Modify motion estimation for SPLITMV mode
1. Search for block8x16/block16x8 uses block8x8's search results.
2. Check block4x4 only if block8x8 is chosen. (This hurts quality,
which will be improved in another check-in.)
3. In block4x4 search, the previous block's result is used as
MV predictor for next block.
This change improves performance.
Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
diff --git a/vp8/encoder/rdopt.c b/vp8/encoder/rdopt.c
index f1e0579..a88f687 100644
--- a/vp8/encoder/rdopt.c
+++ b/vp8/encoder/rdopt.c
@@ -1031,6 +1031,7 @@
typedef struct
{
MV *ref_mv;
+ MV *mvp;
int segment_rd;
int segment_num;
@@ -1044,6 +1045,9 @@
int mvthresh;
int *mdcounts;
+ MV sv_mvp[4]; // save 4 mvp from 8x8
+ int sv_istep[2]; // save 2 initial step_param for 16x8/8x16
+
} BEST_SEG_INFO;
@@ -1129,7 +1133,7 @@
int sseshift;
int num00;
int step_param = 0;
- int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+ int further_steps;
int n;
int thissme;
int bestsme = INT_MAX;
@@ -1141,6 +1145,27 @@
if (best_label_rd < label_mv_thresh)
break;
+ if(cpi->compressor_speed)
+ {
+ if (segmentation == BLOCK_8X16 || segmentation == BLOCK_16X8)
+ {
+ bsi->mvp = &bsi->sv_mvp[i];
+ if (i==1 && segmentation == BLOCK_16X8) bsi->mvp = &bsi->sv_mvp[2];
+
+ step_param = bsi->sv_istep[i];
+ }
+
+ // use previous block's result as next block's MV predictor.
+ if (segmentation == BLOCK_4X4 && i>0)
+ {
+ bsi->mvp = &(x->e_mbd.block[i-1].bmi.mv.as_mv);
+ if (i==4 || i==8 || i==12) bsi->mvp = &(x->e_mbd.block[i-4].bmi.mv.as_mv);
+ step_param = 2;
+ }
+ }
+
+ further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+
{
int sadpb = x->sadperbit4;
@@ -1156,7 +1181,7 @@
else
{
- bestsme = cpi->diamond_search_sad(x, c, e, bsi->ref_mv,
+ bestsme = cpi->diamond_search_sad(x, c, e, bsi->mvp,
&mode_mv[NEW4X4], step_param,
sadpb / 2, &num00, v_fn_ptr, x->mvsadcost, x->mvcost, bsi->ref_mv);
@@ -1171,7 +1196,7 @@
num00--;
else
{
- thissme = cpi->diamond_search_sad(x, c, e, bsi->ref_mv,
+ thissme = cpi->diamond_search_sad(x, c, e, bsi->mvp,
&temp_mv, step_param + n,
sadpb / 2, &num00, v_fn_ptr, x->mvsadcost, x->mvcost, bsi->ref_mv);
@@ -1190,7 +1215,7 @@
// Should we do a full search (best quality only)
if ((cpi->compressor_speed == 0) && (bestsme >> sseshift) > 4000)
{
- thissme = cpi->full_search_sad(x, c, e, bsi->ref_mv,
+ thissme = cpi->full_search_sad(x, c, e, bsi->mvp,
sadpb / 4, 16, v_fn_ptr, x->mvcost, x->mvsadcost,bsi->ref_mv);
if (thissme < bestsme)
@@ -1259,8 +1284,9 @@
segmentyrate += bestlabelyrate;
