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 "common_audio/resampler/push_sinc_resampler.h"
12
13 #include <algorithm>
14 #include <cmath>
15 #include <cstring>
16 #include <memory>
17
18 #include "common_audio/include/audio_util.h"
19 #include "common_audio/resampler/sinusoidal_linear_chirp_source.h"
20 #include "rtc_base/time_utils.h"
21 #include "test/gmock.h"
22 #include "test/gtest.h"
23
24 namespace webrtc {
25 namespace {
26
27 // Almost all conversions have an RMS error of around -14 dbFS.
28 const double kResamplingRMSError = -14.42;
29
30 // Used to convert errors to dbFS.
31 template <typename T>
DBFS(T x)32 T DBFS(T x) {
33 return 20 * std::log10(x);
34 }
35
36 } // namespace
37
38 class PushSincResamplerTest : public ::testing::TestWithParam<
39 ::testing::tuple<int, int, double, double>> {
40 public:
PushSincResamplerTest()41 PushSincResamplerTest()
42 : input_rate_(::testing::get<0>(GetParam())),
43 output_rate_(::testing::get<1>(GetParam())),
44 rms_error_(::testing::get<2>(GetParam())),
45 low_freq_error_(::testing::get<3>(GetParam())) {}
46
~PushSincResamplerTest()47 ~PushSincResamplerTest() override {}
48
49 protected:
50 void ResampleBenchmarkTest(bool int_format);
51 void ResampleTest(bool int_format);
52
53 int input_rate_;
54 int output_rate_;
55 double rms_error_;
56 double low_freq_error_;
57 };
58
59 class ZeroSource : public SincResamplerCallback {
60 public:
Run(size_t frames,float * destination)61 void Run(size_t frames, float* destination) override {
62 std::memset(destination, 0, sizeof(float) * frames);
63 }
64 };
65
ResampleBenchmarkTest(bool int_format)66 void PushSincResamplerTest::ResampleBenchmarkTest(bool int_format) {
67 const size_t input_samples = static_cast<size_t>(input_rate_ / 100);
68 const size_t output_samples = static_cast<size_t>(output_rate_ / 100);
69 const int kResampleIterations = 500000;
70
71 // Source for data to be resampled.
72 ZeroSource resampler_source;
73
74 std::unique_ptr<float[]> resampled_destination(new float[output_samples]);
75 std::unique_ptr<float[]> source(new float[input_samples]);
76 std::unique_ptr<int16_t[]> source_int(new int16_t[input_samples]);
77 std::unique_ptr<int16_t[]> destination_int(new int16_t[output_samples]);
78
79 resampler_source.Run(input_samples, source.get());
80 for (size_t i = 0; i < input_samples; ++i) {
81 source_int[i] = static_cast<int16_t>(floor(32767 * source[i] + 0.5));
82 }
83
84 printf("Benchmarking %d iterations of %d Hz -> %d Hz:\n", kResampleIterations,
85 input_rate_, output_rate_);
86 const double io_ratio = input_rate_ / static_cast<double>(output_rate_);
87 SincResampler sinc_resampler(io_ratio, SincResampler::kDefaultRequestSize,
88 &resampler_source);
89 int64_t start = rtc::TimeNanos();
90 for (int i = 0; i < kResampleIterations; ++i) {
91 sinc_resampler.Resample(output_samples, resampled_destination.get());
92 }
93 double total_time_sinc_us =
94 (rtc::TimeNanos() - start) / rtc::kNumNanosecsPerMicrosec;
95 printf("SincResampler took %.2f us per frame.\n",
96 total_time_sinc_us / kResampleIterations);
97
98 PushSincResampler resampler(input_samples, output_samples);
99 start = rtc::TimeNanos();
100 if (int_format) {
101 for (int i = 0; i < kResampleIterations; ++i) {
102 EXPECT_EQ(output_samples,
103 resampler.Resample(source_int.get(), input_samples,
104 destination_int.get(), output_samples));
105 }
106 } else {
107 for (int i = 0; i < kResampleIterations; ++i) {
108 EXPECT_EQ(output_samples, resampler.Resample(source.get(), input_samples,
109 resampled_destination.get(),
110 output_samples));
111 }
112 }
113 double total_time_us =
114 (rtc::TimeNanos() - start) / rtc::kNumNanosecsPerMicrosec;
115 printf(
116 "PushSincResampler took %.2f us per frame; which is a %.1f%% overhead "
117 "on SincResampler.\n\n",
118 total_time_us / kResampleIterations,
119 (total_time_us - total_time_sinc_us) / total_time_sinc_us * 100);
120 }
121
122 // Disabled because it takes too long to run routinely. Use for performance
123 // benchmarking when needed.
