• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (c) Facebook, Inc. and its affiliates.
2 // All rights reserved.
3 //
4 // Copyright 2019 Google LLC
5 //
6 // This source code is licensed under the BSD-style license found in the
7 // LICENSE file in the root directory of this source tree.
8 
9 #pragma once
10 
11 #include <gtest/gtest.h>
12 
13 #include <algorithm>
14 #include <cmath>
15 #include <cassert>
16 #include <cstddef>
17 #include <cstdlib>
18 #include <functional>
19 #include <limits>
20 #include <random>
21 #include <vector>
22 
23 #include <xnnpack.h>
24 
25 
26 class AveragePoolingOperatorTester {
27  public:
padding_tf_same(bool padding_same)28   inline AveragePoolingOperatorTester& padding_tf_same(bool padding_same) {
29     if (padding_same) {
30       assert(padding_top() == 0);
31       assert(padding_left() == 0);
32       assert(padding_bottom() == 0);
33       assert(padding_right() == 0);
34     }
35     this->padding_tf_same_ = padding_same;
36     return *this;
37   }
38 
padding_tf_same()39   inline bool padding_tf_same() const {
40     return this->padding_tf_same_;
41   }
42 
padding(uint32_t padding)43   inline AveragePoolingOperatorTester& padding(uint32_t padding) {
44     assert(!padding_tf_same());
45     this->padding_top_ = padding;
46     this->padding_right_ = padding;
47     this->padding_bottom_ = padding;
48     this->padding_left_ = padding;
49     return *this;
50   }
51 
padding(uint32_t padding_height,uint32_t padding_width)52   inline AveragePoolingOperatorTester& padding(uint32_t padding_height, uint32_t padding_width) {
53     assert(!padding_tf_same());
54     this->padding_top_ = padding_height;
55     this->padding_right_ = padding_width;
56     this->padding_bottom_ = padding_height;
57     this->padding_left_ = padding_width;
58     return *this;
59   }
60 
padding_height(uint32_t padding_height)61   inline AveragePoolingOperatorTester& padding_height(uint32_t padding_height) {
62     assert(!padding_tf_same());
63     this->padding_top_ = padding_height;
64     this->padding_bottom_ = padding_height;
65     return *this;
66   }
67 
padding_width(uint32_t padding_width)68   inline AveragePoolingOperatorTester& padding_width(uint32_t padding_width) {
69     assert(!padding_tf_same());
70     this->padding_right_ = padding_width;
71     this->padding_left_ = padding_width;
72     return *this;
73   }
74 
padding_top(uint32_t padding_top)75   inline AveragePoolingOperatorTester& padding_top(uint32_t padding_top) {
76     assert(!padding_tf_same());
77     this->padding_top_ = padding_top;
78     return *this;
79   }
80 
padding_top()81   inline uint32_t padding_top() const {
82     if (padding_tf_same()) {
83       const uint32_t total_padding_height =
84         (output_height() - 1) * stride_height() + pooling_height() - input_height();
85       return total_padding_height / 2;
86     } else {
87       return this->padding_top_;
88     }
89   }
90 
padding_left(uint32_t padding_left)91   inline AveragePoolingOperatorTester& padding_left(uint32_t padding_left) {
92     assert(!padding_tf_same());
93     this->padding_left_ = padding_left;
94     return *this;
95   }
96 
padding_left()97   inline uint32_t padding_left() const {
98     if (padding_tf_same()) {
99       const uint32_t total_padding_width =
100         (output_width() - 1) * stride_width() + pooling_width() - input_width();
101       return total_padding_width / 2;
102     } else {
103       return this->padding_left_;
104     }
105   }
106 
padding_bottom(uint32_t padding_bottom)107   inline AveragePoolingOperatorTester& padding_bottom(uint32_t padding_bottom) {
108     assert(!padding_tf_same());
109     this->padding_bottom_ = padding_bottom;
110     return *this;
111   }
112 
padding_bottom()113   inline uint32_t padding_bottom() const {
114     if (padding_tf_same()) {
115       const uint32_t total_padding_height =
116         (output_height() - 1) * stride_height() + pooling_height() - input_height();
117       return total_padding_height - total_padding_height / 2;
118     } else {
119       return this->padding_bottom_;
120     }
121   }
122 
padding_right(uint32_t padding_right)123   inline AveragePoolingOperatorTester& padding_right(uint32_t padding_right) {
124     assert(!padding_tf_same());
125     this->padding_right_ = padding_right;
126     return *this;
127   }
128 
padding_right()129   inline uint32_t padding_right() const {
130     if (padding_tf_same()) {
131       const uint32_t total_padding_width =
132         (output_width() - 1) * stride_width() + pooling_width() - input_width();
133       return total_padding_width - total_padding_width / 2;
134     } else {
135       return this->padding_right_;
136     }
137   }
138 
input_size(size_t input_height,size_t input_width)139   inline AveragePoolingOperatorTester& input_size(size_t input_height, size_t input_width) {
