1 /*****************************************************************************
2 * sofalizer.c : SOFAlizer filter for virtual binaural acoustics
3 *****************************************************************************
4 * Copyright (C) 2013-2015 Andreas Fuchs, Wolfgang Hrauda,
5 * Acoustics Research Institute (ARI), Vienna, Austria
6 *
7 * Authors: Andreas Fuchs <andi.fuchs.mail@gmail.com>
8 * Wolfgang Hrauda <wolfgang.hrauda@gmx.at>
9 *
10 * SOFAlizer project coordinator at ARI, main developer of SOFA:
11 * Piotr Majdak <piotr@majdak.at>
12 *
13 * This program is free software; you can redistribute it and/or modify it
14 * under the terms of the GNU Lesser General Public License as published by
15 * the Free Software Foundation; either version 2.1 of the License, or
16 * (at your option) any later version.
17 *
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU Lesser General Public License for more details.
22 *
23 * You should have received a copy of the GNU Lesser General Public License
24 * along with this program; if not, write to the Free Software Foundation,
25 * Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
26 *****************************************************************************/
27
28 #include <math.h>
29 #include <mysofa.h>
30
31 #include "libavcodec/avfft.h"
32 #include "libavutil/avstring.h"
33 #include "libavutil/channel_layout.h"
34 #include "libavutil/float_dsp.h"
35 #include "libavutil/intmath.h"
36 #include "libavutil/opt.h"
37 #include "avfilter.h"
38 #include "filters.h"
39 #include "internal.h"
40 #include "audio.h"
41
42 #define TIME_DOMAIN 0
43 #define FREQUENCY_DOMAIN 1
44
45 typedef struct MySofa { /* contains data of one SOFA file */
46 struct MYSOFA_HRTF *hrtf;
47 struct MYSOFA_LOOKUP *lookup;
48 struct MYSOFA_NEIGHBORHOOD *neighborhood;
49 int ir_samples; /* length of one impulse response (IR) */
50 int n_samples; /* ir_samples to next power of 2 */
51 float *lir, *rir; /* IRs (time-domain) */
52 float *fir;
53 int max_delay;
54 } MySofa;
55
56 typedef struct VirtualSpeaker {
57 uint8_t set;
58 float azim;
59 float elev;
60 } VirtualSpeaker;
61
62 typedef struct SOFAlizerContext {
63 const AVClass *class;
64
65 char *filename; /* name of SOFA file */
66 MySofa sofa; /* contains data of the SOFA file */
67
68 int sample_rate; /* sample rate from SOFA file */
69 float *speaker_azim; /* azimuth of the virtual loudspeakers */
70 float *speaker_elev; /* elevation of the virtual loudspeakers */
71 char *speakers_pos; /* custom positions of the virtual loudspeakers */
72 float lfe_gain; /* initial gain for the LFE channel */
73 float gain_lfe; /* gain applied to LFE channel */
74 int lfe_channel; /* LFE channel position in channel layout */
75
76 int n_conv; /* number of channels to convolute */
77
78 /* buffer variables (for convolution) */
79 float *ringbuffer[2]; /* buffers input samples, length of one buffer: */
80 /* no. input ch. (incl. LFE) x buffer_length */
81 int write[2]; /* current write position to ringbuffer */
82 int buffer_length; /* is: longest IR plus max. delay in all SOFA files */
83 /* then choose next power of 2 */
84 int n_fft; /* number of samples in one FFT block */
85 int nb_samples;
86
87 /* netCDF variables */
88 int *delay[2]; /* broadband delay for each channel/IR to be convolved */
89
90 float *data_ir[2]; /* IRs for all channels to be convolved */
91 /* (this excludes the LFE) */
92 float *temp_src[2];
93 FFTComplex *temp_fft[2]; /* Array to hold FFT values */
94 FFTComplex *temp_afft[2]; /* Array to accumulate FFT values prior to IFFT */
95
96 /* control variables */
97 float gain; /* filter gain (in dB) */
98 float rotation; /* rotation of virtual loudspeakers (in degrees) */
99 float elevation; /* elevation of virtual loudspeakers (in deg.) */
100 float radius; /* distance virtual loudspeakers to listener (in metres) */
101 int type; /* processing type */
102 int framesize; /* size of buffer */
103 int normalize; /* should all IRs be normalized upon import ? */
104 int interpolate; /* should wanted IRs be interpolated from neighbors ? */
105 int minphase; /* should all IRs be minphased upon import ? */
106 float anglestep; /* neighbor search angle step, in agles */
107 float radstep; /* neighbor search radius step, in meters */
108
109 VirtualSpeaker vspkrpos[64];
110
111 FFTContext *fft[2], *ifft[2];
112 FFTComplex *data_hrtf[2];
113
114 AVFloatDSPContext *fdsp;
115 } SOFAlizerContext;
116
close_sofa(struct MySofa * sofa)117 static int close_sofa(struct MySofa *sofa)
118 {
119 if (sofa->neighborhood)
120 mysofa_neighborhood_free(sofa->neighborhood);
121 sofa->neighborhood = NULL;
122 if (sofa->lookup)
123 mysofa_lookup_free(sofa->lookup);
124 sofa->lookup = NULL;
125 if (sofa->hrtf)
126 mysofa_free(sofa->hrtf);
