1 /*
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 M�ller
5 *
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.
10 *
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.
15 *
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
19 *
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
24 *
25 * lots of code improvements by Frank Klemm ( http://www.uni-jena.de/~pfk/mpp/ )
26 * -- credit him for all the _good_ programming ;)
27 *
28 *
29 * For an explanation of the concepts and the basic algorithms involved, go to:
30 * http://www.replaygain.org/
31 */
32
33 /*
34 * Here's the deal. Call
35 *
36 * InitGainAnalysis ( long samplefreq );
37 *
38 * to initialize everything. Call
39 *
40 * AnalyzeSamples ( const Float_t* left_samples,
41 * const Float_t* right_samples,
42 * size_t num_samples,
43 * int num_channels );
44 *
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.
48 *
49 * GetTitleGain()
50 *
51 * will return the recommended dB level change for all samples analyzed
52 * SINCE THE LAST TIME you called GetTitleGain() OR InitGainAnalysis().
53 *
54 * GetAlbumGain()
55 *
56 * will return the recommended dB level change for all samples analyzed
57 * since InitGainAnalysis() was called and finalized with GetTitleGain().
58 *
59 * Pseudo-code to process an album:
60 *
61 * Float_t l_samples [4096];
62 * Float_t r_samples [4096];
63 * size_t num_samples;
64 * unsigned int num_songs;
65 * unsigned int i;
66 *
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() );
72 * }
73 * fprintf ("Recommended dB change for whole album: %+6.2f dB\n", GetAlbumGain() );
74 */
75
76 /*
77 * So here's the main source of potential code confusion:
78 *
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.
82 *
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
86 * simple routine.
87 *
88 * Optimization/clarity suggestions are welcome.
89 */
90
91 #include <stdio.h>
92 #include <stdlib.h>
93 #include <string.h>
94 #include <math.h>
95
96 #include "gain_analysis.h"
97
98 typedef unsigned short Uint16_t;
99 typedef signed short Int16_t;
100 typedef unsigned int Uint32_t;
101 typedef signed int Int32_t;
102
103 #define YULE_ORDER 10
104 #define BUTTER_ORDER 2
105 #define YULE_FILTER filterYule
106 #define BUTTER_FILTER filterButter
107 #define RMS_PERCENTILE 0.95 // percentile which is louder than the proposed level
108 #define MAX_SAMP_FREQ 48000. // maximum allowed sample frequency [Hz]
109 #define RMS_WINDOW_TIME 0.050 // Time slice size [s]
110 #define STEPS_per_dB 100. // Table entries per dB
111 #define MAX_dB 120. // Table entries for 0...MAX_dB (normal max. values are 70...80 dB)
112
113 #define MAX_ORDER (BUTTER_ORDER > YULE_ORDER ? BUTTER_ORDER : YULE_ORDER)
114 #define MAX_SAMPLES_PER_WINDOW (size_t) (MAX_SAMP_FREQ * RMS_WINDOW_TIME) // max. Samples per Time slice