140     assert(input_height >= 1);
141     assert(input_width >= 1);
142     this->input_height_ = input_height;
143     this->input_width_ = input_width;
144     return *this;
145   }
146 
input_height(size_t input_height)147   inline AveragePoolingOperatorTester& input_height(size_t input_height) {
148     assert(input_height >= 1);
149     this->input_height_ = input_height;
150     return *this;
151   }
152 
input_height()153   inline size_t input_height() const {
154     return this->input_height_;
155   }
156 
input_width(size_t input_width)157   inline AveragePoolingOperatorTester& input_width(size_t input_width) {
158     assert(input_width >= 1);
159     this->input_width_ = input_width;
160     return *this;
161   }
162 
input_width()163   inline size_t input_width() const {
164     return this->input_width_;
165   }
166 
channels(size_t channels)167   inline AveragePoolingOperatorTester& channels(size_t channels) {
168     assert(channels != 0);
169     this->channels_ = channels;
170     return *this;
171   }
172 
channels()173   inline size_t channels() const {
174     return this->channels_;
175   }
176 
batch_size(size_t batch_size)177   inline AveragePoolingOperatorTester& batch_size(size_t batch_size) {
178     assert(batch_size != 0);
179     this->batch_size_ = batch_size;
180     return *this;
181   }
182 
batch_size()183   inline size_t batch_size() const {
184     return this->batch_size_;
185   }
186 
pooling_size(uint32_t pooling_size)187   inline AveragePoolingOperatorTester& pooling_size(uint32_t pooling_size) {
188     assert(pooling_size >= 1);
189     this->pooling_height_ = pooling_size;
190     this->pooling_width_ = pooling_size;
191     return *this;
192   }
193 
pooling_size(uint32_t pooling_height,uint32_t pooling_width)194   inline AveragePoolingOperatorTester& pooling_size(uint32_t pooling_height, uint32_t pooling_width) {
195     assert(pooling_height >= 1);
196     assert(pooling_width >= 1);
197     this->pooling_height_ = pooling_height;
198     this->pooling_width_ = pooling_width;
199     return *this;
200   }
201 
pooling_height(uint32_t pooling_height)202   inline AveragePoolingOperatorTester& pooling_height(uint32_t pooling_height) {
203     assert(pooling_height >= 1);
204     this->pooling_height_ = pooling_height;
205     return *this;
206   }
207 
pooling_height()208   inline uint32_t pooling_height() const {
209     return this->pooling_height_;
210   }
211 
pooling_width(uint32_t pooling_width)212   inline AveragePoolingOperatorTester& pooling_width(uint32_t pooling_width) {
213     assert(pooling_width >= 1);
214     this->pooling_width_ = pooling_width;
215     return *this;
216   }
217 
pooling_width()218   inline uint32_t pooling_width() const {
219     return this->pooling_width_;
220   }
221 
stride(uint32_t stride)222   inline AveragePoolingOperatorTester& stride(uint32_t stride) {
223     assert(stride >= 1);
224     this->stride_height_ = stride;
225     this->stride_width_ = stride;
226     return *this;
227   }
228 
stride(uint32_t stride_height,uint32_t stride_width)229   inline AveragePoolingOperatorTester& stride(uint32_t stride_height, uint32_t stride_width) {
230     assert(stride_height >= 1);
231     assert(stride_width >= 1);
232     this->stride_height_ = stride_height;
233     this->stride_width_ = stride_width;
234     return *this;
235   }
236 
stride_height(uint32_t stride_height)237   inline AveragePoolingOperatorTester& stride_height(uint32_t stride_height) {
238     assert(stride_height >= 1);
239     this->stride_height_ = stride_height;
240     return *this;
241   }
242 
stride_height()243   inline uint32_t stride_height() const {
244     return this->stride_height_;
245   }
246 
stride_width(uint32_t stride_width)247   inline AveragePoolingOperatorTester& stride_width(uint32_t stride_width) {
248     assert(stride_width >= 1);
249     this->stride_width_ = stride_width;
250     return *this;
251   }
252 
stride_width()253   inline uint32_t stride_width() const {
254     return this->stride_width_;
255   }
256 
output_height()257   inline size_t output_height() const {
258     if (padding_tf_same()) {
259       return (input_height() + stride_height() - 1) / stride_height();
260     } else {
261       const size_t padded_input_height = padding_top() + input_height() + padding_bottom();
262       if (padded_input_height <= pooling_height()) {
263         return 1;
264       } else {
265         return (padded_input_height - pooling_height()) / stride_height() + 1;
266       }
267     }
268   }
269 
output_width()270   inline size_t output_width() const {
271     if (padding_tf_same()) {
272       return (input_width() + stride_width() - 1) / stride_width();
273     } else {
274       const size_t padded_input_width = padding_left() + input_width() + padding_right();
