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
|
/*
* 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 <cpuinfo.h>
#include <cassert>
#include <cmath>
#include <cstring>
#include <iomanip>
#include <iostream>
#include <stdexcept>
#include "fbgemm/Fbgemm.h"
namespace fbgemm2 {
template <typename T, typename accT>
PackAWithQuantRowOffset<T, accT>::PackAWithQuantRowOffset(
matrix_op_t trans,
int32_t nRow,
int32_t nCol,
const float* smat,
int32_t ld,
inpType* pmat,
float scale,
int32_t zero_pt,
int32_t groups,
int32_t* row_offset)
: PackMatrix<PackAWithQuantRowOffset<T, accT>, T, accT>(
nRow,
nCol,
pmat,
zero_pt),
trans_(trans),
smat_(smat),
ld_(ld),
scale_(scale),
G_(groups),
row_offset_(row_offset) {
assert(G_ == 1 && "Groups != 1 not supported yet");
rowOffsetAllocatedHere = false;
if (cpuinfo_has_x86_avx512f()) {
BaseType::brow_ = PackingTraits<T, accT, inst_set_t::avx512>::MCB;
BaseType::bcol_ = PackingTraits<T, accT, inst_set_t::avx512>::KCB;
row_interleave_B_ =
PackingTraits<T, accT, inst_set_t::avx512>::ROW_INTERLEAVE;
} else if (cpuinfo_has_x86_avx2()) {
BaseType::brow_ = PackingTraits<T, accT, inst_set_t::avx2>::MCB;
BaseType::bcol_ = PackingTraits<T, accT, inst_set_t::avx2>::KCB;
row_interleave_B_ =
PackingTraits<T, accT, inst_set_t::avx2>::ROW_INTERLEAVE;
} else {
// TODO: Have default slower path
assert(0 && "unknown architecure");
}
if (pmat) {
BaseType::buf_ = pmat;
} else {
BaseType::bufAllocatedHere_ = true;
BaseType::buf_ = (T*)fbgemmAlignedAlloc(
64, BaseType::brow_ * BaseType::bcol_ * sizeof(T));
}
if (!row_offset_) {
rowOffsetAllocatedHere = true;
row_offset_ = reinterpret_cast<int32_t*>(
fbgemmAlignedAlloc(64, BaseType::brow_ * sizeof(accT)));
}
}
template <typename T, typename accT>
void PackAWithQuantRowOffset<T, accT>::pack(const block_type_t& block) {
assert(block.row_start % BaseType::blockRowSize() == 0);
assert(block.col_start % BaseType::blockColSize() == 0);
assert(block.row_size <= BaseType::blockRowSize());
assert(block.col_size <= BaseType::blockColSize());
block_type_t block_p = {block.row_start,
block.row_size,
block.col_start,
(block.col_size + row_interleave_B_ - 1) /
row_interleave_B_ * row_interleave_B_};
assert(block_p.col_size <= BaseType::blockColSize());
BaseType::packedBlock(block_p);
T* out = BaseType::getBuf();
bool tr = (trans_ == matrix_op_t::Transpose);
// accumulate into row offset?
