2 * Copyright (C) 2017 Paul B Mahol
3 * Copyright (C) 2013-2015 Andreas Fuchs, Wolfgang Hrauda
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #include "libavutil/avstring.h"
24 #include "libavutil/channel_layout.h"
25 #include "libavutil/float_dsp.h"
26 #include "libavutil/intmath.h"
27 #include "libavutil/opt.h"
28 #include "libavcodec/avfft.h"
36 #define FREQUENCY_DOMAIN 1
41 typedef struct HeadphoneContext {
60 float lfe_gain, gain_lfe;
72 FFTComplex *temp_fft[2];
73 FFTComplex *temp_afft[2];
75 FFTContext *fft[2], *ifft[2];
76 FFTComplex *data_hrtf[2];
78 float (*scalarproduct_float)(const float *v1, const float *v2, int len);
86 static int parse_channel_name(const char *arg, uint64_t *rchannel)
88 uint64_t layout = av_get_channel_layout(arg);
90 if (av_get_channel_layout_nb_channels(layout) != 1)
91 return AVERROR(EINVAL);
96 static void parse_map(AVFilterContext *ctx)
98 HeadphoneContext *s = ctx->priv;
99 char *arg, *tokenizer, *p;
100 uint64_t used_channels = 0;
103 while ((arg = av_strtok(p, "|", &tokenizer))) {
104 uint64_t out_channel;
107 if (parse_channel_name(arg, &out_channel)) {
108 av_log(ctx, AV_LOG_WARNING, "Failed to parse \'%s\' as channel name.\n", arg);
111 if (used_channels & out_channel) {
112 av_log(ctx, AV_LOG_WARNING, "Ignoring duplicate channel '%s'.\n", arg);
115 used_channels |= out_channel;
116 s->mapping[s->nb_irs] = out_channel;
120 if (s->hrir_fmt == HRIR_MULTI)
121 s->nb_hrir_inputs = 1;
123 s->nb_hrir_inputs = s->nb_irs;
126 typedef struct ThreadData {
133 FFTComplex **temp_fft;
134 FFTComplex **temp_afft;
137 static int headphone_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
139 HeadphoneContext *s = ctx->priv;
140 ThreadData *td = arg;
141 AVFrame *in = td->in, *out = td->out;
143 int *write = &td->write[jobnr];
144 const float *const ir = td->ir[jobnr];
145 int *n_clippings = &td->n_clippings[jobnr];
146 float *ringbuffer = td->ringbuffer[jobnr];
147 float *temp_src = td->temp_src[jobnr];
148 const int ir_len = s->ir_len;
149 const int air_len = s->air_len;
150 const float *src = (const float *)in->data[0];
151 float *dst = (float *)out->data[0];
152 const int in_channels = in->channels;
153 const int buffer_length = s->buffer_length;
154 const uint32_t modulo = (uint32_t)buffer_length - 1;
161 for (l = 0; l < in_channels; l++) {
162 buffer[l] = ringbuffer + l * buffer_length;
165 for (i = 0; i < in->nb_samples; i++) {
166 const float *temp_ir = ir;
169 for (l = 0; l < in_channels; l++) {
170 *(buffer[l] + wr) = src[l];
173 for (l = 0; l < in_channels; l++) {
174 const float *const bptr = buffer[l];
176 if (l == s->lfe_channel) {
177 *dst += *(buffer[s->lfe_channel] + wr) * s->gain_lfe;
182 read = (wr - (ir_len - 1) + buffer_length) & modulo;
184 if (read + ir_len < buffer_length) {
185 memcpy(temp_src, bptr + read, ir_len * sizeof(*temp_src));
