KleidiAI Coverage Report

Directory: ./
Coverage: low: ≥ 0% medium: ≥ 75.0% high: ≥ 90.0%
Coverage Exec / Excl / Total
Lines: 95.0% 38 / 5 / 45
Functions: 90.9% 10 / 0 / 11
Branches: -% 0 / 10 / 10
kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.c
Line Branch Exec Source
1 //
2 // SPDX-FileCopyrightText: Copyright 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_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla.h"
12
13 #include <stddef.h>
14 #include <stdint.h>
15
16 #include "kai/kai_common.h"
17
18 typedef struct {
19 uint16_t maxval;
20 uint16_t minval;
21 const void* A_ptr;
22 const void* B_ptr;
23 size_t N;
24 size_t K;
25 void* output_ptr;
26 uint64_t flags;
27 } KernelArgs;
28
29 static const size_t kai_m_step = 1;
30 static const size_t kai_nr = 2;
31 static const size_t kai_n_step = 8;
32 static const size_t kai_kr = 2;
33 static const size_t kai_sr = 1;
34
35 void kai_kernel_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(KernelArgs* args_ptr);
36
37 uint16_t kai_f16_from_float_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(float value);
38
39 156 size_t kai_get_m_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(void) {
40 156 return kai_m_step;
41 }
42
43 780 size_t kai_get_n_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(void) {
44 780 return kai_n_step * kai_get_sme_vector_length_u16() / kai_kr;
45 }
46
47 156 size_t kai_get_nr_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(void) {
48 156 return kai_nr * kai_get_sme_vector_length_u16() / kai_kr;
49 }
50
51 156 size_t kai_get_kr_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(void) {
52 156 return kai_kr;
53 }
54
55 156 size_t kai_get_sr_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(void) {
56 156 return kai_sr;
57 }
58
59 ✗ size_t kai_get_lhs_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(size_t m_idx, size_t k) {
60 − KAI_ASSUME(m_idx == 0);
61
62 ✗ return m_idx * k;
63 }
64
65 156 static size_t kai_get_rhs_packed_stride_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(size_t k) {
66 312 return kai_get_n_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla() *
67 156 (kai_roundup(k, kai_kr) * sizeof(uint16_t) + sizeof(uint16_t));
68 }
69
70 156 size_t kai_get_rhs_packed_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(size_t n_idx, size_t k) {
71 − KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla() == 0);
72
73 156 const size_t block_idx = n_idx / kai_get_n_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla();
74 312 return block_idx * kai_get_rhs_packed_stride_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(k);
75 156 }
76
77 156 size_t kai_get_dst_offset_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(size_t m_idx, size_t n_idx, size_t dst_stride) {
78 − KAI_ASSUME(m_idx == 0);
79 − KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla() == 0);
80
81 156 return (m_idx * dst_stride) + (n_idx * sizeof(uint16_t));
82 }
83
84 156 size_t kai_get_dst_size_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(size_t m, size_t n) {
85 156 return m * n * sizeof(uint16_t);
86 }
87
88 158 void kai_run_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(
89 size_t m, size_t n, size_t k, const void* lhs, size_t lhs_stride, const void* rhs_packed, void* dst,
90 size_t dst_stride_row, size_t dst_stride_col, float clamp_min, float clamp_max) {
91 158 KAI_UNUSED(dst_stride_row);
92 158 KAI_UNUSED(dst_stride_col);
93 158 KAI_UNUSED(lhs_stride);
94 − KAI_ASSUME(m == 1);
95
96 158 uint64_t flags = 2;
97
98 158 KernelArgs args;
99
100 158 args.minval = kai_f16_from_float_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(clamp_min);
101 158 args.maxval = kai_f16_from_float_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(clamp_max);
102 158 args.A_ptr = lhs;
103 158 args.B_ptr = rhs_packed;
104 158 args.N = n;
105 158 args.K = k;
106 158 args.output_ptr = dst;
107 158 args.flags = flags;
108
109 158 kai_commit_za();
110
111 158 kai_kernel_matmul_clamp_f16_f16_f16p2vlx2b_1x8vl_sme_mla(&args);
112 158 }
113
114 #endif // Architectural features check.
115