virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-08-10 06:10:52 +02:00
Code Issues Releases Activity

avfilter/af_sofalizer: speed up fast convolution

Browse Source 
Do inverse FFT only once per output channel.
This commit is contained in:
Paul B Mahol
2018-12-25 18:46:07 +01:00
parent 7efe84aebd
commit 8d0b8c50bd
1 changed files with 28 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
libavfilter/af_sofalizer.c
View File

@@ -88,7 +88,8 @@ typedef struct SOFAlizerContext {
float *data_ir[2]; /* IRs for all channels to be convolved */ float *data_ir[2]; /* IRs for all channels to be convolved */
/* (this excludes the LFE) */ /* (this excludes the LFE) */
float *temp_src[2]; float *temp_src[2];
FFTComplex *temp_fft[2]; FFTComplex *temp_fft[2]; /* Array to hold FFT values */
FFTComplex *temp_afft[2]; /* Array to accumulate FFT values prior to IFFT */
/* control variables */ /* control variables */
float gain; /* filter gain (in dB) */ float gain; /* filter gain (in dB) */
@@ -322,6 +323,7 @@ typedef struct ThreadData {
float **ringbuffer; float **ringbuffer;
float **temp_src; float **temp_src;
FFTComplex **temp_fft; FFTComplex **temp_fft;
FFTComplex **temp_afft;
} ThreadData; } ThreadData;
static int sofalizer_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) static int sofalizer_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
@@ -430,6 +432,7 @@ static int sofalizer_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr,
/* -1 for AND instead of MODULO (applied to powers of 2): */ /* -1 for AND instead of MODULO (applied to powers of 2): */
const uint32_t modulo = (uint32_t)buffer_length - 1; const uint32_t modulo = (uint32_t)buffer_length - 1;
FFTComplex *fft_in = s->temp_fft[jobnr]; /* temporary array for FFT input/output data */ FFTComplex *fft_in = s->temp_fft[jobnr]; /* temporary array for FFT input/output data */
FFTComplex *fft_acc = s->temp_afft[jobnr];
FFTContext *ifft = s->ifft[jobnr]; FFTContext *ifft = s->ifft[jobnr];
FFTContext *fft = s->fft[jobnr]; FFTContext *fft = s->fft[jobnr];
const int n_conv = s->n_conv; const int n_conv = s->n_conv;
@@ -458,6 +461,9 @@ static int sofalizer_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr,
dst[2 * j] = 0; dst[2 * j] = 0;
} }
/* fill FFT accumulation with 0 */
memset(fft_acc, 0, sizeof(FFTComplex) * n_fft);
for (i = 0; i < n_conv; i++) { for (i = 0; i < n_conv; i++) {
if (i == s->lfe_channel) { /* LFE */ if (i == s->lfe_channel) { /* LFE */
for (j = 0; j < in->nb_samples; j++) { for (j = 0; j < in->nb_samples; j++) {
@@ -490,26 +496,26 @@ static int sofalizer_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr,
/* complex multiplication of input signal and HRTFs */ /* complex multiplication of input signal and HRTFs */
/* output channel (real): */ /* output channel (real): */
fft_in[j].re = re * hcomplex->re - im * hcomplex->im; fft_acc[j].re += re * hcomplex->re - im * hcomplex->im;
/* output channel (imag): */ /* output channel (imag): */
fft_in[j].im = re * hcomplex->im + im * hcomplex->re; fft_acc[j].im += re * hcomplex->im + im * hcomplex->re;
}
} }
/* transform output signal of current channel back to time domain */ /* transform output signal of current channel back to time domain */
av_fft_permute(ifft, fft_in); av_fft_permute(ifft, fft_acc);
av_fft_calc(ifft, fft_in); av_fft_calc(ifft, fft_acc);
for (j = 0; j < in->nb_samples; j++) { for (j = 0; j < in->nb_samples; j++) {
/* write output signal of current channel to output buffer */ /* write output signal of current channel to output buffer */
dst[2 * j] += fft_in[j].re * fft_scale; dst[2 * j] += fft_acc[j].re * fft_scale;
} }
for (j = 0; j < n_samples - 1; j++) { /* overflow length is IR length - 1 */ for (j = 0; j < n_samples - 1; j++) { /* overflow length is IR length - 1 */
/* write the rest of output signal to overflow buffer */ /* write the rest of output signal to overflow buffer */
int write_pos = (wr + j) & modulo; int write_pos = (wr + j) & modulo;
*(ringbuffer + write_pos) += fft_in[in->nb_samples + j].re * fft_scale; *(ringbuffer + write_pos) += fft_acc[in->nb_samples + j].re * fft_scale;
}
} }
/* go through all samples of current output buffer: count clippings */ /* go through all samples of current output buffer: count clippings */
@@ -549,6 +555,7 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
td.delay = s->delay; td.ir = s->data_ir; td.n_clippings = n_clippings; td.delay = s->delay; td.ir = s->data_ir; td.n_clippings = n_clippings;
td.ringbuffer = s->ringbuffer; td.temp_src = s->temp_src; td.ringbuffer = s->ringbuffer; td.temp_src = s->temp_src;
td.temp_fft = s->temp_fft; td.temp_fft = s->temp_fft;
td.temp_afft = s->temp_afft;
if (s->type == TIME_DOMAIN) { if (s->type == TIME_DOMAIN) {
ctx->internal->execute(ctx, sofalizer_convolute, &td, NULL, 2); ctx->internal->execute(ctx, sofalizer_convolute, &td, NULL, 2);
@@ -790,7 +797,10 @@ static int load_data(AVFilterContext *ctx, int azim, int elev, float radius, int
s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float)); s->ringbuffer[1] = av_calloc(s->buffer_length, sizeof(float));
s->temp_fft[0] = av_malloc_array(s->n_fft, sizeof(FFTComplex)); s->temp_fft[0] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
s->temp_fft[1] = av_malloc_array(s->n_fft, sizeof(FFTComplex)); s->temp_fft[1] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
if (!s->temp_fft[0] || !s->temp_fft[1]) { s->temp_afft[0] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
s->temp_afft[1] = av_malloc_array(s->n_fft, sizeof(FFTComplex));
if (!s->temp_fft[0] || !s->temp_fft[1] ||
!s->temp_afft[0] || !s->temp_afft[1]) {
ret = AVERROR(ENOMEM); ret = AVERROR(ENOMEM);
goto fail; goto fail;
} }
@@ -957,6 +967,8 @@ static av_cold void uninit(AVFilterContext *ctx)
av_freep(&s->speaker_elev); av_freep(&s->speaker_elev);
av_freep(&s->temp_src[0]); av_freep(&s->temp_src[0]);
av_freep(&s->temp_src[1]); av_freep(&s->temp_src[1]);
av_freep(&s->temp_afft[0]);
av_freep(&s->temp_afft[1]);
av_freep(&s->temp_fft[0]); av_freep(&s->temp_fft[0]);
av_freep(&s->temp_fft[1]); av_freep(&s->temp_fft[1]);
av_freep(&s->data_hrtf[0]); av_freep(&s->data_hrtf[0]);