From 904837e0cb2f5f01bf5d5901b7aa57a026860ae4 Mon Sep 17 00:00:00 2001 From: Dou Xinpeng <15529241576@163.com> Date: Wed, 25 Sep 2024 11:30:38 +0800 Subject: [PATCH 1/5] cann: fix crash when llama-bench is running on multiple cann devices (#9627) --- ggml/src/ggml-cann/common.h | 1 + 1 file changed, 1 insertion(+) diff --git a/ggml/src/ggml-cann/common.h b/ggml/src/ggml-cann/common.h index e6a570107..edfa49614 100644 --- a/ggml/src/ggml-cann/common.h +++ b/ggml/src/ggml-cann/common.h @@ -227,6 +227,7 @@ struct ggml_backend_cann_context { * @brief Destructor for cleaning up resources. */ ~ggml_backend_cann_context() { + ggml_cann_set_device(device); if (copy_event != nullptr) { ACL_CHECK(aclrtDestroyEvent(copy_event)); } From 3d6bf6919f7b10726421779cd344f2da05421c68 Mon Sep 17 00:00:00 2001 From: Gabe Goodhart Date: Wed, 25 Sep 2024 01:06:52 -0600 Subject: [PATCH 2/5] llama : add IBM Granite MoE architecture (#9438) * feat(gguf-py): Add granitemoe architecture This includes the addition of new tensor names for the new moe layers. These may not be correct at this point due to the need for the hack in gguf_writer.py to double-check the length of the shape for these layers. Branch: GraniteMoE Signed-off-by: Gabe Goodhart * feat(convert_hf_to_gguf): Add GraniteMoeModel GraniteMoe has the same configuration deltas as Granite Branch: GraniteMoE Signed-off-by: Gabe Goodhart * fix(granitemoe convert): Split the double-sized input layer into gate and up After a lot of staring and squinting, it's clear that the standard mixtral expert implementation is equivalent to the vectorized parallel experts in granite. The difference is that in granite, the w1 and w3 are concatenated into a single tensor "input_linear." Rather than reimplementing all of the math on the llama.cpp side, the much simpler route is to just split this tensor during conversion and follow the standard mixtral route. Branch: GraniteMoE Co-Authored-By: alex.brooks@ibm.com Signed-off-by: Gabe Goodhart * feat(granitemoe): Implement granitemoe GraniteMoE follows the mixtral architecture (once the input_linear layers are split into gate_exps/up_exps). The main delta is the addition of the same four multipliers used in Granite. Branch: GraniteMoE Signed-off-by: Gabe Goodhart * Typo fix in docstring Co-Authored-By: ggerganov@gmail.com Co-authored-by: Georgi Gerganov Signed-off-by: Gabe Goodhart * fix(conversion): Simplify tensor name mapping in conversion Branch: GraniteMoE Co-Authored-By: git@compilade.net Signed-off-by: Gabe Goodhart * fix(convert): Remove unused tensor name mappings Branch: GraniteMoE Co-Authored-By: git@compilade.net Signed-off-by: Gabe Goodhart * fix(convert): Sanity check on merged FFN tensor sizes Branch: GraniteMoE Co-Authored-By: git@compilade.net Signed-off-by: Gabe Goodhart * fix: Allow "output" layer in granite moe architecture (convert and cpp) Branch: GraniteMoE Co-Authored-By: git@compilade.net Signed-off-by: Gabe Goodhart * fix(granite): Add missing 'output' tensor for Granite This is a fix for the previous `granite` architecture PR. Recent snapshots have included this (`lm_head.weights`) as part of the architecture Branch: GraniteMoE Signed-off-by: Gabe Goodhart --------- Signed-off-by: Gabe Goodhart Co-authored-by: Georgi Gerganov --- convert_hf_to_gguf.py | 33 +++++++++++++++++++++++++++++++-- gguf-py/gguf/constants.py | 18 ++++++++++++++++++ gguf-py/gguf/tensor_mapping.py | 20 +++++++++++--------- src/llama.cpp | 30 ++++++++++++++++++++++++++++-- 4 files changed, 88 insertions(+), 13 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index ff4c9226f..7be609054 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -4102,16 +4102,45 @@ class GraniteModel(LlamaModel): # consistency if attention_scale := self.hparams.get("attention_multiplier"): self.gguf_writer.add_attention_scale(attention_scale) + logger.info("gguf: (granite) attention_scale = %s", attention_scale) if embedding_scale := self.hparams.get("embedding_multiplier"): self.gguf_writer.add_embedding_scale(embedding_scale) + logger.info("gguf: (granite) embedding_scale = %s", embedding_scale) if residual_scale := self.hparams.get("residual_multiplier"): self.gguf_writer.add_residual_scale(residual_scale) - if logits_scaling := self.hparams.get("logits_scaling"): - self.gguf_writer.add_logit_scale(logits_scaling) + logger.info("gguf: (granite) residual_scale = %s", residual_scale) + if logits_scale := self.hparams.get("logits_scaling"): + self.gguf_writer.add_logit_scale(logits_scale) + logger.info("gguf: (granite) logits_scale = %s", logits_scale) + + +@Model.register("GraniteMoeForCausalLM") +class GraniteMoeModel(GraniteModel): + """Conversion for IBM's GraniteMoeForCausalLM""" + model_arch = gguf.MODEL_ARCH.GRANITE_MOE + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + """In modeling_granitemoe, the JetMoe implementation of parallel experts + is used. This essentially merges w1 and w3 into a single tensor with 2x + the hidden size that is then split during forward. To keep compatibility + with existing mixtral support, we pull them apart here. + """ + + if name.endswith("block_sparse_moe.input_linear.weight"): + ffn_dim = self.hparams["intermediate_size"] + assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size" + gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :] + return [ + (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate), + (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up), + ] + + return super().modify_tensors(data_torch, name, bid) ###### CONVERSION LOGIC ###### + # tree of lazy tensors class LazyTorchTensor(gguf.LazyBase): _tensor_type = torch.Tensor diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index b36a60d49..560eee916 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -235,6 +235,7 @@ class MODEL_ARCH(IntEnum): NEMOTRON = auto() EXAONE = auto() GRANITE = auto() + GRANITE_MOE = auto() class MODEL_TENSOR(IntEnum): @@ -392,6 +393,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.NEMOTRON: "nemotron", MODEL_ARCH.EXAONE: "exaone", MODEL_ARCH.GRANITE: "granite", + MODEL_ARCH.GRANITE_MOE: "granitemoe", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -1232,6 +1234,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_ARCH.GRANITE: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, MODEL_TENSOR.ATTN_NORM, MODEL_TENSOR.ATTN_Q, MODEL_TENSOR.ATTN_K, @@ -1242,6 +1245,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.GRANITE_MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], # TODO } diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 2ebfa2b43..4e850726e 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -251,11 +251,12 @@ class TensorNameMap: ), MODEL_TENSOR.FFN_GATE_INP: ( - "layers.{bid}.feed_forward.gate", # mixtral - "model.layers.{bid}.block_sparse_moe.gate", # mixtral - "model.layers.{bid}.mlp.gate", # qwen2moe olmoe - "transformer.decoder_layer.{bid}.router", # Grok - "transformer.blocks.{bid}.ffn.router.layer", # dbrx + "layers.{bid}.feed_forward.gate", # mixtral + "model.layers.{bid}.block_sparse_moe.gate", # mixtral + "model.layers.{bid}.mlp.gate", # qwen2moe olmoe + "transformer.decoder_layer.{bid}.router", # Grok + "transformer.blocks.{bid}.ffn.router.layer", # dbrx + "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe ), MODEL_TENSOR.FFN_GATE_INP_SHEXP: ( @@ -364,10 +365,11 @@ class TensorNameMap: ), MODEL_TENSOR.FFN_DOWN_EXP: ( - "layers.{bid}.feed_forward.experts.w2", # mixtral (merged) - "transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged) - "transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx - "model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged) + "layers.{bid}.feed_forward.experts.w2", # mixtral (merged) + "transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged) + "transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx + "model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged) + "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe ), MODEL_TENSOR.FFN_DOWN_SHEXP: ( diff --git a/src/llama.cpp b/src/llama.cpp index a718de054..0accb1492 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -215,6 +215,7 @@ enum llm_arch { LLM_ARCH_EXAONE, LLM_ARCH_RWKV6, LLM_ARCH_GRANITE, + LLM_ARCH_GRANITE_MOE, LLM_ARCH_UNKNOWN, }; @@ -266,6 +267,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_EXAONE, "exaone" }, { LLM_ARCH_RWKV6, "rwkv6" }, { LLM_ARCH_GRANITE, "granite" }, + { LLM_ARCH_GRANITE_MOE, "granitemoe" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -1467,6 +1469,7 @@ static const std::map> LLM_TENSOR_NA { { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, @@ -1478,6 +1481,24 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_GRANITE_MOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -2396,7 +2417,7 @@ struct llama_hparams { float f_max_alibi_bias = 0.0f; float f_logit_scale = 0.0f; - // Additional scale factors (Granite) + // Additional scale factors (Granite/Granite MoE) float f_residual_scale = 0.0f; float f_embedding_scale = 0.0f; float f_attention_scale = 0.0f; @@ -6048,6 +6069,7 @@ static void llm_load_hparams( } } break; case LLM_ARCH_GRANITE: + case LLM_ARCH_GRANITE_MOE: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale); @@ -6056,6 +6078,7 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale); switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_3B; break; case 40: model.type = e_model::MODEL_3B; break; // Add additional layer/vocab/etc checks here for other model sizes default: model.type = e_model::MODEL_UNKNOWN; @@ -6810,7 +6833,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp); } - if (model.arch == LLM_ARCH_GRANITE) { + if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) { LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale); LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale); LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale); @@ -6984,6 +7007,7 @@ static bool llm_load_tensors( case LLM_ARCH_REFACT: case LLM_ARCH_MINICPM: case LLM_ARCH_GRANITE: + case LLM_ARCH_GRANITE_MOE: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -15930,6 +15954,7 @@ static struct ggml_cgraph * llama_build_graph( switch (model.arch) { case LLM_ARCH_LLAMA: case LLM_ARCH_GRANITE: + case LLM_ARCH_GRANITE_MOE: { result = llm.build_llama(); } break; @@ -19231,6 +19256,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_DEEPSEEK2: case LLM_ARCH_CHATGLM: case LLM_ARCH_GRANITE: + case LLM_ARCH_GRANITE_MOE: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2 From afbbfaa537a96f562c34df4542930fa951b40d9e Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Wed, 25 Sep 2024 14:05:13 +0200 Subject: [PATCH 3/5] server : add more env vars, improve gen-docs (#9635) * server : add more env vars, improve gen-docs * update server docs * LLAMA_ARG_NO_CONTEXT_SHIFT --- common/arg.cpp | 56 ++++++++-------- examples/gen-docs/gen-docs.cpp | 85 +++++++++++++++-------- examples/server/README.md | 119 +++++++++++++++++++-------------- examples/server/server.cpp | 4 ++ 4 files changed, 157 insertions(+), 107 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index c1ec3c4f9..6880117ed 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -691,7 +691,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, [](gpt_params & params) { params.ctx_shift = false; } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT")); add_opt(llama_arg( {"--chunks"}, "N", format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks), @@ -1102,7 +1102,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; } else { throw std::invalid_argument("invalid value"); } } - ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); + ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_POOLING")); add_opt(llama_arg( {"--attention"}, "{causal,non,causal}", "attention type for embeddings, use model default if unspecified", @@ -1121,77 +1121,77 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; } else { throw std::invalid_argument("invalid value"); } } - )); + ).set_env("LLAMA_ARG_ROPE_SCALING_TYPE")); add_opt(llama_arg( {"--rope-scale"}, "N", "RoPE context scaling factor, expands context by a factor of N", [](gpt_params & params, const std::string & value) { params.rope_freq_scale = 1.0f / std::stof(value); } - )); + ).set_env("LLAMA_ARG_ROPE_SCALE")); add_opt(llama_arg( {"--rope-freq-base"}, "N", "RoPE base frequency, used by NTK-aware scaling (default: loaded from model)", [](gpt_params & params, const std::string & value) { params.rope_freq_base = std::stof(value); } - )); + ).set_env("LLAMA_ARG_ROPE_FREQ_BASE")); add_opt(llama_arg( {"--rope-freq-scale"}, "N", "RoPE frequency scaling factor, expands context by a factor of 1/N", [](gpt_params & params, const std::string & value) { params.rope_freq_scale = std::stof(value); } - )); + ).set_env("LLAMA_ARG_ROPE_FREQ_SCALE")); add_opt(llama_arg( {"--yarn-orig-ctx"}, "N", format("YaRN: original context size of model (default: %d = model training context size)", params.yarn_orig_ctx), [](gpt_params & params, int value) { params.yarn_orig_ctx = value; } - )); + ).set_env("LLAMA_ARG_YARN_ORIG_CTX")); add_opt(llama_arg( {"--yarn-ext-factor"}, "N", format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor), [](gpt_params & params, const std::string & value) { params.yarn_ext_factor = std::stof(value); } - )); + ).set_env("LLAMA_ARG_YARN_EXT_FACTOR")); add_opt(llama_arg( {"--yarn-attn-factor"}, "N", format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor), [](gpt_params & params, const std::string & value) { params.yarn_attn_factor = std::stof(value); } - )); + ).set_env("LLAMA_ARG_YARN_ATTN_FACTOR")); add_opt(llama_arg( {"--yarn-beta-slow"}, "N", format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow), [](gpt_params & params, const std::string & value) { params.yarn_beta_slow = std::stof(value); } - )); + ).set_env("LLAMA_ARG_YARN_BETA_SLOW")); add_opt(llama_arg( {"--yarn-beta-fast"}, "N", format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast), [](gpt_params & params, const std::string & value) { params.yarn_beta_fast = std::stof(value); } - )); + ).set_env("LLAMA_ARG_YARN_BETA_FAST")); add_opt(llama_arg( {"-gan", "--grp-attn-n"}, "N", format("group-attention factor (default: %d)", params.grp_attn_n), [](gpt_params & params, int value) { params.grp_attn_n = value; } - )); + ).set_env("LLAMA_ARG_GRP_ATTN_N")); add_opt(llama_arg( {"-gaw", "--grp-attn-w"}, "N", format("group-attention width (default: %.1f)", (double)params.grp_attn_w), [](gpt_params & params, int value) { params.grp_attn_w = value; } - )); + ).set_env("LLAMA_ARG_GRP_ATTN_W")); add_opt(llama_arg( {"-dkvc", "--dump-kv-cache"}, "verbose print of the KV cache", @@ -1205,7 +1205,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, [](gpt_params & params) { params.no_kv_offload = true; } - )); + ).set_env("LLAMA_ARG_NO_KV_OFFLOAD")); add_opt(llama_arg( {"-ctk", "--cache-type-k"}, "TYPE", format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()), @@ -1213,7 +1213,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, // TODO: get the type right here params.cache_type_k = value; } - )); + ).set_env("LLAMA_ARG_CACHE_TYPE_K")); add_opt(llama_arg( {"-ctv", "--cache-type-v"}, "TYPE", format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()), @@ -1221,7 +1221,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, // TODO: get the type right here params.cache_type_v = value; } - )); + ).set_env("LLAMA_ARG_CACHE_TYPE_V")); add_opt(llama_arg( {"--perplexity", "--all-logits"}, format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"), @@ -1355,7 +1355,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, [](gpt_params & params, const std::string & value) { params.rpc_servers = value; } - )); + ).set_env("LLAMA_ARG_RPC")); #endif add_opt(llama_arg( {"--mlock"}, @@ -1363,14 +1363,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, [](gpt_params & params) { params.use_mlock = true; } - )); + ).set_env("LLAMA_ARG_MLOCK")); add_opt(llama_arg( {"--no-mmap"}, "do not memory-map model (slower load but may reduce pageouts if not using mlock)", [](gpt_params & params) { params.use_mmap = false; } - )); + ).set_env("LLAMA_ARG_NO_MMAP")); add_opt(llama_arg( {"--numa"}, "TYPE", "attempt optimizations that help on some NUMA systems\n" @@ -1385,7 +1385,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } else { throw std::invalid_argument("invalid value"); } } - )); + ).set_env("LLAMA_ARG_NUMA")); add_opt(llama_arg( {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N", "number of layers to store in VRAM", @@ -1433,7 +1433,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting the split mode has no effect.\n"); } } - )); + ).set_env("LLAMA_ARG_SPLIT_MODE")); add_opt(llama_arg( {"-ts", "--tensor-split"}, "N0,N1,N2,...", "fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1", @@ -1460,7 +1460,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting a tensor split has no effect.\n"); } } - )); + ).set_env("LLAMA_ARG_TENSOR_SPLIT")); add_opt(llama_arg( {"-mg", "--main-gpu"}, "INDEX", format("the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu), @@ -1470,7 +1470,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting the main GPU has no effect.\n"); } } - )); + ).set_env("LLAMA_ARG_MAIN_GPU")); add_opt(llama_arg( {"--check-tensors"}, format("check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false"), @@ -1533,7 +1533,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, [](gpt_params & params, const std::string & value) { params.model_alias = value; } - ).set_examples({LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ALIAS")); add_opt(llama_arg( {"-m", "--model"}, "FNAME", ex == LLAMA_EXAMPLE_EXPORT_LORA @@ -1741,7 +1741,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, [](gpt_params & params, const std::string & value) { params.public_path = value; } - ).set_examples({LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH")); add_opt(llama_arg( {"--embedding", "--embeddings"}, format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"), @@ -1779,14 +1779,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, [](gpt_params & params, const std::string & value) { params.ssl_file_key = value; } - ).set_examples({LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_KEY_FILE")); add_opt(llama_arg( {"--ssl-cert-file"}, "FNAME", "path to file a PEM-encoded SSL certificate", [](gpt_params & params, const std::string & value) { params.ssl_file_cert = value; } - ).set_examples({LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE")); add_opt(llama_arg( {"-to", "--timeout"}, "N", format("server read/write timeout in seconds (default: %d)", params.timeout_read), @@ -1794,7 +1794,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, params.timeout_read = value; params.timeout_write = value; } - ).set_examples({LLAMA_EXAMPLE_SERVER})); + ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TIMEOUT")); add_opt(llama_arg( {"--threads-http"}, "N", format("number of threads used to process HTTP requests (default: %d)", params.n_threads_http), diff --git a/examples/gen-docs/gen-docs.cpp b/examples/gen-docs/gen-docs.cpp index b6d4725fd..4b19a9dc2 100644 --- a/examples/gen-docs/gen-docs.cpp +++ b/examples/gen-docs/gen-docs.cpp @@ -6,42 +6,73 @@ // Export usage message (-h) to markdown format +static void write_table_header(std::ofstream & file) { + file << "| Argument | Explanation |\n"; + file << "| -------- | ----------- |\n"; +} + +static void write_table_entry(std::ofstream & file, const llama_arg & opt) { + file << "| `"; + // args + for (const auto & arg : opt.args) { + if (arg == opt.args.front()) { + file << arg; + if (opt.args.size() > 1) file << ", "; + } else { + file << arg << (arg != opt.args.back() ? ", " : ""); + } + } + // value hint + if (opt.value_hint) { + std::string md_value_hint(opt.value_hint); + string_replace_all(md_value_hint, "|", "\\|"); + file << " " << md_value_hint; + } + if (opt.value_hint_2) { + std::string md_value_hint_2(opt.value_hint_2); + string_replace_all(md_value_hint_2, "|", "\\|"); + file << " " << md_value_hint_2; + } + // help text + std::string md_help(opt.help); + string_replace_all(md_help, "\n", "
