Merge "Fix highbd obmc_variance unit test" into nextgenv2
diff --git a/vp10/common/reconinter.c b/vp10/common/reconinter.c
index 53fd1a6..d2fc980 100644
--- a/vp10/common/reconinter.c
+++ b/vp10/common/reconinter.c
@@ -1298,97 +1298,63 @@
#endif // CONFIG_SUPERTX
#if CONFIG_OBMC
-// obmc_mask_N[is_neighbor_predictor][overlap_position]
-static const uint8_t obmc_mask_1[2][1] = {
- { 55},
- { 9}
+// obmc_mask_N[overlap_position]
+static const uint8_t obmc_mask_1[1] = {
+ 55
};
-static const uint8_t obmc_mask_2[2][2] = {
- { 45, 62},
- { 19, 2}
+static const uint8_t obmc_mask_2[2] = {
+ 45, 62
};
-static const uint8_t obmc_mask_4[2][4] = {
- { 39, 50, 59, 64},
- { 25, 14, 5, 0}
+static const uint8_t obmc_mask_4[4] = {
+ 39, 50, 59, 64
};
-static const uint8_t obmc_mask_8[2][8] = {
- { 36, 42, 48, 53, 57, 61, 63, 64},
- { 28, 22, 16, 11, 7, 3, 1, 0}
+static const uint8_t obmc_mask_8[8] = {
+ 36, 42, 48, 53, 57, 61, 63, 64
};
-static const uint8_t obmc_mask_16[2][16] = {
- { 34, 37, 40, 43, 46, 49, 52, 54, 56, 58, 60, 61, 63, 64, 64, 64},
- { 30, 27, 24, 21, 18, 15, 12, 10, 8, 6, 4, 3, 1, 0, 0, 0}
+static const uint8_t obmc_mask_16[16] = {
+ 34, 37, 40, 43, 46, 49, 52, 54, 56, 58, 60, 61, 63, 64, 64, 64
};
-static const uint8_t obmc_mask_32[2][32] = {
- { 33, 35, 36, 38, 40, 41, 43, 44,
- 45, 47, 48, 50, 51, 52, 53, 55,
- 56, 57, 58, 59, 60, 60, 61, 62,
- 62, 63, 63, 64, 64, 64, 64, 64 },
- { 31, 29, 28, 26, 24, 23, 21, 20,
- 19, 17, 16, 14, 13, 12, 11, 9,
- 8, 7, 6, 5, 4, 4, 3, 2,
- 2, 1, 1, 0, 0, 0, 0, 0 }
+static const uint8_t obmc_mask_32[32] = {
+ 33, 35, 36, 38, 40, 41, 43, 44, 45, 47, 48, 50, 51, 52, 53, 55,
+ 56, 57, 58, 59, 60, 60, 61, 62, 62, 63, 63, 64, 64, 64, 64, 64
};
#if CONFIG_EXT_PARTITION
-static const uint8_t obmc_mask_64[2][64] = {
- {
- 33, 34, 35, 35, 36, 37, 38, 39, 40, 40, 41, 42, 43, 44, 44, 44,
- 45, 46, 47, 47, 48, 49, 50, 51, 51, 51, 52, 52, 53, 54, 55, 56,
- 56, 56, 57, 57, 58, 58, 59, 60, 60, 60, 60, 60, 61, 62, 62, 62,
- 62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
- }, {
- 31, 30, 29, 29, 28, 27, 26, 25, 24, 24, 23, 22, 21, 20, 20, 20,
- 19, 18, 17, 17, 16, 15, 14, 13, 13, 13, 12, 12, 11, 10, 9, 8,
- 8, 8, 7, 7, 6, 6, 5, 4, 4, 4, 4, 4, 3, 2, 2, 2,
- 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
- }
+static const uint8_t obmc_mask_64[64] = {
+ 33, 34, 35, 35, 36, 37, 38, 39, 40, 40, 41, 42, 43, 44, 44, 44,
+ 45, 46, 47, 47, 48, 49, 50, 51, 51, 51, 52, 52, 53, 54, 55, 56,
+ 56, 56, 57, 57, 58, 58, 59, 60, 60, 60, 60, 60, 61, 62, 62, 62,
+ 62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
};
#endif // CONFIG_EXT_PARTITION
-void setup_obmc_mask(int length, const uint8_t *mask[2]) {
+const uint8_t* vp10_get_obmc_mask(int length) {
switch (length) {
case 1:
- mask[0] = obmc_mask_1[0];
- mask[1] = obmc_mask_1[1];
- break;
+ return obmc_mask_1;
case 2:
- mask[0] = obmc_mask_2[0];
- mask[1] = obmc_mask_2[1];
- break;
+ return obmc_mask_2;
case 4:
- mask[0] = obmc_mask_4[0];
- mask[1] = obmc_mask_4[1];
- break;
+ return obmc_mask_4;
case 8:
- mask[0] = obmc_mask_8[0];
- mask[1] = obmc_mask_8[1];
- break;
+ return obmc_mask_8;
case 16:
- mask[0] = obmc_mask_16[0];
- mask[1] = obmc_mask_16[1];
- break;
+ return obmc_mask_16;
case 32:
- mask[0] = obmc_mask_32[0];
- mask[1] = obmc_mask_32[1];
- break;
+ return obmc_mask_32;
#if CONFIG_EXT_PARTITION
case 64:
- mask[0] = obmc_mask_64[0];
- mask[1] = obmc_mask_64[1];
- break;
+ return obmc_mask_64;
#endif // CONFIG_EXT_PARTITION
default:
- mask[0] = NULL;
- mask[1] = NULL;
assert(0);
- break;
+ return NULL;
}
}
@@ -1398,168 +1364,101 @@
// xd->plane[].dst.buf
void vp10_build_obmc_inter_prediction(VP10_COMMON *cm,
MACROBLOCKD *xd, int mi_row, int mi_col,
- int use_tmp_dst_buf,
- uint8_t *final_buf[MAX_MB_PLANE],
- int final_stride[MAX_MB_PLANE],
- uint8_t *tmp_buf1[MAX_MB_PLANE],
- int tmp_stride1[MAX_MB_PLANE],
- uint8_t *tmp_buf2[MAX_MB_PLANE],
- int tmp_stride2[MAX_MB_PLANE]) {
- const TileInfo *const tile = &xd->tile;
- BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
- int plane, i, mi_step;
- int above_available = mi_row > tile->mi_row_start;
+ uint8_t *above[MAX_MB_PLANE],
+ int above_stride[MAX_MB_PLANE],
+ uint8_t *left[MAX_MB_PLANE],
+ int left_stride[MAX_MB_PLANE]) {
+ const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+ int plane, i;
#if CONFIG_VP9_HIGHBITDEPTH
- int is_hbd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+ const int is_hbd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
#endif // CONFIG_VP9_HIGHBITDEPTH
- if (use_tmp_dst_buf) {
- for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
- const struct macroblockd_plane *pd = &xd->plane[plane];
- int bw = (xd->n8_w * 8) >> pd->subsampling_x;
- int bh = (xd->n8_h * 8) >> pd->subsampling_y;
- int row;
-#if CONFIG_VP9_HIGHBITDEPTH
- if (is_hbd) {
- uint16_t *final_buf16 = CONVERT_TO_SHORTPTR(final_buf[plane]);
- uint16_t *bmc_buf16 = CONVERT_TO_SHORTPTR(pd->dst.buf);
- for (row = 0; row < bh; ++row)
- memcpy(final_buf16 + row * final_stride[plane],
- bmc_buf16 + row * pd->dst.stride, bw * sizeof(uint16_t));
- } else {
-#endif
- for (row = 0; row < bh; ++row)
- memcpy(final_buf[plane] + row * final_stride[plane],
- pd->dst.buf + row * pd->dst.stride, bw);
-#if CONFIG_VP9_HIGHBITDEPTH
- }
-#endif // CONFIG_VP9_HIGHBITDEPTH
- }
- }
-
// handle above row
- for (i = 0; above_available && i < VPXMIN(xd->n8_w, cm->mi_cols - mi_col);
- i += mi_step) {
- int mi_row_offset = -1;
- int mi_col_offset = i;
- int overlap;
- MODE_INFO *above_mi = xd->mi[mi_col_offset +
- mi_row_offset * xd->mi_stride];
- MB_MODE_INFO *above_mbmi = &above_mi->mbmi;
+ if (xd->up_available) {
+ const int overlap = num_4x4_blocks_high_lookup[bsize] * 2;
+ const int miw = VPXMIN(xd->n8_w, cm->mi_cols - mi_col);
+ const int mi_row_offset = -1;
- mi_step = VPXMIN(xd->n8_w,
- num_8x8_blocks_wide_lookup[above_mbmi->sb_type]);
+ assert(miw > 0);
- if (!is_neighbor_overlappable(above_mbmi))
- continue;
+ i = 0;
+ do { // for each mi in the above row
+ const int mi_col_offset = i;
+ const MB_MODE_INFO *const above_mbmi =
+ &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+ const int mi_step =
+ VPXMIN(xd->n8_w, num_8x8_blocks_wide_lookup[above_mbmi->sb_type]);
- overlap = num_4x4_blocks_high_lookup[bsize] << 1;
-
- for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
- const struct macroblockd_plane *pd = &xd->plane[plane];
- int bw = (mi_step * MI_SIZE) >> pd->subsampling_x;
- int bh = overlap >> pd->subsampling_y;
- int row, col;
- int dst_stride = use_tmp_dst_buf ? final_stride[plane] : pd->dst.stride;
- uint8_t *dst = use_tmp_dst_buf ?
