Skip to content

GitLab

Commit 3426f114 authored by Hui Su's avatar Hui Su Committed by Gerrit Code Review
Browse files

Merge "Refactor vp10_pattern_search" into nextgenv2

parents ab994029 19f3eaa6
Hide whitespace changes
Inline Side-by-side
vp10/encoder/mcomp.c
View file @ 3426f114
......@@ -986,199 +986,75 @@ static INLINE void calc_int_cost_list(const MACROBLOCK *x,
}
}
// Generic pattern search function that searches over multiple scales.
// Each scale can have a different number of candidates and shape of
// candidates as indicated in the num_candidates and candidates arrays
// passed into this function
//
static int vp10_pattern_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,
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;
static INLINE void calc_int_sad_list(const MACROBLOCK *x,
const MV *const ref_mv,
int sadpb,
const vp10_variance_fn_ptr_t *fn_ptr,
const MV *best_mv, int *cost_list,
const int use_mvcost,
const int bestsad) {
static const MV neighbors[4] = {{0, -1}, {1, 0}, {0, 1}, {-1, 0}};
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;
// 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);
const struct buf_2d *const in_what = &x->e_mbd.plane[0].pre[0];
const MV fcenter_mv = {ref_mv->row >> 3, ref_mv->col >> 3};
int i;
const int br = best_mv->row;
const int bc = best_mv->col;
// 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 (cost_list[0] == INT_MAX) {
cost_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 };
cost_list[i + 1] = fn_ptr->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
}
if (best_site == -1) {
continue;
} else {
best_init_s = t;
k = best_site;
} 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))
cost_list[i + 1] = INT_MAX;
else
cost_list[i + 1] = fn_ptr->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
}
}
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 best_site = -1;
s = best_init_s;
do {
// No need to search all 6 points the 1st time if initial search was used
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;
} else {
if (use_mvcost) {
for (i = 0; i < 4; i++) {
const MV this_mv = {br + neighbors[i].row,
bc + neighbors[i].col};
if (cost_list[i + 1] != INT_MAX) {
cost_list[i + 1] +=
mvsad_err_cost(x, &this_mv, &fcenter_mv, sadpb);
}
}
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);
} while (s--);
}
// Returns the one-away integer pel sad values around the best as follows:
// cost_list[0]: cost at the best integer pel
// cost_list[1]: cost at delta {0, -1} (left) from the best integer pel
// cost_list[2]: cost at delta { 1, 0} (bottom) from the best integer pel
// cost_list[3]: cost at delta { 0, 1} (right) from the best integer pel
// cost_list[4]: cost at delta {-1, 0} (top) from the best integer pel
if (cost_list) {
const MV best_mv = { br, bc };
calc_int_cost_list(x, center_mv, sad_per_bit, vfp, &best_mv, cost_list);
}
}
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 cost_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 vp10_pattern_search_sad(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,
const int num_candidates[MAX_PATTERN_SCALES],
const MV candidates[MAX_PATTERN_SCALES]
[MAX_PATTERN_CANDIDATES]) {
// Generic pattern search function that searches over multiple scales.
// Each scale can have a different number of candidates and shape of
// candidates as indicated in the num_candidates and candidates arrays
// passed into this function
//
static int vp10_pattern_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,
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,
......@@ -1186,6 +1062,7 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
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];
const int last_is_4 = num_candidates[0] == 4;
int br, bc;
int bestsad = INT_MAX;
int thissad;
......@@ -1202,8 +1079,9 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
}
// 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) +
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
......@@ -1251,11 +1129,12 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
// 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 && cost_list != NULL);
const int last_s = (last_is_4 && cost_list != NULL);
int best_site = -1;
s = best_init_s;
for (; s >= do_sad; s--) {
for (; s >= last_s; s--) {
// No need to search all points the 1st time if initial search was used
if (!do_init_search || s != best_init_s) {
