• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #pragma once
7 
8 #include <gtest/gtest.h>
9 
10 #include <algorithm>
11 #include <cmath>
12 #include <cassert>
13 #include <cstddef>
14 #include <cstdlib>
15 #include <functional>
16 #include <random>
17 #include <vector>
18 
19 #include <xnnpack.h>
20 
21 
22 class ResizeBilinearOperatorTester {
23  public:
input_size(size_t input_height,size_t input_width)24   inline ResizeBilinearOperatorTester& input_size(size_t input_height, size_t input_width) {
25     assert(input_height >= 1);
26     assert(input_width >= 1);
27     this->input_height_ = input_height;
28     this->input_width_ = input_width;
29     return *this;
30   }
31 
input_height(size_t input_height)32   inline ResizeBilinearOperatorTester& input_height(size_t input_height) {
33     assert(input_height >= 1);
34     this->input_height_ = input_height;
35     return *this;
36   }
37 
input_height()38   inline size_t input_height() const {
39     return this->input_height_;
40   }
41 
input_width(size_t input_width)42   inline ResizeBilinearOperatorTester& input_width(size_t input_width) {
43     assert(input_width >= 1);
44     this->input_width_ = input_width;
45     return *this;
46   }
47 
input_width()48   inline size_t input_width() const {
49     return this->input_width_;
50   }
51 
output_size(size_t output_height,size_t output_width)52   inline ResizeBilinearOperatorTester& output_size(size_t output_height, size_t output_width) {
53     assert(output_height >= 1);
54     assert(output_width >= 1);
55     this->output_height_ = output_height;
56     this->output_width_ = output_width;
57     return *this;
58   }
59 
output_height(size_t output_height)60   inline ResizeBilinearOperatorTester& output_height(size_t output_height) {
61     assert(output_height >= 1);
62     this->output_height_ = output_height;
63     return *this;
64   }
65 
output_height()66   inline size_t output_height() const {
67     return this->output_height_;
68   }
69 
output_width(size_t output_width)70   inline ResizeBilinearOperatorTester& output_width(size_t output_width) {
71     assert(output_width >= 1);
72     this->output_width_ = output_width;
73     return *this;
74   }
75 
output_width()76   inline size_t output_width() const {
77     return this->output_width_;
78   }
79 
height_scale()80   inline float height_scale() const {
81     if (align_corners() && output_height() > 1) {
82       return float(input_height() - 1) / float(output_height() - 1);
83     } else {
84       return float(input_height()) / float(output_height());
85     }
86   }
87 
width_scale()88   inline float width_scale() const {
89     if (align_corners() && output_width() > 1) {
90       return float(input_width() - 1) / float(output_width() - 1);
91     } else {
92       return float(input_width()) / float(output_width());
93     }
94   }
95 
channels(size_t channels)96   inline ResizeBilinearOperatorTester& channels(size_t channels) {
97     assert(channels != 0);
98     this->channels_ = channels;
99     return *this;
100   }
101 
channels()102   inline size_t channels() const {
103     return this->channels_;
104   }
105 
batch_size(size_t batch_size)106   inline ResizeBilinearOperatorTester& batch_size(size_t batch_size) {
107     assert(batch_size != 0);
108     this->batch_size_ = batch_size;
109     return *this;
110   }
111 
batch_size()112   inline size_t batch_size() const {
113     return this->batch_size_;
114   }
115 
input_pixel_stride(size_t input_pixel_stride)116   inline ResizeBilinearOperatorTester& input_pixel_stride(size_t input_pixel_stride) {
117     assert(input_pixel_stride != 0);
118     this->input_pixel_stride_ = input_pixel_stride;
119     return *this;
120   }
121 
input_pixel_stride()122   inline size_t input_pixel_stride() const {
123     if (this->input_pixel_stride_ == 0) {
124       return channels();
125     } else {
126       assert(this->input_pixel_stride_ >= channels());
127       return this->input_pixel_stride_;
128     }
129   }
130 
