encoder: Remove 64x upsampled reference buffers
They do not handle border extension correctly (interpolation and
border extension do not commute unless you upsample into the
border), nor do they handle crop dimensions that are not a multiple
of 8 (the upsampled version is not sufficiently large), in addition
to using massive amounts of memory and being a criminal waste of
cache (1 byte used for every 8 bytes fetched).
This commit reimplements use_upsampled_references by computing the
subpixel samples on the fly. This implementation not only corrects
the border handling, but is also faster, while maintaining the
same quality.
HL AWCY results are basically noise:
PSNR | PSNR HVS | SSIM | MS SSIM | CIEDE 2000
0.0188 | 0.0187 | 0.0045 | 0.0063 | 0.0228
Change-Id: I7527db9f83b87a7bb8b35342f7e6457cd0bef9cd
diff --git a/av1/encoder/mcomp.c b/av1/encoder/mcomp.c
index 79543c4..b49e080 100644
--- a/av1/encoder/mcomp.c
+++ b/av1/encoder/mcomp.c
@@ -228,49 +228,45 @@
#define CHECK_BETTER0(v, r, c) CHECK_BETTER(v, r, c)
-static INLINE const uint8_t *upre(const uint8_t *buf, int stride, int r,
- int c) {
- return &buf[(r)*stride + (c)];
-}
-
/* checks if (r, c) has better score than previous best */
#if CONFIG_EXT_INTER
-#define CHECK_BETTER1(v, r, c) \
- if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
- MV this_mv = { r, c }; \
- thismse = upsampled_pref_error( \
- xd, vfp, src_address, src_stride, upre(y, y_stride, r, c), y_stride, \
- second_pred, mask, mask_stride, invert_mask, w, h, &sse); \
- v = mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit); \
- v += thismse; \
- if (v < besterr) { \
- besterr = v; \
- br = r; \
- bc = c; \
- *distortion = thismse; \
- *sse1 = sse; \
- } \
- } else { \
- v = INT_MAX; \
+#define CHECK_BETTER1(v, r, c) \
+ if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
+ MV this_mv = { r, c }; \
+ thismse = upsampled_pref_error(xd, vfp, src_address, src_stride, \
+ pre(y, y_stride, r, c), y_stride, sp(c), \
+ sp(r), second_pred, mask, mask_stride, \
+ invert_mask, w, h, &sse); \
+ v = mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit); \
+ v += thismse; \
+ if (v < besterr) { \
+ besterr = v; \
+ br = r; \
+ bc = c; \
+ *distortion = thismse; \
+ *sse1 = sse; \
+ } \
+ } else { \
+ v = INT_MAX; \
}
#else
-#define CHECK_BETTER1(v, r, c) \
- if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
- MV this_mv = { r, c }; \
- thismse = upsampled_pref_error(xd, vfp, src_address, src_stride, \
- upre(y, y_stride, r, c), y_stride, \
- second_pred, w, h, &sse); \
- v = mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit); \
- v += thismse; \
- if (v < besterr) { \
- besterr = v; \
- br = r; \
- bc = c; \
- *distortion = thismse; \
- *sse1 = sse; \
- } \
- } else { \
- v = INT_MAX; \
+#define CHECK_BETTER1(v, r, c) \
+ if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
+ MV this_mv = { r, c }; \
+ thismse = upsampled_pref_error(xd, vfp, src_address, src_stride, \
+ pre(y, y_stride, r, c), y_stride, sp(c), \
+ sp(r), second_pred, w, h, &sse); \
+ v = mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit); \
+ v += thismse; \
+ if (v < besterr) { \
+ besterr = v; \
+ br = r; \
+ bc = c; \
+ *distortion = thismse; \
+ *sse1 = sse; \
+ } \
+ } else { \
+ v = INT_MAX; \
}
#endif // CONFIG_EXT_INTER
@@ -700,16 +696,14 @@
};
/* clang-format on */
-static int upsampled_pref_error(const MACROBLOCKD *xd,
