ct1.get_local_id(2), + blocks_per_row_x); + +#pragma unroll + for (int ir = 0; ir < qr; ++ir) { + const int kqs = ir * WARP_SIZE + item_ct1.get_local_id(2); + const int kbxd = kqs / QI8_1; + +#pragma unroll + for (int i = 0; i < mmq_x; i += nwarps) { + const int col_y_eff = dpct::min( + (unsigned int)(col_y_0 + item_ct1.get_local_id(1) + i), + ncols_y - 1); // to prevent out-of-bounds memory accesses + + const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd]; + + const int index_y = (item_ct1.get_local_id(1) + i) * WARP_SIZE + + kqs % WARP_SIZE; + tile_y_qs[index_y] = get_int_from_int8_aligned( + by0->qs, item_ct1.get_local_id(2) % QI8_1); + } + +#pragma unroll + for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) { + const int ids = + (ids0 + item_ct1.get_local_id(1) * QI8_1 + + item_ct1.get_local_id(2) / (WARP_SIZE / QI8_1)) % + mmq_x; + const int kby = item_ct1.get_local_id(2) % (WARP_SIZE / QI8_1); + const int col_y_eff = sycl::min(col_y_0 + ids, ncols_y - 1); + + // if the sum is not needed it's faster to transform the scale to f32 ahead of time + const sycl::half2 *dsi_src = + &y[col_y_eff * blocks_per_col_y + ib0 * (qk / QK8_1) + + ir * (WARP_SIZE / QI8_1) + kby] + .ds; + sycl::half2 *dsi_dst = + &tile_y_ds[ids * (WARP_SIZE / QI8_1) + kby]; + if (need_sum) { + *dsi_dst = *dsi_src; + } else { + float * dfi_dst = (float *) dsi_dst; + *dfi_dst = (*dsi_src)[0]; + } + } + + /* + DPCT1118:9: SYCL group functions and algorithms must be encountered + in converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:56: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + +// #pragma unroll // unrolling this loop causes too much register pressure + for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) { +#pragma unroll + for (int j = 0; j < mmq_x; j += nwarps) { +#pragma unroll + for (int i = 0; i < mmq_y; i += WARP_SIZE) { + sum[i / WARP_SIZE][j / nwarps] += vec_dot( + tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, + tile_y_qs, tile_y_ds, item_ct1.get_local_id(2) + i, + item_ct1.get_local_id(1) + j, k); + } + } + } + + /* + DPCT1118:10: SYCL group functions and algorithms must be encountered + in converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:57: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + } + } + +#pragma unroll + for (int j = 0; j < mmq_x; j += nwarps) { + const int col_dst = col_dst_0 + j + item_ct1.get_local_id(1); + + if (col_dst >= ncols_dst) { + return; + } + +#pragma unroll + for (int i = 0; i < mmq_y; i += WARP_SIZE) { + const int row_dst = row_dst_0 + item_ct1.get_local_id(2) + i; + + if (row_dst >= nrows_dst) { + continue; + } + + dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps]; + } + } +} + +#define MMQ_X_Q4_0_RDNA2 64 +#define MMQ_Y_Q4_0_RDNA2 128 +#define NWARPS_Q4_0_RDNA2 8 +#define MMQ_X_Q4_0_RDNA1 64 +#define MMQ_Y_Q4_0_RDNA1 64 +#define NWARPS_Q4_0_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q4_0_AMPERE 4 +#define MMQ_Y_Q4_0_AMPERE 32 +#define NWARPS_Q4_0_AMPERE 4 +#else +#define MMQ_X_Q4_0_AMPERE 64 +#define MMQ_Y_Q4_0_AMPERE 128 +#define NWARPS_Q4_0_AMPERE 4 +#endif +#define MMQ_X_Q4_0_PASCAL 64 +#define MMQ_Y_Q4_0_PASCAL 64 +#define NWARPS_Q4_0_PASCAL 8 + +template static void + mul_mat_q4_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_0, float *tile_x_d_q4_0, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + + const int mmq_x = MMQ_X_Q4_0_AMPERE; + const int mmq_y = MMQ_Y_Q4_0_AMPERE; + const int nwarps = NWARPS_Q4_0_AMPERE; + allocate_tiles_q4_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_qs_q4_0, tile_x_d_q4_0); + mul_mat_q, VDR_Q4_0_Q8_1_MMQ, + vec_dot_q4_0_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q4_1_RDNA2 64 +#define MMQ_Y_Q4_1_RDNA2 128 +#define NWARPS_Q4_1_RDNA2 8 +#define MMQ_X_Q4_1_RDNA1 64 +#define MMQ_Y_Q4_1_RDNA1 64 +#define NWARPS_Q4_1_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q4_1_AMPERE 4 +#define MMQ_Y_Q4_1_AMPERE 32 +#define NWARPS_Q4_1_AMPERE 4 +#else +#define MMQ_X_Q4_1_AMPERE 64 +#define MMQ_Y_Q4_1_AMPERE 128 +#define NWARPS_Q4_1_AMPERE 4 +#endif +#define MMQ_X_Q4_1_PASCAL 64 +#define MMQ_Y_Q4_1_PASCAL 64 +#define NWARPS_Q4_1_PASCAL 8 + +template static void + mul_mat_q4_1( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_1, + sycl::half2 *tile_x_dm_q4_1, int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q4_1_AMPERE; + const int mmq_y = MMQ_Y_Q4_1_AMPERE; + const int nwarps = NWARPS_Q4_1_AMPERE; + allocate_tiles_q4_1(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_qs_q4_1, tile_x_dm_q4_1); + mul_mat_q, VDR_Q4_1_Q8_1_MMQ, + vec_dot_q4_1_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q5_0_RDNA2 64 +#define MMQ_Y_Q5_0_RDNA2 128 +#define NWARPS_Q5_0_RDNA2 8 +#define MMQ_X_Q5_0_RDNA1 64 +#define MMQ_Y_Q5_0_RDNA1 64 +#define NWARPS_Q5_0_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q5_0_AMPERE 4 +#define MMQ_Y_Q5_0_AMPERE 32 +#define NWARPS_Q5_0_AMPERE 4 +#else +#define MMQ_X_Q5_0_AMPERE 128 +#define MMQ_Y_Q5_0_AMPERE 64 +#define NWARPS_Q5_0_AMPERE 4 +#endif +#define MMQ_X_Q5_0_PASCAL 64 +#define MMQ_Y_Q5_0_PASCAL 64 +#define NWARPS_Q5_0_PASCAL 8 + +template static void + mul_mat_q5_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_0, float *tile_x_d_q5_0, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q5_0_AMPERE; + const int mmq_y = MMQ_Y_Q5_0_AMPERE; + const int nwarps = NWARPS_Q5_0_AMPERE; + allocate_tiles_q5_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q5_0, tile_x_d_q5_0); + mul_mat_q, VDR_Q5_0_Q8_1_MMQ, + vec_dot_q5_0_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q5_1_RDNA2 64 +#define MMQ_Y_Q5_1_RDNA2 128 +#define NWARPS_Q5_1_RDNA2 8 +#define MMQ_X_Q5_1_RDNA1 64 +#define MMQ_Y_Q5_1_RDNA1 64 +#define NWARPS_Q5_1_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q5_1_AMPERE 4 +#define MMQ_Y_Q5_1_AMPERE 32 +#define NWARPS_Q5_1_AMPERE 4 +#else +#define MMQ_X_Q5_1_AMPERE 128 +#define MMQ_Y_Q5_1_AMPERE 64 +#define NWARPS_Q5_1_AMPERE 4 +#endif +#define MMQ_X_Q5_1_PASCAL 64 +#define MMQ_Y_Q5_1_PASCAL 64 +#define NWARPS_Q5_1_PASCAL 8 + +template static void +mul_mat_q5_1( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_1, + sycl::half2 *tile_x_dm_q5_1, int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q5_1_AMPERE; + const int mmq_y = MMQ_Y_Q5_1_AMPERE; + const int nwarps = NWARPS_Q5_1_AMPERE; + allocate_tiles_q5_1(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q5_1, tile_x_dm_q5_1); + mul_mat_q, VDR_Q5_1_Q8_1_MMQ, + vec_dot_q5_1_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q8_0_RDNA2 64 +#define MMQ_Y_Q8_0_RDNA2 128 +#define NWARPS_Q8_0_RDNA2 8 +#define MMQ_X_Q8_0_RDNA1 64 +#define MMQ_Y_Q8_0_RDNA1 64 +#define NWARPS_Q8_0_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q8_0_AMPERE 4 +#define MMQ_Y_Q8_0_AMPERE 32 +#define NWARPS_Q8_0_AMPERE 4 +#else +#define MMQ_X_Q8_0_AMPERE 128 +#define MMQ_Y_Q8_0_AMPERE 64 +#define NWARPS_Q8_0_AMPERE 4 +#endif +#define MMQ_X_Q8_0_PASCAL 64 +#define MMQ_Y_Q8_0_PASCAL 64 +#define NWARPS_Q8_0_PASCAL 8 + +template static void + mul_mat_q8_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q8_0, float *tile_x_d_q8_0, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q8_0_AMPERE; + const int mmq_y = MMQ_Y_Q8_0_AMPERE; + const int nwarps = NWARPS_Q8_0_AMPERE; + allocate_tiles_q8_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_qs_q8_0, tile_x_d_q8_0); + mul_mat_q, VDR_Q8_0_Q8_1_MMQ, + vec_dot_q8_0_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q2_K_RDNA2 64 +#define MMQ_Y_Q2_K_RDNA2 128 +#define NWARPS_Q2_K_RDNA2 8 +#define MMQ_X_Q2_K_RDNA1 128 +#define MMQ_Y_Q2_K_RDNA1 32 +#define NWARPS_Q2_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q2_K_AMPERE 4 +#define MMQ_Y_Q2_K_AMPERE 32 +#define NWARPS_Q2_K_AMPERE 4 +#else +#define MMQ_X_Q2_K_AMPERE 64 +#define MMQ_Y_Q2_K_AMPERE 128 +#define NWARPS_Q2_K_AMPERE 4 +#endif +#define MMQ_X_Q2_K_PASCAL 64 +#define MMQ_Y_Q2_K_PASCAL 64 +#define NWARPS_Q2_K_PASCAL 8 + +template static void +mul_mat_q2_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q2_K, + sycl::half2 *tile_x_dm_q2_K, int *tile_x_sc_q2_K, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q2_K_AMPERE; + const int mmq_y = MMQ_Y_Q2_K_AMPERE; + const int nwarps = NWARPS_Q2_K_AMPERE; + allocate_tiles_q2_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q2_K, tile_x_dm_q2_K, tile_x_sc_q2_K); + mul_mat_q, VDR_Q2_K_Q8_1_MMQ, + vec_dot_q2_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q3_K_RDNA2 128 +#define MMQ_Y_Q3_K_RDNA2 64 +#define NWARPS_Q3_K_RDNA2 8 +#define MMQ_X_Q3_K_RDNA1 32 +#define MMQ_Y_Q3_K_RDNA1 128 +#define NWARPS_Q3_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q3_K_AMPERE 4 +#define MMQ_Y_Q3_K_AMPERE 32 +#define NWARPS_Q3_K_AMPERE 4 +#else +#define MMQ_X_Q3_K_AMPERE 128 +#define MMQ_Y_Q3_K_AMPERE 128 +#define NWARPS_Q3_K_AMPERE 4 +#endif +#define MMQ_X_Q3_K_PASCAL 64 +#define MMQ_Y_Q3_K_PASCAL 64 +#define NWARPS_Q3_K_PASCAL 8 + +template static void +mul_mat_q3_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q3_K, + sycl::half2 *tile_x_dm_q3_K, int *tile_x_qh_q3_K, int *tile_x_sc_q3_K, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q3_K_AMPERE; + const int mmq_y = MMQ_Y_Q3_K_AMPERE; + const int nwarps = NWARPS_Q3_K_AMPERE; + allocate_tiles_q3_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q3_K, tile_x_dm_q3_K, tile_x_qh_q3_K, + tile_x_sc_q3_K); + mul_mat_q, VDR_Q3_K_Q8_1_MMQ, + vec_dot_q3_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q4_K_RDNA2 