virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2025-08-15 14:13:16 +02:00
Code Issues Releases Activity

avcodec/utvideodec: add SIMD support for median prediction for planar formats

Browse Source 
~10% faster overall.

Signed-off-by: Paul B Mahol <onemda@gmail.com>
This commit is contained in:
Paul B Mahol
2016-12-23 15:41:51 +01:00
parent 373f21a911
commit ea93052db3
4 changed files with 140 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
configure vendored
View File

@@ -2531,7 +2531,7 @@ truespeech_decoder_select="bswapdsp"
tscc_decoder_select="zlib" tscc_decoder_select="zlib"
twinvq_decoder_select="mdct lsp sinewin" twinvq_decoder_select="mdct lsp sinewin"
txd_decoder_select="texturedsp" txd_decoder_select="texturedsp"
utvideo_decoder_select="bswapdsp" utvideo_decoder_select="bswapdsp huffyuvdsp"
utvideo_encoder_select="bswapdsp huffman huffyuvencdsp" utvideo_encoder_select="bswapdsp huffman huffyuvencdsp"
vble_decoder_select="huffyuvdsp" vble_decoder_select="huffyuvdsp"
vc1_decoder_select="blockdsp h263_decoder h264qpel intrax8 mpegvideo vc1dsp" vc1_decoder_select="blockdsp h263_decoder h264qpel intrax8 mpegvideo vc1dsp"

2
libavcodec/utvideo.h
View File

@@ -30,6 +30,7 @@
#include "libavutil/common.h" #include "libavutil/common.h"
#include "avcodec.h" #include "avcodec.h"
#include "bswapdsp.h" #include "bswapdsp.h"
#include "huffyuvdsp.h"
#include "huffyuvencdsp.h" #include "huffyuvencdsp.h"
enum { enum {
@@ -69,6 +70,7 @@ typedef struct UtvideoContext {
const AVClass *class; const AVClass *class;
AVCodecContext *avctx; AVCodecContext *avctx;
BswapDSPContext bdsp; BswapDSPContext bdsp;
HuffYUVDSPContext hdspdec;
HuffYUVEncDSPContext hdsp; HuffYUVEncDSPContext hdsp;
uint32_t frame_info_size, flags, frame_info; uint32_t frame_info_size, flags, frame_info;

