From 137435ff154f8d96bb79986dda9ea4dc1b43d2c3 Mon Sep 17 00:00:00 2001 From: Charles Xu Date: Tue, 3 Mar 2026 10:40:26 +0100 Subject: [PATCH] kleidiai : add sme fp16 compute path for q4_0 gemm on aarch64 (#20043) --- ggml/src/ggml-cpu/CMakeLists.txt | 11 +++++--- ggml/src/ggml-cpu/kleidiai/kernels.cpp | 35 +++++++++++++------------- 2 files changed, 25 insertions(+), 21 deletions(-) diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index 3dc948e4d8..6ca3176a2f 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -566,9 +566,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name) # Fetch KleidiAI sources: include(FetchContent) - set(KLEIDIAI_COMMIT_TAG "v1.16.0") + set(KLEIDIAI_COMMIT_TAG "v1.22.0") set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz") - set(KLEIDIAI_ARCHIVE_MD5 "0a9e9008adb6031f9e8cf70dff4a3321") + set(KLEIDIAI_ARCHIVE_MD5 "54049037570ab0ee0a0d126b2ba5ece1") if (POLICY CMP0135) cmake_policy(SET CMP0135 NEW) @@ -608,6 +608,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name) ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/) set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}") @@ -648,7 +649,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name) if (NOT SME_ENABLED MATCHES -1) list(APPEND GGML_KLEIDIAI_SOURCES - ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S @@ -656,10 +656,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name) ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S) - set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2") + set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2+sme2+fp16") endif() if (NOT SVE_ENABLED MATCHES -1) diff --git a/ggml/src/ggml-cpu/kleidiai/kernels.cpp b/ggml/src/ggml-cpu/kleidiai/kernels.cpp index d114f2d49b..40f7c0df65 100644 --- a/ggml/src/ggml-cpu/kleidiai/kernels.cpp +++ b/ggml/src/ggml-cpu/kleidiai/kernels.cpp @@ -1,4 +1,4 @@ -// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// SPDX-FileCopyrightText: Copyright 2025-2026 Arm Limited and/or its affiliates // SPDX-License-Identifier: MIT // @@ -9,7 +9,6 @@ #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h" #include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h" #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h" -#include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h" #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h" #include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h" #include "kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.h" @@ -20,6 +19,7 @@ #include "kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.h" #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.h" #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h" +#include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h" #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h" #include "kai_lhs_quant_pack_qsi8d32p_f32.h" @@ -31,6 +31,7 @@ #include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h" #include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h" #include "kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.h" +#include "kai_lhs_pack_f16pmrx2_f32_neon.h" #include "kai_common.h" @@ -309,24 +310,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = { { /* SME GEMM */ /* .kern_info = */ { - /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, - /* .get_lhs_offset_ex = */ &kernel_offs_fn3, - /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3, - /* .run_kernel_ex = */ &kernel_run_fn11, + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa, + /* .get_lhs_offset_ex = */ &kernel_offs_fn3, + /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3, + /* .run_kernel_ex = */ &kernel_run_fn11, }, /* .gemm_lhs_info = */ { - /* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon, - /* .get_packed_offset_ex = */ &lhs_offs_fn6, - /* .packed_size_ex = */ &lhs_ps_fn6, - /* .pack_func_ex = */ &lhs_pack_float_fn10, + /* .get_offset = */ kai_get_lhs_offset_lhs_pack_f16pmrx2_f32_neon, + /* .get_packed_offset_ex = */ &lhs_offs_fn6, + /* .packed_size_ex = */ &lhs_ps_fn6, + /* .pack_func_ex = */ &lhs_pack_void_fn10, }, /* SME GEMV */ /* .kern_info = */ {