From 072d85c86526065b081a6486509f42fe3d2af40c Mon Sep 17 00:00:00 2001 From: Michael Kozlov Date: Thu, 30 Jan 2025 14:17:12 +0000 Subject: [PATCH 1/3] Optimize F32 <- QAI8DXP (LHS) x QSI4C32P (RHS) for 4x8 SDOT - Update build script - Add to unit test Signed-off-by: Michael Kozlov --- CMakeLists.txt | 2 + kai/ukernels/matmul/BUILD.bazel | 2 + ..._qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c | 185 +++++++++ ..._qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.h | 145 +++++++ ...8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S | 362 ++++++++++++++++++ ...matmul_clamp_f32_qai8dxp_qsi4c32p_test.cpp | 5 +- 6 files changed, 700 insertions(+), 1 deletion(-) create mode 100644 kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c create mode 100644 kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.h create mode 100644 kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S diff --git a/CMakeLists.txt b/CMakeLists.txt index 97ce7eb1..1d85bc8c 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -130,10 +130,12 @@ set(KLEIDIAI_FILES_NEON_DOTPROD_ASM kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod_asm.S kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod_asm.S kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c + kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c ) set(KLEIDIAI_FILES_NEON_DOTPROD diff --git a/kai/ukernels/matmul/BUILD.bazel b/kai/ukernels/matmul/BUILD.bazel index 97f12218..ad4769a3 100644 --- a/kai/ukernels/matmul/BUILD.bazel +++ b/kai/ukernels/matmul/BUILD.bazel @@ -73,6 +73,7 @@ DOTPROD_KERNELS = [ "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod", "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod", "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod", + "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod", "matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod", "matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod", "matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod", @@ -91,6 +92,7 @@ DOTPROD_KERNELS_ASM = [ "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod", "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod", "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod", + "matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod", ] # buildifier: keep sorted diff --git a/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c new file mode 100644 index 00000000..6f6df50b --- /dev/null +++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.c @@ -0,0 +1,185 @@ +// +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// +// SPDX-License-Identifier: Apache-2.0 +// +#if !defined(__aarch64__) && !defined(__ARM_FEATURE_DOTPROD) && !defined(_M_ARM64) +#error "Dotprod extension required to compile this micro-kernel" +#else // Architectural features check. + +#include "kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.h" + +#include +#include + +#include "kai/kai_common.h" + +typedef struct { + float* dst; + const void* lhs_packed; + const void* rhs_packed; + const float* clamp_vals; + size_t dst_stride_row; + size_t m; + size_t n; + size_t num_blocks; + size_t num_subblocks; +} KernelArgs; + +void kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(KernelArgs* args_ptr); + +// Compute args +static const size_t kai_m_step = 4; +static const size_t kai_n_step = 8; +// Packing args +static const size_t kai_mr = 4; +static const size_t kai_nr = 8; +static const size_t kai_kr = 8; +static const size_t kai_sr = 2; +// LHS format args (num. bytes per value, multiplier, zero_point (if asymmetric)) +static const size_t kai_num_bytes_qvalue_lhs = 1; +static