[FFmpeg-devel] [PATCH] avfilter: add arbitrary audio FIR filter
Paul B Mahol
onemda at gmail.com
Tue May 9 01:03:27 EEST 2017
On 5/8/17, Muhammad Faiz <mfcc64 at gmail.com> wrote:
> On Mon, May 8, 2017 at 11:06 PM, Paul B Mahol <onemda at gmail.com> wrote:
>> On 5/8/17, Muhammad Faiz <mfcc64 at gmail.com> wrote:
>>> On Mon, May 8, 2017 at 6:59 PM, Paul B Mahol <onemda at gmail.com> wrote:
>>>> Signed-off-by: Paul B Mahol <onemda at gmail.com>
>>>> ---
>>>> configure | 2 +
>>>> doc/filters.texi | 23 ++
>>>> libavfilter/Makefile | 1 +
>>>> libavfilter/af_afir.c | 544
>>>> +++++++++++++++++++++++++++++++++++++++++++++++
>>>> libavfilter/allfilters.c | 1 +
>>>> 5 files changed, 571 insertions(+)
>>>> create mode 100644 libavfilter/af_afir.c
>>>>
>>>> diff --git a/configure b/configure
>>>> index 2e1786a..a46c375 100755
>>>> --- a/configure
>>>> +++ b/configure
>>>> @@ -3081,6 +3081,8 @@ unix_protocol_select="network"
>>>> # filters
>>>> afftfilt_filter_deps="avcodec"
>>>> afftfilt_filter_select="fft"
>>>> +afir_filter_deps="avcodec"
>>>> +afir_filter_select="fft"
>>>> amovie_filter_deps="avcodec avformat"
>>>> aresample_filter_deps="swresample"
>>>> ass_filter_deps="libass"
>>>> diff --git a/doc/filters.texi b/doc/filters.texi
>>>> index f431274..0efce9a 100644
>>>> --- a/doc/filters.texi
>>>> +++ b/doc/filters.texi
>>>> @@ -878,6 +878,29 @@ afftfilt="1-clip((b/nb)*b,0,1)"
>>>> @end example
>>>> @end itemize
>>>>
>>>> + at section afir
>>>> +
>>>> +Apply an Arbitary Frequency Impulse Response filter.
>>>> +
>>>> +This filter uses second stream as FIR coefficients.
>>>> +If second stream holds single channel, it will be used
>>>> +for all input channels in first stream, otherwise
>>>> +number of channels in second stream must be same as
>>>> +number of channels in first stream.
>>>> +
>>>> +It accepts the following parameters:
>>>> +
>>>> + at table @option
>>>> + at item dry
>>>> +Set dry gain. This sets input gain.
>>>> +
>>>> + at item wet
>>>> +Set wet gain. This sets final output gain.
>>>> +
>>>> + at item length
>>>> +Set Impulse Response filter length. Default is 1, which means whole IR
>>>> is
>>>> processed.
>>>> + at end table
>>>> +
>>>> @anchor{aformat}
>>>> @section aformat
>>>>
>>>> diff --git a/libavfilter/Makefile b/libavfilter/Makefile
>>>> index 0f99086..de5f992 100644
>>>> --- a/libavfilter/Makefile
>>>> +++ b/libavfilter/Makefile
>>>> @@ -37,6 +37,7 @@ OBJS-$(CONFIG_AEMPHASIS_FILTER) +=
>>>> af_aemphasis.o
>>>> OBJS-$(CONFIG_AEVAL_FILTER) += aeval.o
>>>> OBJS-$(CONFIG_AFADE_FILTER) += af_afade.o
>>>> OBJS-$(CONFIG_AFFTFILT_FILTER) += af_afftfilt.o
>>>> window_func.o
>>>> +OBJS-$(CONFIG_AFIR_FILTER) += af_afir.o
>>>> OBJS-$(CONFIG_AFORMAT_FILTER) += af_aformat.o
>>>> OBJS-$(CONFIG_AGATE_FILTER) += af_agate.o
>>>> OBJS-$(CONFIG_AINTERLEAVE_FILTER) += f_interleave.o
>>>> diff --git a/libavfilter/af_afir.c b/libavfilter/af_afir.c
>>>> new file mode 100644
>>>> index 0000000..bc1b6a4
>>>> --- /dev/null
>>>> +++ b/libavfilter/af_afir.c
>>>> @@ -0,0 +1,544 @@
>>>> +/*
>>>> + * Copyright (c) 2017 Paul B Mahol
>>>> + *
>>>> + * This file is part of FFmpeg.
>>>> + *
>>>> + * FFmpeg is free software; you can redistribute it and/or
>>>> + * modify it under the terms of the GNU Lesser General Public
>>>> + * License as published by the Free Software Foundation; either
>>>> + * version 2.1 of the License, or (at your option) any later version.
>>>> + *
>>>> + * FFmpeg is distributed in the hope that it will be useful,
>>>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>>>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
>>>> + * Lesser General Public License for more details.
>>>> + *
>>>> + * You should have received a copy of the GNU Lesser General Public
>>>> + * License along with FFmpeg; if not, write to the Free Software
>>>> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
>>>> 02110-1301 USA
>>>> + */
>>>> +
>>>> +/**
>>>> + * @file
>>>> + * An arbitrary audio FIR filter
>>>> + */
>>>> +
>>>> +#include "libavutil/audio_fifo.h"
>>>> +#include "libavutil/common.h"
>>>> +#include "libavutil/opt.h"
>>>> +#include "libavcodec/avfft.h"
>>>> +
>>>> +#include "audio.h"
>>>> +#include "avfilter.h"
>>>> +#include "formats.h"
>>>> +#include "internal.h"
>>>> +
>>>> +#define MAX_IR_DURATION 30
>>>> +
>>>> +typedef struct AudioFIRContext {
>>>> + const AVClass *class;
>>>> +
>>>> + float wet_gain;
>>>> + float dry_gain;
>>>> + float length;
>>>> +
>>>> + float gain;
>>>> +
>>>> + int eof_coeffs;
>>>> + int have_coeffs;
>>>> + int nb_coeffs;
>>>> + int nb_taps;
>>>> + int part_size;
>>>> + int part_index;
>>>> + int block_length;
>>>> + int nb_partitions;
>>>> + int nb_channels;
>>>> + int ir_length;
>>>> + int fft_length;
>>>> + int nb_coef_channels;
>>>> + int one2many;
>>>> + int nb_samples;
>>>> + int want_skip;
>>>> + int need_padding;
>>>> +
>>>> + RDFTContext **rdft, **irdft;
>>>> + float **sum;
>>>> + float **block;
>>>> + FFTComplex **coeff;
>>>> +
>>>> + AVAudioFifo *fifo[2];
>>>> + AVFrame *in[2];
>>>> + AVFrame *buffer;
>>>> + int64_t pts;
>>>> + int index;
>>>> +} AudioFIRContext;
>>>> +
>>>> +static int fir_channel(AVFilterContext *ctx, void *arg, int ch, int
>>>> nb_jobs)
>>>> +{
>>>> + AudioFIRContext *s = ctx->priv;
>>>> + const FFTComplex *coeff = s->coeff[ch * !s->one2many];
>>>> + const float *src = (const float *)s->in[0]->extended_data[ch];
>>>> + int index1 = (s->index + 1) % 3;
>>>> + int index2 = (s->index + 2) % 3;
>>>> + float *sum = s->sum[ch];
>>>> + AVFrame *out = arg;
>>>> + float *block;
>>>> + float *dst;
>>>> + int n, i, j;
>>>> +
>>>> + memset(sum, 0, sizeof(*sum) * s->fft_length);
>>>> + block = s->block[ch] + s->part_index * s->block_length;
>>>> + memset(block, 0, sizeof(*block) * s->fft_length);
>>>> + for (n = 0; n < s->nb_samples; n++) {
>>>> + block[s->part_size + n] = src[n] * s->dry_gain;
>>>> + }
>>>> +
>>>> + av_rdft_calc(s->rdft[ch], block);
>>>> + block[2 * s->part_size] = block[1];
>>>> + block[1] = 0;
>>>> +
>>>> + j = s->part_index;
>>>> +
>>>> + for (i = 0; i < s->nb_partitions; i++) {
>>>> + const int coffset = i * (s->part_size + 1);
>>>> +
>>>> + block = s->block[ch] + j * s->block_length;
>>>> + for (n = 0; n < s->part_size; n++) {
>>>> + const float cre = coeff[coffset + n].re;
>>>> + const float cim = coeff[coffset + n].im;
>>>> + const float tre = block[2 * n ];
>>>> + const float tim = block[2 * n + 1];
>>>> +
>>>> + sum[2 * n ] += tre * cre - tim * cim;
>>>> + sum[2 * n + 1] += tre * cim + tim * cre;
>>>> + }
>>>> + sum[2 * n] += block[2 * n] * coeff[coffset + n].re;
>>>> +
>>>> + if (j == 0)
>>>> + j = s->nb_partitions;
>>>> + j--;
>>>> + }
>>>> +
>>>> + sum[1] = sum[2 * n];
>>>> + av_rdft_calc(s->irdft[ch], sum);
>>>> +
>>>> + dst = (float *)s->buffer->extended_data[ch] + index1 *
>>>> s->part_size;
>>>> + for (n = 0; n < s->part_size; n++) {
>>>> + dst[n] += sum[n];
>>>> + }
>>>> +
>>>> + dst = (float *)s->buffer->extended_data[ch] + index2 *
>>>> s->part_size;
>>>> +
>>>> + memcpy(dst, sum + s->part_size, s->part_size * sizeof(*dst));
>>>> +
>>>> + dst = (float *)s->buffer->extended_data[ch] + s->index *
>>>> s->part_size;
>>>> +
>>>> + if (out) {
>>>> + float *ptr = (float *)out->extended_data[ch];
>>>> + for (n = 0; n < out->nb_samples; n++) {
>>>> + ptr[n] = dst[n] * s->gain * s->wet_gain;
>>>> + }
>>>> + }
>>>> +
>>>> + return 0;
>>>> +}
>>>> +
>>>> +static int fir_frame(AudioFIRContext *s, AVFilterLink *outlink)
>>>> +{
>>>> + AVFilterContext *ctx = outlink->src;
>>>> + AVFrame *out = NULL;
>>>> + int ret;
>>>> +
>>>> + s->nb_samples = FFMIN(s->part_size,
>>>> av_audio_fifo_size(s->fifo[0]));
>>>> +
>>>> + if (!s->want_skip) {
>>>> + out = ff_get_audio_buffer(outlink, s->nb_samples);
>>>> + if (!out)
>>>> + return AVERROR(ENOMEM);
>>>> + }
>>>> +
>>>> + s->in[0] = ff_get_audio_buffer(ctx->inputs[0], s->nb_samples);
>>>> + if (!s->in[0]) {
>>>> + av_frame_free(&out);
>>>> + return AVERROR(ENOMEM);
>>>> + }
>>>> +
>>>> + av_audio_fifo_peek(s->fifo[0], (void **)s->in[0]->extended_data,
>>>> s->nb_samples);
>>>> +
>>>> + ctx->internal->execute(ctx, fir_channel, out, NULL,
>>>> outlink->channels);
>>>> +
>>>> + s->part_index = (s->part_index + 1) % s->nb_partitions;
>>>> +
>>>> + av_audio_fifo_drain(s->fifo[0], s->nb_samples);
>>>> +
>>>> + if (!s->want_skip) {
>>>> + out->pts = s->pts;
>>>> + if (s->pts != AV_NOPTS_VALUE)
>>>> + s->pts += av_rescale_q(out->nb_samples, (AVRational){1,
>>>> outlink->sample_rate}, outlink->time_base);
>>>> + }
>>>> +
>>>> + s->index++;
>>>> + if (s->index == 3)
>>>> + s->index = 0;
>>>> +
>>>> + av_frame_free(&s->in[0]);
>>>> +
>>>> + if (s->want_skip == 1) {
>>>> + s->want_skip = 0;
>>>> + ret = 0;
>>>> + } else {
>>>> + ret = ff_filter_frame(outlink, out);
>>>> + }
>>>> +
>>>> + return ret;
>>>> +}
>>>> +
>>>> +static int convert_coeffs(AVFilterContext *ctx)
>>>> +{
>>>> + AudioFIRContext *s = ctx->priv;
>>>> + int i, ch, n, N;
>>>> + float power = 0;
>>>> +
>>>> + s->nb_taps = av_audio_fifo_size(s->fifo[1]);
>>>> +
>>>> + for (n = 4; (1 << n) < s->nb_taps; n++);
>>>> + N = FFMIN(n, 16);
>>>
>>> It is nice to allow user set maximum N e.g. for low latency app, user
>>> can set low N with higher nb_partitions.
>>
>> Could be later added, but for low latency, one uses NUPOLS or first
>> partition is done in time domain.
>> Using small N drastically reduces speed.
