forked from google/XNNPACK
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathrdsum-microkernel-tester.h
302 lines (255 loc) · 9.55 KB
/
rdsum-microkernel-tester.h
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
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
// Copyright 2024 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#pragma once
#include <algorithm>
#include <cassert>
#include <cmath>
#include <cstddef>
#include <cstdint>
#include <cstdlib>
#include <limits>
#include <random>
#include <vector>
#include <gtest/gtest.h>
#include <fp16/fp16.h>
#include "xnnpack.h"
#include "xnnpack/aligned-allocator.h"
#include "xnnpack/microfnptr.h"
#include "xnnpack/microparams.h"
#include "xnnpack/requantization.h"
#include "replicable_random_device.h"
class RDSumMicrokernelTester {
public:
RDSumMicrokernelTester& rows(size_t rows) {
assert(rows != 0);
this->rows_ = rows;
return *this;
}
size_t rows() const {
return this->rows_;
}
RDSumMicrokernelTester& channels(size_t channels) {
assert(channels != 0);
this->channels_ = channels;
return *this;
}
size_t channels() const {
return this->channels_;
}
RDSumMicrokernelTester& channel_tile(size_t channel_tile) {
assert(channel_tile != 0);
this->channel_tile_ = channel_tile;
return *this;
}
size_t channel_tile() const {
return this->channel_tile_;
}
RDSumMicrokernelTester& input_stride(size_t input_stride) {
assert(input_stride != 0);
this->input_stride_ = input_stride;
return *this;
}
size_t input_stride() const {
if (this->input_stride_ == 0) {
return channels();
} else {
assert(this->input_stride_ >= channels());
return this->input_stride_;
}
}
RDSumMicrokernelTester& input_scale(float input_scale) {
assert(input_scale > 0.0f);
assert(std::isnormal(input_scale));
this->input_scale_ = input_scale;
return *this;
}
float input_scale() const {
return this->input_scale_;
}
RDSumMicrokernelTester& input_zero_point(uint8_t input_zero_point) {
this->input_zero_point_ = input_zero_point;
return *this;
}
uint8_t input_zero_point() const {
return this->input_zero_point_;
}
RDSumMicrokernelTester& output_scale(float output_scale) {
assert(output_scale > 0.0f);
assert(std::isnormal(output_scale));
this->output_scale_ = output_scale;
return *this;
}
float output_scale() const {
return this->output_scale_;
}
RDSumMicrokernelTester& output_zero_point(uint8_t output_zero_point) {
this->output_zero_point_ = output_zero_point;
return *this;
}
uint8_t output_zero_point() const {
return this->output_zero_point_;
}
RDSumMicrokernelTester& iterations(size_t iterations) {
this->iterations_ = iterations;
return *this;
}
size_t iterations() const {
return this->iterations_;
}
uint8_t qmin() const {
return this->qmin_;
}
uint8_t qmax() const {
return this->qmax_;
}
void Test(xnn_qs8_rdsum_ukernel_fn rdsum,
xnn_init_qs8_rsum_params_fn init_params = nullptr) const {
xnnpack::ReplicableRandomDevice rng;
std::uniform_int_distribution<int32_t> i8dist(
std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
std::vector<int8_t> input((rows() - 1) * input_stride() + channels() + XNN_EXTRA_BYTES);
std::vector<int8_t> zero(channels() + XNN_EXTRA_BYTES, 0);
std::vector<int32_t> output(channels());
std::vector<int32_t> output_ref(channels());
{//for (size_t iteration = 0; iteration < iterations(); iteration++) {
std::generate(input.begin(), input.end(), [&]() { return i8dist(rng); });
std::generate(output.begin(), output.end(), [&]() { return i8dist(rng); });
std::fill(output.begin(), output.end(), 0);
output_ref = output;
// Compute reference results, without clamping.
for (size_t c = 0; c < channels(); c++) {
for (size_t n = 0; n < rows(); n++) {
output_ref[c] += int32_t(input[n * input_stride() + c]);
}
}
// Prepare parameters.
struct xnn_qs8_rsum_params params;
if (init_params) {
init_params(¶ms);
}
// Call optimized micro-kernel.
rdsum(rows(), channels(), input.data(), input_stride(), zero.data(), output.data(), ¶ms);
// Verify results.
for (size_t c = 0; c < channels(); c++) {
EXPECT_EQ(output[c], output_ref[c])
<< "at position " << c << ", rows = " << rows() << ", channels = " << channels();
}
}
}
void Test(xnn_qu8_rdsum_ukernel_fn rdsum,
xnn_init_qs8_rsum_params_fn init_params = nullptr) const {
xnnpack::ReplicableRandomDevice rng;
std::uniform_int_distribution<int32_t> u8dist(
std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max());
std::vector<uint8_t> input((rows() - 1) * input_stride() + channels() + XNN_EXTRA_BYTES);
std::vector<uint8_t> zero(channels() + XNN_EXTRA_BYTES, 0);
std::vector<uint32_t> output(channels());
std::vector<uint32_t> output_ref(channels());
{
std::generate(input.begin(), input.end(), [&]() { return u8dist(rng); });
std::generate(output.begin(), output.end(), [&]() { return u8dist(rng); });
std::fill(output.begin(), output.end(), 0);
output_ref = output;
// Compute reference results, without clamping.
