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
7 * Authors: Andreas Fuchs <andi.fuchs.mail@gmail.com>
8 * Wolfgang Hrauda <wolfgang.hrauda@gmx.at>
10 * SOFAlizer project coordinator at ARI, main developer of SOFA:
11 * Piotr Majdak <piotr@majdak.at>
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.
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.
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 *****************************************************************************/
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"
42 #define FREQUENCY_DOMAIN 1
44 typedef struct MySofa { /* contains data of one SOFA file */
45 struct MYSOFA_EASY *easy;
46 int ir_samples; /* length of one impulse response (IR) */
47 int n_samples; /* ir_samples to next power of 2 */
48 float *lir, *rir; /* IRs (time-domain) */
52 typedef struct VirtualSpeaker {
58 typedef struct SOFAlizerContext {
61 char *filename; /* name of SOFA file */
62 MySofa sofa; /* contains data of the SOFA file */
64 int sample_rate; /* sample rate from SOFA file */
65 float *speaker_azim; /* azimuth of the virtual loudspeakers */
66 float *speaker_elev; /* elevation of the virtual loudspeakers */
67 char *speakers_pos; /* custom positions of the virtual loudspeakers */
68 float lfe_gain; /* initial gain for the LFE channel */
69 float gain_lfe; /* gain applied to LFE channel */
70 int lfe_channel; /* LFE channel position in channel layout */
72 int n_conv; /* number of channels to convolute */
74 /* buffer variables (for convolution) */
75 float *ringbuffer[2]; /* buffers input samples, length of one buffer: */
76 /* no. input ch. (incl. LFE) x buffer_length */
77 int write[2]; /* current write position to ringbuffer */
78 int buffer_length; /* is: longest IR plus max. delay in all SOFA files */
79 /* then choose next power of 2 */
80 int n_fft; /* number of samples in one FFT block */
82 /* netCDF variables */
83 int *delay[2]; /* broadband delay for each channel/IR to be convolved */
85 float *data_ir[2]; /* IRs for all channels to be convolved */
86 /* (this excludes the LFE) */
88 FFTComplex *temp_fft[2];
90 /* control variables */
91 float gain; /* filter gain (in dB) */
92 float rotation; /* rotation of virtual loudspeakers (in degrees) */
93 float elevation; /* elevation of virtual loudspeakers (in deg.) */
94 float radius; /* distance virtual loudspeakers to listener (in metres) */
95 int type; /* processing type */
96 int framesize; /* size of buffer */
98 VirtualSpeaker vspkrpos[64];
100 FFTContext *fft[2], *ifft[2];
101 FFTComplex *data_hrtf[2];
103 AVFloatDSPContext *fdsp;
106 static int close_sofa(struct MySofa *sofa)
108 mysofa_close(sofa->easy);
114 static int preload_sofa(AVFilterContext *ctx, char *filename, int *samplingrate)
116 struct SOFAlizerContext *s = ctx->priv;
117 struct MYSOFA_HRTF *mysofa;
121 mysofa = mysofa_load(filename, &ret);
122 if (ret || !mysofa) {
123 av_log(ctx, AV_LOG_ERROR, "Can't find SOFA-file '%s'\n", filename);
124 return AVERROR(EINVAL);
127 if (mysofa->DataSamplingRate.elements != 1)
128 return AVERROR(EINVAL);
129 *samplingrate = mysofa->DataSamplingRate.values[0];
130 s->sofa.ir_samples = mysofa->N;
131 s->sofa.n_samples = 1 << (32 - ff_clz(s->sofa.ir_samples));
132 license = mysofa_getAttribute(mysofa->attributes, (char *)"License");
134 av_log(ctx, AV_LOG_INFO, "SOFA license: %s\n", license);
