avfilter/avf_showcwt: optimize multiplication with kernel

Paul B Mahol committed 3 years ago

Commit 2f2f76d1 parent 587189c1 file · +61 −12Browse files

Unified · Split

@@ -25,6 +25,7 @@ · 24 unmodified lines
25	25	#include "libavutil/avassert.h"
26	26	#include "libavutil/avstring.h"
27	27	#include "libavutil/channel_layout.h"
	28	+#include "libavutil/float_dsp.h"
28	29	#include "libavutil/cpu.h"
29	30	#include "libavutil/opt.h"
30	31	#include "libavutil/parseutils.h"
@@ -89,6 +90,9 @@ typedef struct ShowCWTContext { · 58 unmodified lines
89	90	AVFrame *outpicref;
90	91	AVFrame *fft_in;
91	92	AVFrame *fft_out;
	93	+ AVFrame *dst_x;
	94	+ AVFrame *src_x;
	95	+ AVFrame *kernel_x;
92	96	AVFrame *ifft_in;
93	97	AVFrame *ifft_out;
94	98	AVFrame *ch_out;
@@ -114,6 +118,8 @@ typedef struct ShowCWTContext { · 19 unmodified lines
114	118	float minimum_frequency;
115	119	float maximum_frequency;
116	120	float deviation;
	121	+
	122	+ AVFloatDSPContext *fdsp;
117	123	} ShowCWTContext;
118	124
119	125	#define OFFSET(x) offsetof(ShowCWTContext, x)
@@ -172,6 +178,9 @@ static av_cold void uninit(AVFilterContext *ctx) · 52 unmodified lines
172	178	av_frame_free(&s->outpicref);
173	179	av_frame_free(&s->fft_in);
174	180	av_frame_free(&s->fft_out);
	181	+ av_frame_free(&s->dst_x);
	182	+ av_frame_free(&s->src_x);
	183	+ av_frame_free(&s->kernel_x);
175	184	av_frame_free(&s->ifft_in);
176	185	av_frame_free(&s->ifft_out);
177	186	av_frame_free(&s->ch_out);
@@ -187,6 +196,8 @@ static av_cold void uninit(AVFilterContext *ctx) · 9 unmodified lines
187	196	av_tx_uninit(&s->ifft[n]);
188	197	av_freep(&s->ifft);
189	198	}
	199	+
	200	+ av_freep(&s->fdsp);
190	201	}
191	202
192	203	static int query_formats(AVFilterContext *ctx)
@@ -483,18 +494,27 @@ static int run_channel_cwt(AVFilterContext ctx, void arg, int jobnr, int nb_jo · 290 unmodified lines
483	494	AVComplexFloat isrc = (AVComplexFloat )s->ifft_in->extended_data[y];
484	495	AVComplexFloat idst = (AVComplexFloat )s->ifft_out->extended_data[y];
485	496	AVComplexFloat chout = ((AVComplexFloat )s->ch_out->extended_data[ch]) + y * ihop_size;
486		− const float kernel = (const float )s->kernel->extended_data[y];
	497	+ AVComplexFloat dstx = (AVComplexFloat )s->dst_x->extended_data[jobnr];
	498	+ AVComplexFloat srcx = (AVComplexFloat )s->src_x->extended_data[jobnr];
	499	+ AVComplexFloat kernelx = (AVComplexFloat )s->kernel_x->extended_data[jobnr];
	500	+ const AVComplexFloat kernel = (const AVComplexFloat )s->kernel->extended_data[y];
487	501	const unsigned index = (const unsigned )s->index;
488	502	const int kernel_start = s->kernel_start[y];
489	503	const int kernel_stop = s->kernel_stop[y];
	504	+ const int kernel_range = kernel_stop - kernel_start;
	505	+
	506	+ memcpy(kernelx, kernel + kernel_start, sizeof(kernel) kernel_range);
	507	+ memcpy(srcx, dst + kernel_start, sizeof(dst) kernel_range);
	508	+
	509	+ s->fdsp->vector_fmul((float )dstx, (const float )srcx,
	510	+ (const float )kernelx, FFALIGN(kernel_range 2, 16));
490	511
491	512	memset(isrc, 0, sizeof(isrc) output_sample_count);
492		− for (int i = kernel_start; i < kernel_stop; i++) {
493		− const unsigned n = index[i];
494		− const float ff = kernel[i];
	513	+ for (int i = 0; i < kernel_range; i++) {
	514	+ const unsigned n = index[i + kernel_start];
495	515
496		− isrc[n].re += ff * dst[i].re;
497		− isrc[n].im += ff * dst[i].im;
	516	+ isrc[n].re += dstx[i].re;
	517	+ isrc[n].im += dstx[i].im;
498	518	}
499	519
500	520	s->itx_fn(s->ifft[jobnr], idst, isrc, sizeof(*isrc));
@@ -515,7 +535,7 @@ static void compute_kernel(AVFilterContext *ctx) · 14 unmodified lines
515	535	unsigned *index = s->index;
516	536
517	537	for (int y = 0; y < fsize; y++) {
518		− float kernel = (float )s->kernel->extended_data[y];
	538	+ AVComplexFloat kernel = (AVComplexFloat )s->kernel->extended_data[y];
