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
|
#include <iostream>
#include <random>
#include "tanh.h"
#include <gtest/gtest.h>
#include "caffe2/core/logging.h"
using namespace dnnlowp;
using namespace std;
TEST(Tanh, TanhUnitTest) {
// For 8bit, we can go down to 0.0145631
// For 9bit in 8bit out, we can go down to 0.00976522
for (double max_abs_err = 0.02; max_abs_err <= 0.1; max_abs_err += 0.01) {
Tanh<uint8_t> tanh_approx(max_abs_err);
LOG(INFO) << "max_abs_err " << max_abs_err << " x_pq "
<< tanh_approx.GetPassRegionEndDequantized() << " x_sq "
<< tanh_approx.GetSaturationRegionBegin();
const int NSAMPLES = 1 << 16;
std::uniform_real_distribution<float> distribution(-5., 5.);
std::default_random_engine generator;
float sq_err_sum = 0, max_err = 0;
for (int i = 0; i < NSAMPLES; ++i) {
float x = distribution(generator);
uint8_t x_q = fbgemm::Quantize<uint8_t>(
x, tanh_approx.GetInputQuantizationParams());
uint8_t y_q = tanh_approx.Compute(x_q);
float y = fbgemm::Dequantize<uint8_t>(
y_q, tanh_approx.GetOutputQuantizationParams());
float err = fabs(tanh(x) - y);
sq_err_sum += err * err;
max_err = std::max(err, max_err);
if (err > max_abs_err) {
LOG(INFO) << "x " << x << " tanh_real " << tanh(x) << " tanh_approx "
<< y << " err " << err << " x_q " << (int)x_q << " y_q "
<< (int)y_q;
}
EXPECT_LE(err, max_abs_err);
}
LOG(INFO) << "avg_l2_err " << std::sqrt(sq_err_sum) / NSAMPLES
<< " max_err " << max_err << endl;
}
}
|
