Do sub-pixel motion search in up-sampled reference frames
Up-sampled the reference frames to 8 times in each dimension using
the 8-tap interpolation filter. In sub-pixel motion search, use the
up-sampled reference frames to find the best matching blocks. This
largely improved the motion search precision, and thus, improved
the compression quality. There was no change in decoder side.
Borg test and speed test results:
1. On derflr set,
Overall PSNR gain: 1.306%, and SSIM gain: 1.512%.
Average speed loss on derf set was 6.0%.
2. On stdhd set,
Overall PSNR gain: 0.754%, and SSIM gain: 0.814%.
On hevchd set,
Overall PSNR gain: 0.465%, and SSIM gain: 0.527%.
Speed loss on HD clips was 3.5%.
Change-Id: I300ebaafff57e88914f3dedc8784cb21d316b04f
diff --git a/vp10/encoder/rdopt.c b/vp10/encoder/rdopt.c
index 03aa9f0..5c74d32 100644
--- a/vp10/encoder/rdopt.c
+++ b/vp10/encoder/rdopt.c
@@ -3929,7 +3929,8 @@
int_mv* ref_mv_sub8x8[2],
#endif
int_mv single_newmv[MAX_REF_FRAMES],
- int *rate_mv) {
+ int *rate_mv,
+ const int block) {
const VP10_COMMON *const cm = &cpi->common;
const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
@@ -4076,6 +4077,40 @@
if (bestsme < INT_MAX) {
int dis; /* TODO: use dis in distortion calculation later. */
unsigned int sse;
+#if CONFIG_AFFINE_MOTION
+ // Use up-sampled reference frames.
+ struct macroblockd_plane *const pd = &xd->plane[0];
+ struct buf_2d backup_pred = pd->pre[0];
+ const YV12_BUFFER_CONFIG *upsampled_ref =
+ get_upsampled_ref(cpi, refs[id]);
+
+ // Set pred for Y plane
+ setup_pred_plane(&pd->pre[0], upsampled_ref->y_buffer,
+ upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3),
+ NULL, pd->subsampling_x, pd->subsampling_y);
+
+ // If bsize < BLOCK_8X8, adjust pred pointer for this block
+ if (bsize < BLOCK_8X8)
+ pd->pre[0].buf =
+ &pd->pre[0].buf[(vp10_raster_block_offset(BLOCK_8X8, block,
+ pd->pre[0].stride)) << 3];
+
+ bestsme = cpi->find_fractional_mv_step(
+ x, &tmp_mv,
+ &ref_mv[id].as_mv,
+ cpi->common.allow_high_precision_mv,
+ x->errorperbit,
+ &cpi->fn_ptr[bsize],
+ 0, cpi->sf.mv.subpel_iters_per_step,
+ NULL,
+ x->nmvjointcost, x->mvcost,
+ &dis, &sse, second_pred,
+ pw, ph, 1);
+
+ // Restore the reference frames.
+ pd->pre[0] = backup_pred;
+#else
+ (void) block;
bestsme = cpi->find_fractional_mv_step(
x, &tmp_mv,
&ref_mv[id].as_mv,
@@ -4087,6 +4122,7 @@
x->nmvjointcost, x->mvcost,
&dis, &sse, second_pred,
pw, ph);
+#endif
}
// Restore the pointer to the first (possibly scaled) prediction buffer.
@@ -4367,6 +4403,43 @@
if (bestsme < INT_MAX) {
int distortion;
+#if CONFIG_AFFINE_MOTION
+ const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
+ const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
+ // Use up-sampled reference frames.
