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 MYSOFA_HRTF *mysofa;
120 mysofa = mysofa_load(filename, &ret);
121 if (ret || !mysofa) {
122 av_log(ctx, AV_LOG_ERROR, "Can't find SOFA-file '%s'\n", filename);
123 return AVERROR(EINVAL);
126 if (mysofa->DataSamplingRate.elements != 1)
127 return AVERROR(EINVAL);
128 av_log(ctx, AV_LOG_DEBUG, "Original IR length: %d.\n", mysofa->N);
129 *samplingrate = mysofa->DataSamplingRate.values[0];
130 license = mysofa_getAttribute(mysofa->attributes, (char *)"License");
132 av_log(ctx, AV_LOG_INFO, "SOFA license: %s\n", license);
138 static int parse_channel_name(char **arg, int *rchannel, char *buf)
140 int len, i, channel_id = 0;
141 int64_t layout, layout0;
143 /* try to parse a channel name, e.g. "FL" */
144 if (av_sscanf(*arg, "%7[A-Z]%n", buf, &len)) {
145 layout0 = layout = av_get_channel_layout(buf);
146 /* channel_id <- first set bit in layout */
147 for (i = 32; i > 0; i >>= 1) {
148 if (layout >= 1LL << i) {
153 /* reject layouts that are not a single channel */
154 if (channel_id >= 64 || layout0 != 1LL << channel_id)
155 return AVERROR(EINVAL);
156 *rchannel = channel_id;
160 return AVERROR(EINVAL);
163 static void parse_speaker_pos(AVFilterContext *ctx, int64_t in_channel_layout)
165 SOFAlizerContext *s = ctx->priv;
166 char *arg, *tokenizer, *p, *args = av_strdup(s->speakers_pos);
172 while ((arg = av_strtok(p, "|", &tokenizer))) {
178 if (parse_channel_name(&arg, &out_ch_id, buf)) {
179 av_log(ctx, AV_LOG_WARNING, "Failed to parse \'%s\' as channel name.\n", buf);
182 if (av_sscanf(arg, "%f %f", &azim, &elev) == 2) {
183 s->vspkrpos[out_ch_id].set = 1;
184 s->vspkrpos[out_ch_id].azim = azim;
185 s->vspkrpos[out_ch_id].elev = elev;
186 } else if (av_sscanf(arg, "%f", &azim) == 1) {
187 s->vspkrpos[out_ch_id].set = 1;
188 s->vspkrpos[out_ch_id].azim = azim;
189 s->vspkrpos[out_ch_id].elev = 0;
196 static int get_speaker_pos(AVFilterContext *ctx,
197 float *speaker_azim, float *speaker_elev)
199 struct SOFAlizerContext *s = ctx->priv;
200 uint64_t channels_layout = ctx->inputs[0]->channel_layout;
201 float azim[16] = { 0 };
202 float elev[16] = { 0 };
203 int m, ch, n_conv = ctx->inputs[0]->channels; /* get no. input channels */
206 return AVERROR(EINVAL);
211 parse_speaker_pos(ctx, channels_layout);
213 /* set speaker positions according to input channel configuration: */
214 for (m = 0, ch = 0; ch < n_conv && m < 64; m++) {
215 uint64_t mask = channels_layout & (1ULL << m);
218 case AV_CH_FRONT_LEFT: azim[ch] = 30; break;
219 case AV_CH_FRONT_RIGHT: azim[ch] = 330; break;
220 case AV_CH_FRONT_CENTER: azim[ch] = 0; break;
221 case AV_CH_LOW_FREQUENCY:
222 case AV_CH_LOW_FREQUENCY_2: s->lfe_channel = ch; break;
223 case AV_CH_BACK_LEFT: azim[ch] = 150; break;
224 case AV_CH_BACK_RIGHT: azim[ch] = 210; break;
225 case AV_CH_BACK_CENTER: azim[ch] = 180; break;
226 case AV_CH_SIDE_LEFT: azim[ch] = 90; break;
227 case AV_CH_SIDE_RIGHT: azim[ch] = 270; break;
228 case AV_CH_FRONT_LEFT_OF_CENTER: azim[ch] = 15; break;
229 case AV_CH_FRONT_RIGHT_OF_CENTER: azim[ch] = 345; break;
