[FFmpeg-cvslog] avfilter/vf_fftfilt: switch to rdft from lavu/tx
Paul B Mahol
git at videolan.org
Mon Feb 14 18:56:06 EET 2022
ffmpeg | branch: master | Paul B Mahol <onemda at gmail.com> | Mon Feb 14 15:21:49 2022 +0100| [8e825ec3ab09d877f12dcf05d76902a8bb9c8b11] | committer: Paul B Mahol
avfilter/vf_fftfilt: switch to rdft from lavu/tx
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=8e825ec3ab09d877f12dcf05d76902a8bb9c8b11
---
configure | 3 -
libavfilter/vf_fftfilt.c | 159 ++++++++++++++++++++++++++++++-----------------
2 files changed, 103 insertions(+), 59 deletions(-)
diff --git a/configure b/configure
index f5f89c79f2..9222de8cd1 100755
--- a/configure
+++ b/configure
@@ -3645,8 +3645,6 @@ drawtext_filter_suggest="libfontconfig libfribidi"
elbg_filter_deps="avcodec"
eq_filter_deps="gpl"
erosion_opencl_filter_deps="opencl"
-fftfilt_filter_deps="avcodec"
-fftfilt_filter_select="rdft"
find_rect_filter_deps="avcodec avformat gpl"
firequalizer_filter_deps="avcodec"
firequalizer_filter_select="rdft"
@@ -7385,7 +7383,6 @@ enabled bm3d_filter && prepend avfilter_deps "avcodec"
enabled cover_rect_filter && prepend avfilter_deps "avformat avcodec"
enabled ebur128_filter && enabled swresample && prepend avfilter_deps "swresample"
enabled elbg_filter && prepend avfilter_deps "avcodec"
-enabled fftfilt_filter && prepend avfilter_deps "avcodec"
enabled find_rect_filter && prepend avfilter_deps "avformat avcodec"
enabled firequalizer_filter && prepend avfilter_deps "avcodec"
enabled mcdeint_filter && prepend avfilter_deps "avcodec"
diff --git a/libavfilter/vf_fftfilt.c b/libavfilter/vf_fftfilt.c
index 37aa45ef8d..b365ed2806 100644
--- a/libavfilter/vf_fftfilt.c
+++ b/libavfilter/vf_fftfilt.c
@@ -29,7 +29,7 @@
#include "libavutil/imgutils.h"
#include "libavutil/opt.h"
#include "libavutil/pixdesc.h"
-#include "libavcodec/avfft.h"
+#include "libavutil/tx.h"
#include "libavutil/eval.h"
#define MAX_THREADS 32
@@ -51,16 +51,24 @@ typedef struct FFTFILTContext {
int planewidth[MAX_PLANES];
int planeheight[MAX_PLANES];
- RDFTContext *hrdft[MAX_THREADS][MAX_PLANES];
- RDFTContext *vrdft[MAX_THREADS][MAX_PLANES];
- RDFTContext *ihrdft[MAX_THREADS][MAX_PLANES];
- RDFTContext *ivrdft[MAX_THREADS][MAX_PLANES];
+ AVTXContext *hrdft[MAX_THREADS][MAX_PLANES];
+ AVTXContext *vrdft[MAX_THREADS][MAX_PLANES];
+ AVTXContext *ihrdft[MAX_THREADS][MAX_PLANES];
+ AVTXContext *ivrdft[MAX_THREADS][MAX_PLANES];
+
+ av_tx_fn htx_fn, ihtx_fn;
+ av_tx_fn vtx_fn, ivtx_fn;
+
int rdft_hbits[MAX_PLANES];
int rdft_vbits[MAX_PLANES];
+ size_t rdft_hstride[MAX_PLANES];
+ size_t rdft_vstride[MAX_PLANES];
size_t rdft_hlen[MAX_PLANES];
size_t rdft_vlen[MAX_PLANES];
- FFTSample *rdft_hdata[MAX_PLANES];
- FFTSample *rdft_vdata[MAX_PLANES];
+ float *rdft_hdata_in[MAX_PLANES];
+ float *rdft_vdata_in[MAX_PLANES];
+ float *rdft_hdata_out[MAX_PLANES];
+ float *rdft_vdata_out[MAX_PLANES];
int dc[MAX_PLANES];
char *weight_str[MAX_PLANES];
@@ -97,14 +105,14 @@ AVFILTER_DEFINE_CLASS(fftfilt);