- &final_buf[plane][(i * MI_SIZE) >> pd->subsampling_x] :
- &pd->dst.buf[(i * MI_SIZE) >> pd->subsampling_x];
- int tmp_stride = tmp_stride1[plane];
- uint8_t *tmp = &tmp_buf1[plane][(i * MI_SIZE) >> pd->subsampling_x];
- const uint8_t *mask[2];
-
- setup_obmc_mask(bh, mask);
+ if (is_neighbor_overlappable(above_mbmi)) {
+ for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+ const struct macroblockd_plane *pd = &xd->plane[plane];
+ const int bw = (mi_step * MI_SIZE) >> pd->subsampling_x;
+ const int bh = overlap >> pd->subsampling_y;
+ const int dst_stride = pd->dst.stride;
+ uint8_t *const dst =
+ &pd->dst.buf[(i * MI_SIZE) >> pd->subsampling_x];
+ const int tmp_stride = above_stride[plane];
+ const uint8_t *const tmp =
+ &above[plane][(i * MI_SIZE) >> pd->subsampling_x];
+ const uint8_t *const mask = vp10_get_obmc_mask(bh);
#if CONFIG_VP9_HIGHBITDEPTH
- if (is_hbd) {
- uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
- uint16_t *tmp16 = CONVERT_TO_SHORTPTR(tmp);
-
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col)
- dst16[col] = ROUND_POWER_OF_TWO(mask[0][row] * dst16[col] +
- mask[1][row] * tmp16[col], 6);
-
- dst16 += dst_stride;
- tmp16 += tmp_stride;
+ if (is_hbd)
+ vpx_highbd_blend_a64_vmask(dst, dst_stride, dst, dst_stride,
+ tmp, tmp_stride, mask, bh, bw, xd->bd);
+ else
+#endif // CONFIG_VP9_HIGHBITDEPTH
+ vpx_blend_a64_vmask(dst, dst_stride, dst, dst_stride,
+ tmp, tmp_stride, mask, bh, bw);
}
- } else {
-#endif // CONFIG_VP9_HIGHBITDEPTH
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col)
- dst[col] = ROUND_POWER_OF_TWO(mask[0][row] * dst[col] +
- mask[1][row] * tmp[col], 6);
- dst += dst_stride;
- tmp += tmp_stride;
}
-#if CONFIG_VP9_HIGHBITDEPTH
- }
-#endif // CONFIG_VP9_HIGHBITDEPTH
- }
- } // each mi in the above row
+ i += mi_step;
+ } while (i < miw);
+ }
- if (mi_col == 0 || (mi_col - 1 < tile->mi_col_start))
- return;
// handle left column
- for (i = 0; i < VPXMIN(xd->n8_h, cm->mi_rows - mi_row);
- i += mi_step) {
- int mi_row_offset = i;
- int mi_col_offset = -1;
- int overlap;
- MODE_INFO *left_mi = xd->mi[mi_col_offset +
- mi_row_offset * xd->mi_stride];
- MB_MODE_INFO *left_mbmi = &left_mi->mbmi;
+ if (xd->left_available) {
+ const int overlap = num_4x4_blocks_wide_lookup[bsize] * 2;
+ const int mih = VPXMIN(xd->n8_h, cm->mi_rows - mi_row);
+ const int mi_col_offset = -1;
- mi_step = VPXMIN(xd->n8_h,
- num_8x8_blocks_high_lookup[left_mbmi->sb_type]);
+ assert(mih > 0);
- if (!is_neighbor_overlappable(left_mbmi))
- continue;
+ i = 0;
+ do { // for each mi in the left column
+ const int mi_row_offset = i;
+ const MB_MODE_INFO *const left_mbmi =
+ &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+ const int mi_step =
+ VPXMIN(xd->n8_h, num_8x8_blocks_high_lookup[left_mbmi->sb_type]);
- overlap = num_4x4_blocks_wide_lookup[bsize] << 1;
-
- for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
- const struct macroblockd_plane *pd = &xd->plane[plane];
- int bw = overlap >> pd->subsampling_x;
- int bh = (mi_step * MI_SIZE) >> pd->subsampling_y;
- int row, col;
- int dst_stride = use_tmp_dst_buf ? final_stride[plane] : pd->dst.stride;
- uint8_t *dst = use_tmp_dst_buf ?
- &final_buf[plane][(i * MI_SIZE * dst_stride) >> pd->subsampling_y] :
- &pd->dst.buf[(i * MI_SIZE * dst_stride) >> pd->subsampling_y];
- int tmp_stride = tmp_stride2[plane];
- uint8_t *tmp = &tmp_buf2[plane]
- [(i * MI_SIZE * tmp_stride) >> pd->subsampling_y];
- const uint8_t *mask[2];
-
- setup_obmc_mask(bw, mask);
+ if (is_neighbor_overlappable(left_mbmi)) {
+ for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+ const struct macroblockd_plane *pd = &xd->plane[plane];
+ const int bw = overlap >> pd->subsampling_x;
+ const int bh = (mi_step * MI_SIZE) >> pd->subsampling_y;
+ const int dst_stride = pd->dst.stride;
+ uint8_t *const dst =
+ &pd->dst.buf[(i * MI_SIZE * dst_stride) >> pd->subsampling_y];
+ const int tmp_stride = left_stride[plane];
+ const uint8_t *const tmp =
+ &left[plane][(i * MI_SIZE * tmp_stride) >> pd->subsampling_y];
+ const uint8_t *const mask = vp10_get_obmc_mask(bw);
#if CONFIG_VP9_HIGHBITDEPTH
- if (is_hbd) {
- uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
- uint16_t *tmp16 = CONVERT_TO_SHORTPTR(tmp);
-
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col)
- dst16[col] = ROUND_POWER_OF_TWO(mask[0][col] * dst16[col] +
- mask[1][col] * tmp16[col], 6);
- dst16 += dst_stride;
- tmp16 += tmp_stride;
+ if (is_hbd)
+ vpx_highbd_blend_a64_hmask(dst, dst_stride, dst, dst_stride,
+ tmp, tmp_stride, mask, bh, bw, xd->bd);
+ else
+#endif // CONFIG_VP9_HIGHBITDEPTH
+ vpx_blend_a64_hmask(dst, dst_stride, dst, dst_stride,
+ tmp, tmp_stride, mask, bh, bw);
}
- } else {
-#endif // CONFIG_VP9_HIGHBITDEPTH
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col)
- dst[col] = ROUND_POWER_OF_TWO(mask[0][col] * dst[col] +
- mask[1][col] * tmp[col], 6);
- dst += dst_stride;
- tmp += tmp_stride;
}
-#if CONFIG_VP9_HIGHBITDEPTH
- }
-#endif // CONFIG_VP9_HIGHBITDEPTH
- }
- } // each mi in the left column
+ i += mi_step;
+ } while (i < mih);
+ }
}
#if CONFIG_EXT_INTER
diff --git a/vp10/common/reconinter.h b/vp10/common/reconinter.h
index ac4a004..c32596e 100644
--- a/vp10/common/reconinter.h
+++ b/vp10/common/reconinter.h
@@ -562,16 +562,13 @@
#endif // CONFIG_EXT_INTERP
#if CONFIG_OBMC
-void setup_obmc_mask(int length, const uint8_t *mask[2]);
+const uint8_t* vp10_get_obmc_mask(int length);
void vp10_build_obmc_inter_prediction(VP10_COMMON *cm,
MACROBLOCKD *xd, int mi_row, int mi_col,
- int use_tmp_dst_buf,
- uint8_t *final_buf[MAX_MB_PLANE],
- int final_stride[MAX_MB_PLANE],
- uint8_t *tmp_buf1[MAX_MB_PLANE],
- int tmp_stride1[MAX_MB_PLANE],
- uint8_t *tmp_buf2[MAX_MB_PLANE],
- int tmp_stride2[MAX_MB_PLANE]);
+ uint8_t *above[MAX_MB_PLANE],
+ int above_stride[MAX_MB_PLANE],
+ uint8_t *left[MAX_MB_PLANE],
+ int left_stride[MAX_MB_PLANE]);
void vp10_build_prediction_by_above_preds(VP10_COMMON *cm,
MACROBLOCKD *xd,
int mi_row, int mi_col,
diff --git a/vp10/decoder/decodeframe.c b/vp10/decoder/decodeframe.c
index c4487a9..2f341b5 100644
--- a/vp10/decoder/decodeframe.c
+++ b/vp10/decoder/decodeframe.c
@@ -1385,7 +1385,7 @@
dst_buf2, dst_stride2);
vp10_setup_dst_planes(xd->plane, get_frame_new_buffer(cm),
mi_row, mi_col);
- vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col, 0, NULL, NULL,
+ vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col,
dst_buf1, dst_stride1,
dst_buf2, dst_stride2);
}
diff --git a/vp10/encoder/block.h b/vp10/encoder/block.h
index 88f1236..4c9f8a5 100644
--- a/vp10/encoder/block.h
+++ b/vp10/encoder/block.h
@@ -165,6 +165,9 @@
// Used to store sub partition's choices.
