Skip to content

Commit

Permalink
vp9: split out loopfilter functions in their own source file.
Browse files Browse the repository at this point in the history
  • Loading branch information
@rbultje
rbultje committed Mar 28, 2017
1 parent f8c0199 commit b823bbc
Show file tree
Hide file tree
Showing 4 changed files with 208 additions and 180 deletions.
2 changes: 1 addition & 1 deletion libavcodec/Makefile
View file
Original file line number Diff line number Diff line change
Expand Up @@ -610,7 +610,7 @@ OBJS-$(CONFIG_VP8_DECODER) += vp8.o vp56rac.o
OBJS-$(CONFIG_VP8_CUVID_DECODER) += cuvid.o
OBJS-$(CONFIG_VP8_MEDIACODEC_DECODER) += mediacodecdec.o
OBJS-$(CONFIG_VP8_VAAPI_ENCODER) += vaapi_encode_vp8.o
OBJS-$(CONFIG_VP9_DECODER) += vp9.o vp9data.o vp9dsp.o \
OBJS-$(CONFIG_VP9_DECODER) += vp9.o vp9data.o vp9dsp.o vp9lpf.o \
vp9block.o vp9prob.o vp9mvs.o vp56rac.o \
vp9dsp_8bpp.o vp9dsp_10bpp.o vp9dsp_12bpp.o
OBJS-$(CONFIG_VP9_CUVID_DECODER) += cuvid.o
Expand Down
181 changes: 2 additions & 179 deletions libavcodec/vp9.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1056,184 +1056,6 @@ static void decode_sb_mem(AVCodecContext *avctx, int row, int col, VP9Filter *lf
}
}

static av_always_inline void filter_plane_cols(VP9Context *s, int col, int ss_h, int ss_v,
uint8_t *lvl, uint8_t (*mask)[4],
uint8_t *dst, ptrdiff_t ls)
{
int y, x, bytesperpixel = s->bytesperpixel;

// filter edges between columns (e.g. block1 | block2)
for (y = 0; y < 8; y += 2 << ss_v, dst += 16 * ls, lvl += 16 << ss_v) {
uint8_t *ptr = dst, *l = lvl, *hmask1 = mask[y], *hmask2 = mask[y + 1 + ss_v];
unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];
unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];
unsigned hm = hm1 | hm2 | hm13 | hm23;

for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8 * bytesperpixel >> ss_h) {
if (col || x > 1) {
if (hm1 & x) {
int L = *l, H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

if (hmask1[0] & x) {
if (hmask2[0] & x) {
av_assert2(l[8 << ss_v] == L);
s->dsp.loop_filter_16[0](ptr, ls, E, I, H);
} else {
s->dsp.loop_filter_8[2][0](ptr, ls, E, I, H);
}
} else if (hm2 & x) {
L = l[8 << ss_v];
H |= (L >> 4) << 8;
E |= s->filter_lut.mblim_lut[L] << 8;
I |= s->filter_lut.lim_lut[L] << 8;
s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]
[!!(hmask2[1] & x)]
[0](ptr, ls, E, I, H);
} else {
s->dsp.loop_filter_8[!!(hmask1[1] & x)]
[0](ptr, ls, E, I, H);
}
} else if (hm2 & x) {
int L = l[8 << ss_v], H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

s->dsp.loop_filter_8[!!(hmask2[1] & x)]
[0](ptr + 8 * ls, ls, E, I, H);
}
}
if (ss_h) {
if (x & 0xAA)
l += 2;
} else {
if (hm13 & x) {
int L = *l, H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

if (hm23 & x) {
L = l[8 << ss_v];
H |= (L >> 4) << 8;
E |= s->filter_lut.mblim_lut[L] << 8;
I |= s->filter_lut.lim_lut[L] << 8;
s->dsp.loop_filter_mix2[0][0][0](ptr + 4 * bytesperpixel, ls, E, I, H);
} else {
s->dsp.loop_filter_8[0][0](ptr + 4 * bytesperpixel, ls, E, I, H);
}
} else if (hm23 & x) {
int L = l[8 << ss_v], H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

s->dsp.loop_filter_8[0][0](ptr + 8 * ls + 4 * bytesperpixel, ls, E, I, H);
}
l++;
}
}
}
}

static av_always_inline void filter_plane_rows(VP9Context *s, int row, int ss_h, int ss_v,
uint8_t *lvl, uint8_t (*mask)[4],
uint8_t *dst, ptrdiff_t ls)
{
int y, x, bytesperpixel = s->bytesperpixel;

