1/* 2 * Copyright (c) 2013 The WebRTC project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include <math.h> 12 13#include "testing/gmock/include/gmock/gmock.h" 14#include "testing/gtest/include/gtest/gtest.h" 15#include "webrtc/common_audio/include/audio_util.h" 16#include "webrtc/common_audio/resampler/push_sinc_resampler.h" 17#include "webrtc/common_audio/resampler/sinusoidal_linear_chirp_source.h" 18#include "webrtc/system_wrappers/interface/scoped_ptr.h" 19#include "webrtc/system_wrappers/interface/tick_util.h" 20#include "webrtc/typedefs.h" 21 22namespace webrtc { 23 24typedef std::tr1::tuple<int, int, double, double> PushSincResamplerTestData; 25class PushSincResamplerTest 26 : public testing::TestWithParam<PushSincResamplerTestData> { 27 public: 28 PushSincResamplerTest() 29 : input_rate_(std::tr1::get<0>(GetParam())), 30 output_rate_(std::tr1::get<1>(GetParam())), 31 rms_error_(std::tr1::get<2>(GetParam())), 32 low_freq_error_(std::tr1::get<3>(GetParam())) { 33 } 34 35 virtual ~PushSincResamplerTest() {} 36 37 protected: 38 void ResampleBenchmarkTest(bool int_format); 39 void ResampleTest(bool int_format); 40 41 int input_rate_; 42 int output_rate_; 43 double rms_error_; 44 double low_freq_error_; 45}; 46 47class ZeroSource : public SincResamplerCallback { 48 public: 49 void Run(int frames, float* destination) { 50 memset(destination, 0, sizeof(float) * frames); 51 } 52}; 53 54void PushSincResamplerTest::ResampleBenchmarkTest(bool int_format) { 55 const int input_samples = input_rate_ / 100; 56 const int output_samples = output_rate_ / 100; 57 const int kResampleIterations = 500000; 58 59 // Source for data to be resampled. 60 ZeroSource resampler_source; 61 62 scoped_ptr<float[]> resampled_destination(new float[output_samples]); 63 scoped_ptr<float[]> source(new float[input_samples]); 64 scoped_ptr<int16_t[]> source_int(new int16_t[input_samples]); 65 scoped_ptr<int16_t[]> destination_int(new int16_t[output_samples]); 66 67 resampler_source.Run(input_samples, source.get()); 68 for (int i = 0; i < input_samples; ++i) { 69 source_int[i] = static_cast<int16_t>(floor(32767 * source[i] + 0.5)); 70 } 71 72 printf("Benchmarking %d iterations of %d Hz -> %d Hz:\n", 73 kResampleIterations, input_rate_, output_rate_); 74 const double io_ratio = input_rate_ / static_cast<double>(output_rate_); 75 SincResampler sinc_resampler(io_ratio, SincResampler::kDefaultRequestSize, 76 &resampler_source); 77 TickTime start = TickTime::Now(); 78 for (int i = 0; i < kResampleIterations; ++i) { 79 sinc_resampler.Resample(output_samples, resampled_destination.get()); 80 } 81 double total_time_sinc_us = (TickTime::Now() - start).Microseconds(); 82 printf("SincResampler took %.2f us per frame.\n", 83 total_time_sinc_us / kResampleIterations); 84 85 PushSincResampler resampler(input_samples, output_samples); 86 start = TickTime::Now(); 87 if (int_format) { 88 for (int i = 0; i < kResampleIterations; ++i) { 89 EXPECT_EQ(output_samples, 90 resampler.Resample(source_int.get(), 91 input_samples, 92 destination_int.get(), 93 output_samples)); 94 } 95 } else { 96 for (int i = 0; i < kResampleIterations; ++i) { 97 EXPECT_EQ(output_samples, 98 resampler.Resample(source.get(), 99 input_samples, 100 resampled_destination.get(), 101 output_samples)); 102 } 103 } 104 double total_time_us = (TickTime::Now() - start).Microseconds(); 105 printf("PushSincResampler took %.2f us per frame; which is a %.1f%% overhead " 106 "on SincResampler.