if (check_bounds(x, br, bc, 1 << s)) {
for (i = 0; i < num_candidates[s]; i++) {
......@@ -1334,9 +1213,9 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
const MV this_mv = {br + candidates[s][i].row,
bc + candidates[s][i].col};
cost_list[i + 1] =
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
CHECK_BETTER
}
} else {
......@@ -1346,9 +1225,9 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
if (!is_mv_in(x, &this_mv))
continue;
cost_list[i + 1] =
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
CHECK_BETTER
}
}
......@@ -1374,9 +1253,9 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
const MV this_mv = {br + candidates[s][next_chkpts_indices[i]].row,
bc + candidates[s][next_chkpts_indices[i]].col};
cost_list[next_chkpts_indices[i] + 1] =
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
CHECK_BETTER
}
} else {
......@@ -1388,9 +1267,9 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
continue;
}
cost_list[next_chkpts_indices[i] + 1] =
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
thissad = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
CHECK_BETTER
}
}
......@@ -1404,47 +1283,19 @@ static int vp10_pattern_search_sad(const MACROBLOCK *x,
}
}
// Returns the one-away integer pel sad values around the best as follows:
// cost_list[0]: sad at the best integer pel
// cost_list[1]: sad at delta {0, -1} (left) from the best integer pel
// cost_list[2]: sad at delta { 1, 0} (bottom) from the best integer pel
// cost_list[3]: sad at delta { 0, 1} (right) from the best integer pel
// cost_list[4]: sad at delta {-1, 0} (top) from the best integer pel
// Returns the one-away integer pel cost/sad around the best as follows:
// cost_list[0]: cost/sad at the best integer pel
// cost_list[1]: cost/sad at delta {0, -1} (left) from the best integer pel
// cost_list[2]: cost/sad at delta { 1, 0} (bottom) from the best integer pel
// cost_list[3]: cost/sad at delta { 0, 1} (right) from the best integer pel
// cost_list[4]: cost/sad at delta {-1, 0} (top) from the best integer pel
if (cost_list) {
static const MV neighbors[4] = {{0, -1}, {1, 0}, {0, 1}, {-1, 0}};
if (cost_list[0] == INT_MAX) {
cost_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 };
cost_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))
cost_list[i + 1] = INT_MAX;
else
cost_list[i + 1] = vfp->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &this_mv),
in_what->stride);
}
}
const MV best_mv = { br, bc };
if (last_is_4) {
calc_int_sad_list(x, center_mv, sad_per_bit, vfp, &best_mv, cost_list,
use_mvcost, bestsad);
} else {
if (use_mvcost) {
for (i = 0; i < 4; i++) {
const MV this_mv = {br + neighbors[i].row,
bc + neighbors[i].col};
if (cost_list[i + 1] != INT_MAX) {
cost_list[i + 1] +=
mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
}
}
}
calc_int_cost_list(x, center_mv, sad_per_bit, vfp, &best_mv, cost_list);
}
}
best_mv->row = br;
......@@ -1515,9 +1366,9 @@ int vp10_hex_search(const MACROBLOCK *x,
{ -1024, 0}},
};
return vp10_pattern_search(x, ref_mv, search_param, sad_per_bit,
do_init_search, cost_list, vfp, use_mvcost,
center_mv, best_mv,
hex_num_candidates, hex_candidates);
do_init_search, cost_list, vfp, use_mvcost,
center_mv, best_mv, hex_num_candidates,
hex_candidates);
}
int vp10_bigdia_search(const MACROBLOCK *x,
......@@ -1556,10 +1407,10 @@ int vp10_bigdia_search(const MACROBLOCK *x,
{{-512, -512}, {0, -1024}, {512, -512}, {1024, 0}, {512, 512}, {0, 1024},
{-512, 512}, {-1024, 0}},
};
return vp10_pattern_search_sad(x, ref_mv, search_param, sad_per_bit,
do_init_search, cost_list, vfp, use_mvcost,
center_mv, best_mv,
bigdia_num_candidates, bigdia_candidates);
return vp10_pattern_search(x, ref_mv, search_param, sad_per_bit,
do_init_search, cost_list, vfp, use_mvcost,
center_mv, best_mv, bigdia_num_candidates,
bigdia_candidates);
}
int vp10_square_search(const MACROBLOCK *x,
......@@ -1599,9 +1450,9 @@ int vp10_square_search(const MACROBLOCK *x,
{0, 1024}, {-1024, 1024}, {-1024, 0}},
};
return vp10_pattern_search(x, ref_mv, search_param, sad_per_bit,
do_init_search, cost_list, vfp, use_mvcost,
center_mv, best_mv,
square_num_candidates, square_candidates);
do_init_search, cost_list, vfp, use_mvcost,
center_mv, best_mv, square_num_candidates,
square_candidates);
}
int vp10_fast_hex_search(const MACROBLOCK *x,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment