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_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme.h" |
11 |
|
|
|
12 |
|
|
#include <stddef.h> |
13 |
|
|
#include <stdint.h> |
14 |
|
|
#include <string.h> |
15 |
|
|
|
16 |
|
|
#include "kai/kai_common.h" |
17 |
|
|
|
18 |
|
|
enum { |
19 |
|
|
NR = 2, |
20 |
|
|
KR = 2, |
21 |
|
|
MAX_N_STEP = NR * ((KAI_SME_VEC_LENGTH_MAX_BYTES / sizeof(uint16_t)) / KR), |
22 |
|
|
}; |
23 |
|
|
|
24 |
|
|
typedef struct { |
25 |
|
|
const void* bias_ptr; |
26 |
|
|
size_t width; |
27 |
|
|
size_t height; |
28 |
|
|
size_t k_chunk_count; |
29 |
|
|
size_t in_stride; |
30 |
|
|
size_t out_stride; |
31 |
|
|
const void* in; |
32 |
|
|
void* out; |
33 |
|
|
const void* pad_row; |
34 |
|
|
} KernelArgs; |
35 |
|
|
|
36 |
|
|
static const size_t kai_num_bytes_input = sizeof(uint16_t); |
37 |
|
|
static const size_t kai_num_bytes_output = sizeof(uint16_t); |
38 |
|
|
static const size_t kai_num_bytes_bias = sizeof(uint16_t); |
39 |
|
|
|
40 |
|
|
void kai_kernel_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(const KernelArgs* args_ptr); |
41 |
|
|
|
42 |
|
113520 |
size_t kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(void) { |
43 |
|
113520 |
return NR * kai_get_sme_vector_length_u16() / KR; |
44 |
|
|
} |
45 |
|
|
|
46 |
|
11352 |
size_t kai_get_rhs_offset_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(size_t n_idx) { |
47 |
|
− |
KAI_ASSUME(n_idx % kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme() == 0); |
48 |
|
|
|
49 |
|
11352 |
return n_idx * kai_num_bytes_input; |
50 |
|
|
} |
51 |
|
|
|
52 |
|
11352 |
size_t kai_get_bias_offset_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(size_t n_idx) { |
53 |
|
11352 |
return n_idx * kai_num_bytes_bias; |
54 |
|
|
} |
55 |
|
|
|
56 |
|
34056 |
static size_t kai_get_rhs_packed_stride_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme( |
57 |
|
|
size_t k_chunk_count, size_t k_chunk_length) { |
58 |
|
68112 |
return kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme() * |
59 |
|
34056 |
(kai_num_bytes_bias + k_chunk_count * kai_roundup(k_chunk_length, KR) * kai_num_bytes_output); |
60 |
|
|
} |
61 |
|
|
|
62 |
|
22704 |
size_t kai_get_rhs_packed_offset_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme( |
63 |
|
|
size_t n_idx, size_t k_chunk_count, size_t k_chunk_length) { |
64 |
|
− |
KAI_ASSUME(n_idx % kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme() == 0); |
65 |
|
|
|
66 |
|
22704 |
const size_t block_idx = n_idx / kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(); |
67 |
|
68112 |
return block_idx * |
68 |
|
22704 |
kai_get_rhs_packed_stride_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(k_chunk_count, k_chunk_length); |
69 |
|
22704 |
} |
70 |
|
|
|
71 |
|
11352 |
size_t kai_get_rhs_packed_size_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme( |
72 |
|
|
size_t n, size_t k_chunk_count, size_t k_chunk_length) { |
73 |
|
11352 |
const size_t n_nr_blocks = kai_roundup(n, kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme()); |
74 |
|
22704 |
return kai_get_rhs_packed_offset_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme( |
75 |
|
11352 |
n_nr_blocks, k_chunk_count, k_chunk_length); |
76 |
|
11352 |
} |
77 |
|
|
|
78 |
|
11352 |
void kai_run_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme( |
79 |
|
|
size_t n, size_t k_chunk_count, size_t k_chunk_length, size_t rhs_stride_row, const void* rhs, const void* bias, |
80 |
|
|
void* rhs_packed) { |
81 |
|
− |
KAI_ASSUME(rhs != NULL); |
82 |
|
− |
KAI_ASSUME(bias != NULL); |
83 |
|
− |
KAI_ASSUME(rhs_packed != NULL); |
84 |
|
|
|
85 |
|
− |
KAI_ASSERT(kai_get_n_step_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme() <= MAX_N_STEP); |
86 |
|
|
static const uint16_t pad_row[MAX_N_STEP] = {0}; |
87 |
|
|
|
88 |
|
11352 |
KernelArgs args; |
89 |
|
11352 |
args.bias_ptr = bias; |
90 |
|
11352 |
args.height = k_chunk_length; |
91 |
|
11352 |
args.width = n; |
92 |
|
11352 |
args.in = rhs; |
93 |
|
11352 |
args.out = rhs_packed; |
94 |
|
11352 |
args.k_chunk_count = k_chunk_count; |
95 |
|
11352 |
args.in_stride = rhs_stride_row; |
96 |
|
11352 |
args.out_stride = |
97 |
|
11352 |
kai_get_rhs_packed_stride_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(k_chunk_count, k_chunk_length); |
98 |
|
11352 |
args.pad_row = pad_row; |
99 |
|
|
|
100 |
|
11352 |
kai_kernel_rhs_imatmul_pack_kxn_x16p2vlx2b_x16_x16_sme(&args); |
101 |
|
11352 |
} |
102 |
|
|
|
103 |
|
|
#endif // Architectural features check. |
104 |
|
|
|