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 ( \
8			!defined(__aarch64__) && !defined(__ARM_FEATURE_FP16_SCALAR_ARITHMETIC) && \
9			!defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)) && \
10			!defined(_M_ARM64)
11			#error This file must be compiled for AArch64, FEAT_FP16.
12			#else // Architectural features check.
13
14			#include "kai_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla.h"
15
16			#include <stddef.h>
17			#include <stdint.h>
18
19			#include "kai/kai_common.h"
20
21			typedef struct {
22			uint16_t maxval;
23			uint16_t minval;
24			unsigned int num_strings;
25			const unsigned int* string_lengths;
26			size_t N;
27			const void* B_ptr;
28			size_t output_offset;
29			size_t input_initial_col;
30			size_t input_offset;
31			void* output_ptr;
32			const void* bias;
33			} KernelArgs;
34
35			void kai_kernel_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(
36			const void* input_ptr, size_t m, KernelArgs* args_ptr, unsigned long flags);
37			uint16_t kai_f16_from_float_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(float value);
38
39			static const size_t kai_mr = 6;
40			static const size_t kai_nr = 32;
41			static const size_t kai_kr = 1;
42			static const size_t kai_sr = 1;
43
44		50	size_t kai_get_m_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(void) {
45		50	return kai_mr;
46			}
47
48		50	size_t kai_get_n_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(void) {
49		50	return kai_nr;
50			}
51
52		17	size_t kai_get_nr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(void) {
53		17	return kai_nr;
54			}
55
56		17	size_t kai_get_kr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(void) {
57		17	return kai_kr;
58			}
59
60		17	size_t kai_get_sr_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(void) {
61		17	return kai_sr;
62			}
63
64		16	size_t kai_get_lhs_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(size_t m_idx, size_t stride) {
65		−	KAI_ASSUME(m_idx % kai_get_m_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla() == 0);
66
67		16	return m_idx * stride;
68			}
69
70		16	size_t kai_get_rhs_packed_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(size_t n_idx, size_t k) {
71		−	KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla() == 0);
72
73		16	return n_idx / kai_nr * (kai_nr * sizeof(uint16_t) + kai_nr * k * sizeof(uint16_t));
74			}
75
76		16	size_t kai_get_dst_offset_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(size_t m_idx, size_t n_idx, size_t stride) {
77		−	KAI_ASSUME(m_idx % kai_get_m_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla() == 0);
78		−	KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla() == 0);
79
80		16	return m_idx * stride + n_idx * sizeof(uint16_t);
81			}
82
83		16	size_t kai_get_dst_size_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(size_t m, size_t n) {
84		16	return m * n * sizeof(uint16_t);
85			}
86
87		17	void kai_run_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(
88			size_t m, size_t n, size_t k, //
89			const void* lhs, size_t lhs_stride, //
90			const void* rhs_packed, //
91			void* dst, size_t dst_stride_row, size_t dst_stride_col, //
92			float clamp_min, float clamp_max) {
93		17	KAI_UNUSED(dst_stride_col);
94
95		17	KernelArgs ka;
96
97		17	unsigned long flags = 0;
98
99		17	unsigned int string_length = k;
100		17	ka.num_strings = 1;
101		17	ka.string_lengths = &string_length;
102		17	ka.N = n;
103		17	ka.B_ptr = rhs_packed;
104		17	ka.bias = NULL;
105
106			// Direct input.
107		17	const void* input_ptr = lhs;
108		17	ka.input_offset = lhs_stride / sizeof(uint16_t);
109		17	ka.input_initial_col = 0;
110
111			// Direct output.
112		17	ka.output_ptr = dst;
113		17	ka.output_offset = dst_stride_row / sizeof(uint16_t);
114
115			// Clamping output.
116		17	flags |= 0x2;
117		17	ka.maxval = kai_f16_from_float_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(clamp_max);
118		17	ka.minval = kai_f16_from_float_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(clamp_min);
119
120		17	kai_kernel_matmul_clamp_f16_f16_f16p32x1b_6x32_neon_mla(input_ptr, m, &ka, flags);
121		17	}
122
123			#endif // Architectural features check.
124