From d67341dc18fc5cc63362880ab2f8f9ecfc7932e7 Mon Sep 17 00:00:00 2001 From: aa956 Date: Thu, 19 Jun 2025 16:01:03 +0300 Subject: [PATCH 1/7] server : add server parameters for draft model cache type (#13782) Co-authored-by: aa956 <27946957+aa956@users.noreply.github.com> --- common/arg.cpp | 26 ++++++++++++++++++++++++++ common/common.h | 3 +++ tools/server/README.md | 2 ++ tools/server/server.cpp | 6 ++---- 4 files changed, 33 insertions(+), 4 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index 231de227a..3dfaa71ef 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -3210,6 +3210,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.speculative.model.path = value; } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT")); + add_opt(common_arg( + {"-ctkd", "--cache-type-k-draft"}, "TYPE", + string_format( + "KV cache data type for K for the draft model\n" + "allowed values: %s\n" + "(default: %s)", + get_all_kv_cache_types().c_str(), + ggml_type_name(params.speculative.cache_type_k) + ), + [](common_params & params, const std::string & value) { + params.speculative.cache_type_k = kv_cache_type_from_str(value); + } + ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT")); + add_opt(common_arg( + {"-ctvd", "--cache-type-v-draft"}, "TYPE", + string_format( + "KV cache data type for V for the draft model\n" + "allowed values: %s\n" + "(default: %s)", + get_all_kv_cache_types().c_str(), + ggml_type_name(params.speculative.cache_type_v) + ), + [](common_params & params, const std::string & value) { + params.speculative.cache_type_v = kv_cache_type_from_str(value); + } + ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT")); add_opt(common_arg( {"-mv", "--model-vocoder"}, "FNAME", diff --git a/common/common.h b/common/common.h index 00b6ca03a..5710c4e97 100644 --- a/common/common.h +++ b/common/common.h @@ -199,6 +199,9 @@ struct common_params_speculative { float p_split = 0.1f; // speculative decoding split probability float p_min = 0.75f; // minimum speculative decoding probability (greedy) + ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K + ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V + struct cpu_params cpuparams; struct cpu_params cpuparams_batch; diff --git a/tools/server/README.md b/tools/server/README.md index 06533c172..43aa65d50 100644 --- a/tools/server/README.md +++ b/tools/server/README.md @@ -187,6 +187,8 @@ The project is under active development, and we are [looking for feedback and co | `-devd, --device-draft ` | comma-separated list of devices to use for offloading the draft model (none = don't offload)
use --list-devices to see a list of available devices | | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model
(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) | | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)
(env: LLAMA_ARG_MODEL_DRAFT) | +| `-ctkd, --cache-type-k-draft TYPE` | KV cache data type for K for speculative decoding model
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(default: f16)
(env: LLAMA_ARG_CACHE_TYPE_K_DRAFT) | +| `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for speculative decoding model
allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1
(default: f16)
(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) | | `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) | | `--tts-use-guide-tokens` | Use guide tokens to improve TTS word recall | | `--embd-bge-small-en-default` | use default bge-small-en-v1.5 model (note: can download weights from the internet) | diff --git a/tools/server/server.cpp b/tools/server/server.cpp index 721d09182..9d55b3338 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -1969,10 +1969,8 @@ struct server_context { params_dft.n_ctx = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx; params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers; params_dft.n_parallel = 1; - - // force F16 KV cache for the draft model for extra performance - params_dft.cache_type_k = GGML_TYPE_F16; - params_dft.cache_type_v = GGML_TYPE_F16; + params_dft.cache_type_k = params_base.speculative.cache_type_k; + params_dft.cache_type_v = params_base.speculative.cache_type_v; llama_init_dft = common_init_from_params(params_dft); From 381174bbdaf10d6a80dc2099f284b20544d86962 Mon Sep 17 00:00:00 2001 From: Alex Trotta <44127594+Ahajha@users.noreply.github.com> Date: Thu, 19 Jun 2025 09:56:12 -0400 Subject: [PATCH 2/7] gguf-py : make sentencepiece optional (#14200) * Make sentencepiece optional * Bump to 0.18.0 * Bump patch instead of minor Co-authored-by: compilade --------- Co-authored-by: compilade --- gguf-py/gguf/vocab.py | 8 +++++++- gguf-py/pyproject.toml | 4 ++-- 2 files changed, 9 insertions(+), 3 deletions(-) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index cca097986..44d066ee7 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -7,7 +7,10 @@ import os from pathlib import Path from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable -from sentencepiece import SentencePieceProcessor +try: + from sentencepiece import SentencePieceProcessor +except ImportError: + SentencePieceProcessor = None import gguf @@ -302,6 +305,9 @@ class SentencePieceVocab(Vocab): name = "spm" def __init__(self, base_path: Path): + if SentencePieceProcessor is None: + raise RuntimeError("sentencepiece is not installed") + added_tokens: dict[str, int] = {} if (fname_tokenizer := base_path / 'tokenizer.model').exists(): # normal location diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index f11351cba..0f3a1eeee 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.17.0" +version = "0.17.1" description = "Read and write ML models in GGUF for GGML" authors = ["GGML "] packages = [ @@ -22,7 +22,7 @@ python = ">=3.8" numpy = ">=1.17" tqdm = ">=4.27" pyyaml = ">=5.1" -sentencepiece = ">=0.1.98,<=0.2.0" +sentencepiece = { version = ">=0.1.98,<=0.2.0", optional = true } PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true } [tool.poetry.dev-dependencies] From 8f71d0f3e86ccbba059350058af8758cafed73e6 Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Thu, 19 Jun 2025 12:24:14 -0700 Subject: [PATCH 3/7] ggml-cpu : remove unnecesary arm feature detection (#14281) Support for Arm runtime feature detection has now been added to GGML_CPU_ALL_VARIANTS. This removes the old and not very functional code. --- ggml/src/ggml-cpu/arch/arm/repack.cpp | 2002 ++++++++++++------------- ggml/src/ggml-cpu/ggml-cpu.c | 95 +- 2 files changed, 1004 insertions(+), 1093 deletions(-) diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp index 9337e01b6..39a0dd301 100644 --- a/ggml/src/ggml-cpu/arch/arm/repack.cpp +++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp @@ -256,45 +256,43 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; - for (int c = 0; c < nc; c += ncols_interleaved) { - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - float32x4_t acc = vdupq_n_f32(0); - for (int b = 0; b < nb; b++) { - int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); - int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); - int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); - int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); - float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); + for (int c = 0; c < nc; c += ncols_interleaved) { + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float32x4_t acc = vdupq_n_f32(0); + for (int b = 0; b < nb; b++) { + int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); + int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); + int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); + int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); + float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); - int8x16_t a0 = vld1q_s8(a_ptr->qs); - int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2); - float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); + int8x16_t a0 = vld1q_s8(a_ptr->qs); + int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2); + float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); - int32x4_t ret = vdupq_n_s32(0); + int32x4_t ret = vdupq_n_s32(0); - ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0); - ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1); - ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2); - ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3); + ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0); + ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1); + ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2); + ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3); - ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0); - ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1); - ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2); - ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3); + ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0); + ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1); + ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2); + ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3); - acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), - vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); - a_ptr++; - b_ptr++; - } - vst1q_f32(s, acc); - s += ncols_interleaved; + acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), + vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); + a_ptr++; + b_ptr++; } - return; + vst1q_f32(s, acc); + s += ncols_interleaved; } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) float sumf[4]; int sumi; @@ -341,50 +339,48 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; + const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx; - for (int c = 0; c < nc; c += ncols_interleaved) { - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - float32x4_t acc = vdupq_n_f32(0); - for (int b = 0; b < nb; b++) { - int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); - int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); - int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); - int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); - float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); + for (int c = 0; c < nc; c += ncols_interleaved) { + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float32x4_t acc = vdupq_n_f32(0); + for (int b = 0; b < nb; b++) { + int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs); + int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16); + int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32); + int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48); + float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d); - int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs); - int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1); - int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2); - int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3); - float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); + int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs); + int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1); + int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2); + int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3); + float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d); - int32x4_t ret0 = vdupq_n_s32(0); - int32x4_t ret1 = vdupq_n_s32(0); + int32x4_t ret0 = vdupq_n_s32(0); + int32x4_t ret1 = vdupq_n_s32(0); - ret0 = vdotq_s32(ret0, b0 << 4, a0); - ret1 = vdotq_s32(ret1, b1 << 4, a0); - ret0 = vdotq_s32(ret0, b2 << 4, a1); - ret1 = vdotq_s32(ret1, b3 << 4, a1); + ret0 = vdotq_s32(ret0, b0 << 4, a0); + ret1 = vdotq_s32(ret1, b1 << 4, a0); + ret0 = vdotq_s32(ret0, b2 << 4, a1); + ret1 = vdotq_s32(ret1, b3 << 4, a1); - ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2); - ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2); - ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3); - ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3); + ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2); + ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2); + ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3); + ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3); - int32x4_t ret = vpaddq_s32(ret0, ret1); + int32x4_t ret = vpaddq_s32(ret0, ret1); - acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), - vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); - a_ptr++; - b_ptr++; - } - vst1q_f32(s, acc); - s += ncols_interleaved; + acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4), + vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd))); + a_ptr++; + b_ptr++; } - return; + vst1q_f32(s, acc); + s += ncols_interleaved; } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) float sumf[4]; int sumi; @@ -432,7 +428,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) #if defined(__ARM_FEATURE_SVE) - if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) { + if (ggml_cpu_get_sve_cnt() == QK8_0) { const void * b_ptr = vx; const void * a_ptr = vy; float * res_ptr = s; @@ -547,54 +543,52 @@ void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - float * res_ptr = s; + const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + float * res_ptr = s; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); - float32x4_t sumf = vdupq_n_f32(0); - for (int l = 0; l < nb; l++) { - uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0); - uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16); - uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32); - uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48); + float32x4_t sumf = vdupq_n_f32(0); + for (int l = 0; l < nb; l++) { + uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0); + uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16); + uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32); + uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48); - int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4); - int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F); - int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4); - int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F); - int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4); - int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F); - int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4); - int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F); + int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4); + int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F); + int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4); + int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F); + int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4); + int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F); + int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4); + int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F); - int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0); - int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16); + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16); - int32x4_t sumi = vdupq_n_s32(0); - sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0); - sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0); - sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1); - sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1); - sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2); - sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2); - sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3); - sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3); + int32x4_t sumi = vdupq_n_s32(0); + sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0); + sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0); + sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1); + sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1); + sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2); + sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2); + sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3); + sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3); - float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d)); - float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); - float32x4_t d = a_d * b_d; + float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d)); + float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + float32x4_t d = a_d * b_d; - sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi)); - } - - vst1q_f32(res_ptr + x * 4, sumf); + sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi)); } - return; + + vst1q_f32(res_ptr + x * 4, sumf); } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) { float sumf[4]; @@ -643,465 +637,463 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(ncols_interleaved); UNUSED(blocklen); -#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - 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_DOTPROD) + 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; - } + __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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) { float sumf[4][4]; @@ -1160,404 +1152,402 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo UNUSED(blocklen); #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); + 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" - ); - return; - } + __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; @@ -1615,7 +1605,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo #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() && ggml_cpu_get_sve_cnt() == QK8_0) { + if (ggml_cpu_get_sve_cnt() == QK8_0) { const void * b_ptr = vx; const void * a_ptr = vy; float * res_ptr = s; @@ -2083,59 +2073,57 @@ void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const UNUSED(blocklen); #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD) - if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { - const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); + const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl); - 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_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + 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_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); - float32x4_t sumf[4]; - for (int m = 0; m < 4; m++) { - sumf[m] = vdupq_n_f32(0); + float32x4_t sumf[4]; + for (int m = 0; m < 4; m++) { + sumf[m] = vdupq_n_f32(0); + } + + for (int l = 0; l < nb; l++) { + float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d)); + float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + + int32x4_t sumi_0 = vdupq_n_s32(0); + int32x4_t sumi_1 = vdupq_n_s32(0); + int32x4_t sumi_2 = vdupq_n_s32(0); + int32x4_t sumi_3 = vdupq_n_s32(0); + + for (int k = 0; k < 4; k++) { + int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0); + int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64); + + uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k); + int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4); + int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF); + + sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3); + sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0); + sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1); + sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2); + sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3); } - for (int l = 0; l < nb; l++) { - float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d)); - float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d)); + sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0)); + sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1)); + sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2)); + sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3)); + } - int32x4_t sumi_0 = vdupq_n_s32(0); - int32x4_t sumi_1 = vdupq_n_s32(0); - int32x4_t sumi_2 = vdupq_n_s32(0); - int32x4_t sumi_3 = vdupq_n_s32(0); - - for (int k = 0; k < 4; k++) { - int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0); - int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64); - - uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k); - int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4); - int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF); - - sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0); - sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1); - sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2); - sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3); - sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0); - sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1); - sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2); - sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3); - } - - sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0)); - sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1)); - sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2)); - sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3)); - } - - for (int m = 0; m < 4; m++) { - vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]); - } + for (int m = 0; m < 4; m++) { + vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]); } } - return; } + return; #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) { float sumf[4][4]; diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index 2c12e493b..1bb9c4e36 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -74,13 +74,8 @@ #if defined(__ARM_ARCH) struct ggml_arm_arch_features_type { - int has_neon; - int has_dotprod; - int has_i8mm; - int has_sve; int sve_cnt; - int has_sme; -} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1}; +} ggml_arm_arch_features = { 0 }; #endif @@ -678,87 +673,15 @@ bool ggml_is_numa(void) { #if defined(__linux__) && defined(__aarch64__) #include -#elif defined(__APPLE__) -#include -#endif - -#if !defined(HWCAP2_I8MM) -#define HWCAP2_I8MM (1 << 13) -#endif - -#if !defined(HWCAP2_SME) -#define HWCAP2_SME (1 << 23) #endif static void ggml_init_arm_arch_features(void) { -#if defined(__linux__) && defined(__aarch64__) - uint32_t hwcap = getauxval(AT_HWCAP); - uint32_t hwcap2 = getauxval(AT_HWCAP2); - - ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD); - ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP); - ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM); - ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE); - ggml_arm_arch_features.has_sme = !!(hwcap2 & HWCAP2_SME); - -#if defined(__ARM_FEATURE_SVE) +#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE) ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL); #endif -#elif defined(__APPLE__) - int oldp = 0; - size_t size = sizeof(oldp); - if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_neon = oldp; - - if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_dotprod = oldp; - - if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_i8mm = oldp; - - if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) { - oldp = 0; - } - ggml_arm_arch_features.has_sme = oldp; - - ggml_arm_arch_features.has_sve = 0; - ggml_arm_arch_features.sve_cnt = 0; -#else -// Run-time CPU feature detection not implemented for this platform, fallback to compile time -#if defined(__ARM_NEON) - ggml_arm_arch_features.has_neon = 1; -#else - ggml_arm_arch_features.has_neon = 0; -#endif - -#if defined(__ARM_FEATURE_MATMUL_INT8) - ggml_arm_arch_features.has_i8mm = 1; -#else - ggml_arm_arch_features.has_i8mm = 0; -#endif - -#if defined(__ARM_FEATURE_SVE) - ggml_arm_arch_features.has_sve = 1; - ggml_arm_arch_features.sve_cnt = 16; -#else - ggml_arm_arch_features.has_sve = 0; - ggml_arm_arch_features.sve_cnt = 0; -#endif - -#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2) - ggml_arm_arch_features.has_sme = 1; -#else - ggml_arm_arch_features.has_sme = 0; -#endif -#endif } -#endif + +#endif // __ARM_ARCH struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) { GGML_ASSERT(!ggml_get_no_alloc(ctx)); @@ -3443,7 +3366,7 @@ int ggml_cpu_has_vxe(void) { int ggml_cpu_has_neon(void) { #if defined(__ARM_ARCH) && defined(__ARM_NEON) - return ggml_arm_arch_features.has_neon; + return 1; #else return 0; #endif @@ -3451,7 +3374,7 @@ int ggml_cpu_has_neon(void) { int ggml_cpu_has_dotprod(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD) - return ggml_arm_arch_features.has_dotprod; + return 1; #else return 0; #endif @@ -3459,7 +3382,7 @@ int ggml_cpu_has_dotprod(void) { int ggml_cpu_has_sve(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE) - return ggml_arm_arch_features.has_sve; + return 1; #else return 0; #endif @@ -3467,7 +3390,7 @@ int ggml_cpu_has_sve(void) { int ggml_cpu_has_matmul_int8(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8) - return ggml_arm_arch_features.has_i8mm; + return 1; #else return 0; #endif @@ -3483,7 +3406,7 @@ int ggml_cpu_get_sve_cnt(void) { int ggml_cpu_has_sme(void) { #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME) - return ggml_arm_arch_features.has_sme; + return 1; #else return 0; #endif From 9eaa51e7f08593f123f00136591179a8f5956ecd Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Fri, 20 Jun 2025 09:50:24 +0800 Subject: [PATCH 4/7] CUDA: add conv_2d_dw (#14265) * CUDA: add conv_2d_dw * better naming * simplify using template * Review: fix operation ordering in ggml-cuda, use __forceinline__, use more const --- ggml/src/ggml-cuda/conv2d-dw.cu | 161 +++++++++++++++++++++++++++++++ ggml/src/ggml-cuda/conv2d-dw.cuh | 5 + ggml/src/ggml-cuda/ggml-cuda.cu | 5 + 3 files changed, 171 insertions(+) create mode 100644 ggml/src/ggml-cuda/conv2d-dw.cu create mode 100644 ggml/src/ggml-cuda/conv2d-dw.cuh diff --git a/ggml/src/ggml-cuda/conv2d-dw.cu b/ggml/src/ggml-cuda/conv2d-dw.cu new file mode 100644 index 000000000..7583233b1 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-dw.cu @@ -0,0 +1,161 @@ +#include "conv2d-dw.cuh" + +struct conv_params { + int in_w, in_h; + int out_w, out_h; + int kernel_w, kernel_h; + int stride_x, stride_y; + int padding_x, padding_y; + int dilation_x, dilation_y; + int channels, batches; +}; + +struct kernel_bounds { + int y_min, y_max; + int x_min, x_max; +}; + +__device__ __forceinline__ kernel_bounds calculate_kernel_bounds(int out_x, int out_y, const conv_params & params) { + kernel_bounds bounds; + bounds.y_min = max(0, (params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y); + bounds.y_max = + min(params.kernel_h, + (params.in_h + params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y); + bounds.x_min = max(0, (params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x); + bounds.x_max = + min(params.kernel_w, + (params.in_w + params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x); + return bounds; +} + +__device__ __forceinline__ int calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) { + return out_coord * stride + kern_coord * dilation - padding; +} + +struct whcn_layout { + __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.in_w * params.in_h) + c * params.in_w * params.in_h + y * params.in_w + x; + } + + __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) { + return c * params.kernel_h * params.kernel_w + ky * params.kernel_w + kx; + } + + __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.out_w * params.out_h) + c * params.out_w * params.out_h + + y * params.out_w + x; + } + + __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y, + int & out_x) { + out_x = global_idx % params.out_w; + out_y = (global_idx / params.out_w) % params.out_h; + c = (global_idx / (params.out_w * params.out_h)) % params.channels; + n = global_idx / (params.out_w * params.out_h * params.channels); + } +}; + +struct cwhn_layout { + __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.in_w * params.in_h) + (y * params.in_w + x) * params.channels + c; + } + + __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) { + return (ky * params.kernel_w + kx) * params.channels + c; + } + + __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) { + return n * (params.channels * params.out_w * params.out_h) + y * (params.out_w * params.channels) + + x * params.channels + c; + } + + __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y, + int & out_x) { + c = global_idx % params.channels; + out_x = (global_idx / params.channels) % params.out_w; + out_y = (global_idx / (params.channels * params.out_w)) % params.out_h; + n = global_idx / (params.channels * params.out_w * params.out_h); + } +}; + +template +__global__ void conv2d_dw_kernel(const T * __restrict__ input, const T * __restrict__ kernel, T * __restrict__ output, + const int in_w, const int in_h, const int out_w, const int out_h, + const int kernel_w, const int kernel_h, const int stride_x, const int stride_y, + const int padding_x, const int padding_y, const int dilation_x, const int dilation_y, + const int channels, const int batches) { + const int global_idx = blockIdx.x * blockDim.x + threadIdx.x; + const int total_elements = batches * channels * out_h * out_w; + + if (global_idx >= total_elements) { + return; + } + + conv_params params = { in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, + stride_y, padding_x, padding_y, dilation_x, dilation_y, channels, batches }; + + int batch_idx, channel_idx, out_y_idx, out_x_idx; + Layout::unpack_indices(global_idx, params, batch_idx, channel_idx, out_y_idx, out_x_idx); + + T accumulator = 0; + kernel_bounds bounds = calculate_kernel_bounds(out_x_idx, out_y_idx, params); + + for (int kern_y = bounds.y_min; kern_y < bounds.y_max; ++kern_y) { + int in_y_idx = calculate_input_coord(out_y_idx, kern_y, params.stride_y, params.dilation_y, params.padding_y); + + for (int kern_x = bounds.x_min; kern_x < bounds.x_max; ++kern_x) { + int in_x_idx = calculate_input_coord(out_x_idx, kern_x, params.stride_x, params.dilation_x, params.padding_x); + + const T input_val = input[Layout::input_index(batch_idx, channel_idx, in_y_idx, in_x_idx, params)]; + const T kernel_val = kernel[Layout::kernel_index(channel_idx, kern_y, kern_x, params)]; + + accumulator += input_val * kernel_val; + } + } + + output[Layout::output_index(batch_idx, channel_idx, out_y_idx, out_x_idx, params)] = accumulator; +} + +void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * kernel = dst->src[0]; + const ggml_tensor * input = dst->src[1]; + + GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); + const float * w_d = (const float *) kernel->data; + const float * x_d = (const float *) input->data; + float * y_d = (float *) dst->data; + + const int32_t * p = (const int32_t *) dst->op_params; + const int stride_x = p[0]; + const int stride_y = p[1]; + const int padding_x = p[2]; + const int padding_y = p[3]; + const int dilation_x = p[4]; + const int dilation_y = p[5]; + + const int in_w = input->ne[0]; + const int in_h = input->ne[1]; + const int kernel_w = kernel->ne[0]; + const int kernel_h = kernel->ne[1]; + const int out_w = dst->ne[0]; + const int out_h = dst->ne[1]; + const int channels = dst->ne[2]; + const int batches = dst->ne[3]; + + cudaStream_t st = ctx.stream(); + + const int total = batches * channels * out_h * out_w; + const int blocks = (total + CUDA_CONV2D_DW_BLOCK_SIZE - 1) / CUDA_CONV2D_DW_BLOCK_SIZE; + + if (ggml_is_contiguous(input)) { + conv2d_dw_kernel<<>>( + x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y, + dilation_x, dilation_y, channels, batches); + } else if (ggml_is_contiguous_channels(input)) { + conv2d_dw_kernel<<>>( + x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y, + dilation_x, dilation_y, channels, batches); + } else { + GGML_ABORT("Unsupported memory layout for conv_2d_dw"); + } +} diff --git a/ggml/src/ggml-cuda/conv2d-dw.cuh b/ggml/src/ggml-cuda/conv2d-dw.cuh new file mode 100644 index 000000000..b5d5a69d3 --- /dev/null +++ b/ggml/src/ggml-cuda/conv2d-dw.cuh @@ -0,0 +1,5 @@ +#pragma once +#include "common.cuh" + +#define CUDA_CONV2D_DW_BLOCK_SIZE 256 +void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 898b24341..80fe05073 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -11,6 +11,7 @@ #include "ggml-cuda/clamp.cuh" #include "ggml-cuda/concat.cuh" #include "ggml-cuda/conv-transpose-1d.cuh" +#include "ggml-cuda/conv2d-dw.cuh" #include "ggml-cuda/convert.cuh" #include "ggml-cuda/count-equal.cuh" #include "ggml-cuda/cpy.cuh" @@ -2310,6 +2311,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_IM2COL: ggml_cuda_op_im2col(ctx, dst); break; + case GGML_OP_CONV_2D_DW: + ggml_cuda_op_conv2d_dw(ctx, dst); + break; case GGML_OP_CONV_TRANSPOSE_1D: ggml_cuda_op_conv_transpose_1d(ctx,dst); break; @@ -3209,6 +3213,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]); } case GGML_OP_IM2COL: + case GGML_OP_CONV_2D_DW: case GGML_OP_POOL_2D: case GGML_OP_SUM: case GGML_OP_SUM_ROWS: From 4c9fdfbe1580a66fd7d77c77418ce2c606a29fdd Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 20 Jun 2025 10:14:14 +0300 Subject: [PATCH 5/7] ubatch : new splitting logic (#14217) ggml-ci --- src/llama-batch.cpp | 924 +++++++++++++++++----------- src/llama-batch.h | 166 ++--- src/llama-context.cpp | 133 ++-- src/llama-context.h | 2 +- src/llama-graph.cpp | 264 +++----- src/llama-graph.h | 6 +- src/llama-hparams.cpp | 4 + src/llama-hparams.h | 2 + src/llama-kv-cache-unified-iswa.cpp | 40 +- src/llama-kv-cache-unified-iswa.h | 7 +- src/llama-kv-cache-unified.cpp | 136 ++-- src/llama-kv-cache-unified.h | 7 +- src/llama-kv-cells.h | 4 +- src/llama-memory-hybrid.cpp | 81 ++- src/llama-memory-hybrid.h | 9 +- src/llama-memory-recurrent.cpp | 76 ++- src/llama-memory-recurrent.h | 7 +- src/llama-memory.h | 7 +- tools/server/server.cpp | 32 - 19 files changed, 992 insertions(+), 915 deletions(-) diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp index 8b6d14fe8..b3c996e18 100644 --- a/src/llama-batch.cpp +++ b/src/llama-batch.cpp @@ -1,7 +1,6 @@ #include "llama-batch.h" #include "llama-impl.h" -#include "llama-cparams.h" #include "llama-vocab.h" #include "llama-memory.h" @@ -10,282 +9,7 @@ #include #include -llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) { - // clear empty sequences - // the previous ubatch is assumed to be gone, - // so nothing should refer to values in these sequences anymore. - for (size_t i = seq.size(); i-- > 0;) { - if (seq[i].length == 0) { - seq.pop_back(); - } else { - break; - } - } - - udatas.push_back({}); - - auto & udata = udatas.back(); - - udata.token.resize(!has_embd ? n_ubatch : 0); - udata.embd.resize(has_embd ? n_embd * n_ubatch : 0); - udata.pos.resize(n_ubatch); - udata.n_seq_id.resize(n_ubatch); - udata.seq_id.resize(n_ubatch); - udata.output.resize(n_ubatch); - - llama_ubatch ubatch = { - /*equal_seqs =*/ true, - /*n_tokens =*/ 0, - /*n_seq_tokens =*/ 0, - /*n_seqs =*/ 0, - /*token =*/ !has_embd ? udata.token.data() : nullptr, - /*embd =*/ has_embd ? udata.embd.data() : nullptr, - /*pos =*/ udata.pos.data(), - /*n_seq_id =*/ udata.n_seq_id.data(), - /*seq_id =*/ udata.seq_id.data(), - /*output =*/ udata.output.data(), - }; - - return ubatch; -} - -void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length) { - GGML_ASSERT(batch != nullptr); - GGML_ASSERT(length <= seq.length); - // Can only add sequences of equal lengths to a batch, - // otherwise it isn't clear to which sequence a token belongs - GGML_ASSERT(seq.n_seq_id == 0 || ubatch.n_seqs == 0 || length == (size_t) ubatch.n_tokens / ubatch.n_seqs); - GGML_ASSERT((seq.n_seq_id != 0) == ubatch.equal_seqs); - // NOTE: loops are separated for cache-friendliness - if (batch->token) { - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - ubatch.token[ubatch.n_tokens + i] = batch->token[ids[seq.offset + i]]; - } - } else { - // simple split - ubatch.token = batch->token + seq.offset; - } - } else { - ubatch.token = nullptr; - } - if (batch->embd) { - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - memcpy( - ubatch.embd + (n_embd * (ubatch.n_tokens + i)), - batch->embd + (n_embd * ids[seq.offset + i]), - n_embd * sizeof(float) - ); - } - } else { - // simple split - ubatch.embd = batch->embd + (n_embd * seq.offset); - } - } else { - ubatch.embd = nullptr; - } - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]]; - } - } else { - // simple split - ubatch.pos = batch->pos + seq.offset; - } - if (ubatch.equal_seqs) { - ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id; - if (seq.seq_id) { - ubatch.seq_id[ubatch.n_seqs] = seq.seq_id; - } - } else { - // simple split - if (batch->n_seq_id) { - ubatch.n_seq_id = batch->n_seq_id + seq.offset; - } else { - for (size_t i = 0; i < length; ++i) { - ubatch.n_seq_id[ubatch.n_seqs + i] = 1; - } - } - if (batch->seq_id) { - ubatch.seq_id = batch->seq_id + seq.offset; - } - } - if (batch->logits) { - if (ubatch.equal_seqs) { - for (size_t i = 0; i < length; ++i) { - size_t id = ids[seq.offset + i]; - int8_t is_output = batch->logits[id]; - ubatch.output[ubatch.n_tokens + i] = is_output; - if (is_output) { out_ids.push_back(id); } - } - } else { - // simple split - ubatch.output = batch->logits + seq.offset; - for (size_t i = 0; i < length; ++i) { - if (ubatch.output[i] != 0) { out_ids.push_back(seq.offset + i); } - } - } - } else { - // only get last output - for (size_t i = 0; i < length; ++i) { - size_t id = ids[seq.offset + i]; - int8_t is_last = id == ids.size() - 1; - ubatch.output[ubatch.n_tokens + i] = is_last; - if (is_last) { out_ids.push_back(id); } - } - } - if (ubatch.n_tokens == 0 && ubatch.n_seqs == 0) { - ubatch.n_seq_tokens = ubatch.equal_seqs ? length : 1; - } - ubatch.n_tokens += length; - ubatch.n_seqs += ubatch.equal_seqs ? 1 : length; // virtual sequences for simple splits - seq.offset += length; - seq.length -= length; - n_tokens -= length; - GGML_ASSERT(ubatch.n_tokens == ubatch.n_seq_tokens * ubatch.n_seqs); -} - -llama_ubatch llama_sbatch::split_simple(size_t n_ubatch) { - n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; - llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); - ubatch.equal_seqs = false; - if (!seq.empty()) { - llama_sbatch_seq & s = seq[0]; - size_t length = s.length < n_ubatch ? s.length : n_ubatch; - GGML_ASSERT(seq.size() == 1 && s.n_seq_id == 0); // don't mix with other splits - add_seq_to_ubatch(ubatch, s, length); - } - return ubatch; -} - -llama_ubatch llama_sbatch::split_equal(size_t n_ubatch) { - n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; - llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); - if (!seq.empty()) { - size_t length = 0; - size_t n_tokens_in_ubatch = 0; - GGML_ASSERT(seq[0].n_seq_id > 0); // should not be mixed with simple splits - // smallest first, because it's easier to split this way; - // starting from the end to pop in constant time. - for (size_t i = seq.size(); i-- > 0;) { - llama_sbatch_seq & s = seq[i]; - GGML_ASSERT(s.length > 0); - if (length == 0) { - length = s.length < n_ubatch ? s.length : n_ubatch; - } - add_seq_to_ubatch(ubatch, s, length); - n_tokens_in_ubatch += length; - // shared prompts can't be mixed with any of their sequences, - // so it's safer to compute them in their own ubatch - if (s.n_seq_id > 1) { break; } - // stop when there isn't enough space for another sequence - if (length + n_tokens_in_ubatch > n_ubatch) { break; } - } - } - return ubatch; -} - -llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) { - n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch; - llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr); - if (!seq.empty()) { - llama_sbatch_seq & s = seq[seq.size() - 1]; - size_t length = s.length < n_ubatch ? s.length : n_ubatch; - GGML_ASSERT(s.n_seq_id > 0); // should not be mixed with simple splits - add_seq_to_ubatch(ubatch, s, length); - } - return ubatch; -} - -llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split) { - GGML_ASSERT(batch.n_tokens >= 0); - this->batch = &batch; - this->n_embd = n_embd; - - n_tokens = batch.n_tokens; - ids.resize(n_tokens); - out_ids.clear(); - // TODO: reserve out_ids and seq - - for (size_t i = 0; i < n_tokens; ++i) { - ids[i] = i; - } - - if (simple_split) { - seq.resize(1); - llama_sbatch_seq & s = seq[0]; - s.n_seq_id = 0; - s.seq_id = nullptr; - s.offset = 0; - s.length = n_tokens; - return; - } - - std::sort(ids.begin(), ids.end(), - [&batch](size_t a, size_t b) { - int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1; - int32_t n_seq_b = batch.n_seq_id ? batch.n_seq_id[b] : 1; - // sort by seq_id, then by pos - if (n_seq_a == n_seq_b) { - if (batch.seq_id) { - for (int32_t i = 0; i < n_seq_a; ++i) { - llama_seq_id seq_id_a = batch.seq_id[a][i]; - llama_seq_id seq_id_b = batch.seq_id[b][i]; - // smaller seq_ids go first - if (seq_id_a != seq_id_b) { - return seq_id_a < seq_id_b; - } - } - } - // when all else is equal, sort by pos - if (batch.pos) { - return batch.pos[a] < batch.pos[b]; - } - // no pos, sort by id - return a < b; - } - // shared prompts go first - return n_seq_a > n_seq_b; - } - ); - - // init seq - llama_sbatch_seq * last_seq = nullptr; - - for (size_t i = 0; i < n_tokens; ++i) { - const size_t bi = ids[i]; - const int32_t n_seqs = batch.n_seq_id[bi]; - llama_seq_id * seq_ids = batch.seq_id[bi]; - if (last_seq != nullptr) { - bool same = n_seqs == last_seq->n_seq_id; - for (int32_t j = 0; same && j < n_seqs; ++j) { - if (seq_ids[j] != last_seq->seq_id[j]) { - same = false; - } - } - if (same) { - last_seq->length += 1; - continue; - } - } - llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1}; - seq.push_back(new_seq); - last_seq = &seq.back(); - } - - // keep shared prompts first at the end, then sort by length descending. - std::sort(seq.begin(), seq.end(), - [](llama_sbatch_seq & a, llama_sbatch_seq & b) { - if (a.n_seq_id == b.n_seq_id) { - return a.length > b.length; - } - return a.n_seq_id < b.n_seq_id; - } - ); -} - -llama_batch_allocr::llama_batch_allocr() { +llama_batch_allocr::llama_batch_allocr(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) { const char * LLAMA_BATCH_DEBUG = getenv("LLAMA_BATCH_DEBUG"); debug = LLAMA_BATCH_DEBUG ? atoi(LLAMA_BATCH_DEBUG) : 0; @@ -294,17 +18,22 @@ llama_batch_allocr::llama_batch_allocr() { for (auto & cur : seq_cpl) { cur.resize(LLAMA_MAX_SEQ); } + + seq_idx.resize(LLAMA_MAX_SEQ, -1); } bool llama_batch_allocr::init( const llama_batch & batch_inp, const llama_vocab & vocab, const llama_memory_i * memory, - bool embd_all) { + uint32_t n_embd, + bool output_all) { clear(); batch = batch_inp; + this->vocab = &vocab; + GGML_ASSERT(batch.n_tokens > 0); // @@ -359,6 +88,7 @@ bool llama_batch_allocr::init( llama_pos p0[LLAMA_MAX_SEQ]; for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { if (!memory) { + // if no memory -> start from 0 p0[s] = 0; } else { p0[s] = memory->seq_pos_max(s) + 1; @@ -370,8 +100,11 @@ bool llama_batch_allocr::init( pos[i] = p0[seq_id]; + // update the starting position for all sequences that are assigned to the this token for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { - p0[batch.seq_id[i][s]] = pos[i] + 1; + const llama_seq_id seq_id = batch.seq_id[i][s]; + + p0[seq_id] = pos[i] + 1; } } @@ -379,7 +112,7 @@ bool llama_batch_allocr::init( } if (!batch.logits) { - if (embd_all) { + if (output_all) { // return the output for all tokens output.resize(batch.n_tokens, true); } else { @@ -389,7 +122,7 @@ bool llama_batch_allocr::init( } batch.logits = output.data(); - } else if (embd_all) { + } else if (output_all) { bool warn = false; for (int32_t i = 0; i < batch.n_tokens; ++i) { @@ -410,6 +143,9 @@ bool llama_batch_allocr::init( // compute stats // + this->n_embd = n_embd; + + // count the outputs in this batch for (int32_t i = 0; i < batch.n_tokens; ++i) { n_outputs += batch.logits[i] != 0; } @@ -417,85 +153,86 @@ bool llama_batch_allocr::init( // determine coupled sequences // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them for (int32_t i = 0; i < batch.n_tokens; ++i) { + const llama_seq_id s0 = batch.seq_id[i][0]; + for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { - seq_pos[batch.seq_id[i][s]].insert(batch.pos[i]); + const llama_seq_id s1 = batch.seq_id[i][s]; + + seq_pos[s1].insert(batch.pos[i]); if (s > 0) { - const llama_seq_id s0 = batch.seq_id[i][0]; - const llama_seq_id s1 = batch.seq_id[i][s]; - // mark that sequence s1 is coupled to s0 seq_cpl[s1][s0] = true; - // note: the other way around is not necessary for now + // note: tracking the other way around is not necessary for now //seq_cpl[s0][s1] = true; } } } + // precompute the sequence sets for each token and determine the unique sequence ids that participate in the batch + { + seq_set_t seq_set_unq; + + for (int32_t i = 0; i < batch.n_tokens; ++i) { + seq_set_t cur; + for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { + const llama_seq_id seq_id = batch.seq_id[i][s]; + + cur .set(seq_id); + seq_set_unq.set(seq_id); + } + + seq_set.push_back(cur); + seq_set_map[cur].push_back(i); + } + + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + if (seq_set_unq.test(s)) { + seq_idx[s] = seq_id_unq.size(); + seq_id_unq.push_back(s); + } + } + } + if (debug > 0) { LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__); - LLAMA_LOG_DEBUG("%s: n_tokens = %d\n", __func__, batch.n_tokens); - LLAMA_LOG_DEBUG("%s: token = %p\n", __func__, (void *) batch.token); - LLAMA_LOG_DEBUG("%s: embd = %p\n", __func__, (void *) batch.embd); - LLAMA_LOG_DEBUG("%s: pos = %p\n", __func__, (void *) batch.pos); - LLAMA_LOG_DEBUG("%s: n_seq_id = %p\n", __func__, (void *) batch.n_seq_id); - LLAMA_LOG_DEBUG("%s: seq_id = %p\n", __func__, (void *) batch.seq_id); - LLAMA_LOG_DEBUG("%s: logits = %p\n", __func__, (void *) batch.logits); - LLAMA_LOG_DEBUG("%s: n_outputs = %d\n", __func__, n_outputs); - if (debug > 1) { - int seq_id_max = 0; - for (int32_t i = 0; i < batch.n_tokens; ++i) { - for (int s = 0; s < batch.n_seq_id[i]; ++s) { - for (int s = 0; s < batch.n_seq_id[i]; ++s) { - seq_id_max = std::max(seq_id_max, batch.seq_id[i][s]); - } + llama_ubatch ubatch { + /*.equal_seqs =*/ false, + /*.n_tokens =*/ (uint32_t) batch.n_tokens, + /*.n_seq_tokens =*/ (uint32_t) 1, + /*.n_seqs =*/ (uint32_t) batch.n_tokens, + /*.n_seqs_unq =*/ (uint32_t) this->seq_id_unq.size(), + /*.token =*/ batch.token, + /*.embd =*/ batch.embd, + /*.pos =*/ batch.pos, + /*.n_seq_id =*/ batch.n_seq_id, + /*.seq_id =*/ batch.seq_id, + /*.seq_id_unq =*/ this->seq_id_unq.data(), + /*.seq_idx =*/ this->seq_idx.data(), + /*.output =*/ batch.logits, + }; + + ubatch_print(ubatch, debug); + + LLAMA_LOG_DEBUG("%s: seq = [\n", __func__); + for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) { + if (seq_pos[s0].empty()) { + continue; + } + + std::stringstream ss; + for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) { + if (seq_cpl[s0][s1]) { + ss << s1 << " "; } } - ++seq_id_max; - LLAMA_LOG_DEBUG("%s: token = [\n", __func__); - for (int32_t i = 0; i < batch.n_tokens; ++i) { - std::vector seq_id(seq_id_max); - - for (int s = 0; s < batch.n_seq_id[i]; ++s) { - seq_id[batch.seq_id[i][s]] = 1; - } - - std::stringstream ss; - for (int s = 0; s < seq_id_max; ++s) { - if (seq_id[s]) { - ss << s%10; - } else { - ss << "."; - } - } - - LLAMA_LOG_DEBUG("%s: %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n", - __func__, i, batch.token[i], vocab.token_to_piece(batch.token[i]).c_str(), - batch.pos[i], batch.n_seq_id[i], ss.str().c_str(), batch.logits[i]); - } - LLAMA_LOG_DEBUG("%s: ]\n", __func__); - - LLAMA_LOG_DEBUG("%s: seq = [\n", __func__); - for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) { - if (seq_pos[s0].empty()) { - continue; - } - - std::stringstream ss; - for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) { - if (seq_cpl[s0][s1]) { - ss << s1 << " "; - } - } - - LLAMA_LOG_DEBUG("%s: %4d: pos = [%4d, %4d], cpl = %s\n", - __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str()); - } - LLAMA_LOG_DEBUG("%s: ]\n", __func__); + LLAMA_LOG_DEBUG("%s: %4d: pos = [%4d, %4d], cpl = %s\n", + __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str()); } + LLAMA_LOG_DEBUG("%s: ]\n", __func__); } // @@ -507,9 +244,22 @@ bool llama_batch_allocr::init( continue; } - if (memory && seq_pos_min(s) != memory->seq_pos_max(s) + 1) { - LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); - return false; + if (memory) { + if (batch.token) { + if (seq_pos_min(s) != memory->seq_pos_max(s) + 1) { + LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); + return false; + } + } else { + assert(batch.embd); + + // for embeddings (typically used as vision input), we allow them to have repeating positions + // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762 + if (seq_pos_min(s) != memory->seq_pos_max(s) && seq_pos_min(s) != memory->seq_pos_max(s) + 1) { + LLAMA_LOG_ERROR("%s: sequence %d does not start from the last position stored in the memory\n", __func__, s); + return false; + } + } } if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) { @@ -532,17 +282,120 @@ bool llama_batch_allocr::init( } } + // disallow partial sequence sub-sets: + // + // invalid: x + // i: 0 1 2 ... + // --------------------------------------- + // seq_id[i][0]: 0 0 1 + // seq_id[i][1]: 1 1 2 + // seq_id[i][2]: 2 + // + // disallow decreasing sequence positions: + // + // invalid: x + // i: 0 1 2 3 4 5 6 ... + // --------------------------------------- + // pos[i]: 4 5 0 1 6 2 3 + // seq_id[i][0]: 0 0 1 1 0 1 0 + // + { + seq_set_t cur_seq_set[LLAMA_MAX_SEQ]; + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + cur_seq_set[s].set(); + } + + llama_pos cur_seq_pos[LLAMA_MAX_SEQ]; + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + cur_seq_pos[s] = -1; + } + + for (int32_t i = 0; i < batch.n_tokens; ++i) { + const llama_pos pos = batch.pos[i]; + + for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) { + const llama_seq_id seq_id = batch.seq_id[i][s]; + + cur_seq_set[seq_id] &= seq_set[i]; + + if (cur_seq_set[seq_id].none()) { + LLAMA_LOG_ERROR("%s: sequence %d belongs to incompatible sequence sets (not allowed)\n", __func__, seq_id); + return false; + } + + if (pos < cur_seq_pos[seq_id]) { + LLAMA_LOG_ERROR("%s: sequence %d positions are decreasing (not allowed)\n", __func__, seq_id); + return false; + } + } + } + } + + split_reset(); + return true; } +llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs) { + const uint32_t n_tokens = n_seq_tokens*n_seqs; + + clear(); + split_reset(); + + ubatches.emplace_back(); + + auto & ubatch = ubatches.back(); + + ubatch.token .resize(n_tokens); + ubatch.embd .clear(); + ubatch.pos .resize(n_tokens); + ubatch.n_seq_id .resize(n_tokens); + ubatch.seq_id .resize(n_tokens); + ubatch.seq_id_unq.resize(0); + ubatch.seq_idx .resize(LLAMA_MAX_SEQ, -1); + ubatch.output .resize(n_tokens); + + for (uint32_t s = 0; s < n_seqs; ++s) { + ubatch.seq_idx[s] = s; + ubatch.seq_id_unq.push_back(s); + } + + llama_ubatch res { + /*.equal_seqs =*/ true, + /*.n_tokens =*/ n_tokens, + /*.n_seq_tokens =*/ n_seq_tokens, + /*.n_seqs =*/ n_seqs, + /*.n_seqs_unq =*/ n_seqs, + + /*.token =*/ ubatch.token.data(), + /*.embd =*/ nullptr, + /*.pos =*/ ubatch.pos.data(), + /*.n_seq_id =*/ ubatch.n_seq_id.data(), + /*.seq_id =*/ ubatch.seq_id.data(), + /*.seq_id_unq =*/ ubatch.seq_id_unq.data(), + /*.seq_idx =*/ ubatch.seq_idx.data(), + /*.output =*/ ubatch.output.data(), + }; + + return res; +} + const llama_batch & llama_batch_allocr::get_batch() const { return batch; } +uint32_t llama_batch_allocr::get_n_tokens() const { + return batch.n_tokens; +} + uint32_t llama_batch_allocr::get_n_outputs() const { return n_outputs; } +std::vector & llama_batch_allocr::get_out_ids() { + return out_ids; +} + llama_pos llama_batch_allocr::seq_pos_min(llama_seq_id seq_id) const { return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].begin(); } @@ -551,14 +404,188 @@ llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const { return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].rbegin(); } +void llama_batch_allocr::split_reset() { + out_ids.clear(); + + used.clear(); + used.resize(get_n_tokens(), false); + + ubatches.clear(); +} + +llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) { + // find the first unused token + uint32_t cur_idx = 0; + while (cur_idx < used.size() && used[cur_idx]) { + ++cur_idx; + } + + // we are done + if (cur_idx >= used.size()) { + return {}; + } + + std::vector idxs; + + while (true) { + idxs.push_back(cur_idx); + + used[cur_idx] = true; + + ++cur_idx; + + if (cur_idx >= used.size()) { + break; + } + + if (idxs.size() >= n_ubatch) { + break; + } + } + + return ubatch_add(idxs, idxs.size(), false); +} + +llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch) { + std::vector cur_seq_set; + + // determine the non-overlapping sequence sets participating in this ubatch + for (int32_t i = 0; i < batch.n_tokens; ++i) { + if (used[i]) { + continue; + } + + bool add = true; + + for (uint32_t s = 0; s < cur_seq_set.size(); ++s) { + // no overlap with existing sequence sets: + if (!