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
|
/*
* 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 <iomanip>
#include <iostream>
#include "fbgemm/Fbgemm.h"
namespace fbgemm {
template <typename T, typename accT>
PackAMatrix<T, accT>::PackAMatrix(
matrix_op_t trans,
int32_t nRow,
int32_t nCol,
const T* smat,
int32_t ld,
inpType* pmat,
int groups)
: PackMatrix<PackAMatrix<T, accT>, T, accT>(nRow, nCol, pmat, groups),
trans_(trans),
smat_(smat),
ld_(ld) {
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 && "unsupported architecure");
}
if (BaseType::numCols() % groups != 0) {
throw std::runtime_error(
"groups = " + std::to_string(groups) +
" does not divide numCols = " + std::to_string(BaseType::numCols()));
}
if (!pmat) {
BaseType::buf_ = (T*)fbgemmAlignedAlloc(
64, BaseType::brow_ * BaseType::bcol_ * sizeof(T));
}
}
template <typename T, typename accT>
void PackAMatrix<T, accT>::pack(const block_type_t& block) {
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_};
BaseType::packedBlock(block_p);
bool tr = (trans_ == matrix_op_t::Transpose);
T* out = BaseType::getBuf();
if (tr) {
// TODO: should print warning because this path is not optimized yet
for (int i = block.row_start; i < block.row_start + block.row_size; ++i) {
int buf_idx = i - block.row_start;
for (int j = block.col_start; j < block.col_start + block.col_size; ++j) {
T val = smat_[i + j * ld_];
out[buf_idx * BaseType::blockColSize() + (j - block.col_start)] = val;
}
// zero fill
// Please note that we zero fill, not zero_pt fill, because for
// requantization original, i.e., not padded, dimensions are used. If we
// were to use padded dimensions for requantization, we would zero_pt
// fill.
// For example, consider the following dot product:
// A = .3(5-15), .3(20-15) //.3 is scale and 15 is zero_pt
// B = .4(1+10), .4(4+10) // .4 is scale and -10 is zero_pt
//
// numElements(A) = 2 and numElements(B) = 2
//
// Dot product is (real): -3*4.4+1.5*5.6 = -4.8
// Dot product is (quantized): 5*1+20*4 = 85
//
// requantization: .3*.4(85 - (5+20)*(-10) - (1+4)*(15) +
// numElements(A)*(15)(-10)) = -4.8
//
// In the above adding one more element zero in the quantized domain,
// i.e., the quantized vectors become:
// A_q = 5, 20, 0
// B_q = 1, 4, 0
//
// and requantization with numElements(A) = 2 will produce the same
// answer (-4.8).
//
// Also in the above adding one more element zero_pt in the quantized
// domain, i.e., the quantized vectors become:
// A_q = 5, 20, 15
// B_q = 1, 4, -10
//
// and requantization with numElements(A) = 3 will produce the same
// answer (-4.8).
for (int j = block.col_size; j < block_p.col_size; ++j) {
out[buf_idx * BaseType::blockColSize() + j] = 0;
}
}
} else {
for (int i = block.row_start; i < block.row_start + block.row_size; ++i) {
int buf_idx = i - block.row_start;
memcpy(
out + buf_idx * BaseType::blockColSize(),
smat_ + i * ld_ + block.col_start,
block.col_size * sizeof(T));
// zero fill
for (int j = block.col_size; j < block_p.col_size; ++j) {
out[buf_idx * BaseType::blockColSize() + j] = 0;
}
}
}
}
template <typename T, typename accT>
int32_t PackAMatrix<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 PackAMatrix<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 class PackAMatrix<uint8_t, int32_t>;
template class PackAMatrix<uint8_t, int16_t>;
} // namespace fbgemm
|