2 * Copyright (c) 2017 Paul B Mahol
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 * An arbitrary audio FIR filter
26 #include "libavutil/audio_fifo.h"
27 #include "libavutil/common.h"
28 #include "libavutil/float_dsp.h"
29 #include "libavutil/opt.h"
30 #include "libavcodec/avfft.h"
38 static void fcmul_add_c(float *sum, const float *t, const float *c, ptrdiff_t len)
42 for (n = 0; n < len; n++) {
43 const float cre = c[2 * n ];
44 const float cim = c[2 * n + 1];
45 const float tre = t[2 * n ];
46 const float tim = t[2 * n + 1];
48 sum[2 * n ] += tre * cre - tim * cim;
49 sum[2 * n + 1] += tre * cim + tim * cre;
52 sum[2 * n] += t[2 * n] * c[2 * n];
55 static int fir_channel(AVFilterContext *ctx, void *arg, int ch, int nb_jobs)
57 AudioFIRContext *s = ctx->priv;
58 const float *src = (const float *)s->in[0]->extended_data[ch];
59 int index1 = (s->index + 1) % 3;
60 int index2 = (s->index + 2) % 3;
61 float *sum = s->sum[ch];
67 memset(sum, 0, sizeof(*sum) * s->fft_length);
68 block = s->block[ch] + s->part_index * s->block_size;
69 memset(block, 0, sizeof(*block) * s->fft_length);
71 s->fdsp->vector_fmul_scalar(block + s->part_size, src, s->dry_gain, FFALIGN(s->nb_samples, 4));
74 av_rdft_calc(s->rdft[ch], block);
75 block[2 * s->part_size] = block[1];
80 for (i = 0; i < s->nb_partitions; i++) {
81 const int coffset = i * s->coeff_size;
82 const FFTComplex *coeff = s->coeff[ch * !s->one2many] + coffset;
84 block = s->block[ch] + j * s->block_size;
85 s->fcmul_add(sum, block, (const float *)coeff, s->part_size);
92 sum[1] = sum[2 * s->part_size];
93 av_rdft_calc(s->irdft[ch], sum);
95 dst = (float *)s->buffer->extended_data[ch] + index1 * s->part_size;
96 for (n = 0; n < s->part_size; n++) {
100 dst = (float *)s->buffer->extended_data[ch] + index2 * s->part_size;
102 memcpy(dst, sum + s->part_size, s->part_size * sizeof(*dst));
104 dst = (float *)s->buffer->extended_data[ch] + s->index * s->part_size;
107 float *ptr = (float *)out->extended_data[ch];
108 s->fdsp->vector_fmul_scalar(ptr, dst, s->gain * s->wet_gain, FFALIGN(out->nb_samples, 4));
115 static int fir_frame(AudioFIRContext *s, AVFilterLink *outlink)
117 AVFilterContext *ctx = outlink->src;
121 s->nb_samples = FFMIN(s->part_size, av_audio_fifo_size(s->fifo[0]));
124 out = ff_get_audio_buffer(outlink, s->nb_samples);
126 return AVERROR(ENOMEM);
129 s->in[0] = ff_get_audio_buffer(ctx->inputs[0], s->nb_samples);
132 return AVERROR(ENOMEM);
135 av_audio_fifo_peek(s->fifo[0], (void **)s->in[0]->extended_data, s->nb_samples);
137 ctx->internal->execute(ctx, fir_channel, out, NULL, outlink->channels);
139 s->part_index = (s->part_index + 1) % s->nb_partitions;
141 av_audio_fifo_drain(s->fifo[0], s->nb_samples);
145 if (s->pts != AV_NOPTS_VALUE)
146 s->pts += av_rescale_q(out->nb_samples, (AVRational){1, outlink->sample_rate}, outlink->time_base);
153 av_frame_free(&s->in[0]);
155 if (s->want_skip == 1) {
159 ret = ff_filter_frame(outlink, out);
