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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
|
// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// pack_avx.h: optimized AVX specializations of the templates in pack.h.
#ifndef GEMMLOWP_INTERNAL_PACK_AVX_H_
#define GEMMLOWP_INTERNAL_PACK_AVX_H_
#include <immintrin.h>
#include "pack.h"
namespace gemmlowp {
// TODO: Add DepthMajorUint8SideMap
typedef SideMap<const std::uint8_t, SideMapOrder::WidthMajor>
WidthMajorUint8SideMap;
template <int Cells>
using WidthMajorSideFormatNCells4x2 =
KernelSideFormat<CellFormat<8, 2, CellOrder::WidthMajor>, Cells>;
template <int Cells>
class PackingRegisterBlock<
WidthMajorUint8SideMap,
PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells>>>
: public PackingRegisterBlockBase<
WidthMajorUint8SideMap,
PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells>>> {
public:
typedef WidthMajorSideFormatNCells4x2<Cells> KernelSideFormat;
typedef typename KernelSideFormat::Cell CellFormat;
static const int kCells = KernelSideFormat::kCells;
static const int kCellWidth = CellFormat::kWidth;
static const int kKernelWidth = CellFormat::kWidth * kCells;
static const int kCellDepth = CellFormat::kDepth;
static const int kCellSize = CellFormat::kSize;
void Pack(PackedSideBlock<KernelSideFormat> *dst, int start_width) {
std::uint8_t *dst_ptr = dst->current_data();
const int width_stride = this->complete_src_.width_stride();
int depth_step = 16;
__m256i one = _mm256_set1_epi16(1);
for (int cell_start_depth = 0; cell_start_depth < kRegisterSize;
cell_start_depth += depth_step) {
for (int cell_start_width = 0; cell_start_width < kKernelWidth;
cell_start_width += kCellWidth) {
std::int32_t *cell_sums_of_each_slice_ptr =
dst->sums_of_each_slice() + start_width + cell_start_width;
const std::uint8_t *src_data =
this->complete_src_.data(cell_start_width, cell_start_depth);
__m128i xmm1 =
_mm_loadu_si128(reinterpret_cast<const __m128i *>(&src_data[0]));
__m128i xmm2 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&src_data[1 * width_stride]));
__m128i xmm3 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&src_data[2 * width_stride]));
__m128i xmm4 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&src_data[3 * width_stride]));
__m128i xmm5 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&src_data[4 * width_stride]));
__m128i xmm6 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&src_data[5 * width_stride]));
__m128i xmm7 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&src_data[6 * width_stride]));
__m128i xmm8 = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&src_data[7 * width_stride]));
__m256i ymm1 = _mm256_set_m128i(xmm5, xmm1);
__m256i ymm2 = _mm256_set_m128i(xmm6, xmm2);
__m256i ymm3 = _mm256_set_m128i(xmm7, xmm3);
__m256i ymm4 = _mm256_set_m128i(xmm8, xmm4);
__m256i ymm5 = _mm256_unpacklo_epi16(ymm1, ymm2);
__m256i ymm6 = _mm256_unpacklo_epi16(ymm3, ymm4);
__m256i ymm9 = _mm256_unpackhi_epi16(ymm1, ymm2);
__m256i ymm10 = _mm256_unpackhi_epi16(ymm3, ymm4);
__m256i ymm7 = _mm256_unpacklo_epi32(ymm5, ymm6);
__m256i ymm8 = _mm256_unpackhi_epi32(ymm5, ymm6);
__m256i ymm13 = _mm256_unpacklo_epi32(ymm9, ymm10);
__m256i ymm14 = _mm256_unpackhi_epi32(ymm9, ymm10);
__m256i ymm11 = _mm256_permute4x64_epi64(ymm7, 0xd8);
__m256i ymm12 = _mm256_permute4x64_epi64(ymm8, 0xd8);
__m256i ymm15 = _mm256_permute4x64_epi64(ymm13, 0xd8);
__m256i ymm16 = _mm256_permute4x64_epi64(ymm14, 0xd8);
__m128i xmm9 = _mm256_castsi256_si128(ymm11);
__m128i xmm10 = _mm256_castsi256_si128(ymm12);
__m128i xmm11 = _mm256_extracti128_si256(ymm11, 1);
__m128i xmm12 = _mm256_extracti128_si256(ymm12, 1);
xmm1 = _mm256_castsi256_si128(ymm15);
xmm2 = _mm256_castsi256_si128(ymm16);
xmm3 = _mm256_extracti128_si256(ymm15, 1);
xmm4 = _mm256_extracti128_si256(ymm16, 1);
_mm_storeu_si128(reinterpret_cast<__m128i *>(&dst_ptr[0]), xmm9);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&dst_ptr[kCellSize * kCells]), xmm11);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&dst_ptr[2 * kCellSize * kCells]),
xmm10);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&dst_ptr[3 * kCellSize * kCells]),
xmm12);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&dst_ptr[4 * kCellSize * kCells]),
xmm1);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&dst_ptr[5 * kCellSize * kCells]),
xmm3);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&dst_ptr[6 * kCellSize * kCells]),
xmm2);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&dst_ptr[7 * kCellSize * kCells]),
xmm4);
ymm6 = _mm256_cvtepu8_epi16(xmm9);
ymm7 = _mm256_madd_epi16(ymm6, one);
