• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2016 Muhammad Faiz <mfcc64@gmail.com>
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg 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  * FFmpeg 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 FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20 
21 #include "libavutil/opt.h"
22 #include "libavutil/eval.h"
23 #include "libavutil/avassert.h"
24 #include "libavcodec/avfft.h"
25 #include "avfilter.h"
26 #include "internal.h"
27 #include "audio.h"
28 
29 #define RDFT_BITS_MIN 4
30 #define RDFT_BITS_MAX 16
31 
32 enum WindowFunc {
33     WFUNC_RECTANGULAR,
34     WFUNC_HANN,
35     WFUNC_HAMMING,
36     WFUNC_BLACKMAN,
37     WFUNC_NUTTALL3,
38     WFUNC_MNUTTALL3,
39     WFUNC_NUTTALL,
40     WFUNC_BNUTTALL,
41     WFUNC_BHARRIS,
42     WFUNC_TUKEY,
43     NB_WFUNC
44 };
45 
46 enum Scale {
47     SCALE_LINLIN,
48     SCALE_LINLOG,
49     SCALE_LOGLIN,
50     SCALE_LOGLOG,
51     NB_SCALE
52 };
53 
54 #define NB_GAIN_ENTRY_MAX 4096
55 typedef struct GainEntry {
56     double  freq;
57     double  gain;
58 } GainEntry;
59 
60 typedef struct OverlapIndex {
61     int buf_idx;
62     int overlap_idx;
63 } OverlapIndex;
64 
65 typedef struct FIREqualizerContext {
66     const AVClass *class;
67 
68     RDFTContext   *analysis_rdft;
69     RDFTContext   *analysis_irdft;
70     RDFTContext   *rdft;
71     RDFTContext   *irdft;
72     FFTContext    *fft_ctx;
73     RDFTContext   *cepstrum_rdft;
74     RDFTContext   *cepstrum_irdft;
75     int           analysis_rdft_len;
76     int           rdft_len;
77     int           cepstrum_len;
78 
79     float         *analysis_buf;
80     float         *dump_buf;
81     float         *kernel_tmp_buf;
82     float         *kernel_buf;
83     float         *cepstrum_buf;
84     float         *conv_buf;
85     OverlapIndex  *conv_idx;
86     int           fir_len;
87     int           nsamples_max;
88     int64_t       next_pts;
89     int           frame_nsamples_max;
90     int           remaining;
91 
92     char          *gain_cmd;
93     char          *gain_entry_cmd;
94     const char    *gain;
95     const char    *gain_entry;
96     double        delay;
97     double        accuracy;
98     int           wfunc;
99     int           fixed;
100     int           multi;
101     int           zero_phase;
102     int           scale;
103     char          *dumpfile;
104     int           dumpscale;
105     int           fft2;
106     int           min_phase;
107 
108     int           nb_gain_entry;
109     int           gain_entry_err;
110     GainEntry     gain_entry_tbl[NB_GAIN_ENTRY_MAX];
111 } FIREqualizerContext;
112 
113 #define OFFSET(x) offsetof(FIREqualizerContext, x)
114 #define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
115 #define TFLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
116 
117 static const AVOption firequalizer_options[] = {
118     { "gain", "set gain curve", OFFSET(gain), AV_OPT_TYPE_STRING, { .str = "gain_interpolate(f)" }, 0, 0, TFLAGS },
119     { "gain_entry", "set gain entry", OFFSET(gain_entry), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, TFLAGS },
120     { "delay", "set delay", OFFSET(delay), AV_OPT_TYPE_DOUBLE, { .dbl = 0.01 }, 0.0, 1e10, FLAGS },
121     { "accuracy", "set accuracy", OFFSET(accuracy), AV_OPT_TYPE_DOUBLE, { .dbl = 5.0 }, 0.0, 1e10, FLAGS },
122     { "wfunc", "set window function", OFFSET(wfunc), AV_OPT_TYPE_INT, { .i64 = WFUNC_HANN }, 0, NB_WFUNC-1, FLAGS, "wfunc" },
123         { "rectangular", "rectangular window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_RECTANGULAR }, 0, 0, FLAGS, "wfunc" },
124         { "hann", "hann window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_HANN }, 0, 0, FLAGS, "wfunc" },
125         { "hamming", "hamming window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_HAMMING }, 0, 0, FLAGS, "wfunc" },
126         { "blackman", "blackman window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_BLACKMAN }, 0, 0, FLAGS, "wfunc" },
127         { "nuttall3", "3-term nuttall window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_NUTTALL3 }, 0, 0, FLAGS, "wfunc" },
128         { "mnuttall3", "minimum 3-term nuttall window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_MNUTTALL3 }, 0, 0, FLAGS, "wfunc" },
129         { "nuttall", "nuttall window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_NUTTALL }, 0, 0, FLAGS, "wfunc" },
130         { "bnuttall", "blackman-nuttall window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_BNUTTALL }, 0, 0, FLAGS, "wfunc" },
131         { "bharris", "blackman-harris window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_BHARRIS }, 0, 0, FLAGS, "wfunc" },
132         { "tukey", "tukey window", 0, AV_OPT_TYPE_CONST, { .i64 = WFUNC_TUKEY }, 0, 0, FLAGS, "wfunc" },
133     { "fixed", "set fixed frame samples", OFFSET(fixed), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
134     { "multi", "set multi channels mode", OFFSET(multi), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
135     { "zero_phase", "set zero phase mode", OFFSET(zero_phase), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