MV pred_mv[MAX_REF_FRAMES];
+ // Store the best motion vector during motion search
+ int_mv best_mv;
+
// Strong color activity detection. Used in RTC coding mode to enhance
// the visual quality at the boundary of moving color objects.
uint8_t color_sensitivity[2];
diff --git a/vp10/encoder/encodeframe.c b/vp10/encoder/encodeframe.c
index 69f766a..ff1ee6b 100644
--- a/vp10/encoder/encodeframe.c
+++ b/vp10/encoder/encodeframe.c
@@ -5091,7 +5091,7 @@
dst_stride2);
vp10_setup_dst_planes(xd->plane, get_frame_new_buffer(cm),
mi_row, mi_col);
- vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col, 0, NULL, NULL,
+ vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col,
dst_buf1, dst_stride1,
dst_buf2, dst_stride2);
}
diff --git a/vp10/encoder/encoder.h b/vp10/encoder/encoder.h
index 519cd03..608cace 100644
--- a/vp10/encoder/encoder.h
+++ b/vp10/encoder/encoder.h
@@ -447,7 +447,7 @@
ActiveMap active_map;
fractional_mv_step_fp *find_fractional_mv_step;
- vp10_full_search_fn_t full_search_sad;
+ vp10_full_search_fn_t full_search_sad; // It is currently unused.
vp10_diamond_search_fn_t diamond_search_sad;
vp10_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
uint64_t time_receive_data;
diff --git a/vp10/encoder/mbgraph.c b/vp10/encoder/mbgraph.c
index 5e66ce5..46cff80 100644
--- a/vp10/encoder/mbgraph.c
+++ b/vp10/encoder/mbgraph.c
@@ -25,7 +25,6 @@
static unsigned int do_16x16_motion_iteration(VP10_COMP *cpi,
const MV *ref_mv,
- MV *dst_mv,
int mb_row,
int mb_col) {
MACROBLOCK *const x = &cpi->td.mb;
@@ -51,8 +50,7 @@
/*cpi->sf.search_method == HEX*/
vp10_hex_search(x, &ref_full, step_param, x->errorperbit, 0,
- cond_cost_list(cpi, cost_list),
- &v_fn_ptr, 0, ref_mv, dst_mv);
+ cond_cost_list(cpi, cost_list), &v_fn_ptr, 0, ref_mv);
// Try sub-pixel MC
// if (bestsme > error_thresh && bestsme < INT_MAX)
@@ -60,7 +58,7 @@
int distortion;
unsigned int sse;
cpi->find_fractional_mv_step(
- x, dst_mv, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
+ x, ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit,
&v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
NULL, NULL,
@@ -74,7 +72,7 @@
#endif // CONFIG_EXT_INTER
xd->mi[0]->mbmi.mode = NEWMV;
- xd->mi[0]->mbmi.mv[0].as_mv = *dst_mv;
+ xd->mi[0]->mbmi.mv[0] = x->best_mv;
#if CONFIG_EXT_INTER
xd->mi[0]->mbmi.ref_frame[1] = NONE;
#endif // CONFIG_EXT_INTER
@@ -92,40 +90,40 @@
}
static int do_16x16_motion_search(VP10_COMP *cpi, const MV *ref_mv,
- int_mv *dst_mv, int mb_row, int mb_col) {
+ int mb_row, int mb_col) {
MACROBLOCK *const x = &cpi->td.mb;
MACROBLOCKD *const xd = &x->e_mbd;
unsigned int err, tmp_err;
- MV tmp_mv;
+ MV best_mv;
// Try zero MV first
// FIXME should really use something like near/nearest MV and/or MV prediction
err = vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
- dst_mv->as_int = 0;
+ best_mv.col = best_mv.row = 0;
// Test last reference frame using the previous best mv as the
// starting point (best reference) for the search
- tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
+ tmp_err = do_16x16_motion_iteration(cpi, ref_mv, mb_row, mb_col);
if (tmp_err < err) {
err = tmp_err;
- dst_mv->as_mv = tmp_mv;
+ best_mv = x->best_mv.as_mv;
}
// If the current best reference mv is not centered on 0,0 then do a 0,0
// based search as well.
if (ref_mv->row != 0 || ref_mv->col != 0) {
unsigned int tmp_err;
- MV zero_ref_mv = {0, 0}, tmp_mv;
+ MV zero_ref_mv = {0, 0};
- tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv,
- mb_row, mb_col);
+ tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, mb_row, mb_col);
if (tmp_err < err) {
- dst_mv->as_mv = tmp_mv;
err = tmp_err;
+ best_mv = x->best_mv.as_mv;
}
}
+ x->best_mv.as_mv = best_mv;
return err;
}
@@ -213,8 +211,8 @@
xd->plane[0].pre[0].stride = golden_ref->y_stride;
g_motion_error = do_16x16_motion_search(cpi,
prev_golden_ref_mv,
- &stats->ref[GOLDEN_FRAME].m.mv,
mb_row, mb_col);
+ stats->ref[GOLDEN_FRAME].m.mv = x->best_mv;
stats->ref[GOLDEN_FRAME].err = g_motion_error;
} else {
stats->ref[GOLDEN_FRAME].err = INT_MAX;
diff --git a/vp10/encoder/mcomp.c b/vp10/encoder/mcomp.c
index e7bd700..afbf3e9 100644
--- a/vp10/encoder/mcomp.c
+++ b/vp10/encoder/mcomp.c
@@ -315,6 +315,7 @@
unsigned int sse; \
unsigned int whichdir; \
int thismse; \
+ MV *bestmv = &x->best_mv.as_mv; \
const unsigned int halfiters = iters_per_step; \
const unsigned int quarteriters = iters_per_step; \
const unsigned int eighthiters = iters_per_step; \
@@ -411,8 +412,8 @@
}
int vp10_find_best_sub_pixel_tree_pruned_evenmore(
- const MACROBLOCK *x,
- MV *bestmv, const MV *ref_mv,
+ MACROBLOCK *x,
+ const MV *ref_mv,
int allow_hp,
int error_per_bit,
const vp10_variance_fn_ptr_t *vfp,
@@ -491,20 +492,20 @@
return besterr;
}
-int vp10_find_best_sub_pixel_tree_pruned_more(const MACROBLOCK *x,
- MV *bestmv, const MV *ref_mv,
- int allow_hp,
- int error_per_bit,
- const vp10_variance_fn_ptr_t *vfp,
- int forced_stop,
- int iters_per_step,
- int *cost_list,
- int *mvjcost, int *mvcost[2],
- int *distortion,
- unsigned int *sse1,
- const uint8_t *second_pred,
- int w, int h,
- int use_upsampled_ref) {
+int vp10_find_best_sub_pixel_tree_pruned_more(MACROBLOCK *x,
+ const MV *ref_mv,
+ int allow_hp,
+ int error_per_bit,
+ const vp10_variance_fn_ptr_t *vfp,
+ int forced_stop,
+ int iters_per_step,
+ int *cost_list,
+ int *mvjcost, int *mvcost[2],
+ int *distortion,
+ unsigned int *sse1,
+ const uint8_t *second_pred,
+ int w, int h,
+ int use_upsampled_ref) {
SETUP_SUBPEL_SEARCH;
(void) use_upsampled_ref;
@@ -568,19 +569,19 @@
return besterr;
}
-int vp10_find_best_sub_pixel_tree_pruned(const MACROBLOCK *x,
- MV *bestmv, const MV *ref_mv,
- int allow_hp,
- int error_per_bit,
- const vp10_variance_fn_ptr_t *vfp,
- int forced_stop,
- int iters_per_step,
- int *cost_list,
- int *mvjcost, int *mvcost[2],
- int *distortion,
- unsigned int *sse1,
- const uint8_t *second_pred,
- int w, int h, int use_upsampled_ref) {
+int vp10_find_best_sub_pixel_tree_pruned(MACROBLOCK *x,
+ const MV *ref_mv,
+ int allow_hp,
+ int error_per_bit,
+ const vp10_variance_fn_ptr_t *vfp,
+ int forced_stop,
+ int iters_per_step,
+ int *cost_list,
+ int *mvjcost, int *mvcost[2],
+ int *distortion,
+ unsigned int *sse1,
+ const uint8_t *second_pred,
+ int w, int h, int use_upsampled_ref) {
SETUP_SUBPEL_SEARCH;
(void) use_upsampled_ref;
@@ -724,19 +725,19 @@
return besterr;
}
-int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
- MV *bestmv, const MV *ref_mv,
- int allow_hp,
- int error_per_bit,
- const vp10_variance_fn_ptr_t *vfp,
- int forced_stop,
- int iters_per_step,
- int *cost_list,
- int *mvjcost, int *mvcost[2],
- int *distortion,
- unsigned int *sse1,
- const uint8_t *second_pred,
- int w, int h, int use_upsampled_ref) {
+int vp10_find_best_sub_pixel_tree(MACROBLOCK *x,
+ const MV *ref_mv,
+ int allow_hp,
+ int error_per_bit,
+ const vp10_variance_fn_ptr_t *vfp,
+ int forced_stop,
+ int iters_per_step,
+ int *cost_list,
+ int *mvjcost, int *mvcost[2],
+ int *distortion,
+ unsigned int *sse1,
+ const uint8_t *second_pred,
+ int w, int h, int use_upsampled_ref) {
const uint8_t *const src_address = x->plane[0].src.buf;
const int src_stride = x->plane[0].src.stride;
const MACROBLOCKD *xd = &x->e_mbd;
@@ -744,6 +745,7 @@
unsigned int sse;
unsigned int thismse;
const int y_stride = xd->plane[0].pre[0].stride;
+ MV *bestmv = &x->best_mv.as_mv;
const int offset = bestmv->row * y_stride + bestmv->col;
const uint8_t *const y = xd->plane[0].pre[0].buf;
@@ -1037,19 +1039,18 @@
// candidates as indicated in the num_candidates and candidates arrays
// passed into this function
//
-static int vp10_pattern_search(const MACROBLOCK *x,
- MV *start_mv,
- int search_param,
- int sad_per_bit,
- int do_init_search,
- int *cost_list,
- const vp10_variance_fn_ptr_t *vfp,
- int use_mvcost,
- const MV *center_mv,
- MV *best_mv,
- const int num_candidates[MAX_PATTERN_SCALES],
- const MV candidates[MAX_PATTERN_SCALES]
- [MAX_PATTERN_CANDIDATES]) {
+static int pattern_search(MACROBLOCK *x,
+ MV *start_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search,
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *vfp,
+ int use_mvcost,
+ const MV *center_mv,
+ const int num_candidates[MAX_PATTERN_SCALES],
+ const MV candidates[MAX_PATTERN_SCALES]
+ [MAX_PATTERN_CANDIDATES]) {
const MACROBLOCKD *const xd = &x->e_mbd;
static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
@@ -1294,8 +1295,8 @@
calc_int_cost_list(x, center_mv, sad_per_bit, vfp, &best_mv, cost_list);
}
}
- best_mv->row = br;
- best_mv->col = bc;
