// Setup cyclic background refresh: set delta q and segmentation map.
void vp9_cyclic_refresh_setup(VP9_COMP *const cpi) {
VP9_COMMON *const cm = &cpi->common;
const RATE_CONTROL *const rc = &cpi->rc;
CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
struct segmentation *const seg = &cm->seg;
const int apply_cyclic_refresh = apply_cyclic_refresh_bitrate(cm, rc);
if (cm->current_video_frame == 0)
cr->low_content_avg = 0.0;
// Don't apply refresh on key frame or temporal enhancement layer frames.
if (!apply_cyclic_refresh ||
(cm->frame_type == KEY_FRAME) ||
(cpi->svc.temporal_layer_id > 0)) {
// Set segmentation map to 0 and disable.
unsigned char *const seg_map = cpi->segmentation_map;
memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
vp9_disable_segmentation(&cm->seg);
if (cm->frame_type == KEY_FRAME) {
memset(cr->last_coded_q_map, MAXQ,
cm->mi_rows * cm->mi_cols * sizeof(*cr->last_coded_q_map));
memset(cr->consec_zero_mv, 0,
cm->mi_rows * cm->mi_cols * sizeof(*cr->consec_zero_mv));
cr->sb_index = 0;
}
return;
} else {
int qindex_delta = 0;
int qindex2;
const double q = vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth);
vpx_clear_system_state();
// Set rate threshold to some multiple (set to 2 for now) of the target
// rate (target is given by sb64_target_rate and scaled by 256).
cr->thresh_rate_sb = ((int64_t)(rc->sb64_target_rate) << 8) << 2;
// Distortion threshold, quadratic in Q, scale factor to be adjusted.
// q will not exceed 457, so (q * q) is within 32bit; see:
// vp9_convert_qindex_to_q(), vp9_ac_quant(), ac_qlookup*[].
cr->thresh_dist_sb = ((int64_t)(q * q)) << 2;
// Set up segmentation.
// Clear down the segment map.
vp9_enable_segmentation(&cm->seg);
vp9_clearall_segfeatures(seg);
// Select delta coding method.
seg->abs_delta = SEGMENT_DELTADATA;
// Note: setting temporal_update has no effect, as the seg-map coding method
// (temporal or spatial) is determined in vp9_choose_segmap_coding_method(),
// based on the coding cost of each method. For error_resilient mode on the
// last_frame_seg_map is set to 0, so if temporal coding is used, it is
// relative to 0 previous map.
// seg->temporal_update = 0;
// Segment BASE "Q" feature is disabled so it defaults to the baseline Q.
vp9_disable_segfeature(seg, CR_SEGMENT_ID_BASE, SEG_LVL_ALT_Q);
// Use segment BOOST1 for in-frame Q adjustment.
vp9_enable_segfeature(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q);
// Use segment BOOST2 for more aggressive in-frame Q adjustment.
vp9_enable_segfeature(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q);
// Set the q delta for segment BOOST1.
qindex_delta = compute_deltaq(cpi, cm->base_qindex, cr->rate_ratio_qdelta);
cr->qindex_delta[1] = qindex_delta;
// Compute rd-mult for segment BOOST1.
qindex2 = clamp(cm->base_qindex + cm->y_dc_delta_q + qindex_delta, 0, MAXQ);
cr->rdmult = vp9_compute_rd_mult(cpi, qindex2);
vp9_set_segdata(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q, qindex_delta);
// Set a more aggressive (higher) q delta for segment BOOST2.
qindex_delta = compute_deltaq(
cpi, cm->base_qindex,
VPXMIN(CR_MAX_RATE_TARGET_RATIO,
0.1 * cr->rate_boost_fac * cr->rate_ratio_qdelta));
cr->qindex_delta[2] = qindex_delta;
vp9_set_segdata(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q, qindex_delta);
// Reset if resoluton change has occurred.
if (cpi->resize_pending != 0)
vp9_cyclic_refresh_reset_resize(cpi);
// Update the segmentation and refresh map.
cyclic_refresh_update_map(cpi);
}
}
static void accumulate_fp_tile_stat(TileDataEnc *tile_data,
TileDataEnc *tile_data_t) {
tile_data->fp_data.intra_factor += tile_data_t->fp_data.intra_factor;
tile_data->fp_data.brightness_factor +=
tile_data_t->fp_data.brightness_factor;
tile_data->fp_data.coded_error += tile_data_t->fp_data.coded_error;
tile_data->fp_data.sr_coded_error += tile_data_t->fp_data.sr_coded_error;
tile_data->fp_data.frame_noise_energy +=
tile_data_t->fp_data.frame_noise_energy;
tile_data->fp_data.intra_error += tile_data_t->fp_data.intra_error;
tile_data->fp_data.intercount += tile_data_t->fp_data.intercount;
tile_data->fp_data.second_ref_count += tile_data_t->fp_data.second_ref_count;
tile_data->fp_data.neutral_count += tile_data_t->fp_data.neutral_count;
tile_data->fp_data.intra_count_low += tile_data_t->fp_data.intra_count_low;
tile_data->fp_data.intra_count_high += tile_data_t->fp_data.intra_count_high;
tile_data->fp_data.intra_skip_count += tile_data_t->fp_data.intra_skip_count;
tile_data->fp_data.mvcount += tile_data_t->fp_data.mvcount;
tile_data->fp_data.sum_mvr += tile_data_t->fp_data.sum_mvr;
tile_data->fp_data.sum_mvr_abs += tile_data_t->fp_data.sum_mvr_abs;
tile_data->fp_data.sum_mvc += tile_data_t->fp_data.sum_mvc;
tile_data->fp_data.sum_mvc_abs += tile_data_t->fp_data.sum_mvc_abs;
tile_data->fp_data.sum_mvrs += tile_data_t->fp_data.sum_mvrs;
tile_data->fp_data.sum_mvcs += tile_data_t->fp_data.sum_mvcs;
tile_data->fp_data.sum_in_vectors += tile_data_t->fp_data.sum_in_vectors;
tile_data->fp_data.intra_smooth_count +=
tile_data_t->fp_data.intra_smooth_count;
tile_data->fp_data.image_data_start_row =
VPXMIN(tile_data->fp_data.image_data_start_row,
tile_data_t->fp_data.image_data_start_row) == INVALID_ROW
? VPXMAX(tile_data->fp_data.image_data_start_row,
tile_data_t->fp_data.image_data_start_row)
: VPXMIN(tile_data->fp_data.image_data_start_row,
tile_data_t->fp_data.image_data_start_row);
}
void vp9_cyclic_refresh_update_sb_postencode(VP9_COMP *const cpi,
const MODE_INFO *const mi,
int mi_row, int mi_col,
BLOCK_SIZE bsize) {
const VP9_COMMON *const cm = &cpi->common;
CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
const int bw = num_8x8_blocks_wide_lookup[bsize];
const int bh = num_8x8_blocks_high_lookup[bsize];
const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
const int block_index = mi_row * cm->mi_cols + mi_col;
int x, y;
for (y = 0; y < ymis; y++)
for (x = 0; x < xmis; x++) {
int map_offset = block_index + y * cm->mi_cols + x;
// Inter skip blocks were clearly not coded at the current qindex, so
// don't update the map for them. For cases where motion is non-zero or
// the reference frame isn't the previous frame, the previous value in
// the map for this spatial location is not entirely correct.