- const aom_variance_fn_ptr_t *vfp,
- const uint8_t *const src, const int src_stride,
- const uint8_t *const y, int y_stride,
- const uint8_t *second_pred,
+static int upsampled_pref_error(
+ const MACROBLOCKD *xd, const aom_variance_fn_ptr_t *vfp,
+ const uint8_t *const src, const int src_stride, const uint8_t *const y,
+ int y_stride, int subpel_x_q3, int subpel_y_q3, const uint8_t *second_pred,
#if CONFIG_EXT_INTER
- const uint8_t *mask, int mask_stride,
- int invert_mask,
+ const uint8_t *mask, int mask_stride, int invert_mask,
#endif
- int w, int h, unsigned int *sse) {
+ int w, int h, unsigned int *sse) {
unsigned int besterr;
#if CONFIG_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
@@ -717,15 +711,17 @@
if (second_pred != NULL) {
#if CONFIG_EXT_INTER
if (mask)
- aom_highbd_comp_mask_upsampled_pred(pred16, second_pred, w, h, y,
- y_stride, mask, mask_stride,
- invert_mask);
+ aom_highbd_comp_mask_upsampled_pred(
+ pred16, second_pred, w, h, subpel_x_q3, subpel_y_q3, y, y_stride,
+ mask, mask_stride, invert_mask, xd->cur_buf->bit_depth);
else
#endif
- aom_highbd_comp_avg_upsampled_pred(pred16, second_pred, w, h, y,
- y_stride);
+ aom_highbd_comp_avg_upsampled_pred(pred16, second_pred, w, h,
+ subpel_x_q3, subpel_y_q3, y,
+ y_stride, xd->cur_buf->bit_depth);
} else {
- aom_highbd_upsampled_pred(pred16, w, h, y, y_stride);
+ aom_highbd_upsampled_pred(pred16, w, h, subpel_x_q3, subpel_y_q3, y,
+ y_stride, xd->cur_buf->bit_depth);
}
besterr = vfp->vf(CONVERT_TO_BYTEPTR(pred16), w, src, src_stride, sse);
@@ -738,13 +734,15 @@
if (second_pred != NULL) {
#if CONFIG_EXT_INTER
if (mask)
- aom_comp_mask_upsampled_pred(pred, second_pred, w, h, y, y_stride, mask,
+ aom_comp_mask_upsampled_pred(pred, second_pred, w, h, subpel_x_q3,
+ subpel_y_q3, y, y_stride, mask,
mask_stride, invert_mask);
else
#endif
- aom_comp_avg_upsampled_pred(pred, second_pred, w, h, y, y_stride);
+ aom_comp_avg_upsampled_pred(pred, second_pred, w, h, subpel_x_q3,
+ subpel_y_q3, y, y_stride);
} else {
- aom_upsampled_pred(pred, w, h, y, y_stride);
+ aom_upsampled_pred(pred, w, h, subpel_x_q3, subpel_y_q3, y, y_stride);
}
besterr = vfp->vf(pred, w, src, src_stride, sse);
@@ -764,12 +762,12 @@
#endif
int w, int h, int offset, int *mvjcost, int *mvcost[2], unsigned int *sse1,
int *distortion) {
- unsigned int besterr = upsampled_pref_error(xd, vfp, src, src_stride,
- y + offset, y_stride, second_pred,
+ unsigned int besterr = upsampled_pref_error(
+ xd, vfp, src, src_stride, y + offset, y_stride, 0, 0, second_pred,
#if CONFIG_EXT_INTER
- mask, mask_stride, invert_mask,
+ mask, mask_stride, invert_mask,
#endif
- w, h, sse1);
+ w, h, sse1);
*distortion = besterr;
besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
return besterr;
@@ -824,7 +822,7 @@
#if CONFIG_EXT_INTER
mask, mask_stride, invert_mask,
#endif
- w, h, (offset * 8), mvjcost, mvcost, sse1, distortion);
+ w, h, offset, mvjcost, mvcost, sse1, distortion);
else
besterr =
setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp, src_address,
@@ -845,17 +843,15 @@
MV this_mv = { tr, tc };
if (use_upsampled_ref) {
- const uint8_t *const pre_address = y + tr * y_stride + tc;
-
thismse = upsampled_pref_error(xd, vfp, src_address, src_stride,
- pre_address, y_stride, second_pred,
+ pre(y, y_stride, tr, tc), y_stride,
+ sp(tc), sp(tr), second_pred,
#if CONFIG_EXT_INTER