const size_t kai_num_bytes_multiplier_lhs = 4; +static const size_t kai_num_bytes_zp_lhs = 4; +// RHS format args (num. bytes per value, multiplier, zero_point (if asymmetric), and reduction sum (if LHS is +// asymmetric)) +static const size_t kai_num_bytes_recip_qvalue_rhs = 2; +static const size_t kai_num_bytes_multiplier_rhs = 2; +static const size_t kai_num_bytes_rsum_rhs = 4; +// DST format args +static const size_t kai_num_bytes_dst_value = 4; +// Extra args +static const size_t kai_num_bytes_bias = 4; +static const size_t kai_k_multiple_of = 32; +static const size_t kai_bl = 32; + +inline static size_t kai_get_k_roundedup(size_t k) { + return kai_roundup(k, kai_k_multiple_of); +} + +inline static size_t kai_get_num_bytes_per_block_rhs(size_t bl) { + KAI_ASSUME((bl % kai_bl) == 0); + size_t num_bytes_per_block_rhs = (bl / kai_num_bytes_recip_qvalue_rhs) + kai_num_bytes_multiplier_rhs; + return num_bytes_per_block_rhs; +} + +inline static size_t kai_get_num_blocks_per_row(size_t k, size_t bl) { + KAI_ASSUME((bl % kai_bl) == 0); + + return kai_roundup(k, bl) / bl; +} + +inline static size_t kai_get_lhs_packed_stride(size_t k) { + const size_t k_internal = kai_get_k_roundedup(k); + size_t lhs_packed_stride = kai_mr * ((k_internal * kai_num_bytes_qvalue_lhs) + kai_num_bytes_multiplier_lhs); + // Since the LHS matrix is asymmetric with per-row quantization, we must include the + // the number of bytes to hold the zero point value + lhs_packed_stride += kai_mr * kai_num_bytes_zp_lhs; + + return lhs_packed_stride; +} + +inline static size_t kai_get_rhs_packed_stride(size_t k, size_t bl) { + KAI_ASSUME((bl % kai_bl) == 0); + + const size_t num_blocks_per_row = kai_get_num_blocks_per_row(k, bl); + const size_t num_bytes_per_block = kai_get_num_bytes_per_block_rhs(bl); + + size_t rhs_packed_stride = kai_nr * (num_bytes_per_block * num_blocks_per_row); + // Since the LHS matrix is quantized asymmetric with per-row quantization, we also include + // the number of bytes for the reduction sum + rhs_packed_stride += kai_nr * kai_num_bytes_rsum_rhs; + // Since the bias is packed with the RHS matrix, the stride is adjusted with the number of bytes of the bias + rhs_packed_stride += kai_nr * kai_num_bytes_bias; + + return rhs_packed_stride; +} + +size_t kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void) { + return kai_m_step; +} + +size_t kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void) { + return kai_n_step; +} + +size_t kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void) { + return kai_mr; +} + +size_t kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void) { + return kai_nr; +} + +size_t kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void) { + return kai_kr; +} + +size_t kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void) { + return kai_sr; +} + +size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(size_t m_idx, size_t k) { + KAI_ASSUME((m_idx % kai_m_step) == 0); + + return (m_idx / kai_mr) * kai_get_lhs_packed_stride(k); +} + +size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t n_idx, size_t k, size_t bl) { + KAI_ASSUME((k % bl) == 0); + KAI_ASSUME((n_idx % kai_n_step) == 0); + + return (n_idx / kai_nr) * kai_get_rhs_packed_stride(k, bl); +} + +size_t kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t m_idx, size_t n_idx, size_t dst_stride) { + KAI_ASSUME((m_idx % kai_m_step) == 0); + KAI_ASSUME((n_idx % kai_n_step) == 0); + + return (n_idx * kai_num_bytes_dst_value) + m_idx * dst_stride; +} + +size_t kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(size_t m, size_t n) { + return m * n * kai_num_bytes_dst_value; +} + +void