ref: 83c76118eb7129f6865a2aa28251aac68785b324
parent: 89ffda0ddf6c8abc044b308cf36dcec883b36977
author: Deb Mukherjee <[email protected]>
date: Tue Sep 9 16:07:11 EDT 2014
Use bigdia search with pruned subpel search Improves function to return sad of integer pels by reusing integer pels already visited in the smallest scale. Turns on BIGDIA search for speed 4. Also, turns on the first version of the pruned subpel search at this speed. derf: -0.32% (speed 4) Speed seems to improve by at least 5% but subject to verification. Change-Id: Iaec8eaffd61d6237ac029e6a2a1b0a88b2a35271
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -320,23 +320,23 @@
switch (whichdir) {case 0:
CHECK_BETTER(left, tr, tc - hstep);
- CHECK_BETTER(up, tr - hstep, tc);
- CHECK_BETTER(diag, tr - hstep, tc - hstep);
+ CHECK_BETTER(down, tr + hstep, tc);
+ CHECK_BETTER(diag, tr + hstep, tc - hstep);
break;
case 1:
CHECK_BETTER(right, tr, tc + hstep);
- CHECK_BETTER(up, tr - hstep, tc);
- CHECK_BETTER(diag, tr - hstep, tc + hstep);
+ CHECK_BETTER(down, tr + hstep, tc);
+ CHECK_BETTER(diag, tr + hstep, tc + hstep);
break;
case 2:
CHECK_BETTER(left, tr, tc - hstep);
- CHECK_BETTER(down, tr + hstep, tc);
- CHECK_BETTER(diag, tr + hstep, tc - hstep);
+ CHECK_BETTER(up, tr - hstep, tc);
+ CHECK_BETTER(diag, tr - hstep, tc - hstep);
break;
case 3:
CHECK_BETTER(right, tr, tc + hstep);
- CHECK_BETTER(down, tr + hstep, tc);
- CHECK_BETTER(diag, tr + hstep, tc + hstep);
+ CHECK_BETTER(up, tr - hstep, tc);
+ CHECK_BETTER(diag, tr - hstep, tc + hstep);
break;
}
} else {@@ -648,11 +648,11 @@
// Returns the one-away integer pel sad values around the best as follows:
// sad_list[0]: sad at the best integer pel
// sad_list[1]: sad at delta {0, -1} (left) from the best integer pel- // sad_list[2]: sad at delta {-1, 0} (top) from the best integer pel+ // sad_list[2]: sad at delta { 1, 0} (bottom) from the best integer pel // sad_list[3]: sad at delta { 0, 1} (right) from the best integer pel- // sad_list[4]: sad at delta { 1, 0} (bottom) from the best integer pel+ // sad_list[4]: sad at delta {-1, 0} (top) from the best integer pel if (sad_list) {- static const MV neighbors[4] = {{0, -1}, {-1, 0}, {0, 1}, {1, 0}};+ static const MV neighbors[4] = {{0, -1}, {1, 0}, {0, 1}, {-1, 0}};sad_list[0] = bestsad;
if (check_bounds(x, br, bc, 1)) { for (i = 0; i < 4; i++) {@@ -660,7 +660,10 @@
bc + neighbors[i].col};
sad_list[i + 1] = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
- in_what->stride);
+ in_what->stride) +
+ (use_mvcost ?
+ mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit) :
+ 0);
}
} else { for (i = 0; i < 4; i++) {@@ -671,8 +674,301 @@
else
sad_list[i + 1] = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
+ in_what->stride) +
+ (use_mvcost ?
+ mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit) :
+ 0);
+ }
+ }
+ }
+ best_mv->row = br;
+ best_mv->col = bc;
+ return bestsad;
+}
+
+// A specialized function where the smallest scale search candidates
+// are 4 1-away neighbors, and sad_list is non-null
+// TODO(debargha): Merge this function with the one above. Also remove
+// use_mvcost option since it is always 1, to save unnecessary branches.