127 sofa->hrtf = NULL;
128 av_freep(&sofa->fir);
129
130 return 0;
131 }
132
preload_sofa(AVFilterContext * ctx,char * filename,int * samplingrate)133 static int preload_sofa(AVFilterContext *ctx, char *filename, int *samplingrate)
134 {
135 struct SOFAlizerContext *s = ctx->priv;
136 struct MYSOFA_HRTF *mysofa;
137 char *license;
138 int ret;
139
140 mysofa = mysofa_load(filename, &ret);
141 s->sofa.hrtf = mysofa;
142 if (ret || !mysofa) {
143 av_log(ctx, AV_LOG_ERROR, "Can't find SOFA-file '%s'\n", filename);
144 return AVERROR(EINVAL);
145 }
146
147 ret = mysofa_check(mysofa);
148 if (ret != MYSOFA_OK) {
149 av_log(ctx, AV_LOG_ERROR, "Selected SOFA file is invalid. Please select valid SOFA file.\n");
150 return ret;
151 }
152
153 if (s->normalize)
154 mysofa_loudness(s->sofa.hrtf);
155
156 if (s->minphase)
157 mysofa_minphase(s->sofa.hrtf, 0.01f);
158
159 mysofa_tocartesian(s->sofa.hrtf);
160
161 s->sofa.lookup = mysofa_lookup_init(s->sofa.hrtf);
162 if (s->sofa.lookup == NULL)
163 return AVERROR(EINVAL);
164
165 if (s->interpolate)
166 s->sofa.neighborhood = mysofa_neighborhood_init_withstepdefine(s->sofa.hrtf,
167 s->sofa.lookup,
168 s->anglestep,
169 s->radstep);
170
171 s->sofa.fir = av_calloc(s->sofa.hrtf->N * s->sofa.hrtf->R, sizeof(*s->sofa.fir));
172 if (!s->sofa.fir)
173 return AVERROR(ENOMEM);
174
175 if (mysofa->DataSamplingRate.elements != 1)
176 return AVERROR(EINVAL);
177 av_log(ctx, AV_LOG_DEBUG, "Original IR length: %d.\n", mysofa->N);
178 *samplingrate = mysofa->DataSamplingRate.values[0];
179 license = mysofa_getAttribute(mysofa->attributes, (char *)"License");
180 if (license)
181 av_log(ctx, AV_LOG_INFO, "SOFA license: %s\n", license);
182
183 return 0;
184 }
185
parse_channel_name(char ** arg,int * rchannel,char * buf)186 static int parse_channel_name(char **arg, int *rchannel, char *buf)
187 {
188 int len, i, channel_id = 0;
189 int64_t layout, layout0;
190
191 /* try to parse a channel name, e.g. "FL" */
192 if (av_sscanf(*arg, "%7[A-Z]%n", buf, &len)) {
193 layout0 = layout = av_get_channel_layout(buf);
194 /* channel_id <- first set bit in layout */
195 for (i = 32; i > 0; i >>= 1) {
196 if (layout >= 1LL << i) {
197 channel_id += i;
198 layout >>= i;
199 }
200 }
201 /* reject layouts that are not a single channel */
202 if (channel_id >= 64 || layout0 != 1LL << channel_id)
203 return AVERROR(EINVAL);
204 *rchannel = channel_id;
205 *arg += len;
206 return 0;
207 }
208 return AVERROR(EINVAL);
209 }
210
parse_speaker_pos(AVFilterContext * ctx,int64_t in_channel_layout)211 static void parse_speaker_pos(AVFilterContext *ctx, int64_t in_channel_layout)
212 {
213 SOFAlizerContext *s = ctx->priv;
214 char *arg, *tokenizer, *p, *args = av_strdup(s->speakers_pos);
215
216 if (!args)
217 return;
218 p = args;
219
220 while ((arg = av_strtok(p, "|", &tokenizer))) {
221 char buf[8];
222 float azim, elev;
223 int out_ch_id;
224
225 p = NULL;
226 if (parse_channel_name(&arg, &out_ch_id, buf)) {
227 av_log(ctx, AV_LOG_WARNING, "Failed to parse \'%s\' as channel name.\n", buf);
228 continue;
229 }
230 if (av_sscanf(arg, "%f %f", &azim, &elev) == 2) {
231 s->vspkrpos[out_ch_id].set = 1;
232 s->vspkrpos[out_ch_id].azim = azim;
233 s->vspkrpos[out_ch_id].elev = elev;
234 } else if (av_sscanf(arg, "%f", &azim) == 1) {
235 s->vspkrpos[out_ch_id].set = 1;
236 s->vspkrpos[out_ch_id].azim = azim;
237 s->vspkrpos[out_ch_id].elev = 0;
238 }
239 }
240
241 av_free(args);
242 }
243
get_speaker_pos(AVFilterContext * ctx,float * speaker_azim,float * speaker_elev)244 static int get_speaker_pos(AVFilterContext *ctx,
245 float *speaker_azim, float *speaker_elev)
246 {
247 struct SOFAlizerContext *s = ctx->priv;
248 uint64_t channels_layout = ctx->inputs[0]->channel_layout;
249 float azim[16] = { 0 };
250 float elev[16] = { 0 };
251 int m, ch, n_conv = ctx->inputs[0]->channels; /* get no. input channels */
252
253 if (n_conv > 16)
254 return AVERROR(EINVAL);
255
256 s->lfe_channel = -1;
257
258 if (s->speakers_pos)
259 parse_speaker_pos(ctx, channels_layout);
260
261 /* set speaker positions according to input channel configuration: */
262 for (m = 0, ch = 0; ch < n_conv && m < 64; m++) {
263 uint64_t mask = channels_layout & (1ULL << m);
264
265 switch (mask) {
266 case AV_CH_FRONT_LEFT: azim[ch] = 30; break;
267 case AV_CH_FRONT_RIGHT: azim[ch] = 330; break;
268 case AV_CH_FRONT_CENTER: azim[ch] = 0; break;
269 case AV_CH_LOW_FREQUENCY:
270 case AV_CH_LOW_FREQUENCY_2: s->lfe_channel = ch; break;
271 case AV_CH_BACK_LEFT: azim[ch] = 150; break;
272 case AV_CH_BACK_RIGHT: azim[ch] = 210; break;
273 case AV_CH_BACK_CENTER: azim[ch] = 180; break;
274 case AV_CH_SIDE_LEFT: azim[ch] = 90; break;
275 case AV_CH_SIDE_RIGHT: azim[ch] = 270; break;
276 case AV_CH_FRONT_LEFT_OF_CENTER: azim[ch] = 15; break;