275       if (padded_input_width <= pooling_width()) {
276         return 1;
277       } else {
278         return (padded_input_width - pooling_width()) / stride_width() + 1;
279       }
280     }
281   }
282 
input_pixel_stride(size_t input_pixel_stride)283   inline AveragePoolingOperatorTester& input_pixel_stride(size_t input_pixel_stride) {
284     assert(input_pixel_stride != 0);
285     this->input_pixel_stride_ = input_pixel_stride;
286     return *this;
287   }
288 
input_pixel_stride()289   inline size_t input_pixel_stride() const {
290     if (this->input_pixel_stride_ == 0) {
291       return channels();
292     } else {
293       assert(this->input_pixel_stride_ >= channels());
294       return this->input_pixel_stride_;
295     }
296   }
297 
output_pixel_stride(size_t output_pixel_stride)298   inline AveragePoolingOperatorTester& output_pixel_stride(size_t output_pixel_stride) {
299     assert(output_pixel_stride != 0);
300     this->output_pixel_stride_ = output_pixel_stride;
301     return *this;
302   }
303 
output_pixel_stride()304   inline size_t output_pixel_stride() const {
305     if (this->output_pixel_stride_ == 0) {
306       return channels();
307     } else {
308       assert(this->output_pixel_stride_ >= channels());
309       return this->output_pixel_stride_;
310     }
311   }
312 
next_input_size(uint32_t next_input_height,uint32_t next_input_width)313   inline AveragePoolingOperatorTester& next_input_size(uint32_t next_input_height, uint32_t next_input_width) {
314     assert(next_input_height >= 1);
315     assert(next_input_width >= 1);
316     this->next_input_height_ = next_input_height;
317     this->next_input_width_ = next_input_width;
318     return *this;
319   }
320 
next_input_height(uint32_t next_input_height)321   inline AveragePoolingOperatorTester& next_input_height(uint32_t next_input_height) {
322     assert(next_input_height >= 1);
323     this->next_input_height_ = next_input_height;
324     return *this;
325   }
326 
next_input_height()327   inline uint32_t next_input_height() const {
328     if (this->next_input_height_ == 0) {
329       return input_height();
330     } else {
331       return this->next_input_height_;
332     }
333   }
334 
next_input_width(uint32_t next_input_width)335   inline AveragePoolingOperatorTester& next_input_width(uint32_t next_input_width) {
336     assert(next_input_width >= 1);
337     this->next_input_width_ = next_input_width;
338     return *this;
339   }
340 
next_input_width()341   inline uint32_t next_input_width() const {
342     if (this->next_input_width_ == 0) {
343       return input_width();
344     } else {
345       return this->next_input_width_;
346     }
347   }
348 
next_output_height()349   inline size_t next_output_height() const {
350     const size_t padded_next_input_height = padding_top() + next_input_height() + padding_bottom();
351     if (padded_next_input_height <= pooling_height()) {
352       return 1;
353     } else {
354       return (padded_next_input_height - pooling_height()) / stride_height() + 1;
355     }
356   }
357 
next_output_width()358   inline size_t next_output_width() const {
359     const size_t padded_next_input_width = padding_left() + next_input_width() + padding_right();
360     if (padded_next_input_width <= pooling_width()) {
361       return 1;
362     } else {
363       return (padded_next_input_width - pooling_width()) / stride_width() + 1;
364     }
365   }
366 
next_batch_size(size_t next_batch_size)367   inline AveragePoolingOperatorTester& next_batch_size(size_t next_batch_size) {
368     assert(next_batch_size >= 1);
369     this->next_batch_size_ = next_batch_size;
370     return *this;
371   }
372 
next_batch_size()373   inline size_t next_batch_size() const {
374     if (this->next_batch_size_ == 0) {
375       return batch_size();
376     } else {
377       return this->next_batch_size_;
378     }
379   }
380 
input_scale(float input_scale)381   inline AveragePoolingOperatorTester& input_scale(float input_scale) {
382     assert(input_scale > 0.0f);
383     assert(std::isnormal(input_scale));
384     this->input_scale_ = input_scale;
385     return *this;
386   }
387 
input_scale()388   inline float input_scale() const {
389     return this->input_scale_;
390   }
391 
input_zero_point(uint8_t input_zero_point)392   inline AveragePoolingOperatorTester& input_zero_point(uint8_t input_zero_point) {
393     this->input_zero_point_ = input_zero_point;
394     return *this;
395   }
396 
input_zero_point()397   inline uint8_t input_zero_point() const {
398     return this->input_zero_point_;
399   }
400 
output_scale(float output_scale)401   inline AveragePoolingOperatorTester& output_scale(float output_scale) {
402     assert(output_scale > 0.0f);
403     assert(std::isnormal(output_scale));
404     this->output_scale_ = output_scale;
