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