[FFmpeg-cvslog] hevc/sao: optimze sao implementation
Mickaël Raulet
git at videolan.org
Fri Jul 18 23:04:18 CEST 2014
ffmpeg | branch: master | Mickaël Raulet <mraulet at insa-rennes.fr> | Wed Jul 16 23:15:32 2014 +0200| [d249e6828e8a84758010ec020a84dfcd156b585e] | committer: Michael Niedermayer
hevc/sao: optimze sao implementation
- adding one extra pixel all around the frame
- do not copy when SAO is not applied
5% improvement
cherry picked from commit 10fc29fc19a12c4d8168fbe1a954b76386db12d0
Signed-off-by: Michael Niedermayer <michaelni at gmx.at>
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=d249e6828e8a84758010ec020a84dfcd156b585e
---
libavcodec/hevc.c | 27 +++++++++---
libavcodec/hevc.h | 1 +
libavcodec/hevc_filter.c | 66 +++++++++++++++++++++---------
libavcodec/hevcdsp.h | 4 +-
libavcodec/hevcdsp_template.c | 91 +++++++++++++++++++++--------------------
5 files changed, 117 insertions(+), 72 deletions(-)
diff --git a/libavcodec/hevc.c b/libavcodec/hevc.c
index afb2baa..d466aa6 100644
--- a/libavcodec/hevc.c
+++ b/libavcodec/hevc.c
@@ -276,6 +276,24 @@ static int decode_lt_rps(HEVCContext *s, LongTermRPS *rps, GetBitContext *gb)
return 0;
}
+static int get_buffer_sao(HEVCContext *s, AVFrame *frame)
+{
+ int ret, i;
+
+ frame->width = s->avctx->width + 2;
+ frame->height = s->avctx->height + 2;
+ if ((ret = ff_get_buffer(s->avctx, frame, AV_GET_BUFFER_FLAG_REF)) < 0)
+ return ret;
+ for (i = 0; frame->data[i]; i++) {
+ int offset = frame->linesize[i] + 1;
+ frame->data[i] += offset;
+ }
+ frame->width = s->avctx->width;
+ frame->height = s->avctx->height;
+
+ return 0;
+}
+
static int set_sps(HEVCContext *s, const HEVCSPS *sps)
{
int ret;
@@ -317,10 +335,8 @@ static int set_sps(HEVCContext *s, const HEVCSPS *sps)
if (sps->sao_enabled) {
av_frame_unref(s->tmp_frame);
- ret = ff_get_buffer(s->avctx, s->tmp_frame, AV_GET_BUFFER_FLAG_REF);
- if (ret < 0)
- goto fail;
- s->frame = s->tmp_frame;
+ ret = get_buffer_sao(s, s->tmp_frame);
+ s->sao_frame = s->tmp_frame;
}
s->sps = sps;
@@ -2582,8 +2598,7 @@ static int hevc_frame_start(HEVCContext *s)
if (s->pps->tiles_enabled_flag)
lc->end_of_tiles_x = s->pps->column_width[0] << s->sps->log2_ctb_size;
- ret = ff_hevc_set_new_ref(s, s->sps->sao_enabled ? &s->sao_frame : &s->frame,
- s->poc);
+ ret = ff_hevc_set_new_ref(s, &s->frame, s->poc);
if (ret < 0)
goto fail;
diff --git a/libavcodec/hevc.h b/libavcodec/hevc.h
index 2a5ce25..c38de1a 100644
--- a/libavcodec/hevc.h
+++ b/libavcodec/hevc.h
@@ -260,6 +260,7 @@ enum SAOType {
SAO_NOT_APPLIED = 0,
SAO_BAND,
SAO_EDGE,
+ SAO_APPLIED
};
enum SAOEOClass {
diff --git a/libavcodec/hevc_filter.c b/libavcodec/hevc_filter.c
index 0d3e238..b0b197c 100644
--- a/libavcodec/hevc_filter.c
+++ b/libavcodec/hevc_filter.c
@@ -142,14 +142,14 @@ static int get_qPy(HEVCContext *s, int xC, int yC)
}
static void copy_CTB(uint8_t *dst, uint8_t *src,
- int width, int height, int stride)
+ int width, int height, int stride_dst, int stride_src)
{
int i;
for (i = 0; i < height; i++) {
memcpy(dst, src, width);
- dst += stride;
- src += stride;
+ dst += stride_dst;
+ src += stride_src;
}
}
@@ -174,7 +174,7 @@ static void restore_tqb_pixels(HEVCContext *s, int x0, int y0, int width, int he