"); + string_replace_all(md_help, "|", "\\|"); + file << "` | " << md_help << " |\n"; +} + +static void write_table(std::ofstream & file, std::vector & opts) { + write_table_header(file); + for (const auto & opt : opts) { + write_table_entry(file, *opt); + } +} + static void export_md(std::string fname, llama_example ex) { std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc); gpt_params params; auto ctx_arg = gpt_params_parser_init(params, ex); - file << "| Argument | Explanation |\n"; - file << "| -------- | ----------- |\n"; + std::vector common_options; + std::vector sparam_options; + std::vector specific_options; for (auto & opt : ctx_arg.options) { - file << "| `"; - // args - for (const auto & arg : opt.args) { - if (arg == opt.args.front()) { - file << arg; - if (opt.args.size() > 1) file << ", "; - } else { - file << arg << (arg != opt.args.back() ? ", " : ""); - } + // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example + if (opt.is_sparam) { + sparam_options.push_back(&opt); + } else if (opt.in_example(ctx_arg.ex)) { + specific_options.push_back(&opt); + } else { + common_options.push_back(&opt); } - // value hint - if (opt.value_hint) { - std::string md_value_hint(opt.value_hint); - string_replace_all(md_value_hint, "|", "\\|"); - file << " " << md_value_hint; - } - if (opt.value_hint_2) { - std::string md_value_hint_2(opt.value_hint_2); - string_replace_all(md_value_hint_2, "|", "\\|"); - file << " " << md_value_hint_2; - } - // help text - std::string md_help(opt.help); - string_replace_all(md_help, "\n", "
"); - string_replace_all(md_help, "|", "\\|"); - file << "` | " << md_help << " |\n"; } + + file << "**Common params**\n\n"; + write_table(file, common_options); + file << "\n\n**Sampling params**\n\n"; + write_table(file, sparam_options); + file << "\n\n**Example-specific params**\n\n"; + write_table(file, specific_options); } int main(int, char **) { diff --git a/examples/server/README.md b/examples/server/README.md index 741950c8a..dfca07f98 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -17,6 +17,8 @@ The project is under active development, and we are [looking for feedback and co ## Usage +**Common params** + | Argument | Explanation | | -------- | ----------- | | `-h, --help, --usage` | print usage and exit | @@ -38,7 +40,6 @@ The project is under active development, and we are [looking for feedback and co | `-b, --batch-size N` | logical maximum batch size (default: 2048)
(env: LLAMA_ARG_BATCH) | | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)
(env: LLAMA_ARG_UBATCH) | | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) | -| `--no-context-shift` | disables context shift on inifinite text generation (default: disabled) | | `-fa, --flash-attn` | enable Flash Attention (default: disabled)
(env: LLAMA_ARG_FLASH_ATTN) | | `-p, --prompt PROMPT` | prompt to start generation with | | `--no-perf` | disable internal libllama performance timings (default: false)
(env: LLAMA_ARG_NO_PERF) | @@ -46,8 +47,56 @@ The project is under active development, and we are [looking for feedback and co | `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) | | `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) | | `--no-escape` | do not process escape sequences | -| `-sp, --special` | special tokens output enabled (default: false) | -| `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) | +| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model
(env: LLAMA_ARG_ROPE_SCALING_TYPE) | +| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N
(env: LLAMA_ARG_ROPE_SCALE) | +| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
(env: LLAMA_ARG_ROPE_FREQ_BASE) | +| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N
(env: LLAMA_ARG_ROPE_FREQ_SCALE) | +| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)
(env: LLAMA_ARG_YARN_ORIG_CTX) | +| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
(env: LLAMA_ARG_YARN_EXT_FACTOR) | +| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0)
(env: LLAMA_ARG_YARN_ATTN_FACTOR) | +| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0)
(env: LLAMA_ARG_YARN_BETA_SLOW) | +| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)
(env: LLAMA_ARG_YARN_BETA_FAST) | +| `-gan, --grp-attn-n N` | group-attention factor (default: 1)
(env: LLAMA_ARG_GRP_ATTN_N) | +| `-gaw, --grp-attn-w N` | group-attention width (default: 512.0)
(env: LLAMA_ARG_GRP_ATTN_W) | +| `-dkvc, --dump-kv-cache` | verbose print of the KV cache | +| `-nkvo, --no-kv-offload` | disable KV offload
(env: LLAMA_ARG_NO_KV_OFFLOAD) | +| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)
(env: LLAMA_ARG_CACHE_TYPE_K) | +| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16)
(env: LLAMA_ARG_CACHE_TYPE_V) | +| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)
(env: LLAMA_ARG_DEFRAG_THOLD) | +| `-np, --parallel N` | number of parallel sequences to decode (default: 1)
(env: LLAMA_ARG_N_PARALLEL) | +| `--mlock` | force system to keep model in RAM rather than swapping or compressing
(env: LLAMA_ARG_MLOCK) | +| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)
(env: LLAMA_ARG_NO_MMAP) | +| `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggerganov/llama.cpp/issues/1437
(env: LLAMA_ARG_NUMA) | +| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM
(env: LLAMA_ARG_N_GPU_LAYERS) | +| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:
- none: use one GPU only
- layer (default): split layers and KV across GPUs
- row: split rows across GPUs
(env: LLAMA_ARG_SPLIT_MODE) | +| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1
(env: LLAMA_ARG_TENSOR_SPLIT) | +| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0)
(env: LLAMA_ARG_MAIN_GPU) | +| `--check-tensors` | check model tensor data for invalid values (default: false) | +| `--override-kv KEY=TYPE:VALUE` | advanced option to override model metadata by key. may be specified multiple times.
types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false | +| `--lora FNAME` | path to LoRA adapter (can be repeated to use multiple adapters) | +| `--lora-scaled FNAME SCALE` | path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters) | +| `--control-vector FNAME` | add a control vector
note: this argument can be repeated to add multiple control vectors | +| `--control-vector-scaled FNAME SCALE` | add a control vector with user defined scaling SCALE
note: this argument can be repeated to add multiple scaled control vectors | +| `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive | +| `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)
(env: LLAMA_ARG_MODEL) | +| `-mu, --model-url MODEL_URL` | model download url (default: unused)
(env: LLAMA_ARG_MODEL_URL) | +| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)
(env: LLAMA_ARG_HF_REPO) | +| `-hff, --hf-file FILE` | Hugging Face model file (default: unused)
(env: LLAMA_ARG_HF_FILE) | +| `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)
(env: HF_TOKEN) | +| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) | +| `--log-disable` | Log disable | +| `--log-file FNAME` | Log to file | +| `--log-colors` | Enable colored logging
(env: LLAMA_LOG_COLORS) | +| `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) | +| `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored.
(env: LLAMA_LOG_VERBOSITY) | +| `--log-prefix` | Enable prefx in log messages
(env: LLAMA_LOG_PREFIX) | +| `--log-timestamps` | Enable timestamps in log messages
(env: LLAMA_LOG_TIMESTAMPS) | + + +**Sampling params** + +| Argument | Explanation | +| -------- | ----------- | | `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) | | `-s, --seed SEED` | RNG seed (default: 4294967295, use random seed for 4294967295) | | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) | @@ -72,54 +121,28 @@ The project is under active development, and we are [looking for feedback and co | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') | | `--grammar-file FNAME` | file to read grammar from | | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead | -| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model | -| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N | -| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model) | -| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N | -| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size) | -| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation) | -| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0) | -| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0) | -| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0) | -| `-gan, --grp-attn-n N` | group-attention factor (default: 1) | -| `-gaw, --grp-attn-w N` | group-attention width (default: 512.0) | -| `-dkvc, --dump-kv-cache` | verbose print of the KV cache | -| `-nkvo, --no-kv-offload` | disable KV offload | -| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16) | -| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) | -| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)
(env: LLAMA_ARG_DEFRAG_THOLD) | -| `-np, --parallel N` | number of parallel sequences to decode (default: 1)
(env: LLAMA_ARG_N_PARALLEL) | + + +**Example-specific params** + +| Argument | Explanation | +| -------- | ----------- | +| `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)
(env: LLAMA_ARG_NO_CONTEXT_SHIFT) | +| `-sp, --special` | special tokens output enabled (default: false) | +| `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) | +| `--pooling {none,mean,cls,last}` | pooling type for embeddings, use model default if unspecified
(env: LLAMA_ARG_POOLING) | | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)
(env: LLAMA_ARG_CONT_BATCHING) | | `-nocb, --no-cont-batching` | disable continuous batching
(env: LLAMA_ARG_NO_CONT_BATCHING) | -| `--mlock` | force system to keep model in RAM rather than swapping or compressing | -| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock) | -| `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggerganov/llama.cpp/issues/1437 | -| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM
(env: LLAMA_ARG_N_GPU_LAYERS) | -| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:
- none: use one GPU only
- layer (default): split layers and KV across GPUs
- row: split rows across GPUs | -| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1 | -| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0) | -| `--check-tensors` | check model tensor data for invalid values (default: false) | -| `--override-kv KEY=TYPE:VALUE` | advanced option to override model metadata by key. may be specified multiple times.
types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false | -| `--lora FNAME` | path to LoRA adapter (can be repeated to use multiple adapters) | -| `--lora-scaled FNAME SCALE` | path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters) | -| `--control-vector FNAME` | add a control vector
note: this argument can be repeated to add multiple control vectors | -| `--control-vector-scaled FNAME SCALE` | add a control vector with user defined scaling SCALE
note: this argument can be repeated to add multiple scaled control vectors | -| `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive | -| `-a, --alias STRING` | set alias for model name (to be used by REST API) | -| `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)
(env: LLAMA_ARG_MODEL) | -| `-mu, --model-url MODEL_URL` | model download url (default: unused)
(env: LLAMA_ARG_MODEL_URL) | -| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)
(env: LLAMA_ARG_HF_REPO) | -| `-hff, --hf-file FILE` | Hugging Face model file (default: unused)
(env: LLAMA_ARG_HF_FILE) | -| `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)
(env: HF_TOKEN) | +| `-a, --alias STRING` | set alias for model name (to be used by REST API)
(env: LLAMA_ARG_ALIAS) | | `--host HOST` | ip address to listen (default: 127.0.0.1)
(env: LLAMA_ARG_HOST) | | `--port PORT` | port to listen (default: 8080)
(env: LLAMA_ARG_PORT) | -| `--path PATH` | path to serve static files from (default: ) | +| `--path PATH` | path to serve static files from (default: )
(env: LLAMA_ARG_STATIC_PATH) | | `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
(env: LLAMA_ARG_EMBEDDINGS) | | `--api-key KEY` | API key to use for authentication (default: none)
(env: LLAMA_API_KEY) | | `--api-key-file FNAME` | path to file containing API keys (default: none) | -| `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key | -| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate | -| `-to, --timeout N` | server read/write timeout in seconds (default: 600) | +| `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key
(env: LLAMA_ARG_SSL_KEY_FILE) | +| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate
(env: LLAMA_ARG_SSL_CERT_FILE) | +| `-to, --timeout N` | server read/write timeout in seconds (default: 600)
(env: LLAMA_ARG_TIMEOUT) | | `--threads-http N` | number of threads used to process HTTP requests (default: -1)
(env: LLAMA_ARG_THREADS_HTTP) | | `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications | | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)
(env: LLAMA_ARG_ENDPOINT_METRICS) | @@ -128,14 +151,6 @@ The project is under active development, and we are [looking for feedback and co | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted:
https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
(env: LLAMA_ARG_CHAT_TEMPLATE) | | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
| | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) | -| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) | -| `--log-disable` | Log disable | -| `--log-file FNAME` | Log to file | -| `--log-colors` | Enable colored logging
(env: LLAMA_LOG_COLORS) | -| `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) | -| `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored.
(env: LLAMA_LOG_VERBOSITY) | -| `--log-prefix` | Enable prefx in log messages
(env: LLAMA_LOG_PREFIX) | -| `--log-timestamps` | Enable timestamps in log messages
(env: LLAMA_LOG_TIMESTAMPS) | Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var. diff --git a/examples/server/server.cpp b/examples/server/server.cpp index e5275a514..61ff09bb2 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2356,6 +2356,10 @@ int main(int argc, char ** argv) { svr.reset(new httplib::Server()); } #else + if (params.ssl_file_key != "" && params.ssl_file_cert != "") { + LOG_ERR("Server is built without SSL support\n"); + return 1; + } svr.reset(new httplib::Server()); #endif From 1e436302188a704ac9ea044af03193648806f19c Mon Sep 17 00:00:00 2001 From: Charles Xu <63788048+chaxu01@users.noreply.github.com> Date: Wed, 25 Sep 2024 15:12:20 +0200 Subject: [PATCH 4/5] ggml : remove assert for AArch64 GEMV and GEMM Q4 kernels (#9217) * ggml : remove assert for AArch64 GEMV and GEMM Q4 kernels * added fallback mechanism when the offline re-quantized model is not optimized for the underlying target. * fix for build errors * remove prints from the low-level code * Rebase to the latest upstream --- ggml/src/ggml-aarch64.c | 3226 +++++++++++++++++++-------------------- 1 file changed, 1591 insertions(+), 1635 deletions(-) diff --git a/ggml/src/ggml-aarch64.c b/ggml/src/ggml-aarch64.c index 2b01b4f93..8912de63d 100644 --- a/ggml/src/ggml-aarch64.c +++ b/ggml/src/ggml-aarch64.c @@ -1,4 +1,7 @@ -// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd. +// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates +// SPDX-License-Identifier: MIT +// + #define GGML_COMMON_IMPL_C #include "ggml-common.h" @@ -595,6 +598,15 @@ size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_ return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8); } +// Return the number of byte lanes in the SVE vector if SVE is supported; otherwise, returns 0 if SVE is not supported. +static int sve_lane_count(void) { +#if defined(__ARM_FEATURE_SVE) + return ggml_sve_cnt_b; +#else + return 0; +#endif +} + void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -614,73 +626,67 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(ncols_interleaved); UNUSED(blocklen); -#if defined(__ARM_FEATURE_SVE) - if (ggml_sve_cnt_b == QK8_0) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) && - "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) + if (ggml_cpu_has_neon()) { + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; - __asm__ __volatile__( - "movi v31.16b, #0x4\n" - "movi v30.16b, #0xf0\n" - "add %x[b_ptr], %x[b_ptr], #0x8\n" - "1:" // Column loop - "add x22, %x[a_ptr], #0x2\n" - "movi v29.16b, #0x0\n" - "mov x21, %x[nb]\n" - "2:" // Block loop - "ldr q28, [%x[b_ptr], #0x0]\n" - "ldr q27, [x22, #0x0]\n" - "movi v26.4s, #0x0\n" - "sub x20, x22, #0x2\n" - "ldr q25, [x22, #0x10]\n" - "ldr q24, [%x[b_ptr], #0x10]\n" - "sub x21, x21, #0x1\n" - "add x22, x22, #0x22\n" - "ldr q23, [%x[b_ptr], #0x20]\n" - "ldr q22, [%x[b_ptr], #0x30]\n" - "ld1r { v21.8h }, [x20]\n" - "ldr q20, [%x[b_ptr], #-0x8]\n" - "sshl v16.16b, v28.16b, v31.16b\n" - "and v28.16b, v28.16b, v30.16b\n" - "sshl v19.16b, v24.16b, v31.16b\n" - "and v24.16b, v24.16b, v30.16b\n" - "add %x[b_ptr], %x[b_ptr], #0x48\n" - "sshl v18.16b, v23.16b, v31.16b\n" - "and v23.16b, v23.16b, v30.16b\n" - ".inst 0x4f9be21a // sdot v26.4s, v16.16b, v27.4b[0]\n" - "sshl v17.16b, v22.16b, v31.16b\n" - "and v22.16b, v22.16b, v30.16b\n" - "fcvtl v21.4s, v21.4h\n" - "fcvtl v16.4s, v20.4h\n" - ".inst 0x4f99e39a // sdot v26.4s, v28.16b, v25.4b[0]\n" - "fmul v16.4s, v16.4s, v21.4s\n" - ".inst 0x4fbbe27a // sdot v26.4s, v19.16b, v27.4b[1]\n" - ".inst 0x4fb9e31a // sdot v26.4s, v24.16b, v25.4b[1]\n" - ".inst 0x4f9bea5a // sdot v26.4s, v18.16b, v27.4b[2]\n" - ".inst 0x4f99eafa // sdot v26.4s, v23.16b, v25.4b[2]\n" - ".inst 0x4fbbea3a // sdot v26.4s, v17.16b, v27.4b[3]\n" - ".inst 0x4fb9eada // sdot v26.4s, v22.16b, v25.4b[3]\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "fmla v29.4s, v26.4s, v16.4s\n" - "cbnz x21, 2b\n" - "sub %x[nc], %x[nc], #0x4\n" - "str q29, [%x[res_ptr], #0x0]\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "cbnz %x[nc], 1b\n" - : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) - : [a_ptr] "r" (a_ptr), [nb] "r" (nb) - : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22" - ); -#else + __asm__ __volatile__( + "movi v31.16b, #0x4\n" + "movi v30.16b, #0xf0\n" + "add %x[b_ptr], %x[b_ptr], #0x8\n" + "1:" // Column loop + "add x22, %x[a_ptr], #0x2\n" + "movi v29.16b, #0x0\n" + "mov x21, %x[nb]\n" + "2:" // Block loop + "ldr q28, [%x[b_ptr], #0x0]\n" + "ldr q27, [x22, #0x0]\n" + "movi v26.4s, #0x0\n" + "sub x20, x22, #0x2\n" + "ldr q25, [x22, #0x10]\n" + "ldr q24, [%x[b_ptr], #0x10]\n" + "sub x21, x21, #0x1\n" + "add x22, x22, #0x22\n" + "ldr q23, [%x[b_ptr], #0x20]\n" + "ldr q22, [%x[b_ptr], #0x30]\n" + "ld1r { v21.8h }, [x20]\n" + "ldr q20, [%x[b_ptr], #-0x8]\n" + "sshl v16.16b, v28.16b, v31.16b\n" + "and v28.16b, v28.16b, v30.16b\n" + "sshl v19.16b, v24.16b, v31.16b\n" + "and v24.16b, v24.16b, v30.16b\n" + "add %x[b_ptr], %x[b_ptr], #0x48\n" + "sshl v18.16b, v23.16b, v31.16b\n" + "and v23.16b, v23.16b, v30.16b\n" + ".inst 0x4f9be21a // sdot v26.4s, v16.16b, v27.4b[0]\n" + "sshl v17.16b, v22.16b, v31.16b\n" + "and v22.16b, v22.16b, v30.16b\n" + "fcvtl v21.4s, v21.4h\n" + "fcvtl v16.4s, v20.4h\n" + ".inst 0x4f99e39a // sdot v26.4s, v28.16b, v25.4b[0]\n" + "fmul v16.4s, v16.4s, v21.4s\n" + ".inst 0x4fbbe27a // sdot v26.4s, v19.16b, v27.4b[1]\n" + ".inst 0x4fb9e31a // sdot v26.4s, v24.16b, v25.4b[1]\n" + ".inst 0x4f9bea5a // sdot v26.4s, v18.16b, v27.4b[2]\n" + ".inst 