>>
>>>
>>>
>>>> + s->ir_length = 1 << n;
>>>> + s->fft_length = (1 << (N + 1)) + 1;
>>>> + s->part_size = 1 << (N - 1);
>>>> + s->block_length = FFALIGN(s->fft_length, 16);
>>>> + s->nb_partitions = (s->nb_taps + s->part_size - 1) / s->part_size;
>>>> + s->nb_coeffs = s->ir_length + s->nb_partitions;
>>>> +
>>>> + for (ch = 0; ch < ctx->inputs[0]->channels; ch++) {
>>>> + s->sum[ch] = av_calloc(s->fft_length, sizeof(**s->sum));
>>>> + if (!s->sum[ch])
>>>> + return AVERROR(ENOMEM);
>>>> + }
>>>> +
>>>> + for (ch = 0; ch < ctx->inputs[1]->channels; ch++) {
>>>> + s->coeff[ch] = av_calloc(s->nb_coeffs, sizeof(**s->coeff));
>>>> + if (!s->coeff[ch])
>>>> + return AVERROR(ENOMEM);
>>>> + }
>>>> +
>>>> + for (ch = 0; ch < ctx->inputs[0]->channels; ch++) {
>>>> + s->block[ch] = av_calloc(s->nb_partitions * s->block_length,
>>>> sizeof(**s->block));
>>>> + if (!s->block[ch])
>>>> + return AVERROR(ENOMEM);
>>>> + }
>>>> +
>>>> + for (ch = 0; ch < ctx->inputs[0]->channels; ch++) {
>>>> + s->rdft[ch] = av_rdft_init(N, DFT_R2C);
>>>> + s->irdft[ch] = av_rdft_init(N, IDFT_C2R);
>>>> + if (!s->rdft[ch] || !s->irdft[ch])
>>>> + return AVERROR(ENOMEM);
>>>> + }
>>>> +
>>>> + s->in[1] = ff_get_audio_buffer(ctx->inputs[1], s->nb_taps);
>>>> + if (!s->in[1])
>>>> + return AVERROR(ENOMEM);
>>>> +
>>>> + s->buffer = ff_get_audio_buffer(ctx->inputs[0], s->part_size * 3);
>>>> + if (!s->buffer)
>>>> + return AVERROR(ENOMEM);
>>>> +
>>>> + av_audio_fifo_read(s->fifo[1], (void **)s->in[1]->extended_data,
>>>> s->nb_taps);
>>>> +
>>>> + for (ch = 0; ch < ctx->inputs[1]->channels; ch++) {
>>>> + float *time = (float *)s->in[1]->extended_data[!s->one2many *
>>>> ch];
>>>> + float *block = s->block[ch];
>>>> + FFTComplex *coeff = s->coeff[ch];
>>>> +
>>>> + for (i = FFMAX(1, s->length * s->nb_taps); i < s->nb_taps; i++)
>>>> + time[i] = 0;
>>>> +
>>>> + for (i = 0; i < s->nb_partitions; i++) {
>>>> + const float scale = 1.f / s->part_size;
>>>> + const int toffset = i * s->part_size;
>>>> + const int coffset = i * (s->part_size + 1);
>>>> + const int boffset = s->part_size;
>>>> + const int remaining = s->nb_taps - (i * s->part_size);
>>>> + const int size = remaining >= s->part_size ? s->part_size :
>>>> remaining;
>>>> +
>>>> + memset(block, 0, sizeof(*block) * s->fft_length);
>>>> + for (n = 0; n < size; n++) {
>>>> + power += time[n + toffset] * time[n + toffset];
>>>> + block[n + boffset] = time[n + toffset];
>>>> + }
>>>> +
>>>> + av_rdft_calc(s->rdft[0], block);
>>>> +
>>>> + coeff[coffset].re = block[0] * scale;
>>>> + coeff[coffset].im = 0;
>>>> + for (n = 1; n < s->part_size; n++) {
>>>> + coeff[coffset + n].re = block[2 * n] * scale;
>>>> + coeff[coffset + n].im = block[2 * n + 1] * scale;
>>>> + }
>>>> + coeff[coffset + s->part_size].re = block[1] * scale;
>>>> + coeff[coffset + s->part_size].im = 0;
>>>> + }
>>>> + }
>>>> +
>>>> + av_frame_free(&s->in[1]);
>>>> + s->gain = 1.f / sqrtf(power);
>>>
>>> I think s->gain is not required at all. The coeffs are already scaled by
>>> scale.
>>
>> Its needed. Various IRs gives different peak values.
>> The calculation is not perfect but it helps.
>
> OK. So, make it optional again (e.g using auto option).
I don't see need for it, without it its always worse.
I updated patch with added SIMD for trivial complex multiplication.
It is faster (not much) then what gcc generates.
More information about the ffmpeg-devel
mailing list