Remove redundant dering passes in CDEF search
Change-Id: I3b932544426cb9aeb63ab4e1f516a4d8195d0702
diff --git a/av1/common/cdef.c b/av1/common/cdef.c
index 71ce202..c578eb9 100644
--- a/av1/common/cdef.c
+++ b/av1/common/cdef.c
@@ -382,7 +382,7 @@
od_dering(dst,
&src[OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER],
dec[pli], dir, NULL, var, pli, dlist, dering_count, threshold,
- clpf_strength, clpf_damping, coeff_shift);
+ clpf_strength, clpf_damping, coeff_shift, 0);
#if CONFIG_AOM_HIGHBITDEPTH
if (cm->use_highbitdepth) {
copy_dering_16bit_to_16bit(
diff --git a/av1/common/od_dering.c b/av1/common/od_dering.c
index 849c9b1..65ea4f2 100644
--- a/av1/common/od_dering.c
+++ b/av1/common/od_dering.c
@@ -243,7 +243,8 @@
int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS], int *dirinit,
int var[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS], int pli,
dering_list *dlist, int dering_count, int threshold,
- int clpf_strength, int clpf_damping, int coeff_shift) {
+ int clpf_strength, int clpf_damping, int coeff_shift,
+ int skip_dering) {
int bi;
int bx;
int by;
@@ -252,45 +253,49 @@
od_filter_dering_direction_4x4, od_filter_dering_direction_8x8
};
bsize = OD_DERING_SIZE_LOG2 - xdec;
- if (pli == 0) {
- if (!dirinit || !*dirinit) {
+ if (!skip_dering) {
+ if (pli == 0) {
+ if (!dirinit || !*dirinit) {
+ for (bi = 0; bi < dering_count; bi++) {
+ by = dlist[bi].by;
+ bx = dlist[bi].bx;
+ dir[by][bx] =
+ od_dir_find8(&in[8 * by * OD_FILT_BSTRIDE + 8 * bx],
+ OD_FILT_BSTRIDE, &var[by][bx], coeff_shift);
+ }
+ if (dirinit) *dirinit = 1;
+ }
for (bi = 0; bi < dering_count; bi++) {
by = dlist[bi].by;
bx = dlist[bi].bx;
- dir[by][bx] = od_dir_find8(&in[8 * by * OD_FILT_BSTRIDE + 8 * bx],
- OD_FILT_BSTRIDE, &var[by][bx], coeff_shift);
+ /* Deringing orthogonal to the direction uses a tighter threshold
+ because we want to be conservative. We've presumably already
+ achieved some deringing, so the amount of change is expected
+ to be low. Also, since we might be filtering across an edge, we
+ want to make sure not to blur it. That being said, we might want
+ to be a little bit more aggressive on pure horizontal/vertical
+ since the ringing there tends to be directional, so it doesn't
+ get removed by the directional filtering. */
+ (filter_dering_direction[bsize - OD_LOG_BSIZE0])(
+ &y[bi << 2 * bsize], 1 << bsize,
+ &in[(by * OD_FILT_BSTRIDE << bsize) + (bx << bsize)],
+ od_adjust_thresh(threshold, var[by][bx]), dir[by][bx]);
}
- if (dirinit) *dirinit = 1;
- }
- for (bi = 0; bi < dering_count; bi++) {
- by = dlist[bi].by;
- bx = dlist[bi].bx;
- /* Deringing orthogonal to the direction uses a tighter threshold
- because we want to be conservative. We've presumably already
- achieved some deringing, so the amount of change is expected
- to be low. Also, since we might be filtering across an edge, we
- want to make sure not to blur it. That being said, we might want
- to be a little bit more aggressive on pure horizontal/vertical