230 case AV_CH_TOP_CENTER: azim[ch] = 0;
231 elev[ch] = 90; break;
232 case AV_CH_TOP_FRONT_LEFT: azim[ch] = 30;
233 elev[ch] = 45; break;
234 case AV_CH_TOP_FRONT_CENTER: azim[ch] = 0;
235 elev[ch] = 45; break;
236 case AV_CH_TOP_FRONT_RIGHT: azim[ch] = 330;
237 elev[ch] = 45; break;
238 case AV_CH_TOP_BACK_LEFT: azim[ch] = 150;
239 elev[ch] = 45; break;
240 case AV_CH_TOP_BACK_RIGHT: azim[ch] = 210;
241 elev[ch] = 45; break;
242 case AV_CH_TOP_BACK_CENTER: azim[ch] = 180;
243 elev[ch] = 45; break;
244 case AV_CH_WIDE_LEFT: azim[ch] = 90; break;
245 case AV_CH_WIDE_RIGHT: azim[ch] = 270; break;
246 case AV_CH_SURROUND_DIRECT_LEFT: azim[ch] = 90; break;
247 case AV_CH_SURROUND_DIRECT_RIGHT: azim[ch] = 270; break;
248 case AV_CH_STEREO_LEFT: azim[ch] = 90; break;
249 case AV_CH_STEREO_RIGHT: azim[ch] = 270; break;
252 return AVERROR(EINVAL);
255 if (s->vspkrpos[m].set) {
256 azim[ch] = s->vspkrpos[m].azim;
257 elev[ch] = s->vspkrpos[m].elev;
264 memcpy(speaker_azim, azim, n_conv * sizeof(float));
265 memcpy(speaker_elev, elev, n_conv * sizeof(float));
271 typedef struct ThreadData {
279 FFTComplex **temp_fft;
282 static int sofalizer_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
284 SOFAlizerContext *s = ctx->priv;
285 ThreadData *td = arg;
286 AVFrame *in = td->in, *out = td->out;
288 int *write = &td->write[jobnr];
289 const int *const delay = td->delay[jobnr];
290 const float *const ir = td->ir[jobnr];
291 int *n_clippings = &td->n_clippings[jobnr];
292 float *ringbuffer = td->ringbuffer[jobnr];
293 float *temp_src = td->temp_src[jobnr];
294 const int ir_samples = s->sofa.ir_samples; /* length of one IR */
295 const int n_samples = s->sofa.n_samples;
296 const float *src = (const float *)in->data[0]; /* get pointer to audio input buffer */
297 float *dst = (float *)out->data[0]; /* get pointer to audio output buffer */
298 const int in_channels = s->n_conv; /* number of input channels */
299 /* ring buffer length is: longest IR plus max. delay -> next power of 2 */
300 const int buffer_length = s->buffer_length;
301 /* -1 for AND instead of MODULO (applied to powers of 2): */
302 const uint32_t modulo = (uint32_t)buffer_length - 1;
303 float *buffer[16]; /* holds ringbuffer for each input channel */
309 for (l = 0; l < in_channels; l++) {
310 /* get starting address of ringbuffer for each input channel */
311 buffer[l] = ringbuffer + l * buffer_length;
314 for (i = 0; i < in->nb_samples; i++) {
315 const float *temp_ir = ir; /* using same set of IRs for each sample */
318 for (l = 0; l < in_channels; l++) {
319 /* write current input sample to ringbuffer (for each channel) */
320 buffer[l][wr] = src[l];
323 /* loop goes through all channels to be convolved */
324 for (l = 0; l < in_channels; l++) {
325 const float *const bptr = buffer[l];
327 if (l == s->lfe_channel) {
328 /* LFE is an input channel but requires no convolution */
329 /* apply gain to LFE signal and add to output buffer */
330 *dst += *(buffer[s->lfe_channel] + wr) * s->gain_lfe;
331 temp_ir += n_samples;
