2 * ReplayGainAnalysis - analyzes input samples and give the recommended dB change
3 * Copyright (C) 2001 David Robinson and Glen Sawyer
4 * Improvements and optimizations added by Frank Klemm, and by Marcel Muller
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 * concept and filter values by David Robinson (David@Robinson.org)
21 * -- blame him if you think the idea is flawed
22 * original coding by Glen Sawyer (mp3gain@hotmail.com)
23 * -- blame him if you think this runs too slowly, or the coding is otherwise flawed
25 * lots of code improvements by Frank Klemm ( http://www.uni-jena.de/~pfk/mpp/ )
26 * -- credit him for all the _good_ programming ;)
29 * For an explanation of the concepts and the basic algorithms involved, go to:
30 * http://www.replaygain.org/
34 * Here's the deal. Call
36 * InitGainAnalysis ( long samplefreq );
38 * to initialize everything. Call
40 * AnalyzeSamples ( const Float_t* left_samples,
41 * const Float_t* right_samples,
45 * as many times as you want, with as many or as few samples as you want.
46 * If mono, pass the sample buffer in through left_samples, leave
47 * right_samples NULL, and make sure num_channels = 1.
51 * will return the recommended dB level change for all samples analyzed
52 * SINCE THE LAST TIME you called GetTitleGain() OR InitGainAnalysis().
56 * will return the recommended dB level change for all samples analyzed
57 * since InitGainAnalysis() was called and finalized with GetTitleGain().
59 * Pseudo-code to process an album:
61 * Float_t l_samples [4096];
62 * Float_t r_samples [4096];
64 * unsigned int num_songs;
67 * InitGainAnalysis ( 44100 );
68 * for ( i = 1; i <= num_songs; i++ ) {
69 * while ( ( num_samples = getSongSamples ( song[i], left_samples, right_samples ) ) > 0 )
70 * AnalyzeSamples ( left_samples, right_samples, num_samples, 2 );
71 * fprintf ("Recommended dB change for song %2d: %+6.2f dB\n", i, GetTitleGain() );
73 * fprintf ("Recommended dB change for whole album: %+6.2f dB\n", GetAlbumGain() );
77 * So here's the main source of potential code confusion:
79 * The filters applied to the incoming samples are IIR filters,
80 * meaning they rely on up to <filter order> number of previous samples
81 * AND up to <filter order> number of previous filtered samples.
83 * I set up the AnalyzeSamples routine to minimize memory usage and interface
84 * complexity. The speed isn't compromised too much (I don't think), but the
85 * internal complexity is higher than it should be for such a relatively
88 * Optimization/clarity suggestions are welcome.
101 #include "gain_analysis.h"
103 /* for each filter: */
104 /* [0] 48 kHz, [1] 44.1 kHz, [2] 32 kHz, [3] 24 kHz, [4] 22050 Hz, [5] 16 kHz, [6] 12 kHz, [7] is 11025 Hz, [8] 8 kHz */
107 #pragma warning ( disable : 4305 )
110 /*lint -save -e736 loss of precision */
111 static const Float_t ABYule[9][2 * YULE_ORDER + 1] = {
112 {0.03857599435200, -3.84664617118067, -0.02160367184185, 7.81501653005538, -0.00123395316851,
113 -11.34170355132042, -0.00009291677959, 13.05504219327545, -0.01655260341619,
114 -12.28759895145294, 0.02161526843274, 9.48293806319790, -0.02074045215285, -5.87257861775999,
115 0.00594298065125, 2.75465861874613, 0.00306428023191, -0.86984376593551, 0.00012025322027,
116 0.13919314567432, 0.00288463683916},
117 {0.05418656406430, -3.47845948550071, -0.02911007808948, 6.36317777566148, -0.00848709379851,
118 -8.54751527471874, -0.00851165645469, 9.47693607801280, -0.00834990904936, -8.81498681370155,
119 0.02245293253339, 6.85401540936998, -0.02596338512915, -4.39470996079559, 0.01624864962975,
120 2.19611684890774, -0.00240879051584, -0.75104302451432, 0.00674613682247, 0.13149317958808,