277 case AV_CH_FRONT_RIGHT_OF_CENTER: azim[ch] = 345; break;
278 case AV_CH_TOP_CENTER: azim[ch] = 0;
279 elev[ch] = 90; break;
280 case AV_CH_TOP_FRONT_LEFT: azim[ch] = 30;
281 elev[ch] = 45; break;
282 case AV_CH_TOP_FRONT_CENTER: azim[ch] = 0;
283 elev[ch] = 45; break;
284 case AV_CH_TOP_FRONT_RIGHT: azim[ch] = 330;
285 elev[ch] = 45; break;
286 case AV_CH_TOP_BACK_LEFT: azim[ch] = 150;
287 elev[ch] = 45; break;
288 case AV_CH_TOP_BACK_RIGHT: azim[ch] = 210;
289 elev[ch] = 45; break;
290 case AV_CH_TOP_BACK_CENTER: azim[ch] = 180;
291 elev[ch] = 45; break;
292 case AV_CH_WIDE_LEFT: azim[ch] = 90; break;
293 case AV_CH_WIDE_RIGHT: azim[ch] = 270; break;
294 case AV_CH_SURROUND_DIRECT_LEFT: azim[ch] = 90; break;
295 case AV_CH_SURROUND_DIRECT_RIGHT: azim[ch] = 270; break;
296 case AV_CH_STEREO_LEFT: azim[ch] = 90; break;
297 case AV_CH_STEREO_RIGHT: azim[ch] = 270; break;
298 case 0: break;
299 default:
300 return AVERROR(EINVAL);
301 }
302
303 if (s->vspkrpos[m].set) {
304 azim[ch] = s->vspkrpos[m].azim;
305 elev[ch] = s->vspkrpos[m].elev;
306 }
307
308 if (mask)
309 ch++;
310 }
311
312 memcpy(speaker_azim, azim, n_conv * sizeof(float));
313 memcpy(speaker_elev, elev, n_conv * sizeof(float));
314
315 return 0;
316
317 }
318
319 typedef struct ThreadData {
320 AVFrame *in, *out;
321 int *write;
322 int **delay;
323 float **ir;
324 int *n_clippings;
325 float **ringbuffer;
326 float **temp_src;
327 FFTComplex **temp_fft;
328 FFTComplex **temp_afft;
329 } ThreadData;
330
sofalizer_convolute(AVFilterContext * ctx,void * arg,int jobnr,int nb_jobs)331 static int sofalizer_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
332 {
333 SOFAlizerContext *s = ctx->priv;
334 ThreadData *td = arg;
335 AVFrame *in = td->in, *out = td->out;
336 int offset = jobnr;
337 int *write = &td->write[jobnr];
338 const int *const delay = td->delay[jobnr];
339 const float *const ir = td->ir[jobnr];
340 int *n_clippings = &td->n_clippings[jobnr];
341 float *ringbuffer = td->ringbuffer[jobnr];
342 float *temp_src = td->temp_src[jobnr];
343 const int ir_samples = s->sofa.ir_samples; /* length of one IR */
344 const int n_samples = s->sofa.n_samples;
345 const int planar = in->format == AV_SAMPLE_FMT_FLTP;
346 const int mult = 1 + !planar;
347 const float *src = (const float *)in->extended_data[0]; /* get pointer to audio input buffer */
348 float *dst = (float *)out->extended_data[jobnr * planar]; /* get pointer to audio output buffer */
349 const int in_channels = s->n_conv; /* number of input channels */
350 /* ring buffer length is: longest IR plus max. delay -> next power of 2 */
351 const int buffer_length = s->buffer_length;
352 /* -1 for AND instead of MODULO (applied to powers of 2): */
353 const uint32_t modulo = (uint32_t)buffer_length - 1;
354 float *buffer[16]; /* holds ringbuffer for each input channel */
355 int wr = *write;
356 int read;
357 int i, l;
358
359 if (!planar)
360 dst += offset;
361
362 for (l = 0; l < in_channels; l++) {
363 /* get starting address of ringbuffer for each input channel */
364 buffer[l] = ringbuffer + l * buffer_length;
365 }
366
367 for (i = 0; i < in->nb_samples; i++) {
368 const float *temp_ir = ir; /* using same set of IRs for each sample */
369
370 dst[0] = 0;
371 if (planar) {
372 for (l = 0; l < in_channels; l++) {
373 const float *srcp = (const float *)in->extended_data[l];
374
375 /* write current input sample to ringbuffer (for each channel) */
376 buffer[l][wr] = srcp[i];
377 }
378 } else {
379 for (l = 0; l < in_channels; l++) {
380 /* write current input sample to ringbuffer (for each channel) */
381 buffer[l][wr] = src[l];
382 }
383 }
384
385 /* loop goes through all channels to be convolved */
386 for (l = 0; l < in_channels; l++) {
387 const float *const bptr = buffer[l];
388
389 if (l == s->lfe_channel) {
390 /* LFE is an input channel but requires no convolution */
391 /* apply gain to LFE signal and add to output buffer */
392 dst[0] += *(buffer[s->lfe_channel] + wr) * s->gain_lfe;
393 temp_ir += n_samples;
394 continue;
395 }
396
397 /* current read position in ringbuffer: input sample write position
398 * - delay for l-th ch. + diff. betw. IR length and buffer length
399 * (mod buffer length) */
400 read = (wr - delay[l] - (ir_samples - 1) + buffer_length) & modulo;
401
402 if (read + ir_samples < buffer_length) {
403 memmove(temp_src, bptr + read, ir_samples * sizeof(*temp_src));
404 } else {
405 int len = FFMIN(n_samples - (read % ir_samples), buffer_length - read);
406
407 memmove(temp_src, bptr + read, len * sizeof(*temp_src));
408 memmove(temp_src + len, bptr, (n_samples - len) * sizeof(*temp_src));
409 }
410
411 /* multiply signal and IR, and add up the results */
412 dst[0] += s->fdsp->scalarproduct_float(temp_ir, temp_src, FFALIGN(ir_samples, 32));
413 temp_ir += n_samples;