187 int len = FFMIN(air_len - (read % ir_len), buffer_length - read);
189 memcpy(temp_src, bptr + read, len * sizeof(*temp_src));
190 memcpy(temp_src + len, bptr, (air_len - len) * sizeof(*temp_src));
193 dst[0] += s->scalarproduct_float(temp_ir, temp_src, FFALIGN(ir_len, 32));
197 if (fabsf(dst[0]) > 1)
202 wr = (wr + 1) & modulo;
210 static int headphone_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
212 HeadphoneContext *s = ctx->priv;
213 ThreadData *td = arg;
214 AVFrame *in = td->in, *out = td->out;
216 int *write = &td->write[jobnr];
217 FFTComplex *hrtf = s->data_hrtf[jobnr];
218 int *n_clippings = &td->n_clippings[jobnr];
219 float *ringbuffer = td->ringbuffer[jobnr];
220 const int ir_len = s->ir_len;
221 const float *src = (const float *)in->data[0];
222 float *dst = (float *)out->data[0];
223 const int in_channels = in->channels;
224 const int buffer_length = s->buffer_length;
225 const uint32_t modulo = (uint32_t)buffer_length - 1;
226 FFTComplex *fft_in = s->temp_fft[jobnr];
227 FFTComplex *fft_acc = s->temp_afft[jobnr];
228 FFTContext *ifft = s->ifft[jobnr];
229 FFTContext *fft = s->fft[jobnr];
230 const int n_fft = s->n_fft;
231 const float fft_scale = 1.0f / s->n_fft;
232 FFTComplex *hrtf_offset;
239 n_read = FFMIN(ir_len, in->nb_samples);
240 for (j = 0; j < n_read; j++) {
241 dst[2 * j] = ringbuffer[wr];
242 ringbuffer[wr] = 0.0;
243 wr = (wr + 1) & modulo;
246 for (j = n_read; j < in->nb_samples; j++) {
250 memset(fft_acc, 0, sizeof(FFTComplex) * n_fft);
252 for (i = 0; i < in_channels; i++) {
253 if (i == s->lfe_channel) {
254 for (j = 0; j < in->nb_samples; j++) {
255 dst[2 * j] += src[i + j * in_channels] * s->gain_lfe;
261 hrtf_offset = hrtf + offset;
263 memset(fft_in, 0, sizeof(FFTComplex) * n_fft);
265 for (j = 0; j < in->nb_samples; j++) {
266 fft_in[j].re = src[j * in_channels + i];
269 av_fft_permute(fft, fft_in);
270 av_fft_calc(fft, fft_in);
271 for (j = 0; j < n_fft; j++) {
272 const FFTComplex *hcomplex = hrtf_offset + j;
273 const float re = fft_in[j].re;
274 const float im = fft_in[j].im;
276 fft_acc[j].re += re * hcomplex->re - im * hcomplex->im;
277 fft_acc[j].im += re * hcomplex->im + im * hcomplex->re;
281 av_fft_permute(ifft, fft_acc);
282 av_fft_calc(ifft, fft_acc);
284 for (j = 0; j < in->nb_samples; j++) {
285 dst[2 * j] += fft_acc[j].re * fft_scale;
288 for (j = 0; j < ir_len - 1; j++) {
289 int write_pos = (wr + j) & modulo;
291 *(ringbuffer + write_pos) += fft_acc[in->nb_samples + j].re * fft_scale;
294 for (i = 0; i < out->nb_samples; i++) {
295 if (fabsf(dst[0]) > 1) {
307 static int check_ir(AVFilterLink *inlink, int input_number)
309 AVFilterContext *ctx = inlink->dst;
310 HeadphoneContext *s = ctx->priv;
311 int ir_len, max_ir_len;
313 ir_len = ff_inlink_queued_samples(inlink);
315 if (ir_len > max_ir_len) {
316 av_log(ctx, AV_LOG_ERROR, "Too big length of IRs: %d > %d.\n", ir_len, max_ir_len);