static inline double lum(void *priv, double x, double y, int plane)
{
FFTFILTContext *s = priv;
- return s->rdft_vdata[plane][(int)x * s->rdft_vlen[plane] + (int)y];
+ return s->rdft_vdata_out[plane][(int)x * s->rdft_vstride[plane] + (int)y];
}
static double weight_Y(void *priv, double x, double y) { return lum(priv, x, y, Y); }
static double weight_U(void *priv, double x, double y) { return lum(priv, x, y, U); }
static double weight_V(void *priv, double x, double y) { return lum(priv, x, y, V); }
-static void copy_rev(FFTSample *dest, int w, int w2)
+static void copy_rev(float *dest, int w, int w2)
{
int i;
@@ -128,16 +136,19 @@ static int rdft_horizontal8(AVFilterContext *ctx, void *arg, int jobnr, int nb_j
for (int i = slice_start; i < slice_end; i++) {
const uint8_t *src = in->data[plane] + i * in->linesize[plane];
- float *hdata = s->rdft_hdata[plane] + i * s->rdft_hlen[plane];
+ float *hdata_in = s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane];
for (int j = 0; j < w; j++)
- hdata[j] = src[j];
+ hdata_in[j] = src[j];
- copy_rev(s->rdft_hdata[plane] + i * s->rdft_hlen[plane], w, s->rdft_hlen[plane]);
+ copy_rev(s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane], w, s->rdft_hlen[plane]);
}
for (int i = slice_start; i < slice_end; i++)
- av_rdft_calc(s->hrdft[jobnr][plane], s->rdft_hdata[plane] + i * s->rdft_hlen[plane]);
+ s->htx_fn(s->hrdft[jobnr][plane],
+ s->rdft_hdata_out[plane] + i * s->rdft_hstride[plane],
+ s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane],
+ sizeof(float));
}
return 0;
@@ -156,16 +167,19 @@ static int rdft_horizontal16(AVFilterContext *ctx, void *arg, int jobnr, int nb_
for (int i = slice_start; i < slice_end; i++) {
const uint16_t *src = (const uint16_t *)(in->data[plane] + i * in->linesize[plane]);
- float *hdata = s->rdft_hdata[plane] + i * s->rdft_hlen[plane];
+ float *hdata_in = s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane];
for (int j = 0; j < w; j++)
- hdata[j] = src[j];
+ hdata_in[j] = src[j];
- copy_rev(s->rdft_hdata[plane] + i * s->rdft_hlen[plane], w, s->rdft_hlen[plane]);
+ copy_rev(s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane], w, s->rdft_hlen[plane]);
}
for (int i = slice_start; i < slice_end; i++)
- av_rdft_calc(s->hrdft[jobnr][plane], s->rdft_hdata[plane] + i * s->rdft_hlen[plane]);
+ s->htx_fn(s->hrdft[jobnr][plane],
+ s->rdft_hdata_out[plane] + i * s->rdft_hstride[plane],
+ s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane],
+ sizeof(float));
}
return 0;
@@ -183,11 +197,14 @@ static int irdft_horizontal8(AVFilterContext *ctx, void *arg, int jobnr, int nb_
const int slice_end = (h * (jobnr+1)) / nb_jobs;
for (int i = slice_start; i < slice_end; i++)
- av_rdft_calc(s->ihrdft[jobnr][plane], s->rdft_hdata[plane] + i * s->rdft_hlen[plane]);
+ s->ihtx_fn(s->ihrdft[jobnr][plane],
+ s->rdft_hdata_out[plane] + i * s->rdft_hstride[plane],
+ s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane],
+ sizeof(float));
for (int i = slice_start; i < slice_end; i++) {
- const float scale = 4.f / (s->rdft_hlen[plane] * s->rdft_vlen[plane]);
- const float *src = s->rdft_hdata[plane] + i * s->rdft_hlen[plane];
+ const float scale = 1.f / (s->rdft_hlen[plane] * s->rdft_vlen[plane]);
+ const float *src = s->rdft_hdata_out[plane] + i * s->rdft_hstride[plane];
uint8_t *dst = out->data[plane] + i * out->linesize[plane];
for (int j = 0; j < w; j++)
@@ -211,11 +228,14 @@ static int irdft_horizontal16(AVFilterContext *ctx, void *arg, int jobnr, int nb
const int slice_end = (h * (jobnr+1)) / nb_jobs;
for (int i = slice_start; i < slice_end; i++)
- av_rdft_calc(s->ihrdft[jobnr][plane], s->rdft_hdata[plane] + i * s->rdft_hlen[plane]);
+ s->ihtx_fn(s->ihrdft[jobnr][plane],
+ s->rdft_hdata_out[plane] + i * s->rdft_hstride[plane],
+ s->rdft_hdata_in[plane] + i * s->rdft_hstride[plane],
+ sizeof(float));
for (int i = slice_start; i < slice_end; i++) {
- const float scale = 4.f / (s->rdft_hlen[plane] * s->rdft_vlen[plane]);
- const float *src = s->rdft_hdata[plane] + i * s->rdft_hlen[plane];
+ const float scale = 1.f / (s->rdft_hlen[plane] * s->rdft_vlen[plane]);
+ const float *src = s->rdft_hdata_out[plane] + i * s->rdft_hstride[plane];
uint16_t *dst = (uint16_t *)(out->data[plane] + i * out->linesize[plane]);
for (int j = 0; j < w; j++)
@@ -303,34 +323,50 @@ static int config_props(AVFilterLink *inlink)
/* RDFT - Array initialization for Horizontal pass*/
s->rdft_hlen[i] = 1 << (32 - ff_clz(w));
+ s->rdft_hstride[i] = FFALIGN(s->rdft_hlen[i] + 2, av_cpu_max_align());
s->rdft_hbits[i] = av_log2(s->rdft_hlen[i]);
- if (!(s->rdft_hdata[i] = av_malloc_array(h, s->rdft_hlen[i] * sizeof(FFTSample))))
+ if (!(s->rdft_hdata_in[i] = av_calloc(h, s->rdft_hstride[i] * sizeof(float))))
+ return AVERROR(ENOMEM);
+
+ if (!(s->rdft_hdata_out[i] = av_calloc(h, s->rdft_hstride[i] * sizeof(float))))
return AVERROR(ENOMEM);
for (int j = 0; j < s->nb_threads; j++) {
- if (!(s->hrdft[j][i] = av_rdft_init(s->rdft_hbits[i], DFT_R2C)))
+ float scale = 1.f, iscale = 1.f;
+
+ av_tx_init(&s->hrdft[j][i], &s->htx_fn, AV_TX_FLOAT_RDFT, 0, 1 << s->rdft_hbits[i], &scale, 0);
+ if (!s->hrdft[j][i])
return AVERROR(ENOMEM);
- if (!(s->ihrdft[j][i] = av_rdft_init(s->rdft_hbits[i], IDFT_C2R)))
+ av_tx_init(&s->ihrdft[j][i], &s->ihtx_fn, AV_TX_FLOAT_RDFT, 1, 1 << s->rdft_hbits[i], &iscale, 0);
+ if (!s->ihrdft[j][i])
return AVERROR(ENOMEM);
}
/* RDFT - Array initialization for Vertical pass*/
s->rdft_vlen[i] = 1 << (32 - ff_clz(h));
+ s->rdft_vstride[i] = FFALIGN(s->rdft_vlen[i] + 2, av_cpu_max_align());
s->rdft_vbits[i] = av_log2(s->rdft_vlen[i]);
- if (!(s->rdft_vdata[i] = av_malloc_array(s->rdft_hlen[i], s->rdft_vlen[i] * sizeof(FFTSample))))
+ if (!(s->rdft_vdata_in[i] = av_calloc(s->rdft_hstride[i], s->rdft_vstride[i] * sizeof(float))))
+ return AVERROR(ENOMEM);
+
+ if (!(s->rdft_vdata_out[i] = av_calloc(s->rdft_hstride[i], s->rdft_vstride[i] * sizeof(float))))