136     { "scale", "set gain scale", OFFSET(scale), AV_OPT_TYPE_INT, { .i64 = SCALE_LINLOG }, 0, NB_SCALE-1, FLAGS, "scale" },
137         { "linlin", "linear-freq linear-gain", 0, AV_OPT_TYPE_CONST, { .i64 = SCALE_LINLIN }, 0, 0, FLAGS, "scale" },
138         { "linlog", "linear-freq logarithmic-gain", 0, AV_OPT_TYPE_CONST, { .i64 = SCALE_LINLOG }, 0, 0, FLAGS, "scale" },
139         { "loglin", "logarithmic-freq linear-gain", 0, AV_OPT_TYPE_CONST, { .i64 = SCALE_LOGLIN }, 0, 0, FLAGS, "scale" },
140         { "loglog", "logarithmic-freq logarithmic-gain", 0, AV_OPT_TYPE_CONST, { .i64 = SCALE_LOGLOG }, 0, 0, FLAGS, "scale" },
141     { "dumpfile", "set dump file", OFFSET(dumpfile), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
142     { "dumpscale", "set dump scale", OFFSET(dumpscale), AV_OPT_TYPE_INT, { .i64 = SCALE_LINLOG }, 0, NB_SCALE-1, FLAGS, "scale" },
143     { "fft2", "set 2-channels fft", OFFSET(fft2), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
144     { "min_phase", "set minimum phase mode", OFFSET(min_phase), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS },
145     { NULL }
146 };
147 
148 AVFILTER_DEFINE_CLASS(firequalizer);
149 
common_uninit(FIREqualizerContext * s)150 static void common_uninit(FIREqualizerContext *s)
151 {
152     av_rdft_end(s->analysis_rdft);
153     av_rdft_end(s->analysis_irdft);
154     av_rdft_end(s->rdft);
155     av_rdft_end(s->irdft);
156     av_fft_end(s->fft_ctx);
157     av_rdft_end(s->cepstrum_rdft);
158     av_rdft_end(s->cepstrum_irdft);
159     s->analysis_rdft = s->analysis_irdft = s->rdft = s->irdft = NULL;
160     s->fft_ctx = NULL;
161     s->cepstrum_rdft = NULL;
162     s->cepstrum_irdft = NULL;
163 
164     av_freep(&s->analysis_buf);
165     av_freep(&s->dump_buf);
166     av_freep(&s->kernel_tmp_buf);
167     av_freep(&s->kernel_buf);
168     av_freep(&s->cepstrum_buf);
169     av_freep(&s->conv_buf);
170     av_freep(&s->conv_idx);
171 }
172 
uninit(AVFilterContext * ctx)173 static av_cold void uninit(AVFilterContext *ctx)
174 {
175     FIREqualizerContext *s = ctx->priv;
176 
177     common_uninit(s);
178     av_freep(&s->gain_cmd);
179     av_freep(&s->gain_entry_cmd);
180 }
181 
query_formats(AVFilterContext * ctx)182 static int query_formats(AVFilterContext *ctx)
183 {
184     AVFilterChannelLayouts *layouts;
185     AVFilterFormats *formats;
186     static const enum AVSampleFormat sample_fmts[] = {
187         AV_SAMPLE_FMT_FLTP,
188         AV_SAMPLE_FMT_NONE
189     };
190     int ret;
191 
192     layouts = ff_all_channel_counts();
193     if (!layouts)
194         return AVERROR(ENOMEM);
195     ret = ff_set_common_channel_layouts(ctx, layouts);
196     if (ret < 0)
197         return ret;
198 
199     formats = ff_make_format_list(sample_fmts);
200     if (!formats)
201         return AVERROR(ENOMEM);
202     ret = ff_set_common_formats(ctx, formats);
203     if (ret < 0)
204         return ret;
205 
206     formats = ff_all_samplerates();
207     if (!formats)
208         return AVERROR(ENOMEM);
209     return ff_set_common_samplerates(ctx, formats);
210 }
211 
fast_convolute(FIREqualizerContext * av_restrict s,const float * av_restrict kernel_buf,float * av_restrict conv_buf,OverlapIndex * av_restrict idx,float * av_restrict data,int nsamples)212 static void fast_convolute(FIREqualizerContext *av_restrict s, const float *av_restrict kernel_buf, float *av_restrict conv_buf,
213                            OverlapIndex *av_restrict idx, float *av_restrict data, int nsamples)
214 {
215     if (nsamples <= s->nsamples_max) {
216         float *buf = conv_buf + idx->buf_idx * s->rdft_len;
217         float *obuf = conv_buf + !idx->buf_idx * s->rdft_len + idx->overlap_idx;
218         int center = s->fir_len/2;
219         int k;
220 
221         memset(buf, 0, center * sizeof(*data));
222         memcpy(buf + center, data, nsamples * sizeof(*data));
223         memset(buf + center + nsamples, 0, (s->rdft_len - nsamples - center) * sizeof(*data));
224         av_rdft_calc(s->rdft, buf);
225 
226         buf[0] *= kernel_buf[0];
227         buf[1] *= kernel_buf[s->rdft_len/2];
228         for (k = 1; k < s->rdft_len/2; k++) {
229             buf[2*k] *= kernel_buf[k];
230             buf[2*k+1] *= kernel_buf[k];
231         }
232 
233         av_rdft_calc(s->irdft, buf);
234         for (k = 0; k < s->rdft_len - idx->overlap_idx; k++)
235             buf[k] += obuf[k];
236         memcpy(data, buf, nsamples * sizeof(*data));
237         idx->buf_idx = !idx->buf_idx;
238         idx->overlap_idx = nsamples;
239     } else {
240         while (nsamples > s->nsamples_max * 2) {
241             fast_convolute(s, kernel_buf, conv_buf, idx, data, s->nsamples_max);
242             data += s->nsamples_max;
243             nsamples -= s->nsamples_max;
244         }
245         fast_convolute(s, kernel_buf, conv_buf, idx, data, nsamples/2);
246         fast_convolute(s, kernel_buf, conv_buf, idx, data + nsamples/2, nsamples - nsamples/2);
247     }
248 }
249 
fast_convolute_nonlinear(FIREqualizerContext * av_restrict s,const float * av_restrict kernel_buf,float * av_restrict conv_buf,OverlapIndex * av_restrict idx,float * av_restrict data,int nsamples)250 static void fast_convolute_nonlinear(FIREqualizerContext *av_restrict s, const float *av_restrict kernel_buf,
