generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Changes to exhaustive motion search.

Browse Source 
This change has been imported from VP9 and
alters the nature and use of exhaustive motion search.

Firstly any exhaustive search is preceded by a normal step search.
The exhaustive search is only carried out if the distortion resulting
from the step search is above a threshold value.

Secondly the simple +/- 64 exhaustive search is replaced by a
multi stage mesh based search where each stage has a range
and step/interval size. Subsequent stages use the best position from
the previous stage as the center of the search but use a reduced range
and interval size.

For example:
  stage 1: Range +/- 64 interval 4
  stage 2: Range +/- 32 interval 2
  stage 3: Range +/- 15 interval 1

This process, especially when it follows on from a normal step
search, has shown itself to be almost as effective as a full range
exhaustive search with step 1 but greatly lowers the computational
complexity such that it can be used in some cases for speeds 0-2.

This patch also removes a double exhaustive search for sub 8x8 blocks
which also contained  a bug (the two searches used different distortion
metrics).

For best quality in my test animation sequence this patch has almost
no impact on quality but improves encode speed by more than 5X.

Restricted use in good quality speeds 0-2 yields significant quality gains
on the animation test of 0.2 - 0.5 db with only a small impact on encode
speed. On most natural video clips, however, where the step search
is performing well, the quality gain and speed impact are small.

Change-Id: Iac24152ae239f42a246f39ee5f00fe62d193cb98
This commit is contained in:
paulwilkins
2015-12-08 15:48:24 +00:00
parent 9d85ce8e0c
commit 4e692bbee2
10 changed files with 253 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files

