generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
Files
7ded038af55f6829de0ac9f49cab90d60a29e94d
vpx/vp10/common/thread_common.c
Geza Lore 7ded038af5 Port interintra experiment from nextgen. 
The interintra experiment, which combines an inter prediction and an
inter prediction have been ported from the nextgen branch. The
experiment is merged into ext_inter, so there is no separate configure
option to enable it.

Change-Id: I0cc20cefd29e9b77ab7bbbb709abc11512320325
2016-02-26 13:01:51 -08:00

577 lines
18 KiB
C
Raw Blame History

/*
* Copyright (c) 2014 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "./vpx_config.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_mem/vpx_mem.h"
#include "vp10/common/entropymode.h"
#include "vp10/common/thread_common.h"
#include "vp10/common/reconinter.h"
#include "vp10/common/loopfilter.h"
#if CONFIG_MULTITHREAD
static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
const int kMaxTryLocks = 4000;
int locked = 0;
int i;
for (i = 0; i < kMaxTryLocks; ++i) {
if (!pthread_mutex_trylock(mutex)) {
locked = 1;
break;
}
}
if (!locked)
pthread_mutex_lock(mutex);
}
#endif // CONFIG_MULTITHREAD
static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) {
#if CONFIG_MULTITHREAD
const int nsync = lf_sync->sync_range;
if (r && !(c & (nsync - 1))) {
pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
mutex_lock(mutex);
while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
}
pthread_mutex_unlock(mutex);
}
#else
(void)lf_sync;
(void)r;
(void)c;
#endif // CONFIG_MULTITHREAD
}
static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c,
const int sb_cols) {
#if CONFIG_MULTITHREAD
const int nsync = lf_sync->sync_range;
int cur;
// Only signal when there are enough filtered SB for next row to run.
int sig = 1;
if (c < sb_cols - 1) {
cur = c;
if (c % nsync)
sig = 0;
} else {
cur = sb_cols + nsync;
}
if (sig) {
mutex_lock(&lf_sync->mutex_[r]);
lf_sync->cur_sb_col[r] = cur;
pthread_cond_signal(&lf_sync->cond_[r]);
pthread_mutex_unlock(&lf_sync->mutex_[r]);
}
#else
(void)lf_sync;
(void)r;
(void)c;
(void)sb_cols;
#endif // CONFIG_MULTITHREAD
}
// Implement row loopfiltering for each thread.
static INLINE
void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer,
VP10_COMMON *const cm,
struct macroblockd_plane planes[MAX_MB_PLANE],
int start, int stop, int y_only,
VP9LfSync *const lf_sync) {
const int num_planes = y_only ? 1 : MAX_MB_PLANE;
const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2;
int mi_row, mi_col;
enum lf_path path;
if (y_only)
path = LF_PATH_444;
else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
path = LF_PATH_420;
else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
path = LF_PATH_444;
else
path = LF_PATH_SLOW;
for (mi_row = start; mi_row < stop;
mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) {
MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
const int r = mi_row >> MI_BLOCK_SIZE_LOG2;
const int c = mi_col >> MI_BLOCK_SIZE_LOG2;
LOOP_FILTER_MASK lfm;
int plane;
sync_read(lf_sync, r, c);
vp10_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
// TODO(JBB): Make setup_mask work for non 420.
vp10_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
&lfm);
vp10_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm);
for (plane = 1; plane < num_planes; ++plane) {
switch (path) {
case LF_PATH_420:
vp10_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm);
break;
case LF_PATH_444:
vp10_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm);
break;
case LF_PATH_SLOW:
vp10_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
mi_row, mi_col);
break;
}
}
sync_write(lf_sync, r, c, sb_cols);
}
}
}
// Row-based multi-threaded loopfilter hook
static int loop_filter_row_worker(VP9LfSync *const lf_sync,
LFWorkerData *const lf_data) {
thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
lf_data->start, lf_data->stop, lf_data->y_only,
lf_sync);
return 1;
}
static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame,
VP10_COMMON *cm,
struct macroblockd_plane planes[MAX_MB_PLANE],
int start, int stop, int y_only,
VPxWorker *workers, int nworkers,
VP9LfSync *lf_sync) {
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
// Number of superblock rows and cols
const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2;
// Decoder may allocate more threads than number of tiles based on user's
// input.