64 +#define MMQ_Y_Q4_K_RDNA2 128 +#define NWARPS_Q4_K_RDNA2 8 +#define MMQ_X_Q4_K_RDNA1 32 +#define MMQ_Y_Q4_K_RDNA1 64 +#define NWARPS_Q4_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q4_K_AMPERE 4 +#define MMQ_Y_Q4_K_AMPERE 32 +#define NWARPS_Q4_K_AMPERE 4 +#else +#define MMQ_X_Q4_K_AMPERE 64 +#define MMQ_Y_Q4_K_AMPERE 128 +#define NWARPS_Q4_K_AMPERE 4 +#endif +#define MMQ_X_Q4_K_PASCAL 64 +#define MMQ_Y_Q4_K_PASCAL 64 +#define NWARPS_Q4_K_PASCAL 8 + +template static void + mul_mat_q4_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q4_K, + sycl::half2 *tile_x_dm_q4_K, int *tile_x_sc_q4_K, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q4_K_AMPERE; + const int mmq_y = MMQ_Y_Q4_K_AMPERE; + const int nwarps = NWARPS_Q4_K_AMPERE; + allocate_tiles_q4_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q4_K, tile_x_dm_q4_K, tile_x_sc_q4_K); + mul_mat_q, VDR_Q4_K_Q8_1_MMQ, + vec_dot_q4_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q5_K_RDNA2 64 +#define MMQ_Y_Q5_K_RDNA2 128 +#define NWARPS_Q5_K_RDNA2 8 +#define MMQ_X_Q5_K_RDNA1 32 +#define MMQ_Y_Q5_K_RDNA1 64 +#define NWARPS_Q5_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q5_K_AMPERE 4 +#define MMQ_Y_Q5_K_AMPERE 32 +#define NWARPS_Q5_K_AMPERE 4 +#else +#define MMQ_X_Q5_K_AMPERE 64 +#define MMQ_Y_Q5_K_AMPERE 128 +#define NWARPS_Q5_K_AMPERE 4 +#endif +#define MMQ_X_Q5_K_PASCAL 64 +#define MMQ_Y_Q5_K_PASCAL 64 +#define NWARPS_Q5_K_PASCAL 8 + +template static void +mul_mat_q5_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_K, + sycl::half2 *tile_x_dm_q5_K, int *tile_x_sc_q5_K, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q5_K_AMPERE; + const int mmq_y = MMQ_Y_Q5_K_AMPERE; + const int nwarps = NWARPS_Q5_K_AMPERE; + allocate_tiles_q5_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q5_K, tile_x_dm_q5_K, tile_x_sc_q5_K); + mul_mat_q, VDR_Q5_K_Q8_1_MMQ, + vec_dot_q5_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q6_K_RDNA2 64 +#define MMQ_Y_Q6_K_RDNA2 128 +#define NWARPS_Q6_K_RDNA2 8 +#define MMQ_X_Q6_K_RDNA1 32 +#define MMQ_Y_Q6_K_RDNA1 64 +#define NWARPS_Q6_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q6_K_AMPERE 4 +#define MMQ_Y_Q6_K_AMPERE 32 +#define NWARPS_Q6_K_AMPERE 4 +#else +#define MMQ_X_Q6_K_AMPERE 64 +#define MMQ_Y_Q6_K_AMPERE 64 +#define NWARPS_Q6_K_AMPERE 4 +#endif +#define MMQ_X_Q6_K_PASCAL 64 +#define MMQ_Y_Q6_K_PASCAL 64 +#define NWARPS_Q6_K_PASCAL 8 + +template static void + mul_mat_q6_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql, sycl::half2 *tile_x_dm, + int *tile_x_sc, int *tile_y_qs, sycl::half2 *tile_y_ds) { + // int * tile_x_ql = nullptr; + // sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + // int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q6_K_AMPERE; + const int mmq_y = MMQ_Y_Q6_K_AMPERE; + const int nwarps = NWARPS_Q6_K_AMPERE; + allocate_tiles_q6_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql, tile_x_dm, tile_x_sc); + mul_mat_q, VDR_Q6_K_Q8_1_MMQ, + vec_dot_q6_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q4_0_RDNA2; + mmq_y = MMQ_Y_Q4_0_RDNA2; + nwarps = NWARPS_Q4_0_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q4_0_RDNA1; + mmq_y = MMQ_Y_Q4_0_RDNA1; + nwarps = NWARPS_Q4_0_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q4_0_AMPERE; + mmq_y = MMQ_Y_Q4_0_AMPERE; + nwarps = NWARPS_Q4_0_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q4_0_PASCAL; + mmq_y = MMQ_Y_Q4_0_PASCAL; + nwarps = NWARPS_Q4_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:20: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q4_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_0_acc_ct1.get_pointer(), + tile_x_d_q4_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:21: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q4_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_0_acc_ct1.get_pointer(), + tile_x_d_q4_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q4_1_RDNA2; + mmq_y = MMQ_Y_Q4_1_RDNA2; + nwarps = NWARPS_Q4_1_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q4_1_RDNA1; + mmq_y = MMQ_Y_Q4_1_RDNA1; + nwarps = NWARPS_Q4_1_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q4_1_AMPERE; + mmq_y = MMQ_Y_Q4_1_AMPERE; + nwarps = NWARPS_Q4_1_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q4_1_PASCAL; + mmq_y = MMQ_Y_Q4_1_PASCAL; + nwarps = NWARPS_Q4_1_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:22: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q4_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_1_acc_ct1.get_pointer(), + tile_x_dm_q4_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:23: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q4_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_1_acc_ct1.get_pointer(), + tile_x_dm_q4_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q5_0_RDNA2; + mmq_y = MMQ_Y_Q5_0_RDNA2; + nwarps = NWARPS_Q5_0_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q5_0_RDNA1; + mmq_y = MMQ_Y_Q5_0_RDNA1; + nwarps = NWARPS_Q5_0_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q5_0_AMPERE; + mmq_y = MMQ_Y_Q5_0_AMPERE; + nwarps = NWARPS_Q5_0_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q5_0_PASCAL; + mmq_y = MMQ_Y_Q5_0_PASCAL; + nwarps = NWARPS_Q5_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:24: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q5_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_0_acc_ct1.get_pointer(), + tile_x_d_q5_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:25: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q5_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_0_acc_ct1.get_pointer(), + tile_x_d_q5_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q5_1_RDNA2; + mmq_y = MMQ_Y_Q5_1_RDNA2; + nwarps = NWARPS_Q5_1_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q5_1_RDNA1; + mmq_y = MMQ_Y_Q5_1_RDNA1; + nwarps = NWARPS_Q5_1_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q5_1_AMPERE; + mmq_y = MMQ_Y_Q5_1_AMPERE; + nwarps = NWARPS_Q5_1_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q5_1_PASCAL; + mmq_y = MMQ_Y_Q5_1_PASCAL; + nwarps = NWARPS_Q5_1_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:26: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q5_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_1_acc_ct1.get_pointer(), + tile_x_dm_q5_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:27: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q5_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_1_acc_ct1.get_pointer(), + tile_x_dm_q5_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q8_0_RDNA2; + mmq_y = MMQ_Y_Q8_0_RDNA2; + nwarps = NWARPS_Q8_0_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q8_0_RDNA1; + mmq_y = MMQ_Y_Q8_0_RDNA1; + nwarps = NWARPS_Q8_0_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q8_0_AMPERE; + mmq_y = MMQ_Y_Q8_0_AMPERE; + nwarps = NWARPS_Q8_0_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q8_0_PASCAL; + mmq_y = MMQ_Y_Q8_0_PASCAL; + nwarps = NWARPS_Q8_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:28: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q8_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q8_0_acc_ct1.get_pointer(), + tile_x_d_q8_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:29: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q8_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q8_0_acc_ct1.get_pointer(), + tile_x_d_q8_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q2_K_RDNA2; + mmq_y = MMQ_Y_Q2_K_RDNA2; + nwarps = NWARPS_Q2_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q2_K_RDNA1; + mmq_y = MMQ_Y_Q2_K_RDNA1; + nwarps = NWARPS_Q2_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q2_K_AMPERE; + mmq_y = MMQ_Y_Q2_K_AMPERE; + nwarps = NWARPS_Q2_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q2_K_PASCAL; + mmq_y = MMQ_Y_Q2_K_PASCAL; + nwarps = NWARPS_Q2_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:30: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K), + cgh); + sycl::local_accessor tile_x_sc_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q2_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q2_K_acc_ct1.get_pointer(), + tile_x_dm_q2_K_acc_ct1.get_pointer(), + tile_x_sc_q2_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:31: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K), + cgh); + sycl::local_accessor tile_x_sc_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q2_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q2_K_acc_ct1.get_pointer(), + tile_x_dm_q2_K_acc_ct1.get_pointer(), + tile_x_sc_q2_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + +#if QK_K == 256 + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q3_K_RDNA2; + mmq_y = MMQ_Y_Q3_K_RDNA2; + nwarps = NWARPS_Q3_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q3_K_RDNA1; + mmq_y = MMQ_Y_Q3_K_RDNA1; + nwarps = NWARPS_Q3_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q3_K_AMPERE; + mmq_y = MMQ_Y_Q3_K_AMPERE; + nwarps = NWARPS_Q3_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q3_K_PASCAL; + mmq_y = MMQ_Y_Q3_K_PASCAL; + nwarps = NWARPS_Q3_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:32: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K), + cgh); + sycl::local_accessor