115 #define PINK_REF 64.82 //298640883795 // calibration value
116
117 Float_t linprebuf [MAX_ORDER * 2];
118 Float_t* linpre; // left input samples, with pre-buffer
119 Float_t lstepbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER];
120 Float_t* lstep; // left "first step" (i.e. post first filter) samples
121 Float_t loutbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER];
122 Float_t* lout; // left "out" (i.e. post second filter) samples
123 Float_t rinprebuf [MAX_ORDER * 2];
124 Float_t* rinpre; // right input samples ...
125 Float_t rstepbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER];
126 Float_t* rstep;
127 Float_t routbuf [MAX_SAMPLES_PER_WINDOW + MAX_ORDER];
128 Float_t* rout;
129 long sampleWindow; // number of samples required to reach number of milliseconds required for RMS window
130 long totsamp;
131 double lsum;
132 double rsum;
133 int freqindex;
134 int first;
135 static Uint32_t A [(size_t)(STEPS_per_dB * MAX_dB)];
136 static Uint32_t B [(size_t)(STEPS_per_dB * MAX_dB)];
137
138 // for each filter:
139 // [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
140
141 #ifdef WIN32
142 #ifndef __GNUC__
143 #pragma warning ( disable : 4305 )
144 #endif
145 #endif
146
147 static const Float_t ABYule[9][2*YULE_ORDER + 1] = {
148 {0.03857599435200, -3.84664617118067, -0.02160367184185, 7.81501653005538, -0.00123395316851,-11.34170355132042, -0.00009291677959, 13.05504219327545, -0.01655260341619,-12.28759895145294, 0.02161526843274, 9.48293806319790, -0.02074045215285, -5.87257861775999, 0.00594298065125, 2.75465861874613, 0.00306428023191, -0.86984376593551, 0.00012025322027, 0.13919314567432, 0.00288463683916 },
149 {0.05418656406430, -3.47845948550071, -0.02911007808948, 6.36317777566148, -0.00848709379851, -8.54751527471874, -0.00851165645469, 9.47693607801280, -0.00834990904936, -8.81498681370155, 0.02245293253339, 6.85401540936998, -0.02596338512915, -4.39470996079559, 0.01624864962975, 2.19611684890774, -0.00240879051584, -0.75104302451432, 0.00674613682247, 0.13149317958808, -0.00187763777362 },
150 {0.15457299681924, -2.37898834973084, -0.09331049056315, 2.84868151156327, -0.06247880153653, -2.64577170229825, 0.02163541888798, 2.23697657451713, -0.05588393329856, -1.67148153367602, 0.04781476674921, 1.00595954808547, 0.00222312597743, -0.45953458054983, 0.03174092540049, 0.16378164858596, -0.01390589421898, -0.05032077717131, 0.00651420667831, 0.02347897407020, -0.00881362733839 },
151 {0.30296907319327, -1.61273165137247, -0.22613988682123, 1.07977492259970, -0.08587323730772, -0.25656257754070, 0.03282930172664, -0.16276719120440, -0.00915702933434, -0.22638893773906, -0.02364141202522, 0.39120800788284, -0.00584456039913, -0.22138138954925, 0.06276101321749, 0.04500235387352, -0.00000828086748, 0.02005851806501, 0.00205861885564, 0.00302439095741, -0.02950134983287 },
152 {0.33642304856132, -1.49858979367799, -0.25572241425570, 0.87350271418188, -0.11828570177555, 0.12205022308084, 0.11921148675203, -0.80774944671438, -0.07834489609479, 0.47854794562326, -0.00469977914380, -0.12453458140019, -0.00589500224440, -0.04067510197014, 0.05724228140351, 0.08333755284107, 0.00832043980773, -0.04237348025746, -0.01635381384540, 0.02977207319925, -0.01760176568150 },
153 {0.44915256608450, -0.62820619233671, -0.14351757464547, 0.29661783706366, -0.22784394429749, -0.37256372942400, -0.01419140100551, 0.00213767857124, 0.04078262797139, -0.42029820170918, -0.12398163381748, 0.22199650564824, 0.04097565135648, 0.00613424350682, 0.10478503600251, 0.06747620744683, -0.01863887810927, 0.05784820375801, -0.03193428438915, 0.03222754072173, 0.00541907748707 },
154 {0.56619470757641, -1.04800335126349, -0.75464456939302, 0.29156311971249, 0.16242137742230, -0.26806001042947, 0.16744243493672, 0.00819999645858, -0.18901604199609, 0.45054734505008, 0.30931782841830, -0.33032403314006, -0.27562961986224, 0.06739368333110, 0.00647310677246, -0.04784254229033, 0.08647503780351, 0.01639907836189, -0.03788984554840, 0.01807364323573, -0.00588215443421 },
155 {0.58100494960553, -0.51035327095184, -0.53174909058578, -0.31863563325245, -0.14289799034253, -0.20256413484477, 0.17520704835522, 0.14728154134330, 0.02377945217615, 0.38952639978999, 0.15558449135573, -0.23313271880868, -0.25344790059353, -0.05246019024463, 0.01628462406333, -0.02505961724053, 0.06920467763959, 0.02442357316099, -0.03721611395801, 0.01818801111503, -0.00749618797172 },
156 {0.53648789255105, -0.25049871956020, -0.42163034350696, -0.43193942311114, -0.00275953611929, -0.03424681017675, 0.04267842219415, -0.04678328784242, -0.10214864179676, 0.26408300200955, 0.14590772289388, 0.15113130533216, -0.02459864859345, -0.17556493366449, -0.11202315195388, -0.18823009262115, -0.04060034127000, 0.05477720428674, 0.04788665548180, 0.04704409688120, -0.02217936801134 }
157 };
158
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 }
169 };
170
171
172 #ifdef WIN32
173 #ifndef __GNUC__
174 #pragma warning ( default : 4305 )
175 #endif
176 #endif
177
178 // When calling these filter procedures, make sure that ip[-order] and op[-order] point to real data!