if ((!is_inter_block(mi) || !mi->skip) &&
mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
cr->last_coded_q_map[map_offset] =
clamp(cm->base_qindex + cr->qindex_delta[mi->segment_id], 0, MAXQ);
} else if (is_inter_block(mi) && mi->skip &&
mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
cr->last_coded_q_map[map_offset] = VPXMIN(
clamp(cm->base_qindex + cr->qindex_delta[mi->segment_id], 0, MAXQ),
cr->last_coded_q_map[map_offset]);
}
}
}
static void set_rt_speed_feature_framesize_dependent(VP9_COMP *cpi,
SPEED_FEATURES *sf, int speed) {
VP9_COMMON *const cm = &cpi->common;
if (speed >= 1) {
if (VPXMIN(cm->width, cm->height) >= 720) {
sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
: DISABLE_ALL_INTER_SPLIT;
} else {
sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
}
}
if (speed >= 2) {
if (VPXMIN(cm->width, cm->height) >= 720) {
sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
: DISABLE_ALL_INTER_SPLIT;
} else {
sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
}
}
if (speed >= 5) {
if (VPXMIN(cm->width, cm->height) >= 720) {
sf->partition_search_breakout_dist_thr = (1 << 25);
} else {
sf->partition_search_breakout_dist_thr = (1 << 23);
}
}
if (speed >= 7) {
sf->encode_breakout_thresh = (VPXMIN(cm->width, cm->height) >= 720) ?
800 : 300;
}
}
// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
// check if we should reset the segment_id, and update the cyclic_refresh map
// and segmentation map.
void vp9_cyclic_refresh_update_segment(VP9_COMP *const cpi,
MB_MODE_INFO *const mbmi,
int mi_row, int mi_col,
BLOCK_SIZE bsize,
int64_t rate,
int64_t dist,
int skip) {
const VP9_COMMON *const cm = &cpi->common;
CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
const int bw = num_8x8_blocks_wide_lookup[bsize];
const int bh = num_8x8_blocks_high_lookup[bsize];
const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
const int block_index = mi_row * cm->mi_cols + mi_col;
const int refresh_this_block = candidate_refresh_aq(cr, mbmi, rate, dist,
bsize);
// Default is to not update the refresh map.
int new_map_value = cr->map[block_index];
int x = 0; int y = 0;
// If this block is labeled for refresh, check if we should reset the
// segment_id.
if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
mbmi->segment_id = refresh_this_block;
// Reset segment_id if it will be skipped.
if (skip)
mbmi->segment_id = CR_SEGMENT_ID_BASE;
}
// Update the cyclic refresh map, to be used for setting segmentation map
// for the next frame. If the block will be refreshed this frame, mark it
// as clean. The magnitude of the -ve influences how long before we consider
// it for refresh again.
if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
new_map_value = -cr->time_for_refresh;
} else if (refresh_this_block) {
// Else if it is accepted as candidate for refresh, and has not already
// been refreshed (marked as 1) then mark it as a candidate for cleanup
// for future time (marked as 0), otherwise don't update it.
if (cr->map[block_index] == 1)
new_map_value = 0;
} else {
// Leave it marked as block that is not candidate for refresh.
new_map_value = 1;
}
// Update entries in the cyclic refresh map with new_map_value, and
// copy mbmi->segment_id into global segmentation map.
for (y = 0; y < ymis; y++)
for (x = 0; x < xmis; x++) {
int map_offset = block_index + y * cm->mi_cols + x;
cr->map[map_offset] = new_map_value;
cpi->segmentation_map[map_offset] = mbmi->segment_id;
}
}
// Select a segment for the current block.
// The choice of segment for a block depends on the ratio of the projected
// bits for the block vs a target average and its spatial complexity.
void vp10_caq_select_segment(VP10_COMP *cpi, MACROBLOCK *mb, BLOCK_SIZE bs,
int mi_row, int mi_col, int projected_rate) {
VP10_COMMON *const cm = &cpi->common;
const int mi_offset = mi_row * cm->mi_cols + mi_col;
const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
const int xmis = VPXMIN(cm->mi_cols - mi_col, num_8x8_blocks_wide_lookup[bs]);
const int ymis = VPXMIN(cm->mi_rows - mi_row, num_8x8_blocks_high_lookup[bs]);
int x, y;
int i;
unsigned char segment;
if (0) {
segment = DEFAULT_AQ2_SEG;
} else {
// Rate depends on fraction of a SB64 in frame (xmis * ymis / bw * bh).
// It is converted to bits * 256 units.
const int target_rate =
(cpi->rc.sb64_target_rate * xmis * ymis * 256) / (bw * bh);
double logvar;
double low_var_thresh;
const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
vpx_clear_system_state();
low_var_thresh = (cpi->oxcf.pass == 2) ? VPXMAX(cpi->twopass.mb_av_energy,
MIN_DEFAULT_LV_THRESH)
: DEFAULT_LV_THRESH;
vp10_setup_src_planes(mb, cpi->Source, mi_row, mi_col);
logvar = vp10_log_block_var(cpi, mb, bs);
segment = AQ_C_SEGMENTS - 1; // Just in case no break out below.
for (i = 0; i < AQ_C_SEGMENTS; ++i) {
// Test rate against a threshold value and variance against a threshold.
// Increasing segment number (higher variance and complexity) = higher Q.
if ((projected_rate < target_rate * aq_c_transitions[aq_strength][i]) &&
(logvar < (low_var_thresh + aq_c_var_thresholds[aq_strength][i]))) {
segment = i;
break;
}
}
}
// Fill in the entires in the segment map corresponding to this SB64.
for (y = 0; y < ymis; y++) {
for (x = 0; x < xmis; x++) {
cpi->segmentation_map[mi_offset + y * cm->mi_cols + x] = segment;
}
}
}
static void read_intra_frame_mode_info(VP10_COMMON *const cm,
MACROBLOCKD *const xd,
int mi_row, int mi_col, vpx_reader *r) {
MODE_INFO *const mi = xd->mi[0];
MB_MODE_INFO *const mbmi = &mi->mbmi;
const MODE_INFO *above_mi = xd->above_mi;
const MODE_INFO *left_mi = xd->left_mi;
const BLOCK_SIZE bsize = mbmi->sb_type;
int i;
const int mi_offset = mi_row * cm->mi_cols + mi_col;
const int bw = xd->plane[0].n4_w >> 1;
const int bh = xd->plane[0].n4_h >> 1;
// TODO(slavarnway): move x_mis, y_mis into xd ?????