251                                      float *av_restrict conv_buf, OverlapIndex *av_restrict idx,
252                                      float *av_restrict data, int nsamples)
253 {
254     if (nsamples <= s->nsamples_max) {
255         float *buf = conv_buf + idx->buf_idx * s->rdft_len;
256         float *obuf = conv_buf + !idx->buf_idx * s->rdft_len + idx->overlap_idx;
257         int k;
258 
259         memcpy(buf, data, nsamples * sizeof(*data));
260         memset(buf + nsamples, 0, (s->rdft_len - nsamples) * sizeof(*data));
261         av_rdft_calc(s->rdft, buf);
262 
263         buf[0] *= kernel_buf[0];
264         buf[1] *= kernel_buf[1];
265         for (k = 2; k < s->rdft_len; k += 2) {
266             float re, im;
267             re = buf[k] * kernel_buf[k] - buf[k+1] * kernel_buf[k+1];
268             im = buf[k] * kernel_buf[k+1] + buf[k+1] * kernel_buf[k];
269             buf[k] = re;
270             buf[k+1] = im;
271         }
272 
273         av_rdft_calc(s->irdft, buf);
274         for (k = 0; k < s->rdft_len - idx->overlap_idx; k++)
275             buf[k] += obuf[k];
276         memcpy(data, buf, nsamples * sizeof(*data));
277         idx->buf_idx = !idx->buf_idx;
278         idx->overlap_idx = nsamples;
279     } else {
280         while (nsamples > s->nsamples_max * 2) {
281             fast_convolute_nonlinear(s, kernel_buf, conv_buf, idx, data, s->nsamples_max);
282             data += s->nsamples_max;
283             nsamples -= s->nsamples_max;
284         }
285         fast_convolute_nonlinear(s, kernel_buf, conv_buf, idx, data, nsamples/2);
286         fast_convolute_nonlinear(s, kernel_buf, conv_buf, idx, data + nsamples/2, nsamples - nsamples/2);
287     }
288 }
289 
fast_convolute2(FIREqualizerContext * av_restrict s,const float * av_restrict kernel_buf,FFTComplex * av_restrict conv_buf,OverlapIndex * av_restrict idx,float * av_restrict data0,float * av_restrict data1,int nsamples)290 static void fast_convolute2(FIREqualizerContext *av_restrict s, const float *av_restrict kernel_buf, FFTComplex *av_restrict conv_buf,
291                             OverlapIndex *av_restrict idx, float *av_restrict data0, float *av_restrict data1, int nsamples)
292 {
293     if (nsamples <= s->nsamples_max) {
294         FFTComplex *buf = conv_buf + idx->buf_idx * s->rdft_len;
295         FFTComplex *obuf = conv_buf + !idx->buf_idx * s->rdft_len + idx->overlap_idx;
296         int center = s->fir_len/2;
297         int k;
298         float tmp;
299 
300         memset(buf, 0, center * sizeof(*buf));
301         for (k = 0; k < nsamples; k++) {
302             buf[center+k].re = data0[k];
303             buf[center+k].im = data1[k];
304         }
305         memset(buf + center + nsamples, 0, (s->rdft_len - nsamples - center) * sizeof(*buf));
306         av_fft_permute(s->fft_ctx, buf);
307         av_fft_calc(s->fft_ctx, buf);
308 
309         /* swap re <-> im, do backward fft using forward fft_ctx */
310         /* normalize with 0.5f */
311         tmp = buf[0].re;
312         buf[0].re = 0.5f * kernel_buf[0] * buf[0].im;
313         buf[0].im = 0.5f * kernel_buf[0] * tmp;
314         for (k = 1; k < s->rdft_len/2; k++) {
315             int m = s->rdft_len - k;
316             tmp = buf[k].re;
317             buf[k].re = 0.5f * kernel_buf[k] * buf[k].im;
318             buf[k].im = 0.5f * kernel_buf[k] * tmp;
319             tmp = buf[m].re;
320             buf[m].re = 0.5f * kernel_buf[k] * buf[m].im;
321             buf[m].im = 0.5f * kernel_buf[k] * tmp;
322         }
323         tmp = buf[k].re;
324         buf[k].re = 0.5f * kernel_buf[k] * buf[k].im;
325         buf[k].im = 0.5f * kernel_buf[k] * tmp;
326 
327         av_fft_permute(s->fft_ctx, buf);
328         av_fft_calc(s->fft_ctx, buf);
329 
330         for (k = 0; k < s->rdft_len - idx->overlap_idx; k++) {
331             buf[k].re += obuf[k].re;
332             buf[k].im += obuf[k].im;
333         }
334 
335         /* swapped re <-> im */
336         for (k = 0; k < nsamples; k++) {
337             data0[k] = buf[k].im;
338             data1[k] = buf[k].re;
339         }
340         idx->buf_idx = !idx->buf_idx;
341         idx->overlap_idx = nsamples;
342     } else {
343         while (nsamples > s->nsamples_max * 2) {
344             fast_convolute2(s, kernel_buf, conv_buf, idx, data0, data1, s->nsamples_max);
345             data0 += s->nsamples_max;
346             data1 += s->nsamples_max;
347             nsamples -= s->nsamples_max;
348         }
349         fast_convolute2(s, kernel_buf, conv_buf, idx, data0, data1, nsamples/2);
350         fast_convolute2(s, kernel_buf, conv_buf, idx, data0 + nsamples/2, data1 + nsamples/2, nsamples - nsamples/2);
351     }
352 }
353 
dump_fir(AVFilterContext * ctx,FILE * fp,int ch)354 static void dump_fir(AVFilterContext *ctx, FILE *fp, int ch)
355 {
356     FIREqualizerContext *s = ctx->priv;
357     int rate = ctx->inputs[0]->sample_rate;
358     int xlog = s->dumpscale == SCALE_LOGLIN || s->dumpscale == SCALE_LOGLOG;
359     int ylog = s->dumpscale == SCALE_LINLOG || s->dumpscale == SCALE_LOGLOG;