+ x->best_mv.as_mv.row = br;
+ x->best_mv.as_mv.col = bc;
return bestsad;
}
@@ -1332,15 +1333,15 @@
x->mvcost, x->errorperbit) : 0);
}
-int vp10_hex_search(const MACROBLOCK *x,
- MV *start_mv,
- int search_param,
- int sad_per_bit,
- int do_init_search,
- int *cost_list,
- const vp10_variance_fn_ptr_t *vfp,
- int use_mvcost,
- const MV *center_mv, MV *best_mv) {
+int vp10_hex_search(MACROBLOCK *x,
+ MV *start_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search,
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *vfp,
+ int use_mvcost,
+ const MV *center_mv) {
// First scale has 8-closest points, the rest have 6 points in hex shape
// at increasing scales
static const int hex_num_candidates[MAX_PATTERN_SCALES] = {
@@ -1361,22 +1362,20 @@
{{-512, -1024}, {512, -1024}, {1024, 0}, {512, 1024}, { -512, 1024},
{ -1024, 0}},
};
- return vp10_pattern_search(x, start_mv, search_param, sad_per_bit,
- do_init_search, cost_list, vfp, use_mvcost,
- center_mv, best_mv, hex_num_candidates,
- hex_candidates);
+ return pattern_search(x, start_mv, search_param, sad_per_bit,
+ do_init_search, cost_list, vfp, use_mvcost,
+ center_mv, hex_num_candidates, hex_candidates);
}
-int vp10_bigdia_search(const MACROBLOCK *x,
- MV *start_mv,
- int search_param,
- int sad_per_bit,
- int do_init_search,
- int *cost_list,
- const vp10_variance_fn_ptr_t *vfp,
- int use_mvcost,
- const MV *center_mv,
- MV *best_mv) {
+static int bigdia_search(MACROBLOCK *x,
+ MV *start_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search,
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *vfp,
+ int use_mvcost,
+ const MV *center_mv) {
// First scale has 4-closest points, the rest have 8 points in diamond
// shape at increasing scales
static const int bigdia_num_candidates[MAX_PATTERN_SCALES] = {
@@ -1403,25 +1402,23 @@
{{-512, -512}, {0, -1024}, {512, -512}, {1024, 0}, {512, 512}, {0, 1024},
{-512, 512}, {-1024, 0}},
};
- return vp10_pattern_search(x, start_mv, search_param, sad_per_bit,
- do_init_search, cost_list, vfp, use_mvcost,
- center_mv, best_mv, bigdia_num_candidates,
- bigdia_candidates);
+ return pattern_search(x, start_mv, search_param, sad_per_bit,
+ do_init_search, cost_list, vfp, use_mvcost,
+ center_mv, bigdia_num_candidates, bigdia_candidates);
}
-int vp10_square_search(const MACROBLOCK *x,
- MV *start_mv,
- int search_param,
- int sad_per_bit,
- int do_init_search,
- int *cost_list,
- const vp10_variance_fn_ptr_t *vfp,
- int use_mvcost,
- const MV *center_mv,
- MV *best_mv) {
+static int square_search(MACROBLOCK *x,
+ MV *start_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search,
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *vfp,
+ int use_mvcost,
+ const MV *center_mv) {
// All scales have 8 closest points in square shape
static const int square_num_candidates[MAX_PATTERN_SCALES] = {
- 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+ 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
};
// Note that the largest candidate step at each scale is 2^scale
static const MV square_candidates[MAX_PATTERN_SCALES]
@@ -1445,40 +1442,37 @@
{{-1024, -1024}, {0, -1024}, {1024, -1024}, {1024, 0}, {1024, 1024},
{0, 1024}, {-1024, 1024}, {-1024, 0}},
};
- return vp10_pattern_search(x, start_mv, search_param, sad_per_bit,
- do_init_search, cost_list, vfp, use_mvcost,
- center_mv, best_mv, square_num_candidates,
- square_candidates);
+ return pattern_search(x, start_mv, search_param, sad_per_bit,
+ do_init_search, cost_list, vfp, use_mvcost,
+ center_mv, square_num_candidates, square_candidates);
}
-int vp10_fast_hex_search(const MACROBLOCK *x,
- MV *ref_mv,
- int search_param,
- int sad_per_bit,
- int do_init_search, // must be zero for fast_hex
- int *cost_list,
- const vp10_variance_fn_ptr_t *vfp,
- int use_mvcost,
- const MV *center_mv,
- MV *best_mv) {
+static int fast_hex_search(MACROBLOCK *x,
+ MV *ref_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search, // must be zero for fast_hex
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *vfp,
+ int use_mvcost,
+ const MV *center_mv) {
return vp10_hex_search(
x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param), sad_per_bit,
- do_init_search, cost_list, vfp, use_mvcost, center_mv, best_mv);
+ do_init_search, cost_list, vfp, use_mvcost, center_mv);
}
-int vp10_fast_dia_search(const MACROBLOCK *x,
- MV *ref_mv,
- int search_param,
- int sad_per_bit,
- int do_init_search,
- int *cost_list,
- const vp10_variance_fn_ptr_t *vfp,
- int use_mvcost,
- const MV *center_mv,
- MV *best_mv) {
- return vp10_bigdia_search(
+static int fast_dia_search(MACROBLOCK *x,
+ MV *ref_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search,
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *vfp,
+ int use_mvcost,
+ const MV *center_mv) {
+ return bigdia_search(
x, ref_mv, VPXMAX(MAX_MVSEARCH_STEPS - 2, search_param), sad_per_bit,
- do_init_search, cost_list, vfp, use_mvcost, center_mv, best_mv);
+ do_init_search, cost_list, vfp, use_mvcost, center_mv);
}
#undef CHECK_BETTER
@@ -1924,12 +1918,12 @@
/* do_refine: If last step (1-away) of n-step search doesn't pick the center
point as the best match, we will do a final 1-away diamond
refining search */
-int vp10_full_pixel_diamond(const VP10_COMP *cpi, MACROBLOCK *x,
- MV *mvp_full, int step_param,
- int sadpb, int further_steps, int do_refine,
- int *cost_list,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *ref_mv, MV *dst_mv) {
+static int full_pixel_diamond(VP10_COMP *cpi, MACROBLOCK *x,
+ MV *mvp_full, int step_param,
+ int sadpb, int further_steps, int do_refine,
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *fn_ptr,
+ const MV *ref_mv) {
MV temp_mv;
int thissme, n, num00 = 0;
int bestsme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
@@ -1937,7 +1931,7 @@
fn_ptr, ref_mv);
if (bestsme < INT_MAX)
bestsme = vp10_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
- *dst_mv = temp_mv;
+ x->best_mv.as_mv = temp_mv;
// If there won't be more n-step search, check to see if refining search is
// needed.
@@ -1962,7 +1956,7 @@
if (thissme < bestsme) {
bestsme = thissme;
- *dst_mv = temp_mv;
+ x->best_mv.as_mv = temp_mv;
}
}
}
@@ -1970,20 +1964,20 @@
// final 1-away diamond refining search
if (do_refine) {
const int search_range = 8;
- MV best_mv = *dst_mv;
+ MV best_mv = x->best_mv.as_mv;
thissme = vp10_refining_search_sad(x, &best_mv, sadpb, search_range,
fn_ptr, ref_mv);
if (thissme < INT_MAX)
thissme = vp10_get_mvpred_var(x, &best_mv, ref_mv, fn_ptr, 1);
if (thissme < bestsme) {
bestsme = thissme;
- *dst_mv = best_mv;
+ x->best_mv.as_mv = best_mv;
}
}
// Return cost list.
if (cost_list) {
- calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
+ calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, &x->best_mv.as_mv, cost_list);
}
return bestsme;
}
@@ -1994,7 +1988,8 @@
// Runs an limited range exhaustive mesh search using a pattern set
// according to the encode speed profile.
static int full_pixel_exhaustive(VP10_COMP *cpi, MACROBLOCK *x,
- MV *centre_mv_full, int sadpb, int *cost_list,
+ const MV *centre_mv_full, int sadpb,
+ int *cost_list,
const vp10_variance_fn_ptr_t *fn_ptr,
const MV *ref_mv, MV *dst_mv) {
const SPEED_FEATURES *const sf = &cpi->sf;
@@ -2053,9 +2048,9 @@
}
int vp10_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
- int sad_per_bit, int distance,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv, MV *best_mv) {
+ int sad_per_bit, int distance,
+ const vp10_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv, MV *best_mv) {
int r, c;
const MACROBLOCKD *const xd = &x->e_mbd;
const struct buf_2d *const what = &x->plane[0].src;
@@ -2086,9 +2081,9 @@
}
int vp10_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv,
- int sad_per_bit, int distance,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv, MV *best_mv) {
+ int sad_per_bit, int distance,
+ const vp10_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv, MV *best_mv) {
int r;
const MACROBLOCKD *const xd = &x->e_mbd;
const struct buf_2d *const what = &x->plane[0].src;
@@ -2151,9 +2146,9 @@
}
int vp10_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv,
- int sad_per_bit, int distance,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv, MV *best_mv) {
+ int sad_per_bit, int distance,
+ const vp10_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv, MV *best_mv) {
int r;
const MACROBLOCKD *const xd = &x->e_mbd;
const struct buf_2d *const what = &x->plane[0].src;
@@ -2318,29 +2313,30 @@
// This function is called when we do joint motion search in comp_inter_inter
// mode.