const int x_mis = VPXMIN(cm->mi_cols - mi_col, bw);
const int y_mis = VPXMIN(cm->mi_rows - mi_row, bh);
mbmi->segment_id = read_intra_segment_id(cm, mi_offset, x_mis, y_mis, r);
mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
mbmi->tx_size = read_tx_size(cm, xd, 1, r);
mbmi->ref_frame[0] = INTRA_FRAME;
mbmi->ref_frame[1] = NONE;
switch (bsize) {
case BLOCK_4X4:
for (i = 0; i < 4; ++i)
mi->bmi[i].as_mode =
read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, i));
mbmi->mode = mi->bmi[3].as_mode;
break;
case BLOCK_4X8:
mi->bmi[0].as_mode = mi->bmi[2].as_mode =
read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 0));
mi->bmi[1].as_mode = mi->bmi[3].as_mode = mbmi->mode =
read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 1));
break;
case BLOCK_8X4:
mi->bmi[0].as_mode = mi->bmi[1].as_mode =
read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 0));
mi->bmi[2].as_mode = mi->bmi[3].as_mode = mbmi->mode =
read_intra_mode(r, get_y_mode_probs(mi, above_mi, left_mi, 2));
break;
default:
mbmi->mode = read_intra_mode(r,
get_y_mode_probs(mi, above_mi, left_mi, 0));
}
mbmi->uv_mode = read_intra_mode(r, vp10_kf_uv_mode_prob[mbmi->mode]);
}
void vpx_reader_fill(vpx_reader *r) {
const uint8_t *const buffer_end = r->buffer_end;
const uint8_t *buffer = r->buffer;
const uint8_t *buffer_start = buffer;
BD_VALUE value = r->value;
int count = r->count;
const size_t bytes_left = buffer_end - buffer;
const size_t bits_left = bytes_left * CHAR_BIT;
int shift = BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
if (r->decrypt_cb) {
size_t n = VPXMIN(sizeof(r->clear_buffer), bytes_left);
r->decrypt_cb(r->decrypt_state, buffer, r->clear_buffer, (int)n);
buffer = r->clear_buffer;
buffer_start = r->clear_buffer;
}
if (bits_left > BD_VALUE_SIZE) {
const int bits = (shift & 0xfffffff8) + CHAR_BIT;
BD_VALUE nv;
BD_VALUE big_endian_values;
memcpy(&big_endian_values, buffer, sizeof(BD_VALUE));
#if SIZE_MAX == 0xffffffffffffffffULL
big_endian_values = HToBE64(big_endian_values);
#else
big_endian_values = HToBE32(big_endian_values);
#endif
nv = big_endian_values >> (BD_VALUE_SIZE - bits);
count += bits;
buffer += (bits >> 3);
value = r->value | (nv << (shift & 0x7));
} else {
static void create_enc_workers(VP9_COMP *cpi, int num_workers) {
VP9_COMMON *const cm = &cpi->common;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
int i;
// Only run once to create threads and allocate thread data.
if (cpi->num_workers == 0) {
int allocated_workers = num_workers;
// While using SVC, we need to allocate threads according to the highest
// resolution. When row based multithreading is enabled, it is OK to
// allocate more threads than the number of max tile columns.
if (cpi->use_svc && !cpi->row_mt) {
int max_tile_cols = get_max_tile_cols(cpi);
allocated_workers = VPXMIN(cpi->oxcf.max_threads, max_tile_cols);
}
CHECK_MEM_ERROR(cm, cpi->workers,
vpx_malloc(allocated_workers * sizeof(*cpi->workers)));
CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
vpx_calloc(allocated_workers, sizeof(*cpi->tile_thr_data)));
for (i = 0; i < allocated_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
winterface->init(worker);
if (i < allocated_workers - 1) {
thread_data->cpi = cpi;
// Allocate thread data.
CHECK_MEM_ERROR(cm, thread_data->td,
vpx_memalign(32, sizeof(*thread_data->td)));
vp9_zero(*thread_data->td);
// Set up pc_tree.
thread_data->td->leaf_tree = NULL;
thread_data->td->pc_tree = NULL;
vp9_setup_pc_tree(cm, thread_data->td);
// Allocate frame counters in thread data.
CHECK_MEM_ERROR(cm, thread_data->td->counts,
vpx_calloc(1, sizeof(*thread_data->td->counts)));
// Create threads
if (!winterface->reset(worker))
vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
"Tile encoder thread creation failed");
} else {
// Main thread acts as a worker and uses the thread data in cpi.
thread_data->cpi = cpi;
thread_data->td = &cpi->td;
}
winterface->sync(worker);
}
}
}
static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
const MV *ref_mv,
MV *dst_mv,
int mb_row,
int mb_col) {
MACROBLOCK *const x = &cpi->td.mb;
MACROBLOCKD *const xd = &x->e_mbd;
MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
const SEARCH_METHODS old_search_method = mv_sf->search_method;
const vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
const int tmp_col_min = x->mv_col_min;
const int tmp_col_max = x->mv_col_max;
const int tmp_row_min = x->mv_row_min;
const int tmp_row_max = x->mv_row_max;
MV ref_full;
int cost_list[5];
// Further step/diamond searches as necessary
int step_param = mv_sf->reduce_first_step_size;
step_param = VPXMIN(step_param, MAX_MVSEARCH_STEPS - 2);
vp9_set_mv_search_range(x, ref_mv);
ref_full.col = ref_mv->col >> 3;
ref_full.row = ref_mv->row >> 3;
mv_sf->search_method = HEX;
vp9_full_pixel_search(cpi, x, BLOCK_16X16, &ref_full, step_param,
x->errorperbit, cond_cost_list(cpi, cost_list), ref_mv,
dst_mv, 0, 0);
mv_sf->search_method = old_search_method;
// Try sub-pixel MC
// if (bestsme > error_thresh && bestsme < INT_MAX)
{
uint32_t distortion;
uint32_t sse;
cpi->find_fractional_mv_step(
x, dst_mv, 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,
&distortion, &sse, NULL, 0, 0);
}
xd->mi[0]->mode = NEWMV;
xd->mi[0]->mv[0].as_mv = *dst_mv;
vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
/* restore UMV window */
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
x->mv_row_min = tmp_row_min;
x->mv_row_max = tmp_row_max;
return vpx_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
xd->plane[0].dst.buf, xd->plane[0].dst.stride);
}
int vp9_compute_rd_mult(const VP9_COMP *cpi, int qindex) {
const int64_t q = vp9_dc_quant(qindex, 0, cpi->common.bit_depth);
#if CONFIG_VP9_HIGHBITDEPTH
int64_t rdmult = 0;
switch (cpi->common.bit_depth) {
case VPX_BITS_8:
rdmult = 88 * q * q / 24;
break;
case VPX_BITS_10:
rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 4);
break;
case VPX_BITS_12:
rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 8);
break;
default:
assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
return -1;
}
#else
int64_t rdmult = 88 * q * q / 24;
#endif // CONFIG_VP9_HIGHBITDEPTH
if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
const int boost_index = VPXMIN(15, (cpi->rc.gfu_boost / 100));
rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
}
if (rdmult < 1)
rdmult = 1;
return (int)rdmult;
}
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]);
}
}
static TX_SIZE read_tx_size(VP10_COMMON *cm, MACROBLOCKD *xd,
int allow_select, vpx_reader *r) {
TX_MODE tx_mode = cm->tx_mode;
BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
if (allow_select && tx_mode == TX_MODE_SELECT && bsize >= BLOCK_8X8)
return read_selected_tx_size(cm, xd, max_tx_size, r);
else
return VPXMIN(max_tx_size, tx_mode_to_biggest_tx_size[tx_mode]);
}
static vpx_codec_err_t vp8_peek_si_internal(const uint8_t *data,
unsigned int data_sz,
vpx_codec_stream_info_t *si,
vpx_decrypt_cb decrypt_cb,
void *decrypt_state)
{
vpx_codec_err_t res = VPX_CODEC_OK;
assert(data != NULL);
if(data + data_sz <= data)
{
res = VPX_CODEC_INVALID_PARAM;
}
else
{
/* Parse uncompresssed part of key frame header.