const int tile_cols = 1 << cm->log2_tile_cols;
const int num_workers = VPXMIN(nworkers, tile_cols);
int i;
if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
num_workers > lf_sync->num_workers) {
vp10_loop_filter_dealloc(lf_sync);
vp10_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
}
// Initialize cur_sb_col to -1 for all SB rows.
memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
// Set up loopfilter thread data.
// The decoder is capping num_workers because it has been observed that using
// more threads on the loopfilter than there are cores will hurt performance
// on Android. This is because the system will only schedule the tile decode
// workers on cores equal to the number of tile columns. Then if the decoder
// tries to use more threads for the loopfilter, it will hurt performance
// because of contention. If the multithreading code changes in the future
// then the number of workers used by the loopfilter should be revisited.
for (i = 0; i < num_workers; ++i) {
VPxWorker *const worker = &workers[i];
LFWorkerData *const lf_data = &lf_sync->lfdata[i];
worker->hook = (VPxWorkerHook)loop_filter_row_worker;
worker->data1 = lf_sync;
worker->data2 = lf_data;
// Loopfilter data
vp10_loop_filter_data_reset(lf_data, frame, cm, planes);
lf_data->start = start + i * MI_BLOCK_SIZE;
lf_data->stop = stop;
lf_data->y_only = y_only;
// Start loopfiltering
if (i == num_workers - 1) {
winterface->execute(worker);
} else {
winterface->launch(worker);
}
}
// Wait till all rows are finished
for (i = 0; i < num_workers; ++i) {
winterface->sync(&workers[i]);
}
}
void vp10_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
VP10_COMMON *cm,
struct macroblockd_plane planes[MAX_MB_PLANE],
int frame_filter_level,
int y_only, int partial_frame,
VPxWorker *workers, int num_workers,
VP9LfSync *lf_sync) {
int start_mi_row, end_mi_row, mi_rows_to_filter;
if (!frame_filter_level) return;
start_mi_row = 0;
mi_rows_to_filter = cm->mi_rows;
if (partial_frame && cm->mi_rows > 8) {
start_mi_row = cm->mi_rows >> 1;
start_mi_row &= 0xfffffff8;
mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
}
end_mi_row = start_mi_row + mi_rows_to_filter;
vp10_loop_filter_frame_init(cm, frame_filter_level);
loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row,
y_only, workers, num_workers, lf_sync);
}
// Set up nsync by width.
static INLINE int get_sync_range(int width) {
// nsync numbers are picked by testing. For example, for 4k
// video, using 4 gives best performance.
if (width < 640)
return 1;
else if (width <= 1280)
return 2;
else if (width <= 4096)
return 4;
else
return 8;
}
// Allocate memory for lf row synchronization
void vp10_loop_filter_alloc(VP9LfSync *lf_sync, VP10_COMMON *cm, int rows,
int width, int num_workers) {
lf_sync->rows = rows;
#if CONFIG_MULTITHREAD
{
int i;
CHECK_MEM_ERROR(cm, lf_sync->mutex_,
vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
if (lf_sync->mutex_) {
for (i = 0; i < rows; ++i) {
pthread_mutex_init(&lf_sync->mutex_[i], NULL);
}
}
CHECK_MEM_ERROR(cm, lf_sync->cond_,
vpx_malloc(sizeof(*lf_sync->cond_) * rows));
if (lf_sync->cond_) {
for (i = 0; i < rows; ++i) {
pthread_cond_init(&lf_sync->cond_[i], NULL);
}
}
}
#endif // CONFIG_MULTITHREAD
CHECK_MEM_ERROR(cm, lf_sync->lfdata,
vpx_malloc(num_workers * sizeof(*lf_sync->lfdata)));
lf_sync->num_workers = num_workers;
CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
// Set up nsync.