+static int vp9_pattern_search_sad(const MACROBLOCK *x,
+ MV *ref_mv,
+ int search_param,
+ int sad_per_bit,
+ int do_init_search,
+ int *sad_list,
+ const vp9_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]) {+ 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,
+ };
+ int i, s, t;
+ const struct buf_2d *const what = &x->plane[0].src;
+ const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+ int br, bc;
+ int bestsad = INT_MAX;
+ int thissad;
+ int k = -1;
+ const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};+ int best_init_s = search_param_to_steps[search_param];
+ // adjust ref_mv to make sure it is within MV range
+ clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
+ br = ref_mv->row;
+ bc = ref_mv->col;
+ if (sad_list != NULL) {+ sad_list[0] = sad_list[1] = sad_list[2] = sad_list[3] = sad_list[4] =
+ INT_MAX;
+ }
+
+ // Work out the start point for the search
+ bestsad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, ref_mv), in_what->stride) +
+ mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+
+ // Search all possible scales upto the search param around the center point
+ // pick the scale of the point that is best as the starting scale of
+ // further steps around it.
+ if (do_init_search) {+ s = best_init_s;
+ best_init_s = -1;
+ for (t = 0; t <= s; ++t) {+ int best_site = -1;
+ if (check_bounds(x, br, bc, 1 << t)) {+ for (i = 0; i < num_candidates[t]; i++) {+ const MV this_mv = {br + candidates[t][i].row,+ bc + candidates[t][i].col};
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ } else {+ for (i = 0; i < num_candidates[t]; i++) {+ const MV this_mv = {br + candidates[t][i].row,+ bc + candidates[t][i].col};
+ if (!is_mv_in(x, &this_mv))
+ continue;
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ }
+ if (best_site == -1) {+ continue;
+ } else {+ best_init_s = t;
+ k = best_site;
+ }
+ }
+ if (best_init_s != -1) {+ br += candidates[best_init_s][k].row;
+ bc += candidates[best_init_s][k].col;
+ }
+ }
+
+ // If the center point is still the best, just skip this and move to
+ // the refinement step.
+ if (best_init_s != -1) {+ int do_sad = (num_candidates[0] == 4 && sad_list != NULL);
+ int best_site = -1;
+ s = best_init_s;
+
+ for (; s >= do_sad; s--) {+ if (!do_init_search || s != best_init_s) {+ if (check_bounds(x, br, bc, 1 << s)) {+ for (i = 0; i < num_candidates[s]; i++) {+ const MV this_mv = {br + candidates[s][i].row,+ bc + candidates[s][i].col};
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ } else {+ for (i = 0; i < num_candidates[s]; i++) {+ const MV this_mv = {br + candidates[s][i].row,+ bc + candidates[s][i].col};
+ if (!is_mv_in(x, &this_mv))
+ continue;
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ }
+
+ if (best_site == -1) {+ continue;
+ } else {+ br += candidates[s][best_site].row;
+ bc += candidates[s][best_site].col;
+ k = best_site;
+ }
+ }
+
+ do {+ int next_chkpts_indices[PATTERN_CANDIDATES_REF];
+ best_site = -1;
+ next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
+ next_chkpts_indices[1] = k;
+ next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
+
+ if (check_bounds(x, br, bc, 1 << s)) {+ for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {+ const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,+ bc + candidates[s][next_chkpts_indices[i]].col};
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ } else {+ for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {+ const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,+ bc + candidates[s][next_chkpts_indices[i]].col};
+ if (!is_mv_in(x, &this_mv))
+ continue;
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ }
+
+ if (best_site != -1) {+ k = next_chkpts_indices[best_site];
+ br += candidates[s][k].row;
+ bc += candidates[s][k].col;
+ }
+ } while (best_site != -1);
+ }
+
+ // Note: If we enter the if below, then sad_list must be non-NULL.
