KleidiAI Coverage Report

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(_M_ARM64)
8			#error This file must be compiled for AArch64.
9			#else // Architectural features check.
10
11			#include "kai_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla.h"
12
13			#include <stddef.h>
14			#include <stdint.h>
15
16			#include "kai/kai_common.h"
17
18			typedef struct {
19			float maxval;
20			float minval;
21			unsigned int num_strings;
22			const unsigned int* string_lengths;
23			size_t N;
24			const void* B_ptr;
25			size_t output_offset;
26			size_t input_initial_col;
27			size_t input_offset;
28			void* output_ptr;
29			const void* bias;
30			} KernelArgs;
31
32			void kai_kernel_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(
33			const void* input_ptr, size_t m, KernelArgs* args_ptr, unsigned long flags);
34
35			static const size_t kai_mr = 6;
36			static const size_t kai_nr = 16;
37			static const size_t kai_kr = 1;
38			static const size_t kai_sr = 1;
39
40		50	size_t kai_get_m_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(void) {
41		50	return kai_mr;
42			}
43
44		50	size_t kai_get_n_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(void) {
45		50	return kai_nr;
46			}
47
48		17	size_t kai_get_nr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(void) {
49		17	return kai_nr;
50			}
51
52		17	size_t kai_get_kr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(void) {
53		17	return kai_kr;
54			}
55
56		17	size_t kai_get_sr_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(void) {
57		17	return kai_sr;
58			}
59
60		16	size_t kai_get_lhs_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(size_t m_idx, size_t stride) {
61		−	KAI_ASSUME(m_idx % kai_get_m_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla() == 0);
62
63		16	return m_idx * stride;
64			}
65
66		16	size_t kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(size_t n_idx, size_t k) {
67		−	KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla() == 0);
68
69		16	return n_idx / kai_nr * (kai_nr * sizeof(float) + kai_nr * k * sizeof(float));
70			}
71
72		16	size_t kai_get_dst_offset_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(size_t m_idx, size_t n_idx, size_t stride) {
73		−	KAI_ASSUME(m_idx % kai_get_m_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla() == 0);
74		−	KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla() == 0);
75
76		16	return m_idx * stride + n_idx * sizeof(float);
77			}
78
79		16	size_t kai_get_dst_size_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(size_t m, size_t n) {
80		16	return m * n * sizeof(float);
81			}
82
83		17	void kai_run_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(
84			size_t m, size_t n, size_t k, //
85			const void* lhs, size_t lhs_stride, //
86			const void* rhs_packed, //
87			void* dst, size_t dst_stride_row, size_t dst_stride_col, //
88			float clamp_min, float clamp_max) {
89		17	KAI_UNUSED(dst_stride_col);
90
91		17	KernelArgs ka;
92
93		17	unsigned long flags = 0;
94
95		17	unsigned int string_length = k;
96		17	ka.num_strings = 1;
97		17	ka.string_lengths = &string_length;
98		17	ka.N = n;
99		17	ka.B_ptr = rhs_packed;
100		17	ka.bias = NULL;
101
102			// Direct input.
103		17	const void* input_ptr = lhs;
104		17	ka.input_offset = lhs_stride / sizeof(float);
105		17	ka.input_initial_col = 0;
106
107			// Direct output.
108		17	ka.output_ptr = dst;
109		17	ka.output_offset = dst_stride_row / sizeof(float);
110
111			// Clamping output.
112		17	flags |= 0x2;
113		17	ka.maxval = clamp_max;
114		17	ka.minval = clamp_min;
115
116		17	kai_kernel_matmul_clamp_f32_f32_f32p16x1b_6x16_neon_mla(input_ptr, m, &ka, flags);
117		17	}
118
119			#endif // Architectural features check.
120