414 }
415
416 /* clippings counter */
417 if (fabsf(dst[0]) > 1)
418 n_clippings[0]++;
419
420 /* move output buffer pointer by +2 to get to next sample of processed channel: */
421 dst += mult;
422 src += in_channels;
423 wr = (wr + 1) & modulo; /* update ringbuffer write position */
424 }
425
426 *write = wr; /* remember write position in ringbuffer for next call */
427
428 return 0;
429 }
430
sofalizer_fast_convolute(AVFilterContext * ctx,void * arg,int jobnr,int nb_jobs)431 static int sofalizer_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
432 {
433 SOFAlizerContext *s = ctx->priv;
434 ThreadData *td = arg;
435 AVFrame *in = td->in, *out = td->out;
436 int offset = jobnr;
437 int *write = &td->write[jobnr];
438 FFTComplex *hrtf = s->data_hrtf[jobnr]; /* get pointers to current HRTF data */
439 int *n_clippings = &td->n_clippings[jobnr];
440 float *ringbuffer = td->ringbuffer[jobnr];
441 const int ir_samples = s->sofa.ir_samples; /* length of one IR */
442 const int planar = in->format == AV_SAMPLE_FMT_FLTP;
443 const int mult = 1 + !planar;
444 float *dst = (float *)out->extended_data[jobnr * planar]; /* get pointer to audio output buffer */
445 const int in_channels = s->n_conv; /* number of input channels */
446 /* ring buffer length is: longest IR plus max. delay -> next power of 2 */
447 const int buffer_length = s->buffer_length;
448 /* -1 for AND instead of MODULO (applied to powers of 2): */
449 const uint32_t modulo = (uint32_t)buffer_length - 1;
450 FFTComplex *fft_in = s->temp_fft[jobnr]; /* temporary array for FFT input/output data */
451 FFTComplex *fft_acc = s->temp_afft[jobnr];
452 FFTContext *ifft = s->ifft[jobnr];
453 FFTContext *fft = s->fft[jobnr];
454 const int n_conv = s->n_conv;
455 const int n_fft = s->n_fft;
456 const float fft_scale = 1.0f / s->n_fft;
457 FFTComplex *hrtf_offset;
458 int wr = *write;
459 int n_read;
460 int i, j;
461
462 if (!planar)
463 dst += offset;
464
465 /* find minimum between number of samples and output buffer length:
466 * (important, if one IR is longer than the output buffer) */
467 n_read = FFMIN(ir_samples, in->nb_samples);
468 for (j = 0; j < n_read; j++) {
469 /* initialize output buf with saved signal from overflow buf */
470 dst[mult * j] = ringbuffer[wr];
471 ringbuffer[wr] = 0.0f; /* re-set read samples to zero */
472 /* update ringbuffer read/write position */
473 wr = (wr + 1) & modulo;
474 }
475
476 /* initialize rest of output buffer with 0 */
477 for (j = n_read; j < in->nb_samples; j++) {
478 dst[mult * j] = 0;
479 }
480
481 /* fill FFT accumulation with 0 */
482 memset(fft_acc, 0, sizeof(FFTComplex) * n_fft);
483
484 for (i = 0; i < n_conv; i++) {
485 const float *src = (const float *)in->extended_data[i * planar]; /* get pointer to audio input buffer */
486
487 if (i == s->lfe_channel) { /* LFE */
488 if (in->format == AV_SAMPLE_FMT_FLT) {
489 for (j = 0; j < in->nb_samples; j++) {
490 /* apply gain to LFE signal and add to output buffer */
491 dst[2 * j] += src[i + j * in_channels] * s->gain_lfe;
492 }
493 } else {
494 for (j = 0; j < in->nb_samples; j++) {
495 /* apply gain to LFE signal and add to output buffer */
496 dst[j] += src[j] * s->gain_lfe;
497 }
498 }
499 continue;
500 }
501
502 /* outer loop: go through all input channels to be convolved */
503 offset = i * n_fft; /* no. samples already processed */
504 hrtf_offset = hrtf + offset;
505
506 /* fill FFT input with 0 (we want to zero-pad) */
507 memset(fft_in, 0, sizeof(FFTComplex) * n_fft);
508
509 if (in->format == AV_SAMPLE_FMT_FLT) {
510 for (j = 0; j < in->nb_samples; j++) {
511 /* prepare input for FFT */
512 /* write all samples of current input channel to FFT input array */
513 fft_in[j].re = src[j * in_channels + i];
514 }
515 } else {
516 for (j = 0; j < in->nb_samples; j++) {
517 /* prepare input for FFT */
518 /* write all samples of current input channel to FFT input array */
519 fft_in[j].re = src[j];
520 }
521 }
522
523 /* transform input signal of current channel to frequency domain */
524 av_fft_permute(fft, fft_in);
525 av_fft_calc(fft, fft_in);
526 for (j = 0; j < n_fft; j++) {
527 const FFTComplex *hcomplex = hrtf_offset + j;
528 const float re = fft_in[j].re;
529 const float im = fft_in[j].im;
530
531 /* complex multiplication of input signal and HRTFs */
532 /* output channel (real): */
533 fft_acc[j].re += re * hcomplex->re - im * hcomplex->im;
534 /* output channel (imag): */
535 fft_acc[j].im += re * hcomplex->im + im * hcomplex->re;
536 }
537 }
538
539 /* transform output signal of current channel back to time domain */
540 av_fft_permute(ifft, fft_acc);
541 av_fft_calc(ifft, fft_acc);
542
543 for (j = 0; j < in->nb_samples; j++) {
544 /* write output signal of current channel to output buffer */
545 dst[mult * j] += fft_acc[j].re * fft_scale;