405     return *this;
406   }
407 
output_scale()408   inline float output_scale() const {
409     return this->output_scale_;
410   }
411 
output_zero_point(uint8_t output_zero_point)412   inline AveragePoolingOperatorTester& output_zero_point(uint8_t output_zero_point) {
413     this->output_zero_point_ = output_zero_point;
414     return *this;
415   }
416 
output_zero_point()417   inline uint8_t output_zero_point() const {
418     return this->output_zero_point_;
419   }
420 
qmin(uint8_t qmin)421   inline AveragePoolingOperatorTester& qmin(uint8_t qmin) {
422     this->qmin_ = qmin;
423     return *this;
424   }
425 
qmin()426   inline uint8_t qmin() const {
427     return this->qmin_;
428   }
429 
qmax(uint8_t qmax)430   inline AveragePoolingOperatorTester& qmax(uint8_t qmax) {
431     this->qmax_ = qmax;
432     return *this;
433   }
434 
qmax()435   inline uint8_t qmax() const {
436     return this->qmax_;
437   }
438 
iterations(size_t iterations)439   inline AveragePoolingOperatorTester& iterations(size_t iterations) {
440     this->iterations_ = iterations;
441     return *this;
442   }
443 
iterations()444   inline size_t iterations() const {
445     return this->iterations_;
446   }
447 
TestQU8()448   void TestQU8() const {
449     std::random_device random_device;
450     auto rng = std::mt19937(random_device());
451     auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), rng);
452 
453     std::vector<uint8_t> input((batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(uint8_t));
454     std::vector<uint8_t> output((batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels());
455     std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels());
456     for (size_t iteration = 0; iteration < iterations(); iteration++) {
457       std::generate(input.begin(), input.end(), std::ref(u8rng));
458       std::fill(output.begin(), output.end(), 0xA5);
459 
460       // Compute reference results.
461       const double scale = double(input_scale()) / (double(output_scale()) * double(pooling_height() * pooling_width()));
462       for (size_t i = 0; i < batch_size(); i++) {
463         for (size_t oy = 0; oy < output_height(); oy++) {
464           for (size_t ox = 0; ox < output_width(); ox++) {
465             for (size_t c = 0; c < channels(); c++) {
466               double acc = 0.0f;
467               for (size_t py = 0; py < pooling_height(); py++) {
468                 const size_t iy = oy * stride_height() + py - padding_top();
469                 for (size_t px = 0; px < pooling_width(); px++) {
470                   const size_t ix = ox * stride_width() + px - padding_left();
471                   if (ix < input_width() && iy < input_height()) {
472                     acc += double(int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]) - int32_t(input_zero_point()));
473                   }
474                 }
475               }
476               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = float(acc * scale + double(output_zero_point()));
477               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] =
478                 std::min<float>(output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c], float(qmax()));
479               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] =
480                 std::max<float>(output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c], float(qmin()));
481             }
482           }
483         }
484       }
485 
486       // Create, setup, run, and destroy Average Pooling operator.
487       ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
488       xnn_operator_t average_pooling_op = nullptr;
489 
490       ASSERT_EQ(xnn_status_success,
491         xnn_create_average_pooling2d_nhwc_qu8(
492           padding_top(), padding_right(), padding_bottom(), padding_left(),
493           pooling_height(), pooling_width(),
494           stride_height(), stride_width(),
495           channels(), input_pixel_stride(), output_pixel_stride(),
496           input_zero_point(), input_scale(),
497           output_zero_point(), output_scale(),
498           qmin(), qmax(),
499           0, &average_pooling_op));
500       ASSERT_NE(nullptr, average_pooling_op);
501 
502       // Smart pointer to automatically delete average_pooling_op.
503       std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_average_pooling_op(average_pooling_op, xnn_delete_operator);
504 
505       ASSERT_EQ(xnn_status_success,
506         xnn_setup_average_pooling2d_nhwc_qu8(
507           average_pooling_op,
508           batch_size(), input_height(), input_width(),
509           input.data(), output.data(),
510           nullptr /* thread pool */));
511 
512       ASSERT_EQ(xnn_status_success,
513         xnn_run_operator(average_pooling_op, nullptr /* thread pool */));
514 
515       // Verify results.