lf_sync->sync_range = get_sync_range(width);
}
// Deallocate lf synchronization related mutex and data
void vp10_loop_filter_dealloc(VP9LfSync *lf_sync) {
if (lf_sync != NULL) {
#if CONFIG_MULTITHREAD
int i;
if (lf_sync->mutex_ != NULL) {
for (i = 0; i < lf_sync->rows; ++i) {
pthread_mutex_destroy(&lf_sync->mutex_[i]);
}
vpx_free(lf_sync->mutex_);
}
if (lf_sync->cond_ != NULL) {
for (i = 0; i < lf_sync->rows; ++i) {
pthread_cond_destroy(&lf_sync->cond_[i]);
}
vpx_free(lf_sync->cond_);
}
#endif // CONFIG_MULTITHREAD
vpx_free(lf_sync->lfdata);
vpx_free(lf_sync->cur_sb_col);
// clear the structure as the source of this call may be a resize in which
// case this call will be followed by an _alloc() which may fail.
vp10_zero(*lf_sync);
}
}
// Accumulate frame counts.
void vp10_accumulate_frame_counts(VP10_COMMON *cm, FRAME_COUNTS *counts,
int is_dec) {
int i, j, k, l, m;
for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
for (j = 0; j < INTRA_MODES; j++)
cm->counts.y_mode[i][j] += counts->y_mode[i][j];
for (i = 0; i < INTRA_MODES; i++)
for (j = 0; j < INTRA_MODES; j++)
cm->counts.uv_mode[i][j] += counts->uv_mode[i][j];
for (i = 0; i < PARTITION_CONTEXTS; i++)
for (j = 0; j < PARTITION_TYPES; j++)
cm->counts.partition[i][j] += counts->partition[i][j];
if (is_dec) {
int n;
for (i = 0; i < TX_SIZES; i++)
for (j = 0; j < PLANE_TYPES; j++)
for (k = 0; k < REF_TYPES; k++)
for (l = 0; l < COEF_BANDS; l++)
for (m = 0; m < COEFF_CONTEXTS; m++) {
cm->counts.eob_branch[i][j][k][l][m] +=
counts->eob_branch[i][j][k][l][m];
for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
cm->counts.coef[i][j][k][l][m][n] +=
counts->coef[i][j][k][l][m][n];
}
} else {
for (i = 0; i < TX_SIZES; i++)
for (j = 0; j < PLANE_TYPES; j++)
for (k = 0; k < REF_TYPES; k++)
for (l = 0; l < COEF_BANDS; l++)
for (m = 0; m < COEFF_CONTEXTS; m++)
cm->counts.eob_branch[i][j][k][l][m] +=
counts->eob_branch[i][j][k][l][m];
// In the encoder, cm->counts.coef is only updated at frame
// level, so not need to accumulate it here.