546 }
547
548 for (j = 0; j < ir_samples - 1; j++) { /* overflow length is IR length - 1 */
549 /* write the rest of output signal to overflow buffer */
550 int write_pos = (wr + j) & modulo;
551
552 *(ringbuffer + write_pos) += fft_acc[in->nb_samples + j].re * fft_scale;
553 }
554
555 /* go through all samples of current output buffer: count clippings */
556 for (i = 0; i < out->nb_samples; i++) {
557 /* clippings counter */
558 if (fabsf(dst[i * mult]) > 1) { /* if current output sample > 1 */
559 n_clippings[0]++;
560 }
561 }
562
563 /* remember read/write position in ringbuffer for next call */
564 *write = wr;
565
566 return 0;
567 }
568
filter_frame(AVFilterLink * inlink,AVFrame * in)569 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
570 {
571 AVFilterContext *ctx = inlink->dst;
572 SOFAlizerContext *s = ctx->priv;
573 AVFilterLink *outlink = ctx->outputs[0];
574 int n_clippings[2] = { 0 };
575 ThreadData td;
576 AVFrame *out;
577
578 out = ff_get_audio_buffer(outlink, in->nb_samples);
579 if (!out) {
580 av_frame_free(&in);
581 return AVERROR(ENOMEM);
582 }
583 av_frame_copy_props(out, in);
584
585 td.in = in; td.out = out; td.write = s->write;
586 td.delay = s->delay; td.ir = s->data_ir; td.n_clippings = n_clippings;
587 td.ringbuffer = s->ringbuffer; td.temp_src = s->temp_src;
588 td.temp_fft = s->temp_fft;
589 td.temp_afft = s->temp_afft;
590
591 if (s->type == TIME_DOMAIN) {
592 ctx->internal->execute(ctx, sofalizer_convolute, &td, NULL, 2);
593 } else if (s->type == FREQUENCY_DOMAIN) {
594 ctx->internal->execute(ctx, sofalizer_fast_convolute, &td, NULL, 2);
595 }
596 emms_c();
597
598 /* display error message if clipping occurred */
599 if (n_clippings[0] + n_clippings[1] > 0) {
600 av_log(ctx, AV_LOG_WARNING, "%d of %d samples clipped. Please reduce gain.\n",
601 n_clippings[0] + n_clippings[1], out->nb_samples * 2);
602 }
603
604 av_frame_free(&in);
605 return ff_filter_frame(outlink, out);
606 }
607
activate(AVFilterContext * ctx)608 static int activate(AVFilterContext *ctx)
609 {
610 AVFilterLink *inlink = ctx->inputs[0];
611 AVFilterLink *outlink = ctx->outputs[0];
612 SOFAlizerContext *s = ctx->priv;
613 AVFrame *in;
614 int ret;
615
616 FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
617
618 if (s->nb_samples)
619 ret = ff_inlink_consume_samples(inlink, s->nb_samples, s->nb_samples, &in);
620 else
621 ret = ff_inlink_consume_frame(inlink, &in);
622 if (ret < 0)
623 return ret;
624 if (ret > 0)
625 return filter_frame(inlink, in);
626
627 FF_FILTER_FORWARD_STATUS(inlink, outlink);
628 FF_FILTER_FORWARD_WANTED(outlink, inlink);
629
630 return FFERROR_NOT_READY;
631 }
632
query_formats(AVFilterContext * ctx)633 static int query_formats(AVFilterContext *ctx)
634 {
635 struct SOFAlizerContext *s = ctx->priv;
636 AVFilterFormats *formats = NULL;
637 AVFilterChannelLayouts *layouts = NULL;
638 int ret, sample_rates[] = { 48000, -1 };
639 static const enum AVSampleFormat sample_fmts[] = {
640 AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
641 AV_SAMPLE_FMT_NONE
642 };
643
644 formats = ff_make_format_list(sample_fmts);
645 if (!formats)
646 return AVERROR(ENOMEM);
647 ret = ff_set_common_formats(ctx, formats);
648 if (ret)
649 return ret;
650
651 layouts = ff_all_channel_layouts();
652 if (!layouts)
653 return AVERROR(ENOMEM);
654
655 ret = ff_channel_layouts_ref(layouts, &ctx->inputs[0]->out_channel_layouts);
656 if (ret)
657 return ret;
658
659 layouts = NULL;
660 ret = ff_add_channel_layout(&layouts, AV_CH_LAYOUT_STEREO);
661 if (ret)
662 return ret;
663
664 ret = ff_channel_layouts_ref(layouts, &ctx->outputs[0]->in_channel_layouts);
665 if (ret)
666 return ret;
667
668 sample_rates[0] = s->sample_rate;
669 formats = ff_make_format_list(sample_rates);
670 if (!formats)
671 return AVERROR(ENOMEM);
672 return ff_set_common_samplerates(ctx, formats);
673 }
674
getfilter_float(AVFilterContext * ctx,float x,float y,float z,float * left,float * right,float * delay_left,float * delay_right)675 static int getfilter_float(AVFilterContext *ctx, float x, float y, float z,
676 float *left, float *right,
677 float *delay_left, float *delay_right)
678 {
679 struct SOFAlizerContext *s = ctx->priv;
680 float c[3], delays[2];
681 float *fl, *fr;
682 int nearest;
683 int *neighbors;
684 float *res;
685
686 c[0] = x, c[1] = y, c[2] = z;
687 nearest = mysofa_lookup(s->sofa.lookup, c);
688 if (nearest < 0)
689 return AVERROR(EINVAL);
690
691 if (s->interpolate) {
692 neighbors = mysofa_neighborhood(s->sofa.neighborhood, nearest);
693 res = mysofa_interpolate(s->sofa.hrtf, c,
694 nearest, neighbors,
695 s->sofa.fir, delays);
696 } else {
697 if (s->sofa.hrtf->DataDelay.elements > s->sofa.hrtf->R) {
698 delays[0] = s->sofa.hrtf->DataDelay.values[nearest * s->sofa.hrtf->R];
