1 /*
2 * Copyright (c) 2018, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <tuple>
13
14 #include "third_party/googletest/src/googletest/include/gtest/gtest.h"
15
16 #include "aom/aom_integer.h"
17 #include "aom_ports/aom_timer.h"
18 #include "av1/encoder/ml.h"
19 #include "config/aom_config.h"
20 #include "config/aom_dsp_rtcd.h"
21 #include "config/av1_rtcd.h"
22 #include "test/util.h"
23 #include "test/register_state_check.h"
24 #include "test/acm_random.h"
25 #include "test/clear_system_state.h"
26
27 namespace {
28 typedef void (*NnPredict_Func)(const float *const input_nodes,
29 const NN_CONFIG *const nn_config,
30 int reduce_prec, float *const output);
31
32 typedef std::tuple<const NnPredict_Func> NnPredictTestParam;
33
34 const float epsilon = 1e-3f; // Error threshold for functional equivalence
35
36 class NnPredictTest : public ::testing::TestWithParam<NnPredictTestParam> {
37 public:
SetUp()38 virtual void SetUp() {
39 const int MAX_NODES2 = NN_MAX_NODES_PER_LAYER * NN_MAX_NODES_PER_LAYER;
40 // Allocate two massive buffers on the heap for edge weights and node bias
41 // Then set-up the double-dimension arrays pointing into the big buffers
42 weights_buf = (float *)aom_malloc(MAX_NODES2 * (NN_MAX_HIDDEN_LAYERS + 1) *
43 sizeof(*weights_buf));
44 bias_buf =
45 (float *)aom_malloc(NN_MAX_NODES_PER_LAYER *
46 (NN_MAX_HIDDEN_LAYERS + 1) * sizeof(*bias_buf));
47 ASSERT_NE(weights_buf, nullptr);
48 ASSERT_NE(bias_buf, nullptr);
49 for (int i = 0; i < NN_MAX_HIDDEN_LAYERS + 1; i++) {
50 weights[i] = &weights_buf[i * MAX_NODES2];
51 bias[i] = &bias_buf[i * NN_MAX_NODES_PER_LAYER];
52 }
53 target_func_ = GET_PARAM(0);
54 }
TearDown()55 virtual void TearDown() {
56 aom_free(weights_buf);
57 aom_free(bias_buf);
58 }
59 void RunNnPredictTest(const NN_CONFIG *const shape);
60 void RunNnPredictSpeedTest(const NN_CONFIG *const shape, const int run_times);
61 void RunNnPredictTest_all(const NN_CONFIG *const shapes,
62 const int num_shapes);
63 void RunNnPredictSpeedTest_all(const NN_CONFIG *const shapes,
64 const int num_shapes, const int run_times);
65
66 private:
67 NnPredict_Func target_func_;
68 libaom_test::ACMRandom rng_;
69 float *weights[NN_MAX_HIDDEN_LAYERS + 1] = { 0 };
70 float *bias[NN_MAX_HIDDEN_LAYERS + 1] = { 0 };
71 float *weights_buf = nullptr, *bias_buf = nullptr;
72 };
73
RunNnPredictTest(const NN_CONFIG * const shape)74 void NnPredictTest::RunNnPredictTest(const NN_CONFIG *const shape) {
75 libaom_test::ClearSystemState();
76 float inputs[NN_MAX_NODES_PER_LAYER] = { 0 };
77 float outputs_test[NN_MAX_NODES_PER_LAYER] = { 0 };
78 float outputs_ref[NN_MAX_NODES_PER_LAYER] = { 0 };
79
80 NN_CONFIG nn_config;
81 memcpy(&nn_config, shape, sizeof(nn_config));
82
83 char shape_str[32] = { 0 };
84 snprintf(shape_str, sizeof(shape_str), "%d", shape->num_inputs);
85 for (int layer = 0; layer < shape->num_hidden_layers; layer++)
86 snprintf(&shape_str[strlen(shape_str)],
87 sizeof(shape_str) - strlen(shape_str), "x%d",
88 shape->num_hidden_nodes[layer]);
89 snprintf(&shape_str[strlen(shape_str)], sizeof(shape_str) - strlen(shape_str),
90 "x%d", shape->num_outputs);
91
92 for (int i = 0; i < NN_MAX_HIDDEN_LAYERS + 1; i++) {
93 nn_config.weights[i] = weights[i];
94 nn_config.bias[i] = bias[i];
95 }
96
97 for (int iter = 0; iter < 10000 && !HasFatalFailure(); ++iter) {
98 for (int node = 0; node < shape->num_inputs; node++) {
99 inputs[node] = ((float)rng_.Rand31() - (1 << 30)) / (1u << 31);
100 }
101 for (int layer = 0; layer < shape->num_hidden_layers; layer++) {
102 for (int node = 0; node < NN_MAX_NODES_PER_LAYER; node++) {
103 bias[layer][node] = ((float)rng_.Rand31() - (1 << 30)) / (1u << 31);
104 }
105 for (int node = 0; node < NN_MAX_NODES_PER_LAYER * NN_MAX_NODES_PER_LAYER;
106 node++) {
107 weights[layer][node] = ((float)rng_.Rand31() - (1 << 30)) / (1u << 31);
108 }
109 }
110 // Now the outputs:
111 int layer = shape->num_hidden_layers;
112 for (int node = 0; node < NN_MAX_NODES_PER_LAYER; node++) {
113 bias[layer][node] = ((float)rng_.Rand31() - (1 << 30)) / (1u << 31);
114 }
115 for (int node = 0; node < NN_MAX_NODES_PER_LAYER * NN_MAX_NODES_PER_LAYER;
116 node++) {