214
vp10/encoder/mcomp.c
View File

@@ -1523,69 +1523,83 @@ int vp10_fast_dia_search(const MACROBLOCK *x,
#undef CHECK_BETTER
int vp10_full_range_search_c(const MACROBLOCK *x,
const search_site_config *cfg,
MV *ref_mv, MV *best_mv,
int search_param, int sad_per_bit, int *num00,
const vp9_variance_fn_ptr_t *fn_ptr,
const MV *center_mv) {
// Exhuastive motion search around a given centre position with a given
// step size.
static int exhuastive_mesh_search(const MACROBLOCK *x,
MV *ref_mv, MV *best_mv,
int range, int step, int sad_per_bit,
const vp9_variance_fn_ptr_t *fn_ptr,
const MV *center_mv) {
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 int range = 64;
const MV fcenter_mv = {center_mv->row >> 3, center_mv->col >> 3};
MV fcenter_mv = {center_mv->row, center_mv->col};
unsigned int best_sad = INT_MAX;
int r, c, i;
int start_col, end_col, start_row, end_row;
int col_step = (step > 1) ? step : 4;
// The cfg and search_param parameters are not used in this search variant
(void)cfg;
(void)search_param;
assert(step >= 1);
clamp_mv(ref_mv, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
*best_mv = *ref_mv;
*num00 = 11;
clamp_mv(&fcenter_mv, x->mv_col_min, x->mv_col_max,
x->mv_row_min, x->mv_row_max);
*best_mv = fcenter_mv;
best_sad = fn_ptr->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);
start_row = VPXMAX(-range, x->mv_row_min - ref_mv->row);
start_col = VPXMAX(-range, x->mv_col_min - ref_mv->col);
end_row = VPXMIN(range, x->mv_row_max - ref_mv->row);
end_col = VPXMIN(range, x->mv_col_max - ref_mv->col);
get_buf_from_mv(in_what, &fcenter_mv), in_what->stride) +
mvsad_err_cost(x, &fcenter_mv, ref_mv, sad_per_bit);
start_row = VPXMAX(-range, x->mv_row_min - fcenter_mv.row);
start_col = VPXMAX(-range, x->mv_col_min - fcenter_mv.col);
end_row = VPXMIN(range, x->mv_row_max - fcenter_mv.row);
end_col = VPXMIN(range, x->mv_col_max - fcenter_mv.col);
for (r = start_row; r <= end_row; ++r) {
for (c = start_col; c <= end_col; c += 4) {
if (c + 3 <= end_col) {
unsigned int sads[4];
const uint8_t *addrs[4];
for (i = 0; i < 4; ++i) {
const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
addrs[i] = get_buf_from_mv(in_what, &mv);
}
fn_ptr->sdx4df(what->buf, what->stride, addrs, in_what->stride, sads);
for (i = 0; i < 4; ++i) {
if (sads[i] < best_sad) {
const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
const unsigned int sad = sads[i] +
mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
if (sad < best_sad) {
best_sad = sad;
*best_mv = mv;
}
for (r = start_row; r <= end_row; r += step) {
for (c = start_col; c <= end_col; c += col_step) {
// Step > 1 means we are not checking every location in this pass.
if (step > 1) {
const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c};
unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &mv), in_what->stride);
if (sad < best_sad) {
sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
if (sad < best_sad) {
best_sad = sad;
*best_mv = mv;
}
}
} else {
for (i = 0; i < end_col - c; ++i) {
const MV mv = {ref_mv->row + r, ref_mv->col + c + i};
unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &mv), in_what->stride);
if (sad < best_sad) {
sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
// 4 sads in a single call if we are checking every location
if (c + 3 <= end_col) {
unsigned int sads[4];
const uint8_t *addrs[4];
for (i = 0; i < 4; ++i) {
const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
addrs[i] = get_buf_from_mv(in_what, &mv);
}
fn_ptr->sdx4df(what->buf, what->stride, addrs,
in_what->stride, sads);
for (i = 0; i < 4; ++i) {
if (sads[i] < best_sad) {
const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
const unsigned int sad = sads[i] +
mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
if (sad < best_sad) {
best_sad = sad;
*best_mv = mv;
}
}
}
} else {
for (i = 0; i < end_col - c; ++i) {
const MV mv = {fcenter_mv.row + r, fcenter_mv.col + c + i};
unsigned int sad = fn_ptr->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &mv), in_what->stride);
if (sad < best_sad) {
best_sad = sad;
*best_mv = mv;
sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
if (sad < best_sad) {
best_sad = sad;
*best_mv = mv;
}
}
}
}
@@ -2014,6 +2028,70 @@ int vp10_full_pixel_diamond(const VP10_COMP *cpi, MACROBLOCK *x,
return bestsme;
}
#define MIN_RANGE 7
#define MAX_RANGE 256
#define MIN_INTERVAL 1
// 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 vp9_variance_fn_ptr_t *fn_ptr,
const MV *ref_mv, MV *dst_mv) {
const SPEED_FEATURES *const sf = &cpi->sf;
MV temp_mv = {centre_mv_full->row, centre_mv_full->col};
MV f_ref_mv = {ref_mv->row >> 3, ref_mv->col >> 3};
int bestsme;
int i;
int interval = sf->mesh_patterns[0].interval;
int range = sf->mesh_patterns[0].range;
int baseline_interval_divisor;
// Keep track of number of exhaustive calls (this frame in this thread).
++(*x->ex_search_count_ptr);
// Trap illegal values for interval and range for this function.
if ((range < MIN_RANGE) || (range > MAX_RANGE) ||
(interval < MIN_INTERVAL) || (interval > range))
return INT_MAX;
baseline_interval_divisor = range / interval;
// Check size of proposed first range against magnitude of the centre
// value used as a starting point.
range = VPXMAX(range, (5 * VPXMAX(abs(temp_mv.row), abs(temp_mv.col))) / 4);
range = VPXMIN(range, MAX_RANGE);
interval = VPXMAX(interval, range / baseline_interval_divisor);
// initial search
bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv, range,
interval, sadpb, fn_ptr, &temp_mv);
if ((interval > MIN_INTERVAL) && (range > MIN_RANGE)) {
// Progressive searches with range and step size decreasing each time
// till we reach a step size of 1. Then break out.
for (i = 1; i < MAX_MESH_STEP; ++i) {
// First pass with coarser step and longer range
bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv,
sf->mesh_patterns[i].range,
sf->mesh_patterns[i].interval,
sadpb, fn_ptr, &temp_mv);
if (sf->mesh_patterns[i].interval == 1)
break;
}
}
if (bestsme < INT_MAX)
bestsme = vp10_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
*dst_mv = temp_mv;
// Return cost list.
if (cost_list) {
calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
}
return bestsme;
}
int vp10_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
int sad_per_bit, int distance,
const vp9_variance_fn_ptr_t *fn_ptr,
@@ -2327,6 +2405,18 @@ int vp10_refining_search_8p_c(const MACROBLOCK *x,
return best_sad;
}
#define MIN_EX_SEARCH_LIMIT 128
static int is_exhaustive_allowed(VP10_COMP *cpi, MACROBLOCK *x) {
const SPEED_FEATURES *const sf = &cpi->sf;
const int max_ex = VPXMAX(MIN_EX_SEARCH_LIMIT,
(*x->m_search_count_ptr * sf->max_exaustive_pct) / 100);
return sf->allow_exhaustive_searches &&
(sf->exhaustive_searches_thresh < INT_MAX) &&
(*x->ex_search_count_ptr <= max_ex) &&
!cpi->rc.is_src_frame_alt_ref;
}
int vp10_full_pixel_search(VP10_COMP *cpi, MACROBLOCK *x,
BLOCK_SIZE bsize, MV *mvp_full,
int step_param, int error_per_bit,
@@ -2345,6 +2435,9 @@ int vp10_full_pixel_search(VP10_COMP *cpi, MACROBLOCK *x,
cost_list[4] = INT_MAX;
}
// Keep track of number of searches (this frame in this thread).
++(*x->m_search_count_ptr);
switch (method) {
case FAST_DIAMOND:
var = vp10_fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
@@ -2370,6 +2463,29 @@ int vp10_full_pixel_search(VP10_COMP *cpi, MACROBLOCK *x,
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);
// Should we allow a follow on exhaustive search?
if (is_exhaustive_allowed(cpi, x)) {
int64_t exhuastive_thr = sf->exhaustive_searches_thresh;
exhuastive_thr >>= 8 - (b_width_log2_lookup[bsize] +
b_height_log2_lookup[bsize]);
// Threshold variance for an exhaustive full search.
if (var > exhuastive_thr) {
int var_ex;
MV tmp_mv_ex;
var_ex = full_pixel_exhaustive(cpi, x, tmp_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;
}
}
}
break;
break;
default:
assert(0 && "Invalid search method.");