519	539	int *kernel_start = s->kernel_start;
520	540	int *kernel_stop = s->kernel_stop;
521	541	float frequency = s->frequency_band[y*2];
@@ -527,18 +547,19 @@ static void compute_kernel(AVFilterContext *ctx) · 5 unmodified lines
527	547
528	548	f = size - fabsf(f - size);
529	549	ff = expf(-ffdeviation) * scale_factor;
530		− kernel[n] = ff;
	550	+ kernel[n].re = ff;
	551	+ kernel[n].im = ff;
531	552	}
532	553
533	554	for (int n = 0; n < size; n++) {
534		− if (kernel[n] != 0.f) {
	555	+ if (kernel[n].re != 0.f) {
535	556	kernel_start[y] = n;
536	557	break;
537	558	}
538	559	}
539	560
540	561	for (int n = 0; n < size; n++) {
541		− if (kernel[size - n - 1] != 0.f) {
	562	+ if (kernel[size - n - 1].re != 0.f) {
542	563	kernel_stop[y] = size - n;
543	564	break;
544	565	}
@@ -561,6 +582,10 @@ static int config_output(AVFilterLink *outlink) · 16 unmodified lines
561	582
562	583	uninit(ctx);
563	584
	585	+ s->fdsp = avpriv_float_dsp_alloc(0);
	586	+ if (!s->fdsp)
	587	+ return AVERROR(ENOMEM);
	588	+
564	589	switch (s->direction) {
565	590	case DIRECTION_LR:
566	591	case DIRECTION_RL:
@@ -622,6 +647,9 @@ static int config_output(AVFilterLink *outlink) · 55 unmodified lines
622	647	s->outpicref = ff_get_video_buffer(outlink, outlink->w, outlink->h);
623	648	s->fft_in = ff_get_audio_buffer(inlink, s->fft_in_size * 2);
624	649	s->fft_out = ff_get_audio_buffer(inlink, s->fft_out_size * 2);
	650	+ s->dst_x = av_frame_alloc();
	651	+ s->src_x = av_frame_alloc();
	652	+ s->kernel_x = av_frame_alloc();
625	653	s->cache = ff_get_audio_buffer(inlink, s->fft_in_size * 2);
626	654	s->ch_out = ff_get_audio_buffer(inlink, s->frequency_band_count * 2 * s->ihop_size);
627	655	s->ifft_in = av_frame_alloc();
@@ -630,7 +658,7 @@ static int config_output(AVFilterLink *outlink) · 2 unmodified lines
630	658	s->index = av_calloc(s->input_padding_size, sizeof(*s->index));
631	659	s->kernel_start = av_calloc(s->frequency_band_count, sizeof(*s->kernel_start));
632	660	s->kernel_stop = av_calloc(s->frequency_band_count, sizeof(*s->kernel_stop));
633		− if (!s->outpicref \|\| !s->fft_in \|\| !s->fft_out \|\|
	661	+ if (!s->outpicref \|\| !s->fft_in \|\| !s->fft_out \|\| !s->src_x \|\| !s->dst_x \|\| !s->kernel_x \|\|
634	662	!s->ifft_in \|\| !s->ifft_out \|\| !s->kernel_start \|\| !s->kernel_stop \|\|
635	663	!s->frequency_band \|\| !s->kernel \|\| !s->cache \|\| !s->index)
636	664	return AVERROR(ENOMEM);
@@ -650,12 +678,33 @@ static int config_output(AVFilterLink *outlink) · 13 unmodified lines
650	678	return ret;
651	679
652	680	s->kernel->format = inlink->format;
653		− s->kernel->nb_samples = s->input_padding_size;
	681	+ s->kernel->nb_samples = s->input_padding_size * 2;
654	682	s->kernel->ch_layout.nb_channels = s->frequency_band_count;
655	683	ret = av_frame_get_buffer(s->kernel, 0);
656	684	if (ret < 0)
657	685	return ret;
658	686
	687	+ s->src_x->format = inlink->format;
	688	+ s->src_x->nb_samples = s->fft_out_size * 2;
	689	+ s->src_x->ch_layout.nb_channels = s->nb_threads;
	690	+ ret = av_frame_get_buffer(s->src_x, 0);
	691	+ if (ret < 0)
	692	+ return ret;
	693	+
	694	+ s->dst_x->format = inlink->format;
	695	+ s->dst_x->nb_samples = s->fft_out_size * 2;
	696	+ s->dst_x->ch_layout.nb_channels = s->nb_threads;
	697	+ ret = av_frame_get_buffer(s->dst_x, 0);
	698	+ if (ret < 0)
	699	+ return ret;
	700	+
	701	+ s->kernel_x->format = inlink->format;
	702	+ s->kernel_x->nb_samples = s->fft_out_size * 2;
	703	+ s->kernel_x->ch_layout.nb_channels = s->nb_threads;
	704	+ ret = av_frame_get_buffer(s->kernel_x, 0);
	705	+ if (ret < 0)
	706	+ return ret;
	707	+
659	708	s->outpicref->sample_aspect_ratio = (AVRational){1,1};
660	709
661	710	for (int y = 0; y < outlink->h; y++) {

avfilter/avf_showcwt: optimize multiplication with kernel

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code