mask, mask_stride, invert_mask,
#endif
w, h, &sse);
} else {
- const uint8_t *const pre_address =
- y + (tr >> 3) * y_stride + (tc >> 3);
+ const uint8_t *const pre_address = pre(y, y_stride, tr, tc);
if (second_pred == NULL)
thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr),
src_address, src_stride, &sse);
@@ -894,16 +890,15 @@
MV this_mv = { tr, tc };
if (use_upsampled_ref) {
- const uint8_t *const pre_address = y + tr * y_stride + tc;
-
thismse = upsampled_pref_error(xd, vfp, src_address, src_stride,
- pre_address, y_stride, second_pred,
+ pre(y, y_stride, tr, tc), y_stride,
+ sp(tc), sp(tr), second_pred,
#if CONFIG_EXT_INTER
mask, mask_stride, invert_mask,
#endif
w, h, &sse);
} else {
- const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
+ const uint8_t *const pre_address = pre(y, y_stride, tr, tc);
if (second_pred == NULL)
thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr), src_address,
@@ -2653,19 +2648,20 @@
#define CHECK_BETTER0(v, r, c) CHECK_BETTER(v, r, c)
#undef CHECK_BETTER1
-#define CHECK_BETTER1(v, r, c) \
- if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
- thismse = upsampled_obmc_pref_error( \
- xd, mask, vfp, z, upre(y, y_stride, r, c), y_stride, w, h, &sse); \
- if ((v = MVC(r, c) + thismse) < besterr) { \
- besterr = v; \
- br = r; \
- bc = c; \
- *distortion = thismse; \
- *sse1 = sse; \
- } \
- } else { \
- v = INT_MAX; \
+#define CHECK_BETTER1(v, r, c) \
+ if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
+ thismse = \
+ upsampled_obmc_pref_error(xd, mask, vfp, z, pre(y, y_stride, r, c), \
+ y_stride, sp(c), sp(r), w, h, &sse); \
+ if ((v = MVC(r, c) + thismse) < besterr) { \
+ besterr = v; \
+ br = r; \
+ bc = c; \
+ *distortion = thismse; \
+ *sse1 = sse; \
+ } \
+ } else { \
+ v = INT_MAX; \
}
static unsigned int setup_obmc_center_error(
@@ -2684,12 +2680,14 @@
const aom_variance_fn_ptr_t *vfp,
const int32_t *const wsrc,
const uint8_t *const y, int y_stride,
- int w, int h, unsigned int *sse) {
+ int subpel_x_q3, int subpel_y_q3, int w,
+ int h, unsigned int *sse) {
unsigned int besterr;
#if CONFIG_HIGHBITDEPTH
if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
DECLARE_ALIGNED(16, uint16_t, pred16[MAX_SB_SQUARE]);
- aom_highbd_upsampled_pred(pred16, w, h, y, y_stride);
+ aom_highbd_upsampled_pred(pred16, w, h, subpel_x_q3, subpel_y_q3, y,
+ y_stride, xd->cur_buf->bit_depth);
besterr = vfp->ovf(CONVERT_TO_BYTEPTR(pred16), w, wsrc, mask, sse);
} else {
@@ -2698,7 +2696,7 @@
DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]);
(void)xd;
#endif // CONFIG_HIGHBITDEPTH
- aom_upsampled_pred(pred, w, h, y, y_stride);
+ aom_upsampled_pred(pred, w, h, subpel_x_q3, subpel_y_q3, y, y_stride);
besterr = vfp->ovf(pred, w, wsrc, mask, sse);
#if CONFIG_HIGHBITDEPTH
@@ -2714,18 +2712,17 @@
int h, int offset, int *mvjcost, int *mvcost[2], unsigned int *sse1,
int *distortion) {
unsigned int besterr = upsampled_obmc_pref_error(
- xd, mask, vfp, wsrc, y + offset, y_stride, w, h, sse1);
+ xd, mask, vfp, wsrc, y + offset, y_stride, 0, 0, w, h, sse1);
*distortion = besterr;
besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
return besterr;
}
int av1_find_best_obmc_sub_pixel_tree_up(
- const AV1_COMP *cpi, MACROBLOCK *x, int mi_row, int mi_col, MV *bestmv,
- const MV *ref_mv, int allow_hp, int error_per_bit,
- const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