output_pixel_stride(size_t output_pixel_stride)131   inline ResizeBilinearOperatorTester& output_pixel_stride(size_t output_pixel_stride) {
132     assert(output_pixel_stride != 0);
133     this->output_pixel_stride_ = output_pixel_stride;
134     return *this;
135   }
136 
output_pixel_stride()137   inline size_t output_pixel_stride() const {
138     if (this->output_pixel_stride_ == 0) {
139       return channels();
140     } else {
141       assert(this->output_pixel_stride_ >= channels());
142       return this->output_pixel_stride_;
143     }
144   }
145 
next_input_size(uint32_t next_input_height,uint32_t next_input_width)146   inline ResizeBilinearOperatorTester& next_input_size(uint32_t next_input_height, uint32_t next_input_width) {
147     assert(next_input_height >= 1);
148     assert(next_input_width >= 1);
149     this->next_input_height_ = next_input_height;
150     this->next_input_width_ = next_input_width;
151     return *this;
152   }
153 
next_input_height(uint32_t next_input_height)154   inline ResizeBilinearOperatorTester& next_input_height(uint32_t next_input_height) {
155     assert(next_input_height >= 1);
156     this->next_input_height_ = next_input_height;
157     return *this;
158   }
159 
next_input_height()160   inline uint32_t next_input_height() const {
161     if (this->next_input_height_ == 0) {
162       return input_height();
163     } else {
164       return this->next_input_height_;
165     }
166   }
167 
next_input_width(uint32_t next_input_width)168   inline ResizeBilinearOperatorTester& next_input_width(uint32_t next_input_width) {
169     assert(next_input_width >= 1);
170     this->next_input_width_ = next_input_width;
171     return *this;
172   }
173 
next_input_width()174   inline uint32_t next_input_width() const {
175     if (this->next_input_width_ == 0) {
176       return input_width();
177     } else {
178       return this->next_input_width_;
179     }
180   }
181 
next_batch_size(size_t next_batch_size)182   inline ResizeBilinearOperatorTester& next_batch_size(size_t next_batch_size) {
183     assert(next_batch_size >= 1);
184     this->next_batch_size_ = next_batch_size;
185     return *this;
186   }
187 
next_batch_size()188   inline size_t next_batch_size() const {
189     if (this->next_batch_size_ == 0) {
190       return batch_size();
191     } else {
192       return this->next_batch_size_;
193     }
194   }
195 
align_corners(bool align_corners)196   inline ResizeBilinearOperatorTester& align_corners(bool align_corners) {
197     this->align_corners_ = align_corners;
198     return *this;
199   }
200 
align_corners()201   inline bool align_corners() const {
202     return this->align_corners_;
203   }
204 
tf_legacy_mode(bool tf_legacy_mode)205   inline ResizeBilinearOperatorTester& tf_legacy_mode(bool tf_legacy_mode) {
206     this->tf_legacy_mode_ = tf_legacy_mode;
207     return *this;
208   }
209 
tf_legacy_mode()210   inline bool tf_legacy_mode() const {
211     return this->tf_legacy_mode_;
212   }
213 
iterations(size_t iterations)214   inline ResizeBilinearOperatorTester& iterations(size_t iterations) {
215     this->iterations_ = iterations;
216     return *this;
217   }
218 
iterations()219   inline size_t iterations() const {
220     return this->iterations_;
221   }
222 
TestF32()223   void TestF32() const {
224     std::random_device random_device;
225     auto rng = std::mt19937(random_device());
226     auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng);
227 
228     std::vector<float> input((batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels() + XNN_EXTRA_BYTES / sizeof(float));
229     std::vector<float> output((batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels());
230     std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels());
231     for (size_t iteration = 0; iteration < iterations(); iteration++) {
232       std::generate(input.begin(), input.end(), std::ref(f32rng));
233       std::fill(output.begin(), output.end(), std::nanf(""));
234 
235       // Compute reference results.