bool row_offset_acc = (block.col_start != 0);
int32_t* row_offset_buf = getRowOffsetBuffer();
float smat_transposed[block.row_size * block.col_size];
if (tr) {
transpose_simd(
block.col_size,
block.row_size,
smat_ + block.col_start * ld_ + block.row_start,
ld_,
smat_transposed,
block.col_size);
}
const float* smat_temp =
tr ? smat_transposed : smat_ + block.row_start * ld_ + block.col_start;
int32_t ld_temp = tr ? block.col_size : ld_;
#if defined(__AVX2__) && defined(__FMA__)
constexpr int VLEN = 8;
__m256 inverse_scale_v = _mm256_set1_ps(1.0f / scale_);
__m256i shuffle_mask_v = _mm256_set_epi8(
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0x0c,
0x08,
0x04,
0x00,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0xff,
0x0c,
0x08,
0x04,
0x00);
__m256i permute_mask_v =
_mm256_set_epi32(0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x04, 0x00);
#endif
for (int i = 0; i < block.row_size; ++i) {
int32_t row_sum = row_offset_acc ? row_offset_buf[i] : 0;
int j = 0;
#if defined(__AVX2__) && defined(__FMA__)
static_assert(
std::is_same<T, uint8_t>::value,
"PackAWithQuantRowOffset<T, accT>::pack only works for T == uint8_t");
for (; j < block.col_size / VLEN * VLEN; j += VLEN) {
__m256 val_v = _mm256_loadu_ps(smat_temp + i * ld_temp + j);
__m256 transformed_v = _mm256_fmadd_ps(
val_v, inverse_scale_v, _mm256_set1_ps(BaseType::zeroPoint()));
__m256 clipped_v = _mm256_max_ps(
_mm256_set1_ps(std::numeric_limits<uint8_t>::min()),
_mm256_min_ps(
transformed_v,
_mm256_set1_ps(std::numeric_limits<uint8_t>::max())));
__m256i res_v = _mm256_cvtps_epi32(clipped_v);
// An instruction sequence to save 8 32-bit integers as 8 8-bit integers
res_v = _mm256_shuffle_epi8(res_v, shuffle_mask_v);
res_v = _mm256_permutevar8x32_epi32(res_v, permute_mask_v);
_mm_storel_epi64(
reinterpret_cast<__m128i*>(out + i * BaseType::blockColSize() + j),
_mm256_castsi256_si128(res_v));
for (int j2 = j; j2 < j + VLEN; ++j2) {
row_sum += out[i * BaseType::blockColSize() + j2];
}
}
#endif
for (; j < block.col_size; ++j) {
float val = smat_temp[i * ld_temp + j];
float transformed = val / scale_ + BaseType::zeroPoint();
float clipped = std::min<float>(
std::max<float>(transformed, std::numeric_limits<uint8_t>::min()),
std::numeric_limits<uint8_t>::max());
T res = round(clipped);
row_sum += res;
out[i * BaseType::blockColSize() + j] = res;
}
// zero fill
// Please see the comment in PackAMatrix.cc on zero vs zero_pt fill.
for (; j < block_p.col_size; ++j) {
out[i * BaseType::blockColSize() + j] = 0;
}
row_offset_buf[i] = row_sum;
}
}
template <typename T, typename accT>
int32_t PackAWithQuantRowOffset<T, accT>::addr(int32_t r, int32_t c) const {
int32_t block_row_id = r / BaseType::blockRowSize();
int32_t brow_offset = (block_row_id * BaseType::blockCols()) *
(BaseType::blockRowSize() * BaseType::blockColSize());
int32_t block_col_id = c / BaseType::blockColSize();
int32_t bcol_offset =
block_col_id * BaseType::blockRowSize() * BaseType::blockColSize();
int32_t block_offset = brow_offset + bcol_offset;
int32_t inblock_offset =
(r % BaseType::blockRowSize()) * BaseType::blockColSize() +
(c % BaseType::blockColSize());
int32_t index = block_offset + inblock_offset;
return index;
}
template <typename T, typename accT>
void PackAWithQuantRowOffset<T, accT>::printPackedMatrix(std::string name) {
std::cout << name << ":"
<< "[" << BaseType::numPackedRows() << ", "
<< BaseType::numPackedCols() << "]" << std::endl;
T* out = BaseType::getBuf();
for (auto r = 0; r < BaseType::numPackedRows(); ++r) {
for (auto c = 0; c < BaseType::numPackedCols(); ++c) {
T val = out[addr(r, c)];
if (std::is_integral<T>::value) {
// cast to int64 because cout doesn't print int8_t type directly
std::cout << std::setw(5) << static_cast<int64_t>(val) << " ";
} else {
std::cout << std::setw(5) << val << " ";
}
}
std::cout << std::endl;
}
std::cout << std::endl;
}
template <typename T, typename accT>
int PackAWithQuantRowOffset<T, accT>::rowOffsetBufferSize() {
if (cpuinfo_initialize()) {
if (cpuinfo_has_x86_avx512f()) {
// TODO: avx512 path
// Currently use avx2 code
return PackingTraits<T, accT, inst_set_t::avx2>::MCB;
} else if (cpuinfo_has_x86_avx2()) {
return PackingTraits<T, accT, inst_set_t::avx2>::MCB;
} else {
assert(0 && "unsupported architecture");
return -1;
}
} else {
throw std::runtime_error("Failed to initialize cpuinfo!");
}
}
template class PackAWithQuantRowOffset<uint8_t, int32_t>;
} // namespace fbgemm2
|