kai_run_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t m, // + size_t n, // + size_t k, // + size_t bl, // + const void* restrict lhs_packed, // + const void* restrict rhs_packed, // + float* restrict dst, // NOLINT(readability-non-const-parameter) + size_t dst_stride_row, // + size_t dst_stride_col, // + float scalar_min, // + float scalar_max) { + KAI_ASSUME(dst_stride_col == sizeof(float)); + KAI_ASSUME((k % bl) == 0); + KAI_ASSUME((bl % kai_bl) == 0); + + if (m == 0) { + return; + } + const size_t num_subblocks = bl / kai_bl; + const size_t num_blocks = kai_get_num_blocks_per_row(k, bl); + const float clamp_vals[2] = {scalar_min, scalar_max}; + + KernelArgs args; + + args.dst = dst; + args.lhs_packed = lhs_packed; + args.rhs_packed = rhs_packed; + args.clamp_vals = clamp_vals; + args.dst_stride_row = dst_stride_row; + args.m = m; + args.n = n; + args.num_blocks = num_blocks; + args.num_subblocks = num_subblocks; + + kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(&args); +} + +#endif // Architectural features check. diff --git a/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.h new file mode 100644 index 00000000..6cd78a6f --- /dev/null +++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.h @@ -0,0 +1,145 @@ +// +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// +// SPDX-License-Identifier: Apache-2.0 +// + +#pragma once + +#include + +#ifdef __cplusplus +extern "C" { +#endif // __cplusplus + +/// Micro-kernel dependencies +/// +/// -# @ref kai_lhs_quant_pack_qai8dxp_f32 to dynamically quantize and pack the LHS matrix in a single step. +/// -# @ref kai_rhs_pack_nxk_qsi4c32p_qsu4c32s1s0 to pack the RHS NxK matrix. +/// -# @ref kai_rhs_pack_kxn_qsi4c32p_qsu4c32s1s0 to pack the RHS KxN matrix. + +/// -------------------------------------------------- + +/// Gets the m step value. +/// The micro-kernel can process any M values. However, the starting M index to +/// be processed must be a multiple of m step. +/// +/// @return the m step value +size_t kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void); + +/// Gets the n step value. +/// The micro-kernel can process any N values. However, the starting N index to +/// be processed must be a multiple of n step. +/// +/// @return the n step +size_t kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void); + +/// Gets the mr value, which must be used to pack the LHS matrix +/// +/// @return the mr value +size_t kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void); + +/// Gets the nr value, which must be used to pack the RHS matrix. +/// +/// @return the nr value +size_t kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void); + +/// Gets the kr value, which must be used to pack the LHS and RHS matrices +/// +/// @return the kr value +size_t kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void); + +/// Gets the sr value, which must be used to pack the LHS and RHS matrices +/// +/// @return the sr value +size_t kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod(void); + +/// Gets the offset in bytes for the packed LHS matrix, +/// which contains the packed Quantized Asymmetric Signed 8-bit with per-row quantization (qai8dx) values. +/// +/// This function should be called before passing the pointer to the packed LHS matrix to the micro-kernel. +/// +/// @param[in] m_idx Row index in the LHS matrix (not packed). It must be a multiple of m_step. +/// @param[in] k Total number of columns in the LHS matrix (not packed). +/// +/// @return the offset in bytes to the packed LHS matrix +size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t m_idx, // + size_t