516       for (size_t i = 0; i < batch_size(); i++) {
517         for (size_t y = 0; y < output_height(); y++) {
518           for (size_t x = 0; x < output_width(); x++) {
519             for (size_t c = 0; c < channels(); c++) {
520               ASSERT_LE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmax()));
521               ASSERT_GE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmin()));
522               ASSERT_NEAR(float(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c])),
523                 output_ref[((i * output_height() + y) * output_width() + x) * channels() + c], 0.80f) <<
524                 "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c;
525             }
526           }
527         }
528       }
529     }
530   }
531 
TestF32()532   void TestF32() const {
533     std::random_device random_device;
534     auto rng = std::mt19937(random_device());
535     auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng);
536 
537     std::vector<float> input((batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(float));
538     std::vector<float> output((batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels());
539     std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels());
540     for (size_t iteration = 0; iteration < iterations(); iteration++) {
541       std::generate(input.begin(), input.end(), std::ref(f32rng));
542       std::fill(output.begin(), output.end(), std::nanf(""));
543 
544       // Compute reference results, without clamping.
545       for (size_t i = 0; i < batch_size(); i++) {
546         for (size_t oy = 0; oy < output_height(); oy++) {
547           for (size_t ox = 0; ox < output_width(); ox++) {
548             for (size_t c = 0; c < channels(); c++) {
549               float acc = 0.0f;
550               int32_t n = 0;
551               for (size_t py = 0; py < pooling_height(); py++) {
552                 const size_t iy = oy * stride_height() + py - padding_top();
553                 for (size_t px = 0; px < pooling_width(); px++) {
554                   const size_t ix = ox * stride_width() + px - padding_left();
555                   if (ix < input_width() && iy < input_height()) {
556                     acc += input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c];
557                     n += 1;
558                   }
559                 }
560               }
561               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = acc / float(n);
562             }
563           }
564         }
565       }
566 
567       // Compute clamping parameters.
568       const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend());
569       const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend());
570       const float accumulated_range = accumulated_max - accumulated_min;
571       const float output_min = accumulated_range == 0.0f ?
572         -std::numeric_limits<float>::infinity() :
573         accumulated_min + accumulated_range / 255.0f * float(qmin());
574       const float output_max = accumulated_range == 0.0f ?
575         +std::numeric_limits<float>::infinity() :
576         accumulated_max - accumulated_range / 255.0f * float(255 - qmax());
577 
578       // Clamp reference results.
579       for (float& value : output_ref) {
580         value = std::max(std::min(value, output_max), output_min);
581       }
582 
583       // Create, setup, run, and destroy Average Pooling operator.
584       ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
585       xnn_operator_t average_pooling_op = nullptr;
586 
587       ASSERT_EQ(xnn_status_success,
588         xnn_create_average_pooling2d_nhwc_f32(
589           padding_top(), padding_right(), padding_bottom(), padding_left(),
590           pooling_height(), pooling_width(),
591           stride_height(), stride_width(),
592           channels(), input_pixel_stride(), output_pixel_stride(),
593           output_min, output_max,
594           0, &average_pooling_op));
595       ASSERT_NE(nullptr, average_pooling_op);
596 
597       // Smart pointer to automatically delete average_pooling_op.
598       std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_average_pooling_op(average_pooling_op, xnn_delete_operator);
599 
600       ASSERT_EQ(xnn_status_success,
601         xnn_setup_average_pooling2d_nhwc_f32(
602           average_pooling_op,
603           batch_size(), input_height(), input_width(),
604           input.data(), output.data(),
605           nullptr /* thread pool */));
606 
607       ASSERT_EQ(xnn_status_success,
608         xnn_run_operator(average_pooling_op, nullptr /* thread pool */));
609 
610       // Verify results.
