diff --git a/CHANGELOG.md b/CHANGELOG.md index 8cb651ef5234aa8109b4a913efb267449b41d5ff..d72d0653553c34faf8b3462bc33fbeb51c842acd 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -10,6 +10,8 @@ KleidiAI follows the [Semantic Versioning](https://semver.org/) specification fo ## Upcoming Release +- New Advanced SIMD micro-kernels: + - Matrix multiplication (MxN) Micro-kernels of QSI8D32 LHS and QAI4C32 RHS with F16 output, optimized for FEAT_DotProd. - Added Convolution example using SME Indirect Matmul Kernels ## v1.9.0 diff --git a/CMakeLists.txt b/CMakeLists.txt index ce3df662924dabeaf9aaa7de2ea6f18632546407..adf696b81fec7e2a6eebd0300fbb34a3396f417d 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -103,8 +103,6 @@ set(KLEIDIAI_FILES_NEON_FP16 ) set(KLEIDIAI_FILES_NEON_FP16_DOTPROD_ASM - kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod_asm.S - kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod_asm.S kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x4_qsi4cxp4x4_1x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi4cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi4cxp4x8_1x4_neon_dotprod_asm.S @@ -117,6 +115,10 @@ set(KLEIDIAI_FILES_NEON_FP16_DOTPROD_ASM kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod_asm.S kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c kai/ukernels/matmul/matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.c + kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod_asm.S + kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c ) set(KLEIDIAI_FILES_NEON_FP16_DOTPROD diff --git a/kai/ukernels/matmul/BUILD.bazel b/kai/ukernels/matmul/BUILD.bazel index 646e9a2fba5b26559de8ab48b1cd0afd6fc48fee..a6209876983e35513861f07609550ee8464fa5bd 100644 --- a/kai/ukernels/matmul/BUILD.bazel +++ b/kai/ukernels/matmul/BUILD.bazel @@ -80,6 +80,7 @@ FP16_DOTPROD_KERNELS_ASM = [ "matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod", "matmul_clamp_f16_qai8dxp_qsi8cxp/kai_matmul_clamp_f16_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod", "matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod", + "matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod", ] # buildifier: keep sorted diff --git a/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c b/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c new file mode 100644 index 0000000000000000000000000000000000000000..abd61f6171a8956b47e15db543ce974f98842f73 --- /dev/null +++ b/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.c @@ -0,0 +1,179 @@ +// +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// +// SPDX-License-Identifier: Apache-2.0 +// +#if !defined(__aarch64__) && !defined(__ARM_FEATURE_DOTPROD) && !defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && \ + !defined(_M_ARM64) +#error "Dotprod extension and fp16 vector arithmetic required to compile this micro-kernel" +#else // Architectural features check. + +#include "kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.h" + +#include +#include + +#include "kai/kai_common.h" + +typedef struct { + void* 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(KernelArgs* args_ptr); + +// Compute args +static const size_t kai_m_step = 8; +static const size_t kai_n_step = 4; +// Packing args +static const size_t kai_mr = 4; +static const size_t kai_nr = 