k); // + +/// Gets the offset in bytes for the packed RHS matrix, +/// which contains the packed Quantized Symmetric Signed 4-bit with per-block (32) quantization (qsi4c32) values. +/// +/// @param[in] n_idx Col index in the RHS matrix (not packed). It must be a multiple of n_step. +/// @param[in] k The common dimension between the LHS and RHS matrix (K). +/// It must be a multiple of the block length (bl). +/// @param[in] bl Block length. It must be a multiple of 32. +/// +/// @return the offset in bytes to the packed RHS matrix +size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t n_idx, // + size_t k, // + size_t bl); // + +/// Gets the offset in bytes for the DST matrix +/// +/// @param[in] m_idx Row index in the DST matrix. It must be a multiple of m_step. +/// @param[in] n_idx Column index in the DST matrix. It must be multiple of n_step. +/// @param[in] dst_stride The number of bytes in in each row of the DST matrix +/// +/// @return the DST offset in bytes +size_t kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t m_idx, // + size_t n_idx, // + size_t dst_stride); // + +/// Gets the size in bytes for the destination (DST) matrix. +/// +/// @param[in] m Number of rows in the destination (DST) matrix. +/// @param[in] n Number of columns in the destination (DST) matrix. +/// +/// @return the destination (DST) matrix size in bytes +size_t kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t m, // + size_t n); // + +/// Runs the matrix multiplication (matmul) micro-kernel followed by a clamp (min-max) operation. +/// +/// LHS matrix: Quantized Asymmetric Signed 8-bit with per-row quantization (qai8dx) and packed. +/// RHS matrix: Quantized Symmetric Signed 4-bit with per-block (32) quantization (qsi4c32) and packed. +/// Output tile: (rows x cols) = m_step x n_step. +/// +/// Note: Please refer to the get functions for m_step and n_step for the exact values. +/// +/// Features used: dotprod +/// +/// @param[in] m The number of output rows written. +/// @param[in] n The number of output columns written. +/// @param[in] k The number of channels. The common dimension between the LHS and RHS matrix. +/// It must be a multiple of the block length (bl). +/// @param[in] bl Block length. Block length. It must be a multiple of 32. +/// @param[in] lhs_packed The LHS packed matrix. The micro-kernel to pack the native LHS matrix is reported at the +/// top of this file. +/// @param[in] rhs_packed The RHS packed matrix. The micro-kernel to pack the native RHS matrix is reported at the +/// top of this file. +/// @param[out] dst The DST matrix. +/// @param[in] dst_stride_row Stride in bytes between two rows of the DST matrix. +/// @param[in] dst_stride_col Stride in bytes between two columns of the DST matrix. It must be sizeof(float) bytes. +/// @param[in] scalar_min Min value used to clamp the final result. +/// @param[in] scalar_max Max value used to clamp the final result. +void kai_run_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod( + size_t m, // + size_t n, // + size_t k, // + size_t bl, // + const void* lhs_packed, // + const void* rhs_packed, // + float* dst, // + size_t dst_stride_row, // + size_t dst_stride_col, // + float scalar_min, // + float scalar_max); // + +#ifdef __cplusplus +} +#endif // __cplusplus diff --git a/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S new file mode 100644 index 00000000..7d9ef2d3 --- /dev/null +++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S @@ -0,0 +1,362 @@ +// +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// +// SPDX-License-Identifier: Apache-2.0 +// + +#if defined(_MSC_VER) + #define