1 /* 2 * Copyright (C) 2012 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17 #include <unistd.h> 18 #include <stdio.h> 19 #include <stdlib.h> 20 #include <fcntl.h> 21 #include <string.h> 22 #include <sys/mman.h> 23 #include <sys/stat.h> 24 #include <errno.h> 25 #include <inttypes.h> 26 #include <time.h> 27 #include <math.h> 28 #include <audio_utils/primitives.h> 29 #include <audio_utils/sndfile.h> 30 #include <android-base/macros.h> 31 #include <utils/Vector.h> 32 #include <media/AudioBufferProvider.h> 33 #include <media/AudioResampler.h> 34 35 using namespace android; 36 37 static bool gVerbose = false; 38 39 static int usage(const char* name) { 40 fprintf(stderr,"Usage: %s [-p] [-f] [-F] [-v] [-c channels]" 41 " [-q {dq|lq|mq|hq|vhq|dlq|dmq|dhq}]" 42 " [-i input-sample-rate] [-o output-sample-rate]" 43 " [-O csv] [-P csv] [<input-file>]" 44 " <output-file>\n", name); 45 fprintf(stderr," -p enable profiling\n"); 46 fprintf(stderr," -f enable filter profiling\n"); 47 fprintf(stderr," -F enable floating point -q {dlq|dmq|dhq} only"); 48 fprintf(stderr," -v verbose : log buffer provider calls\n"); 49 fprintf(stderr," -c # channels (1-2 for lq|mq|hq; 1-8 for dlq|dmq|dhq)\n"); 50 fprintf(stderr," -q resampler quality\n"); 51 fprintf(stderr," dq : default quality\n"); 52 fprintf(stderr," lq : low quality\n"); 53 fprintf(stderr," mq : medium quality\n"); 54 fprintf(stderr," hq : high quality\n"); 55 fprintf(stderr," vhq : very high quality\n"); 56 fprintf(stderr," dlq : dynamic low quality\n"); 57 fprintf(stderr," dmq : dynamic medium quality\n"); 58 fprintf(stderr," dhq : dynamic high quality\n"); 59 fprintf(stderr," -i input file sample rate (ignored if input file is specified)\n"); 60 fprintf(stderr," -o output file sample rate\n"); 61 fprintf(stderr," -O # frames output per call to resample() in CSV format\n"); 62 fprintf(stderr," -P # frames provided per call to resample() in CSV format\n"); 63 return -1; 64 } 65 66 // Convert a list of integers in CSV format to a Vector of those values. 67 // Returns the number of elements in the list, or -1 on error. 68 int parseCSV(const char *string, Vector<int>& values) 69 { 70 // pass 1: count the number of values and do syntax check 71 size_t numValues = 0; 72 bool hadDigit = false; 73 for (const char *p = string; ; ) { 74 switch (*p++) { 75 case '0': case '1': case '2': case '3': case '4': 76 case '5': case '6': case '7': case '8': case '9': 77 hadDigit = true; 78 break; 79 case '\0': 80 if (hadDigit) { 81 // pass 2: allocate and initialize vector of values 82 values.resize(++numValues); 83 values.editItemAt(0) = atoi(p = optarg); 84 for (size_t i = 1; i < numValues; ) { 85 if (*p++ == ',') { 86 values.editItemAt(i++) = atoi(p); 87 } 88 } 89 return numValues; 90 } 91 FALLTHROUGH_INTENDED; 92 case ',': 93 if (hadDigit) { 94 hadDigit = false; 95 numValues++; 96 break; 97 } 98 FALLTHROUGH_INTENDED; 99 default: 100 return -1; 101 } 102 } 103 } 104 105 int main(int argc, char* argv[]) { 106 const char* const progname = argv[0]; 107 bool profileResample = false; 108 bool profileFilter = false; 109 bool useFloat = false; 110 int channels = 1; 111 int input_freq = 0; 112 int output_freq = 0; 113 AudioResampler::src_quality quality = AudioResampler::DEFAULT_QUALITY; 114 Vector<int> Ovalues; 115 Vector<int> Pvalues; 