1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
|
/*
* Copyright (c) Facebook, Inc. and its affiliates.
* All rights reserved.
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#include "fbgemm/FbgemmI8DepthwiseAvx2.h"
#include "fbgemm/Utils.h"
#include "fbgemm/Fbgemm.h"
#include <immintrin.h>
using namespace std;
namespace fbgemm {
// clang-format off
static int masks[8][8] = {
// NOTE: clang-format wants to use a different formatting but the current
// formatting should be easier to read.
{ 0, 0, 0, 0, 0, 0, 0, 0, },
{ -1, 0, 0, 0, 0, 0, 0, 0, },
{ -1, -1, 0, 0, 0, 0, 0, 0, },
{ -1, -1, -1, 0, 0, 0, 0, 0, },
{ -1, -1, -1, -1, 0, 0, 0, 0, },
{ -1, -1, -1, -1, -1, 0, 0, 0, },
{ -1, -1, -1, -1, -1, -1, 0, 0, },
{ -1, -1, -1, -1, -1, -1, -1, 0, },
};
// clang-format on
PackedDepthWiseConvMatrix::PackedDepthWiseConvMatrix(
int K,
int kernel_prod,
const int8_t* smat)
: K_(K), kernel_prod_(kernel_prod) {
// Transpose the input matrix to make packing faster.
int8_t* smat_transposed
= static_cast<int8_t*>(ALIGNED_MALLOC(K * kernel_prod * sizeof(int8_t), 64));
for (int i = 0; i < kernel_prod; ++i) {
for (int j = 0; j < K; ++j) {
smat_transposed[i * K + j] = smat[i + j * kernel_prod];
}
}
// Allocate packed arrays
int kernel_prod_aligned = (kernel_prod + 1) / 2 * 2;
pmat_ = static_cast<int8_t *>(fbgemmAlignedAlloc(64, ((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t)));
//posix_memalign(
// (void**)&pmat_,
// 64,
// ((K + 31) / 32) * kernel_prod_aligned * 32 * sizeof(int8_t));
// Pack input matrix
// The layout is optimized to use vpmaddubsw efficiently (see
// madd_epi16x4_packed function).
// For a group of 32 channels, we have 10 32B SIMD registers.
// Denote ith channel jth filter as (i, j)
// 0th SIMD register:
// (0, 0), (0, 1), (0, 2), (0, 3), ..., (3, 0), (3, 1), (3, 2), (3, 3)
// (16, 0), (16, 1), (16, 2), (16, 3), ..., (19, 0), (19, 1), (19, 2), (19, 3)
// 1st SIMD register:
// (4, 0), (4, 1), (4, 2), (4, 3), ..., (7, 0), (7, 1), (7, 2), (7, 3)
// (20, 0), (20, 1), (20, 2), (20, 3), ..., (23, 0), (23, 1), (23, 2), (23, 3)
// 2nd SIMD register:
// (8, 0), (8, 1), (8, 2), (8, 3), ..., (11, 0), (11, 1), (11, 2), (11, 3)
// (24, 0), (24, 1), (24, 2), (24, 3), ..., (27, 0), (27, 1), (27, 2), (27, 3)
// 3rd SIMD register:
// (12, 0), (12, 1), (12, 2), (12, 3), ..., (15, 0), (15, 1), (15, 2), (15, 3)
// (28, 0), (28, 1), (28, 2), (28, 3), ..., (31, 0), (31, 1), (31, 2), (31, 3)
// 4-7th SIMD register: same as the previous 4 registers but for 4-7th filter
// coefficients
// ...
//
// REMAINDER
// If kernel_prod % 4 == 1 for example when kernel_prod == 9
// 8th SIMD register:
// (0, 8), zero, ..., (7, 8), zero
// (16, 8), zero, ..., (23, 8), zero
// 9th SIMD register:
// (8, 8), zero, ..., (15, 8), zero
// (24, 8), zero, ..., (31, 8), zero
// We use madd_epi16_packed for this case
//
// If kernel_prod % 4 == 2 for example when kernel_prod == 10
// 8th SIMD register:
// (0, 8), (0, 9), ..., (7, 8), (7, 9)
// (16, 8), (16, 9), ..., (23, 8), (23, 9)
// 9th SIMD register:
// (8, 8), (8, 9), ..., (15, 8), (15, 9)
// (24, 8), (24, 9), ..., (31, 8), (31, 9)
//
// If kernel_prod % 4 == 3 for example when kernel_prod == 11
// 8th SIMD register:
// (0, 8), (0, 9), (0, 10), zero, ..., (3, 8), (3, 9), (3, 10), zero
// (16, 8), (16, 9), (16, 10), zero, ..., (19, 8), (19, 9), (19, 10), zero
// 9th SIMD register:
// (4, 8), (4, 9), (4, 10), zero, ..., (7, 8), (7, 9), (7, 10), zero