KAI_ASM_GLOBAL(name) GLOBAL name + #define KAI_ASM_FUNCTION_TYPE(name) + #define KAI_ASM_FUNCTION_LABEL(name) name PROC + #define KAI_ASM_FUNCTION_END(name) ENDP + + #define KAI_ASM_CODE(name) AREA name, CODE, READONLY + #define KAI_ASM_ALIGN + #define KAI_ASM_LABEL(name) name + #define KAI_ASM_INST(hex) DCD hex + #define KAI_ASM_END END +#else + #if defined(__APPLE__) + #define KAI_ASM_GLOBAL(name) .globl _##name + #define KAI_ASM_FUNCTION_TYPE(name) + #define KAI_ASM_FUNCTION_LABEL(name) _##name: + #define KAI_ASM_FUNCTION_END(name) + #else + #define KAI_ASM_GLOBAL(name) .global name + #define KAI_ASM_FUNCTION_TYPE(name) .type name, %function + #define KAI_ASM_FUNCTION_LABEL(name) name: + #define KAI_ASM_FUNCTION_END(name) .size name, .-name + #endif + + #define KAI_ASM_CODE(name) .text + #define KAI_ASM_ALIGN .p2align 4,,11 + #define KAI_ASM_LABEL(name) name: + #define KAI_ASM_INST(hex) .inst hex + #define KAI_ASM_END +#endif + + KAI_ASM_CODE(matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod) + KAI_ASM_ALIGN + + KAI_ASM_GLOBAL(kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod) + +KAI_ASM_FUNCTION_TYPE(kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod) +KAI_ASM_FUNCTION_LABEL(kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod) + stp x20, x21, [sp, -144]! + stp x22, x23, [sp, 16] + stp x24, x25, [sp, 32] + stp x26, x27, [sp, 48] + str x28, [sp, 64] + stp d10, d11, [sp, 72] + stp d12, d13, [sp, 88] + stp d14, d15, [sp, 104] + stp d8, d9, [sp, 120] + mov x16, #0x80 + movi v11.16b, #0xf0 + mov x21, #0x20 + ldr x15, [x0, #0x40] + ldr x20, [x0, #0x28] + ldr x14, [x0, #0x38] + ldr x13, [x0, #0x8] + ldr x12, [x0, #0x10] + ldr x11, [x0, #0x30] + mul x16, x15, x16 + mov x10, x20 + ldr x9, [x0, #0x0] + ldr x28, [x0, #0x20] + ldr x27, [x0, #0x18] + madd x16, x14, x16, x21 + cbz x10, label_12 +KAI_ASM_LABEL(label_1) // Row loop + mov x26, x12 + mov x25, x11 + add x24, x9, x28, LSL #2 +KAI_ASM_LABEL(label_2) // Column loop + movi v0.16b, #0x0 + movi v12.16b, #0x0 + mov x22, x13 + mov x21, x14 + movi v14.16b, #0x0 + movi v13.16b, #0x0 + movi v15.16b, #0x0 + movi v7.16b, #0x0 + movi v8.16b, #0x0 + movi v9.16b, #0x0 +KAI_ASM_LABEL(label_3) // Block loop + movi v5.4s, #0x0 + movi v6.4s, #0x0 + mov x20, x15 + movi v4.4s, #0x0 + movi v1.4s, #0x0 + movi v3.4s, #0x0 + movi v21.4s, #0x0 + movi v10.4s, #0x0 + movi v2.4s, #0x0 +KAI_ASM_LABEL(label_4) // Sub block loop + ldr q28, [x26, #0x0] + ldr q26, [x26, #0x10] + subs x20, x20, #0x1 + ldr q19, [x22, #0x0] + ldr q23, [x26, #0x20] + ldr q31, [x26, #0x30] + ldr q18, [x22, #0x10] + ldr q27, [x26, #0x40] + ldr q25, [x26, #0x50] + shl v30.16b, v28.16b, #0x4 + shl v20.16b, v26.16b, #0x4 + ldr q17, [x22, #0x20] + ldr q24, [x26, #0x60] + shl v16.16b, v23.16b, #0x4 + and v28.16b, v28.16b, v11.16b + ldr q29, [x26, #0x70] + ldr q22, [x22, #0x30] + and v26.16b, v26.16b, v11.16b + and v23.16b, v23.16b, v11.16b + KAI_ASM_INST(0x4f93e3c5) // sdot v5.4s, v30.16b, v19.4b[0] + KAI_ASM_INST(0x4f93e286) // sdot v6.4s, v20.16b, v19.4b[0] + add x26, x26, #0x80 + KAI_ASM_INST(0x4fb3e3c4) // sdot v4.4s, v30.16b, v19.4b[1] + KAI_ASM_INST(0x4fb3e281) // sdot v1.4s, v20.16b, v19.4b[1] + KAI_ASM_INST(0x4f93ebc3) // sdot v3.4s, v30.16b, v19.4b[2] + KAI_ASM_INST(0x4f93ea95) // sdot v21.4s, v20.16b, v19.4b[2] + KAI_ASM_INST(0x4fb3ebca) // sdot v10.4s, v30.16b, v19.4b[3] + ldr q30, [x22, #0x40] + KAI_ASM_INST(0x4fb3ea82) // sdot v2.4s, v20.16b, v19.4b[3] + ldr q20, [x22, #0x50] + ldr q19, [x22, #0x60] + KAI_ASM_INST(0x4f92e205) // sdot v5.4s, v16.16b, v18.4b[0] + KAI_ASM_INST(0x4fb2e204) // sdot v4.4s, v16.16b, v18.4b[1] + KAI_ASM_INST(0x4f92ea03) // sdot