165 static int convert_coeffs(AVFilterContext *ctx)
167 AudioFIRContext *s = ctx->priv;
171 s->nb_taps = av_audio_fifo_size(s->fifo[1]);
173 return AVERROR(EINVAL);
175 for (n = 4; (1 << n) < s->nb_taps; n++);
177 s->ir_length = 1 << n;
178 s->fft_length = (1 << (N + 1)) + 1;
179 s->part_size = 1 << (N - 1);
180 s->block_size = FFALIGN(s->fft_length, 32);
181 s->coeff_size = FFALIGN(s->part_size + 1, 32);
182 s->nb_partitions = (s->nb_taps + s->part_size - 1) / s->part_size;
183 s->nb_coeffs = s->ir_length + s->nb_partitions;
185 for (ch = 0; ch < ctx->inputs[0]->channels; ch++) {
186 s->sum[ch] = av_calloc(s->fft_length, sizeof(**s->sum));
188 return AVERROR(ENOMEM);
191 for (ch = 0; ch < ctx->inputs[1]->channels; ch++) {
192 s->coeff[ch] = av_calloc(s->nb_partitions * s->coeff_size, sizeof(**s->coeff));
194 return AVERROR(ENOMEM);
197 for (ch = 0; ch < ctx->inputs[0]->channels; ch++) {
198 s->block[ch] = av_calloc(s->nb_partitions * s->block_size, sizeof(**s->block));
200 return AVERROR(ENOMEM);
203 for (ch = 0; ch < ctx->inputs[0]->channels; ch++) {
204 s->rdft[ch] = av_rdft_init(N, DFT_R2C);
205 s->irdft[ch] = av_rdft_init(N, IDFT_C2R);
206 if (!s->rdft[ch] || !s->irdft[ch])
207 return AVERROR(ENOMEM);
210 s->in[1] = ff_get_audio_buffer(ctx->inputs[1], s->nb_taps);
212 return AVERROR(ENOMEM);
214 s->buffer = ff_get_audio_buffer(ctx->inputs[0], s->part_size * 3);
216 return AVERROR(ENOMEM);
218 av_audio_fifo_read(s->fifo[1], (void **)s->in[1]->extended_data, s->nb_taps);
220 for (ch = 0; ch < ctx->inputs[1]->channels; ch++) {
221 float *time = (float *)s->in[1]->extended_data[!s->one2many * ch];
222 float *block = s->block[ch];
223 FFTComplex *coeff = s->coeff[ch];
225 power += s->fdsp->scalarproduct_float(time, time, s->nb_taps);
227 for (i = FFMAX(1, s->length * s->nb_taps); i < s->nb_taps; i++)
230 for (i = 0; i < s->nb_partitions; i++) {
231 const float scale = 1.f / s->part_size;
232 const int toffset = i * s->part_size;
233 const int coffset = i * s->coeff_size;
234 const int boffset = s->part_size;
235 const int remaining = s->nb_taps - (i * s->part_size);
236 const int size = remaining >= s->part_size ? s->part_size : remaining;
238 memset(block, 0, sizeof(*block) * s->fft_length);
239 memcpy(block + boffset, time + toffset, size * sizeof(*block));
241 av_rdft_calc(s->rdft[0], block);
243 coeff[coffset].re = block[0] * scale;
244 coeff[coffset].im = 0;
245 for (n = 1; n < s->part_size; n++) {
246 coeff[coffset + n].re = block[2 * n] * scale;
247 coeff[coffset + n].im = block[2 * n + 1] * scale;
249 coeff[coffset + s->part_size].re = block[1] * scale;
250 coeff[coffset + s->part_size].im = 0;
254 av_frame_free(&s->in[1]);
255 s->gain = s->again ? 1.f / sqrtf(power / ctx->inputs[1]->channels) : 1.f;
256 av_log(ctx, AV_LOG_DEBUG, "nb_taps: %d\n", s->nb_taps);
257 av_log(ctx, AV_LOG_DEBUG, "nb_partitions: %d\n", s->nb_partitions);
258 av_log(ctx, AV_LOG_DEBUG, "partition size: %d\n", s->part_size);