117 weights[layer][node] = ((float)rng_.Rand31() - (1 << 30)) / (1u << 31);
118 }
119
120 av1_nn_predict_c(inputs, &nn_config, 0, outputs_ref);
121 target_func_(inputs, &nn_config, 0, outputs_test);
122 libaom_test::ClearSystemState();
123
124 for (int node = 0; node < shape->num_outputs; node++) {
125 if (outputs_ref[node] < epsilon) {
126 ASSERT_LE(outputs_test[node], epsilon)
127 << "Reference output was near-zero, test output was not ("
128 << shape_str << ")";
129 } else {
130 const float error = outputs_ref[node] - outputs_test[node];
131 const float relative_error = fabsf(error / outputs_ref[node]);
132 ASSERT_LE(relative_error, epsilon)
133 << "Excessive relative error between reference and test ("
134 << shape_str << ")";
135 }
136 }
137 }
138 }
139
RunNnPredictSpeedTest(const NN_CONFIG * const shape,const int run_times)140 void NnPredictTest::RunNnPredictSpeedTest(const NN_CONFIG *const shape,
141 const int run_times) {
142 libaom_test::ClearSystemState();
143 float inputs[NN_MAX_NODES_PER_LAYER] = { 0 };
144 float outputs_test[NN_MAX_NODES_PER_LAYER] = { 0 };
145 float outputs_ref[NN_MAX_NODES_PER_LAYER] = { 0 };
146
147 NN_CONFIG nn_config;
148 memcpy(&nn_config, shape, sizeof(nn_config));
149
150 for (int i = 0; i < NN_MAX_HIDDEN_LAYERS; i++) {
151 nn_config.weights[i] = weights[i];
152 nn_config.bias[i] = bias[i];
153 }
154 // Don't bother actually changing the values for inputs/weights/bias: it
155 // shouldn't make any difference for a speed test.
156
157 aom_usec_timer timer;
158 aom_usec_timer_start(&timer);
159 for (int i = 0; i < run_times; ++i) {
160 av1_nn_predict_c(inputs, &nn_config, 0, outputs_ref);
161 }
162 aom_usec_timer_mark(&timer);
163 const double time1 = static_cast<double>(aom_usec_timer_elapsed(&timer));
164 aom_usec_timer_start(&timer);
165 for (int i = 0; i < run_times; ++i) {
166 target_func_(inputs, &nn_config, 0, outputs_test);
167 }
168 aom_usec_timer_mark(&timer);
169 libaom_test::ClearSystemState();
170 const double time2 = static_cast<double>(aom_usec_timer_elapsed(&timer));
171
172 printf("%d", shape->num_inputs);
173 for (int layer = 0; layer < shape->num_hidden_layers; layer++)
174 printf("x%d", shape->num_hidden_nodes[layer]);
175 printf("x%d: ", shape->num_outputs);
176 printf("%7.2f/%7.2fns (%3.2f)\n", time1, time2, time1 / time2);
177 }
178
179 // This is all the neural network shapes observed executed in a few different
180 // runs of the encoder. It also conveniently covers all the kernels
181 // implemented.
182 static const NN_CONFIG shapes[] = {
183 { 10, 16, 1, { 64 }, { 0 }, { 0 } }, { 12, 1, 1, { 12 }, { 0 }, { 0 } },
184 { 12, 1, 1, { 24 }, { 0 }, { 0 } }, { 12, 1, 1, { 32 }, { 0 }, { 0 } },
185 { 18, 4, 1, { 24 }, { 0 }, { 0 } }, { 18, 4, 1, { 32 }, { 0 }, { 0 } },
186 { 4, 1, 1, { 16 }, { 0 }, { 0 } }, { 8, 1, 1, { 16 }, { 0 }, { 0 } },
187 { 8, 4, 1, { 16 }, { 0 }, { 0 } }, { 8, 1, 1, { 24 }, { 0 }, { 0 } },
188 { 8, 1, 1, { 32 }, { 0 }, { 0 } }, { 8, 1, 1, { 64 }, { 0 }, { 0 } },
189 { 9, 3, 1, { 32 }, { 0 }, { 0 } }, { 4, 4, 1, { 8 }, { 0 }, { 0 } },
190 };
191
RunNnPredictTest_all(const NN_CONFIG * const shapes,const int num_shapes)192 void NnPredictTest::RunNnPredictTest_all(const NN_CONFIG *const shapes,
193 const int num_shapes) {
194 for (int i = 0; i < num_shapes; i++) RunNnPredictTest(&shapes[i]);
195 }
196
RunNnPredictSpeedTest_all(const NN_CONFIG * const shapes,const int num_shapes,const int run_times)197 void NnPredictTest::RunNnPredictSpeedTest_all(const NN_CONFIG *const shapes,
198 const int num_shapes,
199 const int run_times) {
200 for (int i = 0; i < num_shapes; i++)
201 NnPredictTest::RunNnPredictSpeedTest(&shapes[i], run_times);
202 }
203
TEST_P(NnPredictTest,RandomValues)204 TEST_P(NnPredictTest, RandomValues) {
205 RunNnPredictTest_all(shapes, sizeof(shapes) / sizeof(*shapes));
206 }
207
TEST_P(NnPredictTest,DISABLED_Speed)208 TEST_P(NnPredictTest, DISABLED_Speed) {
209 RunNnPredictSpeedTest_all(shapes, sizeof(shapes) / sizeof(*shapes), 10000000);
210 }
211
212 #if HAVE_SSE3
213 INSTANTIATE_TEST_SUITE_P(SSE3, NnPredictTest,
214 ::testing::Values(av1_nn_predict_sse3));
215 #endif
216
217 } // namespace
218