// block1
// filter edges between rows (e.g. ------)
// block2
for (y = 0; y < 8; y++, dst += 8 * ls >> ss_v) {
uint8_t *ptr = dst, *l = lvl, *vmask = mask[y];
unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

for (x = 1; vm & ~(x - 1); x <<= (2 << ss_h), ptr += 16 * bytesperpixel, l += 2 << ss_h) {
if (row || y) {
if (vm & x) {
int L = *l, H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

if (vmask[0] & x) {
if (vmask[0] & (x << (1 + ss_h))) {
av_assert2(l[1 + ss_h] == L);
s->dsp.loop_filter_16[1](ptr, ls, E, I, H);
} else {
s->dsp.loop_filter_8[2][1](ptr, ls, E, I, H);
}
} else if (vm & (x << (1 + ss_h))) {
L = l[1 + ss_h];
H |= (L >> 4) << 8;
E |= s->filter_lut.mblim_lut[L] << 8;
I |= s->filter_lut.lim_lut[L] << 8;
s->dsp.loop_filter_mix2[!!(vmask[1] & x)]
[!!(vmask[1] & (x << (1 + ss_h)))]
[1](ptr, ls, E, I, H);
} else {
s->dsp.loop_filter_8[!!(vmask[1] & x)]
[1](ptr, ls, E, I, H);
}
} else if (vm & (x << (1 + ss_h))) {
int L = l[1 + ss_h], H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

s->dsp.loop_filter_8[!!(vmask[1] & (x << (1 + ss_h)))]
[1](ptr + 8 * bytesperpixel, ls, E, I, H);
}
}
if (!ss_v) {
if (vm3 & x) {
int L = *l, H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

if (vm3 & (x << (1 + ss_h))) {
L = l[1 + ss_h];
H |= (L >> 4) << 8;
E |= s->filter_lut.mblim_lut[L] << 8;
I |= s->filter_lut.lim_lut[L] << 8;
s->dsp.loop_filter_mix2[0][0][1](ptr + ls * 4, ls, E, I, H);
} else {
s->dsp.loop_filter_8[0][1](ptr + ls * 4, ls, E, I, H);
}
} else if (vm3 & (x << (1 + ss_h))) {
int L = l[1 + ss_h], H = L >> 4;
int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

s->dsp.loop_filter_8[0][1](ptr + ls * 4 + 8 * bytesperpixel, ls, E, I, H);
}
}
}
if (ss_v) {
if (y & 1)
lvl += 16;
} else {
lvl += 8;
}
}
}

static void loopfilter_sb(AVCodecContext *avctx, VP9Filter *lflvl,
int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)
{
VP9Context *s = avctx->priv_data;
AVFrame *f = s->s.frames[CUR_FRAME].tf.f;
uint8_t *dst = f->data[0] + yoff;
ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];
uint8_t (*uv_masks)[8][4] = lflvl->mask[s->ss_h | s->ss_v];
int p;

/* FIXME: In how far can we interleave the v/h loopfilter calls? E.g.
* if you think of them as acting on a 8x8 block max, we can interleave
* each v/h within the single x loop, but that only works if we work on
* 8 pixel blocks, and we won't always do that (we want at least 16px
* to use SSE2 optimizations, perhaps 32 for AVX2) */

filter_plane_cols(s, col, 0, 0, lflvl->level, lflvl->mask[0][0], dst, ls_y);
filter_plane_rows(s, row, 0, 0, lflvl->level, lflvl->mask[0][1], dst, ls_y);

for (p = 0; p < 2; p++) {
dst = f->data[1 + p] + uvoff;
filter_plane_cols(s, col, s->ss_h, s->ss_v, lflvl->level, uv_masks[0], dst, ls_uv);
filter_plane_rows(s, row, s->ss_h, s->ss_v, lflvl->level, uv_masks[1], dst, ls_uv);
}
}

static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)
{
int sb_start = ( idx * n) >> log2_n;
Expand Down Expand Up @@ -1522,7 +1344,8 @@ FF_ENABLE_DEPRECATION_WARNINGS
for (col = 0; col < s->cols;
col += 8, yoff2 += 64 * bytesperpixel,
uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {
loopfilter_sb(avctx, lflvl_ptr, row, col, yoff2, uvoff2);
ff_vp9_loopfilter_sb(avctx, lflvl_ptr, row, col,
yoff2, uvoff2);
}
}

Expand Down
3 changes: 3 additions & 0 deletions libavcodec/vp9dec.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -203,4 +203,7 @@ void ff_vp9_decode_block(AVCodecContext *ctx, int row, int col,
VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
enum BlockLevel bl, enum BlockPartition bp);

void ff_vp9_loopfilter_sb(AVCodecContext *avctx, VP9Filter *lflvl,
int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff);

#endif /* AVCODEC_VP9DEC_H */
Loading

0 comments on commit b823bbc

Please sign in to comment.