259 av_log(ctx, AV_LOG_DEBUG, "ir_length: %d\n", s->ir_length);
266 static int read_ir(AVFilterLink *link, AVFrame *frame)
268 AVFilterContext *ctx = link->dst;
269 AudioFIRContext *s = ctx->priv;
270 int nb_taps, max_nb_taps;
272 av_audio_fifo_write(s->fifo[1], (void **)frame->extended_data,
274 av_frame_free(&frame);
276 nb_taps = av_audio_fifo_size(s->fifo[1]);
277 max_nb_taps = MAX_IR_DURATION * ctx->outputs[0]->sample_rate;
278 if (nb_taps > max_nb_taps) {
279 av_log(ctx, AV_LOG_ERROR, "Too big number of coefficients: %d > %d.\n", nb_taps, max_nb_taps);
280 return AVERROR(EINVAL);
286 static int filter_frame(AVFilterLink *link, AVFrame *frame)
288 AVFilterContext *ctx = link->dst;
289 AudioFIRContext *s = ctx->priv;
290 AVFilterLink *outlink = ctx->outputs[0];
293 av_audio_fifo_write(s->fifo[0], (void **)frame->extended_data,
295 if (s->pts == AV_NOPTS_VALUE)
298 av_frame_free(&frame);
300 if (!s->have_coeffs && s->eof_coeffs) {
301 ret = convert_coeffs(ctx);
306 if (s->have_coeffs) {
307 while (av_audio_fifo_size(s->fifo[0]) >= s->part_size) {
308 ret = fir_frame(s, outlink);
316 static int request_frame(AVFilterLink *outlink)
318 AVFilterContext *ctx = outlink->src;
319 AudioFIRContext *s = ctx->priv;
322 if (!s->eof_coeffs) {
323 ret = ff_request_frame(ctx->inputs[1]);
324 if (ret == AVERROR_EOF) {
330 ret = ff_request_frame(ctx->inputs[0]);
331 if (ret == AVERROR_EOF && s->have_coeffs) {
332 if (s->need_padding) {
333 AVFrame *silence = ff_get_audio_buffer(outlink, s->part_size);
336 return AVERROR(ENOMEM);
337 av_audio_fifo_write(s->fifo[0], (void **)silence->extended_data,
338 silence->nb_samples);
339 av_frame_free(&silence);
343 while (av_audio_fifo_size(s->fifo[0]) > 0) {
344 ret = fir_frame(s, outlink);
353 static int query_formats(AVFilterContext *ctx)
355 AVFilterFormats *formats;
356 AVFilterChannelLayouts *layouts;
357 static const enum AVSampleFormat sample_fmts[] = {
363 layouts = ff_all_channel_counts();
364 if ((ret = ff_channel_layouts_ref(layouts, &ctx->outputs[0]->in_channel_layouts)) < 0)
367 for (i = 0; i < 2; i++) {
368 layouts = ff_all_channel_counts();
369 if ((ret = ff_channel_layouts_ref(layouts, &ctx->inputs[i]->out_channel_layouts)) < 0)
373 formats = ff_make_format_list(sample_fmts);
374 if ((ret = ff_set_common_formats(ctx, formats)) < 0)
377 formats = ff_all_samplerates();
378 return ff_set_common_samplerates(ctx, formats);
381 static int config_output(AVFilterLink *outlink)
383 AVFilterContext *ctx = outlink->src;
384 AudioFIRContext *s = ctx->priv;
386 if (ctx->inputs[0]->channels != ctx->inputs[1]->channels &&
387 ctx->inputs[1]->channels != 1) {
388 av_log(ctx, AV_LOG_ERROR,
389 "Second input must have same number of channels as first input or "
390 "exactly 1 channel.\n");
391 return AVERROR(EINVAL);
394 s->one2many = ctx->inputs[1]->channels == 1;
395 outlink->sample_rate = ctx->inputs[0]->sample_rate;
396 outlink->time_base = ctx->inputs[0]->time_base;
397 outlink->channel_layout = ctx->inputs[0]->channel_layout;