// for (n = 0; n < UNCONSTRAINED_NODES + 1; n++)
// cm->counts.coef[i][j][k][l][m][n] +=
// counts->coef[i][j][k][l][m][n];
}
for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
for (j = 0; j < SWITCHABLE_FILTERS; j++)
cm->counts.switchable_interp[i][j] += counts->switchable_interp[i][j];
#if CONFIG_REF_MV
for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
for (j = 0; j < 2; ++j)
cm->counts.newmv_mode[i][j] += counts->newmv_mode[i][j];
for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
for (j = 0; j < 2; ++j)
cm->counts.zeromv_mode[i][j] += counts->zeromv_mode[i][j];
for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
for (j = 0; j < 2; ++j)
cm->counts.refmv_mode[i][j] += counts->refmv_mode[i][j];
for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
for (j = 0; j < 2; ++j)
cm->counts.drl_mode0[i][j] += counts->drl_mode0[i][j];
for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
for (j = 0; j < 2; ++j)
cm->counts.drl_mode1[i][j] += counts->drl_mode1[i][j];
#if CONFIG_EXT_INTER
for (j = 0; j < 2; ++j)
cm->counts.new2mv_mode[j] += counts->new2mv_mode[j];
#endif // CONFIG_EXT_INTER
#endif
for (i = 0; i < INTER_MODE_CONTEXTS; i++)
for (j = 0; j < INTER_MODES; j++)
cm->counts.inter_mode[i][j] += counts->inter_mode[i][j];
#if CONFIG_EXT_INTER
for (i = 0; i < INTER_MODE_CONTEXTS; i++)
for (j = 0; j < INTER_COMPOUND_MODES; j++)
cm->counts.inter_compound_mode[i][j] +=
counts->inter_compound_mode[i][j];
for (i = 0; i < BLOCK_SIZES; i++)
for (j = 0; j < 2; j++)
cm->counts.interintra[i][j] += counts->interintra[i][j];
#endif // CONFIG_EXT_INTER
for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
for (j = 0; j < 2; j++)
cm->counts.intra_inter[i][j] += counts->intra_inter[i][j];
for (i = 0; i < COMP_INTER_CONTEXTS; i++)
for (j = 0; j < 2; j++)
cm->counts.comp_inter[i][j] += counts->comp_inter[i][j];
for (i = 0; i < REF_CONTEXTS; i++)
for (j = 0; j < (SINGLE_REFS - 1); j++)
for (k = 0; k < 2; k++)
cm->counts.single_ref[i][j][k] += counts->single_ref[i][j][k];
for (i = 0; i < REF_CONTEXTS; i++)
for (j = 0; j < (COMP_REFS - 1); j++)
for (k = 0; k < 2; k++)
cm->counts.comp_ref[i][j][k] += counts->comp_ref[i][j][k];
for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
for (j = 0; j < TX_SIZES; j++)
cm->counts.tx.p32x32[i][j] += counts->tx.p32x32[i][j];
for (j = 0; j < TX_SIZES - 1; j++)
cm->counts.tx.p16x16[i][j] += counts->tx.p16x16[i][j];
for (j = 0; j < TX_SIZES - 2; j++)
cm->counts.tx.p8x8[i][j] += counts->tx.p8x8[i][j];
}
for (i = 0; i < TX_SIZES; i++)
cm->counts.tx.tx_totals[i] += counts->tx.tx_totals[i];
#if CONFIG_VAR_TX
for (i = 0; i < TXFM_PARTITION_CONTEXTS; ++i)
for (j = 0; j < 2; ++j)
cm->counts.txfm_partition[i][j] += counts->txfm_partition[i][j];
#endif
for (i = 0; i < SKIP_CONTEXTS; i++)
for (j = 0; j < 2; j++)
cm->counts.skip[i][j] += counts->skip[i][j];
#if CONFIG_REF_MV
for (m = 0; m < NMV_CONTEXTS; ++m) {
for (i = 0; i < MV_JOINTS; i++)
cm->counts.mv[m].joints[i] += counts->mv[m].joints[i];
for (k = 0; k < 2; k++) {
nmv_component_counts *comps = &cm->counts.mv[m].comps[k];
nmv_component_counts *comps_t = &counts->mv[m].comps[k];