179
180 // If your compiler complains that "'operation on 'output' may be undefined", you can
181 // either ignore the warnings or uncomment the three "y" lines (and comment out the indicated line)
182
183 static void
filterYule(const Float_t * input,Float_t * output,size_t nSamples,const Float_t * kernel)184 filterYule (const Float_t* input, Float_t* output, size_t nSamples, const Float_t* kernel)
185 {
186
187 while (nSamples--) {
188 *output = 1e-10 /* 1e-10 is a hack to avoid slowdown because of denormals */
189 + input [0] * kernel[0]
190 - output[-1] * kernel[1]
191 + input [-1] * kernel[2]
192 - output[-2] * kernel[3]
193 + input [-2] * kernel[4]
194 - output[-3] * kernel[5]
195 + input [-3] * kernel[6]
196 - output[-4] * kernel[7]
197 + input [-4] * kernel[8]
198 - output[-5] * kernel[9]
199 + input [-5] * kernel[10]
200 - output[-6] * kernel[11]
201 + input [-6] * kernel[12]
202 - output[-7] * kernel[13]
203 + input [-7] * kernel[14]
204 - output[-8] * kernel[15]
205 + input [-8] * kernel[16]
206 - output[-9] * kernel[17]
207 + input [-9] * kernel[18]
208 - output[-10]* kernel[19]
209 + input [-10]* kernel[20];
210 ++output;
211 ++input;
212 }
213 }
214
215 static void
filterButter(const Float_t * input,Float_t * output,size_t nSamples,const Float_t * kernel)216 filterButter (const Float_t* input, Float_t* output, size_t nSamples, const Float_t* kernel)
217 {
218
219 while (nSamples--) {
220 *output =
221 input [0] * kernel[0]
222 - output[-1] * kernel[1]
223 + input [-1] * kernel[2]
224 - output[-2] * kernel[3]
225 + input [-2] * kernel[4];
226 ++output;
227 ++input;
228 }
229 }
230
231
232 // returns a INIT_GAIN_ANALYSIS_OK if successful, INIT_GAIN_ANALYSIS_ERROR if not
233
234 int
ResetSampleFrequency(long samplefreq)235 ResetSampleFrequency ( long samplefreq ) {
236 int i;
237
238 // zero out initial values
239 for ( i = 0; i < MAX_ORDER; i++ )
240 linprebuf[i] = lstepbuf[i] = loutbuf[i] = rinprebuf[i] = rstepbuf[i] = routbuf[i] = 0.;
241
242 switch ( (int)(samplefreq) ) {
243 case 48000: freqindex = 0; break;
244 case 44100: freqindex = 1; break;
245 case 32000: freqindex = 2; break;
246 case 24000: freqindex = 3; break;
247 case 22050: freqindex = 4; break;
248 case 16000: freqindex = 5; break;
249 case 12000: freqindex = 6; break;
250 case 11025: freqindex = 7; break;
251 case 8000: freqindex = 8; break;
252 default: return INIT_GAIN_ANALYSIS_ERROR;
253 }
254
255 sampleWindow = (int) ceil (samplefreq * RMS_WINDOW_TIME);
256
257 lsum = 0.;
258 rsum = 0.;
259 totsamp = 0;
260
261 memset ( A, 0, sizeof(A) );
262
263 return INIT_GAIN_ANALYSIS_OK;
264 }
265
266 int
InitGainAnalysis(long samplefreq)267 InitGainAnalysis ( long samplefreq )
268 {
269 if (ResetSampleFrequency(samplefreq) != INIT_GAIN_ANALYSIS_OK) {
270 return INIT_GAIN_ANALYSIS_ERROR;
271 }
272
273 linpre = linprebuf + MAX_ORDER;
274 rinpre = rinprebuf + MAX_ORDER;
275 lstep = lstepbuf + MAX_ORDER;
276 rstep = rstepbuf + MAX_ORDER;
277 lout = loutbuf + MAX_ORDER;
