kai/ukernels/matmul/pack/kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.c
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | |||
| 7 | #if (!defined(__aarch64__) || !defined(__ARM_FEATURE_SVE2)) && !defined(_M_ARM64) | ||
| 8 | #error This file must be compiled for AArch64, FEAT_SVE2. | ||
| 9 | #else // Architectural features check. | ||
| 10 | |||
| 11 | #include "kai_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme.h" | ||
| 12 | |||
| 13 | #include <stddef.h> | ||
| 14 | #include <stdint.h> | ||
| 15 | #include <string.h> | ||
| 16 | |||
| 17 | #include "kai/kai_common.h" | ||
| 18 | |||
| 19 | enum { | ||
| 20 | NR = 2, | ||
| 21 | KR = 4, | ||
| 22 | MAX_N_STEP = NR * ((KAI_SME_VEC_LENGTH_MAX_BYTES / sizeof(uint8_t)) / KR), | ||
| 23 | }; | ||
| 24 | |||
| 25 | typedef struct { | ||
| 26 | const void* bias_ptr; | ||
| 27 | const void* scale_ptr; | ||
| 28 | int32_t input_zero_point; | ||
| 29 | float scale_multiplier; | ||
| 30 | size_t width; | ||
| 31 | size_t height; | ||
| 32 | size_t in_stride; | ||
| 33 | size_t out_stride; | ||
| 34 | const void* in; | ||
| 35 | void* out; | ||
| 36 | const void* pad_row; | ||
| 37 | } KernelArgs; | ||
| 38 | |||
| 39 | static const size_t kai_num_bytes_input = sizeof(uint8_t); | ||
| 40 | static const size_t kai_num_bytes_output = sizeof(uint8_t); | ||
| 41 | static const size_t kai_num_bytes_bias = sizeof(int32_t); | ||
| 42 | static const size_t kai_num_bytes_scale = sizeof(float); | ||
| 43 | |||
| 44 | void kai_kernel_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(const KernelArgs* args_ptr); | ||
| 45 | |||
| 46 | 9624 | size_t kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(void) { | |
| 47 | 9624 | return NR * kai_get_sme_vector_length_u8() / KR; | |
| 48 | } | ||
| 49 | |||
| 50 | 834 | size_t kai_get_rhs_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(size_t n_idx) { | |
| 51 | − | KAI_ASSUME(n_idx % kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme() == 0); | |
| 52 | |||
| 53 | 834 | return n_idx * kai_num_bytes_input; | |
| 54 | } | ||
| 55 | |||
| 56 | 834 | size_t kai_get_bias_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(size_t n_idx) { | |
| 57 | 834 | return n_idx * kai_num_bytes_bias; | |
| 58 | } | ||
| 59 | |||
| 60 | 834 | size_t kai_get_scale_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(size_t n_idx) { | |
| 61 | 834 | return n_idx * kai_num_bytes_scale; | |
| 62 | } | ||
| 63 | |||
| 64 | 2652 | size_t kai_get_rhs_packed_stride_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(size_t k) { | |
| 65 | 5304 | return kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme() * | |
| 66 | 2652 | (kai_num_bytes_bias + kai_roundup(k, KR) * kai_num_bytes_output + kai_num_bytes_scale); | |
| 67 | } | ||
| 68 | |||
| 69 | 1818 | size_t kai_get_rhs_packed_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(size_t n_idx, size_t k) { | |
| 70 | − | KAI_ASSUME(n_idx % kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme() == 0); | |
| 71 | |||
| 72 | 1818 | const size_t block_idx = n_idx / kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(); | |
| 73 | 3636 | return block_idx * kai_get_rhs_packed_stride_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(k); | |
| 74 | 1818 | } | |
| 75 | |||
| 76 | 834 | size_t kai_get_rhs_packed_size_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(size_t n, size_t k) { | |
| 77 | 834 | const size_t n_nr_blocks = kai_roundup(n, kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme()); | |
| 78 | 1668 | return kai_get_rhs_packed_offset_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(n_nr_blocks, k); | |
| 79 | 834 | } | |
| 80 | |||
| 81 | 834 | void kai_run_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme( | |
| 82 | size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride_row, const void* rhs, | ||
| 83 | const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, | ||
| 84 | const struct kai_rhs_pack_qsi8cx_params* params) { | ||
| 85 | − | KAI_ASSUME(num_groups == 1); | |
| 86 | − | KAI_ASSUME(nr == kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme()); | |
| 87 | − | KAI_ASSUME(kr == KR); | |
| 88 | − | KAI_ASSUME(sr == 1); | |
| 89 | − | KAI_ASSUME(rhs != NULL); | |
| 90 | − | KAI_ASSUME(bias != NULL); | |
| 91 | − | KAI_ASSUME(scale != NULL); | |
| 92 | − | KAI_ASSUME(rhs_packed != NULL); | |
| 93 | − | KAI_ASSUME(extra_bytes == 0); | |
| 94 | − | KAI_ASSUME(params != NULL); | |
| 95 | |||
| 96 | − | KAI_ASSERT(kai_get_n_step_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme() <= MAX_N_STEP); | |
| 97 | static const uint8_t pad_row[MAX_N_STEP] = {0}; | ||
| 98 | |||
| 99 | 834 | KernelArgs args; | |
| 100 | 834 | args.bias_ptr = bias; | |
| 101 | 834 | args.scale_ptr = scale; | |
| 102 | 834 | args.height = k; | |
| 103 | 834 | args.width = n; | |
| 104 | 834 | args.in = rhs; | |
| 105 | 834 | args.out = rhs_packed; | |
| 106 | 834 | args.in_stride = rhs_stride_row; | |
| 107 | 834 | args.out_stride = kai_get_rhs_packed_stride_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(args.height); | |
| 108 | 834 | args.input_zero_point = params->lhs_zero_point; | |
| 109 | 834 | args.scale_multiplier = params->scale_multiplier; | |
| 110 | 834 | args.pad_row = pad_row; | |
| 111 | |||
| 112 | 834 | kai_commit_za(); | |
| 113 | |||
| 114 | 834 | kai_kernel_rhs_pack_kxn_qsi8cxp2vlx4sb_qs8cx_f32_i32_sme(&args); | |
| 115 | 834 | } | |
| 116 | |||
| 117 | #endif // Architectural features check. | ||
| 118 |