335 /* current read position in ringbuffer: input sample write position
336 * - delay for l-th ch. + diff. betw. IR length and buffer length
337 * (mod buffer length) */
338 read = (wr - delay[l] - (n_samples - 1) + buffer_length) & modulo;
340 if (read + n_samples < buffer_length) {
341 memmove(temp_src, bptr + read, n_samples * sizeof(*temp_src));
343 int len = FFMIN(n_samples - (read % n_samples), buffer_length - read);
345 memmove(temp_src, bptr + read, len * sizeof(*temp_src));
346 memmove(temp_src + len, bptr, (n_samples - len) * sizeof(*temp_src));
349 /* multiply signal and IR, and add up the results */
350 dst[0] += s->fdsp->scalarproduct_float(temp_ir, temp_src, FFALIGN(ir_samples, 32));
351 temp_ir += n_samples;
354 /* clippings counter */
355 if (fabsf(dst[0]) > 1)
358 /* move output buffer pointer by +2 to get to next sample of processed channel: */
361 wr = (wr + 1) & modulo; /* update ringbuffer write position */
364 *write = wr; /* remember write position in ringbuffer for next call */
369 static int sofalizer_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
371 SOFAlizerContext *s = ctx->priv;
372 ThreadData *td = arg;
373 AVFrame *in = td->in, *out = td->out;
375 int *write = &td->write[jobnr];
376 FFTComplex *hrtf = s->data_hrtf[jobnr]; /* get pointers to current HRTF data */
377 int *n_clippings = &td->n_clippings[jobnr];
378 float *ringbuffer = td->ringbuffer[jobnr];
379 const int n_samples = s->sofa.n_samples; /* length of one IR */
380 const float *src = (const float *)in->data[0]; /* get pointer to audio input buffer */
381 float *dst = (float *)out->data[0]; /* get pointer to audio output buffer */
382 const int in_channels = s->n_conv; /* number of input channels */
383 /* ring buffer length is: longest IR plus max. delay -> next power of 2 */
384 const int buffer_length = s->buffer_length;
385 /* -1 for AND instead of MODULO (applied to powers of 2): */
386 const uint32_t modulo = (uint32_t)buffer_length - 1;
387 FFTComplex *fft_in = s->temp_fft[jobnr]; /* temporary array for FFT input/output data */
388 FFTContext *ifft = s->ifft[jobnr];
389 FFTContext *fft = s->fft[jobnr];
390 const int n_conv = s->n_conv;
391 const int n_fft = s->n_fft;
392 const float fft_scale = 1.0f / s->n_fft;
393 FFTComplex *hrtf_offset;
400 /* find minimum between number of samples and output buffer length:
401 * (important, if one IR is longer than the output buffer) */
402 n_read = FFMIN(s->sofa.n_samples, in->nb_samples);
403 for (j = 0; j < n_read; j++) {
404 /* initialize output buf with saved signal from overflow buf */
405 dst[2 * j] = ringbuffer[wr];
406 ringbuffer[wr] = 0.0; /* re-set read samples to zero */
407 /* update ringbuffer read/write position */
408 wr = (wr + 1) & modulo;
411 /* initialize rest of output buffer with 0 */
412 for (j = n_read; j < in->nb_samples; j++) {
416 for (i = 0; i < n_conv; i++) {
417 if (i == s->lfe_channel) { /* LFE */
418 for (j = 0; j < in->nb_samples; j++) {
419 /* apply gain to LFE signal and add to output buffer */