122 {0.15457299681924, -2.37898834973084, -0.09331049056315, 2.84868151156327, -0.06247880153653,
123 -2.64577170229825, 0.02163541888798, 2.23697657451713, -0.05588393329856, -1.67148153367602,
124 0.04781476674921, 1.00595954808547, 0.00222312597743, -0.45953458054983, 0.03174092540049,
125 0.16378164858596, -0.01390589421898, -0.05032077717131, 0.00651420667831, 0.02347897407020,
127 {0.30296907319327, -1.61273165137247, -0.22613988682123, 1.07977492259970, -0.08587323730772,
128 -0.25656257754070, 0.03282930172664, -0.16276719120440, -0.00915702933434, -0.22638893773906,
129 -0.02364141202522, 0.39120800788284, -0.00584456039913, -0.22138138954925, 0.06276101321749,
130 0.04500235387352, -0.00000828086748, 0.02005851806501, 0.00205861885564, 0.00302439095741,
132 {0.33642304856132, -1.49858979367799, -0.25572241425570, 0.87350271418188, -0.11828570177555,
133 0.12205022308084, 0.11921148675203, -0.80774944671438, -0.07834489609479, 0.47854794562326,
134 -0.00469977914380, -0.12453458140019, -0.00589500224440, -0.04067510197014, 0.05724228140351,
135 0.08333755284107, 0.00832043980773, -0.04237348025746, -0.01635381384540, 0.02977207319925,
137 {0.44915256608450, -0.62820619233671, -0.14351757464547, 0.29661783706366, -0.22784394429749,
138 -0.37256372942400, -0.01419140100551, 0.00213767857124, 0.04078262797139, -0.42029820170918,
139 -0.12398163381748, 0.22199650564824, 0.04097565135648, 0.00613424350682, 0.10478503600251,
140 0.06747620744683, -0.01863887810927, 0.05784820375801, -0.03193428438915, 0.03222754072173,
142 {0.56619470757641, -1.04800335126349, -0.75464456939302, 0.29156311971249, 0.16242137742230,
143 -0.26806001042947, 0.16744243493672, 0.00819999645858, -0.18901604199609, 0.45054734505008,
144 0.30931782841830, -0.33032403314006, -0.27562961986224, 0.06739368333110, 0.00647310677246,
145 -0.04784254229033, 0.08647503780351, 0.01639907836189, -0.03788984554840, 0.01807364323573,
147 {0.58100494960553, -0.51035327095184, -0.53174909058578, -0.31863563325245, -0.14289799034253,
148 -0.20256413484477, 0.17520704835522, 0.14728154134330, 0.02377945217615, 0.38952639978999,
149 0.15558449135573, -0.23313271880868, -0.25344790059353, -0.05246019024463, 0.01628462406333,
150 -0.02505961724053, 0.06920467763959, 0.02442357316099, -0.03721611395801, 0.01818801111503,
152 {0.53648789255105, -0.25049871956020, -0.42163034350696, -0.43193942311114, -0.00275953611929,
153 -0.03424681017675, 0.04267842219415, -0.04678328784242, -0.10214864179676, 0.26408300200955,
154 0.14590772289388, 0.15113130533216, -0.02459864859345, -0.17556493366449, -0.11202315195388,
155 -0.18823009262115, -0.04060034127000, 0.05477720428674, 0.04788665548180, 0.04704409688120,
159 static const Float_t ABButter[9][2 * BUTTER_ORDER + 1] = {
160 {0.98621192462708, -1.97223372919527, -1.97242384925416, 0.97261396931306, 0.98621192462708},
161 {0.98500175787242, -1.96977855582618, -1.97000351574484, 0.97022847566350, 0.98500175787242},
162 {0.97938932735214, -1.95835380975398, -1.95877865470428, 0.95920349965459, 0.97938932735214},
163 {0.97531843204928, -1.95002759149878, -1.95063686409857, 0.95124613669835, 0.97531843204928},
164 {0.97316523498161, -1.94561023566527, -1.94633046996323, 0.94705070426118, 0.97316523498161},
165 {0.96454515552826, -1.92783286977036, -1.92909031105652, 0.93034775234268, 0.96454515552826},
166 {0.96009142950541, -1.91858953033784, -1.92018285901082, 0.92177618768381, 0.96009142950541},
167 {0.95856916599601, -1.91542108074780, -1.91713833199203, 0.91885558323625, 0.95856916599601},
168 {0.94597685600279, -1.88903307939452, -1.89195371200558, 0.89487434461664, 0.94597685600279}
174 #pragma warning ( default : 4305 )