699 delays[1] = s->sofa.hrtf->DataDelay.values[nearest * s->sofa.hrtf->R + 1];
700 } else {
701 delays[0] = s->sofa.hrtf->DataDelay.values[0];
702 delays[1] = s->sofa.hrtf->DataDelay.values[1];
703 }
704 res = s->sofa.hrtf->DataIR.values + nearest * s->sofa.hrtf->N * s->sofa.hrtf->R;
705 }
706
707 *delay_left = delays[0];
708 *delay_right = delays[1];
709
710 fl = res;
711 fr = res + s->sofa.hrtf->N;
712
713 memcpy(left, fl, sizeof(float) * s->sofa.hrtf->N);
714 memcpy(right, fr, sizeof(float) * s->sofa.hrtf->N);
715
716 return 0;
717 }
718
load_data(AVFilterContext * ctx,int azim,int elev,float radius,int sample_rate)719 static int load_data(AVFilterContext *ctx, int azim, int elev, float radius, int sample_rate)
720 {
721 struct SOFAlizerContext *s = ctx->priv;
722 int n_samples;
723 int ir_samples;
724 int n_conv = s->n_conv; /* no. channels to convolve */
725 int n_fft;
726 float delay_l; /* broadband delay for each IR */
727 float delay_r;
728 int nb_input_channels = ctx->inputs[0]->channels; /* no. input channels */
729 float gain_lin = expf((s->gain - 3 * nb_input_channels) / 20 * M_LN10); /* gain - 3dB/channel */
730 FFTComplex *data_hrtf_l = NULL;
731 FFTComplex *data_hrtf_r = NULL;
732 FFTComplex *fft_in_l = NULL;
733 FFTComplex *fft_in_r = NULL;
734 float *data_ir_l = NULL;
735 float *data_ir_r = NULL;
736 int offset = 0; /* used for faster pointer arithmetics in for-loop */
737 int i, j, azim_orig = azim, elev_orig = elev;
738 int ret = 0;
739 int n_current;
740 int n_max = 0;
741
742 av_log(ctx, AV_LOG_DEBUG, "IR length: %d.\n", s->sofa.hrtf->N);
743 s->sofa.ir_samples = s->sofa.hrtf->N;
744 s->sofa.n_samples = 1 << (32 - ff_clz(s->sofa.ir_samples));
745
746 n_samples = s->sofa.n_samples;
747 ir_samples = s->sofa.ir_samples;
748
749 if (s->type == TIME_DOMAIN) {
750 s->data_ir[0] = av_calloc(n_samples, sizeof(float) * s->n_conv);
751 s->data_ir[1] = av_calloc(n_samples, sizeof(float) * s->n_conv);
752
753 if (!s->data_ir[0] || !s->data_ir[1]) {
754 ret = AVERROR(ENOMEM);
755 goto fail;
756 }
757 }
758
759 s->delay[0] = av_calloc(s->n_conv, sizeof(int));
760 s->delay[1] = av_calloc(s->n_conv, sizeof(int));
761
762 if (!s->delay[0] || !s->delay[1]) {
763 ret = AVERROR(ENOMEM);
764 goto fail;
765 }
766
767 /* get temporary IR for L and R channel */
768 data_ir_l = av_calloc(n_conv * n_samples, sizeof(*data_ir_l));
769 data_ir_r = av_calloc(n_conv * n_samples, sizeof(*data_ir_r));
770 if (!data_ir_r || !data_ir_l) {
771 ret = AVERROR(ENOMEM);
772 goto fail;
773 }
774
775 if (s->type == TIME_DOMAIN) {
776 s->temp_src[0] = av_calloc(n_samples, sizeof(float));
777 s->temp_src[1] = av_calloc(n_samples, sizeof(float));
778 if (!s->temp_src[0] || !s->temp_src[1]) {
779 ret = AVERROR(ENOMEM);
780 goto fail;
781 }
782 }
783
784 s->speaker_azim = av_calloc(s->n_conv, sizeof(*s->speaker_azim));
785 s->speaker_elev = av_calloc(s->n_conv, sizeof(*s->speaker_elev));
786 if (!s->speaker_azim || !s->speaker_elev) {
787 ret = AVERROR(ENOMEM);
788 goto fail;
789 }
790
791 /* get speaker positions */
792 if ((ret = get_speaker_pos(ctx, s->speaker_azim, s->speaker_elev)) < 0) {
793 av_log(ctx, AV_LOG_ERROR, "Couldn't get speaker positions. Input channel configuration not supported.\n");
794 goto fail;
795 }
796
797 for (i = 0; i < s->n_conv; i++) {
798 float coordinates[3];
799
800 /* load and store IRs and corresponding delays */
801 azim = (int)(s->speaker_azim[i] + azim_orig) % 360;
802 elev = (int)(s->speaker_elev[i] + elev_orig) % 90;
803
804 coordinates[0] = azim;
805 coordinates[1] = elev;
806 coordinates[2] = radius;
807
808 mysofa_s2c(coordinates);
809
810 /* get id of IR closest to desired position */
811 ret = getfilter_float(ctx, coordinates[0], coordinates[1], coordinates[2],
812 data_ir_l + n_samples * i,
813 data_ir_r + n_samples * i,
814 &delay_l, &delay_r);
815 if (ret < 0)
816 goto fail;
817
818 s->delay[0][i] = delay_l * sample_rate;
819 s->delay[1][i] = delay_r * sample_rate;
820
821 s->sofa.max_delay = FFMAX3(s->sofa.max_delay, s->delay[0][i], s->delay[1][i]);
822 }
823
824 /* get size of ringbuffer (longest IR plus max. delay) */
825 /* then choose next power of 2 for performance optimization */
826 n_current = n_samples + s->sofa.max_delay;
827 /* length of longest IR plus max. delay */
828 n_max = FFMAX(n_max, n_current);
829
830 /* buffer length is longest IR plus max. delay -> next power of 2
831 (32 - count leading zeros gives required exponent) */
832 s->buffer_length = 1 << (32 - ff_clz(n_max));
833 s->n_fft = n_fft = 1 << (32 - ff_clz(n_max + s->framesize));
834
835 if (s->type == FREQUENCY_DOMAIN) {
836 av_fft_end(s->fft[0]);
837 av_fft_end(s->fft[1]);
838 s->fft[0] = av_fft_init(av_log2(s->n_fft), 0);
839 s->fft[1] = av_fft_init(av_log2(s->n_fft), 0);
840 av_fft_end(s->ifft[0]);