return AVERROR(ENOMEM);
for (int j = 0; j < s->nb_threads; j++) {
- if (!(s->vrdft[j][i] = av_rdft_init(s->rdft_vbits[i], DFT_R2C)))
+ float scale = 1.f, iscale = 1.f;
+
+ av_tx_init(&s->vrdft[j][i], &s->vtx_fn, AV_TX_FLOAT_RDFT, 0, 1 << s->rdft_vbits[i], &scale, 0);
+ if (!s->vrdft[j][i])
return AVERROR(ENOMEM);
- if (!(s->ivrdft[j][i] = av_rdft_init(s->rdft_vbits[i], IDFT_C2R)))
+ av_tx_init(&s->ivrdft[j][i], &s->ivtx_fn, AV_TX_FLOAT_RDFT, 1, 1 << s->rdft_vbits[i], &iscale, 0);
+ if (!s->ivrdft[j][i])
return AVERROR(ENOMEM);
}
}
/*Luminance value - Array initialization*/
for (plane = 0; plane < 3; plane++) {
- if(!(s->weight[plane] = av_malloc_array(s->rdft_hlen[plane], s->rdft_vlen[plane] * sizeof(double))))
+ if(!(s->weight[plane] = av_calloc(s->rdft_hlen[plane], s->rdft_vlen[plane] * sizeof(double))))
return AVERROR(ENOMEM);
if (s->eval_mode == EVAL_MODE_INIT)
@@ -360,7 +396,7 @@ static int multiply_data(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs
/*Change user defined parameters*/
for (int i = slice_start; i < slice_end; i++) {
const double *weight = s->weight[plane] + i * s->rdft_vlen[plane];
- float *vdata = s->rdft_vdata[plane] + i * s->rdft_vlen[plane];
+ float *vdata = s->rdft_vdata_out[plane] + i * s->rdft_vstride[plane];
for (int j = 0; j < s->rdft_vlen[plane]; j++)
vdata[j] *= weight[j];
@@ -377,16 +413,18 @@ static int copy_vertical(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs
for (int plane = 0; plane < s->nb_planes; plane++) {
const int hlen = s->rdft_hlen[plane];
const int vlen = s->rdft_vlen[plane];
+ const int hstride = s->rdft_hstride[plane];
+ const int vstride = s->rdft_vstride[plane];
const int slice_start = (hlen * jobnr) / nb_jobs;
const int slice_end = (hlen * (jobnr+1)) / nb_jobs;
const int h = s->planeheight[plane];
- FFTSample *hdata = s->rdft_hdata[plane];
- FFTSample *vdata = s->rdft_vdata[plane];
+ float *hdata = s->rdft_hdata_out[plane];
+ float *vdata = s->rdft_vdata_in[plane];
for (int i = slice_start; i < slice_end; i++) {
for (int j = 0; j < h; j++)
- vdata[i * vlen + j] = hdata[j * hlen + i];
- copy_rev(vdata + i * vlen, h, vlen);
+ vdata[i * vstride + j] = hdata[j * hstride + i];
+ copy_rev(vdata + i * vstride, h, vlen);
}
}
@@ -403,7 +441,10 @@ static int rdft_vertical(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs
const int slice_end = (height * (jobnr+1)) / nb_jobs;
for (int i = slice_start; i < slice_end; i++)
- av_rdft_calc(s->vrdft[jobnr][plane], s->rdft_vdata[plane] + i * s->rdft_vlen[plane]);
+ s->vtx_fn(s->vrdft[jobnr][plane],
+ s->rdft_vdata_out[plane] + i * s->rdft_vstride[plane],
+ s->rdft_vdata_in[plane] + i * s->rdft_vstride[plane],
+ sizeof(float));
}
return 0;
@@ -419,7 +460,10 @@ static int irdft_vertical(AVFilterContext *ctx, void *arg, int jobnr, int nb_job
const int slice_end = (height * (jobnr+1)) / nb_jobs;
for (int i = slice_start; i < slice_end; i++)
- av_rdft_calc(s->ivrdft[jobnr][plane], s->rdft_vdata[plane] + i * s->rdft_vlen[plane]);