611       for (size_t i = 0; i < batch_size(); i++) {
612         for (size_t y = 0; y < output_height(); y++) {
613           for (size_t x = 0; x < output_width(); x++) {
614             for (size_t c = 0; c < channels(); c++) {
615               ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_max);
616               ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_min);
617               ASSERT_NEAR(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c],
618                   output_ref[((i * output_height() + y) * output_width() + x) * channels() + c],
619                   std::abs(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]) * 1.0e-6f) <<
620                 "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c;
621             }
622           }
623         }
624       }
625     }
626   }
627 
TestSetupQU8()628   void TestSetupQU8() const {
629     std::random_device random_device;
630     auto rng = std::mt19937(random_device());
631     auto u8rng = std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), rng);
632 
633     std::vector<uint8_t> input(XNN_EXTRA_BYTES / sizeof(uint8_t) + std::max(
634       (batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels(),
635       (next_batch_size() * next_input_height() * next_input_width() - 1) * input_pixel_stride() + channels()));
636     std::vector<uint8_t> output(std::max(
637       (batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels(),
638       (next_batch_size() * next_output_height() * next_output_width() - 1) * output_pixel_stride() + channels()));
639     std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels());
640     std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * channels());
641     for (size_t iteration = 0; iteration < iterations(); iteration++) {
642       std::generate(input.begin(), input.end(), std::ref(u8rng));
643       std::fill(output.begin(), output.end(), 0xA5);
644 
645       // Compute reference results.
646       const double scale = double(input_scale()) / (double(output_scale()) * double(pooling_height() * pooling_width()));
647       for (size_t i = 0; i < batch_size(); i++) {
648         for (size_t oy = 0; oy < output_height(); oy++) {
649           for (size_t ox = 0; ox < output_width(); ox++) {
650             for (size_t c = 0; c < channels(); c++) {
651               double acc = 0.0f;
652               for (size_t py = 0; py < pooling_height(); py++) {
653                 const size_t iy = oy * stride_height() + py - padding_top();
654                 for (size_t px = 0; px < pooling_width(); px++) {
655                   const size_t ix = ox * stride_width() + px - padding_left();
656                   if (ix < input_width() && iy < input_height()) {
657                     acc += double(int32_t(input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c]) - int32_t(input_zero_point()));
658                   }
659                 }
660               }
661               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = float(acc * scale + double(output_zero_point()));
662               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] =
663                 std::min<float>(output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c], float(qmax()));
664               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] =
665                 std::max<float>(output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c], float(qmin()));
666             }
667           }
668         }
669       }
670 
671       // Create, setup, and run Average Pooling operator once.
672       ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
673       xnn_operator_t average_pooling_op = nullptr;
674 
675       ASSERT_EQ(xnn_status_success,
676         xnn_create_average_pooling2d_nhwc_qu8(
677           padding_top(), padding_right(), padding_bottom(), padding_left(),
678           pooling_height(), pooling_width(),
679           stride_height(), stride_width(),
680           channels(), input_pixel_stride(), output_pixel_stride(),
681           input_zero_point(), input_scale(),
682           output_zero_point(), output_scale(),
683           qmin(), qmax(),
684           0, &average_pooling_op));
685       ASSERT_NE(nullptr, average_pooling_op);
686 
687       ASSERT_EQ(xnn_status_success,
688         xnn_setup_average_pooling2d_nhwc_qu8(
689           average_pooling_op,
690           batch_size(), input_height(), input_width(),
691           input.data(), output.data(),
692           nullptr /* thread pool */));
693 
694       ASSERT_EQ(xnn_status_success,
695         xnn_run_operator(average_pooling_op, nullptr /* thread pool */));
696 
697       // Verify results of the first run.
698       for (size_t i = 0; i < batch_size(); i++) {
699         for (size_t y = 0; y < output_height(); y++) {
700           for (size_t x = 0; x < output_width(); x++) {
701             for (size_t c = 0; c < channels(); c++) {
702               ASSERT_LE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmax()));
703               ASSERT_GE(uint32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c]), uint32_t(qmin()));
704               ASSERT_NEAR(float(int32_t(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c])),
705                 output_ref[((i * output_height() + y) * output_width() + x) * channels() + c], 0.80f) <<
706                 "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c;
707             }
708           }
709         }
710       }
711 
712       // Re-generate data for the second run.
713       std::generate(input.begin(), input.end(), std::ref(u8rng));
714       std::fill(output.begin(), output.end(), 0xA5);
715 
716       // Compute reference results for the second run.
717       for (size_t i = 0; i < next_batch_size(); i++) {
718         for (size_t oy = 0; oy < next_output_height(); oy++) {
719           for (size_t ox = 0; ox < next_output_width(); ox++) {
720             for (size_t c = 0; c < channels(); c++) {
721               double acc = 0.0f;
722               for (size_t py = 0; py < pooling_height(); py++) {
723                 const size_t iy = oy * stride_height() + py - padding_top();
724                 for (size_t px = 0; px < pooling_width(); px++) {
725                   const size_t ix = ox * stride_width() + px - padding_left();
726                   if (ix < next_input_width() && iy < next_input_height()) {
727                     acc += double(int32_t(input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + c]) - int32_t(input_zero_point()));
728                   }
729                 }
730               }
731               next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] = float(acc * scale + double(output_zero_point()));
732               next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] =
733                 std::min<float>(next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c], float(qmax()));
734               next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] =
735                 std::max<float>(next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c], float(qmin()));
736             }
737           }
738         }
739       }
740 
741       // Setup and run Average Pooling operator the second time, and destroy the operator.