v3.4s, v16.16b, v18.4b[2] + KAI_ASM_INST(0x4fb2ea0a) // sdot v10.4s, v16.16b, v18.4b[3] + shl v16.16b, v31.16b, #0x4 + and v31.16b, v31.16b, v11.16b + KAI_ASM_INST(0x4f92e206) // sdot v6.4s, v16.16b, v18.4b[0] + KAI_ASM_INST(0x4fb2e201) // sdot v1.4s, v16.16b, v18.4b[1] + KAI_ASM_INST(0x4f92ea15) // sdot v21.4s, v16.16b, v18.4b[2] + KAI_ASM_INST(0x4fb2ea02) // sdot v2.4s, v16.16b, v18.4b[3] + ldr q18, [x22, #0x70] + shl v16.16b, v27.16b, #0x4 + and v27.16b, v27.16b, v11.16b + add x22, x22, #0x80 + KAI_ASM_INST(0x4f91e205) // sdot v5.4s, v16.16b, v17.4b[0] + KAI_ASM_INST(0x4fb1e204) // sdot v4.4s, v16.16b, v17.4b[1] + KAI_ASM_INST(0x4f91ea03) // sdot v3.4s, v16.16b, v17.4b[2] + KAI_ASM_INST(0x4fb1ea0a) // sdot v10.4s, v16.16b, v17.4b[3] + shl v16.16b, v25.16b, #0x4 + and v25.16b, v25.16b, v11.16b + KAI_ASM_INST(0x4f91e206) // sdot v6.4s, v16.16b, v17.4b[0] + KAI_ASM_INST(0x4fb1e201) // sdot v1.4s, v16.16b, v17.4b[1] + KAI_ASM_INST(0x4f91ea15) // sdot v21.4s, v16.16b, v17.4b[2] + KAI_ASM_INST(0x4fb1ea02) // sdot v2.4s, v16.16b, v17.4b[3] + shl v17.16b, v24.16b, #0x4 + shl v16.16b, v29.16b, #0x4 + and v24.16b, v24.16b, v11.16b + and v29.16b, v29.16b, v11.16b + KAI_ASM_INST(0x4f96e225) // sdot v5.4s, v17.16b, v22.4b[0] + KAI_ASM_INST(0x4fb6e224) // sdot v4.4s, v17.16b, v22.4b[1] + KAI_ASM_INST(0x4f96ea23) // sdot v3.4s, v17.16b, v22.4b[2] + KAI_ASM_INST(0x4fb6ea2a) // sdot v10.4s, v17.16b, v22.4b[3] + KAI_ASM_INST(0x4f96e206) // sdot v6.4s, v16.16b, v22.4b[0] + KAI_ASM_INST(0x4fb6e201) // sdot v1.4s, v16.16b, v22.4b[1] + KAI_ASM_INST(0x4f96ea15) // sdot v21.4s, v16.16b, v22.4b[2] + KAI_ASM_INST(0x4fb6ea02) // sdot v2.4s, v16.16b, v22.4b[3] + KAI_ASM_INST(0x4f9ee385) // sdot v5.4s, v28.16b, v30.4b[0] + KAI_ASM_INST(0x4fbee384) // sdot v4.4s, v28.16b, v30.4b[1] + KAI_ASM_INST(0x4f9eeb83) // sdot v3.4s, v28.16b, v30.4b[2] + KAI_ASM_INST(0x4fbeeb8a) // sdot v10.4s, v28.16b, v30.4b[3] + KAI_ASM_INST(0x4f9ee346) // sdot v6.4s, v26.16b, v30.4b[0] + KAI_ASM_INST(0x4fbee341) // sdot v1.4s, v26.16b, v30.4b[1] + KAI_ASM_INST(0x4f9eeb55) // sdot v21.4s, v26.16b, v30.4b[2] + KAI_ASM_INST(0x4fbeeb42) // sdot v2.4s, v26.16b, v30.4b[3] + KAI_ASM_INST(0x4f94e2e5) // sdot v5.4s, v23.16b, v20.4b[0] + KAI_ASM_INST(0x4fb4e2e4) // sdot v4.4s, v23.16b, v20.4b[1] + KAI_ASM_INST(0x4f94eae3) // sdot v3.4s, v23.16b, v20.4b[2] + KAI_ASM_INST(0x4fb4eaea) // sdot v10.4s, v23.16b, v20.4b[3] + KAI_ASM_INST(0x4f94e3e6) // sdot v6.4s, v31.16b, v20.4b[0] + KAI_ASM_INST(0x4fb4e3e1) // sdot v1.4s, v31.16b, v20.4b[1] + KAI_ASM_INST(0x4f94ebf5) // sdot v21.4s, v31.16b, v20.4b[2] + KAI_ASM_INST(0x4fb4ebe2) // sdot v2.4s, v31.16b, v20.4b[3] + KAI_ASM_INST(0x4f93e365) // sdot v5.4s, v27.16b, v19.4b[0] + KAI_ASM_INST(0x4fb3e364) // sdot v4.4s, v27.16b, v19.4b[1] + KAI_ASM_INST(0x4f93eb63) // sdot v3.4s, v27.16b, v19.4b[2] + KAI_ASM_INST(0x4fb3eb6a) // sdot v10.4s, v27.16b, v19.4b[3] + KAI_ASM_INST(0x4f93e326) // sdot v6.4s, v25.16b, v19.4b[0] + KAI_ASM_INST(0x4fb3e321) // sdot v1.4s, v25.16b, v19.4b[1] + KAI_ASM_INST(0x4f93eb35) // sdot v21.4s, v25.16b, v19.4b[2] + KAI_ASM_INST(0x4fb3eb22) // sdot v2.4s, v25.16b, v19.4b[3] + KAI_ASM_INST(0x4f92e305) // sdot v5.4s, v24.16b, v18.4b[0] + KAI_ASM_INST(0x4fb2e304) // sdot v4.4s, v24.16b, v18.4b[1] + KAI_ASM_INST(0x4f92eb03) // sdot v3.4s, v24.16b, v18.4b[2] + KAI_ASM_INST(0x4fb2eb0a) // sdot v10.4s, v24.16b, v18.4b[3] + KAI_ASM_INST(0x4f92e3a6) // sdot v6.4s, v29.16b, v18.4b[0] + KAI_ASM_INST(0x4fb2e3a1) // sdot v1.4s, v29.16b, v18.4b[1] + KAI_ASM_INST(0x4f92ebb5) // sdot v21.4s, v29.16b, v18.4b[2] + KAI_ASM_INST(0x4fb2eba2) // sdot v2.4s, v29.16b, v18.4b[3] + bgt