(cur_seq_set[s] & seq_set[i]).none()) { + add = false; + break; + } + } + + if (add) { + cur_seq_set.push_back(seq_set[i]); + + if (cur_seq_set.size() > n_ubatch) { + break; + } + } + } + + const uint32_t n_seqs = cur_seq_set.size(); + + // we are done + if (n_seqs == 0) { + return {}; + } + + // the current batch index of each sequence set + std::vector cur_idx(n_seqs, 0); + + for (uint32_t s = 0; s < n_seqs; ++s) { + while (used[seq_set_map[cur_seq_set[s]][cur_idx[s]]]) { + ++cur_idx[s]; + } + } + + // the list of batch indices for each sequence set + // at the end we will concat these to get the final ubatch + std::vector idxs_per_seq(n_seqs); + + while (true) { + // we can only add new n_seq_tokens tokens if all the sequence sets have at least one more unused token and + // if we haven't reached n_ubatch + bool can_expand = true; + + for (uint32_t s = 0; s < n_seqs; ++s) { + if (cur_idx[s] >= (int32_t) seq_set_map[cur_seq_set[s]].size()) { + can_expand = false; + break; + } + } + + if (!can_expand) { + break; + } + + for (uint32_t s = 0; s < n_seqs; ++s) { + const int32_t idx = seq_set_map[cur_seq_set[s]][cur_idx[s]]; + + idxs_per_seq[s].push_back(idx); + + used[idx] = true; + + ++cur_idx[s]; + } + + if ((idxs_per_seq[0].size() + 1)*n_seqs > n_ubatch) { + break; + } + } + + // concat the per-sequence-set lists + std::vector idxs; + + for (uint32_t s = 0; s < n_seqs; ++s) { + idxs.insert(idxs.end(), idxs_per_seq[s].begin(), idxs_per_seq[s].end()); + } + + return ubatch_add(idxs, n_seqs, true); +} + +llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) { + // find the first unused token + uint32_t cur_idx = 0; + while (cur_idx < used.size() && used[cur_idx]) { + ++cur_idx; + } + + // we are done + if (cur_idx >= used.size()) { + return {}; + } + + // this is the starting sequence set + // we allow adding tokens only if their sequence set is a subset of the current sequence set + auto cur_seq_set = seq_set[cur_idx]; + + std::vector idxs; + + while (true) { + idxs.push_back(cur_idx); + + used[cur_idx] = true; + + if (idxs.size() >= n_ubatch) { + break; + } + + do { + ++cur_idx; + } while (cur_idx < get_n_tokens() && (used[cur_idx] || ((cur_seq_set & seq_set[cur_idx]) != seq_set[cur_idx]))); + + if (cur_idx == get_n_tokens()) { + break; + } + + cur_seq_set = seq_set[cur_idx]; + } + + return ubatch_add(idxs, 1, true); +} + void llama_batch_allocr::clear() { n_outputs = 0; batch = {}; - pos.clear(); - n_seq_id.clear(); - seq_id.clear(); - output.clear(); + + pos .clear(); + n_seq_id .clear(); + seq_id .clear(); + seq_id_unq.clear(); + output .clear(); for (auto & cur : seq_pos) { cur.clear(); @@ -567,6 +594,177 @@ void llama_batch_allocr::clear() { for (auto & cur : seq_cpl) { std::fill(cur.begin(), cur.end(), false); } + + seq_set.clear(); + + seq_set_map.clear(); + + std::fill(seq_idx.begin(), seq_idx.end(), -1); +} + +llama_ubatch llama_batch_allocr::ubatch_add(const std::vector & idxs, uint32_t n_seqs, bool equal_seqs) { + const uint32_t n_tokens = idxs.size(); + + assert(n_tokens%n_seqs == 0); + + ubatches.emplace_back(); + + auto & ubatch = ubatches.back(); + + const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1; + + const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0; + const int64_t n_pos_all = (int64_t) n_tokens*n_pos_cur; + + ubatch.token .resize(n_tokens); + ubatch.embd .resize(n_embd_all); + ubatch.pos .resize(n_pos_all); + ubatch.n_seq_id .resize(n_tokens); + ubatch.seq_id .resize(n_tokens); + ubatch.seq_id_unq.resize(0); + ubatch.seq_idx .resize(LLAMA_MAX_SEQ, -1); + ubatch.output .resize(n_tokens); + + seq_set_t seq_set_unq; + + for (size_t i = 0; i < idxs.size(); ++i) { + if (batch.token) { + ubatch.token[i] = batch.token[idxs[i]]; + } + + if (batch.embd) { + memcpy(ubatch.embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float)); + } + + for (int j = 0; j < n_pos_cur; ++j) { + ubatch.pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]]; + } + + ubatch.n_seq_id[i] = batch.n_seq_id[idxs[i]]; + ubatch.seq_id[i] = batch.seq_id[idxs[i]]; + ubatch.output[i] = batch.logits[idxs[i]]; + + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + seq_set_unq.set(ubatch.seq_id[i][s]); + } + + if (ubatch.output[i]) { + out_ids.push_back(idxs[i]); + } + } + + for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + if (seq_set_unq.test(s)) { + ubatch.seq_idx[s] = ubatch.seq_id_unq.size(); + ubatch.seq_id_unq.push_back(s); + } + } + + llama_ubatch res { + /*.equal_seqs =*/ equal_seqs, + /*.n_tokens =*/ n_tokens, + /*.n_seq_tokens =*/ n_tokens/n_seqs, + /*.n_seqs =*/ n_seqs, + /*.n_seqs_unq =*/ (uint32_t) ubatch.seq_id_unq.size(), + + /*.token =*/ batch.token ? ubatch.token.data() : nullptr, + /*.embd =*/ batch.embd ? ubatch.embd.data() : nullptr, + /*.pos =*/ ubatch.pos.data(), + /*.n_seq_id =*/ ubatch.n_seq_id.data(), + /*.seq_id =*/ ubatch.seq_id.data(), + /*.seq_id_unq =*/ ubatch.seq_id_unq.data(), + /*.seq_idx =*/ ubatch.seq_idx.data(), + /*.output =*/ ubatch.output.data(), + }; + + if (debug > 0) { + LLAMA_LOG_DEBUG("%s: added ubatch %d to split:\n", __func__, (int) ubatches.size() - 1); + + ubatch_print(res, debug); + } + + return res; +} + +void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) { + if (debug > 0) { + LLAMA_LOG_DEBUG("%s: equal_seqs = %d\n", __func__, ubatch.equal_seqs); + LLAMA_LOG_DEBUG("%s: n_tokens = %d\n", __func__, ubatch.n_tokens); + LLAMA_LOG_DEBUG("%s: n_seq_tokens = %d\n", __func__, ubatch.n_seq_tokens); + LLAMA_LOG_DEBUG("%s: n_seqs = %d\n", __func__, ubatch.n_seqs); + LLAMA_LOG_DEBUG("%s: n_seqs_unq = %d\n", __func__, ubatch.n_seqs_unq); + + std::stringstream ss_seq_id_unq; + std::stringstream ss_seq_idx; + + ss_seq_id_unq << "[ "; + ss_seq_idx << "["; + + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + ss_seq_id_unq << ubatch.seq_id_unq[s] << " "; + } + + for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) { + if (ubatch.seq_idx[s] >= 0) { + ss_seq_idx << ubatch.seq_idx[s]%10; + } else { + ss_seq_idx << "."; + } + } + + ss_seq_id_unq << "]"; + ss_seq_idx << "]"; + + LLAMA_LOG_DEBUG("%s: token = %p\n", __func__, (void *) ubatch.token); + LLAMA_LOG_DEBUG("%s: embd = %p\n", __func__, (void *) ubatch.embd); + LLAMA_LOG_DEBUG("%s: pos = %p\n", __func__, (void *) ubatch.pos); + LLAMA_LOG_DEBUG("%s: n_seq_id = %p\n", __func__, (void *) ubatch.n_seq_id); + LLAMA_LOG_DEBUG("%s: seq_id = %p\n", __func__, (void *) ubatch.seq_id); + LLAMA_LOG_DEBUG("%s: seq_id_unq = %s\n", __func__, ss_seq_id_unq.str().c_str()); + LLAMA_LOG_DEBUG("%s: seq_idx = %s\n", __func__, ss_seq_idx.str().c_str()); + LLAMA_LOG_DEBUG("%s: output = %p\n", __func__, (void *) ubatch.output); + LLAMA_LOG_DEBUG("%s: n_outputs = %d\n", __func__, n_outputs); + + if (debug > 1) { + int seq_id_max = 0; + for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + seq_id_max = std::max(seq_id_max, ubatch.seq_id[i][s]); + } + } + } + ++seq_id_max; + + LLAMA_LOG_DEBUG("%s: token = [\n", __func__); + for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { + std::vector seq_id(seq_id_max); + + for (int s = 0; s < ubatch.n_seq_id[i]; ++s) { + seq_id[ubatch.seq_id[i][s]] = 1; + } + + std::stringstream ss; + for (int s = 0; s < seq_id_max; ++s) { + if (seq_id[s]) { + ss << s%10; + } else { + ss << "."; + } + } + + if (ubatch.token) { + LLAMA_LOG_DEBUG("%s: %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n", + __func__, i, ubatch.token[i], vocab->token_to_piece(ubatch.token[i]).c_str(), + ubatch.pos[i], ubatch.n_seq_id[i], ss.str().c_str(), ubatch.output[i]); + } else { + LLAMA_LOG_DEBUG("%s: %4d: [embd], pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n", + __func__, i, ubatch.pos[i], ubatch.n_seq_id[i], ss.str().c_str(), ubatch.output[i]); + } + } + LLAMA_LOG_DEBUG("%s: ]\n", __func__); + } + } } // @@ -577,25 +775,25 @@ struct llama_batch llama_batch_get_one( llama_token * tokens, int32_t n_tokens) { return { - /*n_tokens =*/ n_tokens, - /*tokens =*/ tokens, - /*embd =*/ nullptr, - /*pos =*/ nullptr, - /*n_seq_id =*/ nullptr, - /*seq_id =*/ nullptr, - /*logits =*/ nullptr, + /*n_tokens =*/ n_tokens, + /*tokens =*/ tokens, + /*embd =*/ nullptr, + /*pos =*/ nullptr, + /*n_seq_id =*/ nullptr, + /*seq_id =*/ nullptr, + /*logits =*/ nullptr, }; } struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) { llama_batch batch = { - /*n_tokens =*/ 0, - /*tokens =*/ nullptr, - /*embd =*/ nullptr, - /*pos =*/ nullptr, - /*n_seq_id =*/ nullptr, - /*seq_id =*/ nullptr, - /*logits =*/ nullptr, + /*n_tokens =*/ 0, + /*tokens =*/ nullptr, + /*embd =*/ nullptr, + /*pos =*/ nullptr, + /*n_seq_id =*/ nullptr, + /*seq_id =*/ nullptr, + /*logits =*/ nullptr, }; if (embd) { diff --git a/src/llama-batch.h b/src/llama-batch.h index a555c1572..d2c537618 100644 --- a/src/llama-batch.h +++ b/src/llama-batch.h @@ -2,86 +2,44 @@ #include "llama.h" +#include "llama-cparams.h" + #include #include #include +#include +#include -// very similar to llama_batch, -// but has more metadata about sequences +// keep this struct lightweight +// it points to data in `llama_batch_allocr` struct llama_ubatch { bool equal_seqs; // TODO: whole_seqs for embeddings? uint32_t n_tokens; // total tokens (n_seq_tokens * n_seqs) - uint32_t n_seq_tokens; // tokens per sequence - uint32_t n_seqs; + uint32_t n_seq_tokens; // tokens per sequence set + uint32_t n_seqs; // sequence sets in the ubatch + uint32_t n_seqs_unq; // unique sequence ids in the ubatch - llama_token * token; // [n_tokens] - float * embd; // [n_embd, n_tokens] - llama_pos * pos; // [n_tokens] - int32_t * n_seq_id; // [n_seqs] - llama_seq_id ** seq_id; // [n_seqs] - int8_t * output; // [n_tokens] + // seq_id_unq: unique sequence ids in the ubatch + // seq_idx: indices of the unique sequence ids in the ubatch in [0, n_seqs_unq) + // used for extracting sequence pooled embeddings + + // // size | idx | val + llama_token * token; // [n_tokens] | i | id, token + float * embd; // [n_embd, n_tokens] | i | embd + llama_pos * pos; // [n_tokens] | i | pos + int32_t * n_seq_id; // [n_tokens] | i | - + llama_seq_id ** seq_id; // [n_tokens] | s | s0, s1, seq_id + llama_seq_id * seq_id_unq; // [n_seqs_unq] | s | seq_id + int32_t * seq_idx; // [LLAMA_MAX_SEQ] | - | seq_idx + int8_t * output; // [n_tokens] | i | - }; -struct llama_sbatch_seq { - int32_t n_seq_id; - - llama_seq_id * seq_id; - - size_t offset; - size_t length; -}; - -// sequence-length-aware batch splitting -struct llama_sbatch { - // tokens left in this batch - size_t n_tokens; - - size_t n_embd; - - // sorted indices into the batch - std::vector ids; - // batch indices of the output - std::vector out_ids; - std::vector seq; - - const llama_batch * batch = nullptr; - - // buffers for the ubatches - // TODO: very hacky, this needs a complete rework - struct ubatch_data { - std::vector token; - std::vector embd; - std::vector pos; - std::vector n_seq_id; - std::vector seq_id; - std::vector output; - }; - - std::vector udatas; - - llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false); - - void add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length); - - // simple split, unknown number of sequences of unequal lengths - llama_ubatch split_simple(size_t n_ubatch); - - // make batches of equal-length sequences - llama_ubatch split_equal(size_t n_ubatch); - - // sequence-wise split - llama_ubatch split_seq(size_t n_ubatch); - - llama_sbatch() = default; - llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false); -}; - -// a helper for sanitizing and fulfilling a batch +// a helper for sanitizing, fulfilling and splitting a batch class llama_batch_allocr { public: - llama_batch_allocr(); + llama_batch_allocr(uint32_t n_pos_per_embd); // sanitize and auto-gen missing data in the input batch // memory is optional. if provided will be used to check for sequence continuity and to determine the positions @@ -89,20 +47,57 @@ public: const llama_batch & batch_inp, const llama_vocab & vocab, const llama_memory_i * memory, - bool embd_all); + uint32_t n_embd, + bool output_all); const llama_batch & get_batch() const; + uint32_t get_n_tokens() const; uint32_t get_n_outputs() const; + // the array of output indices in the order they were encountered during the ubatch splitting + std::vector & get_out_ids(); + + // min/max positions of each sequence in the current ubatch llama_pos seq_pos_min(llama_seq_id seq_id) const; llama_pos seq_pos_max(llama_seq_id seq_id) const; + // call once before splitting the batch to reset the internal state + void split_reset(); + + // simple split, unknown number of sequence sets of unequal lengths + llama_ubatch split_simple(uint32_t n_ubatch); + + // make ubatches of equal-length sequences sets + llama_ubatch split_equal(uint32_t n_ubatch); + + // sequence-set-wise split - each ubatch contains a single sequence-set + llama_ubatch split_seq(uint32_t n_ubatch); + + // a helper method for creating a well-defined ubatch of tokens + // TODO: support embeddings if needed in the future + llama_ubatch ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs); + private: void clear(); + // create the next ubatch based on the provided batch indices (idxs) and the number of sequence sets (n_seqs) + // return llama_ubatch.n_tokens == 0 if the entire batch was consumed + llama_ubatch ubatch_add(const std::vector & idxs, uint32_t n_seqs, bool equal_seqs); + + // for debugging, start with LLAMA_BATCH_DEBUG=2 + void ubatch_print(const llama_ubatch & ubatch, int debug); + llama_batch batch; + // only for debugging purposes + const llama_vocab * vocab; + + // TODO: this is more of a temporary solution until we have a better way to handle multiple positions per token/embd + // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762 + const uint32_t n_pos_per_embd; + + uint32_t n_embd; uint32_t n_outputs; std::array seq_id_0 = { 0 }; // default sequence id @@ -110,10 +105,43 @@ private: std::vector pos; std::vector n_seq_id; std::vector seq_id; + std::vector seq_id_unq; + std::vector seq_idx; std::vector output; - std::vector> seq_pos; // seq_pos[s]: the set of positions in sequence s - std::vector> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1 + using pos_set_t = std::set; + using seq_cpl_t = std::vector; + + std::vector seq_pos; // seq_pos[s]: the set of positions in sequence s + std::vector seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1 + + using idx_vec_t = std::vector; + using seq_set_t = std::bitset; + + std::vector seq_set; // seq_set[i]: the sequence set of token i + + std::unordered_map seq_set_map; // the indices at which the sequence set appears + + // batch indices of the output + std::vector out_ids; + + // used[i] indicates if token i has already been used in a previous ubatch + std::vector used; + + // llama_ubatch points to this data: + struct ubatch { + std::vector token; + std::vector embd; + std::vector pos; + std::vector n_seq_id; + std::vector seq_id; + std::vector seq_id_unq; + std::vector seq_idx; + std::vector output; + }; + + // current splitting state: + std::vector ubatches; int debug; }; diff --git a/src/llama-context.cpp b/src/llama-context.cpp index f56a58e9b..5a18a4fb3 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -20,7 +20,7 @@ llama_context::llama_context( const llama_model & model, llama_context_params params) : model(model), - batch_allocr(std::make_unique()) { + balloc(std::make_unique(model.hparams.n_pos_per_embd())) { LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__); t_start_us = model.t_start_us; @@ -722,22 +722,26 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, } int llama_context::encode(const llama_batch & batch_inp) { + GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT + if (batch_inp.n_tokens == 0) { LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); return -1; } + const auto & hparams = model.hparams; + + const int64_t n_embd = hparams.n_embd; + // note: during encode, we always pass the full sequence starting from pos = 0 - if (!batch_allocr->init(batch_inp, model.vocab, nullptr, true)) { + if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, true)) { LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); return -1; } - const llama_batch & batch = batch_allocr->get_batch(); + const uint32_t n_tokens = balloc->get_n_tokens(); - const uint32_t n_tokens = batch.n_tokens; - - GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + const llama_ubatch ubatch = balloc->split_simple(n_tokens); // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens"); @@ -751,14 +755,6 @@ int llama_context::encode(const llama_batch & batch_inp) { n_queued_tokens += n_tokens; - const auto & hparams = model.hparams; - - const int64_t n_embd = hparams.n_embd; - - llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true); - - const llama_ubatch ubatch = sbatch.split_simple(n_tokens); - // reserve