317 return AVERROR(EINVAL);
319 s->hrir_in[input_number].ir_len = ir_len;
320 s->ir_len = FFMAX(ir_len, s->ir_len);
325 static int headphone_frame(HeadphoneContext *s, AVFrame *in, AVFilterLink *outlink)
327 AVFilterContext *ctx = outlink->src;
328 int n_clippings[2] = { 0 };
332 out = ff_get_audio_buffer(outlink, in->nb_samples);
335 return AVERROR(ENOMEM);
339 td.in = in; td.out = out; td.write = s->write;
340 td.ir = s->data_ir; td.n_clippings = n_clippings;
341 td.ringbuffer = s->ringbuffer; td.temp_src = s->temp_src;
342 td.temp_fft = s->temp_fft;
343 td.temp_afft = s->temp_afft;
345 if (s->type == TIME_DOMAIN) {
346 ctx->internal->execute(ctx, headphone_convolute, &td, NULL, 2);
348 ctx->internal->execute(ctx, headphone_fast_convolute, &td, NULL, 2);
352 if (n_clippings[0] + n_clippings[1] > 0) {
353 av_log(ctx, AV_LOG_WARNING, "%d of %d samples clipped. Please reduce gain.\n",
354 n_clippings[0] + n_clippings[1], out->nb_samples * 2);
358 return ff_filter_frame(outlink, out);
361 static int convert_coeffs(AVFilterContext *ctx, AVFilterLink *inlink)
363 struct HeadphoneContext *s = ctx->priv;
364 const int ir_len = s->ir_len;
365 int nb_input_channels = ctx->inputs[0]->channels;
366 float gain_lin = expf((s->gain - 3 * nb_input_channels) / 20 * M_LN10);
372 s->air_len = 1 << (32 - ff_clz(ir_len));
373 if (s->type == TIME_DOMAIN) {
374 s->air_len = FFALIGN(s->air_len, 32);
376 s->buffer_length = 1 << (32 - ff_clz(s->air_len));
377 s->n_fft = n_fft = 1 << (32 - ff_clz(ir_len + s->size));
379 if (s->type == FREQUENCY_DOMAIN) {
380 s->fft[0] = av_fft_init(av_log2(s->n_fft), 0);
381 s->fft[1] = av_fft_init(av_log2(s->n_fft), 0);
382 s->ifft[0] = av_fft_init(av_log2(s->n_fft), 1);
383 s->ifft[1] = av_fft_init(av_log2(s->n_fft), 1);
385 if (!s->fft[0] || !s->fft[1] || !s->ifft[0] || !s->ifft[1]) {
386 av_log(ctx, AV_LOG_ERROR, "Unable to create FFT contexts of size %d.\n", s->n_fft);
387 ret = AVERROR(ENOMEM);
392 if (s->type == TIME_DOMAIN) {
393 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
394 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float) * nb_input_channels);
396 s->ringbuffer[0] = av_calloc(s->buffer_length, sizeof(float));
397 s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float));
398 s->temp_fft[0] = av_calloc(s->n_fft, sizeof(FFTComplex));
399 s->temp_fft[1] = av_calloc(s->n_fft, sizeof(FFTComplex));
400 s->temp_afft[0] = av_calloc(s->n_fft, sizeof(FFTComplex));
401 s->temp_afft[1] = av_calloc(s->n_fft, sizeof(FFTComplex));
402 if (!s->temp_fft[0] || !s->temp_fft[1] ||
403 !s->temp_afft[0] || !s->temp_afft[1]) {
404 ret = AVERROR(ENOMEM);
409 if (!s->ringbuffer[0] || !s->ringbuffer[1]) {
410 ret = AVERROR(ENOMEM);
414 if (s->type == TIME_DOMAIN) {
415 s->temp_src[0] = av_calloc(s->air_len, sizeof(float));
416 s->temp_src[1] = av_calloc(s->air_len, sizeof(float));
418 s->data_ir[0] = av_calloc(nb_input_channels * s->air_len, sizeof(*s->data_ir[0]));