360     int x;
361     int center = s->fir_len / 2;
362     double delay = s->zero_phase ? 0.0 : (double) center / rate;
363     double vx, ya, yb;
364 
365     if (!s->min_phase) {
366         s->analysis_buf[0] *= s->rdft_len/2;
367         for (x = 1; x <= center; x++) {
368             s->analysis_buf[x] *= s->rdft_len/2;
369             s->analysis_buf[s->analysis_rdft_len - x] *= s->rdft_len/2;
370         }
371     } else {
372         for (x = 0; x < s->fir_len; x++)
373             s->analysis_buf[x] *= s->rdft_len/2;
374     }
375 
376     if (ch)
377         fprintf(fp, "\n\n");
378 
379     fprintf(fp, "# time[%d] (time amplitude)\n", ch);
380 
381     if (!s->min_phase) {
382     for (x = center; x > 0; x--)
383         fprintf(fp, "%15.10f %15.10f\n", delay - (double) x / rate, (double) s->analysis_buf[s->analysis_rdft_len - x]);
384 
385     for (x = 0; x <= center; x++)
386         fprintf(fp, "%15.10f %15.10f\n", delay + (double)x / rate , (double) s->analysis_buf[x]);
387     } else {
388         for (x = 0; x < s->fir_len; x++)
389             fprintf(fp, "%15.10f %15.10f\n", (double)x / rate, (double) s->analysis_buf[x]);
390     }
391 
392     av_rdft_calc(s->analysis_rdft, s->analysis_buf);
393 
394     fprintf(fp, "\n\n# freq[%d] (frequency desired_gain actual_gain)\n", ch);
395 
396     for (x = 0; x <= s->analysis_rdft_len/2; x++) {
397         int i = (x == s->analysis_rdft_len/2) ? 1 : 2 * x;
398         vx = (double)x * rate / s->analysis_rdft_len;
399         if (xlog)
400             vx = log2(0.05*vx);
401         ya = s->dump_buf[i];
402         yb = s->min_phase && (i > 1) ? hypotf(s->analysis_buf[i], s->analysis_buf[i+1]) : s->analysis_buf[i];
403         if (s->min_phase)
404             yb = fabs(yb);
405         if (ylog) {
406             ya = 20.0 * log10(fabs(ya));
407             yb = 20.0 * log10(fabs(yb));
408         }
409         fprintf(fp, "%17.10f %17.10f %17.10f\n", vx, ya, yb);
410     }
411 }
412 
entry_func(void * p,double freq,double gain)413 static double entry_func(void *p, double freq, double gain)
414 {
415     AVFilterContext *ctx = p;
416     FIREqualizerContext *s = ctx->priv;
417 
418     if (s->nb_gain_entry >= NB_GAIN_ENTRY_MAX) {
419         av_log(ctx, AV_LOG_ERROR, "entry table overflow.\n");
420         s->gain_entry_err = AVERROR(EINVAL);
421         return 0;
422     }
423 
424     if (isnan(freq)) {
425         av_log(ctx, AV_LOG_ERROR, "nan frequency (%g, %g).\n", freq, gain);
426         s->gain_entry_err = AVERROR(EINVAL);
427         return 0;
428     }
429 
430     if (s->nb_gain_entry > 0 && freq <= s->gain_entry_tbl[s->nb_gain_entry - 1].freq) {
431         av_log(ctx, AV_LOG_ERROR, "unsorted frequency (%g, %g).\n", freq, gain);
432         s->gain_entry_err = AVERROR(EINVAL);
433         return 0;
434     }
435 
436     s->gain_entry_tbl[s->nb_gain_entry].freq = freq;
437     s->gain_entry_tbl[s->nb_gain_entry].gain = gain;
438     s->nb_gain_entry++;
439     return 0;
440 }
441 
gain_entry_compare(const void * key,const void * memb)442 static int gain_entry_compare(const void *key, const void *memb)
443 {
444     const double *freq = key;
445     const GainEntry *entry = memb;
446 
447     if (*freq < entry[0].freq)
448         return -1;
449     if (*freq > entry[1].freq)
450         return 1;
451     return 0;
452 }
453 
gain_interpolate_func(void * p,double freq)454 static double gain_interpolate_func(void *p, double freq)
455 {
456     AVFilterContext *ctx = p;
457     FIREqualizerContext *s = ctx->priv;
458     GainEntry *res;
459     double d0, d1, d;
460 
461     if (isnan(freq))
462         return freq;
463 
464     if (!s->nb_gain_entry)
465         return 0;
466 
467     if (freq <= s->gain_entry_tbl[0].freq)
468         return s->gain_entry_tbl[0].gain;
469 
470     if (freq >= s->gain_entry_tbl[s->nb_gain_entry-1].freq)
471         return s->gain_entry_tbl[s->nb_gain_entry-1].gain;
472 
473     res = bsearch(&freq, &s->gain_entry_tbl, s->nb_gain_entry - 1, sizeof(*res), gain_entry_compare);
474     av_assert0(res);
475 
476     d  = res[1].freq - res[0].freq;
477     d0 = freq - res[0].freq;
478     d1 = res[1].freq - freq;
479 
480     if (d0 && d1)
481         return (d0 * res[1].gain + d1 * res[0].gain) / d;
482 
483     if (d0)
484         return res[1].gain;
485 
486     return res[0].gain;
487 }
488 
cubic_interpolate_func(void * p,double freq)489 static double cubic_interpolate_func(void *p, double freq)
490 {
491     AVFilterContext *ctx = p;
492     FIREqualizerContext *s = ctx->priv;
493     GainEntry *res;
494     double x, x2, x3;
495     double a, b, c, d;
496     double m0, m1, m2, msum, unit;
497 
498     if (!s->nb_gain_entry)
499         return 0;
500 
501     if (freq <= s->gain_entry_tbl[0].freq)
502         return s->gain_entry_tbl[0].gain;
503 
504     if (freq >= s->gain_entry_tbl[s->nb_gain_entry-1].freq)
505         return s->gain_entry_tbl[s->nb_gain_entry-1].gain;
506 
507     res = bsearch(&freq, &s->gain_entry_tbl, s->nb_gain_entry - 1, sizeof(*res), gain_entry_compare);
508     av_assert0(res);
509 
510     unit = res[1].freq - res[0].freq;
511     m0 = res != s->gain_entry_tbl ?