420 dst[2 * j] += src[i + j * in_channels] * s->gain_lfe;
425 /* outer loop: go through all input channels to be convolved */
426 offset = i * n_fft; /* no. samples already processed */
427 hrtf_offset = hrtf + offset;
429 /* fill FFT input with 0 (we want to zero-pad) */
430 memset(fft_in, 0, sizeof(FFTComplex) * n_fft);
432 for (j = 0; j < in->nb_samples; j++) {
433 /* prepare input for FFT */
434 /* write all samples of current input channel to FFT input array */
435 fft_in[j].re = src[j * in_channels + i];
438 /* transform input signal of current channel to frequency domain */
439 av_fft_permute(fft, fft_in);
440 av_fft_calc(fft, fft_in);
441 for (j = 0; j < n_fft; j++) {
442 const FFTComplex *hcomplex = hrtf_offset + j;
443 const float re = fft_in[j].re;
444 const float im = fft_in[j].im;
446 /* complex multiplication of input signal and HRTFs */
447 /* output channel (real): */
448 fft_in[j].re = re * hcomplex->re - im * hcomplex->im;
449 /* output channel (imag): */
450 fft_in[j].im = re * hcomplex->im + im * hcomplex->re;
453 /* transform output signal of current channel back to time domain */
454 av_fft_permute(ifft, fft_in);
455 av_fft_calc(ifft, fft_in);
457 for (j = 0; j < in->nb_samples; j++) {
458 /* write output signal of current channel to output buffer */
459 dst[2 * j] += fft_in[j].re * fft_scale;
462 for (j = 0; j < n_samples - 1; j++) { /* overflow length is IR length - 1 */
463 /* write the rest of output signal to overflow buffer */
464 int write_pos = (wr + j) & modulo;
466 *(ringbuffer + write_pos) += fft_in[in->nb_samples + j].re * fft_scale;
470 /* go through all samples of current output buffer: count clippings */
471 for (i = 0; i < out->nb_samples; i++) {
472 /* clippings counter */
473 if (fabsf(dst[0]) > 1) { /* if current output sample > 1 */
477 /* move output buffer pointer by +2 to get to next sample of processed channel: */
481 /* remember read/write position in ringbuffer for next call */
487 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
489 AVFilterContext *ctx = inlink->dst;
490 SOFAlizerContext *s = ctx->priv;
491 AVFilterLink *outlink = ctx->outputs[0];
492 int n_clippings[2] = { 0 };
496 out = ff_get_audio_buffer(outlink, in->nb_samples);
499 return AVERROR(ENOMEM);
501 av_frame_copy_props(out, in);
503 td.in = in; td.out = out; td.write = s->write;
504 td.delay = s->delay; td.ir = s->data_ir; td.n_clippings = n_clippings;
505 td.ringbuffer = s->ringbuffer; td.temp_src = s->temp_src;
506 td.temp_fft = s->temp_fft;
508 if (s->type == TIME_DOMAIN) {
509 ctx->internal->execute(ctx, sofalizer_convolute, &td, NULL, 2);
511 ctx->internal->execute(ctx, sofalizer_fast_convolute, &td, NULL, 2);
515 /* display error message if clipping occurred */
516 if (n_clippings[0] + n_clippings[1] > 0) {
517 av_log(ctx, AV_LOG_WARNING, "%d of %d samples clipped. Please reduce gain.\n",
518 n_clippings[0] + n_clippings[1], out->nb_samples * 2);
522 return ff_filter_frame(outlink, out);
525 static int query_formats(AVFilterContext *ctx)
527 struct SOFAlizerContext *s = ctx->priv;
528 AVFilterFormats *formats = NULL;