419 s->data_ir[1] = av_calloc(nb_input_channels * s->air_len, sizeof(*s->data_ir[1]));
420 if (!s->data_ir[0] || !s->data_ir[1] || !s->temp_src[0] || !s->temp_src[1]) {
421 ret = AVERROR(ENOMEM);
425 s->data_hrtf[0] = av_calloc(n_fft, sizeof(*s->data_hrtf[0]) * nb_input_channels);
426 s->data_hrtf[1] = av_calloc(n_fft, sizeof(*s->data_hrtf[1]) * nb_input_channels);
427 if (!s->data_hrtf[0] || !s->data_hrtf[1]) {
428 ret = AVERROR(ENOMEM);
433 for (i = 0; i < s->nb_hrir_inputs; av_frame_free(&frame), i++) {
434 int len = s->hrir_in[i].ir_len;
437 ret = ff_inlink_consume_samples(ctx->inputs[i + 1], len, len, &frame);
440 ptr = (float *)frame->extended_data[0];
442 if (s->hrir_fmt == HRIR_STEREO) {
443 int idx = av_get_channel_layout_channel_index(inlink->channel_layout,
447 if (s->type == TIME_DOMAIN) {
448 float *data_ir_l = s->data_ir[0] + idx * s->air_len;
449 float *data_ir_r = s->data_ir[1] + idx * s->air_len;
451 for (j = 0; j < len; j++) {
452 data_ir_l[j] = ptr[len * 2 - j * 2 - 2] * gain_lin;
453 data_ir_r[j] = ptr[len * 2 - j * 2 - 1] * gain_lin;
456 FFTComplex *fft_in_l = s->data_hrtf[0] + idx * n_fft;
457 FFTComplex *fft_in_r = s->data_hrtf[1] + idx * n_fft;
459 for (j = 0; j < len; j++) {
460 fft_in_l[j].re = ptr[j * 2 ] * gain_lin;
461 fft_in_r[j].re = ptr[j * 2 + 1] * gain_lin;
464 av_fft_permute(s->fft[0], fft_in_l);
465 av_fft_calc(s->fft[0], fft_in_l);
466 av_fft_permute(s->fft[0], fft_in_r);
467 av_fft_calc(s->fft[0], fft_in_r);
470 int I, N = ctx->inputs[1]->channels;
472 for (k = 0; k < N / 2; k++) {
473 int idx = av_get_channel_layout_channel_index(inlink->channel_layout,
479 if (s->type == TIME_DOMAIN) {
480 float *data_ir_l = s->data_ir[0] + idx * s->air_len;
481 float *data_ir_r = s->data_ir[1] + idx * s->air_len;
483 for (j = 0; j < len; j++) {
484 data_ir_l[j] = ptr[len * N - j * N - N + I ] * gain_lin;
485 data_ir_r[j] = ptr[len * N - j * N - N + I + 1] * gain_lin;
488 FFTComplex *fft_in_l = s->data_hrtf[0] + idx * n_fft;
489 FFTComplex *fft_in_r = s->data_hrtf[1] + idx * n_fft;
491 for (j = 0; j < len; j++) {
492 fft_in_l[j].re = ptr[j * N + I ] * gain_lin;
493 fft_in_r[j].re = ptr[j * N + I + 1] * gain_lin;
496 av_fft_permute(s->fft[0], fft_in_l);
497 av_fft_calc(s->fft[0], fft_in_l);
498 av_fft_permute(s->fft[0], fft_in_r);
499 av_fft_calc(s->fft[0], fft_in_r);
511 static int activate(AVFilterContext *ctx)
513 HeadphoneContext *s = ctx->priv;
514 AVFilterLink *inlink = ctx->inputs[0];
515 AVFilterLink *outlink = ctx->outputs[0];
519 FF_FILTER_FORWARD_STATUS_BACK_ALL(ctx->outputs[0], ctx);
522 for (i = 0; i < s->nb_hrir_inputs; i++) {
523 AVFilterLink *input = ctx->inputs[i + 1];
525 if (s->hrir_in[i].eof)
528 if ((ret = check_ir(input, i)) < 0)
531 if (ff_outlink_get_status(input) == AVERROR_EOF) {
532 if (!ff_inlink_queued_samples(input)) {