4; +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_sum_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 = 4; +static const size_t kai_num_bytes_offset_rhs = 4; + +// DST format args +static const size_t kai_num_bytes_dst_value = 2; +// Extra args +static const size_t kai_num_bytes_bias = 4; +static const size_t kai_bl = 32; + +inline static size_t kai_get_num_bytes_per_block_lhs(size_t bl) { + return (bl * kai_num_bytes_qvalue_lhs) + kai_num_bytes_multiplier_lhs + kai_num_bytes_sum_lhs; +} + +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 + kai_num_bytes_offset_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, size_t bl) { + return kai_mr * kai_get_num_blocks_per_row(k, bl) * kai_get_num_bytes_per_block_lhs(bl); +} + +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 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(void) { + return kai_m_step; +} + +size_t kai_get_n_step_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(void) { + return kai_n_step; +} + +size_t kai_get_mr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(void) { + return kai_mr; +} + +size_t kai_get_nr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(void) { + return kai_nr; +} + +size_t kai_get_kr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(void) { + return kai_kr; +} + +size_t kai_get_sr_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(void) { + return kai_sr; +} + +size_t kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod( + size_t m_idx, size_t k, size_t bl) { + KAI_ASSUME((m_idx % kai_m_step) == 0); + + return (m_idx / kai_mr) * kai_get_lhs_packed_stride(k, bl); +} + +size_t kai_get_rhs_packed_offset_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(size_t m, size_t n) { + return m * n * kai_num_bytes_dst_value; +} + +void kai_run_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod( + size_t m, // + size_t n, // + size_t k, // + size_t bl, // + const void* restrict lhs_packed, // + const void* restrict rhs_packed, // + void* 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(uint16_t)); + 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(&args); +} + +#endif // Architectural features check. diff --git a/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.h b/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.h new file mode 100644 index 0000000000000000000000000000000000000000..a2beaf40530343ce68e5a41553affc352c31dd02 --- /dev/null +++ b/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.h @@ -0,0 +1,150 @@ +// +// 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_run_lhs_quant_pack_qsi8d32pscalef32_f16_neon to dynamically quantize and pack the LHS matrix in a single +/// step. +/// -# @ref kai_run_rhs_pack_nxk_qai4c32p_qau4c32s0s1_f32_f32_f32_neon to pack the RHS NxK 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod(void); + +/// Gets the offset in bytes for the packed LHS matrix, +/// which contains the packed Quantized Symmetric Signed 8-bit with per-block (multiple of 32) quantization (qsi8d32) +/// 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). +/// 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 LHS matrix +size_t kai_get_lhs_packed_offset_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod( + size_t m_idx, // + size_t k, // + size_t bl); // + +/// Gets the offset in bytes for the packed RHS matrix, +/// which