116 117 int ch; 118 while ((ch = getopt(argc, argv, "pfFvc:q:i:o:O:P:")) != -1) { 119 switch (ch) { 120 case 'p': 121 profileResample = true; 122 break; 123 case 'f': 124 profileFilter = true; 125 break; 126 case 'F': 127 useFloat = true; 128 break; 129 case 'v': 130 gVerbose = true; 131 break; 132 case 'c': 133 channels = atoi(optarg); 134 break; 135 case 'q': 136 if (!strcmp(optarg, "dq")) 137 quality = AudioResampler::DEFAULT_QUALITY; 138 else if (!strcmp(optarg, "lq")) 139 quality = AudioResampler::LOW_QUALITY; 140 else if (!strcmp(optarg, "mq")) 141 quality = AudioResampler::MED_QUALITY; 142 else if (!strcmp(optarg, "hq")) 143 quality = AudioResampler::HIGH_QUALITY; 144 else if (!strcmp(optarg, "vhq")) 145 quality = AudioResampler::VERY_HIGH_QUALITY; 146 else if (!strcmp(optarg, "dlq")) 147 quality = AudioResampler::DYN_LOW_QUALITY; 148 else if (!strcmp(optarg, "dmq")) 149 quality = AudioResampler::DYN_MED_QUALITY; 150 else if (!strcmp(optarg, "dhq")) 151 quality = AudioResampler::DYN_HIGH_QUALITY; 152 else { 153 usage(progname); 154 return -1; 155 } 156 break; 157 case 'i': 158 input_freq = atoi(optarg); 159 break; 160 case 'o': 161 output_freq = atoi(optarg); 162 break; 163 case 'O': 164 if (parseCSV(optarg, Ovalues) < 0) { 165 fprintf(stderr, "incorrect syntax for -O option\n"); 166 return -1; 167 } 168 break; 169 case 'P': 170 if (parseCSV(optarg, Pvalues) < 0) { 171 fprintf(stderr, "incorrect syntax for -P option\n"); 172 return -1; 173 } 174 break; 175 case '?': 176 default: 177 usage(progname); 178 return -1; 179 } 180 } 181 182 if (channels < 1 183 || channels > (quality < AudioResampler::DYN_LOW_QUALITY ? 2 : 8)) { 184 fprintf(stderr, "invalid number of audio channels %d\n", channels); 185 return -1; 186 } 187 if (useFloat && quality < AudioResampler::DYN_LOW_QUALITY) { 188 fprintf(stderr, "float processing is only possible for dynamic resamplers\n"); 189 return -1; 190 } 191 192 argc -= optind; 193 argv += optind; 194 195 const char* file_in = NULL; 196 const char* file_out = NULL; 197 if (argc == 1) { 198 file_out = argv[0]; 199 } else if (argc == 2) { 200 file_in = argv[0]; 201 file_out = argv[1]; 202 } else { 203 usage(progname); 204 return -1; 205 } 206 207 // ---------------------------------------------------------- 208 209 size_t input_size; 210 void* input_vaddr; 211 if (argc == 2) { 212 SF_INFO info; 213 info.format = 0; 214 SNDFILE *sf = sf_open(file_in, SFM_READ, &info); 215 if (sf == NULL) { 216 perror(file_in); 217 return EXIT_FAILURE; 218 } 219 input_size = info.frames * info.channels * sizeof(short); 220 input_vaddr = malloc(input_size); 221 (void) sf_readf_short(sf, (short *) input_vaddr, info.frames); 222 sf_close(sf); 223 channels = info.channels; 224 input_freq = info.samplerate; 225 } else { 226 // data for testing is exactly (input sampling rate/1000)/2 seconds 227 // so 44.1khz input is 22.05 seconds 228 double k = 1000; // Hz / s 229 double time = (input_freq / 2) / k; 230 size_t input_frames = size_t(input_freq * time); 231 input_size = channels * sizeof(int16_t) * input_frames; 232 input_vaddr = malloc(input_size); 233 int16_t* in = (int16_t*)input_vaddr; 234 for (size_t i=0 ; i<input_frames ; i++) { 235 double t = double(i) / input_freq; 236 double y = sin(M_PI * k * t * t); 237 int16_t yi = floor(y * 32767.0 + 