278 rout = routbuf + MAX_ORDER;
279
280 memset ( B, 0, sizeof(B) );
281
282 return INIT_GAIN_ANALYSIS_OK;
283 }
284
285 // returns GAIN_ANALYSIS_OK if successful, GAIN_ANALYSIS_ERROR if not
286
fsqr(const double d)287 static __inline double fsqr(const double d)
288 { return d*d;
289 }
290
291 int
AnalyzeSamples(const Float_t * left_samples,const Float_t * right_samples,size_t num_samples,int num_channels)292 AnalyzeSamples ( const Float_t* left_samples, const Float_t* right_samples, size_t num_samples, int num_channels )
293 {
294 const Float_t* curleft;
295 const Float_t* curright;
296 long batchsamples;
297 long cursamples;
298 long cursamplepos;
299 int i;
300
301 if ( num_samples == 0 )
302 return GAIN_ANALYSIS_OK;
303
304 cursamplepos = 0;
305 batchsamples = num_samples;
306
307 switch ( num_channels) {
308 case 1: right_samples = left_samples;
309 case 2: break;
310 default: return GAIN_ANALYSIS_ERROR;
311 }
312
313 if ( num_samples < MAX_ORDER ) {
314 memcpy ( linprebuf + MAX_ORDER, left_samples , num_samples * sizeof(Float_t) );
315 memcpy ( rinprebuf + MAX_ORDER, right_samples, num_samples * sizeof(Float_t) );
316 }
317 else {
318 memcpy ( linprebuf + MAX_ORDER, left_samples, MAX_ORDER * sizeof(Float_t) );
319 memcpy ( rinprebuf + MAX_ORDER, right_samples, MAX_ORDER * sizeof(Float_t) );
320 }
321
322 while ( batchsamples > 0 ) {
323 cursamples = batchsamples > sampleWindow-totsamp ? sampleWindow - totsamp : batchsamples;
324 if ( cursamplepos < MAX_ORDER ) {
325 curleft = linpre+cursamplepos;
326 curright = rinpre+cursamplepos;
327 if (cursamples > MAX_ORDER - cursamplepos )
328 cursamples = MAX_ORDER - cursamplepos;
329 }
330 else {
331 curleft = left_samples + cursamplepos;
332 curright = right_samples + cursamplepos;
333 }
334
335 YULE_FILTER ( curleft , lstep + totsamp, cursamples, ABYule[freqindex]);
336 YULE_FILTER ( curright, rstep + totsamp, cursamples, ABYule[freqindex]);
337
338 BUTTER_FILTER ( lstep + totsamp, lout + totsamp, cursamples, ABButter[freqindex]);
339 BUTTER_FILTER ( rstep + totsamp, rout + totsamp, cursamples, ABButter[freqindex]);
340
341 curleft = lout + totsamp; // Get the squared values
342 curright = rout + totsamp;
343
344 i = cursamples % 16;
345 while (i--)
346 { lsum += fsqr(*curleft++);
347 rsum += fsqr(*curright++);
348 }
349 i = cursamples / 16;
350 while (i--)
351 { lsum += fsqr(curleft[0])
352 + fsqr(curleft[1])
353 + fsqr(curleft[2])
354 + fsqr(curleft[3])
355 + fsqr(curleft[4])
356 + fsqr(curleft[5])
357 + fsqr(curleft[6])
358 + fsqr(curleft[7])
359 + fsqr(curleft[8])
360 + fsqr(curleft[9])
361 + fsqr(curleft[10])
362 + fsqr(curleft[11])
363 + fsqr(curleft[12])
364 + fsqr(curleft[13])
365 + fsqr(curleft[14])
366 + fsqr(curleft[15]);
367 curleft += 16;
368 rsum += fsqr(curright[0])
369 + fsqr(curright[1])
370 + fsqr(curright[2])
371 + fsqr(curright[3])
372 + fsqr(curright[4])
373 + fsqr(curright[5])
374 + fsqr(curright[6])
375 + fsqr(curright[7])
376 + fsqr(curright[8])