contains the packed Quantized Asymmetric Signed 4-bit with per-block (multiple of 32) quantization (qai4c32) +/// 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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 Symmetric Signed 8-bit with per-block (multiple of 32) quantization (qsi8d32) and packed. +/// RHS matrix: Quantized Asymmetric Signed 4-bit with per-block (multiple of 32) quantization (qai4c32) 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(uint16_t) 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod( + size_t m, // + size_t n, // + size_t k, // + size_t bl, // + const void* lhs_packed, // + const void* rhs_packed, // + void* 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_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod_asm.S b/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod_asm.S new file mode 100644 index 0000000000000000000000000000000000000000..873f2345197c1ca25295f74dc5271d811516bea8 --- /dev/null +++ b/kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod_asm.S @@ -0,0 +1,532 @@ +// +// 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod) + KAI_ASM_ALIGN + + KAI_ASM_GLOBAL(kai_kernel_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod) + +KAI_ASM_FUNCTION_TYPE(kai_kernel_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod) +KAI_ASM_FUNCTION_LABEL(kai_kernel_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_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 x21, #0x20 + movi v11.16b, #0xf0 + mov x6, #0x80 + ldr x20, [x0, #0x28] + ldr x7, [x0, #0x40] + ldr x8, [x0, #0x38] + ldr x17, [x0, #0x8] + ldr x16, [x0, #0x10] + ldr x15, [x0, #0x30] + mov x14, x20 + madd x6, x7, x6, x21 + ldr x13, [x0, #0x0] + ldr x12, [x0, #0x20] + ldr x11, [x0, #0x18] + cmp x14, #0x8 + mul x6, x8, x6 + blt label_11 +KAI_ASM_LABEL(label_1) // Row loop + mov x10, x16 + mov x9, x15 + add x28, x13, x12, LSL #3 +KAI_ASM_LABEL(label_2) // Column loop + mov x23, x17 + movi v0.16b, #0x0 + movi v12.16b, #0x0 + mov x22, x8 + movi v14.16b, #0x0 + movi v13.16b, #0x0 + movi v15.16b, #0x0 + movi v9.16b, #0x0 + movi v5.16b, #0x0 + movi v8.16b, #0x0 + add x21, x23, x6 +KAI_ASM_LABEL(label_3) // Block loop + movi v1.4s, #0x0 + movi v7.4s, #0x0 + mov x20, x7 + movi v3.4s, #0x0 + movi v4.4s, #0x0 + movi v23.4s, #0x0 + movi v6.4s, #0x0 + movi v10.4s, #0x0 + movi v2.4s, #0x0 +KAI_ASM_LABEL(label_4) // Sub block loop + ldr q31, [x10, #0x0] + ldr q22, [x23, #0x0] + subs x20, x20, #0x1 + ldr q30, [x21, #0x0] + ldr q21, [x10, #0x10] + ldr q20, [x23, #0x10] + ldr q19, [x21, #0x10] + ldr q29, [x10, #0x20] + ldr q18, [x23, #0x20] + shl v28.16b, v31.16b, #0x4 + and v31.16b, v31.16b, v11.16b + ldr q17, [x21, #0x20] + ldr q27, [x10, #0x30] + shl v16.16b, v21.16b, #0x4 + and v21.16b, v21.16b, v11.16b + ldr q26, [x23, #0x30] + ldr q25, [x21, #0x30] + add x10, x10, #0x40 + ldr q24, [x23, #0x40] + KAI_ASM_INST(0x4f96e381) // sdot v1.4s, v28.16b, v22.4b[0] + KAI_ASM_INST(0x4fb6e387) // sdot v7.4s, v28.16b, v22.4b[1] + KAI_ASM_INST(0x4f96eb83) // sdot v3.4s, v28.16b, v22.4b[2] + KAI_ASM_INST(0x4fb6eb84) // sdot v4.4s, v28.16b, v22.4b[3] + ldr q22, [x21, #0x40] + KAI_ASM_INST(0x4f9ee397) // sdot v23.4s, v28.16b, v30.4b[0] + KAI_ASM_INST(0x4fbee386) // sdot v6.4s, v28.16b, v30.4b[1] + KAI_ASM_INST(0x4f9eeb8a) // sdot v10.4s, v28.16b, v30.4b[2] + KAI_ASM_INST(0x4fbeeb82) // sdot