512          unit * (res[0].gain - res[-1].gain) / (res[0].freq - res[-1].freq) : 0;
513     m1 = res[1].gain - res[0].gain;
514     m2 = res != s->gain_entry_tbl + s->nb_gain_entry - 2 ?
515          unit * (res[2].gain - res[1].gain) / (res[2].freq - res[1].freq) : 0;
516 
517     msum = fabs(m0) + fabs(m1);
518     m0 = msum > 0 ? (fabs(m0) * m1 + fabs(m1) * m0) / msum : 0;
519     msum = fabs(m1) + fabs(m2);
520     m1 = msum > 0 ? (fabs(m1) * m2 + fabs(m2) * m1) / msum : 0;
521 
522     d = res[0].gain;
523     c = m0;
524     b = 3 * res[1].gain - m1 - 2 * c - 3 * d;
525     a = res[1].gain - b - c - d;
526 
527     x = (freq - res[0].freq) / unit;
528     x2 = x * x;
529     x3 = x2 * x;
530 
531     return a * x3 + b * x2 + c * x + d;
532 }
533 
534 static const char *const var_names[] = {
535     "f",
536     "sr",
537     "ch",
538     "chid",
539     "chs",
540     "chlayout",
541     NULL
542 };
543 
544 enum VarOffset {
545     VAR_F,
546     VAR_SR,
547     VAR_CH,
548     VAR_CHID,
549     VAR_CHS,
550     VAR_CHLAYOUT,
551     VAR_NB
552 };
553 
generate_min_phase_kernel(FIREqualizerContext * s,float * rdft_buf)554 static void generate_min_phase_kernel(FIREqualizerContext *s, float *rdft_buf)
555 {
556     int k, cepstrum_len = s->cepstrum_len, rdft_len = s->rdft_len;
557     double norm = 2.0 / cepstrum_len;
558     double minval = 1e-7 / rdft_len;
559 
560     memset(s->cepstrum_buf, 0, cepstrum_len * sizeof(*s->cepstrum_buf));
561     memcpy(s->cepstrum_buf, rdft_buf, rdft_len/2 * sizeof(*rdft_buf));
562     memcpy(s->cepstrum_buf + cepstrum_len - rdft_len/2, rdft_buf + rdft_len/2, rdft_len/2  * sizeof(*rdft_buf));
563 
564     av_rdft_calc(s->cepstrum_rdft, s->cepstrum_buf);
565 
566     s->cepstrum_buf[0] = log(FFMAX(s->cepstrum_buf[0], minval));
567     s->cepstrum_buf[1] = log(FFMAX(s->cepstrum_buf[1], minval));
568 
569     for (k = 2; k < cepstrum_len; k += 2) {
570         s->cepstrum_buf[k] = log(FFMAX(s->cepstrum_buf[k], minval));
571         s->cepstrum_buf[k+1] = 0;
572     }
573 
574     av_rdft_calc(s->cepstrum_irdft, s->cepstrum_buf);
575 
576     memset(s->cepstrum_buf + cepstrum_len/2 + 1, 0, (cepstrum_len/2 - 1) * sizeof(*s->cepstrum_buf));
577     for (k = 1; k < cepstrum_len/2; k++)
578         s->cepstrum_buf[k] *= 2;
579 
580     av_rdft_calc(s->cepstrum_rdft, s->cepstrum_buf);
581 
582     s->cepstrum_buf[0] = exp(s->cepstrum_buf[0] * norm) * norm;
583     s->cepstrum_buf[1] = exp(s->cepstrum_buf[1] * norm) * norm;
584     for (k = 2; k < cepstrum_len; k += 2) {
585         double mag = exp(s->cepstrum_buf[k] * norm) * norm;
586         double ph = s->cepstrum_buf[k+1] * norm;
587         s->cepstrum_buf[k] = mag * cos(ph);
588         s->cepstrum_buf[k+1] = mag * sin(ph);
589     }
590 
591     av_rdft_calc(s->cepstrum_irdft, s->cepstrum_buf);
592     memset(rdft_buf, 0, s->rdft_len * sizeof(*rdft_buf));
593     memcpy(rdft_buf, s->cepstrum_buf, s->fir_len * sizeof(*rdft_buf));
594 
595     if (s->dumpfile) {
596         memset(s->analysis_buf, 0, s->analysis_rdft_len * sizeof(*s->analysis_buf));
597         memcpy(s->analysis_buf, s->cepstrum_buf, s->fir_len * sizeof(*s->analysis_buf));
598     }
599 
600 }
601 
generate_kernel(AVFilterContext * ctx,const char * gain,const char * gain_entry)602 static int generate_kernel(AVFilterContext *ctx, const char *gain, const char *gain_entry)
603 {
604     FIREqualizerContext *s = ctx->priv;
605     AVFilterLink *inlink = ctx->inputs[0];
606     const char *gain_entry_func_names[] = { "entry", NULL };
607     const char *gain_func_names[] = { "gain_interpolate", "cubic_interpolate", NULL };
608     double (*gain_entry_funcs[])(void *, double, double) = { entry_func, NULL };
609     double (*gain_funcs[])(void *, double) = { gain_interpolate_func, cubic_interpolate_func, NULL };
610     double vars[VAR_NB];
611     AVExpr *gain_expr;
612     int ret, k, center, ch;
613     int xlog = s->scale == SCALE_LOGLIN || s->scale == SCALE_LOGLOG;
614     int ylog = s->scale == SCALE_LINLOG || s->scale == SCALE_LOGLOG;
615     FILE *dump_fp = NULL;
616 
617     s->nb_gain_entry = 0;
618     s->gain_entry_err = 0;
619     if (gain_entry) {
620         double result = 0.0;
621         ret = av_expr_parse_and_eval(&result, gain_entry, NULL, NULL, NULL, NULL,
622                                      gain_entry_func_names, gain_entry_funcs, ctx, 0, ctx);
623         if (ret < 0)
624             return ret;
625         if (s->gain_entry_err < 0)
626             return s->gain_entry_err;
627     }
628 
629     av_log(ctx, AV_LOG_DEBUG, "nb_gain_entry = %d.\n", s->nb_gain_entry);
630 
631     ret = av_expr_parse(&gain_expr, gain, var_names,
632                         gain_func_names, gain_funcs, NULL, NULL, 0, ctx);
633     if (ret < 0)
634         return ret;
635 