\n\n", total_time_us / kResampleIterations, 107 (total_time_us - total_time_sinc_us) / total_time_sinc_us * 100); 108} 109 110// Disabled because it takes too long to run routinely. Use for performance 111// benchmarking when needed. 112TEST_P(PushSincResamplerTest, DISABLED_BenchmarkInt) { 113 ResampleBenchmarkTest(true); 114} 115 116TEST_P(PushSincResamplerTest, DISABLED_BenchmarkFloat) { 117 ResampleBenchmarkTest(false); 118} 119 120// Tests resampling using a given input and output sample rate. 121void PushSincResamplerTest::ResampleTest(bool int_format) { 122 // Make comparisons using one second of data. 123 static const double kTestDurationSecs = 1; 124 // 10 ms blocks. 125 const int kNumBlocks = kTestDurationSecs * 100; 126 const int input_block_size = input_rate_ / 100; 127 const int output_block_size = output_rate_ / 100; 128 const int input_samples = kTestDurationSecs * input_rate_; 129 const int output_samples = kTestDurationSecs * output_rate_; 130 131 // Nyquist frequency for the input sampling rate. 132 const double input_nyquist_freq = 0.5 * input_rate_; 133 134 // Source for data to be resampled. 135 SinusoidalLinearChirpSource resampler_source( 136 input_rate_, input_samples, input_nyquist_freq, 0); 137 138 PushSincResampler resampler(input_block_size, output_block_size); 139 140 // TODO(dalecurtis): If we switch to AVX/SSE optimization, we'll need to 141 // allocate these on 32-byte boundaries and ensure they're sized % 32 bytes. 142 scoped_ptr<float[]> resampled_destination(new float[output_samples]); 143 scoped_ptr<float[]> pure_destination(new float[output_samples]); 144 scoped_ptr<float[]> source(new float[input_samples]); 145 scoped_ptr<int16_t[]> source_int(new int16_t[input_block_size]); 146 scoped_ptr<int16_t[]> destination_int(new int16_t[output_block_size]); 147 148 // The sinc resampler has an implicit delay of approximately half the kernel 149 // size at the input sample rate. By moving to a push model, this delay 150 // becomes explicit and is managed by zero-stuffing in PushSincResampler. We 151 // deal with it in the test by delaying the "pure" source to match. It must be 152 // checked before the first call to Resample(), because ChunkSize() will 153 // change afterwards. 154 const int output_delay_samples = output_block_size - 155 resampler.get_resampler_for_testing()->ChunkSize(); 156 157 // Generate resampled signal. 158 // With the PushSincResampler, we produce the signal block-by-10ms-block 159 // rather than in a single pass, to exercise how it will be used in WebRTC. 160 resampler_source.Run(input_samples, source.get()); 161 if (int_format) { 162 for (int i = 0; i < kNumBlocks; ++i) { 163 ScaleAndRoundToInt16( 164 &source[i * input_block_size], input_block_size, source_int.get()); 165 EXPECT_EQ(output_block_size, 166 resampler.Resample(source_int.get(), 167 input_block_size, 168 destination_int.get(), 169 output_block_size)); 170 ScaleToFloat(destination_int.get(), 171 output_block_size, 172 &resampled_destination[i * output_block_size]); 173 } 174 } else { 175 for (int i = 0; i < kNumBlocks; ++i) { 176 EXPECT_EQ( 177 output_block_size, 178 resampler.Resample(&source[i * input_block_size], 179 input_block_size, 180 &resampled_destination[i * output_block_size], 181 output_block_size)); 182 } 183 } 184 185 // Generate pure signal. 