TEST_P(PushSincResamplerTest,DISABLED_BenchmarkInt)124 TEST_P(PushSincResamplerTest, DISABLED_BenchmarkInt) {
125 ResampleBenchmarkTest(true);
126 }
127
TEST_P(PushSincResamplerTest,DISABLED_BenchmarkFloat)128 TEST_P(PushSincResamplerTest, DISABLED_BenchmarkFloat) {
129 ResampleBenchmarkTest(false);
130 }
131
132 // Tests resampling using a given input and output sample rate.
ResampleTest(bool int_format)133 void PushSincResamplerTest::ResampleTest(bool int_format) {
134 // Make comparisons using one second of data.
135 static const double kTestDurationSecs = 1;
136 // 10 ms blocks.
137 const size_t kNumBlocks = static_cast<size_t>(kTestDurationSecs * 100);
138 const size_t input_block_size = static_cast<size_t>(input_rate_ / 100);
139 const size_t output_block_size = static_cast<size_t>(output_rate_ / 100);
140 const size_t input_samples =
141 static_cast<size_t>(kTestDurationSecs * input_rate_);
142 const size_t output_samples =
143 static_cast<size_t>(kTestDurationSecs * output_rate_);
144
145 // Nyquist frequency for the input sampling rate.
146 const double input_nyquist_freq = 0.5 * input_rate_;
147
148 // Source for data to be resampled.
149 SinusoidalLinearChirpSource resampler_source(input_rate_, input_samples,
150 input_nyquist_freq, 0);
151
152 PushSincResampler resampler(input_block_size, output_block_size);
153
154 // TODO(dalecurtis): If we switch to AVX/SSE optimization, we'll need to
155 // allocate these on 32-byte boundaries and ensure they're sized % 32 bytes.
156 std::unique_ptr<float[]> resampled_destination(new float[output_samples]);
157 std::unique_ptr<float[]> pure_destination(new float[output_samples]);
158 std::unique_ptr<float[]> source(new float[input_samples]);
159 std::unique_ptr<int16_t[]> source_int(new int16_t[input_block_size]);
160 std::unique_ptr<int16_t[]> destination_int(new int16_t[output_block_size]);
161
162 // The sinc resampler has an implicit delay of approximately half the kernel
163 // size at the input sample rate. By moving to a push model, this delay
164 // becomes explicit and is managed by zero-stuffing in PushSincResampler. We
165 // deal with it in the test by delaying the "pure" source to match. It must be
166 // checked before the first call to Resample(), because ChunkSize() will
167 // change afterwards.
168 const size_t output_delay_samples =
169 output_block_size - resampler.get_resampler_for_testing()->ChunkSize();
170
171 // Generate resampled signal.
172 // With the PushSincResampler, we produce the signal block-by-10ms-block
173 // rather than in a single pass, to exercise how it will be used in WebRTC.
174 resampler_source.Run(input_samples, source.get());
175 if (int_format) {
176 for (size_t i = 0; i < kNumBlocks; ++i) {
177 FloatToS16(&source[i * input_block_size], input_block_size,
178 source_int.get());
179 EXPECT_EQ(output_block_size,
180 resampler.Resample(source_int.get(), input_block_size,
181 destination_int.get(), output_block_size));
182 S16ToFloat(destination_int.get(), output_block_size,
183 &resampled_destination[i * output_block_size]);
184 }
185 } else {
186 for (size_t i = 0; i < kNumBlocks; ++i) {
187 EXPECT_EQ(
188 output_block_size,
189 resampler.Resample(&source[i * input_block_size], input_block_size,
190 &resampled_destination[i * output_block_size],
191 output_block_size));
192 }
193 }
194
195 // Generate pure signal.
