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
|
/*
* 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 <random>
#ifdef _OPENMP
#include <omp.h>
#endif
#include <gtest/gtest.h>
#include "TestUtils.h"
#include "bench/BenchUtils.h"
#include "fbgemm/FbgemmFP16.h"
#include "src/RefImplementations.h"
#ifdef USE_IACA
#include "iacaMarks.h"
#endif
using namespace std;
using namespace fbgemm;
namespace {
// The template parameter is transpose of A and B
class FBGemmFP16Test
: public testing::TestWithParam<pair<matrix_op_t, matrix_op_t>> {
protected:
vector<vector<int>> GenShapes() const {
vector<vector<int>> shapes;
random_device r;
default_random_engine generator(r());
uniform_int_distribution<int> dm(1, 256);
uniform_int_distribution<int> dnk(1, 1024);
for (int i = 0; i < 10; i++) {
int m = dm(generator);
int n = dnk(generator);
int k = dnk(generator);
shapes.push_back({m, n, k});
if (m > 10) {
shapes.push_back({(m / 10) * 10, n, k});
}
}
return shapes;
}
};
}; // namespace
INSTANTIATE_TEST_CASE_P(
InstantiationName,
FBGemmFP16Test,
::testing::Values(
pair<matrix_op_t, matrix_op_t>(
matrix_op_t::NoTranspose, matrix_op_t::NoTranspose),
pair<matrix_op_t, matrix_op_t>(
matrix_op_t::NoTranspose, matrix_op_t::Transpose)/*,
pair<matrix_op_t, matrix_op_t>(
matrix_op_t::Transpose, matrix_op_t::NoTranspose),
pair<matrix_op_t, matrix_op_t>(
matrix_op_t::Transpose, matrix_op_t::Transpose)*/));
TEST_P(FBGemmFP16Test, Test) {
auto shapes = GenShapes();
float alpha = 1.f, beta = 0.f;
matrix_op_t atrans, btrans;
tie(atrans, btrans) = GetParam();
for (auto s : shapes) {
int m = s[0];
int n = s[1];
int k = s[2];
cerr << "m = " << m << " n = " << n << " k = " << k;
if (atrans == matrix_op_t::Transpose) {
cerr << " A_transposed";
}
if (btrans == matrix_op_t::Transpose) {
cerr << " B_transposed";
}
cerr << endl;
// initialize with small numbers
aligned_vector<int> Aint(m * k);
aligned_vector<int> Bint(k * n);
randFill(Aint, 0, 4);
randFill(Bint, 0, 4);
aligned_vector<float> A(Aint.begin(), Aint.end());
aligned_vector<float> B(Bint.begin(), Bint.end());
aligned_vector<float> C(m * n, NAN);
aligned_vector<float> A_ref(A), B_ref(B), C_ref(C);
if (atrans == matrix_op_t::Transpose) {
transpose_matrix(A_ref.data(), k, m);
}
if (btrans == matrix_op_t::Transpose) {
transpose_matrix(B_ref.data(), n, k);
}
// Gold via reference sgemm
matmul_fp_ref(m, n, k, k, n, n, A_ref.data(), B_ref.data(), C_ref.data());
// fbgemm fp16
PackedGemmMatrixFP16 Bp(btrans, k, n, alpha, B.data());
#ifdef _OPENMP
#pragma omp parallel
#endif
{
int num_threads = fbgemm_get_num_threads();
int tid = fbgemm_get_thread_num();
cblas_gemm_compute(
atrans, m, A.data(), Bp, beta, C.data(), tid, num_threads);
}
// correctness check
for (int i = 0; i < m; ++i) {
for (int j = 0; j < n; ++j) {
float expected = C_ref[i * n + j];
float actual = C[i * n + j];
EXPECT_EQ(expected, actual)
<< "GEMM results differ at (" << i << ", " << j << "). ref "
<< expected << " FBGemm " << actual;
}
}
}
}
TEST_P(FBGemmFP16Test, Unpack) {
auto shapes = GenShapes();
float alpha = 1.f, beta = 0.f;
matrix_op_t atrans, btrans;
tie(atrans, btrans) = GetParam();
for (auto s : shapes) {
int m = s[0];
int n = s[1];
int k = s[2];
cerr << "m = " << m << " n = " << n << " k = " << k;
if (atrans == matrix_op_t::Transpose) {
cerr << " A_transposed";
}
if (btrans == matrix_op_t::Transpose) {
cerr << " B_transposed";
}
cerr << endl;
// initialize with small numbers
aligned_vector<int> Aint(m * k);
aligned_vector<int> Bint(k * n);
randFill(Aint, 0, 4);
randFill(Bint, 0, 4);
aligned_vector<float> A(Aint.begin(), Aint.end());
aligned_vector<float> B(Bint.begin(), Bint.end());
aligned_vector<float> C(m * n, NAN);
aligned_vector<float> A_ref(A), B_ref(B), C_ref(C);
if (atrans == matrix_op_t::Transpose) {
transpose_matrix(A_ref.data(), k, m);
}
if (btrans == matrix_op_t::Transpose) {
transpose_matrix(B_ref.data(), n, k);
}
// Gold via reference sgemm
matmul_fp_ref(m, n, k, k, n, n, A_ref.data(), B_ref.data(), C_ref.data());
// fbgemm fp16
PackedGemmMatrixFP16 Bp(btrans, k, n, alpha, B.data());
EXPECT_TRUE(Bp.packed());
// Test unpack
aligned_vector<float16> tmp(Bp.matSize());
memcpy(tmp.data(), Bp.pmat(), Bp.matSize() * sizeof(float16));
Bp.unpackFromSrc(btrans, tmp.data());
EXPECT_FALSE(Bp.packed());
memcpy(tmp.data(), Bp.pmat(), Bp.matSize() * sizeof(float16));
for (int i = 0; i < k; ++i) {
for (int j = 0; j < n; ++j) {
EXPECT_EQ(B[i * n + j], cpu_half2float(tmp[i * n + j]));
}
}
// Pack it back
Bp.packFromSrc(btrans, tmp.data());
EXPECT_TRUE(Bp.packed());
#ifdef _OPENMP
#pragma omp parallel
#endif
{
int num_threads = fbgemm_get_num_threads();
int tid = fbgemm_get_thread_num();
cblas_gemm_compute(
atrans, m, A.data(), Bp, beta, C.data(), tid, num_threads);
}
// correctness check
for (int i = 0; i < m; ++i) {
for (int j = 0; j < n; ++j) {
float expected = C_ref[i * n + j];
float actual = C[i * n + j];
EXPECT_EQ(expected, actual)
<< "GEMM results differ at (" << i << ", " << j << "). ref "
<< expected << " FBGemm " << actual;
}
}
}
}
|