* 3 bytes:- including version, frame type and an offset
* 3 bytes:- sync code (0x9d, 0x01, 0x2a)
* 4 bytes:- including image width and height in the lowest 14 bits
* of each 2-byte value.
*/
uint8_t clear_buffer[10];
const uint8_t *clear = data;
if (decrypt_cb)
{
int n = VPXMIN(sizeof(clear_buffer), data_sz);
decrypt_cb(decrypt_state, data, clear_buffer, n);
clear = clear_buffer;
}
si->is_kf = 0;
if (data_sz >= 10 && !(clear[0] & 0x01)) /* I-Frame */
{
si->is_kf = 1;
/* vet via sync code */
if (clear[3] != 0x9d || clear[4] != 0x01 || clear[5] != 0x2a)
return VPX_CODEC_UNSUP_BITSTREAM;
si->w = (clear[6] | (clear[7] << 8)) & 0x3fff;
si->h = (clear[8] | (clear[9] << 8)) & 0x3fff;
/*printf("w=%d, h=%d\n", si->w, si->h);*/
if (!(si->h | si->w))
res = VPX_CODEC_UNSUP_BITSTREAM;
}
else
{
res = VPX_CODEC_UNSUP_BITSTREAM;
}
}
return res;
}
void vp9_encode_tiles_mt(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const int num_workers = VPXMIN(cpi->oxcf.max_threads, tile_cols);
int i;
vp9_init_tile_data(cpi);
create_enc_workers(cpi, num_workers);
for (i = 0; i < num_workers; i++) {
EncWorkerData *thread_data;
thread_data = &cpi->tile_thr_data[i];
// Before encoding a frame, copy the thread data from cpi.
if (thread_data->td != &cpi->td) {
thread_data->td->mb = cpi->td.mb;
thread_data->td->rd_counts = cpi->td.rd_counts;
}
if (thread_data->td->counts != &cpi->common.counts) {
memcpy(thread_data->td->counts, &cpi->common.counts,
sizeof(cpi->common.counts));
}
// Handle use_nonrd_pick_mode case.
if (cpi->sf.use_nonrd_pick_mode) {
MACROBLOCK *const x = &thread_data->td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
PICK_MODE_CONTEXT *ctx = &thread_data->td->pc_root->none;
int j;
for (j = 0; j < MAX_MB_PLANE; ++j) {
p[j].coeff = ctx->coeff_pbuf[j][0];
p[j].qcoeff = ctx->qcoeff_pbuf[j][0];
pd[j].dqcoeff = ctx->dqcoeff_pbuf[j][0];
p[j].eobs = ctx->eobs_pbuf[j][0];
}
}
}
launch_enc_workers(cpi, enc_worker_hook, NULL, num_workers);
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
// Accumulate counters.
if (i < cpi->num_workers - 1) {
vp9_accumulate_frame_counts(&cm->counts, thread_data->td->counts, 0);
accumulate_rd_opt(&cpi->td, thread_data->td);
}
}
}
static int dec_get_segment_id(const VP10_COMMON *cm, const uint8_t *segment_ids,
int mi_offset, int x_mis, int y_mis) {
int x, y, segment_id = INT_MAX;
for (y = 0; y < y_mis; y++)
for (x = 0; x < x_mis; x++)
segment_id =
VPXMIN(segment_id, segment_ids[mi_offset + y * cm->mi_cols + x]);
assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
return segment_id;
}
// Set golden frame update interval, for non-svc 1 pass CBR mode.
void vp9_cyclic_refresh_set_golden_update(VP9_COMP *const cpi) {
RATE_CONTROL *const rc = &cpi->rc;
CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
// Set minimum gf_interval for GF update to a multiple of the refresh period,
// with some max limit. Depending on past encoding stats, GF flag may be
// reset and update may not occur until next baseline_gf_interval.
if (cr->percent_refresh > 0)
rc->baseline_gf_interval = VPXMIN(4 * (100 / cr->percent_refresh), 40);
else
rc->baseline_gf_interval = 40;
if (cpi->oxcf.rc_mode == VPX_VBR) rc->baseline_gf_interval = 20;
}
static int read_inter_segment_id(VP10_COMMON *const cm, MACROBLOCKD *const xd,
int mi_row, int mi_col, vpx_reader *r) {
struct segmentation *const seg = &cm->seg;
MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
int predicted_segment_id, segment_id;
const int mi_offset = mi_row * cm->mi_cols + mi_col;
const int bw = xd->plane[0].n4_w >> 1;
const int bh = xd->plane[0].n4_h >> 1;
// TODO(slavarnway): move x_mis, y_mis into xd ?????
const int x_mis = VPXMIN(cm->mi_cols - mi_col, bw);
const int y_mis = VPXMIN(cm->mi_rows - mi_row, bh);
if (!seg->enabled)
return 0; // Default for disabled segmentation
predicted_segment_id = cm->last_frame_seg_map ?
dec_get_segment_id(cm, cm->last_frame_seg_map, mi_offset, x_mis, y_mis) :
0;
if (!seg->update_map) {
copy_segment_id(cm, cm->last_frame_seg_map, cm->current_frame_seg_map,
mi_offset, x_mis, y_mis);
return predicted_segment_id;
}
if (seg->temporal_update) {
const vpx_prob pred_prob = vp10_get_pred_prob_seg_id(seg, xd);
mbmi->seg_id_predicted = vpx_read(r, pred_prob);
segment_id = mbmi->seg_id_predicted ? predicted_segment_id
: read_segment_id(r, seg);
} else {
segment_id = read_segment_id(r, seg);
}
set_segment_id(cm, mi_offset, x_mis, y_mis, segment_id);
return segment_id;
}
static void encode_txfm_probs(VP9_COMMON *cm, vpx_writer *w,
FRAME_COUNTS *counts) {
// Mode
vpx_write_literal(w, VPXMIN(cm->tx_mode, ALLOW_32X32), 2);
if (cm->tx_mode >= ALLOW_32X32)
vpx_write_bit(w, cm->tx_mode == TX_MODE_SELECT);
// Probabilities
if (cm->tx_mode == TX_MODE_SELECT) {
int i, j;
unsigned int ct_8x8p[TX_SIZES - 3][2];
unsigned int ct_16x16p[TX_SIZES - 2][2];
unsigned int ct_32x32p[TX_SIZES - 1][2];
for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
tx_counts_to_branch_counts_8x8(counts->tx.p8x8[i], ct_8x8p);
for (j = 0; j < TX_SIZES - 3; j++)
vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p8x8[i][j], ct_8x8p[j]);
}
for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
tx_counts_to_branch_counts_16x16(counts->tx.p16x16[i], ct_16x16p);
for (j = 0; j < TX_SIZES - 2; j++)
vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p16x16[i][j],
ct_16x16p[j]);
}
for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
tx_counts_to_branch_counts_32x32(counts->tx.p32x32[i], ct_32x32p);
for (j = 0; j < TX_SIZES - 1; j++)
vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p32x32[i][j],
ct_32x32p[j]);
}
}
}
static vpx_codec_err_t decoder_peek_si_internal(const uint8_t *data,
unsigned int data_sz,
vpx_codec_stream_info_t *si,
int *is_intra_only,
vpx_decrypt_cb decrypt_cb,
void *decrypt_state) {
int intra_only_flag = 0;
uint8_t clear_buffer[9];
if (data + data_sz <= data)
return VPX_CODEC_INVALID_PARAM;
si->is_kf = 0;
si->w = si->h = 0;
if (decrypt_cb) {
data_sz = VPXMIN(sizeof(clear_buffer), data_sz);
decrypt_cb(decrypt_state, data, clear_buffer, data_sz);
data = clear_buffer;
}
{
int show_frame;
int error_resilient;
struct vpx_read_bit_buffer rb = { data, data + data_sz, 0, NULL, NULL };
const int frame_marker = vpx_rb_read_literal(&rb, 2);
const BITSTREAM_PROFILE profile = vp9_read_profile(&rb);
if (frame_marker != VP9_FRAME_MARKER)
return VPX_CODEC_UNSUP_BITSTREAM;
if (profile >= MAX_PROFILES)
return VPX_CODEC_UNSUP_BITSTREAM;
if ((profile >= 2 && data_sz <= 1) || data_sz < 1)
return VPX_CODEC_UNSUP_BITSTREAM;
if (vpx_rb_read_bit(&rb)) { // show an existing frame
vpx_rb_read_literal(&rb, 3); // Frame buffer to show.