398 outlink->channels = ctx->inputs[0]->channels;
400 s->fifo[0] = av_audio_fifo_alloc(ctx->inputs[0]->format, ctx->inputs[0]->channels, 1024);
401 s->fifo[1] = av_audio_fifo_alloc(ctx->inputs[1]->format, ctx->inputs[1]->channels, 1024);
402 if (!s->fifo[0] || !s->fifo[1])
403 return AVERROR(ENOMEM);
405 s->sum = av_calloc(outlink->channels, sizeof(*s->sum));
406 s->coeff = av_calloc(ctx->inputs[1]->channels, sizeof(*s->coeff));
407 s->block = av_calloc(ctx->inputs[0]->channels, sizeof(*s->block));
408 s->rdft = av_calloc(outlink->channels, sizeof(*s->rdft));
409 s->irdft = av_calloc(outlink->channels, sizeof(*s->irdft));
410 if (!s->sum || !s->coeff || !s->block || !s->rdft || !s->irdft)
411 return AVERROR(ENOMEM);
413 s->nb_channels = outlink->channels;
414 s->nb_coef_channels = ctx->inputs[1]->channels;
417 s->pts = AV_NOPTS_VALUE;
422 static av_cold void uninit(AVFilterContext *ctx)
424 AudioFIRContext *s = ctx->priv;
428 for (ch = 0; ch < s->nb_channels; ch++) {
429 av_freep(&s->sum[ch]);
435 for (ch = 0; ch < s->nb_coef_channels; ch++) {
436 av_freep(&s->coeff[ch]);
442 for (ch = 0; ch < s->nb_channels; ch++) {
443 av_freep(&s->block[ch]);
449 for (ch = 0; ch < s->nb_channels; ch++) {
450 av_rdft_end(s->rdft[ch]);
456 for (ch = 0; ch < s->nb_channels; ch++) {
457 av_rdft_end(s->irdft[ch]);
462 av_frame_free(&s->in[0]);
463 av_frame_free(&s->in[1]);
464 av_frame_free(&s->buffer);
466 av_audio_fifo_free(s->fifo[0]);
467 av_audio_fifo_free(s->fifo[1]);
472 static av_cold int init(AVFilterContext *ctx)
474 AudioFIRContext *s = ctx->priv;
476 s->fcmul_add = fcmul_add_c;
478 s->fdsp = avpriv_float_dsp_alloc(0);
480 return AVERROR(ENOMEM);
488 static const AVFilterPad afir_inputs[] = {
491 .type = AVMEDIA_TYPE_AUDIO,
492 .filter_frame = filter_frame,
495 .type = AVMEDIA_TYPE_AUDIO,
496 .filter_frame = read_ir,
501 static const AVFilterPad afir_outputs[] = {
504 .type = AVMEDIA_TYPE_AUDIO,
505 .config_props = config_output,
506 .request_frame = request_frame,
511 #define AF AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
512 #define OFFSET(x) offsetof(AudioFIRContext, x)
514 static const AVOption afir_options[] = {
515 { "dry", "set dry gain", OFFSET(dry_gain), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 1, AF },
516 { "wet", "set wet gain", OFFSET(wet_gain), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 1, AF },
517 { "length", "set IR length", OFFSET(length), AV_OPT_TYPE_FLOAT, {.dbl=1}, 0, 1, AF },
518 { "again", "enable auto gain", OFFSET(again), AV_OPT_TYPE_BOOL, {.i64=1}, 0, 1, AF },
522 AVFILTER_DEFINE_CLASS(afir);
524 AVFilter ff_af_afir = {
526 .description = NULL_IF_CONFIG_SMALL("Apply Finite Impulse Response filter with supplied coefficients in 2nd stream."),
527 .priv_size = sizeof(AudioFIRContext),
528 .priv_class = &afir_class,
529 .query_formats = query_formats,
532 .inputs = afir_inputs,
533 .outputs = afir_outputs,
534 .flags = AVFILTER_FLAG_SLICE_THREADS,