0x4f99eafa // sdot v26.4s, v23.16b, v25.4b[2]\n" + ".inst 0x4fbbea3a // sdot v26.4s, v17.16b, v27.4b[3]\n" + ".inst 0x4fb9eada // sdot v26.4s, v22.16b, v25.4b[3]\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "fmla v29.4s, v26.4s, v16.4s\n" + "cbnz x21, 2b\n" + "sub %x[nc], %x[nc], #0x4\n" + "str q29, [%x[res_ptr], #0x0]\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "cbnz %x[nc], 1b\n" + : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) + : [a_ptr] "r" (a_ptr), [nb] "r" (nb) + : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22" + ); + return; + } +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) float sumf[4]; int sumi; @@ -704,7 +710,6 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * } for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } -#endif } void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { @@ -726,79 +731,72 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(ncols_interleaved); UNUSED(blocklen); -#if defined(__ARM_FEATURE_SVE) - if (ggml_sve_cnt_b == QK8_0) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; - __asm__ __volatile__( - "movi v2.16b, #0x4\n" - "movi v1.16b, #0xf0\n" - "add %x[b_ptr], %x[b_ptr], #0x8\n" - "1:" // Column loop - "add x23, %x[a_ptr], #0x2\n" - "movi v0.16b, #0x0\n" - "mov x22, %x[nb]\n" - "2:" // Block loop - "ldr q31, [%x[b_ptr], #0x0]\n" - "ldr q30, [%x[b_ptr], #0x10]\n" - "mov x21, x23\n" - "movi v29.4s, #0x0\n" - "ldr q28, [%x[b_ptr], #0x20]\n" - "ldr q27, [%x[b_ptr], #0x30]\n" - "movi v26.4s, #0x0\n" - "sub x20, x23, #0x2\n" - "ld1r { v25.8h }, [x20]\n" - "ldr q24, [%x[b_ptr], #-0x8]\n" - "sub x22, x22, #0x1\n" - "add x23, x23, #0x22\n" - "ld1r { v23.2d }, [x21], #0x8\n" - "sshl v22.16b, v31.16b, v2.16b\n" - "sshl v16.16b, v30.16b, v2.16b\n" - "add %x[b_ptr], %x[b_ptr], #0x48\n" - "ld1r { v21.2d }, [x21], #0x8\n" - "sshl v20.16b, v28.16b, v2.16b\n" - "sshl v19.16b, v27.16b, v2.16b\n" - "ld1r { v18.2d }, [x21], #0x8\n" - "ld1r { v17.2d }, [x21], #0x8\n" - "and v31.16b, v31.16b, v1.16b\n" - "and v30.16b, v30.16b, v1.16b\n" - ".inst 0x4e9796dd // sdot v29.4s, v22.16b, v23.16b\n" - ".inst 0x4e97961a // sdot v26.4s, v16.16b, v23.16b\n" - "and v28.16b, v28.16b, v1.16b\n" - "and v27.16b, v27.16b, v1.16b\n" - "fcvtl v25.4s, v25.4h\n" - "fcvtl v16.4s, v24.4h\n" - ".inst 0x4e95969d // sdot v29.4s, v20.16b, v21.16b\n" - ".inst 0x4e95967a // sdot v26.4s, v19.16b, v21.16b\n" - "fmul v16.4s, v16.4s, v25.4s\n" - ".inst 0x4e9297fd // sdot v29.4s, v31.16b, v18.16b\n" - ".inst 0x4e9297da // sdot v26.4s, v30.16b, v18.16b\n" - ".inst 0x4e91979d // sdot v29.4s, v28.16b, v17.16b\n" - ".inst 0x4e91977a // sdot v26.4s, v27.16b, v17.16b\n" - "addp v29.4s, v29.4s, v26.4s\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "fmla v0.4s, v29.4s, v16.4s\n" - "cbnz x22, 2b\n" - "sub %x[nc], %x[nc], #0x4\n" - "str q0, [%x[res_ptr], #0x0]\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "cbnz %x[nc], 1b\n" - : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) - : [a_ptr] "r" (a_ptr), [nb] "r" (nb) - : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23" - ); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); -#else + __asm__ __volatile__( + "movi v2.16b, #0x4\n" + "movi v1.16b, #0xf0\n" + "add %x[b_ptr], %x[b_ptr], #0x8\n" + "1:" // Column loop + "add x23, %x[a_ptr], #0x2\n" + "movi v0.16b, #0x0\n" + "mov x22, %x[nb]\n" + "2:" // Block loop + "ldr q31, [%x[b_ptr], #0x0]\n" + "ldr q30, [%x[b_ptr], #0x10]\n" + "mov x21, x23\n" + "movi v29.4s, #0x0\n" + "ldr q28, [%x[b_ptr], #0x20]\n" + "ldr q27, [%x[b_ptr], #0x30]\n" + "movi v26.4s, #0x0\n" + "sub x20, x23, #0x2\n" + "ld1r { v25.8h }, [x20]\n" + "ldr q24, [%x[b_ptr], #-0x8]\n" + "sub x22, x22, #0x1\n" + "add x23, x23, #0x22\n" + "ld1r { v23.2d }, [x21], #0x8\n" + "sshl v22.16b, v31.16b, v2.16b\n" + "sshl v16.16b, v30.16b, v2.16b\n" + "add %x[b_ptr], %x[b_ptr], #0x48\n" + "ld1r { v21.2d }, [x21], #0x8\n" + "sshl v20.16b, v28.16b, v2.16b\n" + "sshl v19.16b, v27.16b, v2.16b\n" + "ld1r { v18.2d }, [x21], #0x8\n" + "ld1r { v17.2d }, [x21], #0x8\n" + "and v31.16b, v31.16b, v1.16b\n" + "and v30.16b, v30.16b, v1.16b\n" + ".inst 0x4e9796dd // sdot v29.4s, v22.16b, v23.16b\n" + ".inst 0x4e97961a // sdot v26.4s, v16.16b, v23.16b\n" + "and v28.16b, v28.16b, v1.16b\n" + "and v27.16b, v27.16b, v1.16b\n" + "fcvtl v25.4s, v25.4h\n" + "fcvtl v16.4s, v24.4h\n" + ".inst 0x4e95969d // sdot v29.4s, v20.16b, v21.16b\n" + ".inst 0x4e95967a // sdot v26.4s, v19.16b, v21.16b\n" + "fmul v16.4s, v16.4s, v25.4s\n" + ".inst 0x4e9297fd // sdot v29.4s, v31.16b, v18.16b\n" + ".inst 0x4e9297da // sdot v26.4s, v30.16b, v18.16b\n" + ".inst 0x4e91979d // sdot v29.4s, v28.16b, v17.16b\n" + ".inst 0x4e91977a // sdot v26.4s, v27.16b, v17.16b\n" + "addp v29.4s, v29.4s, v26.4s\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "fmla v0.4s, v29.4s, v16.4s\n" + "cbnz x22, 2b\n" + "sub %x[nc], %x[nc], #0x4\n" + "str q0, [%x[res_ptr], #0x0]\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "cbnz %x[nc], 1b\n" + : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) + : [a_ptr] "r" (a_ptr), [nb] "r" (nb) + : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23" + ); + return; + } +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) float sumf[4]; int sumi; @@ -822,7 +820,6 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * } for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } -#endif } void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { @@ -844,8 +841,9 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(ncols_interleaved); UNUSED(blocklen); -#if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - if (ggml_sve_cnt_b == QK8_0) { +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) +#if defined(__ARM_FEATURE_SVE) + if (ggml_cpu_has_sve() && sve_lane_count() == QK8_0) { const void * b_ptr = vx; const void * a_ptr = vy; float * res_ptr = s; @@ -910,24 +908,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * ); return; } - else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { - GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) && - "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal " - "performance"); - } - else if (ggml_cpu_has_neon()) { - GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 " - "quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(ggml_cpu_has_sve() && - "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); +#endif // #if defined(__ARM_FEATURE_SVE) #elif defined(__AVX2__) // Lookup table to convert signed nibbles to signed bytes __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); @@ -1018,31 +999,33 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * _mm256_storeu_ps(s + (y * nr + x * 8), acc_row); } } -#else - float sumf[8]; - int sumi; + return; +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) + { + float sumf[8]; + int sumi; - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); } - sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d); } } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } -#endif } void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { @@ -1065,505 +1048,500 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(ncols_interleaved); UNUSED(blocklen); -#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) - if (ggml_sve_cnt_b == QK8_0) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) + if (ggml_cpu_has_neon()) { + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + size_t res_stride = bs * sizeof(float); + + __asm__ __volatile__( + "mov x10, %x[nr]\n" + "mov x9, #0x88\n" + "cmp x10, #0x10\n" + "mul x9, %x[nb], x9\n" + "blt 4f\n" + "1:" // Row loop + "add x28, %x[b_ptr], #0x8\n" + "mov x27, %x[nc]\n" + "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" + "2:" // Column loop + "add x25, %x[a_ptr], #0x8\n" + "movi v15.16b, #0x0\n" + "movi v19.16b, #0x0\n" + "mov x24, %x[nb]\n" + "add x23, x25, x9\n" + "movi v18.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "add x22, x23, x9\n" + "movi v11.16b, #0x0\n" + "movi v13.16b, #0x0\n" + "add x21, x22, x9\n" + "movi v23.16b, #0x0\n" + "movi v16.16b, #0x0\n" + "movi v25.16b, #0x0\n" + "movi v7.16b, #0x0\n" + "movi v0.16b, #0x0\n" + "movi v4.16b, #0x0\n" + "movi v5.16b, #0x0\n" + "movi v21.16b, #0x0\n" + "movi v8.16b, #0x0\n" + "movi v1.16b, #0x0\n" + "3:" // Block loop + "ldr q3, [x28, #0x0]\n" + "ldr q31, [x25, #0x0]\n" + "movi v28.16b, #0x4\n" + "movi v10.4s, #0x0\n" + "ldr q22, [x28, #0x10]\n" + "ldr q6, [x25, #0x10]\n" + "movi v29.4s, #0x0\n" + "movi v9.4s, #0x0\n" + "ldr q27, [x28, #0x20]\n" + "ldr q30, [x28, #0x30]\n" + "movi v20.4s, #0x0\n" + "movi v24.16b, #0xf0\n" + "ldr d2, [x25, #-0x8]\n" + "ldr d26, [x23, #-0x8]\n" + "sshl v12.16b, v3.16b, v28.16b\n" + "sub x20, x28, #0x8\n" + "ldr d17, [x20, #0x0]\n" + "and v3.16b, v3.16b, v24.16b\n" + "subs x24, x24, #0x1\n" + "add x28, x28, #0x48\n" + ".inst 0x4f9fe18a // sdot v10.4s, v12.16b, v31.4b[0]\n" + ".inst 0x4fbfe19d // sdot v29.4s, v12.16b, v31.4b[1]\n" + ".inst 0x4f9fe989 // sdot v9.4s, v12.16b, v31.4b[2]\n" + ".inst 0x4fbfe994 // sdot v20.4s, v12.16b, v31.4b[3]\n" + "sshl v31.16b, v22.16b, v28.16b\n" + "and v22.16b, v22.16b, v24.16b\n" + "fcvtl v17.4s, v17.4h\n" + "fcvtl v2.4s, v2.4h\n" + "fcvtl v26.4s, v26.4h\n" + ".inst 0x4f86e3ea // sdot v10.4s, v31.16b, v6.4b[0]\n" + ".inst 0x4fa6e3fd // sdot v29.4s, v31.16b, v6.4b[1]\n" + ".inst 0x4f86ebe9 // sdot v9.4s, v31.16b, v6.4b[2]\n" + ".inst 0x4fa6ebf4 // sdot v20.4s, v31.16b, v6.4b[3]\n" + "sshl v6.16b, v27.16b, v28.16b\n" + "sshl v28.16b, v30.16b, v28.16b\n" + "and v27.16b, v27.16b, v24.16b\n" + "and v30.16b, v30.16b, v24.16b\n" + "ldr q24, [x25, #0x20]\n" + ".inst 0x4f98e0ca // sdot v10.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8c9 // sdot v9.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8d4 // sdot v20.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x30]\n" + ".inst 0x4f98e38a // sdot v10.4s, v28.16b, v24.4b[0]\n" + ".inst 0x4fb8e39d // sdot v29.4s, v28.16b, v24.4b[1]\n" + ".inst 0x4f98eb89 // sdot v9.4s, v28.16b, v24.4b[2]\n" + ".inst 0x4fb8eb94 // sdot v20.4s, v28.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x40]\n" + ".inst 0x4f98e06a // sdot v10.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e869 // sdot v9.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e874 // sdot v20.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x50]\n" + ".inst 0x4f98e2ca // sdot v10.4s, v22.16b, v24.4b[0]\n" + ".inst 0x4fb8e2dd // sdot v29.4s, v22.16b, v24.4b[1]\n" + ".inst 0x4f98eac9 // sdot v9.4s, v22.16b, v24.4b[2]\n" + ".inst 0x4fb8ead4 // sdot v20.4s, v22.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x60]\n" + ".inst 0x4f98e36a // sdot v10.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb69 // sdot v9.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb74 // sdot v20.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x25, #0x70]\n" + "add x25, x25, #0x88\n" + ".inst 0x4f98e3ca // sdot v10.4s, v30.16b, v24.4b[0]\n" + ".inst 0x4fb8e3dd // sdot v29.4s, v30.16b, v24.4b[1]\n" + ".inst 0x4f98ebc9 // sdot v9.4s, v30.16b, v24.4b[2]\n" + ".inst 0x4fb8ebd4 // sdot v20.4s, v30.16b, v24.4b[3]\n" + "fmul v24.4s, v17.4s, v2.s[0]\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v15.4s, v10.4s, v24.4s\n" + "ldr q24, [x23, #0x0]\n" + "fmul v10.4s, v17.4s, v2.s[1]\n" + "fmla v19.4s, v29.4s, v10.4s\n" + "ldr q10, [x23, #0x10]\n" + "fmul v29.4s, v17.4s, v2.s[2]\n" + "fmul v2.4s, v17.4s, v2.s[3]\n" + "fmla v18.4s, v9.4s, v29.4s\n" + "movi v9.4s, #0x0\n" + "movi v29.4s, #0x0\n" + ".inst 0x4f98e189 // sdot v9.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e19d // sdot v29.4s, v12.16b, v24.4b[1]\n" + "fmla v14.4s, v20.4s, v2.4s\n" + "movi v20.4s, #0x0\n" + "movi v2.4s, #0x0\n" + ".inst 0x4f98e994 // sdot v20.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x20]\n" + ".inst 0x4f8ae3e9 // sdot v9.4s, v31.16b, v10.4b[0]\n" + ".inst 0x4faae3fd // sdot v29.4s, v31.16b, v10.4b[1]\n" + ".inst 0x4f8aebf4 // sdot v20.4s, v31.16b, v10.4b[2]\n" + ".inst 0x4faaebe2 // sdot v2.4s, v31.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x30]\n" + ".inst 0x4f98e0c9 // sdot v9.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8d4 // sdot v20.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x40]\n" + ".inst 0x4f8ae389 // sdot v9.4s, v28.16b, v10.4b[0]\n" + ".inst 0x4faae39d // sdot v29.4s, v28.16b, v10.4b[1]\n" + ".inst 0x4f8aeb94 // sdot v20.4s, v28.16b, v10.4b[2]\n" + ".inst 0x4faaeb82 // sdot v2.4s, v28.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x50]\n" + ".inst 0x4f98e069 // sdot v9.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e874 // sdot v20.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x23, #0x60]\n" + ".inst 0x4f8ae2c9 // sdot v9.4s, v22.16b, v10.4b[0]\n" + ".inst 0x4faae2dd // sdot v29.4s, v22.16b, v10.4b[1]\n" + ".inst 0x4f8aead4 // sdot v20.4s, v22.16b, v10.4b[2]\n" + ".inst 0x4faaeac2 // sdot v2.4s, v22.16b, v10.4b[3]\n" + "ldr q10, [x23, #0x70]\n" + "add x23, x23, #0x88\n" + ".inst 0x4f98e369 // sdot v9.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb74 // sdot v20.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x0]\n" + ".inst 0x4f8ae3c9 // sdot v9.4s, v30.16b, v10.4b[0]\n" + ".inst 0x4faae3dd // sdot v29.4s, v30.16b, v10.4b[1]\n" + ".inst 0x4f8aebd4 // sdot v20.4s, v30.16b, v10.4b[2]\n" + ".inst 0x4faaebc2 // sdot v2.4s, v30.16b, v10.4b[3]\n" + "fmul v10.4s, v17.4s, v26.s[0]\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "fmla v11.4s, v9.4s, v10.4s\n" + "ldr q9, [x22, #0x10]\n" + "fmul v10.4s, v17.4s, v26.s[1]\n" + "fmla v13.4s, v29.4s, v10.4s\n" + "ldr d29, [x22, #-0x8]\n" + "fmul v10.4s, v17.4s, v26.s[2]\n" + "fmul v26.4s, v17.4s, v26.s[3]\n" + "fcvtl v29.4s, v29.4h\n" + "fmla v23.4s, v20.4s, v10.4s\n" + "movi v20.4s, #0x0\n" + "movi v10.4s, #0x0\n" + "fmla v16.4s, v2.4s, v26.4s\n" + "movi v26.4s, #0x0\n" + "movi v2.4s, #0x0\n" + ".inst 0x4f98e194 // sdot v20.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" + ".inst 0x4f98e99a // sdot v26.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x20]\n" + ".inst 0x4f89e3f4 // sdot v20.4s, v31.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" + ".inst 0x4f89ebfa // sdot v26.4s, v31.16b, v9.4b[2]\n" + ".inst 0x4fa9ebe2 // sdot v2.4s, v31.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x30]\n" + ".inst 0x4f98e0d4 // sdot v20.4s, v6.16b, v24.4b[0]\n" + ".inst 0x4fb8e0ca // sdot v10.4s, v6.16b, v24.4b[1]\n" + ".inst 0x4f98e8da // sdot v26.4s, v6.16b, v24.4b[2]\n" + ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x40]\n" + ".inst 0x4f89e394 // sdot v20.4s, v28.16b, v9.4b[0]\n" + ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" + ".inst 0x4f89eb9a // sdot v26.4s, v28.16b, v9.4b[2]\n" + ".inst 0x4fa9eb82 // sdot v2.4s, v28.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x50]\n" + ".inst 0x4f98e074 // sdot v20.4s, v3.16b, v24.4b[0]\n" + ".inst 0x4fb8e06a // sdot v10.4s, v3.16b, v24.4b[1]\n" + ".inst 0x4f98e87a // sdot v26.4s, v3.16b, v24.4b[2]\n" + ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" + "ldr q24, [x22, #0x60]\n" + ".inst 0x4f89e2d4 // sdot v20.4s, v22.16b, v9.4b[0]\n" + ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" + ".inst 0x4f89eada // sdot v26.4s, v22.16b, v9.4b[2]\n" + ".inst 0x4fa9eac2 // sdot v2.4s, v22.16b, v9.4b[3]\n" + "ldr q9, [x22, #0x70]\n" + "add x22, x22, #0x88\n" + ".inst 0x4f98e374 // sdot v20.4s, v27.16b, v24.4b[0]\n" + ".inst 0x4fb8e36a // sdot v10.4s, v27.16b, v24.4b[1]\n" + ".inst 0x4f98eb7a // sdot v26.4s, v27.16b, v24.4b[2]\n" + ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" + "ldr q24, [x21, #0x0]\n" + ".inst 0x4f89e3d4 // sdot v20.4s, v30.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ca // sdot v10.4s, v30.16b, v9.4b[1]\n" + ".inst 0x4f89ebda // sdot v26.4s, v30.16b, v9.4b[2]\n" + ".inst 0x4fa9ebc2 // sdot v2.4s, v30.16b, v9.4b[3]\n" + "fmul v9.4s, v17.4s, v29.s[0]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "fmla v25.4s, v20.4s, v9.4s\n" + "ldr q9, [x21, #0x10]\n" + "fmul v20.4s, v17.4s, v29.s[1]\n" + "fmla v7.4s, v10.4s, v20.4s\n" + "ldr d20, [x21, #-0x8]\n" + "fmul v10.4s, v17.4s, v29.s[2]\n" + "fmul v29.4s, v17.4s, v29.s[3]\n" + "fcvtl v20.4s, v20.4h\n" + "fmla v0.4s, v26.4s, v10.4s\n" + "movi v26.4s, #0x0\n" + "movi v10.4s, #0x0\n" + "fmla v4.4s, v2.4s, v29.4s\n" + "movi v2.4s, #0x0\n" + "movi v29.4s, #0x0\n" + ".inst 0x4f98e19a // sdot v26.4s, v12.16b, v24.4b[0]\n" + ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" + ".inst 0x4f98e982 // sdot v2.4s, v12.16b, v24.4b[2]\n" + ".inst 0x4fb8e99d // sdot v29.4s, v12.16b, v24.4b[3]\n" + "ldr q12, [x21, #0x20]\n" + "fmul v24.4s, v17.4s, v20.s[0]\n" + ".inst 0x4f89e3fa // sdot v26.4s, v31.16b, v9.4b[0]\n" + ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" + ".inst 0x4f89ebe2 // sdot v2.4s, v31.16b, v9.4b[2]\n" + ".inst 0x4fa9ebfd // sdot v29.4s, v31.16b, v9.4b[3]\n" + "ldr q9, [x21, #0x30]\n" + "fmul v31.4s, v17.4s, v20.s[1]\n" + ".inst 0x4f8ce0da // sdot v26.4s, v6.16b, v12.4b[0]\n" + ".inst 0x4face0ca // sdot v10.4s, v6.16b, v12.4b[1]\n" + ".inst 0x4f8ce8c2 // sdot v2.4s, v6.16b, v12.4b[2]\n" + ".inst 0x4face8dd // sdot v29.4s, v6.16b, v12.4b[3]\n" + "ldr q12, [x21, #0x40]\n" + "fmul v6.4s, v17.4s, v20.s[2]\n" + "fmul v20.4s, v17.4s, v20.s[3]\n" + ".inst 0x4f89e39a // sdot v26.4s, v28.16b, v9.4b[0]\n" + ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" + ".inst 0x4f89eb82 // sdot v2.4s, v28.16b, v9.4b[2]\n" + ".inst 0x4fa9eb9d // sdot v29.4s, v28.16b, v9.4b[3]\n" + "ldr q9, [x21, #0x50]\n" + ".inst 0x4f8ce07a // sdot v26.4s, v3.16b, v12.4b[0]\n" + ".inst 0x4face06a // sdot v10.4s, v3.16b, v12.4b[1]\n" + ".inst 0x4f8ce862 // sdot v2.4s, v3.16b, v12.4b[2]\n" + ".inst 0x4face87d // sdot v29.4s, v3.16b, v12.4b[3]\n" + "ldr q12, [x21, #0x60]\n" + ".inst 0x4f89e2da // sdot v26.4s, v22.16b, v9.4b[0]\n" + ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" + ".inst 0x4f89eac2 // sdot v2.4s, v22.16b, v9.4b[2]\n" + ".inst 0x4fa9eadd // sdot v29.4s, v22.16b, v9.4b[3]\n" + "ldr q17, [x21, #0x70]\n" + "add x21, x21, #0x88\n" + ".inst 0x4f8ce37a // sdot v26.4s, v27.16b, v12.4b[0]\n" + ".inst 0x4face36a // sdot v10.4s, v27.16b, v12.4b[1]\n" + ".inst 0x4f8ceb62 // sdot v2.4s, v27.16b, v12.4b[2]\n" + ".inst 0x4faceb7d // sdot v29.4s, v27.16b, v12.4b[3]\n" + ".inst 0x4f91e3da // sdot v26.4s, v30.16b, v17.4b[0]\n" + ".inst 0x4fb1e3ca // sdot v10.4s, v30.16b, v17.4b[1]\n" + ".inst 0x4f91ebc2 // sdot v2.4s, v30.16b, v17.4b[2]\n" + ".inst 0x4fb1ebdd // sdot v29.4s, v30.16b, v17.4b[3]\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "scvtf v10.4s, v10.4s, #0x4\n" + "fmla v5.4s, v26.4s, v24.4s\n" + "scvtf v2.4s, v2.4s, #0x4\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "fmla v21.4s, v10.4s, v31.4s\n" + "fmla v8.4s, v2.4s, v6.4s\n" + "fmla v1.4s, v29.4s, v20.4s\n" + "bgt 3b\n" + "mov x20, %x[res_ptr]\n" + "subs x27, x27, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "str q15, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q19, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q18, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q14, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q11, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q13, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q23, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q16, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q25, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q7, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q0, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q4, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q5, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q21, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q8, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q1, [x20, #0x0]\n" + "bne 2b\n" + "mov x20, #0x4\n" + "sub x10, x10, #0x10\n" + "cmp x10, #0x10\n" + "mov %x[res_ptr], x26\n" + "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" + "bge 1b\n" + "4:" // Row loop skip + "cbz x10, 9f\n" + "5:" // Row tail: Row loop + "add x24, %x[b_ptr], #0x8\n" + "mov x23, %x[nc]\n" + "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" + "6:" // Row tail: Column loop + "movi v15.16b, #0x0\n" + "movi v19.16b, #0x0\n" + "add x25, %x[a_ptr], #0x8\n" + "mov x21, %x[nb]\n" + "movi v18.