- int *mvjcost, int *mvcost[2], int *distortion, unsigned int *sse1,
- int is_second, int use_upsampled_ref) {
+ MACROBLOCK *x, MV *bestmv, const MV *ref_mv, int allow_hp,
+ int error_per_bit, const aom_variance_fn_ptr_t *vfp, int forced_stop,
+ int iters_per_step, int *mvjcost, int *mvcost[2], int *distortion,
+ unsigned int *sse1, int is_second, int use_upsampled_ref) {
const int32_t *wsrc = x->wsrc_buf;
const int32_t *mask = x->mask_buf;
const int *const z = wsrc;
@@ -2756,27 +2753,11 @@
int y_stride;
const uint8_t *y;
- const struct buf_2d backup_pred = pd->pre[is_second];
int minc, maxc, minr, maxr;
av1_set_subpel_mv_search_range(&x->mv_limits, &minc, &maxc, &minr, &maxr,
ref_mv);
- if (use_upsampled_ref) {
-#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
- int ref = has_second_ref(&xd->mi[0]->mbmi)
- ? xd->mi[0]->mbmi.ref_frame[is_second]
- : xd->mi[0]->mbmi.ref_frame[0];
-#else // !(CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF)
- int ref = xd->mi[0]->mbmi.ref_frame[is_second];
-#endif // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
- const YV12_BUFFER_CONFIG *upsampled_ref = get_upsampled_ref(cpi, ref);
- setup_pred_plane(&pd->pre[is_second], mbmi->sb_type,
- upsampled_ref->y_buffer, upsampled_ref->y_crop_width,
- upsampled_ref->y_crop_height, upsampled_ref->y_stride,
- (mi_row << 3), (mi_col << 3), NULL, pd->subsampling_x,
- pd->subsampling_y);
- }
y = pd->pre[is_second].buf;
y_stride = pd->pre[is_second].stride;
offset = bestmv->row * y_stride + bestmv->col;
@@ -2790,7 +2771,7 @@
if (use_upsampled_ref)
besterr = upsampled_setup_obmc_center_error(
xd, mask, bestmv, ref_mv, error_per_bit, vfp, z, y, y_stride, w, h,
- (offset * 8), mvjcost, mvcost, sse1, distortion);
+ offset, mvjcost, mvcost, sse1, distortion);
else
besterr = setup_obmc_center_error(mask, bestmv, ref_mv, error_per_bit, vfp,
z, y, y_stride, offset, mvjcost, mvcost,
@@ -2803,15 +2784,13 @@
tc = bc + search_step[idx].col;
if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
MV this_mv = { tr, tc };
+ const uint8_t *const pre_address = pre(y, y_stride, tr, tc);
if (use_upsampled_ref) {
- const uint8_t *const pre_address = y + tr * y_stride + tc;
-
- thismse = upsampled_obmc_pref_error(
- xd, mask, vfp, src_address, pre_address, y_stride, w, h, &sse);
+ thismse =
+ upsampled_obmc_pref_error(xd, mask, vfp, src_address, pre_address,
+ y_stride, sp(tc), sp(tr), w, h, &sse);
} else {
- const uint8_t *const pre_address =
- y + (tr >> 3) * y_stride + (tc >> 3);
thismse = vfp->osvf(pre_address, y_stride, sp(tc), sp(tr),
src_address, mask, &sse);
}
@@ -2839,15 +2818,12 @@
MV this_mv = { tr, tc };
if (use_upsampled_ref) {
- const uint8_t *const pre_address = y + tr * y_stride + tc;
-
thismse = upsampled_obmc_pref_error(xd, mask, vfp, src_address,
- pre_address, y_stride, w, h, &sse);
+ pre(y, y_stride, tr, tc), y_stride,
+ sp(tc), sp(tr), w, h, &sse);
} else {
- const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
-
- thismse = vfp->osvf(pre_address, y_stride, sp(tc), sp(tr), src_address,
- mask, &sse);
+ thismse = vfp->osvf(pre(y, y_stride, tr, tc), y_stride, sp(tc), sp(tr),
+ src_address, mask, &sse);
}
cost_array[4] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
@@ -2895,10 +2871,6 @@
bestmv->row = br;
bestmv->col = bc;
- if (use_upsampled_ref) {
- pd->pre[is_second] = backup_pred;
- }
-
return besterr;
}