__m256i sums_of_each_slice_xmm = _mm256_loadu_si256(
reinterpret_cast<const __m256i *>(&cell_sums_of_each_slice_ptr[0]));
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
ymm6 = _mm256_cvtepu8_epi16(xmm11);
ymm7 = _mm256_madd_epi16(ymm6, one);
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
ymm6 = _mm256_cvtepu8_epi16(xmm10);
ymm7 = _mm256_madd_epi16(ymm6, one);
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
ymm6 = _mm256_cvtepu8_epi16(xmm12);
ymm7 = _mm256_madd_epi16(ymm6, one);
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
ymm6 = _mm256_cvtepu8_epi16(xmm1);
ymm7 = _mm256_madd_epi16(ymm6, one);
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
ymm6 = _mm256_cvtepu8_epi16(xmm3);
ymm7 = _mm256_madd_epi16(ymm6, one);
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
ymm6 = _mm256_cvtepu8_epi16(xmm2);
ymm7 = _mm256_madd_epi16(ymm6, one);
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
ymm6 = _mm256_cvtepu8_epi16(xmm4);
ymm7 = _mm256_madd_epi16(ymm6, one);
sums_of_each_slice_xmm = _mm256_add_epi32(sums_of_each_slice_xmm, ymm7);
_mm256_storeu_si256(
reinterpret_cast<__m256i *>(&cell_sums_of_each_slice_ptr[0]),
sums_of_each_slice_xmm);
dst_ptr += kCellSize;
}
dst_ptr += 7 * kCellSize * kCells;
}
dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
}
};
// Pack format for 4x2 rhs format
template <int Cells>
using RhsWidthMajorSideFormatNCells4x2 =
KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, Cells>;
template <int Cells>
class PackingRegisterBlock<
WidthMajorUint8SideMap,
PackedSideBlock<RhsWidthMajorSideFormatNCells4x2<Cells>>>
: public PackingRegisterBlockBase<
WidthMajorUint8SideMap,
PackedSideBlock<RhsWidthMajorSideFormatNCells4x2<Cells>>> {
public:
typedef RhsWidthMajorSideFormatNCells4x2<Cells> KernelSideFormat;
typedef typename KernelSideFormat::Cell CellFormat;
static const int kCells = KernelSideFormat::kCells;
static const int kCellWidth = CellFormat::kWidth;
static const int kKernelWidth = CellFormat::kWidth * kCells;
static const int kCellDepth = CellFormat::kDepth;
static const int kCellSize = CellFormat::kSize;
void Pack(PackedSideBlock<KernelSideFormat> *dst, int start_width) {
std::uint8_t *dst_ptr = dst->current_data();
const int width_stride = this->complete_src_.width_stride();
int depth_step = 8;
__m128i one = _mm_set1_epi16(1);
for (int cell_start_depth = 0; cell_start_depth < kRegisterSize;
cell_start_depth += depth_step) {
for (int cell_start_width = 0; cell_start_width < kKernelWidth;
cell_start_width += kCellWidth) {
std::int32_t *cell_sums_of_each_slice_ptr =
dst->sums_of_each_slice() + start_width + cell_start_width;
const std::uint8_t *src_data =
this->complete_src_.data(cell_start_width, cell_start_depth);
__m128i xmm1 =
_mm_loadl_epi64(reinterpret_cast<const __m128i *>(&src_data[0]));
__m128i xmm2 = _mm_loadl_epi64(
reinterpret_cast<const __m128i *>(&src_data[1 * width_stride]));
__m128i xmm3 = _mm_loadl_epi64(
reinterpret_cast<const __m128i *>(&src_data[2 * width_stride]));
__m128i xmm4 = _mm_loadl_epi64(
reinterpret_cast<const __m128i *>(&src_data[3 * width_stride]));
__m128i xmm5 = _mm_unpacklo_epi16(xmm1, xmm2);
__m128i xmm8 = _mm_shuffle_epi32(xmm5, 0x31);
__m128i xmm6 = _mm_unpacklo_epi16(xmm3, xmm4);
__m128i xmm7 = _mm_shuffle_epi32(xmm6, 0x80);
__m128i xmm9 = _mm_blend_epi16(xmm5, xmm7, 0xcc);
__m128i xmm10 = _mm_blend_epi16(xmm8, xmm6, 0xcc);
_mm_storel_epi64(reinterpret_cast<__m128i *>(&dst_ptr[0]), xmm9);
_mm_storel_epi64(
reinterpret_cast<__m128i *>(&dst_ptr[kCellSize * kCells]), xmm10);
__m128i xmm11 = _mm_shuffle_epi32(xmm9, 0xee);
__m128i xmm12 = _mm_shuffle_epi32(xmm10, 0xee);
_mm_storel_epi64(
reinterpret_cast<__m128i *>(&dst_ptr[2 * kCellSize * kCells]),
xmm11);
_mm_storel_epi64(
reinterpret_cast<__m128i *>(&dst_ptr[3 * kCellSize * kCells]),
xmm12);
xmm1 = _mm_cvtepu8_epi16(xmm9);
xmm2 = _mm_madd_epi16(xmm1, one);
__m128i sums_of_each_slice_xmm = _mm_loadu_si128(
reinterpret_cast<const __m128i *>(&cell_sums_of_each_slice_ptr[0]));
sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
xmm1 = _mm_cvtepu8_epi16(xmm10);
xmm2 = _mm_madd_epi16(xmm1, one);
sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
xmm1 = _mm_cvtepu8_epi16(xmm11);
xmm2 = _mm_madd_epi16(xmm1, one);
sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
xmm1 = _mm_cvtepu8_epi16(xmm12);
xmm2 = _mm_madd_epi16(xmm1, one);
sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
_mm_storeu_si128(
reinterpret_cast<__m128i *>(&cell_sums_of_each_slice_ptr[0]),
sums_of_each_slice_xmm);
dst_ptr += kCellSize;
}
dst_ptr += 3 * kCellSize * kCells;
}
dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
}
};
} // namespace gemmlowp
#endif // GEMMLOWP_INTERNAL_PACK_AVX_H_
|