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "7:" // Row tail: Block loop + "ldr q7, [x24, #0x0]\n" + "ldr q5, [x25, #0x0]\n" + "movi v9.16b, #0x4\n" + "movi v4.4s, #0x0\n" + "ldr q3, [x24, #0x10]\n" + "ldr q2, [x25, #0x10]\n" + "movi v1.4s, #0x0\n" + "movi v0.4s, #0x0\n" + "ldr q13, [x24, #0x20]\n" + "ldr q31, [x25, #0x20]\n" + "movi v30.4s, #0x0\n" + "movi v29.16b, #0xf0\n" + "ldr q28, [x24, #0x30]\n" + "ldr q27, [x25, #0x30]\n" + "sshl v20.16b, v7.16b, v9.16b\n" + "sub x20, x24, #0x8\n" + "ldr q26, [x25, #0x40]\n" + "ldr q25, [x25, #0x50]\n" + "sshl v17.16b, v3.16b, v9.16b\n" + "and v7.16b, v7.16b, v29.16b\n" + "ldr q24, [x25, #0x60]\n" + "ldr q16, [x25, #0x70]\n" + "sshl v22.16b, v13.16b, v9.16b\n" + "and v3.16b, v3.16b, v29.16b\n" + "ldr d21, [x20, #0x0]\n" + "ldr d12, [x25, #-0x8]\n" + ".inst 0x4f85e284 // sdot v4.4s, v20.16b, v5.4b[0]\n" + ".inst 0x4fa5e281 // sdot v1.4s, v20.16b, v5.4b[1]\n" + ".inst 0x4f85ea80 // sdot v0.4s, v20.16b, v5.4b[2]\n" + ".inst 0x4fa5ea9e // sdot v30.4s, v20.16b, v5.4b[3]\n" + "sshl v9.16b, v28.16b, v9.16b\n" + "subs x21, x21, #0x1\n" + "and v13.16b, v13.16b, v29.16b\n" + "and v28.16b, v28.16b, v29.16b\n" + "add x25, x25, #0x88\n" + "add x24, x24, #0x48\n" + "fcvtl v21.4s, v21.4h\n" + "fcvtl v12.4s, v12.4h\n" + ".inst 0x4f82e224 // sdot v4.4s, v17.16b, v2.4b[0]\n" + ".inst 0x4fa2e221 // sdot v1.4s, v17.16b, v2.4b[1]\n" + ".inst 0x4f82ea20 // sdot v0.4s, v17.16b, v2.4b[2]\n" + ".inst 0x4fa2ea3e // sdot v30.4s, v17.16b, v2.4b[3]\n" + "fmul v11.4s, v21.4s, v12.s[0]\n" + "fmul v23.4s, v21.4s, v12.s[1]\n" + "fmul v17.4s, v21.4s, v12.s[2]\n" + ".inst 0x4f9fe2c4 // sdot v4.4s, v22.16b, v31.4b[0]\n" + "fmul v6.4s, v21.4s, v12.s[3]\n" + ".inst 0x4fbfe2c1 // sdot v1.4s, v22.16b, v31.4b[1]\n" + ".inst 0x4f9feac0 // sdot v0.4s, v22.16b, v31.4b[2]\n" + ".inst 0x4fbfeade // sdot v30.4s, v22.16b, v31.4b[3]\n" + ".inst 0x4f9be124 // sdot v4.4s, v9.16b, v27.4b[0]\n" + ".inst 0x4fbbe121 // sdot v1.4s, v9.16b, v27.4b[1]\n" + ".inst 0x4f9be920 // sdot v0.4s, v9.16b, v27.4b[2]\n" + ".inst 0x4fbbe93e // sdot v30.4s, v9.16b, v27.4b[3]\n" + ".inst 0x4f9ae0e4 // sdot v4.4s, v7.16b, v26.4b[0]\n" + ".inst 0x4fbae0e1 // sdot v1.4s, v7.16b, v26.4b[1]\n" + ".inst 0x4f9ae8e0 // sdot v0.4s, v7.16b, v26.4b[2]\n" + ".inst 0x4fbae8fe // sdot v30.4s, v7.16b, v26.4b[3]\n" + ".inst 0x4f99e064 // sdot v4.4s, v3.16b, v25.4b[0]\n" + ".inst 0x4fb9e061 // sdot v1.4s, v3.16b, v25.4b[1]\n" + ".inst 0x4f99e860 // sdot v0.4s, v3.16b, v25.4b[2]\n" + ".inst 0x4fb9e87e // sdot v30.4s, v3.16b, v25.4b[3]\n" + ".inst 0x4f98e1a4 // sdot v4.4s, v13.16b, v24.4b[0]\n" + ".inst 0x4fb8e1a1 // sdot v1.4s, v13.16b, v24.4b[1]\n" + ".inst 0x4f98e9a0 // sdot v0.4s, v13.16b, v24.4b[2]\n" + ".inst 0x4fb8e9be // sdot v30.4s, v13.16b, v24.4b[3]\n" + ".inst 0x4f90e384 // sdot v4.4s, v28.16b, v16.4b[0]\n" + ".inst 0x4fb0e381 // sdot v1.4s, v28.16b, v16.4b[1]\n" + ".inst 0x4f90eb80 // sdot v0.4s, v28.16b, v16.4b[2]\n" + ".inst 0x4fb0eb9e // sdot v30.4s, v28.16b, v16.4b[3]\n" + "scvtf v4.4s, v4.4s, #0x4\n" + "scvtf v1.4s, v1.4s, #0x4\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "fmla v15.4s, v4.4s, v11.4s\n" + "scvtf v30.4s, v30.4s, #0x4\n" + "fmla v19.4s, v1.4s, v23.4s\n" + "fmla v18.4s, v0.4s, v17.4s\n" + "fmla v14.4s, v30.4s, v6.4s\n" + "bgt 7b\n" + "mov x20, %x[res_ptr]\n" + "cmp x10, #0x1\n" + "str q15, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x2\n" + "str q19, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x3\n" + "str q18, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "str q14, [x20, #0x0]\n" + "8:" // Row tail: Accumulator store skip + "subs x23, x23, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "bne 6b\n" + "subs x10, x10, #0x4\n" + "add %x[a_ptr], %x[a_ptr], x9\n" + "mov %x[res_ptr], x22\n" + "bgt 5b\n" + "9:" // Row tail: Row loop skip + : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) + : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + ); + return; } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) && - "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - size_t res_stride = bs * sizeof(float); +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) + { + float sumf[4][4]; + int sumi; - __asm__ __volatile__( - "mov x10, %x[nr]\n" - "mov x9, #0x88\n" - "cmp x10, #0x10\n" - "mul x9, %x[nb], x9\n" - "blt 4f\n" - "1:" // Row loop - "add x28, %x[b_ptr], #0x8\n" - "mov x27, %x[nc]\n" - "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" - "2:" // Column loop - "add x25, %x[a_ptr], #0x8\n" - "movi v15.16b, #0x0\n" - "movi v19.16b, #0x0\n" - "mov x24, %x[nb]\n" - "add x23, x25, x9\n" - "movi v18.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "add x22, x23, x9\n" - "movi v11.16b, #0x0\n" - "movi v13.16b, #0x0\n" - "add x21, x22, x9\n" - "movi v23.16b, #0x0\n" - "movi v16.16b, #0x0\n" - "movi v25.16b, #0x0\n" - "movi v7.16b, #0x0\n" - "movi v0.16b, #0x0\n" - "movi v4.16b, #0x0\n" - "movi v5.16b, #0x0\n" - "movi v21.16b, #0x0\n" - "movi v8.16b, #0x0\n" - "movi v1.16b, #0x0\n" - "3:" // Block loop - "ldr q3, [x28, #0x0]\n" - "ldr q31, [x25, #0x0]\n" - "movi v28.16b, #0x4\n" - "movi v10.4s, #0x0\n" - "ldr q22, [x28, #0x10]\n" - "ldr q6, [x25, #0x10]\n" - "movi v29.4s, #0x0\n" - "movi v9.4s, #0x0\n" - "ldr q27, [x28, #0x20]\n" - "ldr q30, [x28, #0x30]\n" - "movi v20.4s, #0x0\n" - "movi v24.16b, #0xf0\n" - "ldr d2, [x25, #-0x8]\n" - "ldr d26, [x23, #-0x8]\n" - "sshl v12.16b, v3.16b, v28.16b\n" - "sub x20, x28, #0x8\n" - "ldr d17, [x20, #0x0]\n" - "and v3.16b, v3.16b, v24.16b\n" - "subs x24, x24, #0x1\n" - "add x28, x28, #0x48\n" - ".inst 0x4f9fe18a // sdot v10.4s, v12.16b, v31.4b[0]\n" - ".inst 0x4fbfe19d // sdot v29.4s, v12.16b, v31.4b[1]\n" - ".inst 0x4f9fe989 // sdot v9.4s, v12.16b, v31.4b[2]\n" - ".inst 0x4fbfe994 // sdot v20.4s, v12.16b, v31.4b[3]\n" - "sshl v31.16b, v22.16b, v28.16b\n" - "and v22.16b, v22.16b, v24.16b\n" - "fcvtl v17.4s, v17.4h\n" - "fcvtl v2.4s, v2.4h\n" - "fcvtl v26.4s, v26.4h\n" - ".inst 0x4f86e3ea // sdot v10.4s, v31.16b, v6.4b[0]\n" - ".inst 0x4fa6e3fd // sdot v29.4s, v31.16b, v6.4b[1]\n" - ".inst 0x4f86ebe9 // sdot v9.4s, v31.16b, v6.4b[2]\n" - ".inst 0x4fa6ebf4 // sdot v20.4s, v31.16b, v6.4b[3]\n" - "sshl v6.16b, v27.16b, v28.16b\n" - "sshl v28.16b, v30.16b, v28.16b\n" - "and v27.16b, v27.16b, v24.16b\n" - "and v30.16b, v30.16b, v24.16b\n" - "ldr q24, [x25, #0x20]\n" - ".inst 0x4f98e0ca // sdot v10.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8c9 // sdot v9.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8d4 // sdot v20.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x30]\n" - ".inst 0x4f98e38a // sdot v10.4s, v28.16b, v24.4b[0]\n" - ".inst 0x4fb8e39d // sdot v29.4s, v28.16b, v24.4b[1]\n" - ".inst 0x4f98eb89 // sdot v9.4s, v28.16b, v24.4b[2]\n" - ".inst 0x4fb8eb94 // sdot v20.4s, v28.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x40]\n" - ".inst 0x4f98e06a // sdot v10.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e869 // sdot v9.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e874 // sdot v20.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x50]\n" - ".inst 0x4f98e2ca // sdot v10.4s, v22.16b, v24.4b[0]\n" - ".inst 0x4fb8e2dd // sdot v29.4s, v22.16b, v24.4b[1]\n" - ".inst 0x4f98eac9 // sdot v9.4s, v22.16b, v24.4b[2]\n" - ".inst 0x4fb8ead4 // sdot v20.4s, v22.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x60]\n" - ".inst 0x4f98e36a // sdot v10.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb69 // sdot v9.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb74 // sdot v20.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x70]\n" - "add x25, x25, #0x88\n" - ".inst 0x4f98e3ca // sdot v10.4s, v30.16b, v24.4b[0]\n" - ".inst 0x4fb8e3dd // sdot v29.4s, v30.16b, v24.4b[1]\n" - ".inst 0x4f98ebc9 // sdot v9.4s, v30.16b, v24.4b[2]\n" - ".inst 0x4fb8ebd4 // sdot v20.4s, v30.16b, v24.4b[3]\n" - "fmul v24.4s, v17.4s, v2.s[0]\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v15.4s, v10.4s, v24.4s\n" - "ldr q24, [x23, #0x0]\n" - "fmul v10.4s, v17.4s, v2.s[1]\n" - "fmla v19.4s, v29.4s, v10.4s\n" - "ldr q10, [x23, #0x10]\n" - "fmul v29.4s, v17.4s, v2.s[2]\n" - "fmul v2.4s, v17.4s, v2.s[3]\n" - "fmla v18.4s, v9.4s, v29.4s\n" - "movi v9.4s, #0x0\n" - "movi v29.4s, #0x0\n" - ".inst 0x4f98e189 // sdot v9.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e19d // sdot v29.4s, v12.16b, v24.4b[1]\n" - "fmla v14.4s, v20.4s, v2.4s\n" - "movi v20.4s, #0x0\n" - "movi v2.4s, #0x0\n" - ".inst 0x4f98e994 // sdot v20.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x20]\n" - ".inst 0x4f8ae3e9 // sdot v9.4s, v31.16b, v10.4b[0]\n" - ".inst 0x4faae3fd // sdot v29.4s, v31.16b, v10.4b[1]\n" - ".inst 0x4f8aebf4 // sdot v20.4s, v31.16b, v10.4b[2]\n" - ".inst 0x4faaebe2 // sdot v2.4s, v31.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x30]\n" - ".inst 0x4f98e0c9 // sdot v9.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8d4 // sdot v20.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x40]\n" - ".inst 0x4f8ae389 // sdot v9.4s, v28.16b, v10.4b[0]\n" - ".inst 0x4faae39d // sdot v29.4s, v28.16b, v10.4b[1]\n" - ".inst 0x4f8aeb94 // sdot v20.4s, v28.16b, v10.4b[2]\n" - ".inst 0x4faaeb82 // sdot v2.4s, v28.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x50]\n" - ".inst 0x4f98e069 // sdot v9.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e874 // sdot v20.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x60]\n" - ".inst 0x4f8ae2c9 // sdot v9.4s, v22.16b, v10.4b[0]\n" - ".inst 0x4faae2dd // sdot v29.4s, v22.16b, v10.4b[1]\n" - ".inst 0x4f8aead4 // sdot v20.4s, v22.16b, v10.4b[2]\n" - ".inst 0x4faaeac2 // sdot v2.4s, v22.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x70]\n" - "add x23, x23, #0x88\n" - ".inst 0x4f98e369 // sdot v9.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb74 // sdot v20.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x0]\n" - ".inst 0x4f8ae3c9 // sdot v9.4s, v30.16b, v10.4b[0]\n" - ".inst 0x4faae3dd // sdot v29.4s, v30.16b, v10.4b[1]\n" - ".inst 0x4f8aebd4 // sdot v20.4s, v30.16b, v10.4b[2]\n" - ".inst 0x4faaebc2 // sdot v2.4s, v30.16b, v10.4b[3]\n" - "fmul v10.4s, v17.4s, v26.s[0]\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "fmla v11.4s, v9.4s, v10.4s\n" - "ldr q9, [x22, #0x10]\n" - "fmul v10.4s, v17.4s, v26.s[1]\n" - "fmla v13.4s, v29.4s, v10.4s\n" - "ldr d29, [x22, #-0x8]\n" - "fmul v10.4s, v17.4s, v26.s[2]\n" - "fmul v26.4s, v17.4s, v26.s[3]\n" - "fcvtl v29.4s, v29.4h\n" - "fmla v23.4s, v20.4s, v10.4s\n" - "movi v20.4s, #0x0\n" - "movi v10.4s, #0x0\n" - "fmla v16.4s, v2.4s, v26.4s\n" - "movi v26.4s, #0x0\n" - "movi v2.4s, #0x0\n" - ".inst 0x4f98e194 // sdot v20.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" - ".inst 0x4f98e99a // sdot v26.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x20]\n" - ".inst 0x4f89e3f4 // sdot v20.4s, v31.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" - ".inst 0x4f89ebfa // sdot v26.4s, v31.16b, v9.4b[2]\n" - ".inst 0x4fa9ebe2 // sdot v2.4s, v31.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x30]\n" - ".inst 0x4f98e0d4 // sdot v20.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0ca // sdot v10.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8da // sdot v26.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x40]\n" - ".inst 0x4f89e394 // sdot v20.4s, v28.16b, v9.4b[0]\n" - ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" - ".inst 0x4f89eb9a // sdot v26.4s, v28.16b, v9.4b[2]\n" - ".inst 0x4fa9eb82 // sdot v2.4s, v28.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x50]\n" - ".inst 0x4f98e074 // sdot v20.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e06a // sdot v10.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e87a // sdot v26.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x60]\n" - ".inst 0x4f89e2d4 // sdot v20.4s, v22.16b, v9.4b[0]\n" - ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" - ".inst 0x4f89eada // sdot v26.4s, v22.16b, v9.4b[2]\n" - ".inst 0x4fa9eac2 // sdot v2.4s, v22.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x70]\n" - "add x22, x22, #0x88\n" - ".inst 0x4f98e374 // sdot v20.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e36a // sdot v10.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb7a // sdot v26.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x21, #0x0]\n" - ".inst 0x4f89e3d4 // sdot v20.4s, v30.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ca // sdot v10.4s, v30.16b, v9.4b[1]\n" - ".inst 0x4f89ebda // sdot v26.4s, v30.16b, v9.4b[2]\n" - ".inst 0x4fa9ebc2 // sdot v2.4s, v30.16b, v9.4b[3]\n" - "fmul v9.4s, v17.4s, v29.s[0]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "fmla v25.4s, v20.4s, v9.4s\n" - "ldr q9, [x21, #0x10]\n" - "fmul v20.4s, v17.4s, v29.s[1]\n" - "fmla v7.4s, v10.4s, v20.4s\n" - "ldr d20, [x21, #-0x8]\n" - "fmul v10.4s, v17.4s, v29.s[2]\n" - "fmul v29.4s, v17.4s, v29.s[3]\n" - "fcvtl v20.4s, v20.4h\n" - "fmla v0.4s, v26.4s, v10.4s\n" - "movi v26.4s, #0x0\n" - "movi v10.4s, #0x0\n" - "fmla v4.4s, v2.4s, v29.4s\n" - "movi v2.4s, #0x0\n" - "movi v29.4s, #0x0\n" - ".inst 0x4f98e19a // sdot v26.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" - ".inst 0x4f98e982 // sdot v2.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e99d // sdot v29.4s, v12.16b, v24.4b[3]\n" - "ldr q12, [x21, #0x20]\n" - "fmul v24.4s, v17.4s, v20.s[0]\n" - ".inst 0x4f89e3fa // sdot v26.4s, v31.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" - ".inst 0x4f89ebe2 // sdot v2.4s, v31.16b, v9.4b[2]\n" - ".inst 0x4fa9ebfd // sdot v29.4s, v31.16b, v9.4b[3]\n" - "ldr q9, [x21, #0x30]\n" - "fmul v31.4s, v17.4s, v20.s[1]\n" - ".inst 0x4f8ce0da // sdot v26.4s, v6.16b, v12.4b[0]\n" - ".inst 0x4face0ca // sdot v10.4s, v6.16b, v12.4b[1]\n" - ".inst 0x4f8ce8c2 // sdot v2.4s, v6.16b, v12.4b[2]\n" - ".inst 0x4face8dd // sdot v29.4s, v6.16b, v12.4b[3]\n" - "ldr q12, [x21, #0x40]\n" - "fmul v6.4s, v17.4s, v20.s[2]\n" - "fmul v20.4s, v17.4s, v20.s[3]\n" - ".inst 0x4f89e39a // sdot v26.4s, v28.16b, v9.4b[0]\n" - ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" - ".inst 0x4f89eb82 // sdot v2.4s, v28.16b, v9.4b[2]\n" - ".inst 0x4fa9eb9d // sdot v29.4s, v28.16b, v9.4b[3]\n" - "ldr q9, [x21, #0x50]\n" - ".inst 0x4f8ce07a // sdot v26.4s, v3.16b, v12.4b[0]\n" - ".inst 0x4face06a // sdot v10.4s, v3.16b, v12.4b[1]\n" - ".inst 0x4f8ce862 // sdot v2.4s, v3.16b, v12.4b[2]\n" - ".inst 0x4face87d // sdot v29.4s, v3.16b, v12.4b[3]\n" - "ldr q12, [x21, #0x60]\n" - ".inst 0x4f89e2da // sdot v26.4s, v22.16b, v9.4b[0]\n" - ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" - ".inst 0x4f89eac2 // sdot v2.4s, v22.16b, v9.4b[2]\n" - ".inst 0x4fa9eadd // sdot v29.4s, v22.16b, v9.4b[3]\n" - "ldr q17, [x21, #0x70]\n" - "add x21, x21, #0x88\n" - ".inst 0x4f8ce37a // sdot v26.4s, v27.16b, v12.4b[0]\n" - ".inst 0x4face36a // sdot v10.4s, v27.16b, v12.4b[1]\n" - ".inst 0x4f8ceb62 // sdot v2.4s, v27.16b, v12.4b[2]\n" - ".inst 0x4faceb7d // sdot v29.4s, v27.16b, v12.4b[3]\n" - ".inst 0x4f91e3da // sdot v26.4s, v30.16b, v17.4b[0]\n" - ".inst 0x4fb1e3ca // sdot v10.4s, v30.16b, v17.4b[1]\n" - ".inst 0x4f91ebc2 // sdot v2.4s, v30.16b, v17.4b[2]\n" - ".inst 0x4fb1ebdd // sdot v29.4s, v30.16b, v17.4b[3]\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "fmla v5.4s, v26.4s, v24.4s\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "fmla v21.4s, v10.4s, v31.4s\n" - "fmla v8.4s, v2.4s, v6.4s\n" - "fmla v1.4s, v29.4s, v20.4s\n" - "bgt 3b\n" - "mov x20, %x[res_ptr]\n" - "subs x27, x27, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "str q15, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q19, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q18, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q14, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q11, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q13, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q23, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q16, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q25, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q7, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q0, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q4, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q5, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q21, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q8, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q1, [x20, #0x0]\n" - "bne 2b\n" - "mov x20, #0x4\n" - "sub x10, x10, #0x10\n" - "cmp x10, #0x10\n" - "mov %x[res_ptr], x26\n" - "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" - "bge 1b\n" - "4:" // Row loop skip - "cbz x10, 9f\n" - "5:" // Row tail: Row loop - "add x24, %x[b_ptr], #0x8\n" - "mov x23, %x[nc]\n" - "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" - "6:" // Row tail: Column loop - "movi v15.16b, #0x0\n" - "movi v19.16b, #0x0\n" - "add x25, %x[a_ptr], #0x8\n" - "mov x21, %x[nb]\n" - "movi v18.