label_4 + ldr q16, [x26, #0x0] + scvtf v5.4s, v5.4s, #0x4 + scvtf v6.4s, v6.4s, #0x4 + add x26, x26, #0x10 + scvtf v4.4s, v4.4s, #0x4 + scvtf v1.4s, v1.4s, #0x4 + scvtf v3.4s, v3.4s, #0x4 + scvtf v21.4s, v21.4s, #0x4 + shll v17.4s, v16.4h, #0x10 + shll2 v16.4s, v16.8h, #0x10 + scvtf v10.4s, v10.4s, #0x4 + scvtf v2.4s, v2.4s, #0x4 + fmla v0.4s, v5.4s, v17.4s + fmla v12.4s, v6.4s, v16.4s + fmla v14.4s, v4.4s, v17.4s + fmla v13.4s, v1.4s, v16.4s + fmla v15.4s, v3.4s, v17.4s + fmla v7.4s, v21.4s, v16.4s + fmla v8.4s, v10.4s, v17.4s + fmla v9.4s, v2.4s, v16.4s + subs x21, x21, #0x1 + bgt label_3 + ld1 { v23.4s }, [x22] + ldr q22, [x26, #0x0] + add x22, x22, #0x10 + add x20, x27, #0x4 + ldr q21, [x26, #0x10] + ldr q20, [x22, #0x0] + cmp x25, #0x8 + ldr q19, [x26, #0x20] + ldr q18, [x26, #0x30] + add x26, x26, #0x40 + ld1r { v17.4s }, [x27] + ld1r { v16.4s }, [x20] + scvtf v23.4s, v23.4s + fmla v0.4s, v22.4s, v23.s[0] + fmla v12.4s, v21.4s, v23.s[0] + fmla v14.4s, v22.4s, v23.s[1] + fmla v13.4s, v21.4s, v23.s[1] + fmla v15.4s, v22.4s, v23.s[2] + fmla v7.4s, v21.4s, v23.s[2] + fmla v8.4s, v22.4s, v23.s[3] + fmla v9.4s, v21.4s, v23.s[3] + fmul v0.4s, v0.4s, v20.s[0] + fmul v12.4s, v12.4s, v20.s[0] + fmul v14.4s, v14.4s, v20.s[1] + fmul v13.4s, v13.4s, v20.s[1] + fmul v15.4s, v15.4s, v20.s[2] + fmul v7.4s, v7.4s, v20.s[2] + fmul v8.4s, v8.4s, v20.s[3] + fmul v9.4s, v9.4s, v20.s[3] + fadd v0.4s, v0.4s, v19.4s + fadd v12.4s, v12.4s, v18.4s + fadd v14.4s, v14.4s, v19.4s + fadd v13.4s, v13.4s, v18.4s + fadd v15.4s, v15.4s, v19.4s + fadd v7.4s, v7.4s, v18.4s + fadd v8.4s, v8.4s, v19.4s + fadd v9.4s, v9.4s, v18.4s + fmax v0.4s, v0.4s, v17.4s + fmax v12.4s, v12.4s, v17.4s + fmax v14.4s, v14.4s, v17.4s + fmax v13.4s, v13.4s, v17.4s + fmax v15.4s, v15.4s, v17.4s + fmax v7.4s, v7.4s, v17.4s + fmax v8.4s, v8.4s, v17.4s + fmax v9.4s, v9.4s, v17.4s + fmin v0.4s, v0.4s, v16.4s + fmin v12.4s, v12.4s, v16.4s + fmin v14.4s, v14.4s, v16.4s + fmin v13.4s, v13.4s, v16.4s + fmin v15.4s, v15.4s, v16.4s + fmin v7.4s, v7.4s, v16.4s + fmin v8.4s, v8.4s, v16.4s + fmin v9.4s, v9.4s, v16.4s + blt label_6 + mov x20, x9 + cmp x10, #0x1 + str q0, [x20, #0x0] + str q12, [x20, #0x10] + add x20, x20, x28 + ble label_11 + cmp x10, #0x2 + str q14, [x20, #0x0] + str q13, [x20, #0x10] + add x20, x20, x28 + ble label_11 + cmp x10, #0x3 + str q15, [x20, #0x0] + str q7, [x20, #0x10] + add x20, x20, x28 + ble label_11 + str q8, [x20, #0x0] + str q9, [x20, #0x10] + b label_11 +KAI_ASM_LABEL(label_6) // Partial output + mov x23, x9 + cmp x10, #0x1 + add x22, x23, x28 + csel x22, x22, x23, GT + cmp x10, #0x2 + add x21, x23, x28, LSL #1 + csel x21, x21, x22, GT + cmp x10, #0x3 + add x20, x21, x28 + csel x20, x20, x21, GT + tbz x25, #2, label_8 + st1 { v8.4s }, [x20], #0x10 + st1 { v15.4s }, [x21], #0x10 + st1 { v14.4s }, [x22], #0x10 + st1 { v0.4s }, [x23], #0x10 + tbz x25, #1, label_7 + st1 { v9.d }[0], [x20], #0x8 + st1 { v7.d }[0], [x21], #0x8 + st1 { v13.d }[0], [x22], #0x8 + st1 { v12.d }[0], [x23], #0x8 + tbz x25, #0, label_10 + st1 { v9.s }[2], [x20] + st1 { v7.s }[2], [x21] + st1 { v13.s }[2], [x22] + st1 { v12.s }[2], [x23] + b label_10 +KAI_ASM_LABEL(label_7) // Output block 0: partial_1_4 + tbz x25, #0, label_10 + st1 { v9.s }[0], [x20] + st1 { v7.s }[0], [x21] + st1 { v13.s }[0], [x22] + st1 { v12.s }[0], [x23] + b label_10 +KAI_ASM_LABEL(label_8) // Output block 0: partial_2_0 + tbz x25, #1, label_9 + st1 { v8.d }[0], [x20], #0x8 + st1 { v15.d }[0], [x21], #0x8 + st1 { v14.d }[0], [x22], #0x8 + st1 { v0.d }[0], [x23], #0x8 + tbz x25, #0, label_10 + st1 { v8.s }[2], [x20] + st1 { v15.s }[2], [x21] + st1 { v14.s }[2], [x22] + st1 { v0.s }[2], [x23] + b label_10 +KAI_ASM_LABEL(label_9) // Output block 0: partial_1_0 + st1 { v8.s }[0], [x20] + st1 { v15.s }[0], [x21] + st1 { v14.s }[0], [x22] + st1 { v0.s }[0], [x23] +KAI_ASM_LABEL(label_10) // Output block 0: Done +KAI_ASM_LABEL(label_11) // Output stage exit + subs x25, x25, #0x8 + add x9, x9, #0x20 + bgt label_2 + subs x10, x10, #0x4 + add x13, x13, x16 + mov x9, x24 + bgt label_1 +KAI_ASM_LABEL(label_12) // Row loop skip + ldp x22, x23, [sp, 16] + ldp x24, x25, [sp, 32] + ldp x26, x27, [sp, 48] + ldr x28, [sp, 64] + ldp d10, d11, [sp, 72] + ldp d12, d13, [sp, 88] + ldp d14, d15, [sp, 104] + ldp d8, d9, [sp, 120] + ldp x20, x21, [sp], 144 + ret + KAI_ASM_FUNCTION_END(kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod) diff --git a/test/tests/matmul_clamp_f32_qai8dxp_qsi4c32p_test.cpp b/test/tests/matmul_clamp_f32_qai8dxp_qsi4c32p_test.cpp index ab603be0..28540f0d 100644 --- a/test/tests/matmul_clamp_f32_qai8dxp_qsi4c32p_test.cpp +++ b/test/tests/matmul_clamp_f32_qai8dxp_qsi4c32p_test.cpp @@ -24,6 +24,7 @@ #include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h" #include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp1x8_qsi4c32p8x8_1x8x32_neon_dotprod.h" #include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.h" +#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod.h" #include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_16x4x32_neon_i8mm.h" #include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p4x8_8x4x32_neon_i8mm.h" #include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x8_qsi4c32p8x8_4x8x32_neon_i8mm.h" @@ -50,7 +51,7 @@ namespace kai::test { static auto cpu_has_dotprod_and_bf16 = []() { return cpu_has_dotprod() && cpu_has_bf16(); }; static auto cpu_has_i8mm_and_bf16 = []() { return cpu_has_i8mm() && cpu_has_bf16(); }; -static const std::array, 7> +static const std::array, 8> variants_kai_matmul_clamp_f32_qai8dxp_qsi4c32p = { {{UKERNEL_MATMUL_VARIANT(clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod), "kai_matmul_clamp_f32_qai8dxp1x4_qsi4c32p4x4_1x4_neon_dotprod", cpu_has_dotprod_and_bf16}, @@ -60,6 +61,8 @@ static const std::array Date: Fri, 31 Jan 2025 11:27:03 +0000 Subject: [PATCH 2/3] Update changelog with new micro-kernel Signed-off-by: Michael Kozlov --- CHANGELOG.md | 3 +++ 1 file changed, 3 insertions(+) diff --git a/CHANGELOG.md b/CHANGELOG.md index 536c1638..965ff197 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -10,6 +10,9 @@ KleidiAI follows the [Semantic Versioning](https://semver.org/) specification fo ## Upcoming Release +- New Advanced SIMD micro-kernels: + - New 4x8 block size variant of matrix multiplication of QAI8DXP LHS and QSI4C32P RHS with F32 output. + - Optimizations for FEAT_DotProd. - Added demonstration of integration using CMake in F16 Arm® Neon™ matrix multiplication example. ## v1.3.0 -- GitLab From 5c7039f2c41fd178d1095adbe298f924b56ebac1 Mon Sep 17 00:00:00 2001 From: Michael Kozlov Date: Mon, 3 Feb 2025 12:35:50 +0000 Subject: [PATCH 3/3] Replace end directive Signed-off-by: Michael Kozlov --- ...tmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S | 2 ++ 1 file changed, 2 insertions(+) diff --git a/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S index 7d9ef2d3..2e241d55 100644 --- a/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S +++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod_asm.S @@ -360,3 +360,5 @@ KAI_ASM_LABEL(label_12) // Row loop skip ldp x20, x21, [sp], 144 ret KAI_ASM_FUNCTION_END(kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p8x4_4x8_neon_dotprod) + + KAI_ASM_END -- GitLab