140 static int parse_channel_name(char **arg, int *rchannel, char *buf)
142 int len, i, channel_id = 0;
143 int64_t layout, layout0;
145 /* try to parse a channel name, e.g. "FL" */
146 if (av_sscanf(*arg, "%7[A-Z]%n", buf, &len)) {
147 layout0 = layout = av_get_channel_layout(buf);
148 /* channel_id <- first set bit in layout */
149 for (i = 32; i > 0; i >>= 1) {
150 if (layout >= 1LL << i) {
155 /* reject layouts that are not a single channel */
156 if (channel_id >= 64 || layout0 != 1LL << channel_id)
157 return AVERROR(EINVAL);
158 *rchannel = channel_id;
162 return AVERROR(EINVAL);
165 static void parse_speaker_pos(AVFilterContext *ctx, int64_t in_channel_layout)
167 SOFAlizerContext *s = ctx->priv;
168 char *arg, *tokenizer, *p, *args = av_strdup(s->speakers_pos);
174 while ((arg = av_strtok(p, "|", &tokenizer))) {
180 if (parse_channel_name(&arg, &out_ch_id, buf)) {
181 av_log(ctx, AV_LOG_WARNING, "Failed to parse \'%s\' as channel name.\n", buf);
184 if (av_sscanf(arg, "%f %f", &azim, &elev) == 2) {
185 s->vspkrpos[out_ch_id].set = 1;
186 s->vspkrpos[out_ch_id].azim = azim;
187 s->vspkrpos[out_ch_id].elev = elev;
188 } else if (av_sscanf(arg, "%f", &azim) == 1) {
189 s->vspkrpos[out_ch_id].set = 1;
190 s->vspkrpos[out_ch_id].azim = azim;
191 s->vspkrpos[out_ch_id].elev = 0;
198 static int get_speaker_pos(AVFilterContext *ctx,
199 float *speaker_azim, float *speaker_elev)
201 struct SOFAlizerContext *s = ctx->priv;
202 uint64_t channels_layout = ctx->inputs[0]->channel_layout;
203 float azim[16] = { 0 };
204 float elev[16] = { 0 };
205 int m, ch, n_conv = ctx->inputs[0]->channels; /* get no. input channels */
208 return AVERROR(EINVAL);
213 parse_speaker_pos(ctx, channels_layout);
215 /* set speaker positions according to input channel configuration: */
216 for (m = 0, ch = 0; ch < n_conv && m < 64; m++) {
217 uint64_t mask = channels_layout & (1ULL << m);
220 case AV_CH_FRONT_LEFT: azim[ch] = 30; break;
221 case AV_CH_FRONT_RIGHT: azim[ch] = 330; break;
222 case AV_CH_FRONT_CENTER: azim[ch] = 0; break;
223 case AV_CH_LOW_FREQUENCY:
224 case AV_CH_LOW_FREQUENCY_2: s->lfe_channel = ch; break;
225 case AV_CH_BACK_LEFT: azim[ch] = 150; break;
226 case AV_CH_BACK_RIGHT: azim[ch] = 210; break;
227 case AV_CH_BACK_CENTER: azim[ch] = 180; break;
228 case AV_CH_SIDE_LEFT: azim[ch] = 90; break;
229 case AV_CH_SIDE_RIGHT: azim[ch] = 270; break;
230 case AV_CH_FRONT_LEFT_OF_CENTER: azim[ch] = 15; break;
231 case AV_CH_FRONT_RIGHT_OF_CENTER: azim[ch] = 345; break;
232 case AV_CH_TOP_CENTER: azim[ch] = 0;
233 elev[ch] = 90; break;
234 case AV_CH_TOP_FRONT_LEFT: azim[ch] = 30;
235 elev[ch] = 45; break;
236 case AV_CH_TOP_FRONT_CENTER: azim[ch] = 0;
237 elev[ch] = 45; break;
238 case AV_CH_TOP_FRONT_RIGHT: azim[ch] = 330;
239 elev[ch] = 45; break;
240 case AV_CH_TOP_BACK_LEFT: azim[ch] = 150;
241 elev[ch] = 45; break;
242 case AV_CH_TOP_BACK_RIGHT: azim[ch] = 210;
243 elev[ch] = 45; break;
244 case AV_CH_TOP_BACK_CENTER: azim[ch] = 180;
245 elev[ch] = 45; break;
246 case AV_CH_WIDE_LEFT: azim[ch] = 90; break;
247 case AV_CH_WIDE_RIGHT: azim[ch] = 270; break;
248 case AV_CH_SURROUND_DIRECT_LEFT: azim[ch] = 90; break;
249 case AV_CH_SURROUND_DIRECT_RIGHT: azim[ch] = 270; break;
250 case AV_CH_STEREO_LEFT: azim[ch] = 90; break;
251 case AV_CH_STEREO_RIGHT: azim[ch] = 270; break;
254 return AVERROR(EINVAL);
257 if (s->vspkrpos[m].set) {