186 SinusoidalLinearChirpSource pure_source( 187 output_rate_, output_samples, input_nyquist_freq, output_delay_samples); 188 pure_source.Run(output_samples, pure_destination.get()); 189 190 // Range of the Nyquist frequency (0.5 * min(input rate, output_rate)) which 191 // we refer to as low and high. 192 static const double kLowFrequencyNyquistRange = 0.7; 193 static const double kHighFrequencyNyquistRange = 0.9; 194 195 // Calculate Root-Mean-Square-Error and maximum error for the resampling. 196 double sum_of_squares = 0; 197 double low_freq_max_error = 0; 198 double high_freq_max_error = 0; 199 int minimum_rate = std::min(input_rate_, output_rate_); 200 double low_frequency_range = kLowFrequencyNyquistRange * 0.5 * minimum_rate; 201 double high_frequency_range = kHighFrequencyNyquistRange * 0.5 * minimum_rate; 202 203 for (int i = 0; i < output_samples; ++i) { 204 double error = fabs(resampled_destination[i] - pure_destination[i]); 205 206 if (pure_source.Frequency(i) < low_frequency_range) { 207 if (error > low_freq_max_error) 208 low_freq_max_error = error; 209 } else if (pure_source.Frequency(i) < high_frequency_range) { 210 if (error > high_freq_max_error) 211 high_freq_max_error = error; 212 } 213 // TODO(dalecurtis): Sanity check frequencies > kHighFrequencyNyquistRange. 214 215 sum_of_squares += error * error; 216 } 217 218 double rms_error = sqrt(sum_of_squares / output_samples); 219 220 // Convert each error to dbFS. 221 #define DBFS(x) 20 * log10(x) 222 rms_error = DBFS(rms_error); 223 // In order to keep the thresholds in this test identical to SincResamplerTest 224 // we must account for the quantization error introduced by truncating from 225 // float to int. This happens twice (once at input and once at output) and we 226 // allow for the maximum possible error (1 / 32767) for each step. 227 // 228 // The quantization error is insignificant in the RMS calculation so does not 229 // need to be accounted for there. 230 low_freq_max_error = DBFS(low_freq_max_error - 2.0 / 32767); 231 high_freq_max_error = DBFS(high_freq_max_error - 2.0 / 32767); 232 233 EXPECT_LE(rms_error, rms_error_); 234 EXPECT_LE(low_freq_max_error, low_freq_error_); 235 236 // All conversions currently have a high frequency error around -6 dbFS. 237 static const double kHighFrequencyMaxError = -6.02; 238 EXPECT_LE(high_freq_max_error, kHighFrequencyMaxError); 239} 240 241TEST_P(PushSincResamplerTest, ResampleInt) { ResampleTest(true); } 242 243TEST_P(PushSincResamplerTest, ResampleFloat) { ResampleTest(false); } 244 245// Almost all conversions have an RMS error of around -14 dbFS. 246static const double kResamplingRMSError = -14.42; 247 248// Thresholds chosen arbitrarily based on what each resampling reported during 249// testing. All thresholds are in dbFS, http://en.wikipedia.org/wiki/DBFS. 250INSTANTIATE_TEST_CASE_P( 251 PushSincResamplerTest, 252 PushSincResamplerTest, 253 testing::Values( 254 // First run through the rates tested in SincResamplerTest. The 255 // thresholds are identical. 256 // 257 // We don't test rates which fail to provide an integer number of 258 // samples in a 10 ms block (22050 and 11025 Hz). WebRTC doesn't support 259 // these rates in any case (for the same reason). 