196 SinusoidalLinearChirpSource pure_source(
197 output_rate_, output_samples, input_nyquist_freq, output_delay_samples);
198 pure_source.Run(output_samples, pure_destination.get());
199
200 // Range of the Nyquist frequency (0.5 * min(input rate, output_rate)) which
201 // we refer to as low and high.
202 static const double kLowFrequencyNyquistRange = 0.7;
203 static const double kHighFrequencyNyquistRange = 0.9;
204
205 // Calculate Root-Mean-Square-Error and maximum error for the resampling.
206 double sum_of_squares = 0;
207 double low_freq_max_error = 0;
208 double high_freq_max_error = 0;
209 int minimum_rate = std::min(input_rate_, output_rate_);
210 double low_frequency_range = kLowFrequencyNyquistRange * 0.5 * minimum_rate;
211 double high_frequency_range = kHighFrequencyNyquistRange * 0.5 * minimum_rate;
212
213 for (size_t i = 0; i < output_samples; ++i) {
214 double error = fabs(resampled_destination[i] - pure_destination[i]);
215
216 if (pure_source.Frequency(i) < low_frequency_range) {
217 if (error > low_freq_max_error)
218 low_freq_max_error = error;
219 } else if (pure_source.Frequency(i) < high_frequency_range) {
220 if (error > high_freq_max_error)
221 high_freq_max_error = error;
222 }
223 // TODO(dalecurtis): Sanity check frequencies > kHighFrequencyNyquistRange.
224
225 sum_of_squares += error * error;
226 }
227
228 double rms_error = sqrt(sum_of_squares / output_samples);
229
230 rms_error = DBFS(rms_error);
231 // In order to keep the thresholds in this test identical to SincResamplerTest
232 // we must account for the quantization error introduced by truncating from
233 // float to int. This happens twice (once at input and once at output) and we
234 // allow for the maximum possible error (1 / 32767) for each step.
235 //
236 // The quantization error is insignificant in the RMS calculation so does not
237 // need to be accounted for there.
238 low_freq_max_error = DBFS(low_freq_max_error - 2.0 / 32767);
239 high_freq_max_error = DBFS(high_freq_max_error - 2.0 / 32767);
240
241 EXPECT_LE(rms_error, rms_error_);
242 EXPECT_LE(low_freq_max_error, low_freq_error_);
243
244 // All conversions currently have a high frequency error around -6 dbFS.
245 static const double kHighFrequencyMaxError = -6.02;
246 EXPECT_LE(high_freq_max_error, kHighFrequencyMaxError);
247 }
248
TEST_P(PushSincResamplerTest,ResampleInt)249 TEST_P(PushSincResamplerTest, ResampleInt) {
250 ResampleTest(true);
251 }
252
TEST_P(PushSincResamplerTest,ResampleFloat)253 TEST_P(PushSincResamplerTest, ResampleFloat) {
254 ResampleTest(false);
255 }
256
257 // Thresholds chosen arbitrarily based on what each resampling reported during
258 // testing. All thresholds are in dbFS, http://en.wikipedia.org/wiki/DBFS.
259 INSTANTIATE_TEST_SUITE_P(
260 PushSincResamplerTest,
261 PushSincResamplerTest,
262 ::testing::Values(
263 // First run through the rates tested in SincResamplerTest. The
264 // thresholds are identical.
265 //
266 // We don't test rates which fail to provide an integer number of
267 // samples in a 10 ms block (22050 and 11025 Hz). WebRTC doesn't support
268 // these rates in any case (for the same reason).