v2.4s, v28.16b, v30.4b[3] + ldr q28, [x23, #0x50] + ldr q30, [x21, #0x50] + KAI_ASM_INST(0x4f94e201) // sdot v1.4s, v16.16b, v20.4b[0] + KAI_ASM_INST(0x4fb4e207) // sdot v7.4s, v16.16b, v20.4b[1] + KAI_ASM_INST(0x4f94ea03) // sdot v3.4s, v16.16b, v20.4b[2] + KAI_ASM_INST(0x4fb4ea04) // sdot v4.4s, v16.16b, v20.4b[3] + ldr q20, [x23, #0x60] + KAI_ASM_INST(0x4f93e217) // sdot v23.4s, v16.16b, v19.4b[0] + KAI_ASM_INST(0x4fb3e206) // sdot v6.4s, v16.16b, v19.4b[1] + KAI_ASM_INST(0x4f93ea0a) // sdot v10.4s, v16.16b, v19.4b[2] + KAI_ASM_INST(0x4fb3ea02) // sdot v2.4s, v16.16b, v19.4b[3] + ldr q19, [x21, #0x60] + shl v16.16b, v29.16b, #0x4 + and v29.16b, v29.16b, v11.16b + KAI_ASM_INST(0x4f92e201) // sdot v1.4s, v16.16b, v18.4b[0] + KAI_ASM_INST(0x4fb2e207) // sdot v7.4s, v16.16b, v18.4b[1] + KAI_ASM_INST(0x4f92ea03) // sdot v3.4s, v16.16b, v18.4b[2] + KAI_ASM_INST(0x4fb2ea04) // sdot v4.4s, v16.16b, v18.4b[3] + ldr q18, [x23, #0x70] + add x23, x23, #0x80 + KAI_ASM_INST(0x4f91e217) // sdot v23.4s, v16.16b, v17.4b[0] + KAI_ASM_INST(0x4fb1e206) // sdot v6.4s, v16.16b, v17.4b[1] + KAI_ASM_INST(0x4f91ea0a) // sdot v10.4s, v16.16b, v17.4b[2] + KAI_ASM_INST(0x4fb1ea02) // sdot v2.4s, v16.16b, v17.4b[3] + ldr q17, [x21, #0x70] + shl v16.16b, v27.16b, #0x4 + and v27.16b, v27.16b, v11.16b + add x21, x21, #0x80 + KAI_ASM_INST(0x4f9ae201) // sdot v1.4s, v16.16b, v26.4b[0] + KAI_ASM_INST(0x4fbae207) // sdot v7.4s, v16.16b, v26.4b[1] + KAI_ASM_INST(0x4f9aea03) // sdot v3.4s, v16.16b, v26.4b[2] + KAI_ASM_INST(0x4fbaea04) // sdot v4.4s, v16.16b, v26.4b[3] + KAI_ASM_INST(0x4f99e217) // sdot v23.4s, v16.16b, v25.4b[0] + KAI_ASM_INST(0x4fb9e206) // sdot v6.4s, v16.16b, v25.4b[1] + KAI_ASM_INST(0x4f99ea0a) // sdot v10.4s, v16.16b, v25.4b[2] + KAI_ASM_INST(0x4fb9ea02) // sdot v2.4s, v16.16b, v25.4b[3] + KAI_ASM_INST(0x4f98e3e1) // sdot v1.4s, v31.16b, v24.4b[0] + KAI_ASM_INST(0x4fb8e3e7) // sdot v7.4s, v31.16b, v24.4b[1] + KAI_ASM_INST(0x4f98ebe3) // sdot v3.4s, v31.16b, v24.4b[2] + KAI_ASM_INST(0x4fb8ebe4) // sdot v4.4s, v31.16b, v24.4b[3] + KAI_ASM_INST(0x4f96e3f7) // sdot v23.4s, v31.16b, v22.4b[0] + KAI_ASM_INST(0x4fb6e3e6) // sdot v6.4s, v31.16b, v22.4b[1] + KAI_ASM_INST(0x4f96ebea) // sdot v10.4s, v31.16b, v22.4b[2] + KAI_ASM_INST(0x4fb6ebe2) // sdot v2.4s, v31.16b, v22.4b[3] + KAI_ASM_INST(0x4f9ce2a1) // sdot v1.4s, v21.16b, v28.4b[0] + KAI_ASM_INST(0x4fbce2a7) // sdot v7.4s, v21.16b, v28.4b[1] + KAI_ASM_INST(0x4f9ceaa3) // sdot v3.4s, v21.16b, v28.4b[2] + KAI_ASM_INST(0x4fbceaa4) // sdot v4.4s, v21.16b, v28.4b[3] + KAI_ASM_INST(0x4f9ee2b7) // sdot v23.4s, v21.16b, v30.4b[0] + KAI_ASM_INST(0x4fbee2a6) // sdot v6.4s, v21.16b, v30.4b[1] + KAI_ASM_INST(0x4f9eeaaa) // sdot v10.4s, v21.16b, v30.4b[2] + KAI_ASM_INST(0x4fbeeaa2) // sdot v2.4s, v21.16b, v30.4b[3] + KAI_ASM_INST(0x4f94e3a1) // sdot v1.4s, v29.16b, v20.4b[0] + KAI_ASM_INST(0x4fb4e3a7) // sdot v7.4s, v29.16b, v20.4b[1] + KAI_ASM_INST(0x4f94eba3) // sdot v3.4s, v29.16b, v20.4b[2] + KAI_ASM_INST(0x4fb4eba4) // sdot v4.4s, v29.16b, v20.4b[3] + KAI_ASM_INST(0x4f93e3b7) // sdot v23.4s, v29.16b, v19.4b[0] + KAI_ASM_INST(0x4fb3e3a6) // sdot v6.4s, v29.16b, v19.4b[1] + KAI_ASM_INST(0x4f93ebaa) // sdot v10.4s, v29.16b, v19.4b[2] + KAI_ASM_INST(0x4fb3eba2) // sdot v2.4s, v29.16b, v19.4b[3] + KAI_ASM_INST(0x4f92e361) // sdot