tile_x_qh_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh); + sycl::local_accessor tile_x_sc_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q3_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q3_K_acc_ct1.get_pointer(), + tile_x_dm_q3_K_acc_ct1.get_pointer(), + tile_x_qh_q3_K_acc_ct1.get_pointer(), + tile_x_sc_q3_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:33: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K), + cgh); + sycl::local_accessor tile_x_qh_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh); + sycl::local_accessor tile_x_sc_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q3_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q3_K_acc_ct1.get_pointer(), + tile_x_dm_q3_K_acc_ct1.get_pointer(), + tile_x_qh_q3_K_acc_ct1.get_pointer(), + tile_x_sc_q3_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +#endif +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q4_K_RDNA2; + mmq_y = MMQ_Y_Q4_K_RDNA2; + nwarps = NWARPS_Q4_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q4_K_RDNA1; + mmq_y = MMQ_Y_Q4_K_RDNA1; + nwarps = NWARPS_Q4_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q4_K_AMPERE; + mmq_y = MMQ_Y_Q4_K_AMPERE; + nwarps = NWARPS_Q4_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q4_K_PASCAL; + mmq_y = MMQ_Y_Q4_K_PASCAL; + nwarps = NWARPS_Q4_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:34: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K), + cgh); + sycl::local_accessor tile_x_sc_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q4_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q4_K_acc_ct1.get_pointer(), + tile_x_dm_q4_K_acc_ct1.get_pointer(), + tile_x_sc_q4_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:35: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K), + cgh); + sycl::local_accessor tile_x_sc_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q4_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q4_K_acc_ct1.get_pointer(), + tile_x_dm_q4_K_acc_ct1.get_pointer(), + tile_x_sc_q4_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q5_K_RDNA2; + mmq_y = MMQ_Y_Q5_K_RDNA2; + nwarps = NWARPS_Q5_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q5_K_RDNA1; + mmq_y = MMQ_Y_Q5_K_RDNA1; + nwarps = NWARPS_Q5_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q5_K_AMPERE; + mmq_y = MMQ_Y_Q5_K_AMPERE; + nwarps = NWARPS_Q5_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q5_K_PASCAL; + mmq_y = MMQ_Y_Q5_K_PASCAL; + nwarps = NWARPS_Q5_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:36: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K), + cgh); + sycl::local_accessor tile_x_sc_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q5_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_K_acc_ct1.get_pointer(), + tile_x_dm_q5_K_acc_ct1.get_pointer(), + tile_x_sc_q5_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:37: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K), + cgh); + sycl::local_accessor tile_x_sc_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q5_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_K_acc_ct1.get_pointer(), + tile_x_dm_q5_K_acc_ct1.get_pointer(), + tile_x_sc_q5_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + const int compute_capability = ggml_sycl_info().devices[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q6_K_RDNA2; + mmq_y = MMQ_Y_Q6_K_RDNA2; + nwarps = NWARPS_Q6_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q6_K_RDNA1; + mmq_y = MMQ_Y_Q6_K_RDNA1; + nwarps = NWARPS_Q6_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q6_K_AMPERE; + mmq_y = MMQ_Y_Q6_K_AMPERE; + nwarps = NWARPS_Q6_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q6_K_PASCAL; + mmq_y = MMQ_Y_Q6_K_PASCAL; + nwarps = NWARPS_Q6_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:38: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K), + cgh); + sycl::local_accessor tile_x_sc_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q6_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_acc_ct1.get_pointer(), + tile_x_dm_acc_ct1.get_pointer(), + tile_x_sc_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:39: 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. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](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( + sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K), + cgh); + sycl::local_accessor tile_x_sc_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + 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) { + mul_mat_q6_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_acc_ct1.get_pointer(), + tile_x_dm_acc_ct1.get_pointer(), + tile_x_sc_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_op_mul_mat_q( + ggml_backend_sycl_context & ctx, + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream) try { + + const int64_t ne00 = src0->ne[0]; + + const int64_t ne10 = src1->ne[0]; + GGML_ASSERT(ne10 % QK8_1 == 0); + + const int64_t ne0 = dst->ne[0]; + + const int64_t row_diff = row_high - row_low; + + int device_id; + SYCL_CHECK( + CHECK_TRY_ERROR(device_id = get_current_device_id())); + + // the main device has a larger memory buffer to hold the results from all GPUs + // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into + const int64_t nrows_dst = device_id == ctx.device ? ne0 : row_diff; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + ggml_mul_mat_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q4_1: + ggml_mul_mat_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_0: + ggml_mul_mat_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_1: + ggml_mul_mat_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q8_0: + ggml_mul_mat_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q2_K: + ggml_mul_mat_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q3_K: + ggml_mul_mat_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q4_K: + ggml_mul_mat_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_K: + ggml_mul_mat_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q6_K: + ggml_mul_mat_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + (void) src1; + (void) dst; + (void) src1_ddf_i; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} diff --git a/ggml-sycl/mmq.hpp b/ggml-sycl/mmq.hpp new file mode 100644 index 0000000000000..3f5297aaa5373 --- /dev/null +++ b/ggml-sycl/mmq.hpp @@ -0,0 +1,33 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_MMQ_HPP +#define GGML_SYCL_MMQ_HPP + +#include "common.hpp" + +void ggml_sycl_op_mul_mat_q( + ggml_backend_sycl_context & ctx, + const ggml_tensor* src0, + const ggml_tensor* src1, + ggml_tensor* dst, + const char* src0_dd_i, + const float* src1_ddf_i, + const char* src1_ddq_i, + float* dst_dd_i, + const int64_t row_low, + const int64_t row_high, + const int64_t src1_ncols, + const int64_t src1_padded_row_size, + const dpct::queue_ptr& stream); + +#endif // GGML_SYCL_MMQ_HPP diff --git a/ggml-sycl/mmvq.cpp b/ggml-sycl/mmvq.cpp new file mode 100644 index 0000000000000..23227649e2661 --- /dev/null +++ b/ggml-sycl/mmvq.cpp @@ -0,0 +1,1024 @@ +#include "mmvq.hpp" +#include "vecdotq.hpp" + + +template +static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq2_xxs_q8_1(&x[ibx], &y[iby], iqs, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq2_xs_q8_1(&x[ibx], &y[iby], iqs, iq2xs_grid, ksigns64); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq2_s_q8_1(&x[ibx], &y[iby], iqs); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq3_xxs_q8_1(&x[ibx], &y[iby], iqs, iq3xxs_grid, ksigns64); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq3_s_q8_1(&x[ibx], &y[iby], iqs, iq3s_grid); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq1_s_q8_1(&x[ibx], &y[iby], iqs, iq1s_grid_gpu); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq1_m_q8_1(&x[ibx], &y[iby], iqs); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq4_nl_q8_1(&x[ibx], &y[iby], iqs); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + + +template +static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, + const void *__restrict__ vy, + float *__restrict__ dst, const int ncols, + const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; + i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index + + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_iq4_xs_q8_1(&x[ibx], &y[iby], iqs); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_1 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK5_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK5_1 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK8_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + + +static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq2_xxs_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) { + auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0]; + auto ksigns64_ptr_ct1 = &ksigns64[0]; + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq2_xs_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) { + auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0]; + auto ksigns64_ptr_ct1 = &ksigns64[0]; + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq2_s_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) { + auto iq3xxs_grid_ptr_ct1 = &iq3xxs_grid[0]; + auto ksigns64_ptr_ct1 = &ksigns64[0]; + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq3_xxs_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) { + auto