236       const float offset = tf_legacy_mode() ? 0.0f : 0.5f;
237       for (size_t batch_index = 0; batch_index < batch_size(); batch_index++) {
238         for (size_t output_y = 0; output_y < output_height(); output_y++) {
239           const float input_y = (float(output_y) + offset) * height_scale() - offset;
240           const int64_t input_y_top = std::max<int64_t>(int64_t(std::floor(input_y)), 0);
241           const int64_t input_y_bottom = std::min<int64_t>(int64_t(std::ceil(input_y)), input_height() - 1);
242           const float y_alpha = input_y - std::floor(input_y);
243           for (size_t output_x = 0; output_x < output_width(); output_x++) {
244             const float input_x = (float(output_x) + offset) * width_scale() - offset;
245             const int64_t input_x_left = std::max<int64_t>(int64_t(std::floor(input_x)), 0);
246             const int64_t input_x_right = std::min<int64_t>(int64_t(std::ceil(input_x)), input_width() - 1);
247             const float x_alpha = input_x - std::floor(input_x);
248             for (size_t c = 0; c < channels(); c++) {
249               output_ref[((batch_index * output_height() + output_y) * output_width() + output_x) * channels() + c] =
250                 input[((batch_index * input_height() + input_y_top) * input_width() + input_x_left) * input_pixel_stride() + c] * (1.0f - y_alpha) * (1.0f - x_alpha) +
251                 input[((batch_index * input_height() + input_y_top) * input_width() + input_x_right) * input_pixel_stride() + c] * (1.0f - y_alpha) * x_alpha +
252                 input[((batch_index * input_height() + input_y_bottom) * input_width() + input_x_left) * input_pixel_stride() + c] * y_alpha * (1.0f - x_alpha) +
253                 input[((batch_index * input_height() + input_y_bottom) * input_width() + input_x_right) * input_pixel_stride() + c] * y_alpha * x_alpha;
254             }
255           }
256         }
257       }
258 
259       // Create, setup, run, and destroy Resize Bilinear operator.
260       ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
261       xnn_operator_t resize_bilinear_op = nullptr;
262 
263       ASSERT_EQ(xnn_status_success,
264         xnn_create_resize_bilinear2d_nhwc_f32(
265           channels(), input_pixel_stride(), output_pixel_stride(),
266           (align_corners() ? XNN_FLAG_ALIGN_CORNERS : 0) | (tf_legacy_mode() ? XNN_FLAG_TENSORFLOW_LEGACY_MODE : 0),
267           &resize_bilinear_op));
268       ASSERT_NE(nullptr, resize_bilinear_op);
269 
270       // Smart pointer to automatically delete resize_bilinear_op.
271       std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_resize_bilinear_op(resize_bilinear_op, xnn_delete_operator);
272 
273       ASSERT_EQ(xnn_status_success,
274         xnn_setup_resize_bilinear2d_nhwc_f32(
275           resize_bilinear_op,
276           batch_size(), input_height(), input_width(),
277           output_height(), output_width(),
278           input.data(), output.data(),
279           nullptr /* thread pool */));
280 
281       ASSERT_EQ(xnn_status_success,
282         xnn_run_operator(resize_bilinear_op, nullptr /* thread pool */));
283 
284       // Verify results.
285       for (size_t i = 0; i < batch_size(); i++) {
286         for (size_t y = 0; y < output_height(); y++) {
287           for (size_t x = 0; x < output_width(); x++) {
288             for (size_t c = 0; c < channels(); c++) {
289               ASSERT_NEAR(output[((i * output_height() + y) * output_width() + x) * output_pixel_stride() + c],
290                   output_ref[((i * output_height() + y) * output_width() + x) * channels() + c],
291                   std::abs(output_ref[((i * output_height() + y) * output_width() + x) * channels() + c]) * 1.0e-5f) <<
292                 "in batch index " << i << ", pixel (" << y << ", " << x << "), channel " << c;
293             }
294           }
295         }
296       }
297     }
298   }
299 
300   // void TestSetupF32() const {
301   //   std::random_device random_device;
302   //   auto rng = std::mt19937(random_device());
303   //   auto f32rng = std::bind(std::uniform_real_distribution<float>(), rng);
304 
305   //   std::vector<float> input(XNN_EXTRA_BYTES / sizeof(float) + std::max(
306   //     (batch_size() * input_height() * input_width() - 1) * input_pixel_stride() + channels(),
307   //     (next_batch_size() * next_input_height() * next_input_width() - 1) * input_pixel_stride() + channels()));
308   //   std::vector<float> output(std::max(
309   //     (batch_size() * output_height() * output_width() - 1) * output_pixel_stride() + channels(),
310   //     (next_batch_size() * next_output_height() * next_output_width() - 1) * output_pixel_stride() + channels()));
311   //   std::vector<float> output_ref(batch_size() * output_height() * output_width() * channels());
312   //   std::vector<float> next_output_ref(next_batch_size() * next_output_height() * next_output_width() * channels());
313   //   for (size_t iteration = 0; iteration < iterations(); iteration++) {
314   //     std::generate(input.begin(), input.end(), std::ref(f32rng));
315   //     std::fill(output.begin(), output.end(), std::nanf(""));
316 
317   //     // Compute reference results, without clamping.