this_segment_rd += best_label_rd;
- if (this_segment_rd > bsi->segment_rd)
+ if (this_segment_rd >= bsi->segment_rd)
break;
+
} /* for each label */
if (this_segment_rd < bsi->segment_rd)
@@ -1282,6 +1308,21 @@
}
}
}
+
+static __inline
+void vp8_cal_step_param(int sr, int *sp)
+{
+ int step = 0;
+
+ if (sr > MAX_FIRST_STEP) sr = MAX_FIRST_STEP;
+ else if (sr < 1) sr = 1;
+
+ while (sr>>=1)
+ step++;
+
+ *sp = MAX_MVSEARCH_STEPS - 1 - step;
+}
+
static int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x,
MV *best_ref_mv, int best_rd,
int *mdcounts, int *returntotrate,
@@ -1290,14 +1331,12 @@
{
int i;
BEST_SEG_INFO bsi;
- BEST_SEG_INFO bsi_8x8;
- int check_8x16 = 0;
- int check_16x8 = 0;
vpx_memset(&bsi, 0, sizeof(bsi));
bsi.segment_rd = best_rd;
bsi.ref_mv = best_ref_mv;
+ bsi.mvp = best_ref_mv;
bsi.mvthresh = mvthresh;
bsi.mdcounts = mdcounts;
@@ -1305,6 +1344,7 @@
{
bsi.modes[i] = ZERO4X4;
}
+
if(cpi->compressor_speed == 0)
{
/* for now, we will keep the original segmentation order
@@ -1316,12 +1356,46 @@
}
else
{
+ int sr;
+
vp8_rd_check_segment(cpi, x, &bsi, BLOCK_8X8);
+
if (bsi.segment_rd < best_rd)
{
- vp8_rd_check_segment(cpi, x, &bsi, BLOCK_8X16);
- vp8_rd_check_segment(cpi, x, &bsi, BLOCK_16X8);
- vp8_rd_check_segment(cpi, x, &bsi, BLOCK_4X4);
+ bsi.sv_mvp[0] = bsi.mvs[0].as_mv;
+ bsi.sv_mvp[1] = bsi.mvs[2].as_mv;
+ bsi.sv_mvp[2] = bsi.mvs[8].as_mv;
+ bsi.sv_mvp[3] = bsi.mvs[10].as_mv;
+
+ // Use 8x8 result as 16x8/8x16's predictor MV. Adjust search range according to the closeness of 2 MV.
+ //block 8X16
+ {
+ sr = MAXF((abs(bsi.sv_mvp[0].row - bsi.sv_mvp[2].row))>>3, (abs(bsi.sv_mvp[0].col - bsi.sv_mvp[2].col))>>3);
+ vp8_cal_step_param(sr, &bsi.sv_istep[0]);
+
+ sr = MAXF((abs(bsi.sv_mvp[1].row - bsi.sv_mvp[3].row))>>3, (abs(bsi.sv_mvp[1].col - bsi.sv_mvp[3].col))>>3);
+ vp8_cal_step_param(sr, &bsi.sv_istep[1]);
+
+ vp8_rd_check_segment(cpi, x, &bsi, BLOCK_8X16);
+ }
+
+ //block 16X8
+ {
+ sr = MAXF((abs(bsi.sv_mvp[0].row - bsi.sv_mvp[1].row))>>3, (abs(bsi.sv_mvp[0].col - bsi.sv_mvp[1].col))>>3);
+ vp8_cal_step_param(sr, &bsi.sv_istep[0]);
+
+ sr = MAXF((abs(bsi.sv_mvp[2].row - bsi.sv_mvp[3].row))>>3, (abs(bsi.sv_mvp[2].col - bsi.sv_mvp[3].col))>>3);
+ vp8_cal_step_param(sr, &bsi.sv_istep[1]);
+
+ vp8_rd_check_segment(cpi, x, &bsi, BLOCK_16X8);
+ }
+
+ // If 8x8 is better than 16x8/8x16, then do 4x4 search
+ if (bsi.segment_num == BLOCK_8X8) // || (sv_segment_rd8x8-bsi.segment_rd) < sv_segment_rd8x8>>5)
+ {
+ bsi.mvp = &bsi.sv_mvp[0];
+ vp8_rd_check_segment(cpi, x, &bsi, BLOCK_4X4);
+ }
}
}