iq3s_grid_ptr_ct1 = &iq3s_grid[0]; + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq3_s_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) { + auto iq1s_grid_ptr_ct1 = &iq1s_grid_gpu[0]; + auto ksigns64_ptr_ct1 = &ksigns64[0]; + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq1_s_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq1_m_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_NL == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq4_nl_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + 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) + [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q_iq4_xs_q8_1( + vx, vy, dst, ncols, nrows, item_ct1); + }); + }); + } +} + +void ggml_sycl_op_mul_mat_vec_q( + ggml_backend_sycl_context & ctx, + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream) { + + const int64_t ne10 = src1->ne[0]; + GGML_ASSERT(ne10 % QK8_1 == 0); + + const int64_t ne00 = src0->ne[0]; + const int64_t row_diff = row_high - row_low; + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_id())); + + // the main device has a larger memory buffer to hold the results from all GPUs + // nrows_dst == nrows of the matrix that the kernel writes into + const int64_t nrows_dst = id == ctx.device ? ne00 : row_diff; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_1: + mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ1_S: + mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ1_M: + mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_XXS: + mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_XS: + mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_S: + mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ3_XXS: + mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ3_S: + mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ4_NL: + mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ4_XS: + mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + (void) src1; + (void) dst; + (void) src1_ddf_i; + (void) src1_ncols; + (void) src1_padded_row_size; +} diff --git a/ggml-sycl/mmvq.hpp b/ggml-sycl/mmvq.hpp new file mode 100644 index 0000000000000..049b43d453532 --- /dev/null +++ b/ggml-sycl/mmvq.hpp @@ -0,0 +1,27 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_MMVQ_HPP +#define GGML_SYCL_MMVQ_HPP + +#include "common.hpp" + + +void ggml_sycl_op_mul_mat_vec_q( + ggml_backend_sycl_context & ctx, + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream); + +#endif // GGML_SYCL_MMVQ_HPP diff --git a/ggml-sycl/presets.hpp b/ggml-sycl/presets.hpp index dcf0261102e91..5e6b61813ab49 100644 --- a/ggml-sycl/presets.hpp +++ b/ggml-sycl/presets.hpp @@ -18,8 +18,6 @@ #define GGML_SYCL_MAX_DEVICES 48 #define GGML_SYCL_NAME "SYCL" -// FIXME: 1024 from cuda -#define GROUP_SIZE 1024 #define WARP_SIZE 32 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses diff --git a/ggml-sycl/vecdotq.hpp b/ggml-sycl/vecdotq.hpp new file mode 100644 index 0000000000000..5e2e825463cde --- /dev/null +++ b/ggml-sycl/vecdotq.hpp @@ -0,0 +1,1161 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_VECDOTQ_HPP +#define GGML_SYCL_VECDOTQ_HPP + +#include "dpct/helper.hpp" + +typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs); + +static __dpct_inline__ int get_int_from_int8(const int8_t* x8, const int& i32) { + const uint16_t* x16 = + (const uint16_t*)(x8 + sizeof(int) * i32); // assume at least 2 byte + // alignment + + int x32 = 0; + x32 |= x16[0] << 0; + x32 |= x16[1] << 16; + + return x32; +} + +static __dpct_inline__ int get_int_from_uint8( + const uint8_t* x8, + const int& i32) { + const uint16_t* x16 = + (const uint16_t*)(x8 + sizeof(int) * i32); // assume at least 2 byte + // alignment + + int x32 = 0; + x32 |= x16[0] << 0; + x32 |= x16[1] << 16; + + return x32; +} + +static __dpct_inline__ int get_int_from_int8_aligned( + const int8_t* x8, + const int& i32) { + return *( + (const int*)(x8 + sizeof(int) * i32)); // assume at least 4 byte alignment +} + +static __dpct_inline__ int get_int_from_uint8_aligned( + const uint8_t* x8, + const int& i32) { + return *( + (const int*)(x8 + sizeof(int) * i32)); // assume at least 4 byte alignment +} + +static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4, + const uint8_t *values, + int &val1, int &val2) { + + uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32; + aux32 = q4 & 0x0f0f0f0f; + uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8); + uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8); + val1 = v1 | (v2 << 16); + aux32 = (q4 >> 4) & 0x0f0f0f0f; + v1 = values[q8[0]] | (values[q8[1]] << 8); + v2 = values[q8[2]] | (values[q8[3]] << 8); + val2 = v1 | (v2 << 16); +} + +#define VDR_Q2_K_Q8_1_MMVQ 1 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q2_K_q8_1_impl_mmvq( + const int &v, const int *__restrict__ u, const uint8_t *__restrict__ scales, + const sycl::half2 &dm2, const float *__restrict__ d8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR2_K; ++i) { + const int sc = scales[2*i]; + + const int vi = (v >> (2*i)) & 0x03030303; + + sumf_d += + d8[i] * (dpct::dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product + + // fill int with 4x m + int m = sc >> 4; + m |= m << 8; + m |= m << 16; + sumf_m += d8[i] * + dpct::dp4a( + m, u[i], + 0); // multiply constant q2_K part with sum of q8_1 values + } + + const sycl::float2 dm2f = + dm2.convert(); + + return dm2f.x() * sumf_d - dm2f.y() * sumf_m; +} + + +#define VDR_Q3_K_Q8_1_MMVQ 1 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q3_K_q8_1_impl_mmvq( + const int &vl, const int &vh, const int *__restrict__ u, + const uint8_t *__restrict__ scales, const int &scale_offset, + const float &d3, const float *__restrict__ d8) { + + float sumf = 0.0f; + +#pragma unroll + for (int i = 0; i < QR3_K; ++i) { + const int isc = scale_offset + 2*i; + + const int isc_low = isc % (QK_K/32); + const int sc_shift_low = 4 * (isc / (QK_K/32)); + const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF; + + const int isc_high = isc % (QK_K/64); + const int sc_shift_high = 2 * (isc / (QK_K/64)); + const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4; + + const int sc = (sc_low | sc_high) - 32; + + const int vil = (vl >> (2*i)) & 0x03030303; + + const int vih = ((vh >> i) << 2) & 0x04040404; + + const int vi = + dpct::vectorized_binary(vil, vih, dpct::sub_sat()); + + sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product + } + + return d3 * sumf; +} + +#define VDR_Q4_K_Q8_1_MMVQ 2 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_vmmq( + const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, + const sycl::half2 &dm4, const float *__restrict__ d8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR4_K; ++i) { + const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F; + const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F; + + const int dot1 = + dpct::dp4a(v1i, u[2 * i + 1], + dpct::dp4a(v0i, u[2 * i + 0], 0)); // SIMD dot product + const int dot2 = + dpct::dp4a(0x01010101, u[2 * i + 1], + dpct::dp4a(0x01010101, u[2 * i + 0], 0)); // sum of u + + sumf_d += d8[i] * (dot1 * sc[i]); + sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values + } + + const sycl::float2 dm4f = + dm4.convert(); + + return dm4f.x() * sumf_d - dm4f.y() * sumf_m; +} + + +#define VDR_Q5_K_Q8_1_MMVQ 2 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_vmmq( + const int *__restrict__ vl, const int *__restrict__ vh, + const int *__restrict__ u, const uint8_t *__restrict__ sc, + const uint8_t *__restrict__ m, const sycl::half2 &dm5, + const float *__restrict__ d8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR5_K; ++i) { + const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F; + const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F; + + const int vh0i = ((vh[0] >> i) << 4) & 0x10101010; + const int vh1i = ((vh[1] >> i) << 4) & 0x10101010; + + const int v0i = vl0i | vh0i; + const int v1i = vl1i | vh1i; + + const int dot1 = + dpct::dp4a(v0i, u[2 * i + 0], + dpct::dp4a(v1i, u[2 * i + 1], 0)); // SIMD dot product + const int dot2 = + dpct::dp4a(0x01010101, u[2 * i + 0], + dpct::dp4a(0x01010101, u[2 * i + 1], 0)); // sum of u + + sumf_d += d8[i] * (dot1 * sc[i]); + sumf_m += d8[i] * (dot2 * m[i]); + + } + + const sycl::float2 dm5f = + dm5.convert(); + + return dm5f.x() * sumf_d - dm5f.y() * sumf_m; +} + + +#define VDR_Q6_K_Q8_1_MMVQ 1 + +// contiguous v/x values +static __dpct_inline__ float +vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh, + const int *__restrict__ u, + const int8_t *__restrict__ scales, const float &d, + const float *__restrict__ d8) { + + float sumf = 0.0f; + +#pragma unroll + for (int i = 0; i < QR6_K; ++i) { + const int sc = scales[4*i]; + + const int vil = (vl >> (4*i)) & 0x0F0F0F0F; + + const int vih = ((vh >> (4*i)) << 4) & 0x30303030; + + const int vi = dpct::vectorized_binary( + (vil | vih), 0x20202020, dpct::sub_sat()); // vi = (vil | vih) - 32 + + sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product + } + + return d*sumf; +} + +// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called +// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q + +#define VDR_Q4_0_Q8_1_MMVQ 2 +#define VDR_Q4_0_Q8_1_MMQ 4 + +template +static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u, + const float &d4, + const sycl::half2 &ds8) { + int sumi = 0; +#pragma unroll + for (int i = 0; i < vdr; ++i) { + const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; + const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + + // SIMD dot product of quantized values + sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi); + sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); + } + + const sycl::float2 ds8f = + ds8.convert(); + + // second part effectively subtracts 8 from each quant value + return d4 * (sumi * ds8f.x() - (8 * vdr / QI4_0) * ds8f.y()); +} + +#define VDR_Q4_1_Q8_1_MMVQ 2 +#define VDR_Q4_1_Q8_1_MMQ 4 + +template +static __dpct_inline__ float vec_dot_q4_1_q8_1_impl(const int *v, const int *u, + const sycl::half2 &dm4, + const sycl::half2 &ds8) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; + const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + + // SIMD dot product of quantized values + sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi); + sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); + } + +#ifdef GGML_SYCL_F16 + const sycl::float2 tmp = + (dm4 * ds8).convert(); + const float d4d8 = tmp.x(); + const float m4s8 = tmp.y(); +#else + const sycl::float2 dm4f = + dm4.convert(); + const sycl::float2 ds8f = + ds8.convert(); + const float d4d8 = dm4f.x() * ds8f.x(); + const float m4s8 = dm4f.y() * ds8f.y(); +#endif // GGML_SYCL_F16 + + // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it + return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); +} + +#define VDR_Q5_0_Q8_1_MMVQ 2 +#define VDR_Q5_0_Q8_1_MMQ 4 + +template +static __dpct_inline__ float +vec_dot_q5_0_q8_1_impl(const int *vl, const int *vh, const int *u, + const float &d5, const sycl::half2 &ds8) { + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits + vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 + vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 + vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 + vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 + sumi = dpct::dp4a(vi0, u[2 * i + 0], + sumi); // SIMD dot product of quantized values + + int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits + vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 + vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 + vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 + vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 + sumi = dpct::dp4a(vi1, u[2 * i + 1], + sumi); // SIMD dot product of quantized values + } + + const sycl::float2 ds8f = + ds8.convert(); + + // second part effectively subtracts 16 from each quant value + return d5 * (sumi * ds8f.x() - (16 * vdr / QI5_0) * ds8f.y()); +} + +#define VDR_Q5_1_Q8_1_MMVQ 2 +#define VDR_Q5_1_Q8_1_MMQ 4 + +template +static __dpct_inline__ float +vec_dot_q5_1_q8_1_impl(const int *vl, const int *vh, const int *u, + const sycl::half2 &dm5, const sycl::half2 &ds8) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits + vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 + vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 + vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 + vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 + sumi = dpct::dp4a(vi0, u[2 * i + 0], + sumi); // SIMD dot product of quantized values + + int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits + vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 + vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 + vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 + vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 + sumi = dpct::dp4a(vi1, u[2 * i + 1], + sumi); // SIMD dot product of quantized values + } + +#ifdef GGML_SYCL_F16 + const sycl::float2 tmp = + (dm5 * ds8).convert(); + const float d5d8 = tmp.x(); + const float m5s8 = tmp.y(); + + +#else + const sycl::float2 dm5f = + dm5.convert(); + const sycl::float2 ds8f = + ds8.convert(); + const float d5d8 = dm5f.x() * ds8f.x(); + const float m5s8 = dm5f.y() * ds8f.y(); +#endif // GGML_SYCL_F16 + + // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it + return sumi*d5d8 + m5s8 / (QI5_1 / vdr); +} + +#define VDR_Q8_0_Q8_1_MMVQ 2 +#define VDR_Q8_0_Q8_1_MMQ 8 + +template +static __dpct_inline__ float vec_dot_q8_0_q8_1_impl(const int *v, const int *u, + const float &d8_0, + const float &d8_1) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + // SIMD dot product of quantized values + sumi = dpct::dp4a(v[i], u[i], sumi); + } + + return d8_0*d8_1 * sumi; +} + +template +static __dpct_inline__ float vec_dot_q8_1_q8_1_impl(const int *v, const int *u, + const sycl::half2 &dm8, + const sycl::half2 &ds8) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + // SIMD dot product of quantized values + sumi = dpct::dp4a(v[i], u[i], sumi); + } + +#ifdef GGML_SYCL_F16 + const sycl::float2 tmp = + (dm8 * ds8).convert(); + const float d8d8 = tmp.x(); + const float m8s8 = tmp.y(); +#else + const sycl::float2 dm8f = + dm8.convert(); + const sycl::float2 ds8f = + ds8.convert(); + const float d8d8 = dm8f.x() * ds8f.x(); + const float m8s8 = dm8f.y() * ds8f.y(); +#endif // GGML_SYCL_F16 + + // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it + return sumi*d8d8 + m8s8 / (QI8_1 / vdr); +} + +static __dpct_inline__ float +vec_dot_q4_0_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq; + + int v[VDR_Q4_0_Q8_1_MMVQ]; + int u[2*VDR_Q4_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); + } + + return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); +} + +static __dpct_inline__ float +vec_dot_q4_1_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq; + + int v[VDR_Q4_1_Q8_1_MMVQ]; + int u[2*VDR_Q4_1_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1); + } + + return vec_dot_q4_1_q8_1_impl(v, u, bq4_1->dm, bq8_1->ds); +} + +static __dpct_inline__ float +vec_dot_q5_0_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq; + + int vl[VDR_Q5_0_Q8_1_MMVQ]; + int vh[VDR_Q5_0_Q8_1_MMVQ]; + int u[2*VDR_Q5_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) { + vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i); + vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0); + } + + return vec_dot_q5_0_q8_1_impl(vl, vh, u, bq5_0->d, bq8_1->ds); +} + +static __dpct_inline__ float +vec_dot_q5_1_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq; + + int vl[VDR_Q5_1_Q8_1_MMVQ]; + int vh[VDR_Q5_1_Q8_1_MMVQ]; + int u[2*VDR_Q5_1_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) { + vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i); + vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1); + } + + return vec_dot_q5_1_q8_1_impl(vl, vh, u, bq5_1->dm, bq8_1->ds); +} + +static __dpct_inline__ float +vec_dot_q8_0_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq; + + int v[VDR_Q8_0_Q8_1_MMVQ]; + int u[VDR_Q8_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_int8(bq8_0->qs, iqs + i); + u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + } + + return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, + bq8_1->ds[0]); +} + +static __dpct_inline__ float +vec_dot_q2_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q2_K * bq2_K = (const block_q2_K *) vbq; + + const int bq8_offset = QR2_K * (iqs / QI8_1); + const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); + + const uint8_t * scales = bq2_K->scales + scale_offset; + + const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs); + int u[QR2_K]; + float d8[QR2_K]; + +#pragma unroll + for (int i = 0; i < QR2_K; ++ i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + d8[i] = bq8_1[bq8_offset + i].ds[0]; + } + + return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8); +} + +static __dpct_inline__ float +vec_dot_q3_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q3_K * bq3_K = (const block_q3_K *) vbq; + + const int bq8_offset = QR3_K * (iqs / (QI3_K/2)); + const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); + + const float d = bq3_K->d; + + const int vl = get_int_from_uint8(bq3_K->qs, iqs); + + // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted + const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; + + int u[QR3_K]; + float d8[QR3_K]; + +#pragma unroll + for (int i = 0; i < QR3_K; ++i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + d8[i] = bq8_1[bq8_offset + i].ds[0]; + } + + return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8); +} + +static __dpct_inline__ float +vec_dot_q4_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + +#ifndef GGML_QKK_64 + const block_q4_K * bq4_K = (const block_q4_K *) vbq; + + int v[2]; + int u[2*QR4_K]; + float d8[QR4_K]; + + // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6 + const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2)); + + // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12 + // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44 + // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76 + // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108 + + const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); + v[0] = q4[0]; + v[1] = q4[4]; + + const uint16_t * scales = (const uint16_t *)bq4_K->scales; + uint16_t aux[2]; + const int j = bq8_offset/2; + if (j < 2) { + aux[0] = scales[j+0] & 0x3f3f; + aux[1] = scales[j+2] & 0x3f3f; + } else { + aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); + aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); + } + const