318   //     for (size_t batch_index = 0; batch_index < batch_size(); batch_index++) {
319   //       for (size_t output_y = 0; output_y < output_height(); output_y++) {
320   //         for (size_t output_x = 0; output_x < output_width(); output_x++) {
321   //           for (size_t c = 0; c < channels(); c++) {
322   //             float acc = 0.0f;
323   //             size_t n = 0;
324   //             for (size_t py = 0; py < pooling_height(); py++) {
325   //               const size_t iy = output_y * stride_height() + py - padding_top();
326   //               for (size_t px = 0; px < pooling_width(); px++) {
327   //                 const size_t input_x = output_x * stride_width() + px - padding_left();
328   //                 if (input_x < input_width() && iy < input_height()) {
329   //                   acc += input[((batch_index * input_height() + iy) * input_width() + input_x) * input_pixel_stride() + c];
330   //                   n += 1;
331   //                 }
332   //               }
333   //             }
334   //             output_ref[((batch_index * output_height() + output_y) * output_width() + output_x) * channels() + c] = acc / float(n);
335   //           }
336   //         }
337   //       }
338   //     }
339 
340   //     // Compute clamping parameters.
341   //     const float accumulated_min = *std::min_element(output_ref.cbegin(), output_ref.cend());
342   //     const float accumulated_max = *std::max_element(output_ref.cbegin(), output_ref.cend());
343   //     const float accumulated_range = accumulated_max - accumulated_min;
344   //     const float output_min = accumulated_range == 0.0f ?
345   //       -std::numeric_limits<float>::infinity() :
346   //       accumulated_min + accumulated_range / 255.0f * float(qmin());
347   //     const float output_max = accumulated_range == 0.0f ?
348   //       +std::numeric_limits<float>::infinity() :
349   //       accumulated_max - accumulated_range / 255.0f * float(255 - qmax());
350 
351   //     // Clamp reference results.
352   //     for (float& value : output_ref) {
353   //       value = std::max(std::min(value, output_max), output_min);
354   //     }
355 
356   //     // Create, setup, and run Average Pooling operator once.
357   //     ASSERT_EQ(xnn_status_success, xnn_initialize(nullptr /* allocator */));
358   //     xnn_operator_t resize_bilinear_op = nullptr;
359 
360   //     ASSERT_EQ(xnn_status_success,
361   //       xnn_create_average_pooling2d_nhwc_f32(
362   //         padding_top(), padding_right(), padding_bottom(), padding_left(),
363   //         pooling_height(), pooling_width(),
364   //         stride_height(), stride_width(),
365   //         channels(), input_pixel_stride(), output_pixel_stride(),
366   //         output_min, output_max,
367   //         0, &resize_bilinear_op));
368   //     ASSERT_NE(nullptr, resize_bilinear_op);
369 
370   //     ASSERT_EQ(xnn_status_success,
371   //       xnn_setup_average_pooling2d_nhwc_f32(
372   //         resize_bilinear_op,
373   //         batch_size(), input_height(), input_width(),
374   //         input.data(), output.data(),
375   //         nullptr /* thread pool */));
376 
377   //     ASSERT_EQ(xnn_status_success,
378   //       xnn_run_operator(resize_bilinear_op, nullptr /* thread pool */));
379 
380   //     // Verify results of the first run.
381   //     for (size_t batch_index = 0; batch_index < batch_size(); batch_index++) {
382   //       for (size_t y = 0; y < output_height(); y++) {
383   //         for (size_t x = 0; x < output_width(); x++) {
384   //           for (size_t c = 0; c < channels(); c++) {
385   //             ASSERT_LE(output[((batch_index * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_max);
386   //             ASSERT_GE(output[((batch_index * output_height() + y) * output_width() + x) * output_pixel_stride() + c], output_min);
387   //             ASSERT_NEAR(output[((batch_index * output_height() + y) * output_width() + x) * output_pixel_stride() + c],
388   //                 output_ref[((batch_index * output_height() + y) * output_width() + x) * channels() + c],
389   //                 std::abs(output_ref[((batch_index * output_height() + y) * output_width() + x) * channels() + c]) * 1.0e-6f) <<
390   //               "in batch index " << batch_index << ", pixel (" << y << ", " << x << "), channel " << c;
391   //           }
392   //         }
393   //       }
394   //     }
395 
396   //     // Re-generate data for the second run.