-int vp10_refining_search_8p_c(const MACROBLOCK *x,
- MV *ref_mv, int error_per_bit,
- int search_range,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv,
- const uint8_t *second_pred) {
+int vp10_refining_search_8p_c(MACROBLOCK *x,
+ int error_per_bit,
+ int search_range,
+ const vp10_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv,
+ const uint8_t *second_pred) {
const MV neighbors[8] = {{-1, 0}, {0, -1}, {0, 1}, {1, 0},
{-1, -1}, {1, -1}, {-1, 1}, {1, 1}};
const MACROBLOCKD *const xd = &x->e_mbd;
const struct buf_2d *const what = &x->plane[0].src;
const struct buf_2d *const in_what = &xd->plane[0].pre[0];
const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
+ MV *best_mv = &x->best_mv.as_mv;
unsigned int best_sad = fn_ptr->sdaf(what->buf, what->stride,
- get_buf_from_mv(in_what, ref_mv), in_what->stride, second_pred) +
- mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
+ get_buf_from_mv(in_what, best_mv), in_what->stride, second_pred) +
+ mvsad_err_cost(x, best_mv, &fcenter_mv, error_per_bit);
int i, j;
for (i = 0; i < search_range; ++i) {
int best_site = -1;
for (j = 0; j < 8; ++j) {
- const MV mv = {ref_mv->row + neighbors[j].row,
- ref_mv->col + neighbors[j].col};
+ const MV mv = {best_mv->row + neighbors[j].row,
+ best_mv->col + neighbors[j].col};
if (is_mv_in(x, &mv)) {
unsigned int sad = fn_ptr->sdaf(what->buf, what->stride,
@@ -2358,8 +2354,8 @@
if (best_site == -1) {
break;
} else {
- ref_mv->row += neighbors[best_site].row;
- ref_mv->col += neighbors[best_site].col;
+ best_mv->row += neighbors[best_site].row;
+ best_mv->col += neighbors[best_site].col;
}
}
return best_sad;
@@ -2378,15 +2374,15 @@
}
int vp10_full_pixel_search(VP10_COMP *cpi, MACROBLOCK *x,
- BLOCK_SIZE bsize, MV *mvp_full,
- int step_param, int error_per_bit,
- int *cost_list,
- const MV *ref_mv, MV *tmp_mv,
- int var_max, int rd) {
+ BLOCK_SIZE bsize, MV *mvp_full,
+ int step_param, int error_per_bit,
+ int *cost_list, const MV *ref_mv,
+ int var_max, int rd) {
const SPEED_FEATURES *const sf = &cpi->sf;
const SEARCH_METHODS method = sf->mv.search_method;
vp10_variance_fn_ptr_t *fn_ptr = &cpi->fn_ptr[bsize];
int var = 0;
+
if (cost_list) {
cost_list[0] = INT_MAX;
cost_list[1] = INT_MAX;
@@ -2400,29 +2396,29 @@
switch (method) {
case FAST_DIAMOND:
- var = vp10_fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
- cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+ var = fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
+ cost_list, fn_ptr, 1, ref_mv);
break;
case FAST_HEX:
- var = vp10_fast_hex_search(x, mvp_full, step_param, error_per_bit, 0,
- cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+ var = fast_hex_search(x, mvp_full, step_param, error_per_bit, 0,
+ cost_list, fn_ptr, 1, ref_mv);
break;
case HEX:
var = vp10_hex_search(x, mvp_full, step_param, error_per_bit, 1,
- cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+ cost_list, fn_ptr, 1, ref_mv);
break;
case SQUARE:
- var = vp10_square_search(x, mvp_full, step_param, error_per_bit, 1,
- cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+ var = square_search(x, mvp_full, step_param, error_per_bit, 1,
+ cost_list, fn_ptr, 1, ref_mv);
break;
case BIGDIA:
- var = vp10_bigdia_search(x, mvp_full, step_param, error_per_bit, 1,
- cost_list, fn_ptr, 1, ref_mv, tmp_mv);
+ var = bigdia_search(x, mvp_full, step_param, error_per_bit, 1,
+ cost_list, fn_ptr, 1, ref_mv);
break;
case NSTEP:
- var = vp10_full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
- MAX_MVSEARCH_STEPS - 1 - step_param,
- 1, cost_list, fn_ptr, ref_mv, tmp_mv);
+ var = full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
+ MAX_MVSEARCH_STEPS - 1 - step_param,
+ 1, cost_list, fn_ptr, ref_mv);
// Should we allow a follow on exhaustive search?
if (is_exhaustive_allowed(cpi, x)) {
@@ -2434,13 +2430,13 @@
if (var > exhuastive_thr) {
int var_ex;
MV tmp_mv_ex;
- var_ex = full_pixel_exhaustive(cpi, x, tmp_mv,
+ var_ex = full_pixel_exhaustive(cpi, x, &x->best_mv.as_mv,
error_per_bit, cost_list, fn_ptr,
ref_mv, &tmp_mv_ex);
if (var_ex < var) {
var = var_ex;
- *tmp_mv = tmp_mv_ex;
+ x->best_mv.as_mv = tmp_mv_ex;
}
}
}
@@ -2452,7 +2448,7 @@
}
if (method != NSTEP && rd && var < var_max)
- var = vp10_get_mvpred_var(x, tmp_mv, ref_mv, fn_ptr, 1);
+ var = vp10_get_mvpred_var(x, &x->best_mv.as_mv, ref_mv, fn_ptr, 1);
return var;
}
diff --git a/vp10/encoder/mcomp.h b/vp10/encoder/mcomp.h
index 704e26c..f97f6c7 100644
--- a/vp10/encoder/mcomp.h
+++ b/vp10/encoder/mcomp.h
@@ -82,31 +82,24 @@
// Perform integral projection based motion estimation.
unsigned int vp10_int_pro_motion_estimation(const struct VP10_COMP *cpi,
- MACROBLOCK *x,
- BLOCK_SIZE bsize,
- int mi_row, int mi_col);
+ MACROBLOCK *x,
+ BLOCK_SIZE bsize,
+ int mi_row, int mi_col);
-typedef int (integer_mv_pattern_search_fn) (
- const MACROBLOCK *x,
- MV *ref_mv,
- int search_param,
- int error_per_bit,
- int do_init_search,
- int *cost_list,
- const vp10_variance_fn_ptr_t *vf,
- int use_mvcost,
- const MV *center_mv,
- MV *best_mv);
-integer_mv_pattern_search_fn vp10_hex_search;
-integer_mv_pattern_search_fn vp10_bigdia_search;
-integer_mv_pattern_search_fn vp10_square_search;
-integer_mv_pattern_search_fn vp10_fast_hex_search;
-integer_mv_pattern_search_fn vp10_fast_dia_search;
+int vp10_hex_search(MACROBLOCK *x,
+ MV *start_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search,
+ int *cost_list,
+ const vp10_variance_fn_ptr_t *vfp,
+ int use_mvcost,
+ const MV *center_mv);
typedef int (fractional_mv_step_fp) (
- const MACROBLOCK *x,
- MV *bestmv, const MV *ref_mv,
+ MACROBLOCK *x,
+ const MV *ref_mv,
int allow_hp,
int error_per_bit,
const vp10_variance_fn_ptr_t *vfp,
@@ -124,39 +117,32 @@
extern fractional_mv_step_fp vp10_find_best_sub_pixel_tree_pruned_evenmore;
typedef int (*vp10_full_search_fn_t)(const MACROBLOCK *x,
- const MV *ref_mv, int sad_per_bit,
- int distance,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv, MV *best_mv);
+ const MV *ref_mv, int sad_per_bit,
+ int distance,
+ const vp10_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv, MV *best_mv);
-typedef int (*vp10_refining_search_fn_t)(const MACROBLOCK *x,
- MV *ref_mv, int sad_per_bit,
- int distance,
+typedef int (*vp10_diamond_search_fn_t)(const MACROBLOCK *x,
+ const search_site_config *cfg,
+ MV *ref_mv, MV *best_mv,
+ int search_param, int sad_per_bit,
+ int *num00,
const vp10_variance_fn_ptr_t *fn_ptr,
const MV *center_mv);
-typedef int (*vp10_diamond_search_fn_t)(const MACROBLOCK *x,
- const search_site_config *cfg,
- MV *ref_mv, MV *best_mv,
- int search_param, int sad_per_bit,
- int *num00,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv);
-
-int vp10_refining_search_8p_c(const MACROBLOCK *x,
- MV *ref_mv, int error_per_bit,
- int search_range,
- const vp10_variance_fn_ptr_t *fn_ptr,
- const MV *center_mv, const uint8_t *second_pred);
+int vp10_refining_search_8p_c(MACROBLOCK *x,
+ int error_per_bit,
+ int search_range,
+ const vp10_variance_fn_ptr_t *fn_ptr,
+ const MV *center_mv, const uint8_t *second_pred);
struct VP10_COMP;
int vp10_full_pixel_search(struct VP10_COMP *cpi, MACROBLOCK *x,
- BLOCK_SIZE bsize, MV *mvp_full,
- int step_param, int error_per_bit,
- int *cost_list,
- const MV *ref_mv, MV *tmp_mv,
- int var_max, int rd);
+ BLOCK_SIZE bsize, MV *mvp_full,
+ int step_param, int error_per_bit,
+ int *cost_list, const MV *ref_mv,
+ int var_max, int rd);
#if CONFIG_EXT_INTER
int vp10_find_best_masked_sub_pixel_tree(const MACROBLOCK *x,
diff --git a/vp10/encoder/rdopt.c b/vp10/encoder/rdopt.c
index b96e6e4..45c10cd 100644
--- a/vp10/encoder/rdopt.c
+++ b/vp10/encoder/rdopt.c
@@ -15,6 +15,7 @@
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_dsp/blend.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem.h"
#include "vpx_ports/system_state.h"
@@ -4753,7 +4754,7 @@
struct buf_2d ref_yv12[2];
int bestsme = INT_MAX;
int sadpb = x->sadperbit16;
- MV tmp_mv;
+ MV *const best_mv = &x->best_mv.as_mv;
int search_range = 3;
int tmp_col_min = x->mv_col_min;
@@ -4814,23 +4815,22 @@
vp10_set_mv_search_range(x, &ref_mv[id].as_mv);
// Use the mv result from the single mode as mv predictor.