742       ASSERT_EQ(xnn_status_success,
743         xnn_setup_average_pooling2d_nhwc_qu8(
744           average_pooling_op,
745           next_batch_size(), next_input_height(), next_input_width(),
746           input.data(), output.data(),
747           nullptr /* thread pool */));
748 
749       ASSERT_EQ(xnn_status_success,
750         xnn_run_operator(average_pooling_op, nullptr /* thread pool */));
751 
752       ASSERT_EQ(xnn_status_success,
753         xnn_delete_operator(average_pooling_op));
754       average_pooling_op = nullptr;
755 
756       // Verify results of the second run.
757       for (size_t i = 0; i < next_batch_size(); i++) {
758         for (size_t y = 0; y < next_output_height(); y++) {
759           for (size_t x = 0; x < next_output_width(); x++) {
760             for (size_t c = 0; c < channels(); c++) {
761               ASSERT_LE(uint32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), uint32_t(qmax()));
762               ASSERT_GE(uint32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c]), uint32_t(qmin()));
763               ASSERT_NEAR(float(int32_t(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c])),
764                 next_output_ref[((i * next_output_height() + y) * next_output_width() + x) * channels() + c], 0.80f) <<
765                 "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c;
766             }
767           }
768         }
769       }
770     }
771   }
772 
TestSetupF32()773   void TestSetupF32() const {
774     std::random_device random_device;
775     auto rng = std::mt19937(random_device());
776     auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng);
777 
778     std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + std::max(
779       (batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels(),
780       (next_batch_size() * next_input_height() * next_input_width() - 1) * input_pixel_stride() + channels()));
781     std::vector<float> output(std::max(
782       (batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels(),
783       (next_batch_size() * next_output_height() * next_output_width() - 1) * output_pixel_stride() + channels()));
784     std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels());
785     std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * channels());
786     for (size_t iteration = 0; iteration < iterations(); iteration++) {
787       std::generate(input.begin(), input.end(), std::ref(f32rng));
788       std::fill(output.begin(), output.end(), std::nanf(""));
789 
790       // Compute reference results, without clamping.
791       for (size_t i = 0; i < batch_size(); i++) {
792         for (size_t oy = 0; oy < output_height(); oy++) {
793           for (size_t ox = 0; ox < output_width(); ox++) {
794             for (size_t c = 0; c < channels(); c++) {
795               float acc = 0.0f;
796               size_t n = 0;
797               for (size_t py = 0; py < pooling_height(); py++) {
798                 const size_t iy = oy * stride_height() + py - padding_top();
799                 for (size_t px = 0; px < pooling_width(); px++) {
800                   const size_t ix = ox * stride_width() + px - padding_left();
801                   if (ix < input_width() && iy < input_height()) {
802                     acc += input[((i * input_height() + iy) * input_width() + ix) * input_pixel_stride() + c];
803                     n += 1;
804                   }
805                 }
806               }
807               output_ref[((i * output_height() + oy) * output_width() + ox) * channels() + c] = acc / float(n);
808             }
809           }
810         }
811       }
812 
813       // Compute clamping parameters.
814       const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend());
815       const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend());
816       const float accumulated_range = accumulated_max - accumulated_min;
817       const float output_min = accumulated_range == 0.0f ?
818         -std::numeric_limits<float>::infinity() :
819         accumulated_min + accumulated_range / 255.0f * float(qmin());
820       const float output_max = accumulated_range == 0.0f ?
821         +std::numeric_limits<float>::infinity() :
822         accumulated_max - accumulated_range / 255.0f * float(255 - qmax());
823 
824       // Clamp reference results.
825       for (float& value : output_ref) {
826         value = std::max(std::min(value, output_max), output_min);
827       }
828 
829       // Create, setup, and run Average Pooling operator once.
830       ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
831       xnn_operator_t average_pooling_op = nullptr;
832 
833       ASSERT_EQ(xnn_status_success,
834         xnn_create_average_pooling2d_nhwc_f32(
835           padding_top(), padding_right(), padding_bottom(), padding_left(),
836           pooling_height(), pooling_width(),
837           stride_height(), stride_width(),
838           channels(), input_pixel_stride(), output_pixel_stride(),
839           output_min, output_max,
840           0, &average_pooling_op));
841       ASSERT_NE(nullptr, average_pooling_op);
842 
843       ASSERT_EQ(xnn_status_success,
844         xnn_setup_average_pooling2d_nhwc_f32(
845           average_pooling_op,
846           batch_size(), input_height(), input_width(),
847           input.data(), output.data(),
848           nullptr /* thread pool */));
849 
850       ASSERT_EQ(xnn_status_success,
851         xnn_run_operator(average_pooling_op, nullptr /* thread pool */));
852 
853       // Verify results of the first run.