258 azim[ch] = s->vspkrpos[m].azim;
259 elev[ch] = s->vspkrpos[m].elev;
266 memcpy(speaker_azim, azim, n_conv * sizeof(float));
267 memcpy(speaker_elev, elev, n_conv * sizeof(float));
273 typedef struct ThreadData {
281 FFTComplex **temp_fft;
284 static int sofalizer_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
286 SOFAlizerContext *s = ctx->priv;
287 ThreadData *td = arg;
288 AVFrame *in = td->in, *out = td->out;
290 int *write = &td->write[jobnr];
291 const int *const delay = td->delay[jobnr];
292 const float *const ir = td->ir[jobnr];
293 int *n_clippings = &td->n_clippings[jobnr];
294 float *ringbuffer = td->ringbuffer[jobnr];
295 float *temp_src = td->temp_src[jobnr];
296 const int ir_samples = s->sofa.ir_samples; /* length of one IR */
297 const int n_samples = s->sofa.n_samples;
298 const float *src = (const float *)in->data[0]; /* get pointer to audio input buffer */
299 float *dst = (float *)out->data[0]; /* get pointer to audio output buffer */
300 const int in_channels = s->n_conv; /* number of input channels */
301 /* ring buffer length is: longest IR plus max. delay -> next power of 2 */
302 const int buffer_length = s->buffer_length;
303 /* -1 for AND instead of MODULO (applied to powers of 2): */
304 const uint32_t modulo = (uint32_t)buffer_length - 1;
305 float *buffer[16]; /* holds ringbuffer for each input channel */
311 for (l = 0; l < in_channels; l++) {
312 /* get starting address of ringbuffer for each input channel */
313 buffer[l] = ringbuffer + l * buffer_length;
316 for (i = 0; i < in->nb_samples; i++) {
317 const float *temp_ir = ir; /* using same set of IRs for each sample */
320 for (l = 0; l < in_channels; l++) {
321 /* write current input sample to ringbuffer (for each channel) */
322 buffer[l][wr] = src[l];
325 /* loop goes through all channels to be convolved */
326 for (l = 0; l < in_channels; l++) {
327 const float *const bptr = buffer[l];
329 if (l == s->lfe_channel) {
330 /* LFE is an input channel but requires no convolution */
331 /* apply gain to LFE signal and add to output buffer */
332 *dst += *(buffer[s->lfe_channel] + wr) * s->gain_lfe;
333 temp_ir += n_samples;
337 /* current read position in ringbuffer: input sample write position
338 * - delay for l-th ch. + diff. betw. IR length and buffer length
339 * (mod buffer length) */
340 read = (wr - delay[l] - (n_samples - 1) + buffer_length) & modulo;
342 if (read + n_samples < buffer_length) {
343 memmove(temp_src, bptr + read, n_samples * sizeof(*temp_src));
345 int len = FFMIN(n_samples - (read % n_samples), buffer_length - read);
347 memmove(temp_src, bptr + read, len * sizeof(*temp_src));
348 memmove(temp_src + len, bptr, (n_samples - len) * sizeof(*temp_src));
351 /* multiply signal and IR, and add up the results */
352 dst[0] += s->fdsp->scalarproduct_float(temp_ir, temp_src, FFALIGN(ir_samples, 32));
353 temp_ir += n_samples;
356 /* clippings counter */
357 if (fabsf(dst[0]) > 1)
360 /* move output buffer pointer by +2 to get to next sample of processed channel: */
363 wr = (wr + 1) & modulo; /* update ringbuffer write position */
366 *write = wr; /* remember write position in ringbuffer for next call */
371 static int sofalizer_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
373 SOFAlizerContext *s = ctx->priv;
374 ThreadData *td = arg;
375 AVFrame *in = td->in, *out = td->out;
377 int *write = &td->write[jobnr];