uint8_t *src = &s->frame->data[c_idx][ ((y << s->sps->log2_min_pu_size) >> vshift) * stride + (((x << s->sps->log2_min_pu_size) >> hshift) << s->sps->pixel_shift)];
uint8_t *dst = &s->sao_frame->data[c_idx][((y << s->sps->log2_min_pu_size) >> vshift) * stride + (((x << s->sps->log2_min_pu_size) >> hshift) << s->sps->pixel_shift)];
for (n = 0; n < (min_pu_size >> vshift); n++) {
- memcpy(dst, src, len);
+ memcpy(src, dst, len);
src += stride;
dst += stride;
}
@@ -247,29 +247,58 @@ static void sao_filter_CTB(HEVCContext *s, int x, int y)
for (c_idx = 0; c_idx < 3; c_idx++) {
int x0 = x >> s->sps->hshift[c_idx];
int y0 = y >> s->sps->vshift[c_idx];
- int stride = s->frame->linesize[c_idx];
+ int stride_src = s->frame->linesize[c_idx];
+ int stride_dst = s->sao_frame->linesize[c_idx];
int ctb_size_h = (1 << (s->sps->log2_ctb_size)) >> s->sps->hshift[c_idx];
int ctb_size_v = (1 << (s->sps->log2_ctb_size)) >> s->sps->vshift[c_idx];
- int width = FFMIN(ctb_size_h,
- (s->sps->width >> s->sps->hshift[c_idx]) - x0);
- int height = FFMIN(ctb_size_v,
- (s->sps->height >> s->sps->vshift[c_idx]) - y0);
-
- uint8_t *src = &s->frame->data[c_idx][y0 * stride + (x0 << s->sps->pixel_shift)];
- uint8_t *dst = &s->sao_frame->data[c_idx][y0 * stride + (x0 << s->sps->pixel_shift)];
+ int width = FFMIN(ctb_size_h, (s->sps->width >> s->sps->hshift[c_idx]) - x0);
+ int height = FFMIN(ctb_size_v, (s->sps->height >> s->sps->vshift[c_idx]) - y0);
+ uint8_t *src = &s->frame->data[c_idx][y0 * stride_src + (x0 << s->sps->pixel_shift)];
+ uint8_t *dst = &s->sao_frame->data[c_idx][y0 * stride_dst + (x0 << s->sps->pixel_shift)];
switch (sao->type_idx[c_idx]) {
case SAO_BAND:
- s->hevcdsp.sao_band_filter(dst, src,
- stride,
+ copy_CTB(dst, src, width << s->sps->pixel_shift, height, stride_dst, stride_src);
+ s->hevcdsp.sao_band_filter(src, dst,
+ stride_src, stride_dst,
sao,
edges, width,
height, c_idx);
restore_tqb_pixels(s, x, y, width, height, c_idx);
+ sao->type_idx[c_idx] = SAO_APPLIED;
break;
case SAO_EDGE:
- s->hevcdsp.sao_edge_filter[restore](dst, src,
- stride,
+ {
+ uint8_t left_pixels = !edges[0] && (CTB(s->sao, x_ctb-1, y_ctb).type_idx[c_idx] != SAO_APPLIED);
+ if (!edges[1]) {
+ uint8_t top_left = !edges[0] && (CTB(s->sao, x_ctb-1, y_ctb-1).type_idx[c_idx] != SAO_APPLIED);
+ uint8_t top_right = !edges[2] && (CTB(s->sao, x_ctb+1, y_ctb-1).type_idx[c_idx] != SAO_APPLIED);
+ if (CTB(s->sao, x_ctb , y_ctb-1).type_idx[c_idx] == 0)
+ memcpy( dst - stride_dst - (top_left << s->sps->pixel_shift),
+ src - stride_src - (top_left << s->sps->pixel_shift),
+ (top_left + width + top_right) << s->sps->pixel_shift);
+ else {
+ if (top_left)
+ memcpy( dst - stride_dst - (1 << s->sps->pixel_shift),
+ src - stride_src - (1 << s->sps->pixel_shift),
+ 1 << s->sps->pixel_shift);
+ if(top_right)
+ memcpy( dst - stride_dst + (width << s->sps->pixel_shift),
+ src - stride_src + (width << s->sps->pixel_shift),
+ 1 << s->sps->pixel_shift);
+ }
+ }
+ if (!edges[3]) { // bottom and bottom right
+ uint8_t bottom_left = !edges[0] && (CTB(s->sao, x_ctb-1, y_ctb+1).type_idx[c_idx] != SAO_APPLIED);
+ memcpy( dst + height * stride_dst - (bottom_left << s->sps->pixel_shift),
+ src + height * stride_src - (bottom_left << s->sps->pixel_shift),
+ (width + 1 + bottom_left) << s->sps->pixel_shift);