397   //     std::generate(input.begin(), input.end(), std::ref(f32rng));
398   //     std::fill(output.begin(), output.end(), std::nanf(""));
399 
400   //     // Compute reference results for the second run.
401   //     for (size_t batch_index = 0; batch_index < next_batch_size(); batch_index++) {
402   //       for (size_t output_y = 0; output_y < next_output_height(); output_y++) {
403   //         for (size_t output_x = 0; output_x < next_output_width(); output_x++) {
404   //           for (size_t c = 0; c < channels(); c++) {
405   //             float acc = 0.0f;
406   //             int32_t n = 0;
407   //             for (size_t py = 0; py < pooling_height(); py++) {
408   //               const size_t iy = output_y * stride_height() + py - padding_top();
409   //               for (size_t px = 0; px < pooling_width(); px++) {
410   //                 const size_t input_x = output_x * stride_width() + px - padding_left();
411   //                 if (input_x < next_input_width() && iy < next_input_height()) {
412   //                   acc += input[((batch_index * next_input_height() + iy) * next_input_width() + input_x) * input_pixel_stride() + c];
413   //                   n += 1;
414   //                 }
415   //               }
416   //             }
417   //             next_output_ref[((batch_index * next_output_height() + output_y) * next_output_width() + output_x) * channels() + c] =
418   //               std::max(std::min(acc / float(n), output_max), output_min);
419   //           }
420   //         }
421   //       }
422   //     }
423 
424   //     // Setup and run Average Pooling operator the second time, and destroutput_y the operator.
425   //     ASSERT_EQ(xnn_status_success,
426   //       xnn_setup_average_pooling2d_nhwc_f32(
427   //         resize_bilinear_op,
428   //         next_batch_size(), next_input_height(), next_input_width(),
429   //         input.data(), output.data(),
430   //         nullptr /* thread pool */));
431 
432   //     ASSERT_EQ(xnn_status_success,
433   //       xnn_run_operator(resize_bilinear_op, nullptr /* thread pool */));
434 
435   //     ASSERT_EQ(xnn_status_success,
436   //       xnn_delete_operator(resize_bilinear_op));
437   //     resize_bilinear_op = nullptr;
438 
439   //     // Verify results of the second run.
440   //     for (size_t batch_index = 0; batch_index < next_batch_size(); batch_index++) {
441   //       for (size_t y = 0; y < next_output_height(); y++) {
442   //         for (size_t x = 0; x < next_output_width(); x++) {
443   //           for (size_t c = 0; c < channels(); c++) {
444   //             ASSERT_LE(output[((batch_index * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c], output_max);
445   //             ASSERT_GE(output[((batch_index * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c], output_min);
446   //             ASSERT_NEAR(output[((batch_index * next_output_height() + y) * next_output_width() + x) * output_pixel_stride() + c],
447   //                 next_output_ref[((batch_index * next_output_height() + y) * next_output_width() + x) * channels() + c],
448   //                 std::abs(next_output_ref[((batch_index * next_output_height() + y) * next_output_width() + x) * channels() + c]) * 1.0e-6f) <<
449   //               "in batch index " << batch_index << ", pixel (" << y << ", " << x << "), channel " << c;
450   //           }
451   //         }
452   //       }
453   //     }
454   //   }
455   // }
456 
457  private:
458   size_t input_height_{1};
459   size_t input_width_{1};
460   size_t output_height_{1};
461   size_t output_width_{1};
462   size_t channels_{1};
463   size_t batch_size_{1};
464   size_t input_pixel_stride_{0};
465   size_t output_pixel_stride_{0};
466   size_t next_input_height_{0};
467   size_t next_input_width_{0};
468   size_t next_batch_size_{0};
469   bool align_corners_{false};
470   bool tf_legacy_mode_{false};
471   size_t iterations_{1};
472 };
473