378 FFTComplex *hrtf = s->data_hrtf[jobnr]; /* get pointers to current HRTF data */
379 int *n_clippings = &td->n_clippings[jobnr];
380 float *ringbuffer = td->ringbuffer[jobnr];
381 const int n_samples = s->sofa.n_samples; /* length of one IR */
382 const float *src = (const float *)in->data[0]; /* get pointer to audio input buffer */
383 float *dst = (float *)out->data[0]; /* get pointer to audio output buffer */
384 const int in_channels = s->n_conv; /* number of input channels */
385 /* ring buffer length is: longest IR plus max. delay -> next power of 2 */
386 const int buffer_length = s->buffer_length;
387 /* -1 for AND instead of MODULO (applied to powers of 2): */
388 const uint32_t modulo = (uint32_t)buffer_length - 1;
389 FFTComplex *fft_in = s->temp_fft[jobnr]; /* temporary array for FFT input/output data */
390 FFTContext *ifft = s->ifft[jobnr];
391 FFTContext *fft = s->fft[jobnr];
392 const int n_conv = s->n_conv;
393 const int n_fft = s->n_fft;
394 const float fft_scale = 1.0f / s->n_fft;
395 FFTComplex *hrtf_offset;
402 /* find minimum between number of samples and output buffer length:
403 * (important, if one IR is longer than the output buffer) */
404 n_read = FFMIN(s->sofa.n_samples, in->nb_samples);
405 for (j = 0; j < n_read; j++) {
406 /* initialize output buf with saved signal from overflow buf */
407 dst[2 * j] = ringbuffer[wr];
408 ringbuffer[wr] = 0.0; /* re-set read samples to zero */
409 /* update ringbuffer read/write position */
410 wr = (wr + 1) & modulo;
413 /* initialize rest of output buffer with 0 */
414 for (j = n_read; j < in->nb_samples; j++) {
418 for (i = 0; i < n_conv; i++) {
419 if (i == s->lfe_channel) { /* LFE */
420 for (j = 0; j < in->nb_samples; j++) {
421 /* apply gain to LFE signal and add to output buffer */
422 dst[2 * j] += src[i + j * in_channels] * s->gain_lfe;
427 /* outer loop: go through all input channels to be convolved */
428 offset = i * n_fft; /* no. samples already processed */
429 hrtf_offset = hrtf + offset;
431 /* fill FFT input with 0 (we want to zero-pad) */
432 memset(fft_in, 0, sizeof(FFTComplex) * n_fft);
434 for (j = 0; j < in->nb_samples; j++) {
435 /* prepare input for FFT */
436 /* write all samples of current input channel to FFT input array */
437 fft_in[j].re = src[j * in_channels + i];
440 /* transform input signal of current channel to frequency domain */
441 av_fft_permute(fft, fft_in);
442 av_fft_calc(fft, fft_in);
443 for (j = 0; j < n_fft; j++) {
444 const FFTComplex *hcomplex = hrtf_offset + j;
445 const float re = fft_in[j].re;
446 const float im = fft_in[j].im;
448 /* complex multiplication of input signal and HRTFs */
449 /* output channel (real): */
450 fft_in[j].re = re * hcomplex->re - im * hcomplex->im;
451 /* output channel (imag): */
452 fft_in[j].im = re * hcomplex->im + im * hcomplex->re;
455 /* transform output signal of current channel back to time domain */
456 av_fft_permute(ifft, fft_in);
457 av_fft_calc(ifft, fft_in);
459 for (j = 0; j < in->nb_samples; j++) {
460 /* write output signal of current channel to output buffer */
461 dst[2 * j] += fft_in[j].re * fft_scale;
464 for (j = 0; j < n_samples - 1; j++) { /* overflow length is IR length - 1 */
465 /* write the rest of output signal to overflow buffer */
466 int write_pos = (wr + j) & modulo;
468 *(ringbuffer + write_pos) += fft_in[in->nb_samples + j].re * fft_scale;