+ }
+ copy_CTB(dst - (left_pixels << s->sps->pixel_shift),
+ src - (left_pixels << s->sps->pixel_shift),
+ (width + 1 + left_pixels) << s->sps->pixel_shift, height, stride_dst, stride_src);
+ s->hevcdsp.sao_edge_filter[restore](src, dst,
+ stride_src, stride_dst,
sao,
edges, width,
height, c_idx,
@@ -277,10 +306,9 @@ static void sao_filter_CTB(HEVCContext *s, int x, int y)
horiz_edge,
diag_edge);
restore_tqb_pixels(s, x, y, width, height, c_idx);
+ sao->type_idx[c_idx] = SAO_APPLIED;
break;
- default :
- copy_CTB(dst, src, width << s->sps->pixel_shift, height, stride);
- break;
+ }
}
}
}
diff --git a/libavcodec/hevcdsp.h b/libavcodec/hevcdsp.h
index 3b82fab..dfff780 100644
--- a/libavcodec/hevcdsp.h
+++ b/libavcodec/hevcdsp.h
@@ -58,11 +58,11 @@ typedef struct HEVCDSPContext {
void (*transform_dc_add[4])(uint8_t *dst, int16_t *coeffs, ptrdiff_t stride);
- void (*sao_band_filter)(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride,
+ void (*sao_band_filter)(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src,
struct SAOParams *sao, int *borders,
int width, int height, int c_idx);
- void (*sao_edge_filter[2])(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride,
+ void (*sao_edge_filter[2])(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src,
struct SAOParams *sao, int *borders, int _width,
int _height, int c_idx, uint8_t *vert_edge,
uint8_t *horiz_edge, uint8_t *diag_edge);
diff --git a/libavcodec/hevcdsp_template.c b/libavcodec/hevcdsp_template.c
index f0bb256..7840ec7 100644
--- a/libavcodec/hevcdsp_template.c
+++ b/libavcodec/hevcdsp_template.c
@@ -329,7 +329,7 @@ IDCT_DC(32)
#undef ADD_AND_SCALE
static void FUNC(sao_band_filter_0)(uint8_t *_dst, uint8_t *_src,
- ptrdiff_t stride, SAOParams *sao,
+ ptrdiff_t stride_dst, ptrdiff_t stride_src, SAOParams *sao,
int *borders, int width, int height,
int c_idx)
{
@@ -341,22 +341,23 @@ static void FUNC(sao_band_filter_0)(uint8_t *_dst, uint8_t *_src,
int *sao_offset_val = sao->offset_val[c_idx];
int sao_left_class = sao->band_position[c_idx];
- stride /= sizeof(pixel);
+ stride_dst /= sizeof(pixel);
+ stride_src /= sizeof(pixel);
for (k = 0; k < 4; k++)
offset_table[(k + sao_left_class) & 31] = sao_offset_val[k + 1];
for (y = 0; y < height; y++) {
for (x = 0; x < width; x++)
- dst[x] = av_clip_pixel(src[x] + offset_table[av_clip_pixel(src[x] >> shift)]);
- dst += stride;
- src += stride;
+ dst[x] = av_clip_pixel(src[x] + offset_table[src[x] >> shift]);
+ dst += stride_dst;
+ src += stride_src;
}
}
#define CMP(a, b) ((a) > (b) ? 1 : ((a) == (b) ? 0 : -1))
static void FUNC(sao_edge_filter)(uint8_t *_dst, uint8_t *_src,
- ptrdiff_t stride, SAOParams *sao,
+ ptrdiff_t stride_dst, ptrdiff_t stride_src, SAOParams *sao,
int width, int height,
int c_idx, int init_x, int init_y) {
@@ -372,30 +373,32 @@ static void FUNC(sao_edge_filter)(uint8_t *_dst, uint8_t *_src,
pixel *dst = (pixel *)_dst;
pixel *src = (pixel *)_src;
- int y_stride = init_y * stride;
+ int y_stride_src = init_y * stride_src;
+ int y_stride_dst = init_y * stride_dst;
int pos_0_0 = pos[sao_eo_class][0][0];
int pos_0_1 = pos[sao_eo_class][0][1];
int pos_1_0 = pos[sao_eo_class][1][0];
int pos_1_1 = pos[sao_eo_class][1][1];
int x, y;