0.5); 238 for (int j = 0; j < channels; j++) { 239 in[i*channels + j] = yi / (1 + j); 240 } 241 } 242 } 243 size_t input_framesize = channels * sizeof(int16_t); 244 size_t input_frames = input_size / input_framesize; 245 246 // For float processing, convert input int16_t to float array 247 if (useFloat) { 248 void *new_vaddr; 249 250 input_framesize = channels * sizeof(float); 251 input_size = input_frames * input_framesize; 252 new_vaddr = malloc(input_size); 253 memcpy_to_float_from_i16(reinterpret_cast<float*>(new_vaddr), 254 reinterpret_cast<int16_t*>(input_vaddr), input_frames * channels); 255 free(input_vaddr); 256 input_vaddr = new_vaddr; 257 } 258 259 // ---------------------------------------------------------- 260 261 class Provider: public AudioBufferProvider { 262 const void* mAddr; // base address 263 const size_t mNumFrames; // total frames 264 const size_t mFrameSize; // size of each frame in bytes 265 size_t mNextFrame; // index of next frame to provide 266 size_t mUnrel; // number of frames not yet released 267 const Vector<int> mPvalues; // number of frames provided per call 268 size_t mNextPidx; // index of next entry in mPvalues to use 269 public: 270 Provider(const void* addr, size_t frames, size_t frameSize, const Vector<int>& Pvalues) 271 : mAddr(addr), 272 mNumFrames(frames), 273 mFrameSize(frameSize), 274 mNextFrame(0), mUnrel(0), mPvalues(Pvalues), mNextPidx(0) { 275 } 276 virtual status_t getNextBuffer(Buffer* buffer) { 277 size_t requestedFrames = buffer->frameCount; 278 if (requestedFrames > mNumFrames - mNextFrame) { 279 buffer->frameCount = mNumFrames - mNextFrame; 280 } 281 if (!mPvalues.isEmpty()) { 282 size_t provided = mPvalues[mNextPidx++]; 283 printf("mPvalue[%zu]=%zu not %zu\n", mNextPidx-1, provided, buffer->frameCount); 284 if (provided < buffer->frameCount) { 285 buffer->frameCount = provided; 286 } 287 if (mNextPidx >= mPvalues.size()) { 288 mNextPidx = 0; 289 } 290 } 291 if (gVerbose) { 292 printf("getNextBuffer() requested %zu frames out of %zu frames available," 293 " and returned %zu frames\n", 294 requestedFrames, (size_t) (mNumFrames - mNextFrame), buffer->frameCount); 295 } 296 mUnrel = buffer->frameCount; 297 if (buffer->frameCount > 0) { 298 buffer->raw = (char *)mAddr + mFrameSize * mNextFrame; 299 return NO_ERROR; 300 } else { 301 buffer->raw = NULL; 302 return NOT_ENOUGH_DATA; 303 } 304 } 305 virtual void releaseBuffer(Buffer* buffer) { 306 if (buffer->frameCount > mUnrel) { 307 fprintf(stderr, "ERROR releaseBuffer() released %zu frames but only %zu available " 308 "to release\n", buffer->frameCount, mUnrel); 309 mNextFrame += mUnrel; 310 mUnrel = 0; 311 } else { 312 if (gVerbose) { 313 printf("releaseBuffer() released %zu frames out of %zu frames available " 314 "to release\n", buffer->frameCount, mUnrel); 315 } 316 mNextFrame += buffer->frameCount; 317 mUnrel -= buffer->frameCount; 318 } 319 buffer->frameCount = 0; 320 buffer->raw = NULL; 321 } 322 void reset() { 323 mNextFrame = 0; 324 } 325 } provider(input_vaddr, input_frames, input_framesize, Pvalues); 326 327 if (gVerbose) { 328 printf("%zu input frames\n", input_frames); 329 } 330 331 audio_format_t format = useFloat ? AUDIO_FORMAT_PCM_FLOAT : AUDIO_FORMAT_PCM_16_BIT; 332 int output_channels = channels > 2 ? channels : 2; // output is at least stereo samples 333 size_t output_framesize = output_channels * (useFloat ? sizeof(float) : sizeof(int32_t)); 334 size_t output_frames = ((int64_t) input_frames * output_freq) / input_freq; 335 size_t output_size = output_frames * output_framesize; 336 337 if (profileFilter) { 338 // Check how fast sample rate changes are that require filter changes. 