472 /* go through all samples of current output buffer: count clippings */
473 for (i = 0; i < out->nb_samples; i++) {
474 /* clippings counter */
475 if (fabsf(dst[0]) > 1) { /* if current output sample > 1 */
479 /* move output buffer pointer by +2 to get to next sample of processed channel: */
483 /* remember read/write position in ringbuffer for next call */
489 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
491 AVFilterContext *ctx = inlink->dst;
492 SOFAlizerContext *s = ctx->priv;
493 AVFilterLink *outlink = ctx->outputs[0];
494 int n_clippings[2] = { 0 };
498 out = ff_get_audio_buffer(outlink, in->nb_samples);
501 return AVERROR(ENOMEM);
503 av_frame_copy_props(out, in);
505 td.in = in; td.out = out; td.write = s->write;
506 td.delay = s->delay; td.ir = s->data_ir; td.n_clippings = n_clippings;
507 td.ringbuffer = s->ringbuffer; td.temp_src = s->temp_src;
508 td.temp_fft = s->temp_fft;
510 if (s->type == TIME_DOMAIN) {
511 ctx->internal->execute(ctx, sofalizer_convolute, &td, NULL, 2);
513 ctx->internal->execute(ctx, sofalizer_fast_convolute, &td, NULL, 2);
517 /* display error message if clipping occurred */
518 if (n_clippings[0] + n_clippings[1] > 0) {
519 av_log(ctx, AV_LOG_WARNING, "%d of %d samples clipped. Please reduce gain.\n",
520 n_clippings[0] + n_clippings[1], out->nb_samples * 2);
524 return ff_filter_frame(outlink, out);
527 static int query_formats(AVFilterContext *ctx)
529 struct SOFAlizerContext *s = ctx->priv;
530 AVFilterFormats *formats = NULL;
531 AVFilterChannelLayouts *layouts = NULL;
532 int ret, sample_rates[] = { 48000, -1 };
534 ret = ff_add_format(&formats, AV_SAMPLE_FMT_FLT);
537 ret = ff_set_common_formats(ctx, formats);
541 layouts = ff_all_channel_layouts();
543 return AVERROR(ENOMEM);
545 ret = ff_channel_layouts_ref(layouts, &ctx->inputs[0]->out_channel_layouts);
550 ret = ff_add_channel_layout(&layouts, AV_CH_LAYOUT_STEREO);
554 ret = ff_channel_layouts_ref(layouts, &ctx->outputs[0]->in_channel_layouts);
558 sample_rates[0] = s->sample_rate;
559 formats = ff_make_format_list(sample_rates);
561 return AVERROR(ENOMEM);
562 return ff_set_common_samplerates(ctx, formats);
565 static int load_data(AVFilterContext *ctx, int azim, int elev, float radius, int sample_rate)
567 struct SOFAlizerContext *s = ctx->priv;
570 int n_conv = s->n_conv; /* no. channels to convolve */
572 float delay_l; /* broadband delay for each IR */
574 int nb_input_channels = ctx->inputs[0]->channels; /* no. input channels */
575 float gain_lin = expf((s->gain - 3 * nb_input_channels) / 20 * M_LN10); /* gain - 3dB/channel */
576 FFTComplex *data_hrtf_l = NULL;
577 FFTComplex *data_hrtf_r = NULL;
578 FFTComplex *fft_in_l = NULL;
579 FFTComplex *fft_in_r = NULL;
580 float *data_ir_l = NULL;
581 float *data_ir_r = NULL;
582 int offset = 0; /* used for faster pointer arithmetics in for-loop */
583 int i, j, azim_orig = azim, elev_orig = elev;
584 int filter_length, ret = 0;
588 s->sofa.easy = mysofa_open(s->filename, sample_rate, &filter_length, &ret);
589 if (!s->sofa.easy || ret) { /* if an invalid SOFA file has been selected */
590 av_log(ctx, AV_LOG_ERROR, "Selected SOFA file is invalid. Please select valid SOFA file.\n");
591 return AVERROR_INVALIDDATA;
594 n_samples = s->sofa.n_samples;
595 ir_samples = s->sofa.ir_samples;
597 s->data_ir[0] = av_calloc(n_samples, sizeof(float) * s->n_conv);
