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
|
#include <array>
#include <bitset>
#include <iomanip>
#include <limits>
#include <random>
#include <gtest/gtest.h>
#include "caffe2/core/logging.h"
#include "fully_connected_fake_lowp_op.h"
constexpr size_t sz = 10000;
using array = std::array<float, sz>;
constexpr size_t sz_test = 10;
using testarray = std::array<float, sz_test>;
bool isclose(float x, float y) {
if (x == y) {
return true;
}
LOG(INFO) << "Error range: " << fabs((x - y) / x);
return fabs(x - y) < 1e-2 * fabs(x);
}
bool isrelclose(float x, float y) {
if (x == y) {
return true;
}
float relerr = float(1.0) / float(1 << 7);
LOG(INFO) << "Relative error range: " << fabs((x - y) / x) << " " << relerr;
return fabs((x - y) / x) < relerr;
}
template <std::size_t N>
double mse(std::array<float, N>& a1, std::array<float, N>& a2) {
double total = 0.0;
for (auto i = 0; i < N; i++) {
auto diff = a1[i] - a2[i];
total += diff * diff;
}
return sqrt(total) / N;
}
// minimum representable value of (10 bit mantissa) fp16=6e-5
// minimum representable value of (7 bit mantissa) bfp16=5e-4
static void test_case(const float v, const float v_fp16, const float v_bfp16) {
std::array<float, 1> input, output1, output2, output3, output4, output5;
input[0] = v;
caffe2::fp32_to_fp16(input.data(), 1, output1.data());
caffe2::fp32_to_bfp16(input.data(), 1, output2.data());
caffe2::fp32_to_bfp24(input.data(), 1, output3.data());
caffe2::fp32_to_bfp14(input.data(), 1, output4.data());
caffe2::fp32_to_bfp16_round(input.data(), 1, output5.data());
if (!std::isinf(mse<1>(input, output1))) {
// If fp16 doesn't overshoot, the acurracy should be always better
// than the one where we do truncation
EXPECT_TRUE(mse<1>(input, output2) >= mse<1>(input, output1));
}
LOG(INFO) << std::hex << std::showbase << *(int*)(&input[0]) << " "
<< std::setprecision(20) << input[0];
LOG(INFO) << std::hex << std::showbase << *(int*)(&output1[0]) << " "
<< std::setprecision(20) << output1[0];
LOG(INFO) << std::hex << std::showbase << *(int*)(&output2[0]) << " "
<< std::setprecision(20) << output2[0];
LOG(INFO) << std::hex << std::showbase << *(int*)(&output3[0]) << " "
<< std::setprecision(20) << output3[0];
LOG(INFO) << std::hex << std::showbase << *(int*)(&output4[0]) << " "
<< std::setprecision(20) << output4[0];
LOG(INFO) << std::hex << std::showbase << *(int*)(&output5[0]) << " "
<< std::setprecision(20) << output5[0];
EXPECT_TRUE(isclose(output1[0], v_fp16));
EXPECT_TRUE(isclose(output2[0], v_bfp16));
EXPECT_TRUE(isclose(output5[0], output2[0]));
EXPECT_TRUE(isrelclose(output5[0], output2[0]));
}
static void
test_vector_case(const float v, const float v_fp16, const float v_bfp16) {
testarray vv, vfp16, vbfp16, vbfp16s, vbfp24, vbfp14, vbfp16r;
int i;
for (i = 0; i < sz_test; i++) {
vv[i] = v;
}
caffe2::fp32_to_fp16(vv.data(), sz_test, vfp16.data());
caffe2::fp32_to_bfp16(vv.data(), sz_test, vbfp16.data());
caffe2::fp32_to_bfp16_scalar(vv.data(), sz_test, vbfp16s.data());
caffe2::fp32_to_bfp24(vv.data(), sz_test, vbfp24.data());
caffe2::fp32_to_bfp14(vv.data(), sz_test, vbfp14.data());
caffe2::fp32_to_bfp16_round(vv.data(), sz_test, vbfp16r.data());
LOG(INFO) << "vector " << vv[0] << " " << vfp16[0] << " " << vbfp16[0] << " "
<< vbfp24[0] << " " << vbfp14[0];
for (auto ii = 0; i < sz_test; i++) {
EXPECT_TRUE(isclose(vfp16[i], v_fp16));
EXPECT_TRUE(isclose(vbfp16[i], v_bfp16));
EXPECT_TRUE(vbfp16[i] == vbfp16s[i]);
EXPECT_TRUE(isclose(vbfp16[i], vbfp16r[i]));
}
}
TEST(FP16Quant, fp32_to_fp16) {
array input, output1, output2;
std::random_device rd;
std::mt19937 e2(rd());
std::uniform_real_distribution<> dist(0, 1);
test_case(10.0, 10.0, 10.0);
test_vector_case(10.0, 10.0, 10.0);
test_case(65000, 64992, 64768);
test_vector_case(65000, 64992, 64768);
test_case(1.2345678910, 1.234375, 1.234375);
test_vector_case(1.2345678910, 1.234375, 1.234375);
test_case(3.3333333, 3.333984375, 3.328125);
test_vector_case(3.3333333, 3.333984375, 3.328125);
test_case(123e10, std::numeric_limits<float>::infinity(), 122.8360646656e10);
test_vector_case(
123e10, std::numeric_limits<float>::infinity(), 122.8360646656e10);
test_case(65504, 65504, 65280);
test_vector_case(65504, 65506, 65280);
test_case(65504 * 4, std::numeric_limits<float>::infinity(), 261120);
test_vector_case(65504 * 4, std::numeric_limits<float>::infinity(), 261120);
test_case(3.402823466e-37, 0, 3.3942e-37);
test_vector_case(3.402823466e-37, 0, 3.3942e-37);
for (auto i = 0; i < sz; i++) {
input[i] = dist(e2) * 1e8;
}
caffe2::fp32_to_fp16(input.data(), input.size(), output1.data());
caffe2::fp32_to_bfp16(input.data(), input.size(), output2.data());
LOG(INFO) << "None " << mse<sz>(input, input);
LOG(INFO) << "FP16 " << mse<sz>(input, output1);
LOG(INFO) << "BFP16 " << mse<sz>(input, output2);
LOG(INFO) << std::hex << std::showbase << *(int*)(&input[0]) << " "
<< input[0];
LOG(INFO) << std::hex << std::showbase << *(int*)(&output1[0]) << " "
<< output1[0];
LOG(INFO) << std::hex << std::showbase << *(int*)(&output2[0]) << " "
<< output2[0];
}
|
