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_headphone: speed up fast convolution

Browse Source 
Do IFFT only once per output channel.
This commit is contained in:
Paul B Mahol
2018-12-25 19:17:25 +01:00
parent 8d0b8c50bd
commit 6c7eb0708e
1 changed files with 23 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
libavfilter/af_headphone.c
View File

@@ -73,6 +73,7 @@ typedef struct HeadphoneContext {
float *data_ir[2]; float *data_ir[2];
float *temp_src[2]; float *temp_src[2];
FFTComplex *temp_fft[2]; FFTComplex *temp_fft[2];
FFTComplex *temp_afft[2];
FFTContext *fft[2], *ifft[2]; FFTContext *fft[2], *ifft[2];
FFTComplex *data_hrtf[2]; FFTComplex *data_hrtf[2];
@@ -158,6 +159,7 @@ 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 headphone_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs) static int headphone_convolute(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
@@ -251,6 +253,7 @@ static int headphone_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr,
const int buffer_length = s->buffer_length; const int buffer_length = s->buffer_length;
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]; FFTComplex *fft_in = s->temp_fft[jobnr];
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_fft = s->n_fft; const int n_fft = s->n_fft;
@@ -273,6 +276,8 @@ static int headphone_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr,
dst[2 * j] = 0; dst[2 * j] = 0;
} }
memset(fft_acc, 0, sizeof(FFTComplex) * n_fft);
for (i = 0; i < in_channels; i++) { for (i = 0; i < in_channels; i++) {
if (i == s->lfe_channel) { if (i == s->lfe_channel) {
for (j = 0; j < in->nb_samples; j++) { for (j = 0; j < in->nb_samples; j++) {
@@ -297,22 +302,22 @@ static int headphone_fast_convolute(AVFilterContext *ctx, void *arg, int jobnr,
const float re = fft_in[j].re; const float re = fft_in[j].re;
const float im = fft_in[j].im; const float im = fft_in[j].im;
fft_in[j].re = re * hcomplex->re - im * hcomplex->im; fft_acc[j].re += re * hcomplex->re - im * hcomplex->im;
fft_in[j].im = re * hcomplex->im + im * hcomplex->re; fft_acc[j].im += re * hcomplex->im + im * hcomplex->re;
}
} }
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++) {
dst[2 * j] += fft_in[j].re * fft_scale; dst[2 * j] += fft_acc[j].re * fft_scale;
} }
for (j = 0; j < ir_len - 1; j++) { for (j = 0; j < ir_len - 1; j++) {
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;
}
} }
for (i = 0; i < out->nb_samples; i++) { for (i = 0; i < out->nb_samples; i++) {
@@ -364,6 +369,7 @@ static int headphone_frame(HeadphoneContext *s, AVFrame *in, AVFilterLink *outli
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, headphone_convolute, &td, NULL, 2); ctx->internal->execute(ctx, headphone_convolute, &td, NULL, 2);
@@ -439,7 +445,10 @@ static int convert_coeffs(AVFilterContext *ctx, AVFilterLink *inlink)
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_calloc(s->n_fft, sizeof(FFTComplex)); s->temp_fft[0] = av_calloc(s->n_fft, sizeof(FFTComplex));
s->temp_fft[1] = av_calloc(s->n_fft, sizeof(FFTComplex)); s->temp_fft[1] = av_calloc(s->n_fft, sizeof(FFTComplex));
if (!s->temp_fft[0] || !s->temp_fft[1]) { s->temp_afft[0] = av_calloc(s->n_fft, sizeof(FFTComplex));
s->temp_afft[1] = av_calloc(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;
} }
@@ -819,6 +828,8 @@ static av_cold void uninit(AVFilterContext *ctx)
av_freep(&s->temp_src[1]); av_freep(&s->temp_src[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->temp_afft[0]);
av_freep(&s->temp_afft[1]);
av_freep(&s->data_hrtf[0]); av_freep(&s->data_hrtf[0]);
av_freep(&s->data_hrtf[1]); av_freep(&s->data_hrtf[1]);
av_freep(&s->fdsp); av_freep(&s->fdsp);