598 s->data_ir[1] = av_calloc(n_samples, sizeof(float) * s->n_conv);
599 s->delay[0] = av_calloc(s->n_conv, sizeof(int));
600 s->delay[1] = av_calloc(s->n_conv, sizeof(int));
602 if (!s->data_ir[0] || !s->data_ir[1] || !s->delay[0] || !s->delay[1]) {
603 ret = AVERROR(ENOMEM);
607 /* get temporary IR for L and R channel */
608 data_ir_l = av_calloc(n_conv * n_samples, sizeof(*data_ir_l));
609 data_ir_r = av_calloc(n_conv * n_samples, sizeof(*data_ir_r));
610 if (!data_ir_r || !data_ir_l) {
611 ret = AVERROR(ENOMEM);
615 if (s->type == TIME_DOMAIN) {
616 s->temp_src[0] = av_calloc(n_samples, sizeof(float));
617 s->temp_src[1] = av_calloc(n_samples, sizeof(float));
618 if (!s->temp_src[0] || !s->temp_src[1]) {
619 ret = AVERROR(ENOMEM);
624 s->speaker_azim = av_calloc(s->n_conv, sizeof(*s->speaker_azim));
625 s->speaker_elev = av_calloc(s->n_conv, sizeof(*s->speaker_elev));
626 if (!s->speaker_azim || !s->speaker_elev) {
627 ret = AVERROR(ENOMEM);
631 /* get speaker positions */
632 if ((ret = get_speaker_pos(ctx, s->speaker_azim, s->speaker_elev)) < 0) {
633 av_log(ctx, AV_LOG_ERROR, "Couldn't get speaker positions. Input channel configuration not supported.\n");
637 for (i = 0; i < s->n_conv; i++) {
638 float coordinates[3];
640 /* load and store IRs and corresponding delays */
641 azim = (int)(s->speaker_azim[i] + azim_orig) % 360;
642 elev = (int)(s->speaker_elev[i] + elev_orig) % 90;
644 coordinates[0] = azim;
645 coordinates[1] = elev;
646 coordinates[2] = radius;
648 mysofa_s2c(coordinates);
650 /* get id of IR closest to desired position */
651 mysofa_getfilter_float(s->sofa.easy, coordinates[0], coordinates[1], coordinates[2],
652 data_ir_l + n_samples * i,
653 data_ir_r + n_samples * i,
656 s->delay[0][i] = delay_l * sample_rate;
657 s->delay[1][i] = delay_r * sample_rate;
659 s->sofa.max_delay = FFMAX3(s->sofa.max_delay, s->delay[0][i], s->delay[1][i]);
662 /* get size of ringbuffer (longest IR plus max. delay) */
663 /* then choose next power of 2 for performance optimization */
664 n_current = n_samples + s->sofa.max_delay;
665 /* length of longest IR plus max. delay */
666 n_max = FFMAX(n_max, n_current);
668 /* buffer length is longest IR plus max. delay -> next power of 2
669 (32 - count leading zeros gives required exponent) */
670 s->buffer_length = 1 << (32 - ff_clz(n_max));
671 s->n_fft = n_fft = 1 << (32 - ff_clz(n_max + s->framesize));
673 if (s->type == FREQUENCY_DOMAIN) {
674 av_fft_end(s->fft[0]);
675 av_fft_end(s->fft[1]);
676 s->fft[0] = av_fft_init(log2(s->n_fft), 0);
677 s->fft[1] = av_fft_init(log2(s->n_fft), 0);
678 av_fft_end(s->ifft[0]);
679 av_fft_end(s->ifft[1]);
680 s->ifft[0] = av_fft_init(log2(s->n_fft), 1);
681 s->ifft[1] = av_fft_init(log2(s->n_fft), 1);
683 if (!s->fft[0] || !s->fft[1] || !s->ifft[0] || !s->ifft[1]) {
684 av_log(ctx, AV_LOG_ERROR, "Unable to create FFT contexts of size %d.\n", s->n_fft);
685 ret = AVERROR(ENOMEM);
690 if (s->type == TIME_DOMAIN) {
691 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
692 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
694 /* get temporary HRTF memory for L and R channel */
695 data_hrtf_l = av_malloc_array(n_fft, sizeof(*data_hrtf_l) * n_conv);
696 data_hrtf_r = av_malloc_array(n_fft, sizeof(*data_hrtf_r) * n_conv);
697 if (!data_hrtf_r || !data_hrtf_l) {
698 ret = AVERROR(ENOMEM);
702 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float));