output buffer if (output_reserve(n_tokens) < n_tokens) { LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); @@ -817,34 +813,28 @@ int llama_context::encode(const llama_batch & batch_inp) { { // extract sequence embeddings auto & embd_seq_out = embd_seq; - embd_seq_out.clear(); - GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; - // TODO: fix indexing [UBATCH_IDX] - for (uint32_t i = 0; i < n_tokens; i++) { - const llama_seq_id seq_id = ubatch.seq_id[i][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } embd_seq_out[seq_id].resize(n_embd); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_RANK: { // extract the rerank score - n_cls_out floats per sequence auto & embd_seq_out = embd_seq; + const uint32_t n_cls_out = hparams.n_cls_out; - // TODO: fix indexing [UBATCH_IDX] - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; + embd_seq_out[seq_id].resize(n_cls_out); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_id)*sizeof(float), n_cls_out*sizeof(float)); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_UNSPECIFIED: @@ -869,12 +859,16 @@ int llama_context::encode(const llama_batch & batch_inp) { cross.v_embd.resize(cross.n_embd*cross.n_enc); memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd)); + const auto & batch = balloc->get_batch(); + // remember the sequence ids used during the encoding - needed for cross attention later cross.seq_ids_enc.resize(n_tokens); for (uint32_t i = 0; i < n_tokens; i++) { cross.seq_ids_enc[i].clear(); + for (int s = 0; s < batch.n_seq_id[i]; s++) { - llama_seq_id seq_id = batch.seq_id[i][s]; + const llama_seq_id seq_id = batch.seq_id[i][s]; + cross.seq_ids_enc[i].insert(seq_id); } } @@ -884,6 +878,8 @@ int llama_context::encode(const llama_batch & batch_inp) { } int llama_context::decode(const llama_batch & batch_inp) { + GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT + if (!memory) { LLAMA_LOG_DEBUG("%s: cannot decode batches with this context (calling encode() instead)\n", __func__); return encode(batch_inp); @@ -894,29 +890,24 @@ int llama_context::decode(const llama_batch & batch_inp) { return -1; } - // when computing embeddings, all tokens are output - const bool embd_all = cparams.embeddings; - - if (!batch_allocr->init(batch_inp, model.vocab, memory.get(), embd_all)) { - LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); - return -1; - } - - const llama_batch & batch = batch_allocr->get_batch(); - const auto & vocab = model.vocab; const auto & hparams = model.hparams; const int32_t n_vocab = vocab.n_tokens(); const int64_t n_embd = hparams.n_embd; - const uint32_t n_tokens_all = batch.n_tokens; + // when computing embeddings, all tokens are output + const bool output_all = cparams.embeddings; - GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, output_all)) { + LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); + return -1; + } - const uint32_t n_outputs_all = batch_allocr->get_n_outputs(); + const uint32_t n_tokens_all = balloc->get_n_tokens(); + const uint32_t n_outputs_all = balloc->get_n_outputs(); - if (embd_all) { + if (output_all) { // require that all tokens are output if (n_outputs_all != n_tokens_all) { LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %d, n_tokens_all = %d)\n", @@ -945,7 +936,7 @@ int llama_context::decode(const llama_batch & batch_inp) { llama_memory_state_ptr mstate; while (true) { - mstate = memory->init_batch(batch, cparams.n_ubatch, embd_all); + mstate = memory->init_batch(*balloc, cparams.n_ubatch, output_all); if (!mstate) { return -2; } @@ -966,19 +957,19 @@ int llama_context::decode(const llama_batch & batch_inp) { did_optimize = true; if (kv_self_update(true)) { - LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, batch.n_tokens); + LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, balloc->get_n_tokens()); continue; } } - LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, batch.n_tokens); + LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, balloc->get_n_tokens()); return 1; } case LLAMA_MEMORY_STATUS_FAILED_COMPUTE: { - LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, batch.n_tokens); + LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, balloc->get_n_tokens()); return -2; } @@ -1005,7 +996,6 @@ int llama_context::decode(const llama_batch & batch_inp) { if (n_outputs_all == n_tokens_all) { n_outputs_new = ubatch.n_tokens; } else { - GGML_ASSERT(ubatch.output); for (uint32_t i = 0; i < ubatch.n_tokens; i++) { n_outputs_new += (int32_t) (ubatch.output[i] != 0); } @@ -1105,27 +1095,27 @@ int llama_context::decode(const llama_batch & batch_inp) { // extract sequence embeddings (cleared before processing each batch) auto & embd_seq_out = embd_seq; - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; + embd_seq_out[seq_id].resize(n_embd); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_RANK: { - // extract the rerank score - a single float per sequence + // extract the rerank score - n_cls_out floats per sequence auto & embd_seq_out = embd_seq; - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { - continue; - } - embd_seq_out[seq_id].resize(1); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float)); + const uint32_t n_cls_out = hparams.n_cls_out; + + for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) { + const llama_seq_id seq_id = ubatch.seq_id_unq[s]; + const int32_t seq_idx = ubatch.seq_idx[seq_id]; + + embd_seq_out[seq_id].resize(n_cls_out); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float)); } } break; case LLAMA_POOLING_TYPE_UNSPECIFIED: @@ -1145,7 +1135,7 @@ int llama_context::decode(const llama_batch & batch_inp) { if (n_outputs > 0) { bool sorted_output = true; - auto & out_ids = mstate->out_ids(); + auto & out_ids = balloc->get_out_ids(); GGML_ASSERT(out_ids.size() == (size_t) n_outputs); @@ -1318,8 +1308,8 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u this->n_outputs = n_outputs; - llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph - llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + llama_batch_allocr balloc(model.hparams.n_pos_per_embd()); + llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs); auto * gf = graph_init(); auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate); @@ -2039,7 +2029,12 @@ void llama_context::opt_epoch_iter( batch.logits [pos_batch] = true; } - const auto n_tokens_all = batch.n_tokens; + if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, true)) { + LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); + return; + } + + const uint32_t n_tokens_all = balloc->get_n_tokens(); n_queued_tokens += n_tokens_all; @@ -2047,7 +2042,7 @@ void llama_context::opt_epoch_iter( uint32_t n_outputs_all = n_tokens_all; - auto mstate = memory->init_batch(batch, cparams.n_ubatch, true); + auto mstate = memory->init_batch(*balloc, cparams.n_ubatch, true); if (!mstate || mstate->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) { LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__); break; diff --git a/src/llama-context.h b/src/llama-context.h index 040f03ae4..7d300c145 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -247,7 +247,7 @@ private: std::map> embd_seq; // reuse the batch_allocr to avoid unnecessary memory allocations - std::unique_ptr batch_allocr; + std::unique_ptr balloc; uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 65d98cbbb..083366fd6 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -130,110 +130,97 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) { if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { const int64_t n_tokens = ubatch->n_tokens; const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_seqs_unq = ubatch->n_seqs_unq; GGML_ASSERT(mean); GGML_ASSERT(ggml_backend_buffer_is_host(mean->buffer)); float * data = (float *) mean->data; - memset(mean->data, 0, n_tokens * n_tokens * ggml_element_size(mean)); + memset(mean->data, 0, n_tokens*n_seqs_unq*ggml_element_size(mean)); - std::vector sum(n_tokens, 0); + std::vector sums(n_seqs_unq, 0); + for (int i = 0; i < n_tokens; i += n_seq_tokens) { + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; - - // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); - - sum[seq_id] += ubatch->n_seq_tokens; - } - - std::vector div(n_tokens, 0.0f); - for (int i = 0; i < n_tokens; ++i) { - const uint64_t s = sum[i]; - if (s > 0) { - div[i] = 1.0f/float(s); + sums[seq_idx] += ubatch->n_seq_tokens; } } - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + std::vector div(n_seqs_unq, 0.0f); + for (int s = 0; s < n_seqs_unq; ++s) { + const uint64_t sum = sums[s]; + if (sum > 0) { + div[s] = 1.0f/float(sum); + } + } - for (int i = 0; i < n_seq_tokens; ++i) { - data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id]; + for (int i = 0; i < n_tokens; i += n_seq_tokens) { + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; + + for (int j = 0; j < n_seq_tokens; ++j) { + data[seq_idx*n_tokens + i + j] = div[seq_idx]; + } } } } } void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) { - if (cparams.embeddings && ( - cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || - cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) { - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs_unq = ubatch->n_seqs_unq; + if (cparams.embeddings && ( + cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || + cparams.pooling_type == LLAMA_POOLING_TYPE_RANK + )) { GGML_ASSERT(cls); GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); uint32_t * data = (uint32_t *) cls->data; - memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls)); - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + for (int i = 0; i < n_tokens; i += n_seq_tokens) { + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; - // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK"); - - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; - - if (pos == 0) { - data[seq_id] = s*n_seq_tokens + i; - } + data[seq_idx] = i; } } } if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; - GGML_ASSERT(cls); GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); uint32_t * data = (uint32_t *) cls->data; - memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls)); - std::vector last_pos(n_tokens, -1); - std::vector last_row(n_tokens, -1); + std::vector last_pos(n_seqs_unq, -1); + std::vector last_row(n_seqs_unq, -1); - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + for (int i = 0; i < n_tokens; ++i) { + const llama_pos pos = ubatch->pos[i]; - // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST"); + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + const int32_t seq_idx = ubatch->seq_idx[seq_id]; - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; - - if (pos >= last_pos[seq_id]) { - last_pos[seq_id] = pos; - last_row[seq_id] = s*n_seq_tokens + i; + if (pos >= last_pos[seq_idx]) { + last_pos[seq_idx] = pos; + last_row[seq_idx] = i; } } } - for (int i = 0; i < n_tokens; ++i) { - if (last_row[i] >= 0) { - data[i] = last_row[i]; + for (int s = 0; s < n_seqs_unq; ++s) { + if (last_row[s] >= 0) { + data[s] = last_row[s]; } } } @@ -266,89 +253,36 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) { } void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) { - if (kq_mask) { - if (cparams.causal_attn) { - const int64_t n_kv = ubatch->n_tokens; - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_kv = ubatch->n_tokens; + const int64_t n_tokens = ubatch->n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); - float * data = (float *) kq_mask->data; + GGML_ASSERT(kq_mask); + GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); - for (int h = 0; h < 1; ++h) { - for (int s1 = 0; s1 < n_seqs; ++s1) { - const llama_seq_id seq_id = ubatch->seq_id[s1][0]; + float * data = (float *) kq_mask->data; - for (int j = 0; j < n_seq_tokens; ++j) { - const int32_t tj = s1*n_seq_tokens + j; + for (int h = 0; h < 1; ++h) { + for (int i1 = 0; i1 < n_tokens; ++i1) { + const llama_seq_id s1 = ubatch->seq_id[i1][0]; - for (int s0 = 0; s0 < n_seqs; ++s0) { - for (int i = 0; i < n_seq_tokens; ++i) { - const int32_t ti = s0*n_seq_tokens + i; - float f = -INFINITY; + for (int i0 = 0; i0 < n_tokens; ++i0) { + float f = -INFINITY; - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { - if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) { - if (hparams.use_alibi) { - f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); - } else { - f = 0.0f; - } - break; - } - } + for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) { + const llama_seq_id s0 = ubatch->seq_id[i0][0]; - data[h*(n_kv*n_tokens) + tj*n_kv + ti] = f; - } + // TODO: reimplement this like in llama_kv_cache_unified + if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) { + if (hparams.use_alibi) { + f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]); + } else { + f = 0.0f; } + break; } } - } - } else { - const int64_t n_tokens = ubatch->n_tokens; - const int64_t n_seq_tokens = ubatch->n_seq_tokens; - const int64_t n_seqs = ubatch->n_seqs; - const int64_t n_stride = ubatch->n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); - - float * data = (float *) kq_mask->data; - - for (int h = 0; h < 1; ++h) { - for (int s1 = 0; s1 < n_seqs; ++s1) { - const llama_seq_id seq_id = ubatch->seq_id[s1][0]; - - for (int j = 0; j < n_seq_tokens; ++j) { - const int32_t tj = s1*n_seq_tokens + j; - - for (int s0 = 0; s0 < n_seqs; ++s0) { - for (int i = 0; i < n_seq_tokens; ++i) { - const int32_t ti = s0*n_seq_tokens + i; - float f = -INFINITY; - - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { - if (ubatch->seq_id[s0][s] == seq_id) { - if (hparams.use_alibi) { - f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); - } else { - f = 0.0f; - } - break; - } - } - - data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; - } - } - - for (int i = n_tokens; i < n_stride; ++i) { - data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; - } - } - } + data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f; } } } @@ -371,34 +305,36 @@ void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch } void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { - if (cross_kq_mask) { - const int64_t n_enc = cross_kq_mask->ne[0]; - const int64_t n_tokens = ubatch->n_tokens; + GGML_ASSERT(cross_kq_mask); - GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer)); - GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + const int64_t n_enc = cross_kq_mask->ne[0]; + const int64_t n_tokens = ubatch->n_tokens; - float * data = (float *) cross_kq_mask->data; + GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_enc; ++i) { - float f = -INFINITY; - // TODO: fix indexing [UBATCH_IDX] - for (int s = 0; s < ubatch->n_seq_id[j]; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[j][s]; - if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) { - f = 0.0f; - } + float * data = (float *) cross_kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int i = 0; i < n_tokens; ++i) { + for (int j = 0; j < n_enc; ++j) { + float f = -INFINITY; + + for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[i][s]; + + if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) { + f = 0.0f; } - data[h*(n_enc*n_tokens) + j*n_enc + i] = f; } - } - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_enc; ++j) { - data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY; - } + data[h*(n_enc*n_tokens) + i*n_enc + j] = f; + } + } + + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_enc; ++j) { + data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY; } } } @@ -467,10 +403,6 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) : res (std::make_unique()) { } -int64_t llm_graph_context::n_pos_per_embd() const { - return hparams.rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1; -} - void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const { if (cb_func) { cb_func(ubatch, cur, name, il); @@ -915,11 +847,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { } ggml_tensor * llm_graph_context::build_inp_pos() const { - auto inp = std::make_unique(n_pos_per_embd()); + auto inp = std::make_unique(hparams.n_pos_per_embd()); auto & cur = inp->pos; - cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd()); + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd()); ggml_set_input(cur); res->add_input(std::move(inp)); @@ -959,7 +891,7 @@ ggml_tensor * llm_graph_context::build_inp_mean() const { auto & cur = inp->mean; - cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, ubatch.n_seqs_unq); ggml_set_input(cur); res->add_input(std::move(inp)); @@ -972,7 +904,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const { auto & cur = inp->cls; - cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_seqs_unq); ggml_set_input(cur); res->add_input(std::move(inp)); diff --git a/src/llama-graph.h b/src/llama-graph.h index 58845e284..9e62fa607 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -95,14 +95,14 @@ public: class llm_graph_input_pos : public llm_graph_input_i { public: - llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {} + llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {} virtual ~llm_graph_input_pos() = default; void set_input(const llama_ubatch * ubatch) override; ggml_tensor * pos = nullptr; // I32 [n_batch] - const int64_t n_pos_per_embd = 1; + const uint32_t n_pos_per_embd = 1; }; // temperature tuning, used by llama4 @@ -464,8 +464,6 @@ struct llm_graph_context { llm_graph_context(const llm_graph_params & params); - int64_t n_pos_per_embd() const; - void cb(ggml_tensor * cur, const char * name, int il) const; // diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp index b40566ced..bba7a12dc 100644 --- a/src/llama-hparams.cpp +++ b/src/llama-hparams.cpp @@ -90,6 +90,10 @@ bool llama_hparams::is_recurrent(uint32_t il) const { return recurrent_layer_arr[il]; } +uint32_t llama_hparams::n_pos_per_embd() const { + return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1; +} + bool llama_hparams::is_swa(uint32_t il) const { if (il < n_layer) { return swa_layers[il]; diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 82bb5b608..7b315a9a7 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -192,6 +192,8 @@ struct llama_hparams { // whether or not the given layer is recurrent (for hybrid models) bool is_recurrent(uint32_t il) const; + uint32_t n_pos_per_embd() const; + bool is_swa(uint32_t il) const; }; diff --git a/src/llama-kv-cache-unified-iswa.cpp b/src/llama-kv-cache-unified-iswa.cpp index a869b1de8..0ced340de 100644 --- a/src/llama-kv-cache-unified-iswa.cpp +++ b/src/llama-kv-cache-unified-iswa.cpp @@ -95,19 +95,22 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const { return kv_swa->seq_pos_max(seq_id); } -llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) { +llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { GGML_UNUSED(embd_all); // first try simple split do { - auto sbatch = llama_sbatch(batch, hparams.n_embd, true); + balloc.split_reset(); std::vector ubatches; + while (true) { + auto ubatch = balloc.split_simple(n_ubatch); - while (sbatch.n_tokens > 0) { - auto ubatch = sbatch.split_simple(n_ubatch); + if (ubatch.n_tokens == 0) { + break; + } - ubatches.push_back(ubatch); + ubatches.push_back(std::move(ubatch)); // NOLINT } auto heads_base = kv_base->prepare(ubatches); @@ -123,19 +126,22 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch assert(heads_base.size() == heads_swa.size()); return std::make_unique( - this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches)); + this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); } while (false); // if it fails, try equal split do { - auto sbatch = llama_sbatch(batch, hparams.n_embd, false); + balloc.split_reset(); std::vector ubatches; + while (true) { + auto ubatch = balloc.split_equal(n_ubatch); - while (sbatch.n_tokens > 0) { - auto ubatch = sbatch.split_equal(n_ubatch); + if (ubatch.n_tokens == 0) { + break; + } - ubatches.push_back(ubatch); + ubatches.push_back(std::move(ubatch)); // NOLINT } auto heads_base = kv_base->prepare(ubatches); @@ -151,7 +157,7 @@ llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch assert(heads_base.size() == heads_swa.size()); return std::make_unique( - this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches)); + this, std::move(heads_base), std::move(heads_swa), std::move(ubatches)); } while (false); // TODO: if we fail again, we should attempt different splitting strategies @@ -214,15 +220,13 @@ llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state( llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state( llama_kv_cache_unified_iswa * kv, - llama_sbatch sbatch, std::vector heads_base, std::vector heads_swa, std::vector ubatches) : - sbatch(std::move(sbatch)), ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - state_base(new llama_kv_cache_unified_state(kv->get_base(), {}, std::move(heads_base), this->ubatches)), - state_swa (new llama_kv_cache_unified_state(kv->get_swa (), {}, std::move(heads_swa), this->ubatches)), + state_base(new llama_kv_cache_unified_state(kv->get_base(), std::move(heads_base), this->ubatches)), + state_swa (new llama_kv_cache_unified_state(kv->get_swa (), std::move(heads_swa), this->ubatches)), status(llama_memory_status_combine(state_base->get_status(), state_swa->get_status())) { } @@ -252,12 +256,6 @@ bool llama_kv_cache_unified_iswa_state::apply() { return res; } -std::vector & llama_kv_cache_unified_iswa_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_kv_cache_unified_iswa_state::get_status() const { return status; } diff --git a/src/llama-kv-cache-unified-iswa.h b/src/llama-kv-cache-unified-iswa.h index 813eaf39b..071041585 100644 --- a/src/llama-kv-cache-unified-iswa.h +++ b/src/llama-kv-cache-unified-iswa.h @@ -32,7 +32,7 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; @@ -90,7 +90,6 @@ public: // used to create a state from a batch llama_kv_cache_unified_iswa_state( llama_kv_cache_unified_iswa * kv, - llama_sbatch sbatch, std::vector heads_base, std::vector heads_swa, std::vector ubatches); @@ -104,8 +103,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -119,8 +116,6 @@ public: private: //llama_kv_cache_unified_iswa * kv; - llama_sbatch sbatch; - // the index of the next ubatch to process size_t i_next = 0; diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache-unified.cpp index d44122889..6897b7971 100644 --- a/src/llama-kv-cache-unified.cpp +++ b/src/llama-kv-cache-unified.cpp @@ -308,17 +308,23 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const { } llama_memory_state_ptr llama_kv_cache_unified::init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { GGML_UNUSED(embd_all); do { - auto sbatch = llama_sbatch(batch, hparams.n_embd, true); + balloc.split_reset(); std::vector ubatches; - while (sbatch.n_tokens > 0) { - ubatches.push_back(sbatch.split_simple(n_ubatch)); + while (true) { + auto ubatch = balloc.split_simple(n_ubatch); + + if (ubatch.n_tokens == 0) { + break; + } + + ubatches.push_back(std::move(ubatch)); // NOLINT } auto heads = prepare(ubatches); @@ -327,7 +333,7 @@ llama_memory_state_ptr llama_kv_cache_unified::init_batch( } return std::make_unique( - this, std::move(sbatch), std::move(heads), std::move(ubatches)); + this, std::move(heads), std::move(ubatches)); } while (false); return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); @@ -644,12 +650,6 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const { } void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) { - if (debug > 0) { - LLAMA_LOG_DEBUG("%s: ubatch info:\n", __func__); - LLAMA_LOG_DEBUG("%s: n_tokens = %d, equal_seqs = %d\n", __func__, ubatch.n_tokens, ubatch.equal_seqs); - LLAMA_LOG_DEBUG("%s: n_seq_tokens = %d, n_seqs = %d\n", __func__, ubatch.n_seq_tokens, ubatch.n_seqs); - } - // keep track of the max sequence position that we would overwrite with this ubatch // for non-SWA cache, this would be always empty llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ]; @@ -657,27 +657,22 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch seq_pos_max_rm[s] = -1; } - for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { - for (uint32_t j = 0; j < ubatch.n_seq_tokens; ++j) { - const uint32_t idx = s*ubatch.n_seq_tokens + j; + for (uint32_t i = 0; i < ubatch.n_tokens; ++i) { + if (!cells.is_empty(head_cur + i)) { + assert(cells.seq_count(head_cur + i) == 1); - if (!cells.is_empty(head_cur + idx)) { - assert(cells.seq_count(head_cur + idx) == 1); + const llama_seq_id seq_id = cells.seq_get(head_cur + i); + const llama_pos pos = cells.pos_get(head_cur + i); - const llama_seq_id seq_id = cells.seq_get(head_cur + idx); - const llama_pos pos = cells.pos_get(head_cur + idx); + seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos); - seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos); + cells.rm(head_cur + i); + } - cells.rm(head_cur + idx); - } + cells.pos_set(head_cur + i, ubatch.pos[i]); - cells.pos_set(head_cur + idx, ubatch.pos[idx]); - - // TODO: fix indexing [UBATCH_IDX] - for (int32_t i = 0; i < ubatch.n_seq_id[s]; i++) { - cells.seq_add(head_cur + idx, ubatch.seq_id[s][i]); - } + for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) { + cells.seq_add(head_cur + i, ubatch.seq_id[i][s]); } } @@ -696,6 +691,7 @@ void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch seq_rm(s, cells.seq_pos_min(s), seq_pos_max_rm[s] + 1); } } + // move the head at the end of the slot head = head_cur + ubatch.n_tokens; } @@ -792,9 +788,7 @@ ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_ } void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const { - const uint32_t n_tokens = ubatch->n_tokens; - const uint32_t n_seq_tokens = ubatch->n_seq_tokens; - const uint32_t n_seqs = ubatch->n_seqs; + const uint32_t n_tokens = ubatch->n_tokens; GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer)); float * data = (float *) dst->data; @@ -814,52 +808,48 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub // xxxxx----- // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615 for (uint32_t h = 0; h < 1; ++h) { - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch->seq_id[s][0]; + for (uint32_t i = 0; i < n_tokens; ++i) { + const llama_seq_id seq_id = ubatch->seq_id[i][0]; - for (uint32_t j = 0; j < n_seq_tokens; ++j) { - const uint32_t idx = s*n_seq_tokens + j; + const llama_pos p1 = ubatch->pos[i]; - const llama_pos p1 = ubatch->pos[idx]; + for (uint32_t j = 0; j < n_kv; ++j) { + float f = 0.0f; - for (uint32_t i = 0; i < n_kv; ++i) { - float f = 0.0f; + bool masked = false; - bool masked = false; + if (cells.is_empty(j)) { + masked = true; + } else { + const llama_pos p0 = cells.pos_get(j); - if (cells.is_empty(i)) { - masked = true; - } else { - const llama_pos p0 = cells.pos_get(i); + // mask the token if not the same sequence + masked = masked || (!cells.seq_has(j, seq_id)); - // mask the token if not the same sequence - masked = masked || (!cells.seq_has(i, seq_id)); + // mask future tokens + masked = masked || (causal_attn && p0 > p1); - // mask future tokens - masked = masked || (causal_attn && p0 > p1); + // apply SWA if any + masked = masked || (is_masked_swa(p0, p1)); - // apply SWA if any - masked = masked || (is_masked_swa(p0, p1)); - - if (!masked && hparams.use_alibi) { - f = -std::abs(p0 - p1); - } + if (!masked && hparams.use_alibi) { + f = -std::abs(p0 - p1); } - - if (masked) { - f = -INFINITY; - } - - data[h*(n_kv*n_tokens) + idx*n_kv + i] = f; } + + if (masked) { + f = -INFINITY; + } + + data[h*(n_kv*n_tokens) + i*n_kv + j] = f; } } // mask padded tokens if (data) { - for (uint32_t j = n_tokens; j < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++j) { - for (uint32_t i = 0; i < n_kv; ++i) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; + for (uint32_t i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (uint32_t j = 0; j < n_kv; ++j) { + data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; } } } @@ -887,12 +877,12 @@ void llama_kv_cache_unified::set_input_pos_bucket(ggml_tensor * dst, const llama const int32_t n_kv = dst->ne[0]; for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_kv; ++i) { + for (int i = 0; i < n_tokens; ++i) { + for (int j = 0; j < n_kv; ++j) { // the position when the cells is empty is irrelevant - it will be masked out later in the attention - const llama_pos p0 = cells.is_empty(i) ? -1 : cells.pos_get(i); + const llama_pos p0 = cells.is_empty(j) ? -1 : cells.pos_get(j); - data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(p0, ubatch->pos[j], hparams.n_rel_attn_bkts, false); + data[h*(n_kv*n_tokens) + i*n_kv + j] = llama_relative_position_bucket(p0, ubatch->pos[i], hparams.n_rel_attn_bkts, false); } } } @@ -1509,12 +1499,9 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell seq_rm(dest_seq_id, -1, -1); - llama_sbatch sbatch; - llama_ubatch ubatch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false); + llama_batch_allocr balloc(hparams.n_pos_per_embd()); - ubatch.n_tokens = cell_count; - ubatch.n_seq_tokens = cell_count; - ubatch.n_seqs = 1; + llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1); for (uint32_t i = 0; i < cell_count; ++i) { llama_pos pos; @@ -1746,9 +1733,8 @@ llama_kv_cache_unified_state::llama_kv_cache_unified_state( llama_kv_cache_unified_state::llama_kv_cache_unified_state( llama_kv_cache_unified * kv, - llama_sbatch sbatch, llama_kv_cache_unified::ubatch_heads heads, - std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sbatch(std::move(sbatch)), heads(std::move(heads)), ubatches(std::move(ubatches)) { + std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), heads(std::move(heads)), ubatches(std::move(ubatches)) { } llama_kv_cache_unified_state::~llama_kv_cache_unified_state() = default; @@ -1781,12 +1767,6 @@ bool llama_kv_cache_unified_state::apply() { return true; } -std::vector & llama_kv_cache_unified_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_kv_cache_unified_state::get_status() const { return status; } diff --git a/src/llama-kv-cache-unified.h b/src/llama-kv-cache-unified.h index d96571d95..156064004 100644 --- a/src/llama-kv-cache-unified.h +++ b/src/llama-kv-cache-unified.h @@ -57,7 +57,7 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; @@ -231,7 +231,6 @@ public: // used to create a decode state from a batch llama_kv_cache_unified_state( llama_kv_cache_unified * kv, - llama_sbatch sbatch, ubatch_heads heads, std::vector ubatches); @@ -244,8 +243,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -286,8 +283,6 @@ private: // batch processing state // - llama_sbatch sbatch; - // the index of the next ubatch to process size_t i_next = 0; diff --git a/src/llama-kv-cells.h b/src/llama-kv-cells.h index 1d4e70f4d..349e9032e 100644 --- a/src/llama-kv-cells.h +++ b/src/llama-kv-cells.h @@ -384,10 +384,10 @@ private: // std::vector shift; - using bits_t = std::bitset; + using seq_set_t = std::bitset; // the bitset seq[i] tells us which sequences are currently occupying the i-th cell - std::vector seq; + std::vector seq; // the set seq_pos[s] tells us which positions are currently present for sequence s // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp index d4b260db4..1b1668681 100644 --- a/src/llama-memory-hybrid.cpp +++ b/src/llama-memory-hybrid.cpp @@ -32,7 +32,7 @@ llama_memory_hybrid::llama_memory_hybrid( mem_attn(new llama_kv_cache_unified( model, filter_attn == nullptr ? - [&](int32_t il) { return !model.hparams.is_recurrent(il); } + [&](int32_t il) { return !hparams.is_recurrent(il); } : filter_attn, type_k, type_v, @@ -47,7 +47,7 @@ llama_memory_hybrid::llama_memory_hybrid( mem_recr(new llama_memory_recurrent( model, filter_recr == nullptr ? - [&](int32_t il) { return model.hparams.is_recurrent(il); } + [&](int32_t il) { return hparams.is_recurrent(il); } : filter_recr, type_r, type_s, @@ -56,42 +56,49 @@ llama_memory_hybrid::llama_memory_hybrid( n_seq_max )) {} -llama_memory_state_ptr llama_memory_hybrid::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled) { +llama_memory_state_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { + do { + balloc.split_reset(); - // since this includes a recurrent cache, we cannot use split_simple - auto sbatch = llama_sbatch(batch, hparams.n_embd, false); + // follow the recurrent pattern for creating the ubatch splits + std::vector ubatches; - // follow the recurrent pattern for creating the ubatch splits - std::vector ubatches; - while (sbatch.n_tokens > 0) { - llama_ubatch ubatch; + while (true) { + llama_ubatch ubatch; - if (embd_pooled) { - // Pooled embeddings cannot be split across ubatches (yet) - ubatch = sbatch.split_seq(n_ubatch); - } else { - ubatch = sbatch.split_equal(n_ubatch); + if (embd_all) { + // if all tokens are output, split by sequence + ubatch = balloc.split_seq(n_ubatch); + } else { + ubatch = balloc.split_equal(n_ubatch); + } + + if (ubatch.n_tokens == 0) { + break; + } + + ubatches.push_back(std::move(ubatch)); // NOLINT } - ubatches.push_back(ubatch); - } + // prepare the recurrent batches first + if (!mem_recr->prepare(ubatches)) { + // TODO: will the recurrent cache be in an undefined state at this point? + LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + } - // prepare the recurrent batches first - if (!mem_recr->prepare(ubatches)) { - // TODO: will the recurrent cache be in an undefined state at this point? - LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); - } + // prepare the attention cache + auto heads_attn = mem_attn->prepare(ubatches); + if (heads_attn.empty()) { + LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); + } - // prepare the attention cache - auto heads_attn = mem_attn->prepare(ubatches); - if (heads_attn.empty()) { - LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__); - return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); - } + return std::make_unique( + this, std::move(heads_attn), std::move(ubatches)); + } while(false); - return std::make_unique( - this, std::move(sbatch), std::move(heads_attn), std::move(ubatches)); + return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } llama_memory_state_ptr llama_memory_hybrid::init_full() { @@ -188,15 +195,13 @@ llama_memory_hybrid_state::llama_memory_hybrid_state( llama_memory_hybrid_state::llama_memory_hybrid_state( llama_memory_hybrid * mem, - llama_sbatch sbatch, std::vector heads_attn, std::vector ubatches) : - sbatch(std::move(sbatch)), ubatches(std::move(ubatches)), // note: here we copy the ubatches. not sure if this is ideal - state_attn(new llama_kv_cache_unified_state(mem->get_mem_attn(), {}, std::move(heads_attn), this->ubatches)), - state_recr(new llama_memory_recurrent_state(mem->get_mem_recr(), {}, this->ubatches)), - status(LLAMA_MEMORY_STATUS_SUCCESS) { + state_attn(new llama_kv_cache_unified_state(mem->get_mem_attn(), std::move(heads_attn), this->ubatches)), + state_recr(new llama_memory_recurrent_state(mem->get_mem_recr(), this->ubatches)), + status(llama_memory_status_combine(state_attn->get_status(), state_recr->get_status())) { } bool llama_memory_hybrid_state::next() { @@ -223,12 +228,6 @@ bool llama_memory_hybrid_state::apply() { return res; } -std::vector & llama_memory_hybrid_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_memory_hybrid_state::get_status() const { return status; } diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h index b5700c522..4d27ab896 100644 --- a/src/llama-memory-hybrid.h +++ b/src/llama-memory-hybrid.h @@ -50,9 +50,9 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, - bool embd_pooled) override; + bool embd_all) override; llama_memory_state_ptr init_full() override; @@ -107,7 +107,6 @@ public: // init success llama_memory_hybrid_state( llama_memory_hybrid * mem, - llama_sbatch sbatch, std::vector heads_attn, std::vector ubatches); @@ -116,8 +115,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -129,8 +126,6 @@ public: const llama_memory_recurrent_state * get_state_recr() const; private: - llama_sbatch sbatch; - // the index of the next ubatch to process size_t i_next = 0; diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp index c4f9a6f1d..b064da008 100644 --- a/src/llama-memory-recurrent.cpp +++ b/src/llama-memory-recurrent.cpp @@ -362,29 +362,31 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const { return result; } -llama_memory_state_ptr llama_memory_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_all) { - auto sbatch = llama_sbatch(batch, hparams.n_embd, false); - +llama_memory_state_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) { std::vector ubatches; - while (sbatch.n_tokens > 0) { + while (true) { llama_ubatch ubatch; if (embd_all) { // if all tokens are output, split by sequence - ubatch = sbatch.split_seq(n_ubatch); + ubatch = balloc.split_seq(n_ubatch); } else { - ubatch = sbatch.split_equal(n_ubatch); + ubatch = balloc.split_equal(n_ubatch); } - ubatches.push_back(ubatch); + if (ubatch.n_tokens == 0) { + break; + } + + ubatches.push_back(std::move(ubatch)); // NOLINT } if (!prepare(ubatches)) { return std::make_unique(LLAMA_MEMORY_STATUS_FAILED_PREPARE); } - return std::make_unique(this, std::move(sbatch), std::move(ubatches)); + return std::make_unique(this, std::move(ubatches)); } llama_memory_state_ptr llama_memory_recurrent::init_full() { @@ -423,9 +425,8 @@ bool llama_memory_recurrent::prepare(const std::vector & ubatches) } bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { - const uint32_t n_seqs = ubatch.n_seqs; - const uint32_t n_seq_tokens = ubatch.n_seq_tokens; + const uint32_t n_seqs = ubatch.n_seqs; // if we have enough unused cells before the current head -> // better to start searching from the beginning of the cache, hoping to fill it @@ -445,9 +446,11 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // everything should fit if all seq_ids are smaller than the max for (uint32_t s = 0; s < n_seqs; ++s) { - const uint32_t n_seq_id = ubatch.n_seq_id[s]; + const uint32_t i = s*n_seq_tokens; // first token of sequence set s + const uint32_t n_seq_id = ubatch.n_seq_id[i]; + for (uint32_t j = 0; j < n_seq_id; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; + const llama_seq_id seq_id = ubatch.seq_id[i][j]; if (seq_id < 0 || (uint32_t) seq_id >= size) { // too big seq_id @@ -506,7 +509,8 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // find usable cell range for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; + const uint32_t i = s*n_seq_tokens; + const llama_seq_id seq_id = ubatch.seq_id[i][0]; auto & seq_meta = cells[seq_id]; bool has_cell = false; if (seq_meta.tail >= 0) { @@ -530,7 +534,7 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { seq_meta.tail = next_empty_cell; // find next empty cell if (s + 1 < n_seqs) { - for (uint32_t i = 0; i < size; ++i) { + for (uint32_t j = 0; j < size; ++j) { next_empty_cell += 1; if (next_empty_cell >= size) { next_empty_cell -= size; } auto & cell = cells[next_empty_cell]; @@ -544,8 +548,9 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // gather and re-order for (uint32_t s = 0; s < n_seqs; ++s) { + const uint32_t i = s*n_seq_tokens; const int32_t dst_id = s + min; - const int32_t src_id = cells[ubatch.seq_id[s][0]].tail; + const int32_t src_id = cells[ubatch.seq_id[i][0]].tail; if (dst_id != src_id) { auto & dst_cell = cells[dst_id]; auto & src_cell = cells[src_id]; @@ -555,8 +560,8 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { std::swap(dst_cell.seq_id, src_cell.seq_id); // swap tails - for (uint32_t i = 0; i < size; ++i) { - int32_t & tail = cells[i].tail; + for (uint32_t j = 0; j < size; ++j) { + int32_t & tail = cells[j].tail; if (tail == src_id) { tail = dst_id; } else if (tail == dst_id) { @@ -568,7 +573,8 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // update the pos of the used seqs for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; + const uint32_t i = s*n_seq_tokens; + const llama_pos last_pos = ubatch.pos[i + n_seq_tokens - 1]; const int32_t cell_id = s + min; auto & cell = cells[cell_id]; @@ -576,12 +582,12 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) { // What should happen when the pos backtracks or skips a value? // Clearing the state mid-batch would require special-casing which isn't done. LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", - __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); + __func__, last_pos, cell.pos, ubatch.seq_id[i][0], n_seq_tokens); } cell.pos = last_pos; cell.seq_id.clear(); - for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; + for (int32_t j = 0; j < ubatch.n_seq_id[i]; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[i][j]; cell.seq_id.insert(seq_id); cells[seq_id].tail = cell_id; } @@ -827,12 +833,9 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell seq_rm(dest_seq_id, -1, -1); - llama_sbatch sbatch; - llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false); + llama_batch_allocr balloc(hparams.n_pos_per_embd()); - batch.n_tokens = cell_count; - batch.n_seq_tokens = cell_count; - batch.n_seqs = 1; + llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1); for (uint32_t i = 0; i < cell_count; ++i) { llama_pos pos; @@ -846,12 +849,12 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell return false; } - batch.pos[i] = pos; + ubatch.pos[i] = pos; } - batch.n_seq_id[0] = 1; - batch.seq_id[0] = &dest_seq_id; + ubatch.n_seq_id[0] = 1; + ubatch.seq_id[0] = &dest_seq_id; - if (!find_slot(batch)) { + if (!find_slot(ubatch)) { LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); return false; } @@ -859,8 +862,8 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell // DEBUG CHECK: kv.head should be our first cell, kv.head + cell_count - 1 should be our last cell (verify seq_id and pos values) // Assume that this is one contiguous block of cells GGML_ASSERT(head + cell_count <= size); - GGML_ASSERT(cells[head].pos == batch.pos[0]); - GGML_ASSERT(cells[head + cell_count - 1].pos == batch.pos[cell_count - 1]); + GGML_ASSERT(cells[head].pos == ubatch.pos[0]); + GGML_ASSERT(cells[head + cell_count - 1].pos == ubatch.pos[cell_count - 1]); GGML_ASSERT(cells[head].has_seq_id(dest_seq_id)); GGML_ASSERT(cells[head + cell_count - 1].has_seq_id(dest_seq_id)); } else { @@ -1048,8 +1051,7 @@ llama_memory_recurrent_state::llama_memory_recurrent_state( llama_memory_recurrent_state::llama_memory_recurrent_state( llama_memory_recurrent * mem, - llama_sbatch sbatch, - std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), sbatch(std::move(sbatch)), ubatches(std::move(ubatches)) {} + std::vector ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), ubatches(std::move(ubatches)) {} llama_memory_recurrent_state::~llama_memory_recurrent_state() = default; @@ -1071,12 +1073,6 @@ bool llama_memory_recurrent_state::apply() { return true; } -std::vector & llama_memory_recurrent_state::out_ids() { - assert(status == LLAMA_MEMORY_STATUS_SUCCESS); - - return sbatch.out_ids; -} - llama_memory_status llama_memory_recurrent_state::get_status() const { return status; } diff --git a/src/llama-memory-recurrent.h b/src/llama-memory-recurrent.h index 290cc84ab..be58dae7c 100644 --- a/src/llama-memory-recurrent.h +++ b/src/llama-memory-recurrent.h @@ -35,7 +35,7 @@ public: // llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) override; @@ -137,7 +137,6 @@ public: // used to create a state from a batch llama_memory_recurrent_state( llama_memory_recurrent * mem, - llama_sbatch sbatch, std::vector ubatches); virtual ~llama_memory_recurrent_state(); @@ -149,8 +148,6 @@ public: bool next() override; bool apply() override; - std::vector & out_ids() override; - llama_memory_status get_status() const override; const llama_ubatch & get_ubatch() const override; @@ -173,8 +170,6 @@ private: llama_memory_recurrent * mem; - llama_sbatch sbatch; - size_t i_next = 0; std::vector ubatches; diff --git a/src/llama-memory.h b/src/llama-memory.h index 24668f861..d2ef0c2a3 100644 --- a/src/llama-memory.h +++ b/src/llama-memory.h @@ -7,6 +7,8 @@ struct llama_ubatch; +class llama_batch_allocr; + class llama_io_write_i; class llama_io_read_i; @@ -50,9 +52,6 @@ struct llama_memory_state_i { // return false on failure virtual bool apply() = 0; - // TODO: this might get reworked in the future when refactoring llama_batch - virtual std::vector & out_ids() = 0; - // get the current ubatch virtual const llama_ubatch & get_ubatch() const = 0; @@ -71,7 +70,7 @@ struct llama_memory_i { // return a state object containing the ubatches and KV cache state required to process them // check the llama_memory_state_i::get_status() for the result virtual llama_memory_state_ptr init_batch( - const llama_batch & batch, + llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) = 0; diff --git a/tools/server/server.cpp b/tools/server/server.cpp index 9d55b3338..aa18513e3 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -3385,38 +3385,6 @@ struct server_context { llama_set_embeddings(ctx, slot_batched->need_embd()); } - // pad the batch so that batch.n_tokens >= n_slots - // TODO: temporary workaround for https://github.com/ggml-org/llama.cpp/issues/13689 - if (slot_batched->need_embd()) { - const int n_slots = slots.size(); - - if (batch.n_tokens < n_slots) { - std::set seq_ids; - for (int j = 0; j < batch.n_tokens; ++j) { - seq_ids.insert(batch.seq_id[j][0]); - } - - // find unused sequence id - llama_seq_id seq_id = -1; - for (int i = 0; i < n_slots; ++i) { - if (seq_ids.find(i) == seq_ids.end()) { - seq_id = i; - } - } - - const int n_add = n_slots - batch.n_tokens; - - SRV_WRN("adding %d dummy tokens to the batch, seq_id = %d\n", n_add, seq_id); - - for (int j = 0; j < n_add; ++j) { - common_batch_add(batch, 0, j, { seq_id }, true); - } - - slots[seq_id].cache_tokens.clear(); - llama_memory_seq_rm(llama_get_memory(ctx), seq_id, -1, -1); - } - } - int32_t i_next = 0; // process the created batch of tokens From 812939a9e90f99d1bd5bb1bc6b99d12600671d50 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 20 Jun 2025 10:50:27 +0300 Subject: [PATCH 6/7] model : more uniform output id handling (#14275) * model : more uniform output id handling ggml-ci * cont : revert n_outputs < n_tokens optimization ggml-ci * cont : fix out_ids initialization ggml-ci --- src/llama-graph.cpp | 54 +-- src/llama-model.cpp | 847 ++++++++++++++++++++++---------------------- 2 files changed, 459 insertions(+), 442 deletions(-) diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 083366fd6..7e162c555 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -92,36 +92,28 @@ void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) { } void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) { - if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) { - //GGML_ASSERT(out_ids && "every model that can must skip unused outputs"); + GGML_ASSERT(out_ids); - if (!out_ids) { - LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__); - } else { - const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_tokens = ubatch->n_tokens; - GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); - int32_t * data = (int32_t *) out_ids->data; + GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); + int32_t * data = (int32_t *) out_ids->data; - if (n_outputs == n_tokens) { - for (int i = 0; i < n_tokens; ++i) { - data[i] = i; - } - } else if (ubatch->output) { - int32_t n_outputs = 0; - for (int i = 0; i < n_tokens; ++i) { - if (ubatch->output[i]) { - data[n_outputs++] = i; - } - } - // the graph needs to have been passed the correct number of outputs - GGML_ASSERT(n_outputs == n_outputs); - } else if (n_outputs == 1) { - // only keep last output - data[0] = n_tokens - 1; - } else { - GGML_ASSERT(n_outputs == 0); - } + if (n_outputs == n_tokens) { + for (int i = 0; i < n_tokens; ++i) { + data[i] = i; + } + + return; + } + + GGML_ASSERT(ubatch->output); + + int n_outputs = 0; + + for (int i = 0; i < n_tokens; ++i) { + if (ubatch->output[i]) { + data[n_outputs++] = i; } } } @@ -874,6 +866,14 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const { } ggml_tensor * llm_graph_context::build_inp_out_ids() const { + // note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls, + // but this would make the graph topology depend on the number of output tokens, which can interere with + // features that require constant topology such as pipline parallelism + // ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471 + //if (n_outputs < n_tokens) { + // return nullptr; + //} + auto inp = std::make_unique(hparams, cparams, n_outputs); auto & cur = inp->out_ids; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index a5853f8b1..e2c82017f 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -4707,6 +4707,8 @@ struct llm_build_llama : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -4769,9 +4771,7 @@ struct llm_build_llama : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -4867,6 +4867,8 @@ struct llm_build_llama_iswa : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -4943,9 +4945,7 @@ struct llm_build_llama_iswa : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5045,6 +5045,9 @@ struct llm_build_deci : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; const int64_t n_head_kv = hparams.n_head_kv(il); @@ -5118,9 +5121,7 @@ struct llm_build_deci : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5199,6 +5200,8 @@ struct llm_build_baichuan : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5250,9 +5253,7 @@ struct llm_build_baichuan : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5321,6 +5322,8 @@ struct llm_build_xverse : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5365,9 +5368,7 @@ struct llm_build_xverse : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5435,6 +5436,8 @@ struct llm_build_falcon : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * attn_norm; @@ -5490,9 +5493,7 @@ struct llm_build_falcon : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); @@ -5561,6 +5562,8 @@ struct llm_build_grok : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5620,9 +5623,7 @@ struct llm_build_grok : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5721,6 +5722,8 @@ struct llm_build_dbrx : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5771,9 +5774,7 @@ struct llm_build_dbrx : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -5853,6 +5854,8 @@ struct llm_build_starcoder : public llm_graph_context { inpL = ggml_add(ctx0, inpL, pos); cb(inpL, "inpL", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -5885,9 +5888,7 @@ struct llm_build_starcoder : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -5952,6 +5953,8 @@ struct llm_build_refact : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -5984,9 +5987,7 @@ struct llm_build_refact : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -6072,78 +6073,79 @@ struct llm_build_bert : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); - // iterate layers + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * cur = inpL; - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; - // self-attention - if (model.layers[il].wqkv) { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); + // self-attention + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); - if (model.layers[il].bqkv) { - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); } - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - } else { - Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); - Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); - Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + } + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // RoPE + if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); } - if (model.layers[il].attn_q_norm) { - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, il); - } - - if (model.layers[il].attn_k_norm) { - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - // RoPE - if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, gf, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); - - if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6240,56 +6242,57 @@ struct llm_build_neo_bert : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); - // iterate layers + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * cur = inpL; - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; - // pre-norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - // self-attention - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + // self-attention + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - // RoPE - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + // RoPE + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cur = build_attn(inp_attn, gf, - model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); - if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = build_attn(inp_attn, gf, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6354,6 +6357,8 @@ struct llm_build_bloom : public llm_graph_context { LLM_NORM, -1); cb(inpL, "inp_norm", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -6386,9 +6391,7 @@ struct llm_build_bloom : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6465,6 +6468,8 @@ struct llm_build_mpt : public llm_graph_context { cb(inpL, "inpL", -1); } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * attn_norm; @@ -6527,9 +6532,7 @@ struct llm_build_mpt : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -6598,6 +6601,8 @@ struct llm_build_stablelm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -6673,9 +6678,7 @@ struct llm_build_stablelm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); @@ -6750,6 +6753,8 @@ struct llm_build_qwen : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -6796,9 +6801,7 @@ struct llm_build_qwen : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -6867,6 +6870,8 @@ struct llm_build_qwen2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -6916,9 +6921,7 @@ struct llm_build_qwen2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -6988,6 +6991,8 @@ struct llm_build_qwen2vl : public llm_graph_context { int sections[4]; std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7037,9 +7042,7 @@ struct llm_build_qwen2vl : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7106,6 +7109,8 @@ struct llm_build_qwen2moe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7164,9 +7169,7 @@ struct llm_build_qwen2moe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7265,6 +7268,8 @@ struct llm_build_qwen3 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7317,9 +7322,7 @@ struct llm_build_qwen3 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7386,6 +7389,8 @@ struct llm_build_qwen3moe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -7438,9 +7443,7 @@ struct llm_build_qwen3moe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -7516,6 +7519,8 @@ struct llm_build_phi2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { attn_norm_output = build_norm(inpL, model.layers[il].attn_norm, @@ -7578,9 +7583,7 @@ struct llm_build_phi2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); @@ -7652,6 +7655,8 @@ struct llm_build_phi3 : public llm_graph_context { inp_attn = build_attn_inp_kv_unified(); } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { auto * residual = inpL; @@ -7715,9 +7720,7 @@ struct llm_build_phi3 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor* inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); residual = ggml_get_rows(ctx0, residual, inp_out_ids); } @@ -7803,15 +7806,16 @@ struct llm_build_plamo : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_norm", il); - ggml_tensor * attention_norm = cur; + ggml_tensor * sa_inp = cur; // self-attention { @@ -7849,18 +7853,17 @@ struct llm_build_plamo : public llm_graph_context { model.layers[il].wo, NULL, Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - ggml_tensor * sa_out = cur; - cur = attention_norm; - - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); - sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids); + sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } + ggml_tensor * sa_out = cur; + + cur = sa_inp; + // feed-forward network { cur = build_ffn(cur, @@ -7925,6 +7928,8 @@ struct llm_build_gpt2 : public llm_graph_context { inpL = ggml_add(ctx0, inpL, pos); cb(inpL, "inpL", -1); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -7957,9 +7962,7 @@ struct llm_build_gpt2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8029,6 +8032,8 @@ struct llm_build_codeshell : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -8073,9 +8078,7 @@ struct llm_build_codeshell : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8129,128 +8132,128 @@ struct llm_build_codeshell : public llm_graph_context { struct llm_build_orion : public llm_graph_context { llm_build_orion(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); - ggml_tensor * cur; - ggml_tensor * inpL; + ggml_tensor * cur; + ggml_tensor * inpL; - inpL = build_inp_embd(model.tok_embd); + inpL = build_inp_embd(model.tok_embd); - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); - auto * inp_attn = build_attn_inp_kv_unified(); + auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; + ggml_tensor * inp_out_ids = build_inp_out_ids(); - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - // if (model.layers[il].bq) { - // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - // cb(Qcur, "Qcur", il); - // } + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - // if (model.layers[il].bk) { - // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - // cb(Kcur, "Kcur", il); - // } + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + // if (model.layers[il].bq) { + // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + // cb(Qcur, "Qcur", il); + // } - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - // if (model.layers[il].bv) { - // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - // cb(Vcur, "Vcur", il); - // } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + // if (model.layers[il].bk) { + // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + // cb(Kcur, "Kcur", il); + // } - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + // if (model.layers[il].bv) { + // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + // cb(Vcur, "Vcur", il); + // } - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); - cur = build_attn(inp_attn, gf, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } + cur = inpL; - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); + cb(cur, "result_norm", -1); + res->t_embd = cur; - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); + // lm_head + cur = build_lora_mm(model.output, cur); - cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "result_output", -1); + res->t_logits = cur; - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); + ggml_build_forward_expand(gf, cur); } }; @@ -8271,6 +8274,8 @@ struct llm_build_internlm2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -8329,9 +8334,7 @@ struct llm_build_internlm2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -8407,6 +8410,8 @@ struct llm_build_minicpm3 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -8526,15 +8531,13 @@ struct llm_build_minicpm3 : public llm_graph_context { q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } // scale_res - scale the hidden states for residual connection - const float scale_res = scale_depth/sqrtf(float(n_layer)); + const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct? cur = ggml_scale(ctx0, cur, scale_res); cb(cur, "hidden_scaled", il); @@ -8611,6 +8614,8 @@ struct llm_build_gemma : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -8656,9 +8661,7 @@ struct llm_build_gemma : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -8727,6 +8730,8 @@ struct llm_build_gemma2_iswa : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -8771,18 +8776,16 @@ struct llm_build_gemma2_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); cb(sa_out, "sa_out", il); @@ -8861,6 +8864,8 @@ struct llm_build_gemma3_iswa : public llm_graph_context { // TODO: is causal == true correct? might need some changes auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const float freq_base_l = model.get_rope_freq_base (cparams, il); const float freq_scale_l = model.get_rope_freq_scale(cparams, il); @@ -8913,18 +8918,16 @@ struct llm_build_gemma3_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); cb(sa_out, "sa_out", il); @@ -8995,6 +8998,8 @@ struct llm_build_starcoder2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9053,9 +9058,7 @@ struct llm_build_starcoder2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -9118,6 +9121,8 @@ struct llm_build_mamba : public llm_graph_context { auto * rs_inp = build_rs_inp(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, @@ -9127,9 +9132,7 @@ struct llm_build_mamba : public llm_graph_context { cur = build_mamba_layer(rs_inp, gf, cur, ubatch, il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -9311,13 +9314,15 @@ struct llm_build_command_r : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); - for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { // norm cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; // self-attention @@ -9385,9 +9390,7 @@ struct llm_build_command_r : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); @@ -9458,6 +9461,8 @@ struct llm_build_cohere2_iswa : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified_iswa(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const bool is_swa = hparams.is_swa(il); @@ -9520,9 +9525,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); @@ -9593,6 +9596,8 @@ struct llm_build_olmo : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9651,9 +9656,7 @@ struct llm_build_olmo : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -9721,6 +9724,8 @@ struct llm_build_olmo2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9771,18 +9776,16 @@ struct llm_build_olmo2 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); cb(cur, "attn_post_norm", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -9850,6 +9853,8 @@ struct llm_build_olmoe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -9904,9 +9909,7 @@ struct llm_build_olmoe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -9976,6 +9979,8 @@ struct llm_build_openelm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const int64_t n_head = hparams.n_head(il); const int64_t n_head_kv = hparams.n_head_kv(il); @@ -10037,11 +10042,9 @@ struct llm_build_openelm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { residual = ggml_get_rows(ctx0, residual, inp_out_ids); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); } ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); @@ -10107,6 +10110,8 @@ struct llm_build_gptneox : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -10151,9 +10156,7 @@ struct llm_build_gptneox : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -10255,6 +10258,8 @@ struct llm_build_arctic : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10301,9 +10306,7 @@ struct llm_build_arctic : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -10395,6 +10398,8 @@ struct llm_build_deepseek : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10456,14 +10461,11 @@ struct llm_build_deepseek : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -10571,6 +10573,8 @@ struct llm_build_deepseek2 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10720,9 +10724,7 @@ struct llm_build_deepseek2 : public llm_graph_context { } } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -10818,6 +10820,8 @@ struct llm_build_bitnet : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -10900,9 +10904,7 @@ struct llm_build_bitnet : public llm_graph_context { cb(cur, "attn_o_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -10977,6 +10979,8 @@ struct llm_build_t5_enc : public llm_graph_context { auto * inp_attn = build_attn_inp_no_cache(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11010,9 +11014,7 @@ struct llm_build_t5_enc : public llm_graph_context { cb(cur, "kqv_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11083,6 +11085,8 @@ struct llm_build_t5_dec : public llm_graph_context { auto * inp_attn_self = build_attn_inp_kv_unified(); auto * inp_attn_cross = build_attn_inp_cross(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11174,11 +11178,8 @@ struct llm_build_t5_dec : public llm_graph_context { //cb(cur, "kqv_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); } @@ -11248,6 +11249,8 @@ struct llm_build_jais : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { cur = build_norm(inpL, model.layers[il].attn_norm, @@ -11280,9 +11283,7 @@ struct llm_build_jais : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); } @@ -11346,6 +11347,8 @@ struct llm_build_chatglm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11412,9 +11415,7 @@ struct llm_build_chatglm : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11479,6 +11480,8 @@ struct llm_build_glm4 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11545,9 +11548,7 @@ struct llm_build_glm4 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11630,6 +11631,8 @@ struct llm_build_nemotron : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11689,9 +11692,7 @@ struct llm_build_nemotron : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -11759,6 +11760,8 @@ struct llm_build_exaone : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -11820,9 +11823,7 @@ struct llm_build_exaone : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -12098,6 +12099,8 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12139,13 +12142,16 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base { ); ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); - x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); - cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); } cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6); @@ -12193,6 +12199,8 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12217,11 +12225,12 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); cb(ffn_inp, "ffn_inp", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); } // feed-forward network @@ -12447,6 +12456,8 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12488,12 +12499,14 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base { ); ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); - x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); } cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7); @@ -12538,6 +12551,8 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { const auto n_seq_tokens = ubatch.n_seq_tokens; const auto n_seqs = ubatch.n_seqs; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { const llama_layer * layer = &model.layers[il]; inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); @@ -12562,11 +12577,12 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base { ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); cb(ffn_inp, "ffn_inp", il); - if (il == n_layer - 1) { - // skip computing output for unused tokens - struct ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); } // feed-forward network @@ -12635,6 +12651,9 @@ struct llm_build_granite : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -12697,9 +12716,7 @@ struct llm_build_granite : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -12818,6 +12835,8 @@ struct llm_build_chameleon : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -12894,21 +12913,19 @@ struct llm_build_chameleon : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, nullptr, Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - - if (hparams.swin_norm) { - cur = build_norm(cur, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - } } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); cb(ffn_inp, "ffn_inp", il); @@ -13149,6 +13166,8 @@ struct llm_build_plm : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13252,9 +13271,7 @@ struct llm_build_plm : public llm_graph_context { q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -13314,6 +13331,8 @@ struct llm_build_bailingmoe : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13375,9 +13394,7 @@ struct llm_build_bailingmoe : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -13463,6 +13480,8 @@ struct llm_build_dots1 : public llm_graph_context { auto * inp_attn = build_attn_inp_kv_unified(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13515,9 +13534,7 @@ struct llm_build_dots1 : public llm_graph_context { Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } @@ -13615,6 +13632,8 @@ struct llm_build_arcee : public llm_graph_context { const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + ggml_tensor * inp_out_ids = build_inp_out_ids(); + for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -13677,9 +13696,7 @@ struct llm_build_arcee : public llm_graph_context { cb(cur, "attn_out", il); } - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); + if (il == n_layer - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } From 9230dbe2c757e2d5071329095727d0fa9d4b85c4 Mon Sep 17 00:00:00 2001 From: Charles Xu Date: Fri, 20 Jun 2025 09:51:01 +0200 Subject: [PATCH 7/7] ggml: Update KleidiAI to v1.9.0 (#14277) --- ggml/src/ggml-cpu/CMakeLists.txt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index df0034057..52cae778c 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -465,9 +465,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name) # Fetch KleidiAI sources: include(FetchContent) - set(KLEIDIAI_COMMIT_TAG "v1.6.0") + set(KLEIDIAI_COMMIT_TAG "v1.9.0") set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz") - set(KLEIDIAI_ARCHIVE_MD5 "75b4ad68f25ab673dcc01065e5a0b05f") + set(KLEIDIAI_ARCHIVE_MD5 "2a8e1bb55d201557553545536489a017") if (POLICY CMP0135) cmake_policy(SET CMP0135 NEW)