v1.4s, v27.16b, v18.4b[0] + KAI_ASM_INST(0x4fb2e367) // sdot v7.4s, v27.16b, v18.4b[1] + KAI_ASM_INST(0x4f92eb63) // sdot v3.4s, v27.16b, v18.4b[2] + KAI_ASM_INST(0x4fb2eb64) // sdot v4.4s, v27.16b, v18.4b[3] + KAI_ASM_INST(0x4f91e377) // sdot v23.4s, v27.16b, v17.4b[0] + KAI_ASM_INST(0x4fb1e366) // sdot v6.4s, v27.16b, v17.4b[1] + KAI_ASM_INST(0x4f91eb6a) // sdot v10.4s, v27.16b, v17.4b[2] + KAI_ASM_INST(0x4fb1eb62) // sdot v2.4s, v27.16b, v17.4b[3] + bgt label_4 + ldr q21, [x10, #0x0] + ld1 { v17.4s }, [x23] + add x23, x23, #0x10 + scvtf v1.4s, v1.4s + ldr q20, [x10, #0x10] + ldr q16, [x23, #0x0] + scvtf v7.4s, v7.4s + scvtf v3.4s, v3.4s + scvtf v4.4s, v4.4s + add x10, x10, #0x20 + add x23, x23, #0x10 + fmla v0.4s, v21.4s, v17.s[0] + fmla v12.4s, v21.4s, v17.s[1] + fmla v14.4s, v21.4s, v17.s[2] + fmla v13.4s, v21.4s, v17.s[3] + fmul v19.4s, v20.4s, v16.s[0] + fmul v18.4s, v20.4s, v16.s[1] + fmul v17.4s, v20.4s, v16.s[2] + fmul v16.4s, v20.4s, v16.s[3] + fmla v0.4s, v1.4s, v19.4s + fmla v12.4s, v7.4s, v18.4s + fmla v14.4s, v3.4s, v17.4s + fmla v13.4s, v4.4s, v16.4s + ld1 { v17.4s }, [x21] + add x21, x21, #0x10 + scvtf v23.4s, v23.4s + scvtf v6.4s, v6.4s + ldr q16, [x21, #0x0] + scvtf v10.4s, v10.4s + scvtf v2.4s, v2.4s + add x21, x21, #0x10 + fmla v15.4s, v21.4s, v17.s[0] + fmla v9.4s, v21.4s, v17.s[1] + fmla v5.4s, v21.4s, v17.s[2] + fmla v8.4s, v21.4s, v17.s[3] + fmul v19.4s, v20.4s, v16.s[0] + fmul v18.4s, v20.4s, v16.s[1] + fmul v17.4s, v20.4s, v16.s[2] + fmul v16.4s, v20.4s, v16.s[3] + fmla v15.4s, v23.4s, v19.4s + fmla v9.4s, v6.4s, v18.4s + fmla v5.4s, v10.4s, v17.4s + fmla v8.4s, v2.4s, v16.4s + subs x22, x22, #0x1 + bgt label_3 + ldr q18, [x10, #0x0] + ld1r { v17.4s }, [x11] + add x20, x11, #0x4 + cmp x9, #0x4 + ld1r { v16.4s }, [x20] + add x10, x10, #0x10 + fadd v0.4s, v0.4s, v18.4s + fadd v12.4s, v12.4s, v18.4s + fadd v14.4s, v14.4s, v18.4s + fadd v13.4s, v13.4s, v18.4s + fadd v15.4s, v15.4s, v18.4s + fadd v9.4s, v9.4s, v18.4s + fadd v5.4s, v5.4s, v18.4s + fadd v8.4s, v8.4s, 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 v9.4s, v9.4s, v17.4s + fmax v5.4s, v5.4s, v17.4s + fmax v8.4s, v8.4s, 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 v9.4s, v9.4s, v16.4s + fmin v5.4s, v5.4s, v16.4s + fmin v8.4s, v8.4s, v16.4s + fcvtn v23.4h, v0.4s + fcvtn v22.4h, v12.4s + fcvtn v21.4h, v14.4s + fcvtn v20.4h, v13.4s + fcvtn v19.4h, v15.4s + fcvtn v18.4h, v9.4s + fcvtn v17.4h, v5.4s + fcvtn v16.4h, v8.4s + blt label_7 + mov x20, x13 + str d23, [x20, #0x0] + add x20, x20, x12 + str d22, [x20, #0x0] + add x20, x20, x12 + str d21, [x20, #0x0] + add x20, x20, x12 + str d20, [x20, #0x0] + add x20, x20, x12 + str d19, [x20, #0x0] + add x20, x20, x12 + str d18, [x20, #0x0] + add x20, x20, x12 + str d17, [x20, #0x0] + add x20, x20, x12 + str d16, [x20, #0x0] + b label_10 +KAI_ASM_LABEL(label_7) // Partial output + mov x27, x13 + add x26, x27, x12, LSL #2 + add x25, x26, x12, LSL #1 + add x24, x26, x12 + add x23, x25, x12 + add x22, x27, x12, LSL #1 + add x21, x27, x12 + add x20, x22, x12 + tbz x9, #1, label_8 + st1 { v16.s }[0], [x23], #0x4 + st1 { v17.s }[0], [x25], #0x4 + st1 { v18.s }[0], [x24], #0x4 + st1 { v19.s }[0], [x26], #0x4 + st1 { v20.s }[0], [x20], #0x4 + st1 { v21.s }[0], [x22], #0x4 + st1 { v22.s }[0], [x21], #0x4 + st1 { v23.s }[0], [x27], #0x4 + tbz x9, #0, label_9 + st1 { v16.h }[2], [x23] + st1 { v17.h }[2], [x25] + st1 { v18.h }[2], [x24] + st1 { v19.h }[2], [x26] + st1 { v20.h }[2], [x20] + st1 { v21.h }[2], [x22] + st1 { v22.h }[2], [x21] + st1 { v23.h }[2], [x27] + b label_9 +KAI_ASM_LABEL(label_8) // Output block 0: partial_1_0 + st1 { v16.h }[0], [x23] + st1 { v17.h }[0], [x25] + st1 { v18.h }[0], [x24] + st1 { v19.h }[0], [x26] + st1 { v20.h }[0], [x20] + st1 { v21.h }[0], [x22] + st1 { v22.h }[0], [x21] + st1 { v23.h }[0], [x27] +KAI_ASM_LABEL(label_9) // Output block 0: Done +KAI_ASM_LABEL(label_10) // Output stage exit + subs x9, x9, #0x4 + add x13, x13, #0x8 + bgt label_2 + mov x20, #0x2 + sub x14, x14, #0x8 + cmp x14, #0x8 + mov x13, x28 + madd x17, x20, x6, x17 + bge label_1 +KAI_ASM_LABEL(label_11) // Row loop skip + cbz x14, label_21 +KAI_ASM_LABEL(label_12) // Row tail: Row loop + mov x26, x16 + mov x25, x15 + add x24, x13, x12, LSL #2 +KAI_ASM_LABEL(label_13) // Row tail: Column loop + movi v0.16b, #0x0 + movi v12.16b, #0x0 + mov x23, x17 + mov x21, x8 + movi v14.16b, #0x0 + movi v13.16b, #0x0 +KAI_ASM_LABEL(label_14) // Row tail: Block loop + movi v1.4s, #0x0 + movi v7.4s, #0x0 + mov x20, x7 + movi v3.4s, #0x0 + movi v4.4s, #0x0 +KAI_ASM_LABEL(label_15) // Row tail: Sub block loop + ldr q31, [x26, #0x0] + ldr q30, [x23, #0x0] + subs x20, x20, #0x1 + ldr q29, [x26, #0x10] + ldr q28, [x23, #0x10] + ldr q27, [x26, #0x20] + ldr q26, [x23, #0x20] + ldr q25, [x26, #0x30] + ldr q24, [x23, #0x30] + shl v23.16b, v31.16b, #0x4 + and v31.16b, v31.16b, v11.16b + ldr q22, [x23, #0x40] + ldr q21, [x23, #0x50] + shl v20.16b, v29.16b, #0x4 + and v29.16b, v29.16b, v11.16b + ldr q19, [x23, #0x60] + ldr q18, [x23, #0x70] + shl v17.16b, v27.16b, #0x4 + and v27.16b, v27.16b, v11.16b + KAI_ASM_INST(0x4f9ee2e1) // sdot v1.4s, v23.16b, v30.4b[0] + KAI_ASM_INST(0x4fbee2e7) // sdot v7.4s, v23.16b, v30.4b[1] + shl v16.16b, v25.16b, #0x4 + add x26, x26, #0x40 + KAI_ASM_INST(0x4f9eeae3) // sdot v3.4s, v23.16b, v30.4b[2] + KAI_ASM_INST(0x4fbeeae4) // sdot v4.4s, v23.16b, v30.4b[3] + and v25.16b, v25.16b, v11.16b + add x23, x23, #0x80 + KAI_ASM_INST(0x4f9ce281) // sdot v1.4s, v20.16b, v28.4b[0] + KAI_ASM_INST(0x4fbce287) // sdot v7.4s, v20.16b, v28.4b[1] + KAI_ASM_INST(0x4f9cea83) // sdot v3.4s, v20.16b, v28.4b[2] + KAI_ASM_INST(0x4fbcea84) // sdot v4.4s, v20.16b, v28.4b[3] + KAI_ASM_INST(0x4f9ae221) // sdot v1.4s, v17.16b, v26.4b[0] + KAI_ASM_INST(0x4fbae227) // sdot v7.4s, v17.16b, v26.4b[1] + KAI_ASM_INST(0x4f9aea23) // sdot v3.4s, v17.16b, v26.4b[2] + KAI_ASM_INST(0x4fbaea24) // sdot v4.4s, v17.16b, v26.4b[3] + KAI_ASM_INST(0x4f98e201) // sdot v1.4s, v16.16b, v24.4b[0] + KAI_ASM_INST(0x4fb8e207) // sdot v7.4s, v16.16b, v24.4b[1] + KAI_ASM_INST(0x4f98ea03) // sdot v3.4s, v16.16b, v24.4b[2] + KAI_ASM_INST(0x4fb8ea04) // sdot v4.4s, v16.16b, v24.4b[3] + KAI_ASM_INST(0x4f96e3e1) // sdot v1.4s, v31.16b, v22.4b[0] + KAI_ASM_INST(0x4fb6e3e7) // sdot v7.4s, v31.16b, v22.4b[1] + KAI_ASM_INST(0x4f96ebe3) // sdot v3.4s, v31.16b, v22.4b[2] + KAI_ASM_INST(0x4fb6ebe4) // sdot v4.4s, v31.16b, v22.4b[3] + KAI_ASM_INST(0x4f95e3a1) // sdot v1.4s, v29.16b, v21.4b[0] + KAI_ASM_INST(0x4fb5e3a7) // sdot v7.4s, v29.16b, v21.4b[1] + KAI_ASM_INST(0x4f95eba3) // sdot v3.4s, v29.16b, v21.4b[2] + KAI_ASM_INST(0x4fb5eba4) // sdot v4.4s, v29.16b, v21.4b[3] + KAI_ASM_INST(0x4f93e361) // sdot