529 AVFilterChannelLayouts *layouts = NULL;
530 int ret, sample_rates[] = { 48000, -1 };
532 ret = ff_add_format(&formats, AV_SAMPLE_FMT_FLT);
535 ret = ff_set_common_formats(ctx, formats);
539 layouts = ff_all_channel_layouts();
541 return AVERROR(ENOMEM);
543 ret = ff_channel_layouts_ref(layouts, &ctx->inputs[0]->out_channel_layouts);
548 ret = ff_add_channel_layout(&layouts, AV_CH_LAYOUT_STEREO);
552 ret = ff_channel_layouts_ref(layouts, &ctx->outputs[0]->in_channel_layouts);
556 sample_rates[0] = s->sample_rate;
557 formats = ff_make_format_list(sample_rates);
559 return AVERROR(ENOMEM);
560 return ff_set_common_samplerates(ctx, formats);
563 static int load_data(AVFilterContext *ctx, int azim, int elev, float radius, int sample_rate)
565 struct SOFAlizerContext *s = ctx->priv;
568 int n_conv = s->n_conv; /* no. channels to convolve */
570 float delay_l; /* broadband delay for each IR */
572 int nb_input_channels = ctx->inputs[0]->channels; /* no. input channels */
573 float gain_lin = expf((s->gain - 3 * nb_input_channels) / 20 * M_LN10); /* gain - 3dB/channel */
574 FFTComplex *data_hrtf_l = NULL;
575 FFTComplex *data_hrtf_r = NULL;
576 FFTComplex *fft_in_l = NULL;
577 FFTComplex *fft_in_r = NULL;
578 float *data_ir_l = NULL;
579 float *data_ir_r = NULL;
580 int offset = 0; /* used for faster pointer arithmetics in for-loop */
581 int i, j, azim_orig = azim, elev_orig = elev;
582 int filter_length, ret = 0;
586 s->sofa.easy = mysofa_open(s->filename, sample_rate, &filter_length, &ret);
587 if (!s->sofa.easy || ret) { /* if an invalid SOFA file has been selected */
588 av_log(ctx, AV_LOG_ERROR, "Selected SOFA file is invalid. Please select valid SOFA file.\n");
589 return AVERROR_INVALIDDATA;
592 av_log(ctx, AV_LOG_DEBUG, "IR length: %d.\n", s->sofa.easy->hrtf->N);
593 s->sofa.ir_samples = s->sofa.easy->hrtf->N;
594 s->sofa.n_samples = 1 << (32 - ff_clz(s->sofa.ir_samples));
596 n_samples = s->sofa.n_samples;
597 ir_samples = s->sofa.ir_samples;
599 s->data_ir[0] = av_calloc(n_samples, sizeof(float) * s->n_conv);
600 s->data_ir[1] = av_calloc(n_samples, sizeof(float) * s->n_conv);
601 s->delay[0] = av_calloc(s->n_conv, sizeof(int));
602 s->delay[1] = av_calloc(s->n_conv, sizeof(int));
604 if (!s->data_ir[0] || !s->data_ir[1] || !s->delay[0] || !s->delay[1]) {
605 ret = AVERROR(ENOMEM);
609 /* get temporary IR for L and R channel */
610 data_ir_l = av_calloc(n_conv * n_samples, sizeof(*data_ir_l));
611 data_ir_r = av_calloc(n_conv * n_samples, sizeof(*data_ir_r));
612 if (!data_ir_r || !data_ir_l) {
613 ret = AVERROR(ENOMEM);
617 if (s->type == TIME_DOMAIN) {
618 s->temp_src[0] = av_calloc(n_samples, sizeof(float));
619 s->temp_src[1] = av_calloc(n_samples, sizeof(float));
620 if (!s->temp_src[0] || !s->temp_src[1]) {
621 ret = AVERROR(ENOMEM);
626 s->speaker_azim = av_calloc(s->n_conv, sizeof(*s->speaker_azim));
627 s->speaker_elev = av_calloc(s->n_conv, sizeof(*s->speaker_elev));
628 if (!s->speaker_azim || !s->speaker_elev) {
629 ret = AVERROR(ENOMEM);
633 /* get speaker positions */