uint8_t * sc = (const uint8_t *)aux; + const uint8_t * m = sc + 2; + + for (int i = 0; i < QR4_K; ++i) { + const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; + d8[i] = bq8i->ds[0]; + + const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); + u[2*i+0] = q8[0]; + u[2*i+1] = q8[4]; + } + + return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8); + +#else + +#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics + const block_q4_K * bq4_K = (const block_q4_K *) vbq; + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + + uint16_t aux16[2]; + const uint8_t * s = (const uint8_t *)aux16; + + const uint16_t * a = (const uint16_t *)bq4_K->scales; + aux16[0] = a[0] & 0x0f0f; + aux16[1] = (a[0] >> 4) & 0x0f0f; + + const float dall = bq4_K->dm[0]; + const float dmin = bq4_K->dm[1]; + + const float d8_1 = bq8_1[0].ds[0]; + const float d8_2 = bq8_1[1].ds[1]; + + const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); + const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); + const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); + const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); + + const int * q4 = (const int *)bq4_K->qs + (iqs/2); + const int v1 = q4[0]; + const int v2 = q4[4]; + + const int dot1 = dpct::dp4a(ui2, v2 & 0x0f0f0f0f, dpct::dp4a(ui1, v1 & 0x0f0f0f0f, 0)); + const int dot2 = dpct::dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, dpct::dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); + const int dot3 = dpct::dp4a(0x01010101, ui2, dpct::dp4a(0x01010101, ui1, 0)); + const int dot4 = dpct::dp4a(0x01010101, ui4, dpct::dp4a(0x01010101, ui3, 0)); + + sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]); + sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]); + + return dall * sumf_d - dmin * sumf_m; + +#else + bad_arch(); +#endif // __SYCL_ARCH__ >= VER_4VEC + +#endif +} + +static __dpct_inline__ float +vec_dot_q5_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + +#ifndef GGML_QKK_64 + const block_q5_K * bq5_K = (const block_q5_K *) vbq; + + int vl[2]; + int vh[2]; + int u[2*QR5_K]; + float d8[QR5_K]; + + const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2)); + const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); + const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4)); + + vl[0] = ql[0]; + vl[1] = ql[4]; + + vh[0] = qh[0] >> bq8_offset; + vh[1] = qh[4] >> bq8_offset; + + const uint16_t * scales = (const uint16_t *)bq5_K->scales; + uint16_t aux[2]; + const int j = bq8_offset/2; + if (j < 2) { + aux[0] = scales[j+0] & 0x3f3f; + aux[1] = scales[j+2] & 0x3f3f; + } else { + aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); + aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); + } + const uint8_t * sc = (const uint8_t *)aux; + const uint8_t * m = sc + 2; + +#pragma unroll + for (int i = 0; i < QR5_K; ++i) { + const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; + d8[i] = bq8i->ds[0]; + + const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); + u[2*i+0] = q8[0]; + u[2*i+1] = q8[4]; + } + + return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8); + +#else + +#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics + const block_q5_K * bq5_K = (const block_q5_K *) vbq; + + const int8_t * s = bq5_K->scales; + + const float d = bq5_K->d; + + const float d8_1 = bq8_1[0].ds[0]; + const float d8_2 = bq8_1[1].ds[1]; + + const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); + const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); + const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); + const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); + + const int * ql = (const int *)bq5_K->qs + (iqs/2); + const int vl1 = ql[0]; + const int vl2 = ql[4]; + + const int step = 4 * (iqs/2); // 0, 4, 8, 12 + const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6 + const int in = step%8; // 0, 4, 0, 4 + const int vh = (*((const int *)(bq5_K->qh + in))) >> im; + + const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f); + const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f); + const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f); + const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f); + + const float sumf_d = d8_1 * (dpct::dp4a(ui1, v1, 0) * s[0] + dpct::dp4a(ui2, v2, 0) * s[1]) + + d8_2 * (dpct::dp4a(ui3, v3, 0) * s[2] + dpct::dp4a(ui4, v4, 0) * s[3]); + + return d * sumf_d; + +#else + bad_arch(); +#endif // __SYCL_ARCH__ >= VER_4VEC + +#endif +} + +static __dpct_inline__ float +vec_dot_q6_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q6_K * bq6_K = (const block_q6_K *) vbq; + + const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4); + const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8); + const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4)); + + const int vl = get_int_from_uint8(bq6_K->ql, iqs); + const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; + + const int8_t * scales = bq6_K->scales + scale_offset; + + int u[QR6_K]; + float d8[QR6_K]; + +#pragma unroll + for (int i = 0; i < QR6_K; ++i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); + d8[i] = bq8_1[bq8_offset + 2 * i].ds[0]; + } + + return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); +} + + +static __dpct_inline__ float +vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint64_t *iq2xxs_grid, const uint8_t *ksigns_iq2xs, + const uint8_t *kmask_iq2xs) { +#if QK_K == 256 + const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq; + +#if QR2_XXS == 8 + const int ib32 = iqs; + const uint16_t * q2 = bq2->qs + 4*ib32; + const uint8_t * aux8 = (const uint8_t *)q2; + const int8_t * q8 = bq8_1[ib32].qs; + uint32_t aux32 = q2[2] | (q2[3] << 16); + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); + const uint8_t signs = ksigns_iq2xs[aux32 & 127]; + for (int j = 0; j < 8; ++j) { + sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1); + } + q8 += 8; + aux32 >>= 7; + } + const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.25f; + return d * sumi; +#else + // iqs is 0...15 + const int ib32 = iqs/2; + const int il = iqs%2; + const uint16_t * q2 = bq2->qs + 4*ib32; + const uint8_t * aux8 = (const uint8_t *)q2; + const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); + const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); + const uint32_t aux32 = q2[2] | (q2[3] << 16); + const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * bq8_1[ib32].ds[0] * 0.25f; + const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; + const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; + const int8_t * q8 = bq8_1[ib32].qs + 16*il; + int sumi1 = 0, sumi2 = 0; + for (int j = 0; j < 8; ++j) { + sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1); + sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1); + } + return d * (sumi1 + sumi2); +#endif +#else + assert(false); + return 0.f; +#endif +} + +static __dpct_inline__ float +vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint64_t *iq2xs_grid, const uint64_t *ksigns64) { +#if DPCT_COMPATIBILITY_TEMP >= \ + MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq; + + const int ib32 = iqs; + const uint16_t * q2 = bq2->qs + 4*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + const uint8_t ls1 = bq2->scales[ib32] & 0xf; + const uint8_t ls2 = bq2->scales[ib32] >> 4; + int sumi1 = 0; + for (int l = 0; l < 2; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); + const int grid_l = dpct::vectorized_binary( + grid[0] ^ signs[0], signs[0], std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid[1] ^ signs[1], signs[1], std::minus<>()); + sumi1 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi1); + sumi1 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi1); + q8 += 8; + } + int sumi2 = 0; + for (int l = 2; l < 4; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511)); + const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9)); + const int grid_l = dpct::vectorized_binary( + grid[0] ^ signs[0], signs[0], std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid[1] ^ signs[1], signs[1], std::minus<>()); + sumi2 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi2); + sumi2 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi2); + q8 += 8; + } + const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f; + return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); +#else + assert(false); + return 0.f; +#endif +#else + assert(false); + return 0.f; +#endif +} + +static __dpct_inline__ float +vec_dot_iq2_s_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { +#if QK_K == 256 + const block_iq2_s * bq2 = (const block_iq2_s *) vbq; + + const int ib32 = iqs; + const int8_t * q8 = bq8_1[ib32].qs; + const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32; + const uint8_t ls1 = bq2->scales[ib32] & 0xf; + const uint8_t ls2 = bq2->scales[ib32] >> 4; + int sumi1 = 0; + for (int l = 0; l < 2; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); + const uint32_t signs0 = dpct::vectorized_binary( + ((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201, + std::equal_to<>()); + const uint32_t signs1 = dpct::vectorized_binary( + ((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201, + std::equal_to<>()); + const int grid_l = dpct::vectorized_binary( + grid[0] ^ signs0, signs0, std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid[1] ^ signs1, signs1, std::minus<>()); + sumi1 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi1); + sumi1 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi1); + q8 += 8; + } + int sumi2 = 0; + for (int l = 2; l < 4; ++l) { + const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300))); + const uint32_t signs0 = dpct::vectorized_binary( + ((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201, + std::equal_to<>()); + const uint32_t signs1 = dpct::vectorized_binary( + ((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201, + std::equal_to<>()); + const int grid_l = dpct::vectorized_binary( + grid[0] ^ signs0, signs0, std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid[1] ^ signs1, signs1, std::minus<>()); + sumi2 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi2); + sumi2 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi2); + q8 += 8; + } + const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f; + return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); +#else + assert(false); +#endif +} + +static __dpct_inline__ float +vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint32_t *iq3xxs_grid, const uint64_t *ksigns64) { +#if DPCT_COMPATIBILITY_TEMP >= \ + MIN_CC_DP4A // lowest compute capability for integer intrinsics +#if QK_K == 256 + const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq; + + const int ib32 = iqs; + const uint8_t * q3 = bq2->qs + 8*ib32; + const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + uint32_t aux32 = gas[0] | (gas[1] << 16); + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0]; + const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1]; + const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127)); + const int grid_l = dpct::vectorized_binary( + grid1[0] ^ signs[0], signs[0], std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid2[0] ^ signs[1], signs[1], std::minus<>()); + sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi); + sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi); + q8 += 8; + aux32 >>= 7; + } + const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.5f; + return d * sumi; +#else + assert(false); + return 0.f; +#endif +#else + assert(false); + return 0.f; +#endif +} + +static __dpct_inline__ float +vec_dot_iq3_s_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint32_t *iq3s_grid) { +#if QK_K == 256 + const block_iq3_s * bq2 = (const block_iq3_s *) vbq; + + const int ib32 = iqs; + const uint8_t * qs = bq2->qs + 8*ib32; + const int8_t * q8 = bq8_1[ib32].qs; + int sumi = 0; + for (int l = 0; l < 4; ++l) { + const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256)); + const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256)); + uint32_t signs0 = dpct::vectorized_binary( + ((bq2->signs[4 * ib32 + l] & 0xf) * 0x01010101) & 0x08040201, + 0x08040201, std::equal_to<>()); + uint32_t signs1 = dpct::vectorized_binary( + ((bq2->signs[4 * ib32 + l] >> 4) * 0x01010101) & 0x08040201, + 0x08040201, std::equal_to<>()); + const int grid_l = dpct::vectorized_binary( + grid1[0] ^ signs0, signs0, std::minus<>()); + const int grid_h = dpct::vectorized_binary( + grid2[0] ^ signs1, signs1, std::minus<>()); + sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi); + sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi); + q8 += 8; + } + const float d = + (float)bq2->d * + (1 + 2 * ((bq2->scales[ib32 / 2] >> 4 * (ib32 % 2)) & 0xf)) * + bq8_1[ib32].ds[0]; + return d * sumi; +#else + assert(false); +#endif +} + +static __dpct_inline__ float +vec_dot_iq1_s_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs, + const uint32_t *iq1s_grid_gpu) { +#if QK_K == 256 + const block_iq1_s * bq1 = (const block_iq1_s *) vbq; + + const int ib32 = iqs; + int sumi = 0; + const int * q8 = (const int *)bq8_1[ib32].qs; + for (int l = 0; l < 4; ++l) { + const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8))); + int grid0 = grid[0] & 0x0f0f0f0f; + int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; + sumi = dpct::dp4a(q8[2 * l + 1], grid1, + dpct::dp4a(q8[2 * l + 0], grid0, sumi)); + } + + const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA; + const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1); + const float d = d1q * bq8_1[ib32].ds[0]; + const float m = d1q * bq8_1[ib32].ds[1]; + return d * sumi + m * delta; +#else + assert(false); +#endif +} + +static __dpct_inline__ float +vec_dot_iq1_m_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { +#if QK_K == 256 + const block_iq1_m * bq1 = (const block_iq1_m *) vbq; + + const int ib32 = iqs; + int sumi[2] = {0, 0}; + float sumf[2] = {0.f, 0.f}; + + const int * q8 = (const int *)bq8_1[ib32].qs; + for (int l = 0; l < 4; ++l) { + const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8))); + int grid0 = grid[0] & 0x0f0f0f0f; + int grid1 = (grid[0] >> 4) & 0x0f0f0f0f; + sumi[l / 2] = dpct::dp4a(q8[2 * l + 1], grid1, + dpct::dp4a(q8[2 * l + 0], grid0, sumi[l / 2])); + const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA; + const int sumy = dpct::dp4a(q8[2 * l + 1], 0x01010101, + dpct::dp4a(q8[2 * l + 0], 0x01010101, 0)); + sumf[l/2] += delta*sumy; + } + + iq1m_scale_t scale; + const uint16_t * sc = (const uint16_t *)bq1->scales; + scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); + const float d = (float)scale.f16 * bq8_1[ib32].ds[0]; + return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1)); +#else + assert(false); +#endif +} + + +static __dpct_inline__ float +vec_dot_iq4_nl_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_iq4_nl * bq = (const block_iq4_nl *) vbq; + + const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs; + const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs; + + const uint8_t * values = (const uint8_t *)kvalues_iq4nl; + + int v1, v2; + int sumi1 = 0, sumi2 = 0; + for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) { + const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16); + get_int_from_table_16(aux, values, v1, v2); + sumi1 = dpct::dp4a(v1, q8[l + 0], sumi1); + sumi2 = dpct::dp4a(v2, q8[l + 4], sumi2); + } + + const float d = (float)bq->d * bq8_1->ds[0]; + return d * (sumi1 + sumi2); +} + + +static __dpct_inline__ float +vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + +#if QK_K == 256 + const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq; + const uint8_t * values = (const uint8_t *)kvalues_iq4nl; + + // iqs is 0...7 + const int ib32 = iqs; + const int32_t * q8 = (const int *)bq8_1[ib32].qs; + const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32; + const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4); + const float d = (float)bq4->d * (ls - 32) * bq8_1[ib32].ds[0]; + int v1, v2; + int sumi1 = 0, sumi2 = 0; + for (int j = 0; j < 4; ++j) { + get_int_from_table_16(q4[j], values, v1, v2); + sumi1 = dpct::dp4a(v1, q8[j + 0], sumi1); + sumi2 = dpct::dp4a(v2, q8[j + 4], sumi2); + } + return d * (sumi1 + sumi2); +#else + assert(false); +#endif +} + +#endif // GGML_SYCL_VECDOTQ_HPP From a04a953cab583f0229e7b4b506346e3e9a85c8d0 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 19 Jun 2024 13:04:36 +0300 Subject: [PATCH 3/4] codecov : remove (#8004) --- .github/labeler.yml | 1 - .github/workflows/code-coverage.yml | 40 ----------------------------- codecov.yml | 14 ---------- 3 files changed, 55 deletions(-) delete mode 100644 .github/workflows/code-coverage.yml delete mode 100644 codecov.yml diff --git a/.github/labeler.yml b/.github/labeler.yml index 97d739b5811e8..5c12bab735e9c 100644 --- a/.github/labeler.yml +++ b/.github/labeler.yml @@ -42,7 +42,6 @@ build: - cmake/** - CMakeLists.txt - CMakePresets.json - - codecov.yml examples: - changed-files: - any-glob-to-any-file: examples/** diff --git a/.github/workflows/code-coverage.yml b/.github/workflows/code-coverage.yml deleted file mode 100644 index f12c558f81bae..0000000000000 --- a/.github/workflows/code-coverage.yml +++ /dev/null @@ -1,40 +0,0 @@ -name: Code Coverage -on: [push, pull_request] - -env: - GGML_NLOOP: 3 - GGML_N_THREADS: 1 - -concurrency: - group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} - cancel-in-progress: true - -jobs: - run: - runs-on: ubuntu-20.04 - steps: - - name: Checkout - uses: actions/checkout@v4 - - - name: Dependencies - run: | - sudo apt-get update - sudo apt-get install build-essential gcc-8 lcov - - - name: Build - run: CC=gcc-8 make -j LLAMA_CODE_COVERAGE=1 tests - - - name: Run tests - run: CC=gcc-8 make test - - - name: Generate coverage report - run: | - make coverage - make lcov-report - - - name: Upload coverage to Codecov - uses: codecov/codecov-action@v3 - env: - CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }} - with: - files: lcov-report/coverage.info diff --git a/codecov.yml b/codecov.yml deleted file mode 100644 index a301c5b2c7694..0000000000000 --- a/codecov.yml +++ /dev/null @@ -1,14 +0,0 @@ -comment: off - -coverage: - status: - project: - default: - target: auto - threshold: 0 - base: auto - patch: - default: - target: auto - threshold: 0 - base: auto From 9c77ec1d74874ee22bdef8f110e8e8d41389abf2 Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 19 Jun 2024 15:04:15 +0200 Subject: [PATCH 4/4] ggml : synchronize threads using barriers (#7993) --- .github/workflows/server.yml | 14 +++ ggml.c | 217 +++++++++++------------------------ 2 files changed, 81 insertions(+), 150 deletions(-) diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml index 1fee9ac281943..6155e94156e42 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -87,8 +87,22 @@ jobs: exit 1 fi + - name: Build (no OpenMP) + id: cmake_build_no_openmp + if: ${{ matrix.sanitizer == 'THREAD' }} + run: | + cmake -B build \ + -DLLAMA_NATIVE=OFF \ + -DLLAMA_BUILD_SERVER=ON \ + -DLLAMA_CURL=ON \ + -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \ + -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \ + -DLLAMA_OPENMP=OFF ; + cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server + - name: Build id: cmake_build + if: ${{ matrix.sanitizer != 'THREAD' }} run: | cmake -B build \ -DLLAMA_NATIVE=OFF \ diff --git a/ggml.c b/ggml.c index d5d33c2ba1029..778ca3fdf1f8f 100644 --- a/ggml.c +++ b/ggml.c @@ -1753,9 +1753,8 @@ struct ggml_compute_state_shared { int n_threads; // synchronization primitives - atomic_int n_active; // num active threads - atomic_int node_n; // active graph node - atomic_int node_task; // active graph node task phase + atomic_int n_barrier; + atomic_int n_barrier_passed; ggml_abort_callback abort_callback; // abort ggml_graph_compute when true void* abort_callback_data; @@ -18972,47 +18971,49 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_ return n_tasks; } -static void ggml_graph_compute_thread_sync_node(int * node_n, struct ggml_compute_state * state, const bool do_yield) { - // wait for other threads to finish - const int last_node_n = * node_n; - - while (true) { - if (do_yield) { - sched_yield(); - } - - *node_n = atomic_load(&state->shared->node_n); - if (*node_n != last_node_n) { - break; - } - -#if defined(__SSE3__) - // Tell the processor we're spinning. It's a processor hint for spinlocks. - _mm_pause(); -#endif +#ifdef GGML_USE_OPENMP +static void ggml_barrier(struct ggml_compute_state * state) { + if (state->shared->n_threads == 1) { + return; } + + #pragma omp barrier } +#else +static void ggml_barrier(struct ggml_compute_state * state) { + if (state->shared->n_threads == 1) { + return; + } -static void ggml_graph_compute_thread_sync_task(int * task_phase, struct ggml_compute_state * state, const bool do_yield) { - // wait for other threads to finish - const int last_task_phase = *task_phase; + atomic_int * n_barrier = &state->shared->n_barrier; + atomic_int * n_barrier_passed = &state->shared->n_barrier_passed; - while (true) { - if (do_yield) { - sched_yield(); - } + int n_threads = state->shared->n_threads; + int passed_old = atomic_load(n_barrier_passed); - *task_phase = atomic_load(&state->shared->node_task); - if (*task_phase != last_task_phase) { - break; + if (atomic_fetch_add(n_barrier, 1) == n_threads - 1) { + // last thread + atomic_store(n_barrier, 0); + atomic_fetch_add(n_barrier_passed, 1); + } else { + // wait for other threads + //while (atomic_load(n_barrier_passed) == passed_old) { + //} + const int n_spin_before_sleep = 100000; + while (true) { + for (int i = 0; i < n_spin_before_sleep; i++) { + if (atomic_load(n_barrier_passed) != passed_old) { + return; + } + #if defined(__SSE3__) + _mm_pause(); + #endif + } + sched_yield(); } - -#if defined(__SSE3__) - // Tell the processor we're spinning. It's a processor hint for spinlocks. - _mm_pause(); -#endif } } +#endif static thread_ret_t ggml_graph_compute_thread(void * data) { struct ggml_compute_state * state = (struct ggml_compute_state *) data; @@ -19020,136 +19021,54 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { const struct ggml_cgraph * cgraph = state->shared->cgraph; const struct ggml_cplan * cplan = state->shared->cplan; - const int n_threads = state->shared->n_threads; + const int ith = state->ith; + const int n_threads = state->shared->n_threads; - set_numa_thread_affinity(state->ith); + set_numa_thread_affinity(ith); - int node_n = -1; - int task_phase = GGML_TASK_TYPE_FINALIZE; + struct ggml_compute_params params = { + /*.type =*/ GGML_TASK_TYPE_INIT, + /*.ith =*/ ith, + /*.nth =*/ state->shared->n_threads, + /*.wsize =*/ cplan->work_size, + /*.wdata =*/ cplan->work_data, + }; - while (true) { + for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) { if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) { - state->shared->node_n += 1; state->ec = GGML_STATUS_ABORTED; return 0; } - if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { - // all other threads are finished and spinning - // do finalize and init here so we don't have synchronize again - struct ggml_compute_params params = { - /*.type =*/ GGML_TASK_TYPE_FINALIZE, - /*.ith =*/ 0, - /*.nth =*/ 0, - /*.wsize =*/ cplan->work_size, - /*.wdata =*/ cplan->work_data, - }; - - if (node_n != -1) { - /* FINALIZE */ - struct ggml_tensor * node = cgraph->nodes[node_n]; - if (GGML_OP_HAS_FINALIZE[node->op]) { - params.nth = ggml_get_n_tasks(node, n_threads, state->shared->n_threads); - ggml_compute_forward(¶ms, node, state); - } - ggml_graph_compute_perf_stats_node(node, state->shared); - } - - // distribute new work or execute it direct if 1T - while (++node_n < cgraph->n_nodes) { - GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes); - struct ggml_tensor * node = cgraph->nodes[node_n]; - const int n_tasks = ggml_get_n_tasks(node, n_threads, state->shared->n_threads); - - state->shared->perf_node_start_cycles = ggml_perf_cycles(); - state->shared->perf_node_start_time_us = ggml_perf_time_us(); - - params.nth = n_tasks; - - if (n_tasks == 1) { - /* INIT */ - if (GGML_OP_HAS_INIT[node->op]) { - params.type = GGML_TASK_TYPE_INIT; - ggml_compute_forward(¶ms, node, state); - } - - // TODO: maybe push node_n to the atomic but if other threads see n_tasks is 1, - // they do something more efficient than spinning (?) - params.type = GGML_TASK_TYPE_COMPUTE; - ggml_compute_forward(¶ms, node, state); - - if (GGML_OP_HAS_FINALIZE[node->op]) { - params.type = GGML_TASK_TYPE_FINALIZE; - ggml_compute_forward(¶ms, node, state); - } - - ggml_graph_compute_perf_stats_node(node, state->shared); - } else { - break; - } - - if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) { - break; - } - } - - task_phase = GGML_TASK_TYPE_INIT; - atomic_store(&state->shared->n_active, n_threads); - atomic_store(&state->shared->node_n, node_n); - atomic_store(&state->shared->node_task, task_phase); - } else { - ggml_graph_compute_thread_sync_node(&node_n, state, false); - ggml_graph_compute_thread_sync_task(&task_phase, state, false); - } - - // check if we should stop - if (node_n >= cgraph->n_nodes) break; - - /* INIT & COMPUTE */ struct ggml_tensor * node = cgraph->nodes[node_n]; const int n_tasks = ggml_get_n_tasks(node, n_threads, state->shared->n_threads); - struct ggml_compute_params params = { - /*.type =*/ GGML_TASK_TYPE_INIT, - /*.ith =*/ state->ith, - /*.nth =*/ n_tasks, - /*.wsize =*/ cplan->work_size, - /*.wdata =*/ cplan->work_data, - }; + params.nth = n_tasks; - if (state->ith < n_tasks) { - if (GGML_OP_HAS_INIT[node->op]) { + /* INIT */ + if (GGML_OP_HAS_INIT[node->op]) { + if (ith < n_tasks) { + params.type = GGML_TASK_TYPE_INIT; ggml_compute_forward(¶ms, node, state); } + ggml_barrier(state); } - if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { - task_phase = GGML_TASK_TYPE_COMPUTE; - atomic_store(&state->shared->n_active, n_threads); - atomic_store(&state->shared->node_task, task_phase); - } - else { - // TODO: this sched_yield can have significant impact on the performance - either positive or negative - // depending on the workload and the operating system. - // since it is not clear what is the best approach, it should potentially become user-configurable - // ref: https://github.com/ggerganov/ggml/issues/291 - // UPD: adding the do_yield flag seems to resolve the issue universally - const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT; - ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield); - } - - if (state->ith < n_tasks) { + /* COMPUTE */ + if (ith < n_tasks) { params.type = GGML_TASK_TYPE_COMPUTE; ggml_compute_forward(¶ms, node, state); } - if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { - task_phase = GGML_TASK_TYPE_FINALIZE; - atomic_store(&state->shared->n_active, n_threads); - atomic_store(&state->shared->node_task, task_phase); - } - else { - ggml_graph_compute_thread_sync_task(&task_phase, state, false); + ggml_barrier(state); + + /* FINALIZE */ + if (GGML_OP_HAS_FINALIZE[node->op]) { + if (params.ith == 0) { + params.type = GGML_TASK_TYPE_FINALIZE; + ggml_compute_forward(¶ms, node, state); + } + ggml_barrier(state); } } @@ -19336,7 +19255,6 @@ static enum ggml_status ggml_graph_compute_parallel(struct ggml_compute_state * // update the number of threads from the actual number of threads that we got from OpenMP n_threads = omp_get_num_threads(); workers[0].shared->n_threads = n_threads; - workers[0].shared->n_active = n_threads; } ggml_graph_compute_thread(&workers[omp_get_thread_num()]); } @@ -19399,9 +19317,8 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl /*.perf_node_start_cycles =*/ 0, /*.perf_node_start_time_us =*/ 0, /*.n_threads =*/ n_threads, - /*.n_active =*/ n_threads, - /*.node_n =*/ -1, - /*.node_task =*/ GGML_TASK_TYPE_FINALIZE, + /*.n_barrier =*/ 0, + /*.n_barrier_passed =*/ 0, /*.abort_callback =*/ NULL, /*.abort_callback_data =*/ NULL, /*.current_chunk; =*/ 0,