- tmp_mv = frame_mv[refs[id]].as_mv;
+ *best_mv = frame_mv[refs[id]].as_mv;
- tmp_mv.col >>= 3;
- tmp_mv.row >>= 3;
+ best_mv->col >>= 3;
+ best_mv->row >>= 3;
#if CONFIG_REF_MV
vp10_set_mvcost(x, refs[id]);
#endif
// Small-range full-pixel motion search.
- bestsme = vp10_refining_search_8p_c(x, &tmp_mv, sadpb,
- search_range,
- &cpi->fn_ptr[bsize],
- &ref_mv[id].as_mv, second_pred);
+ bestsme = vp10_refining_search_8p_c(x, sadpb, search_range,
+ &cpi->fn_ptr[bsize],
+ &ref_mv[id].as_mv, second_pred);
if (bestsme < INT_MAX)
- bestsme = vp10_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
- second_pred, &cpi->fn_ptr[bsize], 1);
+ bestsme = vp10_get_mvpred_av_var(x, best_mv, &ref_mv[id].as_mv,
+ second_pred, &cpi->fn_ptr[bsize], 1);
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
@@ -4859,8 +4859,7 @@
pd->pre[0].stride)) << 3];
bestsme = cpi->find_fractional_mv_step(
- x, &tmp_mv,
- &ref_mv[id].as_mv,
+ x, &ref_mv[id].as_mv,
cpi->common.allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
@@ -4875,8 +4874,7 @@
} else {
(void) block;
bestsme = cpi->find_fractional_mv_step(
- x, &tmp_mv,
- &ref_mv[id].as_mv,
+ x, &ref_mv[id].as_mv,
cpi->common.allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
@@ -4893,7 +4891,7 @@
xd->plane[0].pre[0] = ref_yv12[0];
if (bestsme < last_besterr[id]) {
- frame_mv[refs[id]].as_mv = tmp_mv;
+ frame_mv[refs[id]].as_mv = *best_mv;
last_besterr[id] = bestsme;
} else {
break;
@@ -5196,11 +5194,6 @@
run_mv_search)
#endif // CONFIG_EXT_INTER
) {
-#if CONFIG_EXT_INTER
- MV *const new_mv = &mode_mv[this_mode][0].as_mv;
-#else
- MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
-#endif // CONFIG_EXT_INTER
int step_param = 0;
int bestsme = INT_MAX;
int sadpb = x->sadperbit4;
@@ -5268,8 +5261,7 @@
bestsme = vp10_full_pixel_search(
cpi, x, bsize, &mvp_full, step_param, sadpb,
cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
- &bsi->ref_mv[0]->as_mv, new_mv,
- INT_MAX, 1);
+ &bsi->ref_mv[0]->as_mv, INT_MAX, 1);
if (bestsme < INT_MAX) {
int distortion;
@@ -5294,9 +5286,7 @@
pd->pre[0].stride)) << 3];
cpi->find_fractional_mv_step(
- x,
- new_mv,
- &bsi->ref_mv[0]->as_mv,
+ x, &bsi->ref_mv[0]->as_mv,
cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
@@ -5311,9 +5301,7 @@
pd->pre[0] = backup_pred;
} else {
cpi->find_fractional_mv_step(
- x,
- new_mv,
- &bsi->ref_mv[0]->as_mv,
+ x, &bsi->ref_mv[0]->as_mv,
cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
@@ -5327,14 +5315,20 @@
// save motion search result for use in compound prediction
#if CONFIG_EXT_INTER
- seg_mvs[i][mv_idx][mbmi->ref_frame[0]].as_mv = *new_mv;
+ seg_mvs[i][mv_idx][mbmi->ref_frame[0]].as_mv = x->best_mv.as_mv;
#else
- seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
+ seg_mvs[i][mbmi->ref_frame[0]].as_mv = x->best_mv.as_mv;
#endif // CONFIG_EXT_INTER
}
if (cpi->sf.adaptive_motion_search)
- x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
+ x->pred_mv[mbmi->ref_frame[0]] = x->best_mv.as_mv;
+
+#if CONFIG_EXT_INTER
+ mode_mv[this_mode][0] = x->best_mv;
+#else
+ mode_mv[NEWMV][0] = x->best_mv;
+#endif // CONFIG_EXT_INTER
// restore src pointers
mi_buf_restore(x, orig_src, orig_pre);
@@ -5903,7 +5897,7 @@
int ref_idx,
int mv_idx,
#endif // CONFIG_EXT_INTER
- int_mv *tmp_mv, int *rate_mv) {
+ int *rate_mv) {
MACROBLOCKD *xd = &x->e_mbd;
const VP10_COMMON *cm = &cpi->common;
MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
@@ -5985,7 +5979,7 @@
if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
x->pred_mv[ref].row = 0;
x->pred_mv[ref].col = 0;
- tmp_mv->as_int = INVALID_MV;
+ x->best_mv.as_int = INVALID_MV;
if (scaled_ref_frame) {
int i;
@@ -6005,7 +5999,7 @@
bestsme = vp10_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
cond_cost_list(cpi, cost_list),
- &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
+ &ref_mv, INT_MAX, 1);
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
@@ -6027,7 +6021,7 @@
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,
+ bestsme = cpi->find_fractional_mv_step(x, &ref_mv,
cm->allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
@@ -6041,7 +6035,7 @@
// Restore the reference frames.
pd->pre[ref_idx] = backup_pred;
} else {
- cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
+ cpi->find_fractional_mv_step(x, &ref_mv,
cm->allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
@@ -6052,11 +6046,11 @@
&dis, &x->pred_sse[ref], NULL, 0, 0, 0);
}
}
- *rate_mv = vp10_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
+ *rate_mv = vp10_mv_bit_cost(&x->best_mv.as_mv, &ref_mv,
x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
if (cpi->sf.adaptive_motion_search)
- x->pred_mv[ref] = tmp_mv->as_mv;
+ x->pred_mv[ref] = x->best_mv.as_mv;
if (scaled_ref_frame) {
int i;
@@ -6990,34 +6984,32 @@
}
#endif // CONFIG_EXT_INTER
} else {
- int_mv tmp_mv;
-
#if CONFIG_EXT_INTER
if (is_comp_interintra_pred) {
- tmp_mv = single_newmvs[mv_idx][refs[0]];
+ x->best_mv = single_newmvs[mv_idx][refs[0]];
rate_mv = single_newmvs_rate[mv_idx][refs[0]];
} else {
single_motion_search(cpi, x, bsize, mi_row, mi_col,
- 0, mv_idx, &tmp_mv, &rate_mv);
- single_newmvs[mv_idx][refs[0]] = tmp_mv;
+ 0, mv_idx, &rate_mv);
+ single_newmvs[mv_idx][refs[0]] = x->best_mv;
single_newmvs_rate[mv_idx][refs[0]] = rate_mv;
}
#else
- single_motion_search(cpi, x, bsize, mi_row, mi_col, &tmp_mv, &rate_mv);
- single_newmv[refs[0]] = tmp_mv;
+ single_motion_search(cpi, x, bsize, mi_row, mi_col, &rate_mv);
+ single_newmv[refs[0]] = x->best_mv;
#endif // CONFIG_EXT_INTER
- if (tmp_mv.as_int == INVALID_MV)
+ if (x->best_mv.as_int == INVALID_MV)
return INT64_MAX;
- frame_mv[refs[0]] = tmp_mv;
- xd->mi[0]->bmi[0].as_mv[0] = tmp_mv;
+ frame_mv[refs[0]] = x->best_mv;
+ xd->mi[0]->bmi[0].as_mv[0] = x->best_mv;
// Estimate the rate implications of a new mv but discount this
// under certain circumstances where we want to help initiate a weak
// motion field, where the distortion gain for a single block may not
// be enough to overcome the cost of a new mv.