+ struct macroblockd_plane *const pd = &xd->plane[0];
+ struct buf_2d backup_pred = pd->pre[0];
+ const YV12_BUFFER_CONFIG *upsampled_ref =
+ get_upsampled_ref(cpi, mbmi->ref_frame[0]);
+
+ // Set pred for Y plane
+ setup_pred_plane(&pd->pre[0], upsampled_ref->y_buffer,
+ upsampled_ref->y_stride,
+ (mi_row << 3), (mi_col << 3),
+ NULL, pd->subsampling_x, pd->subsampling_y);
+
+ // adjust pred pointer for this block
+ pd->pre[0].buf =
+ &pd->pre[0].buf[(vp10_raster_block_offset(BLOCK_8X8, i,
+ pd->pre[0].stride)) << 3];
+
+ cpi->find_fractional_mv_step(
+ x,
+ new_mv,
+ &bsi->ref_mv[0]->as_mv,
+ cm->allow_high_precision_mv,
+ x->errorperbit, &cpi->fn_ptr[bsize],
+ cpi->sf.mv.subpel_force_stop,
+ cpi->sf.mv.subpel_iters_per_step,
+ cond_cost_list(cpi, cost_list),
+ x->nmvjointcost, x->mvcost,
+ &distortion,
+ &x->pred_sse[mbmi->ref_frame[0]],
+ NULL, pw, ph, 1);
+
+ // Restore the reference frames.
+ pd->pre[0] = backup_pred;
+#else
cpi->find_fractional_mv_step(
x,
new_mv,
@@ -4380,6 +4453,7 @@
&distortion,
&x->pred_sse[mbmi->ref_frame[0]],
NULL, 0, 0);
+#endif
// save motion search result for use in compound prediction
#if CONFIG_EXT_INTER
@@ -4426,7 +4500,7 @@
#else
seg_mvs[i],
#endif // CONFIG_EXT_INTER
- &rate_mv);
+ &rate_mv, i);
#if CONFIG_EXT_INTER
compound_seg_newmvs[i][0].as_int =
frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
@@ -4975,6 +5049,33 @@
if (bestsme < INT_MAX) {
int dis; /* TODO: use dis in distortion calculation later. */
+#if CONFIG_AFFINE_MOTION
+ const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
+ const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
+ // Use up-sampled reference frames.
+ struct macroblockd_plane *const pd = &xd->plane[0];
+ struct buf_2d backup_pred = pd->pre[0];
+ const YV12_BUFFER_CONFIG *upsampled_ref = get_upsampled_ref(cpi, ref);
+
+ // Set pred for Y plane
+ setup_pred_plane(&pd->pre[0], upsampled_ref->y_buffer,
+ upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3),
+ NULL, pd->subsampling_x, pd->subsampling_y);
+
+ bestsme = cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
+ cm->allow_high_precision_mv,
+ x->errorperbit,
+ &cpi->fn_ptr[bsize],
+ cpi->sf.mv.subpel_force_stop,
+ cpi->sf.mv.subpel_iters_per_step,
+ cond_cost_list(cpi, cost_list),
+ x->nmvjointcost, x->mvcost,
+ &dis, &x->pred_sse[ref], NULL,
+ pw, ph, 1);
+
+ // Restore the reference frames.
+ pd->pre[0] = backup_pred;
+#else
cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
cm->allow_high_precision_mv,
x->errorperbit,
@@ -4984,6 +5085,7 @@
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&dis, &x->pred_sse[ref], NULL, 0, 0);
+#endif
}
*rate_mv = vp10_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
@@ -5328,7 +5430,7 @@
if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
joint_motion_search(cpi, x, bsize, frame_mv,
- mi_row, mi_col, NULL, single_newmv, &rate_mv);
+ mi_row, mi_col, NULL, single_newmv, &rate_mv, 0);
} else {
rate_mv = vp10_mv_bit_cost(&frame_mv[refs[0]].as_mv,
&x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
@@ -5358,7 +5460,7 @@
if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
joint_motion_search(cpi, x, bsize, frame_mv,
mi_row, mi_col,
- single_newmv, &rate_mv);
+ single_newmv, &rate_mv, 0);
} else {
rate_mv = vp10_mv_bit_cost(&frame_mv[refs[0]].as_mv,
&x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
