KleidiAI Coverage Report

Directory: ./
Coverage: low: ≥ 0% medium: ≥ 75.0% high: ≥ 90.0%
Coverage Exec / Excl / Total
Lines: 100.0% 39 / 4 / 43
Functions: 100.0% 9 / 0 / 9
Branches: -% 0 / 8 / 8
kai/ukernels/matmul/imatmul_clamp_f16_f16p_f16p/kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.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 #include "kai_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa.h"
11
12 #include <stddef.h>
13 #include <stdint.h>
14
15 #include "kai/kai_common.h"
16
17 typedef struct {
18 const void* A;
19 const void* B;
20 void* C;
21 uint64_t ldcb;
22 uint64_t M;
23 uint64_t N;
24 uint64_t K;
25 uint16_t min;
26 uint16_t max;
27 void* accumulator_buffer;
28 uint64_t flags;
29 } KernelArgs;
30
31 static const size_t kai_mr = 2;
32 static const size_t kai_nr = 2;
33 static const size_t kai_kr = 2;
34
35 void kai_kernel_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(KernelArgs* args);
36 uint16_t kai_f16_from_float_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(float value);
37
38 // Returns a constant value specific to this kernel that's relative to vector length
39 14076 static size_t kai_get_kernel_vec_length_constant(void) {
40 14076 const size_t kernel_vec_length_constant = kai_get_sme_vector_length_u16() / kai_kr;
41 28152 return kernel_vec_length_constant;
42 14076 }
43
44 4692 size_t kai_get_m_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(void) {
45 4692 return kai_mr * kai_get_kernel_vec_length_constant();
46 }
47
48 9384 size_t kai_get_n_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(void) {
49 9384 return kai_nr * kai_get_kernel_vec_length_constant();
50 }
51
52 2346 size_t kai_get_lhs_packed_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(
53 size_t m_idx, size_t k_chunk_count, size_t k_chunk_length) {
54 − KAI_ASSUME(m_idx % kai_get_m_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa() == 0);
55 2346 return m_idx * k_chunk_count * kai_roundup(k_chunk_length, kai_kr) * sizeof(uint16_t);
56 }
57
58 2346 static size_t kai_get_rhs_packed_stride_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(
59 size_t k_chunk_count, size_t k_chunk_length) {
60 4692 return kai_get_n_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa() *
61 2346 (sizeof(uint16_t) + k_chunk_count * kai_roundup(k_chunk_length, kai_kr) * sizeof(uint16_t));
62 }
63
64 2346 size_t kai_get_rhs_packed_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(
65 size_t n_idx, size_t k_chunk_count, size_t k_chunk_length) {
66 − KAI_ASSUME(n_idx % kai_get_n_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa() == 0);
67 2346 const size_t block_idx = n_idx / kai_get_n_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa();
68 7038 return block_idx *
69 2346 kai_get_rhs_packed_stride_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(
70 2346 k_chunk_count, k_chunk_length);
71 2346 }
72
73 2346 size_t kai_get_dst_offset_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(
74 size_t m_idx, size_t n_idx, size_t dst_stride_row) {
75 − KAI_ASSUME(m_idx % kai_get_m_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa() == 0);
76 − KAI_ASSUME(n_idx % kai_get_n_step_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa() == 0);
77
78 2346 return m_idx * dst_stride_row + n_idx * sizeof(uint16_t);
79 }
80
81 2346 size_t kai_get_dst_size_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(size_t m, size_t n) {
82 2346 return m * n * sizeof(uint16_t);
83 }
84
85 2347 void kai_run_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(
86 size_t m, size_t n, size_t k_chunk_count, size_t k_chunk_length, const void* lhs_packed, const void* rhs_packed,
87 void* dst, size_t dst_stride_row, float clamp_min, float clamp_max) {
88 2347 KernelArgs args;
89
90 2347 args.A = lhs_packed;
91 2347 args.B = rhs_packed;
92 2347 args.C = dst;
93 2347 args.ldcb = dst_stride_row;
94 2347 args.M = m;
95 2347 args.N = n;
96 2347 args.K = k_chunk_count * kai_roundup(k_chunk_length, kai_kr);
97 2347 args.min = kai_f16_from_float_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(clamp_min);
98 2347 args.max = kai_f16_from_float_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(clamp_max);
99 2347 args.accumulator_buffer = NULL;
100 2347 args.flags = 0;
101
102 2347 kai_commit_za();
103
104 2347 kai_kernel_imatmul_clamp_f16_f16p2vlx2_f16p2vlx2b_2vlx2vl_sme_mopa(&args);
105 2347 }
106
107 #endif // Architectural features check.
108