377 + fsqr(curright[9])
378 + fsqr(curright[10])
379 + fsqr(curright[11])
380 + fsqr(curright[12])
381 + fsqr(curright[13])
382 + fsqr(curright[14])
383 + fsqr(curright[15]);
384 curright += 16;
385 }
386
387 batchsamples -= cursamples;
388 cursamplepos += cursamples;
389 totsamp += cursamples;
390 if ( totsamp == sampleWindow ) { // Get the Root Mean Square (RMS) for this set of samples
391 double val = STEPS_per_dB * 10. * log10 ( (lsum+rsum) / totsamp * 0.5 + 1.e-37 );
392 int ival = (int) val;
393 if ( ival < 0 ) ival = 0;
394 if ( ival >= (int)(sizeof(A)/sizeof(*A)) ) ival = sizeof(A)/sizeof(*A) - 1;
395 A [ival]++;
396 lsum = rsum = 0.;
397 memmove ( loutbuf , loutbuf + totsamp, MAX_ORDER * sizeof(Float_t) );
398 memmove ( routbuf , routbuf + totsamp, MAX_ORDER * sizeof(Float_t) );
399 memmove ( lstepbuf, lstepbuf + totsamp, MAX_ORDER * sizeof(Float_t) );
400 memmove ( rstepbuf, rstepbuf + totsamp, MAX_ORDER * sizeof(Float_t) );
401 totsamp = 0;
402 }
403 if ( totsamp > sampleWindow ) // somehow I really screwed up: Error in programming! Contact author about totsamp > sampleWindow
404 return GAIN_ANALYSIS_ERROR;
405 }
406 if ( num_samples < MAX_ORDER ) {
407 memmove ( linprebuf, linprebuf + num_samples, (MAX_ORDER-num_samples) * sizeof(Float_t) );
408 memmove ( rinprebuf, rinprebuf + num_samples, (MAX_ORDER-num_samples) * sizeof(Float_t) );
409 memcpy ( linprebuf + MAX_ORDER - num_samples, left_samples, num_samples * sizeof(Float_t) );
410 memcpy ( rinprebuf + MAX_ORDER - num_samples, right_samples, num_samples * sizeof(Float_t) );
411 }
412 else {
413 memcpy ( linprebuf, left_samples + num_samples - MAX_ORDER, MAX_ORDER * sizeof(Float_t) );
414 memcpy ( rinprebuf, right_samples + num_samples - MAX_ORDER, MAX_ORDER * sizeof(Float_t) );
415 }
416
417 return GAIN_ANALYSIS_OK;
418 }
419
420
421 static Float_t
analyzeResult(Uint32_t * Array,size_t len)422 analyzeResult ( Uint32_t* Array, size_t len )
423 {
424 Uint32_t elems;
425 Int32_t upper;
426 size_t i;
427
428 elems = 0;
429 for ( i = 0; i < len; i++ )
430 elems += Array[i];
431 if ( elems == 0 )
432 return GAIN_NOT_ENOUGH_SAMPLES;
433
434 upper = (Int32_t) ceil (elems * (1. - RMS_PERCENTILE));
435 for ( i = len; i-- > 0; ) {
436 if ( (upper -= Array[i]) <= 0 )
437 break;
438 }
439
440 return (Float_t) ((Float_t)PINK_REF - (Float_t)i / (Float_t)STEPS_per_dB);
441 }
442
443
444 Float_t
GetTitleGain(void)445 GetTitleGain ( void )
446 {
447 Float_t retval;
448 int i;
449
450 retval = analyzeResult ( A, sizeof(A)/sizeof(*A) );
451
452 for ( i = 0; i < (int)(sizeof(A)/sizeof(*A)); i++ ) {
453 B[i] += A[i];
454 A[i] = 0;
455 }
456
457 for ( i = 0; i < MAX_ORDER; i++ )
458 linprebuf[i] = lstepbuf[i] = loutbuf[i] = rinprebuf[i] = rstepbuf[i] = routbuf[i] = 0.f;
459
460 totsamp = 0;
461 lsum = rsum = 0.;
462 return retval;
463 }
464
465
466 Float_t
GetAlbumGain(void)467 GetAlbumGain ( void )
468 {
469 return analyzeResult ( B, sizeof(B)/sizeof(*B) );
470 }
471
472 /* end of gain_analysis.c */
473