636     if (s->dumpfile && (!s->dump_buf || !s->analysis_rdft || !(dump_fp = fopen(s->dumpfile, "w"))))
637         av_log(ctx, AV_LOG_WARNING, "dumping failed.\n");
638 
639     vars[VAR_CHS] = inlink->channels;
640     vars[VAR_CHLAYOUT] = inlink->channel_layout;
641     vars[VAR_SR] = inlink->sample_rate;
642     for (ch = 0; ch < inlink->channels; ch++) {
643         float *rdft_buf = s->kernel_tmp_buf + ch * s->rdft_len;
644         double result;
645         vars[VAR_CH] = ch;
646         vars[VAR_CHID] = av_channel_layout_extract_channel(inlink->channel_layout, ch);
647         vars[VAR_F] = 0.0;
648         if (xlog)
649             vars[VAR_F] = log2(0.05 * vars[VAR_F]);
650         result = av_expr_eval(gain_expr, vars, ctx);
651         s->analysis_buf[0] = ylog ? pow(10.0, 0.05 * result) : result;
652 
653         vars[VAR_F] = 0.5 * inlink->sample_rate;
654         if (xlog)
655             vars[VAR_F] = log2(0.05 * vars[VAR_F]);
656         result = av_expr_eval(gain_expr, vars, ctx);
657         s->analysis_buf[1] = ylog ? pow(10.0, 0.05 * result) : result;
658 
659         for (k = 1; k < s->analysis_rdft_len/2; k++) {
660             vars[VAR_F] = k * ((double)inlink->sample_rate /(double)s->analysis_rdft_len);
661             if (xlog)
662                 vars[VAR_F] = log2(0.05 * vars[VAR_F]);
663             result = av_expr_eval(gain_expr, vars, ctx);
664             s->analysis_buf[2*k] = ylog ? pow(10.0, 0.05 * result) : s->min_phase ? fabs(result) : result;
665             s->analysis_buf[2*k+1] = 0.0;
666         }
667 
668         if (s->dump_buf)
669             memcpy(s->dump_buf, s->analysis_buf, s->analysis_rdft_len * sizeof(*s->analysis_buf));
670 
671         av_rdft_calc(s->analysis_irdft, s->analysis_buf);
672         center = s->fir_len / 2;
673 
674         for (k = 0; k <= center; k++) {
675             double u = k * (M_PI/center);
676             double win;
677             switch (s->wfunc) {
678             case WFUNC_RECTANGULAR:
679                 win = 1.0;
680                 break;
681             case WFUNC_HANN:
682                 win = 0.5 + 0.5 * cos(u);
683                 break;
684             case WFUNC_HAMMING:
685                 win = 0.53836 + 0.46164 * cos(u);
686                 break;
687             case WFUNC_BLACKMAN:
688                 win = 0.42 + 0.5 * cos(u) + 0.08 * cos(2*u);
689                 break;
690             case WFUNC_NUTTALL3:
691                 win = 0.40897 + 0.5 * cos(u) + 0.09103 * cos(2*u);
692                 break;
693             case WFUNC_MNUTTALL3:
694                 win = 0.4243801 + 0.4973406 * cos(u) + 0.0782793 * cos(2*u);
695                 break;
696             case WFUNC_NUTTALL:
697                 win = 0.355768 + 0.487396 * cos(u) + 0.144232 * cos(2*u) + 0.012604 * cos(3*u);
698                 break;
699             case WFUNC_BNUTTALL:
700                 win = 0.3635819 + 0.4891775 * cos(u) + 0.1365995 * cos(2*u) + 0.0106411 * cos(3*u);
701                 break;
702             case WFUNC_BHARRIS:
703                 win = 0.35875 + 0.48829 * cos(u) + 0.14128 * cos(2*u) + 0.01168 * cos(3*u);
704                 break;
705             case WFUNC_TUKEY:
706                 win = (u <= 0.5 * M_PI) ? 1.0 : (0.5 + 0.5 * cos(2*u - M_PI));
707                 break;
708             default:
709                 av_assert0(0);
710             }
711             s->analysis_buf[k] *= (2.0/s->analysis_rdft_len) * (2.0/s->rdft_len) * win;
712             if (k)
713                 s->analysis_buf[s->analysis_rdft_len - k] = s->analysis_buf[k];
714         }
715 
716         memset(s->analysis_buf + center + 1, 0, (s->analysis_rdft_len - s->fir_len) * sizeof(*s->analysis_buf));
717         memcpy(rdft_buf, s->analysis_buf, s->rdft_len/2 * sizeof(*s->analysis_buf));
718         memcpy(rdft_buf + s->rdft_len/2, s->analysis_buf + s->analysis_rdft_len - s->rdft_len/2, s->rdft_len/2 * sizeof(*s->analysis_buf));
719         if (s->min_phase)
720             generate_min_phase_kernel(s, rdft_buf);
721         av_rdft_calc(s->rdft, rdft_buf);
722 
723         for (k = 0; k < s->rdft_len; k++) {
724             if (isnan(rdft_buf[k]) || isinf(rdft_buf[k])) {
725                 av_log(ctx, AV_LOG_ERROR, "filter kernel contains nan or infinity.\n");
726                 av_expr_free(gain_expr);
727                 if (dump_fp)
728                     fclose(dump_fp);
729                 return AVERROR(EINVAL);
730             }
731         }
732 
733         if (!s->min_phase) {
734             rdft_buf[s->rdft_len-1] = rdft_buf[1];
735             for (k = 0; k < s->rdft_len/2; k++)
736                 rdft_buf[k] = rdft_buf[2*k];
737             rdft_buf[s->rdft_len/2] = rdft_buf[s->rdft_len-1];
738         }
739 
740         if (dump_fp)
741             dump_fir(ctx, dump_fp, ch);
742 