177 /* When calling this procedure, make sure that ip[-order] and op[-order] point to real data! */
180 filterYule(const Float_t * input, Float_t * output, size_t nSamples, const Float_t * const kernel)
182 /*register double y; */
185 *output = 1e-10 /* 1e-10 is a hack to avoid slowdown because of denormals */
186 + input[0] * kernel[0]
187 - output[-1] * kernel[1]
188 + input[-1] * kernel[2]
189 - output[-2] * kernel[3]
190 + input[-2] * kernel[4]
191 - output[-3] * kernel[5]
192 + input[-3] * kernel[6]
193 - output[-4] * kernel[7]
194 + input[-4] * kernel[8]
195 - output[-5] * kernel[9]
196 + input[-5] * kernel[10]
197 - output[-6] * kernel[11]
198 + input[-6] * kernel[12]
199 - output[-7] * kernel[13]
200 + input[-7] * kernel[14]
201 - output[-8] * kernel[15]
202 + input[-8] * kernel[16]
203 - output[-9] * kernel[17]
204 + input[-9] * kernel[18]
205 - output[-10] * kernel[19]
206 + input[-10] * kernel[20];
209 /* *output++ = (Float_t)y; */
214 filterButter(const Float_t * input, Float_t * output, size_t nSamples, const Float_t * const kernel)
215 { /*register double y; */
218 *output = input[0] * kernel[0]
219 - output[-1] * kernel[1]
220 + input[-1] * kernel[2]
221 - output[-2] * kernel[3]
222 + input[-2] * kernel[4];
225 /* *output++ = (Float_t)y; */
231 static int ResetSampleFrequency(replaygain_t * rgData, long samplefreq);
233 /* returns a INIT_GAIN_ANALYSIS_OK if successful, INIT_GAIN_ANALYSIS_ERROR if not */
236 ResetSampleFrequency(replaygain_t * rgData, long samplefreq)
240 /* zero out initial values */
241 for (i = 0; i < MAX_ORDER; i++)
242 rgData->linprebuf[i] = rgData->lstepbuf[i]
244 = rgData->rinprebuf[i]
245 = rgData->rstepbuf[i]
246 = rgData->routbuf[i] = 0.;
248 switch ((int) (samplefreq)) {
250 rgData->freqindex = 0;
253 rgData->freqindex = 1;
256 rgData->freqindex = 2;
259 rgData->freqindex = 3;
262 rgData->freqindex = 4;
265 rgData->freqindex = 5;
268 rgData->freqindex = 6;
271 rgData->freqindex = 7;
274 rgData->freqindex = 8;
277 return INIT_GAIN_ANALYSIS_ERROR;
280 rgData->sampleWindow =
281 (samplefreq * RMS_WINDOW_TIME_NUMERATOR + RMS_WINDOW_TIME_DENOMINATOR -
282 1) / RMS_WINDOW_TIME_DENOMINATOR;
288 memset(rgData->A, 0, sizeof(rgData->A));
290 return INIT_GAIN_ANALYSIS_OK;
294 InitGainAnalysis(replaygain_t * rgData, long samplefreq)
296 if (ResetSampleFrequency(rgData, samplefreq) != INIT_GAIN_ANALYSIS_OK) {
297 return INIT_GAIN_ANALYSIS_ERROR;
300 rgData->linpre = rgData->linprebuf + MAX_ORDER;
301 rgData->rinpre = rgData->rinprebuf + MAX_ORDER;
302 rgData->lstep = rgData->lstepbuf + MAX_ORDER;
303 rgData->rstep = rgData->rstepbuf + MAX_ORDER;
304 rgData->lout = rgData->loutbuf + MAX_ORDER;
305 rgData->rout = rgData->routbuf + MAX_ORDER;
307 memset(rgData->B, 0, sizeof(rgData->B));
309 return INIT_GAIN_ANALYSIS_OK;
312 /* returns GAIN_ANALYSIS_OK if successful, GAIN_ANALYSIS_ERROR if not */
321 AnalyzeSamples(replaygain_t * rgData, const Float_t * left_samples, const Float_t * right_samples,
322 size_t num_samples, int num_channels)
324 const Float_t *curleft;
325 const Float_t *curright;
331 if (num_samples == 0)
332 return GAIN_ANALYSIS_OK;
335 batchsamples = (long) num_samples;
337 switch (num_channels) {
339 right_samples = left_samples;
344 return GAIN_ANALYSIS_ERROR;
347 if (num_samples < MAX_ORDER) {
348 memcpy(rgData->linprebuf + MAX_ORDER, left_samples, num_samples * sizeof(Float_t));
349 memcpy(rgData->rinprebuf + MAX_ORDER, right_samples, num_samples * sizeof(Float_t));
352 memcpy(rgData->linprebuf + MAX_ORDER, left_samples, MAX_ORDER * sizeof(Float_t));
353 memcpy(rgData->rinprebuf + MAX_ORDER, right_samples, MAX_ORDER * sizeof(Float_t));
356 while (batchsamples > 0) {
357 cursamples = batchsamples > rgData->sampleWindow - rgData->totsamp ?