return VPX_CODEC_OK;
}
if (data_sz <= 8)
return VPX_CODEC_UNSUP_BITSTREAM;
si->is_kf = !vpx_rb_read_bit(&rb);
show_frame = vpx_rb_read_bit(&rb);
error_resilient = vpx_rb_read_bit(&rb);
if (si->is_kf) {
if (!vp9_read_sync_code(&rb))
return VPX_CODEC_UNSUP_BITSTREAM;
if (!parse_bitdepth_colorspace_sampling(profile, &rb))
return VPX_CODEC_UNSUP_BITSTREAM;
vp9_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
} else {
intra_only_flag = show_frame ? 0 : vpx_rb_read_bit(&rb);
rb.bit_offset += error_resilient ? 0 : 2; // reset_frame_context
if (intra_only_flag) {
if (!vp9_read_sync_code(&rb))
return VPX_CODEC_UNSUP_BITSTREAM;
if (profile > PROFILE_0) {
if (!parse_bitdepth_colorspace_sampling(profile, &rb))
return VPX_CODEC_UNSUP_BITSTREAM;
}
rb.bit_offset += REF_FRAMES; // refresh_frame_flags
vp9_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
}
}
}
if (is_intra_only != NULL)
*is_intra_only = intra_only_flag;
return VPX_CODEC_OK;
}
// Update the layer context from a change_config() call.
void vp9_update_layer_context_change_config(VP9_COMP *const cpi,
const int target_bandwidth) {
SVC *const svc = &cpi->svc;
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
const RATE_CONTROL *const rc = &cpi->rc;
int sl, tl, layer = 0, spatial_layer_target;
float bitrate_alloc = 1.0;
if (svc->temporal_layering_mode != VP9E_TEMPORAL_LAYERING_MODE_NOLAYERING) {
for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
layer = LAYER_IDS_TO_IDX(sl, tl, oxcf->ts_number_layers);
svc->layer_context[layer].target_bandwidth =
oxcf->layer_target_bitrate[layer];
}
layer = LAYER_IDS_TO_IDX(sl, ((oxcf->ts_number_layers - 1) < 0 ?
0 : (oxcf->ts_number_layers - 1)), oxcf->ts_number_layers);
spatial_layer_target =
svc->layer_context[layer].target_bandwidth =
oxcf->layer_target_bitrate[layer];
for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
LAYER_CONTEXT *const lc =
&svc->layer_context[sl * oxcf->ts_number_layers + tl];
RATE_CONTROL *const lrc = &lc->rc;
lc->spatial_layer_target_bandwidth = spatial_layer_target;
bitrate_alloc = (float)lc->target_bandwidth / spatial_layer_target;
lrc->starting_buffer_level =
(int64_t)(rc->starting_buffer_level * bitrate_alloc);
lrc->optimal_buffer_level =
(int64_t)(rc->optimal_buffer_level * bitrate_alloc);
lrc->maximum_buffer_size =
(int64_t)(rc->maximum_buffer_size * bitrate_alloc);
lrc->bits_off_target =
VPXMIN(lrc->bits_off_target, lrc->maximum_buffer_size);
lrc->buffer_level = VPXMIN(lrc->buffer_level, lrc->maximum_buffer_size);
lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[tl];
lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
lrc->max_frame_bandwidth = rc->max_frame_bandwidth;
lrc->worst_quality = rc->worst_quality;
lrc->best_quality = rc->best_quality;
}
}
} else {
int layer_end;
if (svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) {
layer_end = svc->number_temporal_layers;
} else {
layer_end = svc->number_spatial_layers;
}
for (layer = 0; layer < layer_end; ++layer) {
LAYER_CONTEXT *const lc = &svc->layer_context[layer];
RATE_CONTROL *const lrc = &lc->rc;
lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
bitrate_alloc = (float)lc->target_bandwidth / target_bandwidth;
// Update buffer-related quantities.
lrc->starting_buffer_level =
(int64_t)(rc->starting_buffer_level * bitrate_alloc);
lrc->optimal_buffer_level =
(int64_t)(rc->optimal_buffer_level * bitrate_alloc);
lrc->maximum_buffer_size =
(int64_t)(rc->maximum_buffer_size * bitrate_alloc);
lrc->bits_off_target = VPXMIN(lrc->bits_off_target,
lrc->maximum_buffer_size);
lrc->buffer_level = VPXMIN(lrc->buffer_level, lrc->maximum_buffer_size);
// Update framerate-related quantities.
if (svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) {
lc->framerate = cpi->framerate / oxcf->ts_rate_decimator[layer];
} else {
lc->framerate = cpi->framerate;
}
lrc->avg_frame_bandwidth = (int)(lc->target_bandwidth / lc->framerate);
lrc->max_frame_bandwidth = rc->max_frame_bandwidth;
// Update qp-related quantities.