743         if (!s->multi)
744             break;
745     }
746 
747     memcpy(s->kernel_buf, s->kernel_tmp_buf, (s->multi ? inlink->channels : 1) * s->rdft_len * sizeof(*s->kernel_buf));
748     av_expr_free(gain_expr);
749     if (dump_fp)
750         fclose(dump_fp);
751     return 0;
752 }
753 
754 #define SELECT_GAIN(s) (s->gain_cmd ? s->gain_cmd : s->gain)
755 #define SELECT_GAIN_ENTRY(s) (s->gain_entry_cmd ? s->gain_entry_cmd : s->gain_entry)
756 
config_input(AVFilterLink * inlink)757 static int config_input(AVFilterLink *inlink)
758 {
759     AVFilterContext *ctx = inlink->dst;
760     FIREqualizerContext *s = ctx->priv;
761     int rdft_bits;
762 
763     common_uninit(s);
764 
765     s->next_pts = 0;
766     s->frame_nsamples_max = 0;
767 
768     s->fir_len = FFMAX(2 * (int)(inlink->sample_rate * s->delay) + 1, 3);
769     s->remaining = s->fir_len - 1;
770 
771     for (rdft_bits = RDFT_BITS_MIN; rdft_bits <= RDFT_BITS_MAX; rdft_bits++) {
772         s->rdft_len = 1 << rdft_bits;
773         s->nsamples_max = s->rdft_len - s->fir_len + 1;
774         if (s->nsamples_max * 2 >= s->fir_len)
775             break;
776     }
777 
778     if (rdft_bits > RDFT_BITS_MAX) {
779         av_log(ctx, AV_LOG_ERROR, "too large delay, please decrease it.\n");
780         return AVERROR(EINVAL);
781     }
782 
783     if (!(s->rdft = av_rdft_init(rdft_bits, DFT_R2C)) || !(s->irdft = av_rdft_init(rdft_bits, IDFT_C2R)))
784         return AVERROR(ENOMEM);
785 
786     if (s->fft2 && !s->multi && inlink->channels > 1 && !(s->fft_ctx = av_fft_init(rdft_bits, 0)))
787         return AVERROR(ENOMEM);
788 
789     if (s->min_phase) {
790         int cepstrum_bits = rdft_bits + 2;
791         if (cepstrum_bits > RDFT_BITS_MAX) {
792             av_log(ctx, AV_LOG_ERROR, "too large delay, please decrease it.\n");
793             return AVERROR(EINVAL);
794         }
795 
796         cepstrum_bits = FFMIN(RDFT_BITS_MAX, cepstrum_bits + 1);
797         s->cepstrum_rdft = av_rdft_init(cepstrum_bits, DFT_R2C);
798         s->cepstrum_irdft = av_rdft_init(cepstrum_bits, IDFT_C2R);
799         if (!s->cepstrum_rdft || !s->cepstrum_irdft)
800             return AVERROR(ENOMEM);
801 
802         s->cepstrum_len = 1 << cepstrum_bits;
803         s->cepstrum_buf = av_malloc_array(s->cepstrum_len, sizeof(*s->cepstrum_buf));
804         if (!s->cepstrum_buf)
805             return AVERROR(ENOMEM);
806     }
807 
808     for ( ; rdft_bits <= RDFT_BITS_MAX; rdft_bits++) {
809         s->analysis_rdft_len = 1 << rdft_bits;
810         if (inlink->sample_rate <= s->accuracy * s->analysis_rdft_len)
811             break;
812     }
813 
814     if (rdft_bits > RDFT_BITS_MAX) {
815         av_log(ctx, AV_LOG_ERROR, "too small accuracy, please increase it.\n");
816         return AVERROR(EINVAL);
817     }
818 
819     if (!(s->analysis_irdft = av_rdft_init(rdft_bits, IDFT_C2R)))
820         return AVERROR(ENOMEM);
821 
822     if (s->dumpfile) {
823         s->analysis_rdft = av_rdft_init(rdft_bits, DFT_R2C);
824         s->dump_buf = av_malloc_array(s->analysis_rdft_len, sizeof(*s->dump_buf));
825     }
826 
827     s->analysis_buf = av_malloc_array(s->analysis_rdft_len, sizeof(*s->analysis_buf));
828     s->kernel_tmp_buf = av_malloc_array(s->rdft_len * (s->multi ? inlink->channels : 1), sizeof(*s->kernel_tmp_buf));
829     s->kernel_buf = av_malloc_array(s->rdft_len * (s->multi ? inlink->channels : 1), sizeof(*s->kernel_buf));
830     s->conv_buf   = av_calloc(2 * s->rdft_len * inlink->channels, sizeof(*s->conv_buf));
831     s->conv_idx   = av_calloc(inlink->channels, sizeof(*s->conv_idx));
832     if (!s->analysis_buf || !s->kernel_tmp_buf || !s->kernel_buf || !s->conv_buf || !s->conv_idx)
833         return AVERROR(ENOMEM);
834 
835     av_log(ctx, AV_LOG_DEBUG, "sample_rate = %d, channels = %d, analysis_rdft_len = %d, rdft_len = %d, fir_len = %d, nsamples_max = %d.\n",
836            inlink->sample_rate, inlink->channels, s->analysis_rdft_len, s->rdft_len, s->fir_len, s->nsamples_max);
837 
838     if (s->fixed)
839         inlink->min_samples = inlink->max_samples = inlink->partial_buf_size = s->nsamples_max;
840 
841     return generate_kernel(ctx, SELECT_GAIN(s), SELECT_GAIN_ENTRY(s));
842 }
843 
filter_frame(AVFilterLink * inlink,AVFrame * frame)844 static int filter_frame(AVFilterLink *inlink, AVFrame *frame)
845 {
846     AVFilterContext *ctx = inlink->dst;
847     FIREqualizerContext *s = ctx->priv;
848     int ch;
849 
850     if (!s->min_phase) {
851         for (ch = 0; ch + 1 < inlink->channels && s->fft_ctx; ch += 2) {
852             fast_convolute2(s, s->kernel_buf, (FFTComplex *)(s->conv_buf + 2 * ch * s->rdft_len),
853                             s->conv_idx + ch, (float *) frame->extended_data[ch],
854                             (float *) frame->extended_data[ch+1], frame->nb_samples);