339 // The delta sample rate changes must indicate a downsampling ratio, 340 // and must be larger than 10% changes. 341 // 342 // On fast devices, filters should be generated between 0.1ms - 1ms. 343 // (single threaded). 344 AudioResampler* resampler = AudioResampler::create(format, channels, 345 8000, quality); 346 int looplimit = 100; 347 timespec start, end; 348 clock_gettime(CLOCK_MONOTONIC, &start); 349 for (int i = 0; i < looplimit; ++i) { 350 resampler->setSampleRate(9000); 351 resampler->setSampleRate(12000); 352 resampler->setSampleRate(20000); 353 resampler->setSampleRate(30000); 354 } 355 clock_gettime(CLOCK_MONOTONIC, &end); 356 int64_t start_ns = start.tv_sec * 1000000000LL + start.tv_nsec; 357 int64_t end_ns = end.tv_sec * 1000000000LL + end.tv_nsec; 358 int64_t time = end_ns - start_ns; 359 printf("%.2f sample rate changes with filter calculation/sec\n", 360 looplimit * 4 / (time / 1e9)); 361 362 // Check how fast sample rate changes are without filter changes. 363 // This should be very fast, probably 0.1us - 1us per sample rate 364 // change. 365 resampler->setSampleRate(1000); 366 looplimit = 1000; 367 clock_gettime(CLOCK_MONOTONIC, &start); 368 for (int i = 0; i < looplimit; ++i) { 369 resampler->setSampleRate(1000+i); 370 } 371 clock_gettime(CLOCK_MONOTONIC, &end); 372 start_ns = start.tv_sec * 1000000000LL + start.tv_nsec; 373 end_ns = end.tv_sec * 1000000000LL + end.tv_nsec; 374 time = end_ns - start_ns; 375 printf("%.2f sample rate changes without filter calculation/sec\n", 376 looplimit / (time / 1e9)); 377 resampler->reset(); 378 delete resampler; 379 } 380 381 void* output_vaddr = malloc(output_size); 382 AudioResampler* resampler = AudioResampler::create(format, channels, 383 output_freq, quality); 384 385 resampler->setSampleRate(input_freq); 386 resampler->setVolume(AudioResampler::UNITY_GAIN_FLOAT, AudioResampler::UNITY_GAIN_FLOAT); 387 388 if (profileResample) { 389 /* 390 * For profiling on mobile devices, upon experimentation 391 * it is better to run a few trials with a shorter loop limit, 392 * and take the minimum time. 393 * 394 * Long tests can cause CPU temperature to build up and thermal throttling 395 * to reduce CPU frequency. 396 * 397 * For frequency checks (index=0, or 1, etc.): 398 * "cat /sys/devices/system/cpu/cpu${index}/cpufreq/scaling_*_freq" 399 * 400 * For temperature checks (index=0, or 1, etc.): 401 * "cat /sys/class/thermal/thermal_zone${index}/temp" 402 * 403 * Another way to avoid thermal throttling is to fix the CPU frequency 404 * at a lower level which prevents excessive temperatures. 405 */ 406 const int trials = 4; 407 const int looplimit = 4; 408 timespec start, end; 409 int64_t time = 0; 410 411 for (int n = 0; n < trials; ++n) { 412 clock_gettime(CLOCK_MONOTONIC, &start); 413 for (int i = 0; i < looplimit; ++i) { 414 resampler->resample((int*) output_vaddr, output_frames, &provider); 415 provider.reset(); // during benchmarking reset only the provider 416 } 417 clock_gettime(CLOCK_MONOTONIC, &end); 418 int64_t start_ns = start.tv_sec * 1000000000LL + start.tv_nsec; 419 int64_t end_ns = end.tv_sec * 1000000000LL + end.tv_nsec; 420 int64_t diff_ns = end_ns - start_ns; 421 if (n == 0 || diff_ns < time) { 422 time = diff_ns; // save the best out of our trials. 