lrc->worst_quality = rc->worst_quality;
lrc->best_quality = rc->best_quality;
}
}
}
void vp9_compute_skin_sb(VP9_COMP *const cpi, BLOCK_SIZE bsize, int mi_row,
int mi_col) {
int i, j, num_bl;
VP9_COMMON *const cm = &cpi->common;
const uint8_t *src_y = cpi->Source->y_buffer;
const uint8_t *src_u = cpi->Source->u_buffer;
const uint8_t *src_v = cpi->Source->v_buffer;
const int src_ystride = cpi->Source->y_stride;
const int src_uvstride = cpi->Source->uv_stride;
const int y_bsize = 4 << b_width_log2_lookup[bsize];
const int uv_bsize = y_bsize >> 1;
const int shy = (y_bsize == 8) ? 3 : 4;
const int shuv = shy - 1;
const int fac = y_bsize / 8;
const int y_shift = src_ystride * (mi_row << 3) + (mi_col << 3);
const int uv_shift = src_uvstride * (mi_row << 2) + (mi_col << 2);
const int mi_row_limit = VPXMIN(mi_row + 8, cm->mi_rows - 2);
const int mi_col_limit = VPXMIN(mi_col + 8, cm->mi_cols - 2);
src_y += y_shift;
src_u += uv_shift;
src_v += uv_shift;
for (i = mi_row; i < mi_row_limit; i += fac) {
num_bl = 0;
for (j = mi_col; j < mi_col_limit; j += fac) {
int consec_zeromv = 0;
int bl_index = i * cm->mi_cols + j;
int bl_index1 = bl_index + 1;
int bl_index2 = bl_index + cm->mi_cols;
int bl_index3 = bl_index2 + 1;
// Don't detect skin on the boundary.
if (i == 0 || j == 0) continue;
if (bsize == BLOCK_8X8)
consec_zeromv = cpi->consec_zero_mv[bl_index];
else
consec_zeromv = VPXMIN(cpi->consec_zero_mv[bl_index],
VPXMIN(cpi->consec_zero_mv[bl_index1],
VPXMIN(cpi->consec_zero_mv[bl_index2],
cpi->consec_zero_mv[bl_index3])));
cpi->skin_map[bl_index] =
vp9_compute_skin_block(src_y, src_u, src_v, src_ystride, src_uvstride,
bsize, consec_zeromv, 0);
num_bl++;
src_y += y_bsize;
src_u += uv_bsize;
src_v += uv_bsize;
}
src_y += (src_ystride << shy) - (num_bl << shy);
src_u += (src_uvstride << shuv) - (num_bl << shuv);
src_v += (src_uvstride << shuv) - (num_bl << shuv);
}
// Remove isolated skin blocks (none of its neighbors are skin) and isolated
// non-skin blocks (all of its neighbors are skin).
// Skip 4 corner blocks which have only 3 neighbors to remove isolated skin
// blocks. Skip superblock borders to remove isolated non-skin blocks.
for (i = mi_row; i < mi_row_limit; i += fac) {
for (j = mi_col; j < mi_col_limit; j += fac) {
int bl_index = i * cm->mi_cols + j;
int num_neighbor = 0;
int mi, mj;
int non_skin_threshold = 8;
// Skip 4 corners.
if ((i == mi_row && (j == mi_col || j == mi_col_limit - fac)) ||
(i == mi_row_limit - fac && (j == mi_col || j == mi_col_limit - fac)))
continue;
// There are only 5 neighbors for non-skin blocks on the border.
if (i == mi_row || i == mi_row_limit - fac || j == mi_col ||
j == mi_col_limit - fac)
non_skin_threshold = 5;
for (mi = -fac; mi <= fac; mi += fac) {
for (mj = -fac; mj <= fac; mj += fac) {
if (i + mi >= mi_row && i + mi < mi_row_limit && j + mj >= mi_col &&
j + mj < mi_col_limit) {
int bl_neighbor_index = (i + mi) * cm->mi_cols + j + mj;
if (cpi->skin_map[bl_neighbor_index]) num_neighbor++;
}
}
}
if (cpi->skin_map[bl_index] && num_neighbor < 2)
cpi->skin_map[bl_index] = 0;
if (!cpi->skin_map[bl_index] && num_neighbor == non_skin_threshold)
cpi->skin_map[bl_index] = 1;
}
}
}
void vp9_encode_tiles_mt(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
const int num_workers = VPXMIN(cpi->oxcf.max_threads, tile_cols);
int i;
vp9_init_tile_data(cpi);
// Only run once to create threads and allocate thread data.
if (cpi->num_workers == 0) {
int allocated_workers = num_workers;
// While using SVC, we need to allocate threads according to the highest
// resolution.
if (cpi->use_svc) {
int max_tile_cols = get_max_tile_cols(cpi);
allocated_workers = VPXMIN(cpi->oxcf.max_threads, max_tile_cols);
}
CHECK_MEM_ERROR(cm, cpi->workers,
vpx_malloc(allocated_workers * sizeof(*cpi->workers)));
CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
vpx_calloc(allocated_workers,
sizeof(*cpi->tile_thr_data)));
for (i = 0; i < allocated_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
winterface->init(worker);
if (i < allocated_workers - 1) {
thread_data->cpi = cpi;
// Allocate thread data.
CHECK_MEM_ERROR(cm, thread_data->td,
vpx_memalign(32, sizeof(*thread_data->td)));
vp9_zero(*thread_data->td);
// Set up pc_tree.
thread_data->td->leaf_tree = NULL;
thread_data->td->pc_tree = NULL;
vp9_setup_pc_tree(cm, thread_data->td);
// Allocate frame counters in thread data.
CHECK_MEM_ERROR(cm, thread_data->td->counts,
vpx_calloc(1, sizeof(*thread_data->td->counts)));
// Create threads
if (!winterface->reset(worker))
vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
"Tile encoder thread creation failed");
} else {
// Main thread acts as a worker and uses the thread data in cpi.
thread_data->cpi = cpi;
thread_data->td = &cpi->td;
}
winterface->sync(worker);
}
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data;
worker->hook = (VPxWorkerHook)enc_worker_hook;
worker->data1 = &cpi->tile_thr_data[i];
worker->data2 = NULL;
thread_data = (EncWorkerData*)worker->data1;
// Before encoding a frame, copy the thread data from cpi.
if (thread_data->td != &cpi->td) {
thread_data->td->mb = cpi->td.mb;
thread_data->td->rd_counts = cpi->td.rd_counts;
}
if (thread_data->td->counts != &cpi->common.counts) {
memcpy(thread_data->td->counts, &cpi->common.counts,
sizeof(cpi->common.counts));
}
// Handle use_nonrd_pick_mode case.
if (cpi->sf.use_nonrd_pick_mode) {
MACROBLOCK *const x = &thread_data->td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
PICK_MODE_CONTEXT *ctx = &thread_data->td->pc_root->none;
int j;
for (j = 0; j < MAX_MB_PLANE; ++j) {
p[j].coeff = ctx->coeff_pbuf[j][0];
p[j].qcoeff = ctx->qcoeff_pbuf[j][0];
pd[j].dqcoeff = ctx->dqcoeff_pbuf[j][0];
p[j].eobs = ctx->eobs_pbuf[j][0];
}
}
}
// Encode a frame
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Set the starting tile for each thread.
thread_data->start = i;
if (i == cpi->num_workers - 1)
winterface->execute(worker);
else
winterface->launch(worker);
}
// Encoding ends.