533 av_log(ctx, AV_LOG_ERROR, "No samples provided for "
534 "HRIR stream %d.\n", i);
535 return AVERROR_INVALIDDATA;
537 s->hrir_in[i].eof = 1;
539 if (ff_outlink_frame_wanted(ctx->outputs[0]))
540 ff_inlink_request_frame(input);
548 ret = convert_coeffs(ctx, inlink);
551 } else if (!s->have_hrirs)
554 if ((ret = ff_inlink_consume_samples(ctx->inputs[0], s->size, s->size, &in)) > 0) {
555 ret = headphone_frame(s, in, outlink);
563 FF_FILTER_FORWARD_STATUS(ctx->inputs[0], ctx->outputs[0]);
564 if (ff_outlink_frame_wanted(ctx->outputs[0]))
565 ff_inlink_request_frame(ctx->inputs[0]);
570 static int query_formats(AVFilterContext *ctx)
572 struct HeadphoneContext *s = ctx->priv;
573 AVFilterFormats *formats = NULL;
574 AVFilterChannelLayouts *layouts = NULL;
575 AVFilterChannelLayouts *stereo_layout = NULL;
576 AVFilterChannelLayouts *hrir_layouts = NULL;
579 ret = ff_add_format(&formats, AV_SAMPLE_FMT_FLT);
582 ret = ff_set_common_formats(ctx, formats);
586 layouts = ff_all_channel_layouts();
588 return AVERROR(ENOMEM);
590 ret = ff_channel_layouts_ref(layouts, &ctx->inputs[0]->outcfg.channel_layouts);
594 ret = ff_add_channel_layout(&stereo_layout, AV_CH_LAYOUT_STEREO);
597 ret = ff_channel_layouts_ref(stereo_layout, &ctx->outputs[0]->incfg.channel_layouts);
601 if (s->hrir_fmt == HRIR_MULTI) {
602 hrir_layouts = ff_all_channel_counts();
604 return AVERROR(ENOMEM);
605 ret = ff_channel_layouts_ref(hrir_layouts, &ctx->inputs[1]->outcfg.channel_layouts);
609 for (i = 1; i <= s->nb_hrir_inputs; i++) {
610 ret = ff_channel_layouts_ref(stereo_layout, &ctx->inputs[i]->outcfg.channel_layouts);
616 formats = ff_all_samplerates();
618 return AVERROR(ENOMEM);
619 return ff_set_common_samplerates(ctx, formats);
622 static int config_input(AVFilterLink *inlink)
624 AVFilterContext *ctx = inlink->dst;
625 HeadphoneContext *s = ctx->priv;
627 if (s->nb_irs < inlink->channels) {
628 av_log(ctx, AV_LOG_ERROR, "Number of HRIRs must be >= %d.\n", inlink->channels);
629 return AVERROR(EINVAL);
632 s->lfe_channel = av_get_channel_layout_channel_index(inlink->channel_layout,
633 AV_CH_LOW_FREQUENCY);
637 static av_cold int init(AVFilterContext *ctx)
639 HeadphoneContext *s = ctx->priv;
644 .type = AVMEDIA_TYPE_AUDIO,
645 .config_props = config_input,
647 if ((ret = ff_insert_inpad(ctx, 0, &pad)) < 0)
651 av_log(ctx, AV_LOG_ERROR, "Valid mapping must be set.\n");
652 return AVERROR(EINVAL);
657 for (i = 0; i < s->nb_hrir_inputs; i++) {
658 char *name = av_asprintf("hrir%d", i);
661 .type = AVMEDIA_TYPE_AUDIO,
664 return AVERROR(ENOMEM);
665 if ((ret = ff_insert_inpad(ctx, i + 1, &pad)) < 0) {
671 if (s->type == TIME_DOMAIN) {
672 AVFloatDSPContext *fdsp = avpriv_float_dsp_alloc(0);
674 return AVERROR(ENOMEM);
675 s->scalarproduct_float = fdsp->scalarproduct_float;
682 static int config_output(AVFilterLink *outlink)
684 AVFilterContext *ctx = outlink->src;
685 HeadphoneContext *s = ctx->priv;