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "7:" // Row tail: Block loop - "ldr q7, [x24, #0x0]\n" - "ldr q5, [x25, #0x0]\n" - "movi v9.16b, #0x4\n" - "movi v4.4s, #0x0\n" - "ldr q3, [x24, #0x10]\n" - "ldr q2, [x25, #0x10]\n" - "movi v1.4s, #0x0\n" - "movi v0.4s, #0x0\n" - "ldr q13, [x24, #0x20]\n" - "ldr q31, [x25, #0x20]\n" - "movi v30.4s, #0x0\n" - "movi v29.16b, #0xf0\n" - "ldr q28, [x24, #0x30]\n" - "ldr q27, [x25, #0x30]\n" - "sshl v20.16b, v7.16b, v9.16b\n" - "sub x20, x24, #0x8\n" - "ldr q26, [x25, #0x40]\n" - "ldr q25, [x25, #0x50]\n" - "sshl v17.16b, v3.16b, v9.16b\n" - "and v7.16b, v7.16b, v29.16b\n" - "ldr q24, [x25, #0x60]\n" - "ldr q16, [x25, #0x70]\n" - "sshl v22.16b, v13.16b, v9.16b\n" - "and v3.16b, v3.16b, v29.16b\n" - "ldr d21, [x20, #0x0]\n" - "ldr d12, [x25, #-0x8]\n" - ".inst 0x4f85e284 // sdot v4.4s, v20.16b, v5.4b[0]\n" - ".inst 0x4fa5e281 // sdot v1.4s, v20.16b, v5.4b[1]\n" - ".inst 0x4f85ea80 // sdot v0.4s, v20.16b, v5.4b[2]\n" - ".inst 0x4fa5ea9e // sdot v30.4s, v20.16b, v5.4b[3]\n" - "sshl v9.16b, v28.16b, v9.16b\n" - "subs x21, x21, #0x1\n" - "and v13.16b, v13.16b, v29.16b\n" - "and v28.16b, v28.16b, v29.16b\n" - "add x25, x25, #0x88\n" - "add x24, x24, #0x48\n" - "fcvtl v21.4s, v21.4h\n" - "fcvtl v12.4s, v12.4h\n" - ".inst 0x4f82e224 // sdot v4.4s, v17.16b, v2.4b[0]\n" - ".inst 0x4fa2e221 // sdot v1.4s, v17.16b, v2.4b[1]\n" - ".inst 0x4f82ea20 // sdot v0.4s, v17.16b, v2.4b[2]\n" - ".inst 0x4fa2ea3e // sdot v30.4s, v17.16b, v2.4b[3]\n" - "fmul v11.4s, v21.4s, v12.s[0]\n" - "fmul v23.4s, v21.4s, v12.s[1]\n" - "fmul v17.4s, v21.4s, v12.s[2]\n" - ".inst 0x4f9fe2c4 // sdot v4.4s, v22.16b, v31.4b[0]\n" - "fmul v6.4s, v21.4s, v12.s[3]\n" - ".inst 0x4fbfe2c1 // sdot v1.4s, v22.16b, v31.4b[1]\n" - ".inst 0x4f9feac0 // sdot v0.4s, v22.16b, v31.4b[2]\n" - ".inst 0x4fbfeade // sdot v30.4s, v22.16b, v31.4b[3]\n" - ".inst 0x4f9be124 // sdot v4.4s, v9.16b, v27.4b[0]\n" - ".inst 0x4fbbe121 // sdot v1.4s, v9.16b, v27.4b[1]\n" - ".inst 0x4f9be920 // sdot v0.4s, v9.16b, v27.4b[2]\n" - ".inst 0x4fbbe93e // sdot v30.4s, v9.16b, v27.4b[3]\n" - ".inst 0x4f9ae0e4 // sdot v4.4s, v7.16b, v26.4b[0]\n" - ".inst 0x4fbae0e1 // sdot v1.4s, v7.16b, v26.4b[1]\n" - ".inst 0x4f9ae8e0 // sdot v0.4s, v7.16b, v26.4b[2]\n" - ".inst 0x4fbae8fe // sdot v30.4s, v7.16b, v26.4b[3]\n" - ".inst 0x4f99e064 // sdot v4.4s, v3.16b, v25.4b[0]\n" - ".inst 0x4fb9e061 // sdot v1.4s, v3.16b, v25.4b[1]\n" - ".inst 0x4f99e860 // sdot v0.4s, v3.16b, v25.4b[2]\n" - ".inst 0x4fb9e87e // sdot v30.4s, v3.16b, v25.4b[3]\n" - ".inst 0x4f98e1a4 // sdot v4.4s, v13.16b, v24.4b[0]\n" - ".inst 0x4fb8e1a1 // sdot v1.4s, v13.16b, v24.4b[1]\n" - ".inst 0x4f98e9a0 // sdot v0.4s, v13.16b, v24.4b[2]\n" - ".inst 0x4fb8e9be // sdot v30.4s, v13.16b, v24.4b[3]\n" - ".inst 0x4f90e384 // sdot v4.4s, v28.16b, v16.4b[0]\n" - ".inst 0x4fb0e381 // sdot v1.4s, v28.16b, v16.4b[1]\n" - ".inst 0x4f90eb80 // sdot v0.4s, v28.16b, v16.4b[2]\n" - ".inst 0x4fb0eb9e // sdot v30.4s, v28.16b, v16.4b[3]\n" - "scvtf v4.4s, v4.4s, #0x4\n" - "scvtf v1.4s, v1.4s, #0x4\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "fmla v15.4s, v4.4s, v11.4s\n" - "scvtf v30.4s, v30.4s, #0x4\n" - "fmla v19.4s, v1.4s, v23.4s\n" - "fmla v18.4s, v0.4s, v17.4s\n" - "fmla v14.4s, v30.4s, v6.4s\n" - "bgt 7b\n" - "mov x20, %x[res_ptr]\n" - "cmp x10, #0x1\n" - "str q15, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x2\n" - "str q19, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x3\n" - "str q18, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "str q14, [x20, #0x0]\n" - "8:" // Row tail: Accumulator store skip - "subs x23, x23, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "bne 6b\n" - "subs x10, x10, #0x4\n" - "add %x[a_ptr], %x[a_ptr], x9\n" - "mov %x[res_ptr], x22\n" - "bgt 5b\n" - "9:" // Row tail: Row loop skip - : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) - : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" - ); -#else - float sumf[4][4]; - int sumi; - - for (int y = 0; y < nr / 4; y++) { - const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; - } - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); } - sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); } } } - } - for (int m = 0; m < 4; m++) { - for (int j = 0; j < ncols_interleaved; j++) - s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } } } } -#endif } void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { @@ -1586,413 +1564,406 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(ncols_interleaved); UNUSED(blocklen); -#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) - if (ggml_sve_cnt_b == QK8_0) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - size_t res_stride = bs * sizeof(float); +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) + if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { + const void * b_ptr = vx; + const void * a_ptr = vy; + float * res_ptr = s; + size_t res_stride = bs * sizeof(float); - __asm__ __volatile__( - "mov x10, %x[nr]\n" - "mov x9, #0x88\n" - "cmp x10, #0x10\n" - "mul x9, %x[nb], x9\n" - "blt 4f\n" - "1:" // Row loop - "add x28, %x[b_ptr], #0x8\n" - "mov x27, %x[nc]\n" - "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" - "2:" // Column loop - "add x25, %x[a_ptr], #0x8\n" - "movi v2.16b, #0x0\n" - "movi v10.16b, #0x0\n" - "mov x24, %x[nb]\n" - "add x23, x25, x9\n" - "movi v12.16b, #0x0\n" - "movi v28.16b, #0x0\n" - "add x22, x23, x9\n" - "movi v11.16b, #0x0\n" - "movi v13.16b, #0x0\n" - "add x21, x22, x9\n" - "movi v22.16b, #0x0\n" - "movi v23.16b, #0x0\n" - "movi v25.16b, #0x0\n" - "movi v5.16b, #0x0\n" - "movi v7.16b, #0x0\n" - "movi v4.16b, #0x0\n" - "movi v6.16b, #0x0\n" - "movi v30.16b, #0x0\n" - "movi v24.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "3:" // Block loop - "ldr q21, [x28, #0x0]\n" - "ldr q16, [x28, #0x10]\n" - "movi v1.16b, #0x4\n" - "movi v19.4s, #0x0\n" - "ldr q27, [x25, #0x0]\n" - "ldr q15, [x25, #0x10]\n" - "movi v26.4s, #0x0\n" - "movi v18.4s, #0x0\n" - "ldr q29, [x28, #0x20]\n" - "ldr q3, [x28, #0x30]\n" - "movi v17.4s, #0x0\n" - "movi v0.16b, #0xf0\n" - "ldr d20, [x25, #-0x8]\n" - "ldr d9, [x23, #-0x8]\n" - "sshl v8.16b, v21.16b, v1.16b\n" - "sshl v31.16b, v16.16b, v1.16b\n" - "and v21.16b, v21.16b, v0.16b\n" - "and v16.16b, v16.16b, v0.16b\n" - "sub x20, x28, #0x8\n" - "subs x24, x24, #0x1\n" - "add x28, x28, #0x48\n" - ".inst 0x4e88a773 // smmla v19.4s, v27.16b, v8.16b\n" - ".inst 0x4e9fa77a // smmla v26.4s, v27.16b, v31.16b\n" - "ldr q27, [x25, #0x20]\n" - ".inst 0x4e88a5f2 // smmla v18.4s, v15.16b, v8.16b\n" - ".inst 0x4e9fa5f1 // smmla v17.4s, v15.16b, v31.16b\n" - "sshl v15.16b, v29.16b, v1.16b\n" - "sshl v1.16b, v3.16b, v1.16b\n" - "and v29.16b, v29.16b, v0.16b\n" - "and v3.16b, v3.16b, v0.16b\n" - "ldr q0, [x25, #0x30]\n" - "fcvtl v20.4s, v20.4h\n" - ".inst 0x4e8fa773 // smmla v19.4s, v27.16b, v15.16b\n" - "fcvtl v9.4s, v9.4h\n" - ".inst 0x4e81a77a // smmla v26.4s, v27.16b, v1.16b\n" - "ldr q27, [x25, #0x40]\n" - ".inst 0x4e8fa412 // smmla v18.4s, v0.16b, v15.16b\n" - ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n" - "ldr q0, [x25, #0x50]\n" - ".inst 0x4e95a773 // smmla v19.4s, v27.16b, v21.16b\n" - ".inst 0x4e90a77a // smmla v26.4s, v27.16b, v16.16b\n" - "ldr q27, [x25, #0x60]\n" - ".inst 0x4e95a412 // smmla v18.4s, v0.16b, v21.16b\n" - ".inst 0x4e90a411 // smmla v17.4s, v0.16b, v16.16b\n" - "ldr q0, [x25, #0x70]\n" - "add x25, x25, #0x88\n" - ".inst 0x4e9da773 // smmla v19.4s, v27.16b, v29.16b\n" - ".inst 0x4e83a77a // smmla v26.4s, v27.16b, v3.16b\n" - "ldr d27, [x20, #0x0]\n" - ".inst 0x4e9da412 // smmla v18.4s, v0.16b, v29.16b\n" - ".inst 0x4e83a411 // smmla v17.4s, v0.16b, v3.16b\n" - "fcvtl v27.4s, v27.4h\n" - "uzp1 v0.2d, v19.2d, v26.2d\n" - "uzp2 v26.2d, v19.2d, v26.2d\n" - "fmul v19.4s, v27.4s, v20.s[0]\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "fmla v2.4s, v0.4s, v19.4s\n" - "ldr q19, [x23, #0x0]\n" - "uzp1 v0.2d, v18.2d, v17.2d\n" - "uzp2 v18.2d, v18.2d, v17.2d\n" - "fmul v17.4s, v27.4s, v20.s[1]\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "fmla v10.4s, v26.4s, v17.4s\n" - "ldr q17, [x23, #0x10]\n" - "fmul v26.4s, v27.4s, v20.s[2]\n" - "fmul v20.4s, v27.4s, v20.s[3]\n" - "fmla v12.4s, v0.4s, v26.4s\n" - "ldr d0, [x22, #-0x8]\n" - "ldr d26, [x21, #-0x8]\n" - "fcvtl v0.4s, v0.4h\n" - "fmla v28.4s, v18.4s, v20.4s\n" - "movi v20.4s, #0x0\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" - ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" - "ldr q19, [x23, #0x20]\n" - "fcvtl v26.4s, v26.4h\n" - ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" - ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" - "ldr q19, [x23, #0x40]\n" - ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" - ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" - "ldr q19, [x23, #0x60]\n" - ".inst 0x4e9da674 // smmla v20.4s, v19.16b, v29.16b\n" - ".inst 0x4e83a672 // smmla v18.4s, v19.16b, v3.16b\n" - "uzp1 v19.2d, v20.2d, v18.2d\n" - "scvtf v19.4s, v19.4s, #0x4\n" - "uzp2 v20.2d, v20.2d, v18.2d\n" - "fmul v18.4s, v27.4s, v9.s[0]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v11.4s, v19.4s, v18.4s\n" - "ldr q18, [x22, #0x0]\n" - "fmul v19.4s, v27.4s, v9.s[1]\n" - "fmla v13.4s, v20.4s, v19.4s\n" - "movi v19.4s, #0x0\n" - "movi v20.4s, #0x0\n" - ".inst 0x4e88a633 // smmla v19.4s, v17.16b, v8.16b\n" - ".inst 0x4e9fa634 // smmla v20.4s, v17.16b, v31.16b\n" - "ldr q17, [x23, #0x30]\n" - ".inst 0x4e8fa633 // smmla v19.4s, v17.16b, v15.16b\n" - ".inst 0x4e81a634 // smmla v20.4s, v17.16b, v1.16b\n" - "ldr q17, [x23, #0x50]\n" - ".inst 0x4e95a633 // smmla v19.4s, v17.16b, v21.16b\n" - ".inst 0x4e90a634 // smmla v20.4s, v17.16b, v16.16b\n" - "ldr q17, [x23, #0x70]\n" - "add x23, x23, #0x88\n" - ".inst 0x4e9da633 // smmla v19.4s, v17.16b, v29.16b\n" - ".inst 0x4e83a634 // smmla v20.4s, v17.16b, v3.16b\n" - "uzp1 v17.2d, v19.2d, v20.2d\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "uzp2 v20.2d, v19.2d, v20.2d\n" - "fmul v19.4s, v27.4s, v9.s[2]\n" - "fmul v9.4s, v27.4s, v9.s[3]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v22.4s, v17.4s, v19.4s\n" - "ldr q17, [x22, #0x10]\n" - "movi v19.4s, #0x0\n" - ".inst 0x4e88a653 // smmla v19.4s, v18.16b, v8.16b\n" - "fmla v23.4s, v20.4s, v9.4s\n" - "movi v20.4s, #0x0\n" - "movi v9.4s, #0x0\n" - ".inst 0x4e9fa654 // smmla v20.4s, v18.16b, v31.16b\n" - "ldr q18, [x22, #0x20]\n" - ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" - ".inst 0x4e8fa653 // smmla v19.4s, v18.16b, v15.16b\n" - ".inst 0x4e81a654 // smmla v20.4s, v18.16b, v1.16b\n" - "ldr q18, [x22, #0x40]\n" - ".inst 0x4e95a653 // smmla v19.4s, v18.16b, v21.16b\n" - ".inst 0x4e90a654 // smmla v20.4s, v18.16b, v16.16b\n" - "ldr q18, [x22, #0x60]\n" - ".inst 0x4e9da653 // smmla v19.4s, v18.16b, v29.16b\n" - ".inst 0x4e83a654 // smmla v20.4s, v18.16b, v3.16b\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e9fa632 // smmla v18.4s, v17.16b, v31.16b\n" - "ldr q17, [x22, #0x30]\n" - ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" - ".inst 0x4e81a632 // smmla v18.4s, v17.16b, v1.16b\n" - "ldr q17, [x22, #0x50]\n" - ".inst 0x4e95a629 // smmla v9.4s, v17.16b, v21.16b\n" - ".inst 0x4e90a632 // smmla v18.4s, v17.16b, v16.16b\n" - "ldr q17, [x22, #0x70]\n" - "add x22, x22, #0x88\n" - ".inst 0x4e9da629 // smmla v9.4s, v17.16b, v29.16b\n" - ".inst 0x4e83a632 // smmla v18.4s, v17.16b, v3.16b\n" - "uzp1 v17.2d, v19.2d, v20.2d\n" - "uzp2 v20.2d, v19.2d, v20.2d\n" - "fmul v19.4s, v27.4s, v0.s[0]\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v25.4s, v17.4s, v19.4s\n" - "ldr q19, [x21, #0x0]\n" - "fmul v17.4s, v27.4s, v0.s[1]\n" - "fmla v5.4s, v20.4s, v17.4s\n" - "ldr q17, [x21, #0x10]\n" - "uzp1 v20.2d, v9.2d, v18.2d\n" - "uzp2 v9.2d, v9.2d, v18.2d\n" - "fmul v18.4s, v27.4s, v0.s[2]\n" - "fmul v0.4s, v27.4s, v0.s[3]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "fmla v7.4s, v20.4s, v18.4s\n" - "movi v20.4s, #0x0\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" - ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" - "ldr q19, [x21, #0x20]\n" - "fmla v4.4s, v9.4s, v0.4s\n" - "movi v9.4s, #0x0\n" - "movi v0.4s, #0x0\n" - ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" - "fmul v8.4s, v27.4s, v26.s[0]\n" - ".inst 0x4e9fa620 // smmla v0.4s, v17.16b, v31.16b\n" - "ldr q17, [x21, #0x30]\n" - ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" - "fmul v31.4s, v27.4s, v26.s[1]\n" - ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" - "ldr q19, [x21, #0x40]\n" - ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" - "fmul v15.4s, v27.4s, v26.s[2]\n" - "fmul v27.4s, v27.4s, v26.s[3]\n" - ".inst 0x4e81a620 // smmla v0.4s, v17.16b, v1.16b\n" - "ldr q1, [x21, #0x50]\n" - ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" - ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" - "ldr q26, [x21, #0x60]\n" - ".inst 0x4e95a429 // smmla v9.4s, v1.16b, v21.16b\n" - ".inst 0x4e90a420 // smmla v0.4s, v1.16b, v16.16b\n" - "ldr q21, [x21, #0x70]\n" - "add x21, x21, #0x88\n" - ".inst 0x4e9da754 // smmla v20.4s, v26.16b, v29.16b\n" - ".inst 0x4e83a752 // smmla v18.4s, v26.16b, v3.16b\n" - ".inst 0x4e9da6a9 // smmla v9.4s, v21.16b, v29.16b\n" - ".inst 0x4e83a6a0 // smmla v0.4s, v21.16b, v3.16b\n" - "uzp1 v29.2d, v20.2d, v18.2d\n" - "uzp2 v21.2d, v20.2d, v18.2d\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "uzp1 v18.2d, v9.2d, v0.2d\n" - "uzp2 v16.2d, v9.2d, v0.2d\n" - "scvtf v21.4s, v21.4s, #0x4\n" - "fmla v6.4s, v29.4s, v8.4s\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "scvtf v16.4s, v16.4s, #0x4\n" - "fmla v30.4s, v21.4s, v31.4s\n" - "fmla v24.4s, v18.4s, v15.4s\n" - "fmla v14.4s, v16.4s, v27.4s\n" - "bgt 3b\n" - "mov x20, %x[res_ptr]\n" - "subs x27, x27, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "str q2, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q10, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q12, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q28, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q11, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q13, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q22, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q23, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q25, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q5, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q7, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q4, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q6, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q30, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q24, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q14, [x20, #0x0]\n" - "bne 2b\n" - "mov x20, #0x4\n" - "sub x10, x10, #0x10\n" - "cmp x10, #0x10\n" - "mov %x[res_ptr], x26\n" - "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" - "bge 1b\n" - "4:" // Row loop skip - "cbz x10, 9f\n" - "5:" // Row tail: Row loop - "add x24, %x[b_ptr], #0x8\n" - "mov x23, %x[nc]\n" - "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" - "6:" // Row tail: Column loop - "movi v2.16b, #0x0\n" - "movi v10.16b, #0x0\n" - "add x25, %x[a_ptr], #0x8\n" - "mov x21, %x[nb]\n" - "movi v12.16b, #0x0\n" - "movi v28.16b, #0x0\n" - "7:" // Row tail: Block loop - "ldr q6, [x24, #0x0]\n" - "ldr q5, [x24, #0x10]\n" - "movi v17.16b, #0x4\n" - "movi v8.4s, #0x0\n" - "ldr q4, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "movi v27.4s, #0x0\n" - "movi v0.4s, #0x0\n" - "ldr q31, [x24, #0x20]\n" - "ldr q14, [x24, #0x30]\n" - "movi v29.4s, #0x0\n" - "movi v22.16b, #0xf0\n" - "ldr q11, [x25, #0x20]\n" - "ldr q23, [x25, #0x30]\n" - "sshl v21.16b, v6.16b, v17.16b\n" - "sshl v16.16b, v5.16b, v17.16b\n" - "ldr q20, [x25, #0x40]\n" - "ldr q26, [x25, #0x50]\n" - "and v6.16b, v6.16b, v22.16b\n" - "and v5.16b, v5.16b, v22.16b\n" - "ldr q25, [x25, #0x60]\n" - "ldr q3, [x25, #0x70]\n" - "sshl v19.16b, v31.16b, v17.16b\n" - "sshl v18.16b, v14.16b, v17.16b\n" - "ldr d17, [x25, #-0x8]\n" - ".inst 0x4e95a488 // smmla v8.4s, v4.16b, v21.16b\n" - ".inst 0x4e90a49b // smmla v27.4s, v4.16b, v16.16b\n" - "and v31.16b, v31.16b, v22.16b\n" - ".inst 0x4e95a5a0 // smmla v0.4s, v13.16b, v21.16b\n" - ".inst 0x4e90a5bd // smmla v29.4s, v13.16b, v16.16b\n" - "and v14.16b, v14.16b, v22.16b\n" - "sub x20, x24, #0x8\n" - "ldr d16, [x20, #0x0]\n" - "subs x21, x21, #0x1\n" - "add x25, x25, #0x88\n" - "fcvtl v17.4s, v17.4h\n" - "add x24, x24, #0x48\n" - ".inst 0x4e93a568 // smmla v8.4s, v11.16b, v19.16b\n" - ".inst 0x4e92a57b // smmla v27.4s, v11.16b, v18.16b\n" - ".inst 0x4e93a6e0 // smmla v0.4s, v23.16b, v19.16b\n" - ".inst 0x4e92a6fd // smmla v29.4s, v23.16b, v18.16b\n" - "fcvtl v16.4s, v16.4h\n" - ".inst 0x4e86a688 // smmla v8.4s, v20.16b, v6.16b\n" - ".inst 0x4e85a69b // smmla v27.4s, v20.16b, v5.16b\n" - "fmul v23.4s, v16.4s, v17.s[0]\n" - "fmul v21.4s, v16.4s, v17.s[1]\n" - "fmul v1.4s, v16.4s, v17.s[2]\n" - "fmul v20.4s, v16.4s, v17.s[3]\n" - ".inst 0x4e86a740 // smmla v0.4s, v26.16b, v6.16b\n" - ".inst 0x4e85a75d // smmla v29.4s, v26.16b, v5.16b\n" - ".inst 0x4e9fa728 // smmla v8.4s, v25.16b, v31.16b\n" - ".inst 0x4e8ea73b // smmla v27.4s, v25.16b, v14.16b\n" - ".inst 0x4e9fa460 // smmla v0.4s, v3.16b, v31.16b\n" - ".inst 0x4e8ea47d // smmla v29.4s, v3.16b, v14.16b\n" - "uzp1 v19.2d, v8.2d, v27.2d\n" - "uzp2 v18.2d, v8.2d, v27.2d\n" - "scvtf v19.4s, v19.4s, #0x4\n" - "uzp1 v17.2d, v0.2d, v29.2d\n" - "uzp2 v16.2d, v0.2d, v29.2d\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "fmla v2.4s, v19.4s, v23.4s\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "scvtf v16.4s, v16.4s, #0x4\n" - "fmla v10.4s, v18.4s, v21.4s\n" - "fmla v12.4s, v17.4s, v1.4s\n" - "fmla v28.4s, v16.4s, v20.4s\n" - "bgt 7b\n" - "mov x20, %x[res_ptr]\n" - "cmp x10, #0x1\n" - "str q2, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x2\n" - "str q10, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x3\n" - "str q12, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "str q28, [x20, #0x0]\n" - "8:" // Row tail: Accumulator store skip - "subs x23, x23, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "bne 6b\n" - "subs x10, x10, #0x4\n" - "add %x[a_ptr], %x[a_ptr], x9\n" - "mov %x[res_ptr], x22\n" - "bgt 5b\n" - "9:" // Row tail: Row loop skip - : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) - : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" - ); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); -#else + __asm__ __volatile__( + "mov x10, %x[nr]\n" + "mov x9, #0x88\n" + "cmp x10, #0x10\n" + "mul x9, %x[nb], x9\n" + "blt 4f\n" + "1:" // Row loop + "add x28, %x[b_ptr], #0x8\n" + "mov x27, %x[nc]\n" + "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" + "2:" // Column loop + "add x25, %x[a_ptr], #0x8\n" + "movi v2.16b, #0x0\n" + "movi v10.16b, #0x0\n" + "mov x24, %x[nb]\n" + "add x23, x25, x9\n" + "movi v12.16b, #0x0\n" + "movi v28.16b, #0x0\n" + "add x22, x23, x9\n" + "movi v11.16b, #0x0\n" + "movi v13.16b, #0x0\n" + "add x21, x22, x9\n" + "movi v22.16b, #0x0\n" + "movi v23.16b, #0x0\n" + "movi v25.16b, #0x0\n" + "movi v5.16b, #0x0\n" + "movi v7.16b, #0x0\n" + "movi v4.16b, #0x0\n" + "movi v6.16b, #0x0\n" + "movi v30.16b, #0x0\n" + "movi v24.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "3:" // Block loop + "ldr q21, [x28, #0x0]\n" + "ldr q16, [x28, #0x10]\n" + "movi v1.16b, #0x4\n" + "movi v19.4s, #0x0\n" + "ldr q27, [x25, #0x0]\n" + "ldr q15, [x25, #0x10]\n" + "movi v26.4s, #0x0\n" + "movi v18.4s, #0x0\n" + "ldr q29, [x28, #0x20]\n" + "ldr q3, [x28, #0x30]\n" + "movi v17.4s, #0x0\n" + "movi v0.16b, #0xf0\n" + "ldr d20, [x25, #-0x8]\n" + "ldr d9, [x23, #-0x8]\n" + "sshl v8.16b, v21.16b, v1.16b\n" + "sshl v31.16b, v16.16b, v1.16b\n" + "and v21.16b, v21.16b, v0.16b\n" + "and v16.16b, v16.16b, v0.16b\n" + "sub x20, x28, #0x8\n" + "subs x24, x24, #0x1\n" + "add x28, x28, #0x48\n" + ".inst 0x4e88a773 // smmla v19.4s, v27.16b, v8.16b\n" + ".inst 0x4e9fa77a // smmla v26.4s, v27.16b, v31.16b\n" + "ldr q27, [x25, #0x20]\n" + ".inst 0x4e88a5f2 // smmla v18.4s, v15.16b, v8.16b\n" + ".inst 0x4e9fa5f1 // smmla v17.4s, v15.16b, v31.16b\n" + "sshl v15.16b, v29.16b, v1.16b\n" + "sshl v1.16b, v3.16b, v1.16b\n" + "and v29.16b, v29.16b, v0.16b\n" + "and v3.16b, v3.16b, v0.16b\n" + "ldr q0, [x25, #0x30]\n" + "fcvtl v20.4s, v20.4h\n" + ".inst 0x4e8fa773 // smmla v19.4s, v27.16b, v15.16b\n" + "fcvtl v9.4s, v9.4h\n" + ".inst 0x4e81a77a // smmla v26.4s, v27.16b, v1.16b\n" + "ldr q27, [x25, #0x40]\n" + ".inst 0x4e8fa412 // smmla v18.4s, v0.16b, v15.16b\n" + ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n" + "ldr q0, [x25, #0x50]\n" + ".inst 0x4e95a773 // smmla v19.4s, v27.16b, v21.16b\n" + ".inst 0x4e90a77a // smmla v26.4s, v27.16b, v16.16b\n" + "ldr q27, [x25, #0x60]\n" + ".inst 0x4e95a412 // smmla v18.4s, v0.16b, v21.16b\n" + ".inst 0x4e90a411 // smmla v17.4s, v0.16b, v16.16b\n" + "ldr q0, [x25, #0x70]\n" + "add x25, x25, #0x88\n" + ".inst 0x4e9da773 // smmla v19.4s, v27.16b, v29.16b\n" + ".inst 0x4e83a77a // smmla v26.4s, v27.16b, v3.16b\n" + "ldr d27, [x20, #0x0]\n" + ".inst 0x4e9da412 // smmla v18.4s, v0.16b, v29.16b\n" + ".inst 0x4e83a411 // smmla v17.4s, v0.16b, v3.16b\n" + "fcvtl v27.4s, v27.4h\n" + "uzp1 v0.2d, v19.2d, v26.2d\n" + "uzp2 v26.2d, v19.2d, v26.2d\n" + "fmul v19.4s, v27.4s, v20.s[0]\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "scvtf v26.4s, v26.4s, #0x4\n" + "fmla v2.4s, v0.4s, v19.4s\n" + "ldr q19, [x23, #0x0]\n" + "uzp1 v0.2d, v18.2d, v17.2d\n" + "uzp2 v18.2d, v18.2d, v17.2d\n" + "fmul v17.4s, v27.4s, v20.s[1]\n" + "scvtf v0.4s, v0.4s, #0x4\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "fmla v10.4s, v26.4s, v17.4s\n" + "ldr q17, [x23, #0x10]\n" + "fmul v26.4s, v27.4s, v20.s[2]\n" + "fmul v20.4s, v27.4s, v20.s[3]\n" + "fmla v12.4s, v0.4s, v26.4s\n" + "ldr d0, [x22, #-0x8]\n" + "ldr d26, [x21, #-0x8]\n" + "fcvtl v0.4s, v0.4h\n" + "fmla v28.4s, v18.4s, v20.4s\n" + "movi v20.4s, #0x0\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" + ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" + "ldr q19, [x23, #0x20]\n" + "fcvtl v26.4s, v26.4h\n" + ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" + ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" + "ldr q19, [x23, #0x40]\n" + ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" + ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" + "ldr q19, [x23, #0x60]\n" + ".inst 0x4e9da674 // smmla v20.4s, v19.16b, v29.16b\n" + ".inst 0x4e83a672 // smmla