703 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float));
704 s->temp_fft[0] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
705 s->temp_fft[1] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
706 if (!s->temp_fft[0] || !s->temp_fft[1]) {
707 ret = AVERROR(ENOMEM);
712 if (!s->ringbuffer[0] || !s->ringbuffer[1]) {
713 ret = AVERROR(ENOMEM);
717 if (s->type == FREQUENCY_DOMAIN) {
718 fft_in_l = av_calloc(n_fft, sizeof(*fft_in_l));
719 fft_in_r = av_calloc(n_fft, sizeof(*fft_in_r));
720 if (!fft_in_l || !fft_in_r) {
721 ret = AVERROR(ENOMEM);
726 for (i = 0; i < s->n_conv; i++) {
729 offset = i * n_samples; /* no. samples already written */
731 lir = data_ir_l + offset;
732 rir = data_ir_r + offset;
734 if (s->type == TIME_DOMAIN) {
735 for (j = 0; j < ir_samples; j++) {
736 /* load reversed IRs of the specified source position
737 * sample-by-sample for left and right ear; and apply gain */
738 s->data_ir[0][offset + j] = lir[ir_samples - 1 - j] * gain_lin;
739 s->data_ir[1][offset + j] = rir[ir_samples - 1 - j] * gain_lin;
742 memset(fft_in_l, 0, n_fft * sizeof(*fft_in_l));
743 memset(fft_in_r, 0, n_fft * sizeof(*fft_in_r));
745 offset = i * n_fft; /* no. samples already written */
746 for (j = 0; j < ir_samples; j++) {
747 /* load non-reversed IRs of the specified source position
748 * sample-by-sample and apply gain,
749 * L channel is loaded to real part, R channel to imag part,
750 * IRs ared shifted by L and R delay */
751 fft_in_l[s->delay[0][i] + j].re = lir[j] * gain_lin;
752 fft_in_r[s->delay[1][i] + j].re = rir[j] * gain_lin;
755 /* actually transform to frequency domain (IRs -> HRTFs) */
756 av_fft_permute(s->fft[0], fft_in_l);
757 av_fft_calc(s->fft[0], fft_in_l);
758 memcpy(data_hrtf_l + offset, fft_in_l, n_fft * sizeof(*fft_in_l));
759 av_fft_permute(s->fft[0], fft_in_r);
760 av_fft_calc(s->fft[0], fft_in_r);
761 memcpy(data_hrtf_r + offset, fft_in_r, n_fft * sizeof(*fft_in_r));
765 if (s->type == FREQUENCY_DOMAIN) {
766 s->data_hrtf[0] = av_malloc_array(n_fft * s->n_conv, sizeof(FFTComplex));
767 s->data_hrtf[1] = av_malloc_array(n_fft * s->n_conv, sizeof(FFTComplex));
768 if (!s->data_hrtf[0] || !s->data_hrtf[1]) {
769 ret = AVERROR(ENOMEM);
773 memcpy(s->data_hrtf[0], data_hrtf_l, /* copy HRTF data to */
774 sizeof(FFTComplex) * n_conv * n_fft); /* filter struct */
775 memcpy(s->data_hrtf[1], data_hrtf_r,
776 sizeof(FFTComplex) * n_conv * n_fft);
780 av_freep(&data_hrtf_l); /* free temporary HRTF memory */
781 av_freep(&data_hrtf_r);
783 av_freep(&data_ir_l); /* free temprary IR memory */
784 av_freep(&data_ir_r);
786 av_freep(&fft_in_l); /* free temporary FFT memory */
792 static av_cold int init(AVFilterContext *ctx)
794 SOFAlizerContext *s = ctx->priv;
798 av_log(ctx, AV_LOG_ERROR, "Valid SOFA filename must be set.\n");
799 return AVERROR(EINVAL);
802 /* preload SOFA file, */
803 ret = preload_sofa(ctx, s->filename, &s->sample_rate);
805 /* file loading error */
806 av_log(ctx, AV_LOG_ERROR, "Error while loading SOFA file: '%s'\n", s->filename);
807 } else { /* no file loading error, resampling not required */
808 av_log(ctx, AV_LOG_DEBUG, "File '%s' loaded.\n", s->filename);
812 av_log(ctx, AV_LOG_ERROR, "No valid SOFA file could be loaded. Please specify valid SOFA file.\n");
816 s->fdsp = avpriv_float_dsp_alloc(0);
818 return AVERROR(ENOMEM);
823 static int config_input(AVFilterLink *inlink)