164
libavcodec/utvideodec.c
View File

@@ -372,8 +372,111 @@ static void restore_rgb_planes10(AVFrame *frame, int width, int height)
} }
} }
static void restore_median(uint8_t *src, int step, int stride, #undef A
int width, int height, int slices, int rmode) #undef B
#undef C
static void restore_median_planar(UtvideoContext *c, uint8_t *src, int stride,
int width, int height, int slices, int rmode)
{
int i, j, slice;
int A, B, C;
uint8_t *bsrc;
int slice_start, slice_height;
const int cmask = ~rmode;
for (slice = 0; slice < slices; slice++) {
slice_start = ((slice * height) / slices) & cmask;
slice_height = ((((slice + 1) * height) / slices) & cmask) -
slice_start;
if (!slice_height)
continue;
bsrc = src + slice_start * stride;
// first line - left neighbour prediction
bsrc[0] += 0x80;
c->hdspdec.add_hfyu_left_pred(bsrc, bsrc, width, 0);
bsrc += stride;
if (slice_height <= 1)
continue;
// second line - first element has top prediction, the rest uses median
C = bsrc[-stride];
bsrc[0] += C;
A = bsrc[0];
for (i = 1; i < width; i++) {
B = bsrc[i - stride];
bsrc[i] += mid_pred(A, B, (uint8_t)(A + B - C));
C = B;
A = bsrc[i];
}
bsrc += stride;
// the rest of lines use continuous median prediction
for (j = 2; j < slice_height; j++) {
c->hdspdec.add_hfyu_median_pred(bsrc, bsrc - stride,
bsrc, width, &B, &C);
bsrc += stride;
}
}
}
/* UtVideo interlaced mode treats every two lines as a single one,
* so restoring function should take care of possible padding between
* two parts of the same "line".
*/
static void restore_median_planar_il(UtvideoContext *c, uint8_t *src, int stride,
int width, int height, int slices, int rmode)
{
int i, j, slice;
int A, B, C;
uint8_t *bsrc;
int slice_start, slice_height;
const int cmask = ~(rmode ? 3 : 1);
const int stride2 = stride << 1;
for (slice = 0; slice < slices; slice++) {
slice_start = ((slice * height) / slices) & cmask;
slice_height = ((((slice + 1) * height) / slices) & cmask) -
slice_start;
slice_height >>= 1;
if (!slice_height)
continue;
bsrc = src + slice_start * stride;
// first line - left neighbour prediction
bsrc[0] += 0x80;
A = c->hdspdec.add_hfyu_left_pred(bsrc, bsrc, width, 0);
c->hdspdec.add_hfyu_left_pred(bsrc + stride, bsrc + stride, width, A);
bsrc += stride2;
if (slice_height <= 1)
continue;
// second line - first element has top prediction, the rest uses median
C = bsrc[-stride2];
bsrc[0] += C;
A = bsrc[0];
for (i = 1; i < width; i++) {
B = bsrc[i - stride2];
bsrc[i] += mid_pred(A, B, (uint8_t)(A + B - C));
C = B;
A = bsrc[i];
}
c->hdspdec.add_hfyu_median_pred(bsrc + stride, bsrc - stride,
bsrc + stride, width, &B, &C);
bsrc += stride2;
// the rest of lines use continuous median prediction
for (j = 2; j < slice_height; j++) {
c->hdspdec.add_hfyu_median_pred(bsrc, bsrc - stride2,
bsrc, width, &B, &C);
c->hdspdec.add_hfyu_median_pred(bsrc + stride, bsrc - stride,
bsrc + stride, width, &B, &C);
bsrc += stride2;
}
}
}
static void restore_median_packed(uint8_t *src, int step, int stride,
int width, int height, int slices, int rmode)
{ {
int i, j, slice; int i, j, slice;
int A, B, C; int A, B, C;
@@ -428,8 +531,8 @@ static void restore_median(uint8_t *src, int step, int stride,
* so restoring function should take care of possible padding between * so restoring function should take care of possible padding between
* two parts of the same "line". * two parts of the same "line".
*/ */
static void restore_median_il(uint8_t *src, int step, int stride, static void restore_median_packed_il(uint8_t *src, int step, int stride,
int width, int height, int slices, int rmode) int width, int height, int slices, int rmode)
{ {
int i, j, slice; int i, j, slice;
int A, B, C; int A, B, C;
@@ -608,14 +711,14 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
return ret; return ret;
if (c->frame_pred == PRED_MEDIAN) { if (c->frame_pred == PRED_MEDIAN) {
if (!c->interlaced) { if (!c->interlaced) {
restore_median(frame.f->data[0] + ff_ut_rgb_order[i], restore_median_packed(frame.f->data[0] + ff_ut_rgb_order[i],
c->planes, frame.f->linesize[0], avctx->width, c->planes, frame.f->linesize[0], avctx->width,
avctx->height, c->slices, 0); avctx->height, c->slices, 0);
} else { } else {
restore_median_il(frame.f->data[0] + ff_ut_rgb_order[i], restore_median_packed_il(frame.f->data[0] + ff_ut_rgb_order[i],
c->planes, frame.f->linesize[0], c->planes, frame.f->linesize[0],
avctx->width, avctx->height, c->slices, avctx->width, avctx->height, c->slices,
0); 0);
} }
} }
} }
@@ -644,14 +747,14 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
return ret; return ret;
if (c->frame_pred == PRED_MEDIAN) { if (c->frame_pred == PRED_MEDIAN) {
if (!c->interlaced) { if (!c->interlaced) {
restore_median(frame.f->data[i], 1, frame.f->linesize[i], restore_median_planar(c, frame.f->data[i], frame.f->linesize[i],
avctx->width >> !!i, avctx->height >> !!i, avctx->width >> !!i, avctx->height >> !!i,
c->slices, !i); c->slices, !i);
} else { } else {
restore_median_il(frame.f->data[i], 1, frame.f->linesize[i], restore_median_planar_il(c, frame.f->data[i], frame.f->linesize[i],
avctx->width >> !!i, avctx->width >> !!i,
avctx->height >> !!i, avctx->height >> !!i,
c->slices, !i); c->slices, !i);
} }
} }
} }
@@ -665,13 +768,13 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
return ret; return ret;
if (c->frame_pred == PRED_MEDIAN) { if (c->frame_pred == PRED_MEDIAN) {
if (!c->interlaced) { if (!c->interlaced) {
restore_median(frame.f->data[i], 1, frame.f->linesize[i], restore_median_planar(c, frame.f->data[i], frame.f->linesize[i],
avctx->width >> !!i, avctx->height, avctx->width >> !!i, avctx->height,
c->slices, 0); c->slices, 0);
} else { } else {
restore_median_il(frame.f->data[i], 1, frame.f->linesize[i], restore_median_planar_il(c, frame.f->data[i], frame.f->linesize[i],
avctx->width >> !!i, avctx->height, avctx->width >> !!i, avctx->height,
c->slices, 0); c->slices, 0);
} }
} }
} }
@@ -685,13 +788,13 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
return ret; return ret;
if (c->frame_pred == PRED_MEDIAN) { if (c->frame_pred == PRED_MEDIAN) {
if (!c->interlaced) { if (!c->interlaced) {
restore_median(frame.f->data[i], 1, frame.f->linesize[i], restore_median_planar(c, frame.f->data[i], frame.f->linesize[i],
avctx->width, avctx->height, avctx->width, avctx->height,
c->slices, 0); c->slices, 0);
} else { } else {
restore_median_il(frame.f->data[i], 1, frame.f->linesize[i], restore_median_planar_il(c, frame.f->data[i], frame.f->linesize[i],
avctx->width, avctx->height, avctx->width, avctx->height,
c->slices, 0); c->slices, 0);
} }
} }
} }
@@ -724,6 +827,7 @@ static av_cold int decode_init(AVCodecContext *avctx)
c->avctx = avctx; c->avctx = avctx;
ff_bswapdsp_init(&c->bdsp); ff_bswapdsp_init(&c->bdsp);
ff_huffyuvdsp_init(&c->hdspdec);
if (avctx->extradata_size >= 16) { if (avctx->extradata_size >= 16) {
av_log(avctx, AV_LOG_DEBUG, "Encoder version %d.%d.%d.%d\n", av_log(avctx, AV_LOG_DEBUG, "Encoder version %d.%d.%d.%d\n",

3
libavcodec/utvideoenc.c
View File

@@ -293,6 +293,9 @@ static void left_predict(uint8_t *src, uint8_t *dst, int stride,
} }
} }
#undef A
#undef B
/* Write data to a plane with median prediction */ /* Write data to a plane with median prediction */
static void median_predict(UtvideoContext *c, uint8_t *src, uint8_t *dst, int stride, static void median_predict(UtvideoContext *c, uint8_t *src, uint8_t *dst, int stride,
int width, int height) int width, int height)