634 if ((ret = get_speaker_pos(ctx, s->speaker_azim, s->speaker_elev)) < 0) {
635 av_log(ctx, AV_LOG_ERROR, "Couldn't get speaker positions. Input channel configuration not supported.\n");
639 for (i = 0; i < s->n_conv; i++) {
640 float coordinates[3];
642 /* load and store IRs and corresponding delays */
643 azim = (int)(s->speaker_azim[i] + azim_orig) % 360;
644 elev = (int)(s->speaker_elev[i] + elev_orig) % 90;
646 coordinates[0] = azim;
647 coordinates[1] = elev;
648 coordinates[2] = radius;
650 mysofa_s2c(coordinates);
652 /* get id of IR closest to desired position */
653 mysofa_getfilter_float(s->sofa.easy, coordinates[0], coordinates[1], coordinates[2],
654 data_ir_l + n_samples * i,
655 data_ir_r + n_samples * i,
658 s->delay[0][i] = delay_l * sample_rate;
659 s->delay[1][i] = delay_r * sample_rate;
661 s->sofa.max_delay = FFMAX3(s->sofa.max_delay, s->delay[0][i], s->delay[1][i]);
664 /* get size of ringbuffer (longest IR plus max. delay) */
665 /* then choose next power of 2 for performance optimization */
666 n_current = n_samples + s->sofa.max_delay;
667 /* length of longest IR plus max. delay */
668 n_max = FFMAX(n_max, n_current);
670 /* buffer length is longest IR plus max. delay -> next power of 2
671 (32 - count leading zeros gives required exponent) */
672 s->buffer_length = 1 << (32 - ff_clz(n_max));
673 s->n_fft = n_fft = 1 << (32 - ff_clz(n_max + s->framesize));
675 if (s->type == FREQUENCY_DOMAIN) {
676 av_fft_end(s->fft[0]);
677 av_fft_end(s->fft[1]);
678 s->fft[0] = av_fft_init(log2(s->n_fft), 0);
679 s->fft[1] = av_fft_init(log2(s->n_fft), 0);
680 av_fft_end(s->ifft[0]);
681 av_fft_end(s->ifft[1]);
682 s->ifft[0] = av_fft_init(log2(s->n_fft), 1);
683 s->ifft[1] = av_fft_init(log2(s->n_fft), 1);
685 if (!s->fft[0] || !s->fft[1] || !s->ifft[0] || !s->ifft[1]) {
686 av_log(ctx, AV_LOG_ERROR, "Unable to create FFT contexts of size %d.\n", s->n_fft);
687 ret = AVERROR(ENOMEM);
692 if (s->type == TIME_DOMAIN) {
693 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
694 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
696 /* get temporary HRTF memory for L and R channel */
697 data_hrtf_l = av_malloc_array(n_fft, sizeof(*data_hrtf_l) * n_conv);
698 data_hrtf_r = av_malloc_array(n_fft, sizeof(*data_hrtf_r) * n_conv);
699 if (!data_hrtf_r || !data_hrtf_l) {
700 ret = AVERROR(ENOMEM);
704 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float));
705 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float));
706 s->temp_fft[0] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
707 s->temp_fft[1] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
708 if (!s->temp_fft[0] || !s->temp_fft[1]) {
709 ret = AVERROR(ENOMEM);
714 if (!s->ringbuffer[0] || !s->ringbuffer[1]) {
715 ret = AVERROR(ENOMEM);
719 if (s->type == FREQUENCY_DOMAIN) {
720 fft_in_l = av_calloc(n_fft, sizeof(*fft_in_l));
721 fft_in_r = av_calloc(n_fft, sizeof(*fft_in_r));
722 if (!fft_in_l || !fft_in_r) {
723 ret = AVERROR(ENOMEM);
728 for (i = 0; i < s->n_conv; i++) {
731 offset = i * n_samples; /* no. samples already written */
