File size: 9,151 Bytes
8b7c501
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
// Copyright 2022 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.

#include <algorithm>
#include <array>
#include <cstddef>
#include <cstdint>
#include <limits>
#include <memory>
#include <random>
#include <vector>

#include <gtest/gtest.h>

#include <xnnpack.h>
#include <xnnpack/node-type.h>
#include <xnnpack/operator-utils.h>
#include <xnnpack/operator.h>
#include <xnnpack/subgraph.h>

class AveragePoolingTestF32 : public ::testing::Test {
protected:
  AveragePoolingTestF32()
  {
    random_device = std::make_unique<std::random_device>();
    rng = std::mt19937((*random_device)());
    input_size_dist = std::uniform_int_distribution<uint32_t>(10, 15);
    pooling_size_dist = std::uniform_int_distribution<uint32_t>(2, 5);
    stride_dist = std::uniform_int_distribution<uint32_t>(1, 2);
    batch_size = input_size_dist(rng);
    input_height = input_size_dist(rng);
    input_width = input_size_dist(rng);
    channels = input_size_dist(rng);
    pooling_height = pooling_size_dist(rng);
    pooling_width = pooling_size_dist(rng);
    // Avoid padding == pooling dimension because it will result in NaNs and cause comparison to fail.
    input_padding_top = std::uniform_int_distribution<uint32_t>(0, pooling_height - 1)(rng);
    input_padding_right = std::uniform_int_distribution<uint32_t>(0, pooling_width - 1)(rng);
    input_padding_bottom = std::uniform_int_distribution<uint32_t>(0, pooling_height - 1)(rng);
    input_padding_left = std::uniform_int_distribution<uint32_t>(0, pooling_width - 1)(rng);
    stride_height = stride_dist(rng);
    stride_width = stride_dist(rng);
    output_height = xnn_compute_convolution_output_dimension(
      input_padding_top + input_height + input_padding_bottom, pooling_height, 1, stride_height);
    output_width = xnn_compute_convolution_output_dimension(
      input_padding_left + input_width + input_padding_right, pooling_width, 1, stride_width);
    output_min = std::uniform_real_distribution<float>(-255.0f, 0.0f)(rng);
    output_max = std::uniform_real_distribution<float>(0.1f, 255.0f)(rng);
    input_dims = {batch_size, input_height, input_width, channels};
    output_dims = {batch_size, output_height, output_width, channels};
    input = std::vector<float>(XNN_EXTRA_BYTES / sizeof(float) + batch_size * input_height * input_width * channels);
    operator_output = std::vector<float>(batch_size * output_height * output_width * channels);
    subgraph_output = std::vector<float>(batch_size * output_height * output_width * channels);
  }

  std::unique_ptr<std::random_device> random_device;
  std::mt19937 rng;
  std::uniform_int_distribution<uint32_t> input_size_dist;
  std::uniform_int_distribution<uint32_t> pooling_size_dist;
  std::uniform_int_distribution<uint32_t> stride_dist;
  uint32_t batch_size;
  uint32_t input_height;
  uint32_t input_width;
  uint32_t channels;
  uint32_t pooling_height;
  uint32_t pooling_width;
  uint32_t output_height;
  uint32_t output_width;
  uint32_t stride_height;
  uint32_t stride_width;
  std::array<size_t, 4> input_dims;
  std::array<size_t, 4> output_dims;
  uint32_t input_padding_top;
  uint32_t input_padding_right;
  uint32_t input_padding_bottom;
  uint32_t input_padding_left;
  float output_min;
  float output_max;

  uint32_t input_id;
  uint32_t output_id;

  std::vector<float> input;
  std::vector<float> operator_output;
  std::vector<float> subgraph_output;
};

TEST_F(AveragePoolingTestF32, define)
{
  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);

  input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr, 0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(), nullptr, 1,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  ASSERT_EQ(
    xnn_status_success,
    xnn_define_average_pooling_2d(
      subgraph, input_padding_top, input_padding_right, input_padding_bottom, input_padding_left, pooling_height,
      pooling_width, stride_height, stride_width, output_min, output_max, input_id, output_id,
      /*flags=*/0));

  ASSERT_EQ(subgraph->num_nodes, 1);
  const struct xnn_node* node = &subgraph->nodes[0];
  ASSERT_EQ(node->type, xnn_node_type_average_pooling_2d);
  ASSERT_EQ(node->compute_type, xnn_compute_type_fp32);