+ s->ivtx_fn(s->ivrdft[jobnr][plane],
+ s->rdft_vdata_in[plane] + i * s->rdft_vstride[plane],
+ s->rdft_vdata_out[plane] + i * s->rdft_vstride[plane],
+ sizeof(float));
}
return 0;
@@ -431,16 +475,17 @@ static int copy_horizontal(AVFilterContext *ctx, void *arg, int jobnr, int nb_jo
for (int plane = 0; plane < s->nb_planes; plane++) {
const int hlen = s->rdft_hlen[plane];
- const int vlen = s->rdft_vlen[plane];
+ const int hstride = s->rdft_hstride[plane];
+ const int vstride = s->rdft_vstride[plane];
const int slice_start = (hlen * jobnr) / nb_jobs;
const int slice_end = (hlen * (jobnr+1)) / nb_jobs;
const int h = s->planeheight[plane];
- FFTSample *hdata = s->rdft_hdata[plane];
- FFTSample *vdata = s->rdft_vdata[plane];
+ float *hdata = s->rdft_hdata_in[plane];
+ float *vdata = s->rdft_vdata_in[plane];
for (int i = slice_start; i < slice_end; i++)
for (int j = 0; j < h; j++)
- hdata[j * hlen + i] = vdata[i * vlen + j];
+ hdata[j * hstride + i] = vdata[i * vstride + j];
}
return 0;
@@ -461,11 +506,6 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
av_frame_copy_props(out, in);
- for (int plane = 0; plane < s->nb_planes; plane++) {
- if (s->eval_mode == EVAL_MODE_FRAME)
- do_eval(s, inlink, plane);
- }
-
ff_filter_execute(ctx, s->rdft_horizontal, in, NULL,
FFMIN(s->planeheight[1], s->nb_threads));
@@ -475,11 +515,16 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
ff_filter_execute(ctx, rdft_vertical, NULL, NULL,
FFMIN(s->planeheight[1], s->nb_threads));
+ for (int plane = 0; plane < s->nb_planes; plane++) {
+ if (s->eval_mode == EVAL_MODE_FRAME)
+ do_eval(s, inlink, plane);
+ }
+
ff_filter_execute(ctx, multiply_data, NULL, NULL,
FFMIN(s->planeheight[1], s->nb_threads));
for (int plane = 0; plane < s->nb_planes; plane++)
- s->rdft_vdata[plane][0] += s->rdft_hlen[plane] * s->rdft_vlen[plane] * s->dc[plane];
+ s->rdft_vdata_out[plane][0] += s->rdft_hlen[plane] * s->rdft_vlen[plane] * s->dc[plane] * (1 << (s->depth - 8));
ff_filter_execute(ctx, irdft_vertical, NULL, NULL,
FFMIN(s->planeheight[1], s->nb_threads));
@@ -497,17 +542,19 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *in)
static av_cold void uninit(AVFilterContext *ctx)
{
FFTFILTContext *s = ctx->priv;
- int i;
- for (i = 0; i < MAX_PLANES; i++) {
- av_free(s->rdft_hdata[i]);
- av_free(s->rdft_vdata[i]);
+
+ for (int i = 0; i < MAX_PLANES; i++) {
+ av_freep(&s->rdft_hdata_in[i]);
+ av_freep(&s->rdft_vdata_in[i]);
+ av_freep(&s->rdft_hdata_out[i]);
+ av_freep(&s->rdft_vdata_out[i]);
av_expr_free(s->weight_expr[i]);
- av_free(s->weight[i]);
+ av_freep(&s->weight[i]);
for (int j = 0; j < s->nb_threads; j++) {
- av_rdft_end(s->hrdft[j][i]);
- av_rdft_end(s->ihrdft[j][i]);
- av_rdft_end(s->vrdft[j][i]);
- av_rdft_end(s->ivrdft[j][i]);
+ av_tx_uninit(&s->hrdft[j][i]);
+ av_tx_uninit(&s->ihrdft[j][i]);
+ av_tx_uninit(&s->vrdft[j][i]);
+ av_tx_uninit(&s->ivrdft[j][i]);
}
}
}
More information about the ffmpeg-cvslog
mailing list