825 AVFilterContext *ctx = inlink->dst;
826 SOFAlizerContext *s = ctx->priv;
829 if (s->type == FREQUENCY_DOMAIN) {
830 inlink->partial_buf_size =
831 inlink->min_samples =
832 inlink->max_samples = s->framesize;
835 /* gain -3 dB per channel, -6 dB to get LFE on a similar level */
836 s->gain_lfe = expf((s->gain - 3 * inlink->channels - 6 + s->lfe_gain) / 20 * M_LN10);
838 s->n_conv = inlink->channels;
840 /* load IRs to data_ir[0] and data_ir[1] for required directions */
841 if ((ret = load_data(ctx, s->rotation, s->elevation, s->radius, inlink->sample_rate)) < 0)
844 av_log(ctx, AV_LOG_DEBUG, "Samplerate: %d Channels to convolute: %d, Length of ringbuffer: %d x %d\n",
845 inlink->sample_rate, s->n_conv, inlink->channels, s->buffer_length);
850 static av_cold void uninit(AVFilterContext *ctx)
852 SOFAlizerContext *s = ctx->priv;
854 close_sofa(&s->sofa);
855 av_fft_end(s->ifft[0]);
856 av_fft_end(s->ifft[1]);
857 av_fft_end(s->fft[0]);
858 av_fft_end(s->fft[1]);
859 av_freep(&s->delay[0]);
860 av_freep(&s->delay[1]);
861 av_freep(&s->data_ir[0]);
862 av_freep(&s->data_ir[1]);
863 av_freep(&s->ringbuffer[0]);
864 av_freep(&s->ringbuffer[1]);
865 av_freep(&s->speaker_azim);
866 av_freep(&s->speaker_elev);
867 av_freep(&s->temp_src[0]);
868 av_freep(&s->temp_src[1]);
869 av_freep(&s->temp_fft[0]);
870 av_freep(&s->temp_fft[1]);
871 av_freep(&s->data_hrtf[0]);
872 av_freep(&s->data_hrtf[1]);
876 #define OFFSET(x) offsetof(SOFAlizerContext, x)
877 #define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
879 static const AVOption sofalizer_options[] = {
880 { "sofa", "sofa filename", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
881 { "gain", "set gain in dB", OFFSET(gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -20, 40, .flags = FLAGS },
882 { "rotation", "set rotation" , OFFSET(rotation), AV_OPT_TYPE_FLOAT, {.dbl=0}, -360, 360, .flags = FLAGS },
883 { "elevation", "set elevation", OFFSET(elevation), AV_OPT_TYPE_FLOAT, {.dbl=0}, -90, 90, .flags = FLAGS },
884 { "radius", "set radius", OFFSET(radius), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 5, .flags = FLAGS },
885 { "type", "set processing", OFFSET(type), AV_OPT_TYPE_INT, {.i64=1}, 0, 1, .flags = FLAGS, "type" },
886 { "time", "time domain", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, .flags = FLAGS, "type" },
887 { "freq", "frequency domain", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, .flags = FLAGS, "type" },
888 { "speakers", "set speaker custom positions", OFFSET(speakers_pos), AV_OPT_TYPE_STRING, {.str=0}, 0, 0, .flags = FLAGS },
889 { "lfegain", "set lfe gain", OFFSET(lfe_gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -9, 9, .flags = FLAGS },
890 { "framesize", "set frame size", OFFSET(framesize), AV_OPT_TYPE_INT, {.i64=1024},1024,96000, .flags = FLAGS },
894 AVFILTER_DEFINE_CLASS(sofalizer);
896 static const AVFilterPad inputs[] = {
899 .type = AVMEDIA_TYPE_AUDIO,
900 .config_props = config_input,
901 .filter_frame = filter_frame,
906 static const AVFilterPad outputs[] = {
909 .type = AVMEDIA_TYPE_AUDIO,
914 AVFilter ff_af_sofalizer = {
916 .description = NULL_IF_CONFIG_SMALL("SOFAlizer (Spatially Oriented Format for Acoustics)."),
917 .priv_size = sizeof(SOFAlizerContext),
918 .priv_class = &sofalizer_class,
921 .query_formats = query_formats,
924 .flags = AVFILTER_FLAG_SLICE_THREADS,