- int y_stride_0_1 = (init_y + pos_0_1) * stride;
- int y_stride_1_1 = (init_y + pos_1_1) * stride;
+ int y_stride_0_1 = (init_y + pos_0_1) * stride_src;
+ int y_stride_1_1 = (init_y + pos_1_1) * stride_src;
for (y = init_y; y < height; y++) {
for (x = init_x; x < width; x++) {
- int diff0 = CMP(src[x + y_stride], src[x + pos_0_0 + y_stride_0_1]);
- int diff1 = CMP(src[x + y_stride], src[x + pos_1_0 + y_stride_1_1]);
- int offset_val = edge_idx[2 + diff0 + diff1];
- dst[x + y_stride] = av_clip_pixel(src[x + y_stride] + sao_offset_val[offset_val]);
+ int diff0 = CMP(src[x + y_stride_src], src[x + pos_0_0 + y_stride_0_1]);
+ int diff1 = CMP(src[x + y_stride_src], src[x + pos_1_0 + y_stride_1_1]);
+ int offset_val = edge_idx[2 + diff0 + diff1];
+ dst[x + y_stride_dst] = av_clip_pixel(src[x + y_stride_src] + sao_offset_val[offset_val]);
}
- y_stride += stride;
- y_stride_0_1 += stride;
- y_stride_1_1 += stride;
+ y_stride_src += stride_src;
+ y_stride_dst += stride_dst;
+ y_stride_0_1 += stride_src;
+ y_stride_1_1 += stride_src;
}
}
static void FUNC(sao_edge_filter_0)(uint8_t *_dst, uint8_t *_src,
- ptrdiff_t stride, SAOParams *sao,
+ ptrdiff_t stride_dst, ptrdiff_t stride_src, SAOParams *sao,
int *borders, int _width, int _height,
int c_idx, uint8_t *vert_edge,
uint8_t *horiz_edge, uint8_t *diag_edge)
@@ -407,24 +410,22 @@ static void FUNC(sao_edge_filter_0)(uint8_t *_dst, uint8_t *_src,
int sao_eo_class = sao->eo_class[c_idx];
int init_x = 0, init_y = 0, width = _width, height = _height;
- stride /= sizeof(pixel);
+ stride_dst /= sizeof(pixel);
+ stride_src /= sizeof(pixel);
if (sao_eo_class != SAO_EO_VERT) {
if (borders[0]) {
int offset_val = sao_offset_val[0];
- int y_stride = 0;
for (y = 0; y < height; y++) {
- dst[y_stride] = av_clip_pixel(src[y_stride] + offset_val);
- y_stride += stride;
+ dst[y * stride_dst] = av_clip_pixel(src[y * stride_src] + offset_val);
}
init_x = 1;
}
if (borders[2]) {
int offset_val = sao_offset_val[0];
- int x_stride = width - 1;
+ int offset = width - 1;
for (x = 0; x < height; x++) {
- dst[x_stride] = av_clip_pixel(src[x_stride] + offset_val);
- x_stride += stride;
+ dst[x * stride_dst + offset] = av_clip_pixel(src[x * stride_src + offset] + offset_val);
}
width--;
}
@@ -437,19 +438,20 @@ static void FUNC(sao_edge_filter_0)(uint8_t *_dst, uint8_t *_src,
init_y = 1;
}
if (borders[3]) {
- int offset_val = sao_offset_val[0];
- int y_stride = stride * (height - 1);
+ int offset_val = sao_offset_val[0];
+ int y_stride_dst = stride_dst * (height - 1);
+ int y_stride_src = stride_src * (height - 1);
for (x = init_x; x < width; x++)
- dst[x + y_stride] = av_clip_pixel(src[x + y_stride] + offset_val);
+ dst[x + y_stride_dst] = av_clip_pixel(src[x + y_stride_src] + offset_val);
height--;
}
}
- FUNC(sao_edge_filter)((uint8_t *)dst, (uint8_t *)src, stride, sao, width, height, c_idx, init_x, init_y);
+ FUNC(sao_edge_filter)((uint8_t *)dst, (uint8_t *)src, stride_dst, stride_src, sao, width, height, c_idx, init_x, init_y);
}
static void FUNC(sao_edge_filter_1)(uint8_t *_dst, uint8_t *_src,
- ptrdiff_t stride, SAOParams *sao,
+ ptrdiff_t stride_dst, ptrdiff_t stride_src, SAOParams *sao,
int *borders, int _width, int _height,
int c_idx, uint8_t *vert_edge,
uint8_t *horiz_edge, uint8_t *diag_edge)