- if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
+ if (discount_newmv_test(cpi, this_mode, x->best_mv, mode_mv, refs[0])) {
rate_mv = VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
}
}
@@ -7780,8 +7772,7 @@
vp10_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
#endif // CONFIG_EXT_INTER
}
- vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col, 0,
- NULL, NULL,
+ vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col,
dst_buf1, dst_stride1,
dst_buf2, dst_stride2);
model_rd_for_sb(cpi, bsize, x, xd, 0, MAX_MB_PLANE - 1,
@@ -8399,6 +8390,18 @@
}
#endif // CONFIG_EXT_INTRA
+#if CONFIG_OBMC
+static void calc_target_weighted_pred(
+ const VP10_COMMON *cm,
+ const MACROBLOCK *x,
+ const MACROBLOCKD *xd,
+ int mi_row, int mi_col,
+ const uint8_t *above, int above_stride,
+ const uint8_t *left, int left_stride,
+ int32_t *mask_buf,
+ int32_t *wsrc_buf);
+#endif // CONFIG_OBMC
+
void vp10_rd_pick_inter_mode_sb(VP10_COMP *cpi,
TileDataEnc *tile_data,
MACROBLOCK *x,
@@ -9579,7 +9582,7 @@
vp10_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
#if CONFIG_OBMC
if (mbmi->motion_variation == OBMC_CAUSAL)
- vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col, 0, NULL, NULL,
+ vp10_build_obmc_inter_prediction(cm, xd, mi_row, mi_col,
dst_buf1, dst_stride1,
dst_buf2, dst_stride2);
#endif // CONFIG_OBMC
@@ -10980,189 +10983,225 @@
}
#if CONFIG_OBMC
-void calc_target_weighted_pred(VP10_COMMON *cm,
- MACROBLOCK *x,
- MACROBLOCKD *xd,
- int mi_row, int mi_col,
- uint8_t *above_buf, int above_stride,
- uint8_t *left_buf, int left_stride,
- int32_t *mask_buf,
- int32_t *weighted_src_buf) {
- BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
- int row, col, i, mi_step;
- int bw = 8 * xd->n8_w;
- int bh = 8 * xd->n8_h;
+// This function has a structure similar to vp10_build_obmc_inter_prediction
+//
+// The OBMC predictor is computed as:
+//
+// PObmc(x,y) =
+// VPX_BLEND_A64(Mh(x),
+// VPX_BLEND_A64(Mv(y), P(x,y), PAbove(x,y)),
+// PLeft(x, y))
+//
+// Scaling up by VPX_BLEND_A64_MAX_ALPHA ** 2 and omitting the intermediate
+// rounding, this can be written as:
+//
+// VPX_BLEND_A64_MAX_ALPHA * VPX_BLEND_A64_MAX_ALPHA * Pobmc(x,y) =
+// Mh(x) * Mv(y) * P(x,y) +
+// Mh(x) * Cv(y) * Pabove(x,y) +
+// VPX_BLEND_A64_MAX_ALPHA * Ch(x) * PLeft(x, y)
+//
+// Where :
+//
+// Cv(y) = VPX_BLEND_A64_MAX_ALPHA - Mv(y)
+// Ch(y) = VPX_BLEND_A64_MAX_ALPHA - Mh(y)
+//
+// This function computes 'wsrc' and 'mask' as:
+//
+// wsrc(x, y) =
+// VPX_BLEND_A64_MAX_ALPHA * VPX_BLEND_A64_MAX_ALPHA * src(x, y) -
+// Mh(x) * Cv(y) * Pabove(x,y) +
+// VPX_BLEND_A64_MAX_ALPHA * Ch(x) * PLeft(x, y)
+//
+// mask(x, y) = Mh(x) * Mv(y)
+//
+// These can then be used to efficiently approximate the error for any
+// predictor P in the context of the provided neighbouring predictors by
+// computing:
+//
+// error(x, y) =
+// wsrc(x, y) - mask(x, y) * P(x, y) / (VPX_BLEND_A64_MAX_ALPHA ** 2)
+//
+static void calc_target_weighted_pred(
+ const VP10_COMMON *cm,
+ const MACROBLOCK *x,
+ const MACROBLOCKD *xd,
+ int mi_row, int mi_col,
+ const uint8_t *above, int above_stride,
+ const uint8_t *left, int left_stride,
+ int32_t *mask_buf,
+ int32_t *wsrc_buf) {
+ const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+ int row, col, i;
+ const int bw = 8 * xd->n8_w;
+ const int bh = 8 * xd->n8_h;
+ const int wsrc_stride = bw;
const int mask_stride = bw;
- const int weighted_src_stride = bw;
- int32_t *dst = weighted_src_buf;
- int32_t *mask2d = mask_buf;
- uint8_t *src;
+ const int src_scale = VPX_BLEND_A64_MAX_ALPHA * VPX_BLEND_A64_MAX_ALPHA;
#if CONFIG_VP9_HIGHBITDEPTH
- int is_hbd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+ const int is_hbd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+#else
+ const int is_hbd = 0;
#endif // CONFIG_VP9_HIGHBITDEPTH
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col) {
- dst[col] = 0;
- mask2d[col] = 64;
- }
- dst += weighted_src_stride;
- mask2d += mask_stride;
- }
+
+ // plane 0 should not be subsampled
+ assert(xd->plane[0].subsampling_x == 0);
+ assert(xd->plane[0].subsampling_y == 0);
+
+ vp10_zero_array(wsrc_buf, bw * bh);
+ for (i = 0; i < bw * bh; ++i)
+ mask_buf[i] = VPX_BLEND_A64_MAX_ALPHA;
// handle above row
if (xd->up_available) {
- for (i = 0; i < VPXMIN(xd->n8_w, cm->mi_cols - mi_col); i += mi_step) {
- int mi_row_offset = -1;
- int mi_col_offset = i;
- MODE_INFO *above_mi = xd->mi[mi_col_offset +
- mi_row_offset * xd->mi_stride];
- MB_MODE_INFO *above_mbmi = &above_mi->mbmi;
- int overlap = num_4x4_blocks_high_lookup[bsize] << 1;
+ const int overlap = num_4x4_blocks_high_lookup[bsize] * 2;
+ const int miw = VPXMIN(xd->n8_w, cm->mi_cols - mi_col);
+ const int mi_row_offset = -1;
+ const uint8_t *const mask1d = vp10_get_obmc_mask(overlap);
- mi_step = VPXMIN(xd->n8_w,
- num_8x8_blocks_wide_lookup[above_mbmi->sb_type]);
+ assert(miw > 0);
+
+ i = 0;
+ do { // for each mi in the above row
+ const int mi_col_offset = i;
+ const MB_MODE_INFO *const above_mbmi =
+ &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+ const int mi_step =
+ VPXMIN(xd->n8_w, num_8x8_blocks_wide_lookup[above_mbmi->sb_type]);
+ const int neighbor_bw = mi_step * MI_SIZE;
if (is_neighbor_overlappable(above_mbmi)) {
- const struct macroblockd_plane *pd = &xd->plane[0];
- int bw = (mi_step * MI_SIZE) >> pd->subsampling_x;
- int bh = overlap >> pd->subsampling_y;
- int dst_stride = weighted_src_stride;
- int32_t *dst = weighted_src_buf + (i * MI_SIZE >> pd->subsampling_x);
- int tmp_stride = above_stride;
- uint8_t *tmp = above_buf + (i * MI_SIZE >> pd->subsampling_x);
- int mask2d_stride = mask_stride;
- int32_t *mask2d = mask_buf + (i * MI_SIZE >> pd->subsampling_x);
- const uint8_t *mask1d[2];
+ const int tmp_stride = above_stride;
+ int32_t *wsrc = wsrc_buf + (i * MI_SIZE);
+ int32_t *mask = mask_buf + (i * MI_SIZE);
- setup_obmc_mask(bh, mask1d);
+ if (!is_hbd) {
+ const uint8_t *tmp = above;
-#if CONFIG_VP9_HIGHBITDEPTH
- if (is_hbd) {
- uint16_t *tmp16 = CONVERT_TO_SHORTPTR(tmp);
-
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col) {
- dst[col] = mask1d[1][row] * tmp16[col];
- mask2d[col] = mask1d[0][row];
+ for (row = 0; row < overlap; ++row) {
+ const uint8_t m0 = mask1d[row];
+ const uint8_t m1 = VPX_BLEND_A64_MAX_ALPHA - m0;
+ for (col = 0; col < neighbor_bw; ++col) {
+ wsrc[col] = m1 * tmp[col];
+ mask[col] = m0;
}
- dst += dst_stride;
- tmp16 += tmp_stride;
- mask2d += mask2d_stride;
+ wsrc += wsrc_stride;
+ mask += mask_stride;
+ tmp += tmp_stride;
}
- } else {
-#endif // CONFIG_VP9_HIGHBITDEPTH
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col) {
- dst[col] = mask1d[1][row] * tmp[col];
- mask2d[col] = mask1d[0][row];
- }
- dst += dst_stride;
- tmp += tmp_stride;
- mask2d += mask2d_stride;
- }
#if CONFIG_VP9_HIGHBITDEPTH
- }
+ } else {
+ const uint16_t *tmp = CONVERT_TO_SHORTPTR(above);
+
+ for (row = 0; row < overlap; ++row) {
+ const uint8_t m0 = mask1d[row];
+ const uint8_t m1 = VPX_BLEND_A64_MAX_ALPHA - m0;
+ for (col = 0; col < neighbor_bw; ++col) {