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
winterface->sync(worker);
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Accumulate counters.
if (i < cpi->num_workers - 1) {
vp9_accumulate_frame_counts(cm, thread_data->td->counts, 0);
accumulate_rd_opt(&cpi->td, thread_data->td);
}
}
}
int vp8_decode_frame(VP8D_COMP *pbi) {
vp8_reader *const bc = &pbi->mbc[8];
VP8_COMMON *const pc = &pbi->common;
MACROBLOCKD *const xd = &pbi->mb;
const unsigned char *data = pbi->fragments.ptrs[0];
const unsigned int data_sz = pbi->fragments.sizes[0];
const unsigned char *data_end = data + data_sz;
ptrdiff_t first_partition_length_in_bytes;
int i, j, k, l;
const int *const mb_feature_data_bits = vp8_mb_feature_data_bits;
int corrupt_tokens = 0;
int prev_independent_partitions = pbi->independent_partitions;
YV12_BUFFER_CONFIG *yv12_fb_new = pbi->dec_fb_ref[INTRA_FRAME];
/* start with no corruption of current frame */
xd->corrupted = 0;
yv12_fb_new->corrupted = 0;
if (data_end - data < 3) {
if (!pbi->ec_active) {
vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
"Truncated packet");
}
/* Declare the missing frame as an inter frame since it will
be handled as an inter frame when we have estimated its
motion vectors. */
pc->frame_type = INTER_FRAME;
pc->version = 0;
pc->show_frame = 1;
first_partition_length_in_bytes = 0;
} else {
unsigned char clear_buffer[10];
const unsigned char *clear = data;
if (pbi->decrypt_cb) {
int n = (int)VPXMIN(sizeof(clear_buffer), data_sz);
pbi->decrypt_cb(pbi->decrypt_state, data, clear_buffer, n);
clear = clear_buffer;
}
pc->frame_type = (FRAME_TYPE)(clear[0] & 1);
pc->version = (clear[0] >> 1) & 7;
pc->show_frame = (clear[0] >> 4) & 1;
first_partition_length_in_bytes =
(clear[0] | (clear[1] << 8) | (clear[2] << 16)) >> 5;
if (!pbi->ec_active && (data + first_partition_length_in_bytes > data_end ||
data + first_partition_length_in_bytes < data)) {
vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
"Truncated packet or corrupt partition 0 length");
}
data += 3;
clear += 3;
vp8_setup_version(pc);
if (pc->frame_type == KEY_FRAME) {
/* vet via sync code */
/* When error concealment is enabled we should only check the sync
* code if we have enough bits available
*/
if (data + 3 < data_end) {
if (clear[0] != 0x9d || clear[1] != 0x01 || clear[2] != 0x2a) {
vpx_internal_error(&pc->error, VPX_CODEC_UNSUP_BITSTREAM,
"Invalid frame sync code");
}
}
/* If error concealment is enabled we should only parse the new size
* if we have enough data. Otherwise we will end up with the wrong
* size.
*/
if (data + 6 < data_end) {
pc->Width = (clear[3] | (clear[4] << 8)) & 0x3fff;
pc->horiz_scale = clear[4] >> 6;
pc->Height = (clear[5] | (clear[6] << 8)) & 0x3fff;
pc->vert_scale = clear[6] >> 6;
data += 7;
} else if (!pbi->ec_active) {
vpx_internal_error(&pc->error, VPX_CODEC_CORRUPT_FRAME,
"Truncated key frame header");
} else {
/* Error concealment is active, clear the frame. */
data = data_end;
}
} else {
void vp10_encode_tiles_mt(VP10_COMP *cpi) {
VP10_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
const int num_workers = VPXMIN(cpi->oxcf.max_threads, tile_cols);
int i;
vp10_init_tile_data(cpi);
// Only run once to create threads and allocate thread data.
if (cpi->num_workers == 0) {
int allocated_workers = num_workers;
CHECK_MEM_ERROR(cm, cpi->workers,
vpx_malloc(allocated_workers * sizeof(*cpi->workers)));
CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
vpx_calloc(allocated_workers,
sizeof(*cpi->tile_thr_data)));
for (i = 0; i < allocated_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
winterface->init(worker);
if (i < allocated_workers - 1) {
thread_data->cpi = cpi;
// Allocate thread data.
CHECK_MEM_ERROR(cm, thread_data->td,
vpx_memalign(32, sizeof(*thread_data->td)));
vp10_zero(*thread_data->td);
// Set up pc_tree.
thread_data->td->leaf_tree = NULL;
thread_data->td->pc_tree = NULL;
vp10_setup_pc_tree(cm, thread_data->td);
// Allocate frame counters in thread data.
CHECK_MEM_ERROR(cm, thread_data->td->counts,
vpx_calloc(1, sizeof(*thread_data->td->counts)));
// Create threads
if (!winterface->reset(worker))
vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
"Tile encoder thread creation failed");
} else {
// Main thread acts as a worker and uses the thread data in cpi.
thread_data->cpi = cpi;
thread_data->td = &cpi->td;
}
winterface->sync(worker);
}
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data;
worker->hook = (VPxWorkerHook)enc_worker_hook;
worker->data1 = &cpi->tile_thr_data[i];
worker->data2 = NULL;
thread_data = (EncWorkerData*)worker->data1;
// Before encoding a frame, copy the thread data from cpi.
if (thread_data->td != &cpi->td) {
thread_data->td->mb = cpi->td.mb;
thread_data->td->rd_counts = cpi->td.rd_counts;
}
if (thread_data->td->counts != &cpi->common.counts) {
memcpy(thread_data->td->counts, &cpi->common.counts,
sizeof(cpi->common.counts));
}
}
// Encode a frame
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Set the starting tile for each thread.
thread_data->start = i;
if (i == cpi->num_workers - 1)
winterface->execute(worker);
else
winterface->launch(worker);
}
// Encoding ends.
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
winterface->sync(worker);
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Accumulate counters.
if (i < cpi->num_workers - 1) {
vp10_accumulate_frame_counts(cm, thread_data->td->counts, 0);
accumulate_rd_opt(&cpi->td, thread_data->td);
}
}
}
// Update the segmentation map, and related quantities: cyclic refresh map,
// refresh sb_index, and target number of blocks to be refreshed.
// The map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or to
// 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock.
// Blocks labeled as BOOST1 may later get set to BOOST2 (during the
// encoding of the superblock).
static void cyclic_refresh_update_map(VP9_COMP *const cpi) {
VP9_COMMON *const cm = &cpi->common;
CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
unsigned char *const seg_map = cpi->segmentation_map;
int i, block_count, bl_index, sb_rows, sb_cols, sbs_in_frame;
int xmis, ymis, x, y;
int consec_zero_mv_thresh = 0;
int qindex_thresh = 0;
int count_sel = 0;
int count_tot = 0;
memset(seg_map, CR_SEGMENT_ID_BASE, cm->mi_rows * cm->mi_cols);
sb_cols = (cm->mi_cols + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
sb_rows = (cm->mi_rows + MI_BLOCK_SIZE - 1) / MI_BLOCK_SIZE;
sbs_in_frame = sb_cols * sb_rows;
// Number of target blocks to get the q delta (segment 1).
block_count = cr->percent_refresh * cm->mi_rows * cm->mi_cols / 100;
// Set the segmentation map: cycle through the superblocks, starting at
// cr->mb_index, and stopping when either block_count blocks have been found
// to be refreshed, or we have passed through whole frame.