269
270 // To 44.1kHz
271 ::testing::make_tuple(8000, 44100, kResamplingRMSError, -62.73),
272 ::testing::make_tuple(16000, 44100, kResamplingRMSError, -62.54),
273 ::testing::make_tuple(32000, 44100, kResamplingRMSError, -63.32),
274 ::testing::make_tuple(44100, 44100, kResamplingRMSError, -73.53),
275 ::testing::make_tuple(48000, 44100, -15.01, -64.04),
276 ::testing::make_tuple(96000, 44100, -18.49, -25.51),
277 ::testing::make_tuple(192000, 44100, -20.50, -13.31),
278
279 // To 48kHz
280 ::testing::make_tuple(8000, 48000, kResamplingRMSError, -63.43),
281 ::testing::make_tuple(16000, 48000, kResamplingRMSError, -63.96),
282 ::testing::make_tuple(32000, 48000, kResamplingRMSError, -64.04),
283 ::testing::make_tuple(44100, 48000, kResamplingRMSError, -62.63),
284 ::testing::make_tuple(48000, 48000, kResamplingRMSError, -73.52),
285 ::testing::make_tuple(96000, 48000, -18.40, -28.44),
286 ::testing::make_tuple(192000, 48000, -20.43, -14.11),
287
288 // To 96kHz
289 ::testing::make_tuple(8000, 96000, kResamplingRMSError, -63.19),
290 ::testing::make_tuple(16000, 96000, kResamplingRMSError, -63.39),
291 ::testing::make_tuple(32000, 96000, kResamplingRMSError, -63.95),
292 ::testing::make_tuple(44100, 96000, kResamplingRMSError, -62.63),
293 ::testing::make_tuple(48000, 96000, kResamplingRMSError, -73.52),
294 ::testing::make_tuple(96000, 96000, kResamplingRMSError, -73.52),
295 ::testing::make_tuple(192000, 96000, kResamplingRMSError, -28.41),
296
297 // To 192kHz
298 ::testing::make_tuple(8000, 192000, kResamplingRMSError, -63.10),
299 ::testing::make_tuple(16000, 192000, kResamplingRMSError, -63.14),
300 ::testing::make_tuple(32000, 192000, kResamplingRMSError, -63.38),
301 ::testing::make_tuple(44100, 192000, kResamplingRMSError, -62.63),
302 ::testing::make_tuple(48000, 192000, kResamplingRMSError, -73.44),
303 ::testing::make_tuple(96000, 192000, kResamplingRMSError, -73.52),
304 ::testing::make_tuple(192000, 192000, kResamplingRMSError, -73.52),
305
306 // Next run through some additional cases interesting for WebRTC.
307 // We skip some extreme downsampled cases (192 -> {8, 16}, 96 -> 8)
308 // because they violate |kHighFrequencyMaxError|, which is not
309 // unexpected. It's very unlikely that we'll see these conversions in
310 // practice anyway.
311
312 // To 8 kHz
313 ::testing::make_tuple(8000, 8000, kResamplingRMSError, -75.50),
314 ::testing::make_tuple(16000, 8000, -18.56, -28.79),
315 ::testing::make_tuple(32000, 8000, -20.36, -14.13),
316 ::testing::make_tuple(44100, 8000, -21.00, -11.39),
317 ::testing::make_tuple(48000, 8000, -20.96, -11.04),
318
319 // To 16 kHz
320 ::testing::make_tuple(8000, 16000, kResamplingRMSError, -70.30),
321 ::testing::make_tuple(16000, 16000, kResamplingRMSError, -75.51),
322 ::testing::make_tuple(32000, 16000, -18.48, -28.59),
323 ::testing::make_tuple(44100, 16000, -19.30, -19.67),
324 ::testing::make_tuple(48000, 16000, -19.81, -18.11),
325 ::testing::make_tuple(96000, 16000, -20.95, -10.9596),
326
327 // To 32 kHz
328 ::testing::make_tuple(8000, 32000, kResamplingRMSError, -70.30),
329 ::testing::make_tuple(16000, 32000, kResamplingRMSError, -75.51),
330 ::testing::make_tuple(32000, 32000, kResamplingRMSError, -75.51),
331 ::testing::make_tuple(44100, 32000, -16.44, -51.0349),
332 ::testing::make_tuple(48000, 32000, -16.90, -43.9967),
333 ::testing::make_tuple(96000, 32000, -19.61, -18.04),
334 ::testing::make_tuple(192000, 32000, -21.02, -10.94)));
335
336 } // namespace webrtc
337