@@ -461,24 +463,22 @@ static void FUNC(sao_edge_filter_1)(uint8_t *_dst, uint8_t *_src,
int sao_eo_class = sao->eo_class[c_idx];
int init_x = 0, init_y = 0, width = _width, height = _height;
- stride /= sizeof(pixel);
+ stride_dst /= sizeof(pixel);
+ stride_src /= sizeof(pixel);
if (sao_eo_class != SAO_EO_VERT) {
if (borders[0]) {
int offset_val = sao_offset_val[0];
- int y_stride = 0;
for (y = 0; y < height; y++) {
- dst[y_stride] = av_clip_pixel(src[y_stride] + offset_val);
- y_stride += stride;
+ dst[y * stride_dst] = av_clip_pixel(src[y * stride_src] + offset_val);
}
init_x = 1;
}
if (borders[2]) {
int offset_val = sao_offset_val[0];
- int x_stride = width - 1;
+ int offset = width - 1;
for (x = 0; x < height; x++) {
- dst[x_stride] = av_clip_pixel(src[x_stride] + offset_val);
- x_stride += stride;
+ dst[x * stride_dst + offset] = av_clip_pixel(src[x * stride_src + offset] + offset_val);
}
width--;
}
@@ -491,15 +491,16 @@ static void FUNC(sao_edge_filter_1)(uint8_t *_dst, uint8_t *_src,
init_y = 1;
}
if (borders[3]) {
- int offset_val = sao_offset_val[0];
- int y_stride = stride * (height - 1);
+ int offset_val = sao_offset_val[0];
+ int y_stride_dst = stride_dst * (height - 1);
+ int y_stride_src = stride_src * (height - 1);
for (x = init_x; x < width; x++)
- dst[x + y_stride] = av_clip_pixel(src[x + y_stride] + offset_val);
+ dst[x + y_stride_dst] = av_clip_pixel(src[x + y_stride_src] + offset_val);
height--;
}
}
- FUNC(sao_edge_filter)((uint8_t *)dst, (uint8_t *)src, stride, sao, width, height, c_idx, init_x, init_y);
+ FUNC(sao_edge_filter)((uint8_t *)dst, (uint8_t *)src, stride_dst, stride_src, sao, width, height, c_idx, init_x, init_y);
{
int save_upper_left = !diag_edge[0] && sao_eo_class == SAO_EO_135D && !borders[0] && !borders[1];
@@ -510,11 +511,11 @@ static void FUNC(sao_edge_filter_1)(uint8_t *_dst, uint8_t *_src,
// Restore pixels that can't be modified
if(vert_edge[0] && sao_eo_class != SAO_EO_VERT) {
for(y = init_y+save_upper_left; y< height-save_lower_left; y++)
- dst[y*stride] = src[y*stride];
+ dst[y*stride_dst] = src[y*stride_src];
}
if(vert_edge[1] && sao_eo_class != SAO_EO_VERT) {
for(y = init_y+save_upper_right; y< height-save_lower_right; y++)
- dst[y*stride+width-1] = src[y*stride+width-1];
+ dst[y*stride_dst+width-1] = src[y*stride_src+width-1];
}
if(horiz_edge[0] && sao_eo_class != SAO_EO_HORIZ) {
@@ -523,16 +524,16 @@ static void FUNC(sao_edge_filter_1)(uint8_t *_dst, uint8_t *_src,
}
if(horiz_edge[1] && sao_eo_class != SAO_EO_HORIZ) {
for(x = init_x+save_lower_left; x < width-save_lower_right; x++)
- dst[(height-1)*stride+x] = src[(height-1)*stride+x];
+ dst[(height-1)*stride_dst+x] = src[(height-1)*stride_src+x];
}
if(diag_edge[0] && sao_eo_class == SAO_EO_135D)
dst[0] = src[0];
if(diag_edge[1] && sao_eo_class == SAO_EO_45D)
dst[width-1] = src[width-1];
if(diag_edge[2] && sao_eo_class == SAO_EO_135D)
- dst[stride*(height-1)+width-1] = src[stride*(height-1)+width-1];
+ dst[stride_dst*(height-1)+width-1] = src[stride_src*(height-1)+width-1];
if(diag_edge[3] && sao_eo_class == SAO_EO_45D)
- dst[stride*(height-1)] = src[stride*(height-1)];
+ dst[stride_dst*(height-1)] = src[stride_src*(height-1)];
}
}
More information about the ffmpeg-cvslog
mailing list