423 } 424 } 425 // Mfrms/s is "Millions of output frames per second". 426 printf("quality: %d channels: %d msec: %" PRId64 " Mfrms/s: %.2lf\n", 427 quality, channels, time/1000000, output_frames * looplimit / (time / 1e9) / 1e6); 428 resampler->reset(); 429 430 // TODO fix legacy bug: reset does not clear buffers. 431 // delete and recreate resampler here. 432 delete resampler; 433 resampler = AudioResampler::create(format, channels, 434 output_freq, quality); 435 resampler->setSampleRate(input_freq); 436 resampler->setVolume(AudioResampler::UNITY_GAIN_FLOAT, AudioResampler::UNITY_GAIN_FLOAT); 437 } 438 439 memset(output_vaddr, 0, output_size); 440 if (gVerbose) { 441 printf("resample() %zu output frames\n", output_frames); 442 } 443 if (Ovalues.isEmpty()) { 444 Ovalues.push(output_frames); 445 } 446 for (size_t i = 0, j = 0; i < output_frames; ) { 447 size_t thisFrames = Ovalues[j++]; 448 if (j >= Ovalues.size()) { 449 j = 0; 450 } 451 if (thisFrames == 0 || thisFrames > output_frames - i) { 452 thisFrames = output_frames - i; 453 } 454 resampler->resample((int*) output_vaddr + output_channels*i, thisFrames, &provider); 455 i += thisFrames; 456 } 457 if (gVerbose) { 458 printf("resample() complete\n"); 459 } 460 resampler->reset(); 461 if (gVerbose) { 462 printf("reset() complete\n"); 463 } 464 delete resampler; 465 resampler = NULL; 466 467 // For float processing, convert output format from float to Q4.27, 468 // which is then converted to int16_t for final storage. 469 if (useFloat) { 470 memcpy_to_q4_27_from_float(reinterpret_cast<int32_t*>(output_vaddr), 471 reinterpret_cast<float*>(output_vaddr), output_frames * output_channels); 472 } 473 474 // mono takes left channel only (out of stereo output pair) 475 // stereo and multichannel preserve all channels. 476 int32_t* out = (int32_t*) output_vaddr; 477 int16_t* convert = (int16_t*) malloc(output_frames * channels * sizeof(int16_t)); 478 479 const int volumeShift = 12; // shift requirement for Q4.27 to Q.15 480 // round to half towards zero and saturate at int16 (non-dithered) 481 const int roundVal = (1<<(volumeShift-1)) - 1; // volumePrecision > 0 482 483 for (size_t i = 0; i < output_frames; i++) { 484 for (int j = 0; j < channels; j++) { 485 int32_t s = out[i * output_channels + j] + roundVal; // add offset here 486 if (s < 0) { 487 s = (s + 1) >> volumeShift; // round to 0 488 if (s < -32768) { 489 s = -32768; 490 } 491 } else { 492 s = s >> volumeShift; 493 if (s > 32767) { 494 s = 32767; 495 } 496 } 497 convert[i * channels + j] = int16_t(s); 498 } 499 } 500 501 // write output to disk 502 SF_INFO info; 503 info.frames = 0; 504 info.samplerate = output_freq; 505 info.channels = channels; 506 info.format = SF_FORMAT_WAV | SF_FORMAT_PCM_16; 507 SNDFILE *sf = sf_open(file_out, SFM_WRITE, &info); 508 if (sf == NULL) { 509 perror(file_out); 510 return EXIT_FAILURE; 511 } 512 (void) sf_writef_short(sf, convert, output_frames); 513 sf_close(sf); 514 515 return EXIT_SUCCESS; 516 } 517