for (size_t c = 0; c < channels(); c++) {
for (size_t n = 0; n < rows(); n++) {
output_ref[c] += uint32_t(input[n * input_stride() + c]);
}
}
// Prepare parameters.
struct xnn_qs8_rsum_params params;
if (init_params) {
init_params(¶ms);
}
// Call optimized micro-kernel.
rdsum(rows(), channels(), input.data(), input_stride(), zero.data(), output.data(), ¶ms);
// Verify results.
for (size_t c = 0; c < channels(); c++) {
EXPECT_EQ(output[c], output_ref[c])
<< "at position " << c << ", rows = " << rows() << ", channels = " << channels();
}
}
}
void Test(xnn_f16_f32acc_rdsum_ukernel_fn rdsum, xnn_init_f16_f32acc_scale_params_fn init_params) const {
xnnpack::ReplicableRandomDevice rng;
std::uniform_real_distribution<float> f32dist(0.01f, 1.0f);
std::vector<xnn_float16> input((rows() - 1) * input_stride() + channels() + XNN_EXTRA_BYTES / sizeof(xnn_float16));
std::vector<xnn_float16> zero(channels() + XNN_EXTRA_BYTES / sizeof(xnn_float16), 0);
std::vector<float> output(channels());
std::vector<float> output_ref(channels());
for (size_t iteration = 0; iteration < iterations(); iteration++) {
std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
std::generate(output.begin(), output.end(), [&]() { return f32dist(rng); });
for (size_t i = 0; i < output.size(); ++i) {
output_ref[i] = output[i];
}
// Compute reference results, without clamping.
for (size_t c = 0; c < channels(); c++) {
float acc = 0.0f;
for (size_t n = 0; n < rows(); n++) {
acc += input[n * input_stride() + c];
}
output_ref[c] += acc / float(rows());
}
// Prepare parameters.
struct xnn_f16_f32acc_scale_params params;
init_params(¶ms, 1.f / float(rows()));
// Call optimized micro-kernel.
rdsum(rows(), channels(), input.data(), input_stride() * sizeof(xnn_float16), zero.data(), output.data(), ¶ms);
// Verify results.
for (size_t c = 0; c < channels(); c++) {
EXPECT_NEAR(output[c], output_ref[c], std::abs(output_ref[c]) * 1.0e-5f)
<< "at position " << c << ", rows = " << rows() << ", channels = " << channels();
}
}
}
void Test(xnn_f32_rdsum_ukernel_fn rdsum, xnn_init_f32_scaleminmax_params_fn init_params) const {
xnnpack::ReplicableRandomDevice rng;
std::uniform_real_distribution<float> f32dist;
std::vector<float> input((rows() - 1) * input_stride() + channels() + XNN_EXTRA_BYTES / sizeof(float));
std::vector<float> zero(channels() + XNN_EXTRA_BYTES / sizeof(float));
std::vector<float> output(channels());
std::vector<float> output_ref(channels());
for (size_t iteration = 0; iteration < iterations(); iteration++) {
std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
std::generate(output.begin(), output.end(), [&]() { return f32dist(rng); });
output_ref = output;
// Prepare parameters.
struct xnn_f32_scaleminmax_params params;
auto input_min = std::min_element(input.begin(), input.end());
auto input_max = std::max_element(input.begin(), input.end());
float mi = *input_min + (*input_max - *input_min) * 0.05;
float ma = *input_max - (*input_min - *input_max) * 0.05;
init_params(¶ms, 1.0f / float(rows()), mi, ma);
// Compute reference results.
for (size_t c = 0; c < channels(); c++) {
float acc = 0.0f;
for (size_t n = 0; n < rows(); n++) {
acc += input[n * input_stride() + c];
}
output_ref[c] += std::max(std::min(acc / float(rows()), ma), mi);
}
// Call optimized micro-kernel.
rdsum(rows(), channels(), input.data(), input_stride() * sizeof(float), zero.data(), output.data(), ¶ms);
// Verify results.
for (size_t c = 0; c < channels(); c++) {
EXPECT_NEAR(output[c], output_ref[c], std::abs(output_ref[c]) * 1.0e-6f)
<< "at position " << c << ", rows = " << rows() << ", channels = " << channels();
}
}
}
private:
size_t rows_{1};
size_t channels_{1};
size_t channel_tile_{1};
size_t input_stride_{0};
float input_scale_{1.25f};
float output_scale_{0.75f};
uint8_t input_zero_point_{121};
uint8_t output_zero_point_{133};
size_t iterations_{3};
uint8_t qmin_{0};
uint8_t qmax_{255};
};