// (20, 8), (20, 9), (20, 10), zero, ..., (23, 8), (23, 9), (23, 10), zero
// 10th SIMD register:
// (8, 8), (8, 9), (8, 10), zero, ..., (11, 8), (11, 9), (11, 10), zero
// (24, 8), (24, 9), (24, 10), zero, ..., (27, 8), (27, 9), (27, 10), zero
// 11th SIMD register:
// (12, 8), (12, 9), (12, 10), zero, ..., (15, 8), (15, 9), (15, 10), zero
// (28, 8), (28, 9), (28, 10), zero, ..., (31, 8), (31, 9), (31, 10), zero
for (int k1 = 0; k1 < K; k1 += 32) {
__m256i* b_v = new __m256i[kernel_prod];
int remainder = K - k1;
if (remainder < 32) {
__m256i mask_v = _mm256_loadu_si256(
reinterpret_cast<const __m256i*>(masks[remainder / 4]));
for (int i = 0; i < kernel_prod; ++i) {
b_v[i] = _mm256_maskload_epi32(
reinterpret_cast<const int*>(smat_transposed + i * K + k1), mask_v);
}
} else {
for (int i = 0; i < kernel_prod; ++i) {
b_v[i] = _mm256_lddqu_si256(
reinterpret_cast<const __m256i*>(smat_transposed + i * K + k1));
}
}
// Interleave 2 SIMD registers
__m256i* b_interleaved_epi16 = new __m256i[kernel_prod_aligned];
__m256i zero_v = _mm256_setzero_si256();
for (int i = 0; i < kernel_prod_aligned / 2; ++i) {
if (2 * i + 1 >= kernel_prod) {
b_interleaved_epi16[2 * i] = _mm256_unpacklo_epi8(b_v[2 * i], zero_v);
b_interleaved_epi16[2 * i + 1] =
_mm256_unpackhi_epi8(b_v[2 * i], zero_v);
} else {
b_interleaved_epi16[2 * i] =
_mm256_unpacklo_epi8(b_v[2 * i], b_v[2 * i + 1]);
b_interleaved_epi16[2 * i + 1] =
_mm256_unpackhi_epi8(b_v[2 * i], b_v[2 * i + 1]);
}
}
// Interleave 4 SIMD registers
__m256i* b_interleaved_epi32 = new __m256i[kernel_prod_aligned];
for (int i = 0; i < kernel_prod_aligned / 4; ++i) {
b_interleaved_epi32[4 * i] = _mm256_unpacklo_epi16(
b_interleaved_epi16[4 * i], b_interleaved_epi16[4 * i + 2]);
b_interleaved_epi32[4 * i + 1] = _mm256_unpackhi_epi16(
b_interleaved_epi16[4 * i], b_interleaved_epi16[4 * i + 2]);
b_interleaved_epi32[4 * i + 2] = _mm256_unpacklo_epi16(
b_interleaved_epi16[4 * i + 1], b_interleaved_epi16[4 * i + 3]);
b_interleaved_epi32[4 * i + 3] = _mm256_unpackhi_epi16(
b_interleaved_epi16[4 * i + 1], b_interleaved_epi16[4 * i + 3]);
}
for (int i = kernel_prod_aligned / 4 * 4; i < kernel_prod_aligned; ++i) {
b_interleaved_epi32[i] = b_interleaved_epi16[i];
}
for (int i = 0; i < kernel_prod_aligned; ++i) {
_mm256_storeu_si256(
reinterpret_cast<__m256i*>(
&pmat_[((k1 / 32) * kernel_prod_aligned + i) * 32]),
b_interleaved_epi32[i]);
}
delete[] b_v;
delete[] b_interleaved_epi16;
delete[] b_interleaved_epi32;
}
FREE(smat_transposed);
}
int PackedDepthWiseConvMatrix::addr(int r, int c) {
int kernel_prod_aligned = (kernel_prod_ + 1) / 2 * 2;
if (c >= kernel_prod_ / 4 * 4 &&
(kernel_prod_ % 4 == 1 || kernel_prod_ % 4 == 2)) {
int kBlock = r / 32;
int reg_idx = (r % 16) / 8 + c / 4 * 4;
int blk_idx = kBlock * kernel_prod_aligned + reg_idx;
int r_ = r % 8;
int c_ = c % 4;
int in_blk_idx = (r % 32) / 16 * 16 + 2 * r_ + c_;
return blk_idx * 32 + in_blk_idx;
} else {
int kBlock = r / 32;
int reg_idx = (r % 16) / 4 + c / 4 * 4;
int blk_idx = kBlock * kernel_prod_aligned + reg_idx;
int r_ = r % 4;
int c_ = c % 4;
int in_blk_idx = (r % 32) / 16 * 16 + 4 * r_ + c_;
return blk_idx * 32 + in_blk_idx;
}
}
void PackedDepthWiseConvMatrix::unpack(int8_t* unpacked_data) {
for (int r = 0; r < K_; ++r) {
for (int c = 0; c < kernel_prod_; ++c) {
unpacked_data[r * kernel_prod_ + c] = pmat_[addr(r, c)];
}
}
}
PackedDepthWiseConvMatrix::~PackedDepthWiseConvMatrix() {
free(pmat_);
}
} // namespace fbgemm
|
