Adjust the sample selection in Warped Motion
This CL follows the original one"Sample selection in warped motion", and
aims to reduce the number of samples gathered. We only search 1 above row
and 1 left column instead of 1/2 bsize range in each direction. The gain
is a little less, but not much.
Borg test result:
avg_psnr ovr_psnr ssim
cam_lowres: -0.225 -0.228 -0.231
lowres: -0.081 -0.075 -0.114
The changes are wrapped in WARPED_MOTION_SORT_SAMPLES macro.
Change-Id: Ifb499c80470b2ffe7a40f68fc401589faada7730
diff --git a/av1/common/mvref_common.c b/av1/common/mvref_common.c
index 92a3df3..6c0cc215 100644
--- a/av1/common/mvref_common.c
+++ b/av1/common/mvref_common.c
@@ -1099,71 +1099,49 @@
int ref_frame = mbmi0->ref_frame[0];
int up_available = xd->up_available;
int left_available = xd->left_available;
- int i, mi_step = 1, np = 0, n, j, k;
+ int i, mi_step = 1, np = 0;
int global_offset_c = mi_col * MI_SIZE;
int global_offset_r = mi_row * MI_SIZE;
const TileInfo *const tile = &xd->tile;
- // Search nb range in the unit of mi
- int bs =
- (AOMMAX(xd->n8_w, xd->n8_h) > 1) ? (AOMMAX(xd->n8_w, xd->n8_h) >> 1) : 1;
- int marked[16 * 32]; // max array size for 128x128
int do_tl = 1;
int do_tr = 1;
- // scan the above rows
+ // scan the nearest above rows
if (up_available) {
- for (n = 0; n < bs; n++) {
- int mi_row_offset = -1 * (n + 1);
+ int mi_row_offset = -1;
+ MODE_INFO *mi = xd->mi[mi_row_offset * xd->mi_stride];
+ MB_MODE_INFO *mbmi = &mi->mbmi;
+ uint8_t n8_w = mi_size_wide[mbmi->sb_type];
- if (!n) {
- MODE_INFO *mi = xd->mi[mi_row_offset * xd->mi_stride];
- MB_MODE_INFO *mbmi = &mi->mbmi;
- uint8_t n8_w = mi_size_wide[mbmi->sb_type];
+ if (xd->n8_w <= n8_w) {
+ // Handle "current block width <= above block width" case.
+ int col_offset = -mi_col % n8_w;
- // Handle "current block width <= above block width" case.
- if (xd->n8_w <= n8_w) {
- int col_offset = -mi_col % n8_w;
+ if (col_offset < 0) do_tl = 0;
+ if (col_offset + n8_w > xd->n8_w) do_tr = 0;
- if (col_offset < 0) do_tl = 0;
- if (col_offset + n8_w > xd->n8_w) do_tr = 0;
-
- if (mbmi->ref_frame[0] == ref_frame &&
- mbmi->ref_frame[1] == NONE_FRAME) {
- record_samples(mbmi, pts, pts_inref, pts_mv, global_offset_r,
- global_offset_c, 0, -1, col_offset, 1);
- pts += 2;
- pts_inref += 2;
- pts_mv += 2;
- np++;
- }
- break;
- }
+ if (mbmi->ref_frame[0] == ref_frame && mbmi->ref_frame[1] == NONE_FRAME) {
+ record_samples(mbmi, pts, pts_inref, pts_mv, global_offset_r,
+ global_offset_c, 0, -1, col_offset, 1);
+ pts += 2;
+ pts_inref += 2;
+ pts_mv += 2;
+ np++;
}
-
+ } else {
// Handle "current block width > above block width" case.
- if (!n) memset(marked, 0, bs * xd->n8_w * sizeof(*marked));
-
for (i = 0; i < AOMMIN(xd->n8_w, cm->mi_cols - mi_col); i += mi_step) {
int mi_col_offset = i;
- MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
- MB_MODE_INFO *mbmi = &mi->mbmi;
- uint8_t n8_w = mi_size_wide[mbmi->sb_type];
- uint8_t n8_h = mi_size_high[mbmi->sb_type];
-
+ mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+ mbmi = &mi->mbmi;
+ n8_w = mi_size_wide[mbmi->sb_type];
mi_step = AOMMIN(xd->n8_w, n8_w);
- // Processed already
- if (marked[n * xd->n8_w + i]) continue;
-
- for (j = 0; j < AOMMIN(bs, n8_h); j++)
- for (k = 0; k < AOMMIN(xd->n8_w, n8_w); k++)
- marked[(n + j) * xd->n8_w + i + k] = 1;
-
if (mbmi->ref_frame[0] == ref_frame &&
mbmi->ref_frame[1] == NONE_FRAME) {
record_samples(mbmi, pts, pts_inref, pts_mv, global_offset_r,
- global_offset_c, -n, -1, i, 1);
+ global_offset_c, 0, -1, i, 1);
pts += 2;
pts_inref += 2;
pts_mv += 2;
@@ -1174,58 +1152,41 @@
}
assert(2 * np <= SAMPLES_ARRAY_SIZE);
- // scan the left columns
+ // scan the nearest left columns
if (left_available) {
- for (n = 0; n < bs; n++) {
- int mi_col_offset = -1 * (n + 1);
+ int mi_col_offset = -1;
- if (!n) {
- MODE_INFO *mi = xd->mi[mi_col_offset];
- MB_MODE_INFO *mbmi = &mi->mbmi;
- uint8_t n8_h = mi_size_high[mbmi->sb_type];
+ MODE_INFO *mi = xd->mi[mi_col_offset];
+ MB_MODE_INFO *mbmi = &mi->mbmi;
+ uint8_t n8_h = mi_size_high[mbmi->sb_type];
- // Handle "current block height <= above block height" case.
