KleidiAI Coverage Report

Directory: ./
File: kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4c32p/kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.c
Date: 2025-10-20 13:18:31
Coverage Exec Excl Total
Lines: 98.4% 62 11 74
Functions: 100.0% 16 0 16
Branches: 75.0% 3 22 26
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 #if !defined(__aarch64__) && !defined(__ARM_FEATURE_DOTPROD) && !defined(_M_ARM64)
7 #error "Dotprod extension required to compile this micro-kernel"
8 #else // Architectural features check.
9
10 #include "kai_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
11
12 #include <stddef.h>
13 #include <stdint.h>
14
15 #include "kai/kai_common.h"
16
17 typedef struct {
18 float* dst;
19 const void* lhs_packed;
20 const void* rhs_packed;
21 const float* clamp_vals;
22 size_t dst_stride_row;
23 size_t m;
24 size_t n;
25 size_t num_blocks;
26 size_t num_subblocks;
27 } KernelArgs;
28
29 void kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(KernelArgs* args_ptr);
30 void kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_opt32_neon_dotprod(KernelArgs* args_ptr);
31
32 // Compute args
33 static const size_t kai_m_step = 16;
34 static const size_t kai_n_step = 4;
35 // Packing args
36 static const size_t kai_mr = 4;
37 static const size_t kai_nr = 4;
38 static const size_t kai_kr = 8;
39 static const size_t kai_sr = 2;
40 // LHS format args (num. bytes per value, multiplier, zero_point (if asymmetric))
41 static const size_t kai_num_bytes_qvalue_lhs = 1;
42 static const size_t kai_num_bytes_multiplier_lhs = 4;
43 static const size_t kai_num_bytes_zp_lhs = 4;
44 // RHS format args (num. bytes per value, multiplier, zero_point (if asymmetric), and reduction sum (if LHS is
45 // asymmetric))
46 static const size_t kai_num_bytes_recip_qvalue_rhs = 2;
47 static const size_t kai_num_bytes_multiplier_rhs = 2;
48 static const size_t kai_num_bytes_rsum_rhs = 4;
49 // DST format args
50 static const size_t kai_num_bytes_dst_value = 4;
51 // Extra args
52 static const size_t kai_num_bytes_bias = 4;
53 static const size_t kai_k_multiple_of = 32;
54 static const size_t kai_bl = 32;
55
56 108 inline static size_t kai_get_k_roundedup(size_t k) {
57 108 return kai_roundup(k, kai_k_multiple_of);
58 }
59
60 108 inline static size_t kai_get_num_bytes_per_block_rhs(size_t bl) {
61 − KAI_ASSUME((bl % kai_bl) == 0);
62 108 size_t num_bytes_per_block_rhs = (bl / kai_num_bytes_recip_qvalue_rhs) + kai_num_bytes_multiplier_rhs;
63 216 return num_bytes_per_block_rhs;
64 108 }
65
66 271 inline static size_t kai_get_num_blocks_per_row(size_t k, size_t bl) {
67 − KAI_ASSUME((bl % kai_bl) == 0);
68
69 271 return kai_roundup(k, bl) / bl;
70 }
71
72 108 inline static size_t kai_get_lhs_packed_stride(size_t k) {
73 108 const size_t k_internal = kai_get_k_roundedup(k);
74 108 size_t lhs_packed_stride = kai_mr * ((k_internal * kai_num_bytes_qvalue_lhs) + kai_num_bytes_multiplier_lhs);
75 // Since the LHS matrix is asymmetric with per-row quantization, we must include the
76 // the number of bytes to hold the zero point value
77 108 lhs_packed_stride += kai_mr * kai_num_bytes_zp_lhs;
78
79 216 return lhs_packed_stride;
80 108 }
81
82 108 inline static size_t kai_get_rhs_packed_stride(size_t k, size_t bl) {
83 − KAI_ASSUME((bl % kai_bl) == 0);
84
85 108 const size_t num_blocks_per_row = kai_get_num_blocks_per_row(k, bl);
86 108 const size_t num_bytes_per_block = kai_get_num_bytes_per_block_rhs(bl);
87
88 108 size_t rhs_packed_stride = kai_nr * (num_bytes_per_block * num_blocks_per_row);