+ if (s == 0) {+ sad_list[0] = bestsad;
+ if (!do_init_search || s != best_init_s) {+ if (check_bounds(x, br, bc, 1 << s)) {+ for (i = 0; i < num_candidates[s]; i++) {+ const MV this_mv = {br + candidates[s][i].row,+ bc + candidates[s][i].col};
+ sad_list[i + 1] =
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ } else {+ for (i = 0; i < num_candidates[s]; i++) {+ const MV this_mv = {br + candidates[s][i].row,+ bc + candidates[s][i].col};
+ if (!is_mv_in(x, &this_mv))
+ continue;
+ sad_list[i + 1] =
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ }
+
+ if (best_site != -1) {+ br += candidates[s][best_site].row;
+ bc += candidates[s][best_site].col;
+ k = best_site;
+ }
+ }
+ while (best_site != -1) {+ int next_chkpts_indices[PATTERN_CANDIDATES_REF];
+ best_site = -1;
+ next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
+ next_chkpts_indices[1] = k;
+ next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
+ sad_list[1] = sad_list[2] = sad_list[3] = sad_list[4] = INT_MAX;
+ sad_list[((k + 2) % 4) + 1] = sad_list[0];
+ sad_list[0] = bestsad;
+
+ if (check_bounds(x, br, bc, 1 << s)) {+ for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {+ const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,+ bc + candidates[s][next_chkpts_indices[i]].col};
+ sad_list[next_chkpts_indices[i] + 1] =
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ } else {+ for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {+ const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,+ bc + candidates[s][next_chkpts_indices[i]].col};
+ if (!is_mv_in(x, &this_mv)) {+ sad_list[next_chkpts_indices[i] + 1] = INT_MAX;
+ continue;
+ }
+ sad_list[next_chkpts_indices[i] + 1] =
+ thissad = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ CHECK_BETTER
+ }
+ }
+
+ if (best_site != -1) {+ k = next_chkpts_indices[best_site];
+ br += candidates[s][k].row;
+ bc += candidates[s][k].col;
+ }
+ }
+ }
+ }
+
+ // Returns the one-away integer pel sad values around the best as follows:
+ // sad_list[0]: sad at the best integer pel
+ // sad_list[1]: sad at delta {0, -1} (left) from the best integer pel+ // sad_list[2]: sad at delta { 1, 0} (bottom) from the best integer pel+ // sad_list[3]: sad at delta { 0, 1} (right) from the best integer pel+ // sad_list[4]: sad at delta {-1, 0} (top) from the best integer pel+ if (sad_list) {+ static const MV neighbors[4] = {{0, -1}, {1, 0}, {0, 1}, {-1, 0}};+ if (sad_list[0] == INT_MAX) {+ sad_list[0] = bestsad;
+ if (check_bounds(x, br, bc, 1)) {+ for (i = 0; i < 4; i++) {+ const MV this_mv = {br + neighbors[i].row,+ bc + neighbors[i].col};
+ sad_list[i + 1] = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
in_what->stride);
+ }
+ } else {+ for (i = 0; i < 4; i++) {+ const MV this_mv = {br + neighbors[i].row,+ bc + neighbors[i].col};
+ if (!is_mv_in(x, &this_mv))
+ sad_list[i + 1] = INT_MAX;
+ else
+ sad_list[i + 1] = vfp->sdf(what->buf, what->stride,
+ get_buf_from_mv(in_what, &this_mv),
+ in_what->stride);
+ }
}
+ } else {+ if (use_mvcost) {+ for (i = 0; i < 4; i++) {+ const MV this_mv = {br + neighbors[i].row,+ bc + neighbors[i].col};
+ if (sad_list[i + 1] != INT_MAX) {+ sad_list[i + 1] +=
+ mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+ }
+ }
+ }
}
}
best_mv->row = br;
@@ -784,10 +1080,10 @@
{{-512, -512}, {0, -1024}, {512, -512}, {1024, 0}, {512, 512}, {0, 1024}, {-512, 512}, {-1024, 0}},};
- return vp9_pattern_search(x, ref_mv, search_param, sad_per_bit,
- do_init_search, sad_list, vfp, use_mvcost,
- center_mv, best_mv,
- bigdia_num_candidates, bigdia_candidates);
+ return vp9_pattern_search_sad(x, ref_mv, search_param, sad_per_bit,
+ do_init_search, sad_list, vfp, use_mvcost,
+ center_mv, best_mv,
+ bigdia_num_candidates, bigdia_candidates);
}
int vp9_square_search(const MACROBLOCK *x,
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -160,6 +160,8 @@
sf->use_square_partition_only = 1;
sf->tx_size_search_method = USE_LARGESTALL;
sf->disable_split_mask = DISABLE_ALL_SPLIT;
+ sf->mv.search_method = BIGDIA;
+ sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
sf->adaptive_rd_thresh = 4;
sf->mode_search_skip_flags |= FLAG_SKIP_COMP_REFMISMATCH |
FLAG_EARLY_TERMINATE;