for (i = 0; i < 2; i++) {
comps->sign[i] += comps_t->sign[i];
comps->class0_hp[i] += comps_t->class0_hp[i];
comps->hp[i] += comps_t->hp[i];
}
for (i = 0; i < MV_CLASSES; i++)
comps->classes[i] += comps_t->classes[i];
for (i = 0; i < CLASS0_SIZE; i++) {
comps->class0[i] += comps_t->class0[i];
for (j = 0; j < MV_FP_SIZE; j++)
comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
}
for (i = 0; i < MV_OFFSET_BITS; i++)
for (j = 0; j < 2; j++)
comps->bits[i][j] += comps_t->bits[i][j];
for (i = 0; i < MV_FP_SIZE; i++)
comps->fp[i] += comps_t->fp[i];
}
}
#else
for (i = 0; i < MV_JOINTS; i++)
cm->counts.mv.joints[i] += counts->mv.joints[i];
for (k = 0; k < 2; k++) {
nmv_component_counts *comps = &cm->counts.mv.comps[k];
nmv_component_counts *comps_t = &counts->mv.comps[k];
for (i = 0; i < 2; i++) {
comps->sign[i] += comps_t->sign[i];
comps->class0_hp[i] += comps_t->class0_hp[i];
comps->hp[i] += comps_t->hp[i];
}
for (i = 0; i < MV_CLASSES; i++)
comps->classes[i] += comps_t->classes[i];
for (i = 0; i < CLASS0_SIZE; i++) {
comps->class0[i] += comps_t->class0[i];
for (j = 0; j < MV_FP_SIZE; j++)
comps->class0_fp[i][j] += comps_t->class0_fp[i][j];
}
for (i = 0; i < MV_OFFSET_BITS; i++)
for (j = 0; j < 2; j++)
comps->bits[i][j] += comps_t->bits[i][j];
for (i = 0; i < MV_FP_SIZE; i++)
comps->fp[i] += comps_t->fp[i];
}
#endif
#if CONFIG_EXT_TX
for (i = 0; i < EXT_TX_SIZES; i++) {
int s, k;
for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
if (use_inter_ext_tx_for_txsize[s][i]) {
for (k = 0; k < TX_TYPES; k++)
cm->counts.inter_ext_tx[s][i][k] += counts->inter_ext_tx[s][i][k];
}
}
for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
if (use_intra_ext_tx_for_txsize[s][i]) {
int j;
for (j = 0; j < INTRA_MODES; ++j)
for (k = 0; k < TX_TYPES; k++)
cm->counts.intra_ext_tx[s][i][j][k] +=
counts->intra_ext_tx[s][i][j][k];
}
}
}
#else
for (i = 0; i < EXT_TX_SIZES; i++) {
int j;
for (j = 0; j < TX_TYPES; ++j)
for (k = 0; k < TX_TYPES; k++)
cm->counts.intra_ext_tx[i][j][k] += counts->intra_ext_tx[i][j][k];
}
for (i = 0; i < EXT_TX_SIZES; i++) {
for (k = 0; k < TX_TYPES; k++)
cm->counts.inter_ext_tx[i][k] += counts->inter_ext_tx[i][k];
}
#endif // CONFIG_EXT_TX
#if CONFIG_SUPERTX
for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; i++)
for (j = 0; j < TX_SIZES; j++)
for (k = 0; k < 2; k++)
cm->counts.supertx[i][j][k] += counts->supertx[i][j][k];
for (i = 0; i < TX_SIZES; i++)
cm->counts.supertx_size[i] += counts->supertx_size[i];
#endif // CONFIG_SUPERTX
for (i = 0; i < PREDICTION_PROBS; i++)
for (j = 0; j < 2; j++)
cm->counts.seg.pred[i][j] += counts->seg.pred[i][j];
for (i = 0; i < MAX_SEGMENTS; i++) {
cm->counts.seg.tree_total[i] += counts->seg.tree_total[i];
cm->counts.seg.tree_mispred[i] += counts->seg.tree_mispred[i];
}
#if CONFIG_EXT_INTRA
for (i = 0; i < PLANE_TYPES; ++i)
for (j = 0; j < 2; ++j)
cm->counts.ext_intra[i][j] += counts->ext_intra[i][j];
for (i = 0; i < INTRA_FILTERS + 1; ++i)
for (j = 0; j < INTRA_FILTERS; ++j)
cm->counts.intra_filter[i][j] += counts->intra_filter[i][j];
#endif // CONFIG_EXT_INTRA
}
Reference in New Issue View Git Blame Copy Permalink