260 261 // To 44.1kHz 262 std::tr1::make_tuple(8000, 44100, kResamplingRMSError, -62.73), 263 std::tr1::make_tuple(16000, 44100, kResamplingRMSError, -62.54), 264 std::tr1::make_tuple(32000, 44100, kResamplingRMSError, -63.32), 265 std::tr1::make_tuple(44100, 44100, kResamplingRMSError, -73.53), 266 std::tr1::make_tuple(48000, 44100, -15.01, -64.04), 267 std::tr1::make_tuple(96000, 44100, -18.49, -25.51), 268 std::tr1::make_tuple(192000, 44100, -20.50, -13.31), 269 270 // To 48kHz 271 std::tr1::make_tuple(8000, 48000, kResamplingRMSError, -63.43), 272 std::tr1::make_tuple(16000, 48000, kResamplingRMSError, -63.96), 273 std::tr1::make_tuple(32000, 48000, kResamplingRMSError, -64.04), 274 std::tr1::make_tuple(44100, 48000, kResamplingRMSError, -62.63), 275 std::tr1::make_tuple(48000, 48000, kResamplingRMSError, -73.52), 276 std::tr1::make_tuple(96000, 48000, -18.40, -28.44), 277 std::tr1::make_tuple(192000, 48000, -20.43, -14.11), 278 279 // To 96kHz 280 std::tr1::make_tuple(8000, 96000, kResamplingRMSError, -63.19), 281 std::tr1::make_tuple(16000, 96000, kResamplingRMSError, -63.39), 282 std::tr1::make_tuple(32000, 96000, kResamplingRMSError, -63.95), 283 std::tr1::make_tuple(44100, 96000, kResamplingRMSError, -62.63), 284 std::tr1::make_tuple(48000, 96000, kResamplingRMSError, -73.52), 285 std::tr1::make_tuple(96000, 96000, kResamplingRMSError, -73.52), 286 std::tr1::make_tuple(192000, 96000, kResamplingRMSError, -28.41), 287 288 // To 192kHz 289 std::tr1::make_tuple(8000, 192000, kResamplingRMSError, -63.10), 290 std::tr1::make_tuple(16000, 192000, kResamplingRMSError, -63.14), 291 std::tr1::make_tuple(32000, 192000, kResamplingRMSError, -63.38), 292 std::tr1::make_tuple(44100, 192000, kResamplingRMSError, -62.63), 293 std::tr1::make_tuple(48000, 192000, kResamplingRMSError, -73.44), 294 std::tr1::make_tuple(96000, 192000, kResamplingRMSError, -73.52), 295 std::tr1::make_tuple(192000, 192000, kResamplingRMSError, -73.52), 296 297 // Next run through some additional cases interesting for WebRTC. 298 // We skip some extreme downsampled cases (192 -> {8, 16}, 96 -> 8) 299 // because they violate |kHighFrequencyMaxError|, which is not 300 // unexpected. It's very unlikely that we'll see these conversions in 301 // practice anyway. 302 303 // To 8 kHz 304 std::tr1::make_tuple(8000, 8000, kResamplingRMSError, -75.50), 305 std::tr1::make_tuple(16000, 8000, -18.56, -28.79), 306 std::tr1::make_tuple(32000, 8000, -20.36, -14.13), 307 std::tr1::make_tuple(44100, 8000, -21.00, -11.39), 308 std::tr1::make_tuple(48000, 8000, -20.96, -11.04), 309 310 // To 16 kHz 311 std::tr1::make_tuple(8000, 16000, kResamplingRMSError, -70.30), 312 std::tr1::make_tuple(16000, 16000, kResamplingRMSError, -75.51), 313 std::tr1::make_tuple(32000, 16000, -18.48, -28.59), 314 std::tr1::make_tuple(44100, 16000, -19.30, -19.67), 315 std::tr1::make_tuple(48000, 16000, -19.81, -18.11), 316 std::tr1::make_tuple(96000, 16000, -20.95, -10.96), 317 318 // To 32 kHz 319 std::tr1::make_tuple(8000, 32000, kResamplingRMSError, -70.30), 320 std::tr1::make_tuple(16000, 32000, kResamplingRMSError, -75.51), 321 std::tr1::make_tuple(32000, 32000, kResamplingRMSError, -75.51), 322 std::tr1::make_tuple(44100, 32000, -16.44, -51.10), 323 std::tr1::make_tuple(48000, 32000, -16.90, -44.03), 324 std::tr1::make_tuple(96000, 32000, -19.61, -18.04), 325 std::tr1::make_tuple(192000, 32000, -21.02, -10.94))); 326 327} // namespace webrtc 328