686 AVFilterLink *inlink = ctx->inputs[0];
688 if (s->hrir_fmt == HRIR_MULTI) {
689 AVFilterLink *hrir_link = ctx->inputs[1];
691 if (hrir_link->channels < inlink->channels * 2) {
692 av_log(ctx, AV_LOG_ERROR, "Number of channels in HRIR stream must be >= %d.\n", inlink->channels * 2);
693 return AVERROR(EINVAL);
697 s->gain_lfe = expf((s->gain - 3 * inlink->channels + s->lfe_gain) / 20 * M_LN10);
702 static av_cold void uninit(AVFilterContext *ctx)
704 HeadphoneContext *s = ctx->priv;
706 av_fft_end(s->ifft[0]);
707 av_fft_end(s->ifft[1]);
708 av_fft_end(s->fft[0]);
709 av_fft_end(s->fft[1]);
710 av_freep(&s->data_ir[0]);
711 av_freep(&s->data_ir[1]);
712 av_freep(&s->ringbuffer[0]);
713 av_freep(&s->ringbuffer[1]);
714 av_freep(&s->temp_src[0]);
715 av_freep(&s->temp_src[1]);
716 av_freep(&s->temp_fft[0]);
717 av_freep(&s->temp_fft[1]);
718 av_freep(&s->temp_afft[0]);
719 av_freep(&s->temp_afft[1]);
720 av_freep(&s->data_hrtf[0]);
721 av_freep(&s->data_hrtf[1]);
723 for (unsigned i = 1; i < ctx->nb_inputs; i++)
724 av_freep(&ctx->input_pads[i].name);
727 #define OFFSET(x) offsetof(HeadphoneContext, x)
728 #define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
730 static const AVOption headphone_options[] = {
731 { "map", "set channels convolution mappings", OFFSET(map), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
732 { "gain", "set gain in dB", OFFSET(gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -20, 40, .flags = FLAGS },
733 { "lfe", "set lfe gain in dB", OFFSET(lfe_gain), AV_OPT_TYPE_FLOAT, {.dbl=0}, -20, 40, .flags = FLAGS },
734 { "type", "set processing", OFFSET(type), AV_OPT_TYPE_INT, {.i64=1}, 0, 1, .flags = FLAGS, "type" },
735 { "time", "time domain", 0, AV_OPT_TYPE_CONST, {.i64=0}, 0, 0, .flags = FLAGS, "type" },
736 { "freq", "frequency domain", 0, AV_OPT_TYPE_CONST, {.i64=1}, 0, 0, .flags = FLAGS, "type" },
737 { "size", "set frame size", OFFSET(size), AV_OPT_TYPE_INT, {.i64=1024},1024,96000, .flags = FLAGS },
738 { "hrir", "set hrir format", OFFSET(hrir_fmt), AV_OPT_TYPE_INT, {.i64=HRIR_STEREO}, 0, 1, .flags = FLAGS, "hrir" },
739 { "stereo", "hrir files have exactly 2 channels", 0, AV_OPT_TYPE_CONST, {.i64=HRIR_STEREO}, 0, 0, .flags = FLAGS, "hrir" },
740 { "multich", "single multichannel hrir file", 0, AV_OPT_TYPE_CONST, {.i64=HRIR_MULTI}, 0, 0, .flags = FLAGS, "hrir" },
744 AVFILTER_DEFINE_CLASS(headphone);
746 static const AVFilterPad outputs[] = {
749 .type = AVMEDIA_TYPE_AUDIO,
750 .config_props = config_output,
755 AVFilter ff_af_headphone = {
757 .description = NULL_IF_CONFIG_SMALL("Apply headphone binaural spatialization with HRTFs in additional streams."),
758 .priv_size = sizeof(HeadphoneContext),
759 .priv_class = &headphone_class,
762 .query_formats = query_formats,
763 .activate = activate,
766 .flags = AVFILTER_FLAG_SLICE_THREADS | AVFILTER_FLAG_DYNAMIC_INPUTS,