358 rgData->sampleWindow - rgData->totsamp : batchsamples;
359 if (cursamplepos < MAX_ORDER) {
360 curleft = rgData->linpre + cursamplepos;
361 curright = rgData->rinpre + cursamplepos;
362 if (cursamples > MAX_ORDER - cursamplepos)
363 cursamples = MAX_ORDER - cursamplepos;
366 curleft = left_samples + cursamplepos;
367 curright = right_samples + cursamplepos;
370 YULE_FILTER(curleft, rgData->lstep + rgData->totsamp, cursamples,
371 ABYule[rgData->freqindex]);
372 YULE_FILTER(curright, rgData->rstep + rgData->totsamp, cursamples,
373 ABYule[rgData->freqindex]);
375 BUTTER_FILTER(rgData->lstep + rgData->totsamp, rgData->lout + rgData->totsamp, cursamples,
376 ABButter[rgData->freqindex]);
377 BUTTER_FILTER(rgData->rstep + rgData->totsamp, rgData->rout + rgData->totsamp, cursamples,
378 ABButter[rgData->freqindex]);
380 curleft = rgData->lout + rgData->totsamp; /* Get the squared values */
381 curright = rgData->rout + rgData->totsamp;
385 rgData->lsum += fsqr(*curleft++);
386 rgData->rsum += fsqr(*curright++);
390 rgData->lsum += fsqr(curleft[0])
399 rgData->rsum += fsqr(curright[0])
410 batchsamples -= cursamples;
411 cursamplepos += cursamples;
412 rgData->totsamp += cursamples;
413 if (rgData->totsamp == rgData->sampleWindow) { /* Get the Root Mean Square (RMS) for this set of samples */
415 STEPS_per_dB * 10. * log10((rgData->lsum + rgData->rsum) / rgData->totsamp * 0.5 +
417 size_t ival = (val <= 0) ? 0 : (size_t) val;
418 if (ival >= sizeof(rgData->A) / sizeof(*(rgData->A)))
419 ival = sizeof(rgData->A) / sizeof(*(rgData->A)) - 1;
421 rgData->lsum = rgData->rsum = 0.;
422 memmove(rgData->loutbuf, rgData->loutbuf + rgData->totsamp,
423 MAX_ORDER * sizeof(Float_t));
424 memmove(rgData->routbuf, rgData->routbuf + rgData->totsamp,
425 MAX_ORDER * sizeof(Float_t));
426 memmove(rgData->lstepbuf, rgData->lstepbuf + rgData->totsamp,
427 MAX_ORDER * sizeof(Float_t));
428 memmove(rgData->rstepbuf, rgData->rstepbuf + rgData->totsamp,
429 MAX_ORDER * sizeof(Float_t));
432 if (rgData->totsamp > rgData->sampleWindow) /* somehow I really screwed up: Error in programming! Contact author about totsamp > sampleWindow */
433 return GAIN_ANALYSIS_ERROR;
435 if (num_samples < MAX_ORDER) {
436 memmove(rgData->linprebuf, rgData->linprebuf + num_samples,
437 (MAX_ORDER - num_samples) * sizeof(Float_t));
438 memmove(rgData->rinprebuf, rgData->rinprebuf + num_samples,
439 (MAX_ORDER - num_samples) * sizeof(Float_t));
440 memcpy(rgData->linprebuf + MAX_ORDER - num_samples, left_samples,
441 num_samples * sizeof(Float_t));
442 memcpy(rgData->rinprebuf + MAX_ORDER - num_samples, right_samples,
443 num_samples * sizeof(Float_t));
446 memcpy(rgData->linprebuf, left_samples + num_samples - MAX_ORDER,
447 MAX_ORDER * sizeof(Float_t));
448 memcpy(rgData->rinprebuf, right_samples + num_samples - MAX_ORDER,
449 MAX_ORDER * sizeof(Float_t));
452 return GAIN_ANALYSIS_OK;
457 analyzeResult(uint32_t const *Array, size_t len)
465 for (i = 0; i < len; i++)
468 return GAIN_NOT_ENOUGH_SAMPLES;
470 upper = (uint32_t) ceil(elems * (1. - RMS_PERCENTILE));
472 for (i = len; i-- > 0;) {
479 return (Float_t) ((Float_t) PINK_REF - (Float_t) i / (Float_t) STEPS_per_dB);
484 GetTitleGain(replaygain_t * rgData)
489 retval = analyzeResult(rgData->A, sizeof(rgData->A) / sizeof(*(rgData->A)));
491 for (i = 0; i < sizeof(rgData->A) / sizeof(*(rgData->A)); i++) {
492 rgData->B[i] += rgData->A[i];
496 for (i = 0; i < MAX_ORDER; i++)
497 rgData->linprebuf[i] = rgData->lstepbuf[i]
499 = rgData->rinprebuf[i]
500 = rgData->rstepbuf[i]
501 = rgData->routbuf[i] = 0.f;
504 rgData->lsum = rgData->rsum = 0.;
509 static Float_t GetAlbumGain(replaygain_t const* rgData);
512 GetAlbumGain(replaygain_t const* rgData)
514 return analyzeResult(rgData->B, sizeof(rgData->B) / sizeof(*(rgData->B)));
518 /* end of gain_analysis.c */