89 // Since the LHS matrix is quantized asymmetric with per-row quantization, we also include
90 // the number of bytes for the reduction sum
91 108 rhs_packed_stride += kai_nr * kai_num_bytes_rsum_rhs;
92 // Since the bias is packed with the RHS matrix, the stride is adjusted with the number of bytes of the bias
93 108 rhs_packed_stride += kai_nr * kai_num_bytes_bias;
94
95 216 return rhs_packed_stride;
96 108 }
97
98 112 size_t kai_get_m_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(void) {
99 112 return kai_m_step;
100 }
101
102 112 size_t kai_get_n_step_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(void) {
103 112 return kai_n_step;
104 }
105
106 112 size_t kai_get_mr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(void) {
107 112 return kai_mr;
108 }
109
110 112 size_t kai_get_nr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(void) {
111 112 return kai_nr;
112 }
113
114 112 size_t kai_get_kr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(void) {
115 112 return kai_kr;
116 }
117
118 112 size_t kai_get_sr_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(void) {
119 112 return kai_sr;
120 }
121
122 108 size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(size_t m_idx, size_t k) {
123 − KAI_ASSUME((m_idx % kai_m_step) == 0);
124
125 108 return (m_idx / kai_mr) * kai_get_lhs_packed_stride(k);
126 }
127
128 108 size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(
129 size_t n_idx, size_t k, size_t bl) {
130 − KAI_ASSUME((k % bl) == 0);
131 − KAI_ASSUME((n_idx % kai_n_step) == 0);
132
133 108 return (n_idx / kai_nr) * kai_get_rhs_packed_stride(k, bl);
134 }
135
136 108 size_t kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(
137 size_t m_idx, size_t n_idx, size_t dst_stride) {
138 − KAI_ASSUME((m_idx % kai_m_step) == 0);
139 − KAI_ASSUME((n_idx % kai_n_step) == 0);
140
141 108 return (n_idx * kai_num_bytes_dst_value) + m_idx * dst_stride;
142 }
143
144 108 size_t kai_get_dst_size_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(size_t m, size_t n) {
145 108 return m * n * kai_num_bytes_dst_value;
146 }
147
148 163 void kai_run_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(
149 size_t m, //
150 size_t n, //
151 size_t k, //
152 size_t bl, //
153 const void* restrict lhs_packed, //
154 const void* restrict rhs_packed, //
155 float* restrict dst, // NOLINT(readability-non-const-parameter)
156 size_t dst_stride_row, //
157 size_t dst_stride_col, //
158 float scalar_min, //
159 float scalar_max) {
160 − KAI_ASSUME(dst_stride_col == sizeof(float));
161 − KAI_ASSUME((k % bl) == 0);
162 − KAI_ASSUME((bl % kai_bl) == 0);
163
164
1/2
✓ Branch 0 taken 163 times.
✗ Branch 1 not taken.
 163 if (m == 0) {
165 ✗ return;
166 }
167 163 const size_t num_subblocks = bl / kai_bl;
168 163 const size_t num_blocks = kai_get_num_blocks_per_row(k, bl);
169 163 const float clamp_vals[2] = {scalar_min, scalar_max};
170
171 163 KernelArgs args;
172
173 163 args.dst = dst;
174 163 args.lhs_packed = lhs_packed;
175 163 args.rhs_packed = rhs_packed;
176 163 args.clamp_vals = clamp_vals;
177 163 args.dst_stride_row = dst_stride_row;
178 163 args.m = m;
179 163 args.n = n;
180 163 args.num_blocks = num_blocks;
181 163 args.num_subblocks = num_subblocks;
182
183
2/2
✓ Branch 0 taken 82 times.
✓ Branch 1 taken 81 times.
 163 if (bl == 32) {
184 82 kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_opt32_neon_dotprod(&args);
185 82 } else {
186 81 kai_kernel_matmul_clamp_f32_qai8dxp4x4_qsi4c32p4x4_16x4_neon_dotprod(&args);
187 }
188 163 }
189
190 #endif // Architectural features check.
191