841 av_fft_end(s->ifft[1]);
842 s->ifft[0] = av_fft_init(av_log2(s->n_fft), 1);
843 s->ifft[1] = av_fft_init(av_log2(s->n_fft), 1);
844
845 if (!s->fft[0] || !s->fft[1] || !s->ifft[0] || !s->ifft[1]) {
846 av_log(ctx, AV_LOG_ERROR, "Unable to create FFT contexts of size %d.\n", s->n_fft);
847 ret = AVERROR(ENOMEM);
848 goto fail;
849 }
850 }
851
852 if (s->type == TIME_DOMAIN) {
853 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
854 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
855 } else if (s->type == FREQUENCY_DOMAIN) {
856 /* get temporary HRTF memory for L and R channel */
857 data_hrtf_l = av_malloc_array(n_fft, sizeof(*data_hrtf_l) * n_conv);
858 data_hrtf_r = av_malloc_array(n_fft, sizeof(*data_hrtf_r) * n_conv);
859 if (!data_hrtf_r || !data_hrtf_l) {
860 ret = AVERROR(ENOMEM);
861 goto fail;
862 }
863
864 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float));
865 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float));
866 s->temp_fft[0] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
867 s->temp_fft[1] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
868 s->temp_afft[0] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
869 s->temp_afft[1] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
870 if (!s->temp_fft[0] || !s->temp_fft[1] ||
871 !s->temp_afft[0] || !s->temp_afft[1]) {
872 ret = AVERROR(ENOMEM);
873 goto fail;
874 }
875 }
876
877 if (!s->ringbuffer[0] || !s->ringbuffer[1]) {
878 ret = AVERROR(ENOMEM);
879 goto fail;
880 }
881
882 if (s->type == FREQUENCY_DOMAIN) {
883 fft_in_l = av_calloc(n_fft, sizeof(*fft_in_l));
884 fft_in_r = av_calloc(n_fft, sizeof(*fft_in_r));
885 if (!fft_in_l || !fft_in_r) {
886 ret = AVERROR(ENOMEM);
887 goto fail;
888 }
889 }
890
891 for (i = 0; i < s->n_conv; i++) {
892 float *lir, *rir;
893
894 offset = i * n_samples; /* no. samples already written */
895
896 lir = data_ir_l + offset;
897 rir = data_ir_r + offset;
898
899 if (s->type == TIME_DOMAIN) {
900 for (j = 0; j < ir_samples; j++) {
901 /* load reversed IRs of the specified source position
902 * sample-by-sample for left and right ear; and apply gain */
903 s->data_ir[0][offset + j] = lir[ir_samples - 1 - j] * gain_lin;
904 s->data_ir[1][offset + j] = rir[ir_samples - 1 - j] * gain_lin;
905 }
906 } else if (s->type == FREQUENCY_DOMAIN) {
907 memset(fft_in_l, 0, n_fft * sizeof(*fft_in_l));
908 memset(fft_in_r, 0, n_fft * sizeof(*fft_in_r));
909
910 offset = i * n_fft; /* no. samples already written */
911 for (j = 0; j < ir_samples; j++) {
912 /* load non-reversed IRs of the specified source position
913 * sample-by-sample and apply gain,
914 * L channel is loaded to real part, R channel to imag part,
915 * IRs are shifted by L and R delay */
916 fft_in_l[s->delay[0][i] + j].re = lir[j] * gain_lin;
917 fft_in_r[s->delay[1][i] + j].re = rir[j] * gain_lin;
918 }
919
920 /* actually transform to frequency domain (IRs -> HRTFs) */
921 av_fft_permute(s->fft[0], fft_in_l);
922 av_fft_calc(s->fft[0], fft_in_l);
923 memcpy(data_hrtf_l + offset, fft_in_l, n_fft * sizeof(*fft_in_l));
924 av_fft_permute(s->fft[0], fft_in_r);
925 av_fft_calc(s->fft[0], fft_in_r);
926 memcpy(data_hrtf_r + offset, fft_in_r, n_fft * sizeof(*fft_in_r));
927 }
928 }
929
930 if (s->type == FREQUENCY_DOMAIN) {
931 s->data_hrtf[0] = av_malloc_array(n_fft * s->n_conv, sizeof(FFTComplex));
932 s->data_hrtf[1] = av_malloc_array(n_fft * s->n_conv, sizeof(FFTComplex));
933 if (!s->data_hrtf[0] || !s->data_hrtf[1]) {
934 ret = AVERROR(ENOMEM);
935 goto fail;
936 }
937
938 memcpy(s->data_hrtf[0], data_hrtf_l, /* copy HRTF data to */
939 sizeof(FFTComplex) * n_conv * n_fft); /* filter struct */
940 memcpy(s->data_hrtf[1], data_hrtf_r,
941 sizeof(FFTComplex) * n_conv * n_fft);
942 }
943
944 fail:
945 av_freep(&data_hrtf_l); /* free temporary HRTF memory */
946 av_freep(&data_hrtf_r);
947
948 av_freep(&data_ir_l); /* free temprary IR memory */
949 av_freep(&data_ir_r);
950
951 av_freep(&fft_in_l); /* free temporary FFT memory */
952 av_freep(&fft_in_r);
953
954 return ret;
955 }
956
init(AVFilterContext * ctx)957 static av_cold int init(AVFilterContext *ctx)
958 {
959 SOFAlizerContext *s = ctx->priv;
960 int ret;
961
962 if (!s->filename) {
963 av_log(ctx, AV_LOG_ERROR, "Valid SOFA filename must be set.\n");
964 return AVERROR(EINVAL);
965 }
966
967 /* preload SOFA file, */
968 ret = preload_sofa(ctx, s->filename, &s->sample_rate);
969 if (ret) {
970 /* file loading error */
971 av_log(ctx, AV_LOG_ERROR, "Error while loading SOFA file: '%s'\n", s->filename);
972 } else { /* no file loading error, resampling not required */
973 av_log(ctx, AV_LOG_DEBUG, "File '%s' loaded.\n", s->filename);
974 }
975
976 if (ret) {
977 av_log(ctx, AV_LOG_ERROR, "No valid SOFA file could be loaded. Please specify valid SOFA file.\n");