v18.4s, v19.16b, v3.16b\n" + "uzp1 v19.2d, v20.2d, v18.2d\n" + "scvtf v19.4s, v19.4s, #0x4\n" + "uzp2 v20.2d, v20.2d, v18.2d\n" + "fmul v18.4s, v27.4s, v9.s[0]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v11.4s, v19.4s, v18.4s\n" + "ldr q18, [x22, #0x0]\n" + "fmul v19.4s, v27.4s, v9.s[1]\n" + "fmla v13.4s, v20.4s, v19.4s\n" + "movi v19.4s, #0x0\n" + "movi v20.4s, #0x0\n" + ".inst 0x4e88a633 // smmla v19.4s, v17.16b, v8.16b\n" + ".inst 0x4e9fa634 // smmla v20.4s, v17.16b, v31.16b\n" + "ldr q17, [x23, #0x30]\n" + ".inst 0x4e8fa633 // smmla v19.4s, v17.16b, v15.16b\n" + ".inst 0x4e81a634 // smmla v20.4s, v17.16b, v1.16b\n" + "ldr q17, [x23, #0x50]\n" + ".inst 0x4e95a633 // smmla v19.4s, v17.16b, v21.16b\n" + ".inst 0x4e90a634 // smmla v20.4s, v17.16b, v16.16b\n" + "ldr q17, [x23, #0x70]\n" + "add x23, x23, #0x88\n" + ".inst 0x4e9da633 // smmla v19.4s, v17.16b, v29.16b\n" + ".inst 0x4e83a634 // smmla v20.4s, v17.16b, v3.16b\n" + "uzp1 v17.2d, v19.2d, v20.2d\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "uzp2 v20.2d, v19.2d, v20.2d\n" + "fmul v19.4s, v27.4s, v9.s[2]\n" + "fmul v9.4s, v27.4s, v9.s[3]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v22.4s, v17.4s, v19.4s\n" + "ldr q17, [x22, #0x10]\n" + "movi v19.4s, #0x0\n" + ".inst 0x4e88a653 // smmla v19.4s, v18.16b, v8.16b\n" + "fmla v23.4s, v20.4s, v9.4s\n" + "movi v20.4s, #0x0\n" + "movi v9.4s, #0x0\n" + ".inst 0x4e9fa654 // smmla v20.4s, v18.16b, v31.16b\n" + "ldr q18, [x22, #0x20]\n" + ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" + ".inst 0x4e8fa653 // smmla v19.4s, v18.16b, v15.16b\n" + ".inst 0x4e81a654 // smmla v20.4s, v18.16b, v1.16b\n" + "ldr q18, [x22, #0x40]\n" + ".inst 0x4e95a653 // smmla v19.4s, v18.16b, v21.16b\n" + ".inst 0x4e90a654 // smmla v20.4s, v18.16b, v16.16b\n" + "ldr q18, [x22, #0x60]\n" + ".inst 0x4e9da653 // smmla v19.4s, v18.16b, v29.16b\n" + ".inst 0x4e83a654 // smmla v20.4s, v18.16b, v3.16b\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e9fa632 // smmla v18.4s, v17.16b, v31.16b\n" + "ldr q17, [x22, #0x30]\n" + ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" + ".inst 0x4e81a632 // smmla v18.4s, v17.16b, v1.16b\n" + "ldr q17, [x22, #0x50]\n" + ".inst 0x4e95a629 // smmla v9.4s, v17.16b, v21.16b\n" + ".inst 0x4e90a632 // smmla v18.4s, v17.16b, v16.16b\n" + "ldr q17, [x22, #0x70]\n" + "add x22, x22, #0x88\n" + ".inst 0x4e9da629 // smmla v9.4s, v17.16b, v29.16b\n" + ".inst 0x4e83a632 // smmla v18.4s, v17.16b, v3.16b\n" + "uzp1 v17.2d, v19.2d, v20.2d\n" + "uzp2 v20.2d, v19.2d, v20.2d\n" + "fmul v19.4s, v27.4s, v0.s[0]\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "fmla v25.4s, v17.4s, v19.4s\n" + "ldr q19, [x21, #0x0]\n" + "fmul v17.4s, v27.4s, v0.s[1]\n" + "fmla v5.4s, v20.4s, v17.4s\n" + "ldr q17, [x21, #0x10]\n" + "uzp1 v20.2d, v9.2d, v18.2d\n" + "uzp2 v9.2d, v9.2d, v18.2d\n" + "fmul v18.4s, v27.4s, v0.s[2]\n" + "fmul v0.4s, v27.4s, v0.s[3]\n" + "scvtf v20.4s, v20.4s, #0x4\n" + "scvtf v9.4s, v9.4s, #0x4\n" + "fmla v7.4s, v20.4s, v18.4s\n" + "movi v20.4s, #0x0\n" + "movi v18.4s, #0x0\n" + ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" + ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" + "ldr q19, [x21, #0x20]\n" + "fmla v4.4s, v9.4s, v0.4s\n" + "movi v9.4s, #0x0\n" + "movi v0.4s, #0x0\n" + ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" + "fmul v8.4s, v27.4s, v26.s[0]\n" + ".inst 0x4e9fa620 // smmla v0.4s, v17.16b, v31.16b\n" + "ldr q17, [x21, #0x30]\n" + ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" + "fmul v31.4s, v27.4s, v26.s[1]\n" + ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" + "ldr q19, [x21, #0x40]\n" + ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" + "fmul v15.4s, v27.4s, v26.s[2]\n" + "fmul v27.4s, v27.4s, v26.s[3]\n" + ".inst 0x4e81a620 // smmla v0.4s, v17.16b, v1.16b\n" + "ldr q1, [x21, #0x50]\n" + ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" + ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" + "ldr q26, [x21, #0x60]\n" + ".inst 0x4e95a429 // smmla v9.4s, v1.16b, v21.16b\n" + ".inst 0x4e90a420 // smmla v0.4s, v1.16b, v16.16b\n" + "ldr q21, [x21, #0x70]\n" + "add x21, x21, #0x88\n" + ".inst 0x4e9da754 // smmla v20.4s, v26.16b, v29.16b\n" + ".inst 0x4e83a752 // smmla v18.4s, v26.16b, v3.16b\n" + ".inst 0x4e9da6a9 // smmla v9.4s, v21.16b, v29.16b\n" + ".inst 0x4e83a6a0 // smmla v0.4s, v21.16b, v3.16b\n" + "uzp1 v29.2d, v20.2d, v18.2d\n" + "uzp2 v21.2d, v20.2d, v18.2d\n" + "scvtf v29.4s, v29.4s, #0x4\n" + "uzp1 v18.2d, v9.2d, v0.2d\n" + "uzp2 v16.2d, v9.2d, v0.2d\n" + "scvtf v21.4s, v21.4s, #0x4\n" + "fmla v6.4s, v29.4s, v8.4s\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "scvtf v16.4s, v16.4s, #0x4\n" + "fmla v30.4s, v21.4s, v31.4s\n" + "fmla v24.4s, v18.4s, v15.4s\n" + "fmla v14.4s, v16.4s, v27.4s\n" + "bgt 3b\n" + "mov x20, %x[res_ptr]\n" + "subs x27, x27, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "str q2, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q10, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q12, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q28, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q11, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q13, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q22, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q23, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q25, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q5, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q7, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q4, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q6, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q30, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q24, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "str q14, [x20, #0x0]\n" + "bne 2b\n" + "mov x20, #0x4\n" + "sub x10, x10, #0x10\n" + "cmp x10, #0x10\n" + "mov %x[res_ptr], x26\n" + "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" + "bge 1b\n" + "4:" // Row loop skip + "cbz x10, 9f\n" + "5:" // Row tail: Row loop + "add x24, %x[b_ptr], #0x8\n" + "mov x23, %x[nc]\n" + "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" + "6:" // Row tail: Column loop + "movi v2.16b, #0x0\n" + "movi v10.16b, #0x0\n" + "add x25, %x[a_ptr], #0x8\n" + "mov x21, %x[nb]\n" + "movi v12.16b, #0x0\n" + "movi v28.16b, #0x0\n" + "7:" // Row tail: Block loop + "ldr q6, [x24, #0x0]\n" + "ldr q5, [x24, #0x10]\n" + "movi v17.16b, #0x4\n" + "movi v8.4s, #0x0\n" + "ldr q4, [x25, #0x0]\n" + "ldr q13, [x25, #0x10]\n" + "movi v27.4s, #0x0\n" + "movi v0.4s, #0x0\n" + "ldr q31, [x24, #0x20]\n" + "ldr q14, [x24, #0x30]\n" + "movi v29.4s, #0x0\n" + "movi v22.16b, #0xf0\n" + "ldr q11, [x25, #0x20]\n" + "ldr q23, [x25, #0x30]\n" + "sshl v21.16b, v6.16b, v17.16b\n" + "sshl v16.16b, v5.16b, v17.16b\n" + "ldr q20, [x25, #0x40]\n" + "ldr q26, [x25, #0x50]\n" + "and v6.16b, v6.16b, v22.16b\n" + "and v5.16b, v5.16b, v22.16b\n" + "ldr q25, [x25, #0x60]\n" + "ldr q3, [x25, #0x70]\n" + "sshl v19.16b, v31.16b, v17.16b\n" + "sshl v18.16b, v14.16b, v17.16b\n" + "ldr d17, [x25, #-0x8]\n" + ".inst 0x4e95a488 // smmla v8.4s, v4.16b, v21.16b\n" + ".inst 0x4e90a49b // smmla v27.4s, v4.16b, v16.16b\n" + "and v31.16b, v31.16b, v22.16b\n" + ".inst 0x4e95a5a0 // smmla v0.4s, v13.16b, v21.16b\n" + ".inst 0x4e90a5bd // smmla v29.4s, v13.16b, v16.16b\n" + "and v14.16b, v14.16b, v22.16b\n" + "sub x20, x24, #0x8\n" + "ldr d16, [x20, #0x0]\n" + "subs x21, x21, #0x1\n" + "add x25, x25, #0x88\n" + "fcvtl v17.4s, v17.4h\n" + "add x24, x24, #0x48\n" + ".inst 0x4e93a568 // smmla v8.4s, v11.16b, v19.16b\n" + ".inst 0x4e92a57b // smmla v27.4s, v11.16b, v18.16b\n" + ".inst 0x4e93a6e0 // smmla v0.4s, v23.16b, v19.16b\n" + ".inst 0x4e92a6fd // smmla v29.4s, v23.16b, v18.16b\n" + "fcvtl v16.4s, v16.4h\n" + ".inst 0x4e86a688 // smmla v8.4s, v20.16b, v6.16b\n" + ".inst 0x4e85a69b // smmla v27.4s, v20.16b, v5.16b\n" + "fmul v23.4s, v16.4s, v17.s[0]\n" + "fmul v21.4s, v16.4s, v17.s[1]\n" + "fmul v1.4s, v16.4s, v17.s[2]\n" + "fmul v20.4s, v16.4s, v17.s[3]\n" + ".inst 0x4e86a740 // smmla v0.4s, v26.16b, v6.16b\n" + ".inst 0x4e85a75d // smmla v29.4s, v26.16b, v5.16b\n" + ".inst 0x4e9fa728 // smmla v8.4s, v25.16b, v31.16b\n" + ".inst 0x4e8ea73b // smmla v27.4s, v25.16b, v14.16b\n" + ".inst 0x4e9fa460 // smmla v0.4s, v3.16b, v31.16b\n" + ".inst 0x4e8ea47d // smmla v29.4s, v3.16b, v14.16b\n" + "uzp1 v19.2d, v8.2d, v27.2d\n" + "uzp2 v18.2d, v8.2d, v27.2d\n" + "scvtf v19.4s, v19.4s, #0x4\n" + "uzp1 v17.2d, v0.2d, v29.2d\n" + "uzp2 v16.2d, v0.2d, v29.2d\n" + "scvtf v18.4s, v18.4s, #0x4\n" + "fmla v2.4s, v19.4s, v23.4s\n" + "scvtf v17.4s, v17.4s, #0x4\n" + "scvtf v16.4s, v16.4s, #0x4\n" + "fmla v10.4s, v18.4s, v21.4s\n" + "fmla v12.4s, v17.4s, v1.4s\n" + "fmla v28.4s, v16.4s, v20.4s\n" + "bgt 7b\n" + "mov x20, %x[res_ptr]\n" + "cmp x10, #0x1\n" + "str q2, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x2\n" + "str q10, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "cmp x10, #0x3\n" + "str q12, [x20, #0x0]\n" + "add x20, x20, %x[res_stride]\n" + "ble 8f\n" + "str q28, [x20, #0x0]\n" + "8:" // Row tail: Accumulator store skip + "subs x23, x23, #0x4\n" + "add %x[res_ptr], %x[res_ptr], #0x10\n" + "bne 6b\n" + "subs x10, x10, #0x4\n" + "add %x[a_ptr], %x[a_ptr], x9\n" + "mov %x[res_ptr], x22\n" + "bgt 5b\n" + "9:" // Row tail: Row loop skip + : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) + : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + ); + return; + } +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) float sumf[4][4]; int sumi; @@ -2012,7 +1983,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + - (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; } sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]); } @@ -2025,7 +1996,6 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * } } } -#endif } void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { @@ -2048,8 +2018,9 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * UNUSED(ncols_interleaved); UNUSED(blocklen); -#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - if (ggml_sve_cnt_b == QK8_0) { +#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) +#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) + if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && sve_lane_count() == QK8_0) { const void * b_ptr = vx; const void * a_ptr = vy; float * res_ptr = s; @@ -2459,154 +2430,329 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * ); return; } - else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { - GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) && - "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal " - "performance"); - } - else if (ggml_cpu_has_neon()) { - GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 " - "quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(ggml_cpu_has_sve() && - "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); +#endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) #elif defined(__AVX2__) || defined(__AVX512F__) - const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx; - const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy; - int64_t b_nb = n / QK4_0; - int64_t y = 0; - // Mask to mask out nibbles from packed bytes - const __m256i m4b = _mm256_set1_epi8(0x0F); - const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3); - // Lookup table to convert signed nibbles to signed bytes - __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); - signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); - // Permute mask used for easier vector processing at later stages - __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4); - int64_t xstart = 0; - int anr = nr - nr%16; // Used to align nr with boundary of 16 -#ifdef __AVX512F__ - int anc = nc - nc%16; // Used to align nc with boundary of 16 - // Mask to mask out nibbles from packed bytes expanded to 512 bit length - const __m512i m4bexpanded = _mm512_set1_epi8(0x0F); - // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length - __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1); + { + const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx; + const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy; + int64_t b_nb = n / QK4_0; + int64_t y = 0; + // Mask to mask out nibbles from packed bytes + const __m256i m4b = _mm256_set1_epi8(0x0F); + const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3); + // Lookup table to convert signed nibbles to signed bytes + __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); + // Permute mask used for easier vector processing at later stages + __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4); + int64_t xstart = 0; + int anr = nr - nr%16; // Used to align nr with boundary of 16 + #ifdef __AVX512F__ + int anc = nc - nc%16; // Used to align nc with boundary of 16 + // Mask to mask out nibbles from packed bytes expanded to 512 bit length + const __m512i m4bexpanded = _mm512_set1_epi8(0x0F); + // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length + __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1); - // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation - for (; y < anr / 4; y += 4) { + // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation + for (; y < anr / 4; y += 4) { - const block_q8_0x4 * a_ptrs[4]; + const block_q8_0x4 * a_ptrs[4]; - a_ptrs[0] = a_ptr_start + (y * nb); - for (int i = 0; i < 3; ++i) { - a_ptrs[i + 1] = a_ptrs[i] + nb; - } - - // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation - for (int64_t x = 0; x < anc / 8; x += 2) { - - const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); - const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); - - // Master FP accumulators - __m512 acc_rows[16]; - for (int i = 0; i < 16; i++) { - acc_rows[i] = _mm512_setzero_ps(); + a_ptrs[0] = a_ptr_start + (y * nb); + for (int i = 0; i < 3; ++i) { + a_ptrs[i + 1] = a_ptrs[i] + nb; } - for (int64_t b = 0; b < nb; b++) { - // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); + // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation + for (int64_t x = 0; x < anc / 8; x += 2) { - const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); - const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); - const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); - const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); + const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); + const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); - // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + // Master FP accumulators + __m512 acc_rows[16]; + for (int i = 0; i < 16; i++) { + acc_rows[i] = _mm512_setzero_ps(); + } - const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); - const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); + for (int64_t b = 0; b < nb; b++) { + // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); - const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); - const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); - const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); - const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); + const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); + const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); + const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); + const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); - // 4-bit -> 8-bit - Sign is maintained - const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) - const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) + // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); - const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) - const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) + const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); + const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); - const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) - const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) + const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); + const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); + const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); + const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); - const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) - const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) + // 4-bit -> 8-bit - Sign is maintained + const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) + const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) - // Shuffle pattern one - right side input - const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) - const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) + const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) + const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) - const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) - const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) + const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) + const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) - const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) - const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) + const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) + const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) - const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) - const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) + // Shuffle pattern one - right side input + const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) + const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) - // Shuffle pattern two - right side input + const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) + const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) - const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) - const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) + const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) + const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) - const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) - const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) + const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) + const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) - const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) - const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + // Shuffle pattern two - right side input - const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) - const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) + const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) - // Scale values - Load the weight scale values of two block_q4_0x8 - const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) + const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) - // Process LHS in pairs of rows - for (int rp = 0; rp < 4; rp++) { + const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) + const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + + const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) + const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + + // Scale values - Load the weight scale values of two block_q4_0x8 + const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + + // Process LHS in pairs of rows + for (int rp = 0; rp < 4; rp++) { + + // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector + __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); + __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0); + __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17); + __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); + __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0); + __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17); + __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); + __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0); + __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17); + __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); + __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0); + __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17); + + __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1); + __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1); + __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1); + __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1); + __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1); + __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1); + __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1); + __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1); + + // Shuffle pattern one - left side input + + const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) + const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + + const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) + const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + + const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) + const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + + const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) + const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + + // Shuffle pattern two - left side input + + const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) + const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + + const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) + const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + + const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) + const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + + const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) + const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + + // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane + // Resembles MMLAs into 2x2 matrices in ARM Version + __m512i iacc_mat_00_sp1 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1)); + __m512i iacc_mat_01_sp1 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1)); + __m512i iacc_mat_10_sp1 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1)); + __m512i iacc_mat_11_sp1 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1)); + __m512i iacc_mat_00_sp2 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2)); + __m512i iacc_mat_01_sp2 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2)); + __m512i iacc_mat_10_sp2 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2)); + __m512i iacc_mat_11_sp2 = + _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2)); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); + __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); + __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); + __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + + + // Straighten out to make 4 row vectors + __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78)); + __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01); + __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78)); + __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11); + + // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes + const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68); + const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16); + + // Multiply with appropiate scales and accumulate + acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); + acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); + acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); + acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); + } + } + + // Store the accumulated values + for (int i = 0; i < 16; i++) { + _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); + } + } + } + // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation + for (; y < nr / 4; y ++) { + + const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); + + // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation + for (int64_t x = 0; x < anc / 8; x += 2) { + + const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); + const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); + + // Master FP accumulators + __m512 acc_rows[4]; + for (int i = 0; i < 4; i++) { + acc_rows[i] = _mm512_setzero_ps(); + } + + for (int64_t b = 0; b < nb; b++) { + // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); + + const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); + const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); + const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); + const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); + + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + + const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); + const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); + + const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); + const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); + const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); + const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); + + // 4-bit -> 8-bit - Sign is maintained + const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) + const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) + + const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) + const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) + + const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) + const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) + + const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) + const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) + + // Shuffle pattern one - right side input + const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) + const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) + + const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) + const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) + + const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) + const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) + + const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) + const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) + + // Shuffle pattern two - right side input + + const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) + const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) + + const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) + const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) + + const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) + const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + + const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) + const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + + + // Scale values - Load the weight scale values of two block_q4_0x8 + const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector - __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); + __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0); __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17); - __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); + __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0); __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17); - __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); + __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0); __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17); - __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); + __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0); __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17); @@ -2680,314 +2826,286 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11); // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68); + const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68); const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16); // Multiply with appropiate scales and accumulate - acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); - acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); - acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); - acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); + acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); + acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); + acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); + acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); + } + + // Store the accumulated values + for (int i = 0; i < 4; i++) { + _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); } } + } + if (anc != nc) { + xstart = anc/8; + y = 0; + } + #endif // __AVX512F__ - // Store the accumulated values - for (int i = 0; i < 16; i++) { - _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); + // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation + + for (; y < anr / 4; y += 4) { + const block_q8_0x4 * a_ptrs[4]; + + a_ptrs[0] = a_ptr_start + (y * nb); + for (int i = 0; i < 3; ++i) { + a_ptrs[i + 1] = a_ptrs[i] + nb; + } + + // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation + for (int64_t x = xstart; x < nc / 8; x++) { + + const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb); + + // Master FP accumulators + __m256 acc_rows[16]; + for (int i = 0; i < 16; i++) { + acc_rows[i] = _mm256_setzero_ps(); + } + + for (int64_t b = 0; b < nb; b++) { + // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); + + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + + // 4-bit -> 8-bit - Sign is maintained + const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) + const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) + + const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) + const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) + + const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) + const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) + + const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) + const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) + + // Shuffle pattern one - right side input + const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) + const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) + + const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) + const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) + + const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) + const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) + + const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) + const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) + + // Shuffle pattern two - right side input + + const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) + const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) + + const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) + const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) + + const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) + const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) + + const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) + const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) + + // Scale values - Load the wight scale values of block_q4_0x8 + const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + + // Process LHS in groups of four + for (int rp = 0; rp < 4; rp++) { + // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated into a 256 bit vector + __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); + __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0); + __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17); + __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); + __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0); + __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17); + __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); + __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0); + __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17); + __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); + __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0); + __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17); + + // Shuffle pattern one - left side input + const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) + const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + + const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) + const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + + const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) + const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + + const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) + const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + + // Shuffle pattern two - left side input + const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) + const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + + const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) + const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + + const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) + const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + + const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) + const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + + // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane + // Resembles MMLAs into 2x2 matrices in ARM Version + __m256i iacc_mat_00_sp1 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1)); + __m256i iacc_mat_01_sp1 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1)); + __m256i iacc_mat_10_sp1 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1)); + __m256i iacc_mat_11_sp1 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1)); + __m256i iacc_mat_00_sp2 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2)); + __m256i iacc_mat_01_sp2 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2)); + __m256i iacc_mat_10_sp2 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2)); + __m256i iacc_mat_11_sp2 = + _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2)); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); + __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); + __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); + __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + + // Straighten out to make 4 row vectors + __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204); + __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204); + __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204); + __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204); + + // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes + const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask); + + // Multiply with appropiate scales and accumulate + acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); + acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); + acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); + acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); + } + } + + // Store the accumulated values + for (int i = 0; i < 16; i++) { + _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); + } } } - } - // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation - for (; y < nr / 4; y ++) { - const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); + // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation + for (; y < nr / 4; y ++) { - // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation - for (int64_t x = 0; x < anc / 8; x += 2) { + const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); - const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); - const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); + // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + for (int64_t x = xstart; x < nc / 8; x++) { - // Master FP accumulators - __m512 acc_rows[4]; - for (int i = 0; i < 4; i++) { - acc_rows[i] = _mm512_setzero_ps(); - } + const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb); - for (int64_t b = 0; b < nb; b++) { - // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); + // Master FP accumulators + __m256 acc_rows[4]; + for (int i = 0; i < 4; i++) { + acc_rows[i] = _mm256_setzero_ps(); + } - const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); - const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); - const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); - const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); + for (int64_t b = 0; b < nb; b++) { + // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); - // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); - const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); - const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); + // 4-bit -> 8-bit - Sign is maintained + const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) + const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) - const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); - const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); - const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); - const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); + const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) + const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) - // 4-bit -> 8-bit - Sign is maintained - const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) - const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) + const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) + const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) - const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) - const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) + const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) + const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) - const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) - const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) + // Shuffle pattern one - right side input + const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) + const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) - const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) - const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) + const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) + const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) - // Shuffle pattern one - right side input - const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) - const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) + const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) + const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) - const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) - const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) + const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) + const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) - const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) - const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) + // Shuffle pattern two - right side input - const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) - const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) + const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) + const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) - // Shuffle pattern two - right side input + const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) + const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) - const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) - const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) + const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) + const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) - const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) - const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) + const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) + const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) - const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) - const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + // Scale values - Load the wight scale values of block_q4_0x8 + const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); - const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) - const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) - - - // Scale values - Load the weight scale values of two block_q4_0x8 - const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); - - // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 - // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector - __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); - __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0); - __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17); - __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); - __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0); - __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17); - __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); - __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0); - __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17); - __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); - __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0); - __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17); - - __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1); - __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1); - __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1); - __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1); - __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1); - __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1); - __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1); - __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1); - - // Shuffle pattern one - left side input - - const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) - const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) - - const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) - const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) - - const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) - const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) - - const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) - const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) - - // Shuffle pattern two - left side input - - const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) - const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) - - const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) - const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) - - const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) - const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) - - const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) - const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) - - // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane - // Resembles MMLAs into 2x2 matrices in ARM Version - __m512i iacc_mat_00_sp1 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1)); - __m512i iacc_mat_01_sp1 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1)); - __m512i iacc_mat_10_sp1 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1)); - __m512i iacc_mat_11_sp1 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1)); - __m512i iacc_mat_00_sp2 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2)); - __m512i iacc_mat_01_sp2 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2)); - __m512i iacc_mat_10_sp2 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2)); - __m512i iacc_mat_11_sp2 = - _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2)); - - // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block - __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); - __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); - __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); - __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); - - - // Straighten out to make 4 row vectors - __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78)); - __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01); - __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78)); - __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11); - - // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68); - const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16); - - // Multiply with appropiate scales and accumulate - acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); - acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); - acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); - acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); - } - - // Store the accumulated values - for (int i = 0; i < 4; i++) { - _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); - } - } - } - if (anc != nc) { - xstart = anc/8; - y = 0; - } -#endif // __AVX512F__ - - // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation - - for (; y < anr / 4; y += 4) { - const block_q8_0x4 * a_ptrs[4]; - - a_ptrs[0] = a_ptr_start + (y * nb); - for (int i = 0; i < 3; ++i) { - a_ptrs[i + 1] = a_ptrs[i] + nb; - } - - // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation - for (int64_t x = xstart; x < nc / 8; x++) { - - const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb); - - // Master FP accumulators - __m256 acc_rows[16]; - for (int i = 0; i < 16; i++) { - acc_rows[i] = _mm256_setzero_ps(); - } - - for (int64_t b = 0; b < nb; b++) { - // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); - - // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); - - // 4-bit -> 8-bit - Sign is maintained - const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) - const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) - - const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) - const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) - - const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) - const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) - - const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) - const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) - - // Shuffle pattern one - right side input - const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) - const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) - - const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) - const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) - - const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) - const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) - - const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) - const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) - - // Shuffle pattern two - right side input - - const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) - const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) - - const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) - const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) - - const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) - const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) - - const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) - const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) - - // Scale values - Load the wight scale values of block_q4_0x8 - const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); - - // Process LHS in groups of four - for (int rp = 0; rp < 4; rp++) { // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 // Loaded as set of 128 bit vectors and repeated into a 256 bit vector - __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); + __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0); __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17); - __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); + __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0); __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17); - __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); + __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0); __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17); - __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); + __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0); __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17); // Shuffle pattern one - left side input + const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) @@ -3001,6 +3119,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) // Shuffle pattern two - left side input + const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) @@ -3038,6 +3157,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + // Straighten out to make 4 row vectors __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204); __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204); @@ -3045,187 +3165,24 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204); // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask); + const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask); // Multiply with appropiate scales and accumulate - acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); - acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); - acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); - acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); + acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); + acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); + acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); + acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); + } + + // Store the accumulated values + for (int i = 0; i < 4; i++) { + _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); } } - - // Store the accumulated values - for (int i = 0; i < 16; i++) { - _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); - } } + return; } - - // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation - for (; y < nr / 4; y ++) { - - const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); - - // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 - for (int64_t x = xstart; x < nc / 8; x++) { - - const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb); - - // Master FP accumulators - __m256 acc_rows[4]; - for (int i = 0; i < 4; i++) { - acc_rows[i] = _mm256_setzero_ps(); - } - - for (int64_t b = 0; b < nb; b++) { - // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); - - // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); - - // 4-bit -> 8-bit - Sign is maintained - const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) - const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) - - const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) - const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) - - const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) - const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) - - const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) - const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) - - // Shuffle pattern one - right side input - const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) - const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) - - const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) - const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) - - const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) - const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) - - const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) - const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) - - // Shuffle pattern two - right side input - - const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) - const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) - - const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) - const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) - - const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) - const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) - - const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) - const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) - - // Scale values - Load the wight scale values of block_q4_0x8 - const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); - - // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 - // Loaded as set of 128 bit vectors and repeated into a 256 bit vector - __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); - __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0); - __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17); - __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); - __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0); - __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17); - __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); - __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0); - __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17); - __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); - __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0); - __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17); - - // Shuffle pattern one - left side input - - const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) - const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) - - const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) - const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) - - const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) - const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) - - const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) - const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) - - // Shuffle pattern two - left side input - - const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) - const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) - - const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) - const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) - - const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) - const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) - - const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) - const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) - - // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane - // Resembles MMLAs into 2x2 matrices in ARM Version - __m256i iacc_mat_00_sp1 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1)); - __m256i iacc_mat_01_sp1 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1)); - __m256i iacc_mat_10_sp1 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1)); - __m256i iacc_mat_11_sp1 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1)); - __m256i iacc_mat_00_sp2 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2)); - __m256i iacc_mat_01_sp2 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2)); - __m256i iacc_mat_10_sp2 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2)); - __m256i iacc_mat_11_sp2 = - _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2)); - - // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block - __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); - __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); - __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); - __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); - - - // Straighten out to make 4 row vectors - __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204); - __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204); - __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204); - __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204); - - // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask); - - // Multiply with appropiate scales and accumulate - acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); - acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); - acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); - acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); - } - - // Store the accumulated values - for (int i = 0; i < 4; i++) { - _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); - } - } - } -#else +#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) float sumf[4][8]; int sumi; @@ -3258,5 +3215,4 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * } } } -#endif } From ea9c32be71b91b42ecc538bd902e93cbb5fb36cb Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Wed, 25 Sep 2024 17:26:01 +0200 Subject: [PATCH 5/5] ci : fix docker build number and tag name (#9638) * ci : fix docker build number and tag name * fine-grant permissions --- .dockerignore | 2 +- .github/workflows/docker.yml | 58 +++++++++++++++++++++++------------- 2 files changed, 39 insertions(+), 21 deletions(-) diff --git a/.dockerignore b/.dockerignore index 8916e2a66..064b7c7be 100644 --- a/.dockerignore +++ b/.dockerignore @@ -1,7 +1,7 @@ *.o *.a .cache/ -.git/ +# Do not ignore .git directory, otherwise the reported build number will always be 0 .github/ .gitignore .vs/ diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 9044cd78b..a4ac9b217 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -15,11 +15,17 @@ on: branches: - master paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal'] + workflow_dispatch: # allows manual triggering, useful for debugging concurrency: group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} cancel-in-progress: true +# Fine-grant permission +# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token +permissions: + packages: write + jobs: push_to_registry: name: Push Docker image to Docker Hub @@ -46,6 +52,8 @@ jobs: steps: - name: Check out the repo uses: actions/checkout@v4 + with: + fetch-depth: 0 # preserve git history, so we can determine the build number - name: Set up QEMU uses: docker/setup-qemu-action@v2 @@ -60,6 +68,34 @@ jobs: username: ${{ github.repository_owner }} password: ${{ secrets.GITHUB_TOKEN }} + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}" # to lower case + REPO_NAME="${{ github.event.repository.name }}" + + # determine tag name postfix (build number, commit hash) + if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then + TAG_POSTFIX="b${BUILD_NUMBER}" + else + SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-') + TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}" + fi + + # list all tags possible + TAGS="" + TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}," + TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}" + + echo "output_tags=$TAGS" >> $GITHUB_OUTPUT + echo "output_tags=$TAGS" # print out for debugging + env: + GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }} + GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}' + # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example - name: Free Disk Space (Ubuntu) uses: jlumbroso/free-disk-space@main @@ -77,25 +113,6 @@ jobs: docker-images: true swap-storage: true - - name: Determine tag name - id: tag - shell: bash - run: | - BUILD_NUMBER="$(git rev-list --count HEAD)" - SHORT_HASH="$(git rev-parse --short=7 HEAD)" - if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then - echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Downcase github.repository_owner - run: | - echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV - env: - GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}' - - name: Build and push Docker image (tagged + versioned) if: github.event_name == 'push' uses: docker/build-push-action@v6 @@ -103,5 +120,6 @@ jobs: context: . push: true platforms: ${{ matrix.config.platforms }} - tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}" + # tag list is generated from step above + tags: ${{ steps.tag.outputs.output_tags }} file: ${{ matrix.config.dockerfile }}