+ wsrc[col] = m1 * tmp[col];
+ mask[col] = m0;
+ }
+ wsrc += wsrc_stride;
+ mask += mask_stride;
+ tmp += tmp_stride;
+ }
#endif // CONFIG_VP9_HIGHBITDEPTH
+ }
}
- } // each mi in the above row
+
+ above += neighbor_bw;
+ i += mi_step;
+ } while (i < miw);
+ }
+
+ for (i = 0; i < bw * bh; ++i) {
+ wsrc_buf[i] *= VPX_BLEND_A64_MAX_ALPHA;
+ mask_buf[i] *= VPX_BLEND_A64_MAX_ALPHA;
}
// handle left column
- dst = weighted_src_buf;
- mask2d = mask_buf;
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col) {
- dst[col] = dst[col] << 6;
- mask2d[col] = mask2d[col] << 6;
- }
- dst += weighted_src_stride;
- mask2d += mask_stride;
- }
-
if (xd->left_available) {
- for (i = 0; i < VPXMIN(xd->n8_h, cm->mi_rows - mi_row); i += mi_step) {
- int mi_row_offset = i;
- int mi_col_offset = -1;
- int overlap = num_4x4_blocks_wide_lookup[bsize] << 1;
- MODE_INFO *left_mi = xd->mi[mi_col_offset +
- mi_row_offset * xd->mi_stride];
- MB_MODE_INFO *left_mbmi = &left_mi->mbmi;
+ const int overlap = num_4x4_blocks_wide_lookup[bsize] * 2;
+ const int mih = VPXMIN(xd->n8_h, cm->mi_rows - mi_row);
+ const int mi_col_offset = -1;
+ const uint8_t *const mask1d = vp10_get_obmc_mask(overlap);
- mi_step = VPXMIN(xd->n8_h,
- num_8x8_blocks_high_lookup[left_mbmi->sb_type]);
+ assert(mih > 0);
+
+ i = 0;
+ do { // for each mi in the left column
+ const int mi_row_offset = i;
+ const MB_MODE_INFO *const left_mbmi =
+ &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+ const int mi_step =
+ VPXMIN(xd->n8_h, num_8x8_blocks_high_lookup[left_mbmi->sb_type]);
+ const int neighbor_bh = mi_step * MI_SIZE;
if (is_neighbor_overlappable(left_mbmi)) {
- const struct macroblockd_plane *pd = &xd->plane[0];
- int bw = overlap >> pd->subsampling_x;
- int bh = (mi_step * MI_SIZE) >> pd->subsampling_y;
- int dst_stride = weighted_src_stride;
- int32_t *dst = weighted_src_buf +
- (i * MI_SIZE * dst_stride >> pd->subsampling_y);
- int tmp_stride = left_stride;
- uint8_t *tmp = left_buf +
- (i * MI_SIZE * tmp_stride >> pd->subsampling_y);
- int mask2d_stride = mask_stride;
- int32_t *mask2d = mask_buf +
- (i * MI_SIZE * mask2d_stride >> pd->subsampling_y);
- const uint8_t *mask1d[2];
+ const int tmp_stride = left_stride;
+ int32_t *wsrc = wsrc_buf + (i * MI_SIZE * wsrc_stride);
+ int32_t *mask = mask_buf + (i * MI_SIZE * mask_stride);
- setup_obmc_mask(bw, mask1d);
+ if (!is_hbd) {
+ const uint8_t *tmp = left;
-#if CONFIG_VP9_HIGHBITDEPTH
- if (is_hbd) {
- uint16_t *tmp16 = CONVERT_TO_SHORTPTR(tmp);
-
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col) {
- dst[col] = (dst[col] >> 6) * mask1d[0][col] +
- (tmp16[col] << 6) * mask1d[1][col];
- mask2d[col] = (mask2d[col] >> 6) * mask1d[0][col];
+ for (row = 0; row < neighbor_bh; ++row) {
+ for (col = 0; col < overlap; ++col) {
+ const uint8_t m0 = mask1d[col];
+ const uint8_t m1 = VPX_BLEND_A64_MAX_ALPHA - m0;
+ wsrc[col] = (wsrc[col] >> VPX_BLEND_A64_ROUND_BITS) * m0 +
+ (tmp[col] << VPX_BLEND_A64_ROUND_BITS) * m1;
+ mask[col] = (mask[col] >> VPX_BLEND_A64_ROUND_BITS) * m0;
}
- dst += dst_stride;
- tmp16 += tmp_stride;
- mask2d += mask2d_stride;
+ wsrc += wsrc_stride;
+ mask += mask_stride;
+ tmp += tmp_stride;
}
- } else {
-#endif // CONFIG_VP9_HIGHBITDEPTH
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col) {
- dst[col] = (dst[col] >> 6) * mask1d[0][col] +
- (tmp[col] << 6) * mask1d[1][col];
- mask2d[col] = (mask2d[col] >> 6) * mask1d[0][col];
- }
- dst += dst_stride;
- tmp += tmp_stride;
- mask2d += mask2d_stride;
- }
#if CONFIG_VP9_HIGHBITDEPTH
+ } else {
+ const uint16_t *tmp = CONVERT_TO_SHORTPTR(left);
+
+ for (row = 0; row < neighbor_bh; ++row) {
+ for (col = 0; col < overlap; ++col) {
+ const uint8_t m0 = mask1d[col];
+ const uint8_t m1 = VPX_BLEND_A64_MAX_ALPHA - m0;
+ wsrc[col] = (wsrc[col] >> VPX_BLEND_A64_ROUND_BITS) * m0 +
+ (tmp[col] << VPX_BLEND_A64_ROUND_BITS) * m1;
+ mask[col] = (mask[col] >> VPX_BLEND_A64_ROUND_BITS) * m0;
+ }
+ wsrc += wsrc_stride;
+ mask += mask_stride;
+ tmp += tmp_stride;
+ }
+#endif // CONFIG_VP9_HIGHBITDEPTH
}
-#endif // CONFIG_VP9_HIGHBITDEPTH
}
- } // each mi in the left column
+
+ left += neighbor_bh * left_stride;
+ i += mi_step;
+ } while (i < mih);
}
- dst = weighted_src_buf;
- src = x->plane[0].src.buf;
-#if CONFIG_VP9_HIGHBITDEPTH
- if (is_hbd) {
- uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+ if (!is_hbd) {
+ const uint8_t *src = x->plane[0].src.buf;
for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col)
- dst[col] = (src16[col] << 12) - dst[col];
- dst += weighted_src_stride;
- src16 += x->plane[0].src.stride;
+ for (col = 0; col < bw; ++col) {
+ wsrc_buf[col] = src[col] * src_scale - wsrc_buf[col];
+ }
+ wsrc_buf += wsrc_stride;
+ src += x->plane[0].src.stride;
}
- } else {
-#endif // CONFIG_VP9_HIGHBITDEPTH
- for (row = 0; row < bh; ++row) {
- for (col = 0; col < bw; ++col)
- dst[col] = (src[col] << 12) - dst[col];
- dst += weighted_src_stride;
- src += x->plane[0].src.stride;
- }
#if CONFIG_VP9_HIGHBITDEPTH
- }
+ } else {
+ const uint16_t *src = CONVERT_TO_SHORTPTR(x->plane[0].src.buf);
+
+ for (row = 0; row < bh; ++row) {
+ for (col = 0; col < bw; ++col) {
+ wsrc_buf[col] = src[col] * src_scale - wsrc_buf[col];
+ }
+ wsrc_buf += wsrc_stride;
+ src += x->plane[0].src.stride;
+ }
#endif // CONFIG_VP9_HIGHBITDEPTH
+ }
}
#endif // CONFIG_OBMC
diff --git a/vp10/encoder/rdopt.h b/vp10/encoder/rdopt.h
index b660e23..da70a22 100644
--- a/vp10/encoder/rdopt.h
+++ b/vp10/encoder/rdopt.h
@@ -90,16 +90,6 @@
int use_fast_coef_casting);
#endif // CONFIG_SUPERTX
-#if CONFIG_OBMC
-void calc_target_weighted_pred(VP10_COMMON *cm,
- MACROBLOCK *x,
- MACROBLOCKD *xd,
- int mi_row, int mi_col,
- uint8_t *above_buf, int above_stride,
- uint8_t *left_buf, int left_stride,
- int32_t *mask_buf, int32_t *weighted_src_buf);
-#endif // CONFIG_OBMC
-
#ifdef __cplusplus
} // extern "C"
#endif
diff --git a/vp10/encoder/temporal_filter.c b/vp10/encoder/temporal_filter.c
index 39be057..d125dae 100644
--- a/vp10/encoder/temporal_filter.c
+++ b/vp10/encoder/temporal_filter.c
@@ -288,7 +288,6 @@
MV best_ref_mv1 = {0, 0};
MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
- MV *ref_mv = &x->e_mbd.mi[0]->bmi[0].as_mv[0].as_mv;
// Save input state
struct buf_2d src = x->plane[0].src;
@@ -315,12 +314,11 @@
// Ignore mv costing by sending NULL pointer instead of cost arrays
vp10_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1,
- cond_cost_list(cpi, cost_list),
- &cpi->fn_ptr[BLOCK_16X16], 0, &best_ref_mv1, ref_mv);
+ cond_cost_list(cpi, cost_list),
+ &cpi->fn_ptr[BLOCK_16X16], 0, &best_ref_mv1);
// Ignore mv costing by sending NULL pointer instead of cost array
- bestsme = cpi->find_fractional_mv_step(x, ref_mv,
- &best_ref_mv1,
+ bestsme = cpi->find_fractional_mv_step(x, &best_ref_mv1,
cpi->common.allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[BLOCK_16X16],
@@ -329,6 +327,8 @@
NULL, NULL,
&distortion, &sse, NULL, 0, 0, 0);
+ x->e_mbd.mi[0]->bmi[0].as_mv[0] = x->best_mv;
+
// Restore input state
x->plane[0].src = src;
xd->plane[0].pre[0] = pre;