978 return ret;
979 }
980
981 s->fdsp = avpriv_float_dsp_alloc(0);
982 if (!s->fdsp)
983 return AVERROR(ENOMEM);
984
985 return 0;
986 }
987
config_input(AVFilterLink * inlink)988 static int config_input(AVFilterLink *inlink)
989 {
990 AVFilterContext *ctx = inlink->dst;
991 SOFAlizerContext *s = ctx->priv;
992 int ret;
993
994 if (s->type == FREQUENCY_DOMAIN)
995 s->nb_samples = s->framesize;
996
997 /* gain -3 dB per channel */
998 s->gain_lfe = expf((s->gain - 3 * inlink->channels + s->lfe_gain) / 20 * M_LN10);
999
1000 s->n_conv = inlink->channels;
1001
1002 /* load IRs to data_ir[0] and data_ir[1] for required directions */
1003 if ((ret = load_data(ctx, s->rotation, s->elevation, s->radius, inlink->sample_rate)) < 0)
1004 return ret;
1005
1006 av_log(ctx, AV_LOG_DEBUG, "Samplerate: %d Channels to convolute: %d, Length of ringbuffer: %d x %d\n",
1007 inlink->sample_rate, s->n_conv, inlink->channels, s->buffer_length);
1008
1009 return 0;
1010 }
1011
uninit(AVFilterContext * ctx)1012 static av_cold void uninit(AVFilterContext *ctx)
1013 {
1014 SOFAlizerContext *s = ctx->priv;
1015
1016 close_sofa(&s->sofa);
1017 av_fft_end(s->ifft[0]);
1018 av_fft_end(s->ifft[1]);
1019 av_fft_end(s->fft[0]);
1020 av_fft_end(s->fft[1]);
1021 s->ifft[0] = NULL;
1022 s->ifft[1] = NULL;
1023 s->fft[0] = NULL;
1024 s->fft[1] = NULL;
1025 av_freep(&s->delay[0]);
1026 av_freep(&s->delay[1]);
1027 av_freep(&s->data_ir[0]);
1028 av_freep(&s->data_ir[1]);
1029 av_freep(&s->ringbuffer[0]);
1030 av_freep(&s->ringbuffer[1]);
1031 av_freep(&s->speaker_azim);
1032 av_freep(&s->speaker_elev);
1033 av_freep(&s->temp_src[0]);
1034 av_freep(&s->temp_src[1]);
1035 av_freep(&s->temp_afft[0]);
1036 av_freep(&s->temp_afft[1]);
1037 av_freep(&s->temp_fft[0]);
1038 av_freep(&s->temp_fft[1]);
1039 av_freep(&s->data_hrtf[0]);
1040 av_freep(&s->data_hrtf[1]);
1041 av_freep(&s->fdsp);
1042 }
1043
1044 #define OFFSET(x) offsetof(SOFAlizerContext, x)
1045 #define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
1046
1047 static const AVOption sofalizer_options[] = {
1048 { "sofa", "sofa filename", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
1049 { "gain", "set gain in dB", OFFSET(gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -20, 40, .flags = FLAGS },
1050 { "rotation", "set rotation" , OFFSET(rotation), AV_OPT_TYPE_FLOAT, {.dbl=0}, -360, 360, .flags = FLAGS },
1051 { "elevation", "set elevation", OFFSET(elevation), AV_OPT_TYPE_FLOAT, {.dbl=0}, -90, 90, .flags = FLAGS },
1052 { "radius", "set radius", OFFSET(radius), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 5, .flags = FLAGS },
1053 { "type", "set processing", OFFSET(type), AV_OPT_TYPE_INT, {.i64=1}, 0, 1, .flags = FLAGS, "type" },
1054 { "time", "time domain", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, .flags = FLAGS, "type" },
1055 { "freq", "frequency domain", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, .flags = FLAGS, "type" },
1056 { "speakers", "set speaker custom positions", OFFSET(speakers_pos), AV_OPT_TYPE_STRING, {.str=0}, 0, 0, .flags = FLAGS },
1057 { "lfegain", "set lfe gain", OFFSET(lfe_gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -20,40, .flags = FLAGS },
1058 { "framesize", "set frame size", OFFSET(framesize), AV_OPT_TYPE_INT, {.i64=1024},1024,96000, .flags = FLAGS },
1059 { "normalize", "normalize IRs", OFFSET(normalize), AV_OPT_TYPE_BOOL, {.i64=1}, 0, 1, .flags = FLAGS },
1060 { "interpolate","interpolate IRs from neighbors", OFFSET(interpolate),AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, .flags = FLAGS },
1061 { "minphase", "minphase IRs", OFFSET(minphase), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, .flags = FLAGS },
1062 { "anglestep", "set neighbor search angle step", OFFSET(anglestep), AV_OPT_TYPE_FLOAT, {.dbl=.5}, 0.01, 10, .flags = FLAGS },
1063 { "radstep", "set neighbor search radius step", OFFSET(radstep), AV_OPT_TYPE_FLOAT, {.dbl=.01}, 0.01, 1, .flags = FLAGS },
1064 { NULL }
1065 };
1066
1067 AVFILTER_DEFINE_CLASS(sofalizer);
1068
1069 static const AVFilterPad inputs[] = {
1070 {
1071 .name = "default",
1072 .type = AVMEDIA_TYPE_AUDIO,
1073 .config_props = config_input,
1074 },
1075 { NULL }
1076 };
1077
1078 static const AVFilterPad outputs[] = {
1079 {
1080 .name = "default",
1081 .type = AVMEDIA_TYPE_AUDIO,
1082 },
1083 { NULL }
1084 };
1085
1086 AVFilter ff_af_sofalizer = {
1087 .name = "sofalizer",
1088 .description = NULL_IF_CONFIG_SMALL("SOFAlizer (Spatially Oriented Format for Acoustics)."),
1089 .priv_size = sizeof(SOFAlizerContext),
1090 .priv_class = &sofalizer_class,
1091 .init = init,
1092 .activate = activate,
1093 .uninit = uninit,
1094 .query_formats = query_formats,
1095 .inputs = inputs,
1096 .outputs = outputs,
1097 .flags = AVFILTER_FLAG_SLICE_THREADS,
1098 };
1099