- if (xd->n8_h <= n8_h) {
- int row_offset = -mi_row % n8_h;
+ if (xd->n8_h <= n8_h) {
+ // Handle "current block height <= above block height" case.
+ int row_offset = -mi_row % n8_h;
- if (row_offset < 0) do_tl = 0;
+ if (row_offset < 0) do_tl = 0;
- if (mbmi->ref_frame[0] == ref_frame &&
- mbmi->ref_frame[1] == NONE_FRAME) {
- record_samples(mbmi, pts, pts_inref, pts_mv, global_offset_r,
- global_offset_c, row_offset, 1, 0, -1);
- pts += 2;
- pts_inref += 2;
- pts_mv += 2;
- np++;
- }
- break;
- }
+ if (mbmi->ref_frame[0] == ref_frame && mbmi->ref_frame[1] == NONE_FRAME) {
+ record_samples(mbmi, pts, pts_inref, pts_mv, global_offset_r,
+ global_offset_c, row_offset, 1, 0, -1);
+ pts += 2;
+ pts_inref += 2;
+ pts_mv += 2;
+ np++;
}
-
+ } else {
// Handle "current block height > above block height" case.
- if (!n) memset(marked, 0, bs * xd->n8_h * sizeof(*marked));
-
for (i = 0; i < AOMMIN(xd->n8_h, cm->mi_rows - mi_row); i += mi_step) {
int mi_row_offset = i;
- MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
- MB_MODE_INFO *mbmi = &mi->mbmi;
- uint8_t n8_w = mi_size_wide[mbmi->sb_type];
- uint8_t n8_h = mi_size_high[mbmi->sb_type];
-
+ mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+ mbmi = &mi->mbmi;
+ n8_h = mi_size_high[mbmi->sb_type];
mi_step = AOMMIN(xd->n8_h, n8_h);
- // Processed already
- if (marked[n * xd->n8_h + i]) continue;
-
- for (j = 0; j < AOMMIN(bs, n8_w); j++)
- for (k = 0; k < AOMMIN(xd->n8_h, n8_h); k++)
- marked[(n + j) * xd->n8_h + i + k] = 1;
-
if (mbmi->ref_frame[0] == ref_frame &&
mbmi->ref_frame[1] == NONE_FRAME) {
record_samples(mbmi, pts, pts_inref, pts_mv, global_offset_r,
- global_offset_c, i, 1, -n, -1);
+ global_offset_c, i, 1, 0, -1);
pts += 2;
pts_inref += 2;
pts_mv += 2;
diff --git a/av1/common/warped_motion.h b/av1/common/warped_motion.h
index 845eb9a..e05f6a8 100644
--- a/av1/common/warped_motion.h
+++ b/av1/common/warped_motion.h
@@ -30,10 +30,9 @@
#define LEAST_SQUARES_SAMPLES_MAX (1 << LEAST_SQUARES_SAMPLES_MAX_BITS)
#if WARPED_MOTION_SORT_SAMPLES
-// #define SAMPLES_ARRAY_SIZE (LEAST_SQUARES_SAMPLES_MAX * 2)
-// Search half bsize on the top and half bsize on the left, 1 upper-left block,
+// Search 1 row on the top and 1 column on the left, 1 upper-left block,
// 1 upper-right block.
-#define SAMPLES_ARRAY_SIZE ((MAX_MIB_SIZE * MAX_MIB_SIZE + 2) * 2)
+#define SAMPLES_ARRAY_SIZE ((MAX_MIB_SIZE * 2 + 2) * 2)
#else
#define SAMPLES_ARRAY_SIZE (LEAST_SQUARES_SAMPLES_MAX * 2)
#endif // WARPED_MOTION_SORT_SAMPLES