854       for (size_t i = 0; i < batch_size(); i++) {
855         for (size_t y = 0; y < output_height(); y++) {
856           for (size_t x = 0; x < output_width(); x++) {
857             for (size_t c = 0; c < channels(); c++) {
858               ASSERT_LE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_max);
859               ASSERT_GE(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_min);
860               ASSERT_NEAR(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c],
861                   output_ref[((i * output_height() + y) * output_width() + x) * channels() + c],
862                   std::abs(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]) * 1.0e-6f) <<
863                 "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c;
864             }
865           }
866         }
867       }
868 
869       // Re-generate data for the second run.
870       std::generate(input.begin(), input.end(), std::ref(f32rng));
871       std::fill(output.begin(), output.end(), std::nanf(""));
872 
873       // Compute reference results for the second run.
874       for (size_t i = 0; i < next_batch_size(); i++) {
875         for (size_t oy = 0; oy < next_output_height(); oy++) {
876           for (size_t ox = 0; ox < next_output_width(); ox++) {
877             for (size_t c = 0; c < channels(); c++) {
878               float acc = 0.0f;
879               int32_t n = 0;
880               for (size_t py = 0; py < pooling_height(); py++) {
881                 const size_t iy = oy * stride_height() + py - padding_top();
882                 for (size_t px = 0; px < pooling_width(); px++) {
883                   const size_t ix = ox * stride_width() + px - padding_left();
884                   if (ix < next_input_width() && iy < next_input_height()) {
885                     acc += input[((i * next_input_height() + iy) * next_input_width() + ix) * input_pixel_stride() + c];
886                     n += 1;
887                   }
888                 }
889               }
890               next_output_ref[((i * next_output_height() + oy) * next_output_width() + ox) * channels() + c] =
891                 std::max(std::min(acc / float(n), output_max), output_min);
892             }
893           }
894         }
895       }
896 
897       // Setup and run Average Pooling operator the second time, and destroy the operator.
898       ASSERT_EQ(xnn_status_success,
899         xnn_setup_average_pooling2d_nhwc_f32(
900           average_pooling_op,
901           next_batch_size(), next_input_height(), next_input_width(),
902           input.data(), output.data(),
903           nullptr /* thread pool */));
904 
905       ASSERT_EQ(xnn_status_success,
906         xnn_run_operator(average_pooling_op, nullptr /* thread pool */));
907 
908       ASSERT_EQ(xnn_status_success,
909         xnn_delete_operator(average_pooling_op));
910       average_pooling_op = nullptr;
911 
912       // Verify results of the second run.
913       for (size_t i = 0; i < next_batch_size(); i++) {
914         for (size_t y = 0; y < next_output_height(); y++) {
915           for (size_t x = 0; x < next_output_width(); x++) {
916             for (size_t c = 0; c < channels(); c++) {
917               ASSERT_LE(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c], output_max);
918               ASSERT_GE(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c], output_min);
919               ASSERT_NEAR(output[((i * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c],
920                   next_output_ref[((i * next_output_height() + y) * next_output_width() + x) * channels() + c],
921                   std::abs(next_output_ref[((i * next_output_height() + y) * next_output_width() + x) * channels() + c]) * 1.0e-6f) <<
922                 "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c;
923             }
924           }
925         }
926       }
927     }
928   }
929 
930  private:
931   uint32_t padding_top_{0};
932   uint32_t padding_right_{0};
933   uint32_t padding_bottom_{0};
934   uint32_t padding_left_{0};
935   bool padding_tf_same_{false};
936   size_t input_height_{1};
937   size_t input_width_{1};
938   size_t channels_{1};
939   size_t batch_size_{1};
940   size_t input_pixel_stride_{0};
941   size_t output_pixel_stride_{0};
942   uint32_t pooling_height_{1};
943   uint32_t pooling_width_{1};
944   uint32_t stride_height_{1};
945   uint32_t stride_width_{1};
946   size_t next_input_height_{0};
947   size_t next_input_width_{0};
948   size_t next_batch_size_{0};
949   float input_scale_{1.0f};
950   float output_scale_{1.0f};
951   uint8_t input_zero_point_{121};
952   uint8_t output_zero_point_{133};
953   uint8_t qmin_{0};
954   uint8_t qmax_{255};
955   size_t iterations_{1};
956 };
957