  ASSERT_EQ(node->params.pooling_2d.padding_top, input_padding_top);
  ASSERT_EQ(node->params.pooling_2d.padding_right, input_padding_right);
  ASSERT_EQ(node->params.pooling_2d.padding_bottom, input_padding_bottom);
  ASSERT_EQ(node->params.pooling_2d.padding_left, input_padding_left);
  ASSERT_EQ(node->params.pooling_2d.pooling_height, pooling_height);
  ASSERT_EQ(node->params.pooling_2d.pooling_width, pooling_width);
  ASSERT_EQ(node->params.pooling_2d.stride_height, stride_height);
  ASSERT_EQ(node->params.pooling_2d.stride_width, stride_width);
  ASSERT_EQ(node->activation.output_min, output_min);
  ASSERT_EQ(node->activation.output_max, output_max);
  ASSERT_EQ(node->num_inputs, 1);
  ASSERT_EQ(node->inputs[0], input_id);
  ASSERT_EQ(node->num_outputs, 1);
  ASSERT_EQ(node->outputs[0], output_id);
  ASSERT_EQ(node->flags, 0);
}

TEST_F(AveragePoolingTestF32, matches_operator_api)
{
  std::uniform_real_distribution<float> f32dist;
  std::generate(input.begin(), input.end(), [&]() { return f32dist(rng); });
  std::fill(operator_output.begin(), operator_output.end(), nanf(""));
  std::fill(subgraph_output.begin(), subgraph_output.end(), nanf(""));

  ASSERT_EQ(xnn_status_success, xnn_initialize(/*allocator=*/nullptr));

  // Call operator API.
  xnn_operator_t op = nullptr;
  const xnn_status status = xnn_create_average_pooling2d_nhwc_f32(
    input_padding_top, input_padding_right, input_padding_bottom, input_padding_left, pooling_height, pooling_width,
    stride_height, stride_width, channels, channels, channels, output_min, output_max, /*flags=*/0, &op);
  if (status == xnn_status_unsupported_hardware) {
    GTEST_SKIP();
  }

  ASSERT_EQ(xnn_status_success, status);
  ASSERT_NE(nullptr, op);
  std::unique_ptr<xnn_operator, decltype(&xnn_delete_operator)> auto_op(op, xnn_delete_operator);

  ASSERT_EQ(
    xnn_status_success,
    xnn_reshape_average_pooling2d_nhwc_f32(
      op, batch_size, input_height, input_width, /*output_height_out=*/nullptr, /*output_width_out=*/nullptr,
      /*threadpool=*/nullptr));
  ASSERT_EQ(xnn_status_success, xnn_setup_average_pooling2d_nhwc_f32(op, input.data(), operator_output.data()));

  ASSERT_EQ(xnn_status_success, xnn_run_operator(op, /*threadpool=*/nullptr));

  // Call subgraph API.
  xnn_subgraph_t subgraph = nullptr;
  ASSERT_EQ(xnn_status_success, xnn_create_subgraph(/*external_value_ids=*/2, /*flags=*/0, &subgraph));
  std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> auto_subgraph(subgraph, xnn_delete_subgraph);
  input_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success, xnn_define_tensor_value(
                          subgraph, xnn_datatype_fp32, input_dims.size(), input_dims.data(), nullptr, /*external_id=*/0,
                          /*flags=*/XNN_VALUE_FLAG_EXTERNAL_INPUT, &input_id));
  ASSERT_NE(input_id, XNN_INVALID_NODE_ID);

  output_id = XNN_INVALID_NODE_ID;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_tensor_value(
      subgraph, xnn_datatype_fp32, output_dims.size(), output_dims.data(), nullptr, /*external_id=*/1,
      /*flags=*/XNN_VALUE_FLAG_EXTERNAL_OUTPUT, &output_id));
  ASSERT_NE(output_id, XNN_INVALID_NODE_ID);

  xnn_runtime_t runtime = nullptr;
  ASSERT_EQ(
    xnn_status_success,
    xnn_define_average_pooling_2d(
      subgraph, input_padding_top, input_padding_right, input_padding_bottom, input_padding_left, pooling_height,
      pooling_width, stride_height, stride_width, output_min, output_max, input_id, output_id,
      /*flags=*/0));
  ASSERT_EQ(xnn_status_success, xnn_create_runtime_v3(subgraph, nullptr, nullptr, /*flags=*/0, &runtime));
  ASSERT_NE(nullptr, runtime);
  std::unique_ptr<xnn_runtime, decltype(&xnn_delete_runtime)> auto_runtime(runtime, xnn_delete_runtime);
  std::array<xnn_external_value, 2> external = {
    xnn_external_value{input_id, input.data()}, xnn_external_value{output_id, subgraph_output.data()}};
  ASSERT_EQ(xnn_status_success, xnn_setup_runtime(runtime, external.size(), external.data()));
  ASSERT_EQ(xnn_status_success, xnn_invoke_runtime(runtime));

  ASSERT_EQ(subgraph_output, operator_output);
}