733 lir = data_ir_l + offset;
734 rir = data_ir_r + offset;
736 if (s->type == TIME_DOMAIN) {
737 for (j = 0; j < ir_samples; j++) {
738 /* load reversed IRs of the specified source position
739 * sample-by-sample for left and right ear; and apply gain */
740 s->data_ir[0][offset + j] = lir[ir_samples - 1 - j] * gain_lin;
741 s->data_ir[1][offset + j] = rir[ir_samples - 1 - j] * gain_lin;
744 memset(fft_in_l, 0, n_fft * sizeof(*fft_in_l));
745 memset(fft_in_r, 0, n_fft * sizeof(*fft_in_r));
747 offset = i * n_fft; /* no. samples already written */
748 for (j = 0; j < ir_samples; j++) {
749 /* load non-reversed IRs of the specified source position
750 * sample-by-sample and apply gain,
751 * L channel is loaded to real part, R channel to imag part,
752 * IRs ared shifted by L and R delay */
753 fft_in_l[s->delay[0][i] + j].re = lir[j] * gain_lin;
754 fft_in_r[s->delay[1][i] + j].re = rir[j] * gain_lin;
757 /* actually transform to frequency domain (IRs -> HRTFs) */
758 av_fft_permute(s->fft[0], fft_in_l);
759 av_fft_calc(s->fft[0], fft_in_l);
760 memcpy(data_hrtf_l + offset, fft_in_l, n_fft * sizeof(*fft_in_l));
761 av_fft_permute(s->fft[0], fft_in_r);
762 av_fft_calc(s->fft[0], fft_in_r);
763 memcpy(data_hrtf_r + offset, fft_in_r, n_fft * sizeof(*fft_in_r));
767 if (s->type == FREQUENCY_DOMAIN) {
768 s->data_hrtf[0] = av_malloc_array(n_fft * s->n_conv, sizeof(FFTComplex));
769 s->data_hrtf[1] = av_malloc_array(n_fft * s->n_conv, sizeof(FFTComplex));
770 if (!s->data_hrtf[0] || !s->data_hrtf[1]) {
771 ret = AVERROR(ENOMEM);
775 memcpy(s->data_hrtf[0], data_hrtf_l, /* copy HRTF data to */
776 sizeof(FFTComplex) * n_conv * n_fft); /* filter struct */
777 memcpy(s->data_hrtf[1], data_hrtf_r,
778 sizeof(FFTComplex) * n_conv * n_fft);
782 av_freep(&data_hrtf_l); /* free temporary HRTF memory */
783 av_freep(&data_hrtf_r);
785 av_freep(&data_ir_l); /* free temprary IR memory */
786 av_freep(&data_ir_r);
788 av_freep(&fft_in_l); /* free temporary FFT memory */
794 static av_cold int init(AVFilterContext *ctx)
796 SOFAlizerContext *s = ctx->priv;
800 av_log(ctx, AV_LOG_ERROR, "Valid SOFA filename must be set.\n");
801 return AVERROR(EINVAL);
804 /* preload SOFA file, */
805 ret = preload_sofa(ctx, s->filename, &s->sample_rate);
807 /* file loading error */
808 av_log(ctx, AV_LOG_ERROR, "Error while loading SOFA file: '%s'\n", s->filename);
809 } else { /* no file loading error, resampling not required */
810 av_log(ctx, AV_LOG_DEBUG, "File '%s' loaded.\n", s->filename);
814 av_log(ctx, AV_LOG_ERROR, "No valid SOFA file could be loaded. Please specify valid SOFA file.\n");
818 s->fdsp = avpriv_float_dsp_alloc(0);
820 return AVERROR(ENOMEM);
825 static int config_input(AVFilterLink *inlink)
827 AVFilterContext *ctx = inlink->dst;
828 SOFAlizerContext *s = ctx->priv;
831 if (s->type == FREQUENCY_DOMAIN) {
832 inlink->partial_buf_size =
833 inlink->min_samples =
834 inlink->max_samples = s->framesize;
837 /* gain -3 dB per channel, -6 dB to get LFE on a similar level */