- since the ringing there tends to be directional, so it doesn't
- get removed by the directional filtering. */
- (filter_dering_direction[bsize - OD_LOG_BSIZE0])(
- &y[bi << 2 * bsize], 1 << bsize,
- &in[(by * OD_FILT_BSTRIDE << bsize) + (bx << bsize)],
- od_adjust_thresh(threshold, var[by][bx]), dir[by][bx]);
- }
- } else {
- for (bi = 0; bi < dering_count; bi++) {
- by = dlist[bi].by;
- bx = dlist[bi].bx;
- (filter_dering_direction[bsize - OD_LOG_BSIZE0])(
- &y[bi << 2 * bsize], 1 << bsize,
- &in[(by * OD_FILT_BSTRIDE << bsize) + (bx << bsize)], threshold,
- dir[by][bx]);
+ } else {
+ for (bi = 0; bi < dering_count; bi++) {
+ by = dlist[bi].by;
+ bx = dlist[bi].bx;
+ (filter_dering_direction[bsize - OD_LOG_BSIZE0])(
+ &y[bi << 2 * bsize], 1 << bsize,
+ &in[(by * OD_FILT_BSTRIDE << bsize) + (bx << bsize)], threshold,
+ dir[by][bx]);
+ }
}
}
if (!clpf_strength) return;
- copy_dering_16bit_to_16bit(in, OD_FILT_BSTRIDE, y, dlist, dering_count,
- bsize);
+ if (threshold && !skip_dering)
+ copy_dering_16bit_to_16bit(in, OD_FILT_BSTRIDE, y, dlist, dering_count,
+ bsize);
for (bi = 0; bi < dering_count; bi++) {
by = dlist[bi].by;
bx = dlist[bi].bx;
diff --git a/av1/common/od_dering.h b/av1/common/od_dering.h
index 60d17bc..a3efec2 100644
--- a/av1/common/od_dering.h
+++ b/av1/common/od_dering.h
@@ -50,7 +50,8 @@
int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS], int *dirinit,
int var[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS], int pli,
dering_list *dlist, int dering_count, int threshold,
- int clpf_strength, int clpf_damping, int coeff_shift);
+ int clpf_strength, int clpf_damping, int coeff_shift,
+ int skip_dering);
int od_filter_dering_direction_4x4_c(uint16_t *y, int ystride,
const uint16_t *in, int threshold,
int dir);
diff --git a/av1/encoder/pickcdef.c b/av1/encoder/pickcdef.c
index 9a9813c..bcb82b7 100644
--- a/av1/encoder/pickcdef.c
+++ b/av1/encoder/pickcdef.c
@@ -257,15 +257,20 @@
(nvb << bsize[pli]) + OD_FILT_VBORDER * (sbr != nvsb - 1) + yoff;
int xsize =
(nhb << bsize[pli]) + OD_FILT_HBORDER * (sbc != nhsb - 1) + xoff;
- copy_sb16_16(&in[(-yoff * OD_FILT_BSTRIDE - xoff)], OD_FILT_BSTRIDE,
- src[pli], (sbr * MAX_MIB_SIZE << bsize[pli]) - yoff,
- (sbc * MAX_MIB_SIZE << bsize[pli]) - xoff, stride[pli],
- ysize, xsize);
clpf_strength = gi % CLPF_STRENGTHS;
+ if (clpf_strength == 0)
+ copy_sb16_16(&in[(-yoff * OD_FILT_BSTRIDE - xoff)], OD_FILT_BSTRIDE,
+ src[pli], (sbr * MAX_MIB_SIZE << bsize[pli]) - yoff,
+ (sbc * MAX_MIB_SIZE << bsize[pli]) - xoff, stride[pli],
+ ysize, xsize);
od_dering(tmp_dst, in, dec[pli], dir, &dirinit, var, pli, dlist,
dering_count, threshold,
clpf_strength + (clpf_strength == 3), clpf_damping,
- coeff_shift);
+ coeff_shift, clpf_strength != 0);
+ if (clpf_strength == 0) {
+ copy_dering_16bit_to_16bit(in, OD_FILT_BSTRIDE, tmp_dst, dlist,
+ dering_count, bsize[pli]);
+ }
mse[pli][sb_count][gi] = compute_dering_mse(
ref_coeff[pli] +
(sbr * MAX_MIB_SIZE << bsize[pli]) * stride[pli] +