v1.4s, v27.16b, v19.4b[0] + KAI_ASM_INST(0x4fb3e367) // sdot v7.4s, v27.16b, v19.4b[1] + KAI_ASM_INST(0x4f93eb63) // sdot v3.4s, v27.16b, v19.4b[2] + KAI_ASM_INST(0x4fb3eb64) // sdot v4.4s, v27.16b, v19.4b[3] + KAI_ASM_INST(0x4f92e321) // sdot v1.4s, v25.16b, v18.4b[0] + KAI_ASM_INST(0x4fb2e327) // sdot v7.4s, v25.16b, v18.4b[1] + KAI_ASM_INST(0x4f92eb23) // sdot v3.4s, v25.16b, v18.4b[2] + KAI_ASM_INST(0x4fb2eb24) // sdot v4.4s, v25.16b, v18.4b[3] + bgt label_15 + ldr q18, [x26, #0x0] + ld1 { v17.4s }, [x23] + add x23, x23, #0x10 + scvtf v1.4s, v1.4s + ldr q20, [x26, #0x10] + ldr q16, [x23, #0x0] + scvtf v7.4s, v7.4s + scvtf v3.4s, v3.4s + scvtf v4.4s, v4.4s + add x26, x26, #0x20 + add x23, x23, #0x10 + fmla v0.4s, v18.4s, v17.s[0] + fmla v12.4s, v18.4s, v17.s[1] + fmla v14.4s, v18.4s, v17.s[2] + fmla v13.4s, v18.4s, v17.s[3] + fmul v19.4s, v20.4s, v16.s[0] + fmul v18.4s, v20.4s, v16.s[1] + fmul v17.4s, v20.4s, v16.s[2] + fmul v16.4s, v20.4s, v16.s[3] + fmla v0.4s, v1.4s, v19.4s + fmla v12.4s, v7.4s, v18.4s + fmla v14.4s, v3.4s, v17.4s + fmla v13.4s, v4.4s, v16.4s + subs x21, x21, #0x1 + bgt label_14 + ldr q18, [x26, #0x0] + ld1r { v17.4s }, [x11] + add x20, x11, #0x4 + cmp x25, #0x4 + ld1r { v16.4s }, [x20] + add x26, x26, #0x10 + fadd v0.4s, v0.4s, v18.4s + fadd v12.4s, v12.4s, v18.4s + fadd v14.4s, v14.4s, v18.4s + fadd v13.4s, v13.4s, 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 + 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 + fcvtn v19.4h, v0.4s + fcvtn v18.4h, v12.4s + fcvtn v17.4h, v14.4s + fcvtn v16.4h, v13.4s + blt label_17 + mov x20, x13 + cmp x14, #0x1 + str d19, [x20, #0x0] + add x20, x20, x12 + ble label_20 + cmp x14, #0x2 + str d18, [x20, #0x0] + add x20, x20, x12 + ble label_20 + cmp x14, #0x3 + str d17, [x20, #0x0] + add x20, x20, x12 + ble label_20 + str d16, [x20, #0x0] + b label_20 +KAI_ASM_LABEL(label_17) // Row tail: Partial output + mov x23, x13 + cmp x14, #0x1 + add x22, x23, x12 + csel x22, x22, x23, GT + cmp x14, #0x2 + add x21, x23, x12, LSL #1 + csel x21, x21, x22, GT + cmp x14, #0x3 + add x20, x21, x12 + csel x20, x20, x21, GT + tbz x25, #1, label_18 + st1 { v16.s }[0], [x20], #0x4 + st1 { v17.s }[0], [x21], #0x4 + st1 { v18.s }[0], [x22], #0x4 + st1 { v19.s }[0], [x23], #0x4 + tbz x25, #0, label_19 + st1 { v16.h }[2], [x20] + st1 { v17.h }[2], [x21] + st1 { v18.h }[2], [x22] + st1 { v19.h }[2], [x23] + b label_19 +KAI_ASM_LABEL(label_18) // Row tail: Output block 0: partial_1_0 + st1 { v16.h }[0], [x20] + st1 { v17.h }[0], [x21] + st1 { v18.h }[0], [x22] + st1 { v19.h }[0], [x23] +KAI_ASM_LABEL(label_19) // Row tail: Output block 0: Done +KAI_ASM_LABEL(label_20) // Row tail: Output stage exit + subs x25, x25, #0x4 + add x13, x13, #0x8 + bgt label_13 + subs x14, x14, #0x4 + add x17, x17, x6 + mov x13, x24 + bgt label_12 +KAI_ASM_LABEL(label_21) // Row tail: 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_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod) + + KAI_ASM_END diff --git a/test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp b/test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp index 5ba1a68758867f07b0673697b6a7ea4d2b1df06f..72a535c5d29e8760a8a0513419dbade22f659102 100644 --- a/test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp +++ b/test/tests/matmul_clamp_f16_qsi8d32p_qai4c32p_test.cpp @@ -15,6 +15,7 @@ #include #include "kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod.h" +#include "kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod.h" #include "kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm.h" #include "kai/ukernels/matmul/matmul_clamp_f16_qsi8d32p_qai4c32p/kai_matmul_clamp_f16_qsi8d32p_qai4c32p_interface.h" #include "kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32pscalef32_f16_neon.h" @@ -38,12 +39,14 @@ namespace kai::test { -static const std::array, 2> +static const std::array, 3> variants_kai_matmul_clamp_f16_qsi8d32p_qai4c32p = { {{UKERNEL_MATMUL_VARIANT(clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod), "kai_matmul_clamp_f16_qsi8d32p1x8_qai4c32p4x8_1x4_neon_dotprod", cpu_has_dotprod_and_fp16}, {UKERNEL_MATMUL_VARIANT(clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm), - "kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm", cpu_has_i8mm_and_fp16}}}; + "kai_matmul_clamp_f16_qsi8d32p4x8_qai4c32p4x8_8x4_neon_i8mm", cpu_has_i8mm_and_fp16}, + {UKERNEL_MATMUL_VARIANT(clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod), + "kai_matmul_clamp_f16_qsi8d32p4x4_qai4c32p4x4_8x4_neon_dotprod", cpu_has_dotprod_and_fp16}}}; class MatMulTest_f16_qsi8d32p_qai4c32p : public ::testing::TestWithParam {}; @@ -121,7 +124,7 @@ TEST_P(MatMulTest_f16_qsi8d32p_qai4c32p, EndToEnd) { const auto lhs_start_row = rect.start_row(); const auto imp_packed_lhs_size = kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32pscalef32_f16_neon(M, K, bl, mr, kr, sr); - Buffer imp_packed_lhs(imp_packed_lhs_size); + Buffer imp_packed_lhs(imp_packed_lhs_size, 0); auto lhs_stride = K * sizeof(uint16_t); auto lhs_offset = kai_get_lhs_offset_lhs_quant_pack_qsi8d32pscalef32_f16_neon(lhs_start_row, lhs_stride); @@ -135,6 +138,31 @@ TEST_P(MatMulTest_f16_qsi8d32p_qai4c32p, EndToEnd) { rect.height() /* m */, K, bl, mr, kr, sr, 0, ref_lhs_f16.data() + lhs_offset, lhs_stride, imp_packed_lhs.data() + lhs_packed_offset); + // Verify LHS quant and pack int8 kernel behaves same for int4 and int8, when the block-depth is same for different + // values of kr, sr. + if (sr == 2) { + const size_t kr_qsi8 = kr / sr; + const size_t sr_qsi8 = 1; + const auto imp_packed_lhs_qsi8_size = + kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32pscalef32_f16_neon(M, K, bl, mr, kr_qsi8, sr_qsi8); + Buffer imp_packed_lhs_qsi8(imp_packed_lhs_qsi8_size, 0); + + auto lhs_qsi8_packed_offset = kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32pscalef32_f16_neon( + lhs_start_row, K, bl, mr, kr_qsi8, sr_qsi8); + + ASSERT_EQ(lhs_qsi8_packed_offset, lhs_matmul_offset); + + kai_run_lhs_quant_pack_qsi8d32pscalef32_f16_neon( + rect.height() /* m */, K, bl, mr, kr_qsi8, sr_qsi8, 0, ref_lhs_f16.data() + lhs_offset, lhs_stride, + imp_packed_lhs_qsi8.data() + lhs_qsi8_packed_offset); + + const auto* imp_packed_lhs_ptr = reinterpret_cast(imp_packed_lhs.data()); + const auto* imp_packed_lhs_qsi8_ptr = reinterpret_cast(imp_packed_lhs_qsi8.data()); + for (size_t i = 0; i < imp_packed_lhs_qsi8_size; i++) { + ASSERT_EQ(imp_packed_lhs_ptr[i], imp_packed_lhs_qsi8_ptr[i]); + } + } + // Prepare the offsets as the RHS packing kernel expects the scaled zero-points in float. const size_t num_blocks_per_row = round_up_division(K, bl); const size_t ref_zp_size = N * num_blocks_per_row;