838 s->gain_lfe = expf((s->gain - 3 * inlink->channels - 6 + s->lfe_gain) / 20 * M_LN10);
840 s->n_conv = inlink->channels;
842 /* load IRs to data_ir[0] and data_ir[1] for required directions */
843 if ((ret = load_data(ctx, s->rotation, s->elevation, s->radius, inlink->sample_rate)) < 0)
846 av_log(ctx, AV_LOG_DEBUG, "Samplerate: %d Channels to convolute: %d, Length of ringbuffer: %d x %d\n",
847 inlink->sample_rate, s->n_conv, inlink->channels, s->buffer_length);
852 static av_cold void uninit(AVFilterContext *ctx)
854 SOFAlizerContext *s = ctx->priv;
856 close_sofa(&s->sofa);
857 av_fft_end(s->ifft[0]);
858 av_fft_end(s->ifft[1]);
859 av_fft_end(s->fft[0]);
860 av_fft_end(s->fft[1]);
865 av_freep(&s->delay[0]);
866 av_freep(&s->delay[1]);
867 av_freep(&s->data_ir[0]);
868 av_freep(&s->data_ir[1]);
869 av_freep(&s->ringbuffer[0]);
870 av_freep(&s->ringbuffer[1]);
871 av_freep(&s->speaker_azim);
872 av_freep(&s->speaker_elev);
873 av_freep(&s->temp_src[0]);
874 av_freep(&s->temp_src[1]);
875 av_freep(&s->temp_fft[0]);
876 av_freep(&s->temp_fft[1]);
877 av_freep(&s->data_hrtf[0]);
878 av_freep(&s->data_hrtf[1]);
882 #define OFFSET(x) offsetof(SOFAlizerContext, x)
883 #define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
885 static const AVOption sofalizer_options[] = {
886 { "sofa", "sofa filename", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
887 { "gain", "set gain in dB", OFFSET(gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -20, 40, .flags = FLAGS },
888 { "rotation", "set rotation" , OFFSET(rotation), AV_OPT_TYPE_FLOAT, {.dbl=0}, -360, 360, .flags = FLAGS },
889 { "elevation", "set elevation", OFFSET(elevation), AV_OPT_TYPE_FLOAT, {.dbl=0}, -90, 90, .flags = FLAGS },
890 { "radius", "set radius", OFFSET(radius), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 5, .flags = FLAGS },
891 { "type", "set processing", OFFSET(type), AV_OPT_TYPE_INT, {.i64=1}, 0, 1, .flags = FLAGS, "type" },
892 { "time", "time domain", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, .flags = FLAGS, "type" },
893 { "freq", "frequency domain", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, .flags = FLAGS, "type" },
894 { "speakers", "set speaker custom positions", OFFSET(speakers_pos), AV_OPT_TYPE_STRING, {.str=0}, 0, 0, .flags = FLAGS },
895 { "lfegain", "set lfe gain", OFFSET(lfe_gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -20,40, .flags = FLAGS },
896 { "framesize", "set frame size", OFFSET(framesize), AV_OPT_TYPE_INT, {.i64=1024},1024,96000, .flags = FLAGS },
900 AVFILTER_DEFINE_CLASS(sofalizer);
902 static const AVFilterPad inputs[] = {
905 .type = AVMEDIA_TYPE_AUDIO,
906 .config_props = config_input,
907 .filter_frame = filter_frame,
912 static const AVFilterPad outputs[] = {
915 .type = AVMEDIA_TYPE_AUDIO,
920 AVFilter ff_af_sofalizer = {
922 .description = NULL_IF_CONFIG_SMALL("SOFAlizer (Spatially Oriented Format for Acoustics)."),
923 .priv_size = sizeof(SOFAlizerContext),
924 .priv_class = &sofalizer_class,
927 .query_formats = query_formats,
930 .flags = AVFILTER_FLAG_SLICE_THREADS,