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_rhs_pack_kxn_x16p32x1b_x16_x16_neon.h"
12
13			#include <stddef.h>
14			#include <stdint.h>
15
16			#include "kai/kai_common.h"
17
18			enum {
19			NR = 32,
20			KR = 1,
21			};
22
23			typedef struct {
24			const void* bias_ptr;
25			size_t width;
26			size_t height;
27			size_t in_stride;
28			size_t out_stride;
29			const void* in;
30			void* out;
31			} KernelArgs;
32
33			static const size_t kai_num_bytes_input = sizeof(uint16_t);
34			static const size_t kai_num_bytes_output = sizeof(uint16_t);
35			static const size_t kai_num_bytes_bias = sizeof(uint16_t);
36
37			void kai_kernel_rhs_pack_kxn_x16p32x1b_x16_x16_neon(const KernelArgs* args_ptr);
38
39		374	size_t kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon(void) {
40		374	return NR;
41			}
42
43		34	size_t kai_get_rhs_offset_rhs_pack_kxn_x16p32x1b_x16_x16_neon(size_t n_idx) {
44		−	KAI_ASSUME(n_idx % kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon() == 0);
45
46		34	return n_idx * kai_num_bytes_input;
47			}
48
49		34	size_t kai_get_bias_offset_rhs_pack_kxn_x16p32x1b_x16_x16_neon(size_t n_idx) {
50		34	return n_idx * kai_num_bytes_bias;
51			}
52
53		102	size_t kai_get_rhs_packed_stride_rhs_pack_kxn_x16p32x1b_x16_x16_neon(size_t k) {
54		204	return kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon() *
55		102	(kai_num_bytes_bias + kai_roundup(k, KR) * kai_num_bytes_output);
56			}
57
58		68	size_t kai_get_rhs_packed_offset_rhs_pack_kxn_x16p32x1b_x16_x16_neon(size_t n_idx, size_t k) {
59		−	KAI_ASSUME(n_idx % kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon() == 0);
60
61		68	const size_t block_idx = n_idx / kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon();
62		136	return block_idx * kai_get_rhs_packed_stride_rhs_pack_kxn_x16p32x1b_x16_x16_neon(k);
63		68	}
64
65		34	size_t kai_get_rhs_packed_size_rhs_pack_kxn_x16p32x1b_x16_x16_neon(size_t n, size_t k) {
66		34	const size_t n_nr_blocks = kai_roundup(n, kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon());
67		68	return kai_get_rhs_packed_offset_rhs_pack_kxn_x16p32x1b_x16_x16_neon(n_nr_blocks, k);
68		34	}
69
70		34	void kai_run_rhs_pack_kxn_x16p32x1b_x16_x16_neon(
71			size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride_row, const void* rhs,
72			const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params) {
73		−	KAI_ASSUME(num_groups == 1);
74		−	KAI_ASSUME(nr == kai_get_n_step_rhs_pack_kxn_x16p32x1b_x16_x16_neon());
75		−	KAI_ASSUME(kr == KR);
76		−	KAI_ASSUME(sr == 1);
77		−	KAI_ASSUME(rhs != NULL);
78		−	KAI_ASSUME(bias != NULL);
79		−	KAI_ASSUME(scale == NULL);
80		−	KAI_ASSUME(rhs_packed != NULL);
81		−	KAI_ASSUME(extra_bytes == 0);
82		−	KAI_ASSUME(params == NULL);
83
84		34	KernelArgs args;
85		34	args.bias_ptr = bias;
86		34	args.height = k;
87		34	args.width = n;
88		34	args.in = rhs;
89		34	args.out = rhs_packed;
90		34	args.in_stride = rhs_stride_row;
91		34	args.out_stride = kai_get_rhs_packed_stride_rhs_pack_kxn_x16p32x1b_x16_x16_neon(args.height);
92
93		34	kai_kernel_rhs_pack_kxn_x16p32x1b_x16_x16_neon(&args);
94		34	}
95
96			#endif // Architectural features check.
97