855         }
856 
857         for ( ; ch < inlink->channels; ch++) {
858             fast_convolute(s, s->kernel_buf + (s->multi ? ch * s->rdft_len : 0),
859                         s->conv_buf + 2 * ch * s->rdft_len, s->conv_idx + ch,
860                         (float *) frame->extended_data[ch], frame->nb_samples);
861         }
862     } else {
863         for (ch = 0; ch < inlink->channels; ch++) {
864             fast_convolute_nonlinear(s, s->kernel_buf + (s->multi ? ch * s->rdft_len : 0),
865                                      s->conv_buf + 2 * ch * s->rdft_len, s->conv_idx + ch,
866                                      (float *) frame->extended_data[ch], frame->nb_samples);
867         }
868     }
869 
870     s->next_pts = AV_NOPTS_VALUE;
871     if (frame->pts != AV_NOPTS_VALUE) {
872         s->next_pts = frame->pts + av_rescale_q(frame->nb_samples, av_make_q(1, inlink->sample_rate), inlink->time_base);
873         if (s->zero_phase && !s->min_phase)
874             frame->pts -= av_rescale_q(s->fir_len/2, av_make_q(1, inlink->sample_rate), inlink->time_base);
875     }
876     s->frame_nsamples_max = FFMAX(s->frame_nsamples_max, frame->nb_samples);
877     return ff_filter_frame(ctx->outputs[0], frame);
878 }
879 
request_frame(AVFilterLink * outlink)880 static int request_frame(AVFilterLink *outlink)
881 {
882     AVFilterContext *ctx = outlink->src;
883     FIREqualizerContext *s= ctx->priv;
884     int ret;
885 
886     ret = ff_request_frame(ctx->inputs[0]);
887     if (ret == AVERROR_EOF && s->remaining > 0 && s->frame_nsamples_max > 0) {
888         AVFrame *frame = ff_get_audio_buffer(outlink, FFMIN(s->remaining, s->frame_nsamples_max));
889 
890         if (!frame)
891             return AVERROR(ENOMEM);
892 
893         av_samples_set_silence(frame->extended_data, 0, frame->nb_samples, outlink->channels, frame->format);
894         frame->pts = s->next_pts;
895         s->remaining -= frame->nb_samples;
896         ret = filter_frame(ctx->inputs[0], frame);
897     }
898 
899     return ret;
900 }
901 
process_command(AVFilterContext * ctx,const char * cmd,const char * args,char * res,int res_len,int flags)902 static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
903                            char *res, int res_len, int flags)
904 {
905     FIREqualizerContext *s = ctx->priv;
906     int ret = AVERROR(ENOSYS);
907 
908     if (!strcmp(cmd, "gain")) {
909         char *gain_cmd;
910 
911         if (SELECT_GAIN(s) && !strcmp(SELECT_GAIN(s), args)) {
912             av_log(ctx, AV_LOG_DEBUG, "equal gain, do not rebuild.\n");
913             return 0;
914         }
915 
916         gain_cmd = av_strdup(args);
917         if (!gain_cmd)
918             return AVERROR(ENOMEM);
919 
920         ret = generate_kernel(ctx, gain_cmd, SELECT_GAIN_ENTRY(s));
921         if (ret >= 0) {
922             av_freep(&s->gain_cmd);
923             s->gain_cmd = gain_cmd;
924         } else {
925             av_freep(&gain_cmd);
926         }
927     } else if (!strcmp(cmd, "gain_entry")) {
928         char *gain_entry_cmd;
929 
930         if (SELECT_GAIN_ENTRY(s) && !strcmp(SELECT_GAIN_ENTRY(s), args)) {
931             av_log(ctx, AV_LOG_DEBUG, "equal gain_entry, do not rebuild.\n");
932             return 0;
933         }
934 
935         gain_entry_cmd = av_strdup(args);
936         if (!gain_entry_cmd)
937             return AVERROR(ENOMEM);
938 
939         ret = generate_kernel(ctx, SELECT_GAIN(s), gain_entry_cmd);
940         if (ret >= 0) {
941             av_freep(&s->gain_entry_cmd);
942             s->gain_entry_cmd = gain_entry_cmd;
943         } else {
944             av_freep(&gain_entry_cmd);
945         }
946     }
947 
948     return ret;
949 }
950 
951 static const AVFilterPad firequalizer_inputs[] = {
952     {
953         .name           = "default",
954         .config_props   = config_input,
955         .filter_frame   = filter_frame,
956         .type           = AVMEDIA_TYPE_AUDIO,
957         .needs_writable = 1,
958     },
959     { NULL }
960 };
961 
962 static const AVFilterPad firequalizer_outputs[] = {
963     {
964         .name           = "default",
965         .request_frame  = request_frame,
966         .type           = AVMEDIA_TYPE_AUDIO,
967     },
968     { NULL }
969 };
970 
971 AVFilter ff_af_firequalizer = {
972     .name               = "firequalizer",
973     .description        = NULL_IF_CONFIG_SMALL("Finite Impulse Response Equalizer."),
974     .uninit             = uninit,
975     .query_formats      = query_formats,
976     .process_command    = process_command,
977     .priv_size          = sizeof(FIREqualizerContext),
978     .inputs             = firequalizer_inputs,
979     .outputs            = firequalizer_outputs,
980     .priv_class         = &firequalizer_class,
981 };
982