assert(cr->sb_index < sbs_in_frame);
i = cr->sb_index;
cr->target_num_seg_blocks = 0;
if (cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
consec_zero_mv_thresh = 100;
if (cpi->noise_estimate.enabled && cpi->noise_estimate.level >= kMedium)
consec_zero_mv_thresh = 80;
}
qindex_thresh =
cpi->oxcf.content == VP9E_CONTENT_SCREEN
? vp9_get_qindex(&cm->seg, CR_SEGMENT_ID_BOOST2, cm->base_qindex)
: vp9_get_qindex(&cm->seg, CR_SEGMENT_ID_BOOST1, cm->base_qindex);
do {
int sum_map = 0;
// Get the mi_row/mi_col corresponding to superblock index i.
int sb_row_index = (i / sb_cols);
int sb_col_index = i - sb_row_index * sb_cols;
int mi_row = sb_row_index * MI_BLOCK_SIZE;
int mi_col = sb_col_index * MI_BLOCK_SIZE;
assert(mi_row >= 0 && mi_row < cm->mi_rows);
assert(mi_col >= 0 && mi_col < cm->mi_cols);
bl_index = mi_row * cm->mi_cols + mi_col;
// Loop through all 8x8 blocks in superblock and update map.
xmis =
VPXMIN(cm->mi_cols - mi_col, num_8x8_blocks_wide_lookup[BLOCK_64X64]);
ymis =
VPXMIN(cm->mi_rows - mi_row, num_8x8_blocks_high_lookup[BLOCK_64X64]);
for (y = 0; y < ymis; y++) {
for (x = 0; x < xmis; x++) {
const int bl_index2 = bl_index + y * cm->mi_cols + x;
// If the block is as a candidate for clean up then mark it
// for possible boost/refresh (segment 1). The segment id may get
// reset to 0 later if block gets coded anything other than ZEROMV.
if (cr->map[bl_index2] == 0) {
count_tot++;
if (cr->last_coded_q_map[bl_index2] > qindex_thresh ||
cr->consec_zero_mv[bl_index2] < consec_zero_mv_thresh) {
sum_map++;
count_sel++;
}
} else if (cr->map[bl_index2] < 0) {
cr->map[bl_index2]++;
}
}
}
// Enforce constant segment over superblock.
// If segment is at least half of superblock, set to 1.
if (sum_map >= xmis * ymis / 2) {
for (y = 0; y < ymis; y++)
for (x = 0; x < xmis; x++) {
seg_map[bl_index + y * cm->mi_cols + x] = CR_SEGMENT_ID_BOOST1;
}
cr->target_num_seg_blocks += xmis * ymis;
}
i++;
if (i == sbs_in_frame) {
i = 0;
}
} while (cr->target_num_seg_blocks < block_count && i != cr->sb_index);
cr->sb_index = i;
cr->reduce_refresh = 0;
if (count_sel < (3 * count_tot) >> 2)
cr->reduce_refresh = 1;
}
static void set_good_speed_feature_framesize_dependent(VP9_COMP *cpi,
SPEED_FEATURES *sf,
int speed) {
VP9_COMMON *const cm = &cpi->common;
if (speed >= 1) {
if (VPXMIN(cm->width, cm->height) >= 720) {
sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
: DISABLE_ALL_INTER_SPLIT;
sf->partition_search_breakout_dist_thr = (1 << 23);
} else {
sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
sf->partition_search_breakout_dist_thr = (1 << 21);
}
}
if (speed >= 2) {
if (VPXMIN(cm->width, cm->height) >= 720) {
sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
: DISABLE_ALL_INTER_SPLIT;
sf->adaptive_pred_interp_filter = 0;
sf->partition_search_breakout_dist_thr = (1 << 24);
sf->partition_search_breakout_rate_thr = 120;
} else {
sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
sf->partition_search_breakout_dist_thr = (1 << 22);
sf->partition_search_breakout_rate_thr = 100;
}
sf->rd_auto_partition_min_limit = set_partition_min_limit(cm);
}
if (speed >= 3) {
if (VPXMIN(cm->width, cm->height) >= 720) {
sf->disable_split_mask = DISABLE_ALL_SPLIT;
sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
sf->partition_search_breakout_dist_thr = (1 << 25);
sf->partition_search_breakout_rate_thr = 200;
} else {
sf->max_intra_bsize = BLOCK_32X32;
sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0;
sf->partition_search_breakout_dist_thr = (1 << 23);
sf->partition_search_breakout_rate_thr = 120;
}
}
// If this is a two pass clip that fits the criteria for animated or
// graphics content then reset disable_split_mask for speeds 1-4.
// Also if the image edge is internal to the coded area.
if ((speed >= 1) && (cpi->oxcf.pass == 2) &&
((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
(vp9_internal_image_edge(cpi)))) {
sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
}
if (speed >= 4) {
if (VPXMIN(cm->width, cm->height) >= 720) {
sf->partition_search_breakout_dist_thr = (1 << 26);
} else {
sf->partition_search_breakout_dist_thr = (1 << 24);
}
sf->disable_split_mask = DISABLE_ALL_SPLIT;
}
}
void vp10_dering_frame(YV12_BUFFER_CONFIG *frame, VP10_COMMON *cm,
MACROBLOCKD *xd, int global_level) {
int r, c;
int sbr, sbc;
int nhsb, nvsb;
od_dering_in *src[3];
unsigned char *bskip;
int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = {{0}};
int stride;
int bsize[3];
int dec[3];
int pli;
int coeff_shift = VPXMAX(cm->bit_depth - 8, 0);
nvsb = (cm->mi_rows + MI_BLOCK_SIZE - 1)/MI_BLOCK_SIZE;
nhsb = (cm->mi_cols + MI_BLOCK_SIZE - 1)/MI_BLOCK_SIZE;
bskip = vpx_malloc(sizeof(*bskip)*cm->mi_rows*cm->mi_cols);
vp10_setup_dst_planes(xd->plane, frame, 0, 0);
for (pli = 0; pli < 3; pli++) {
dec[pli] = xd->plane[pli].subsampling_x;
bsize[pli] = 8 >> dec[pli];
}
stride = bsize[0]*cm->mi_cols;
for (pli = 0; pli < 3; pli++) {
src[pli] = vpx_malloc(sizeof(*src)*cm->mi_rows*cm->mi_cols*64);
for (r = 0; r < bsize[pli]*cm->mi_rows; ++r) {
for (c = 0; c < bsize[pli]*cm->mi_cols; ++c) {
#if CONFIG_VPX_HIGHBITDEPTH
if (cm->use_highbitdepth) {
src[pli][r * stride + c] =
CONVERT_TO_SHORTPTR(xd->plane[pli].dst.buf)
[r * xd->plane[pli].dst.stride + c];
} else {
#endif
src[pli][r * stride + c] =
xd->plane[pli].dst.buf[r * xd->plane[pli].dst.stride + c];
#if CONFIG_VPX_HIGHBITDEPTH
}
#endif
}
}
}
for (r = 0; r < cm->mi_rows; ++r) {
for (c = 0; c < cm->mi_cols; ++c) {
const MB_MODE_INFO *mbmi =
&cm->mi_grid_visible[r * cm->mi_stride + c]->mbmi;
bskip[r * cm->mi_cols + c] = mbmi->skip;
}
}
for (sbr = 0; sbr < nvsb; sbr++) {
for (sbc = 0; sbc < nhsb; sbc++) {
int level;
int nhb, nvb;
nhb = VPXMIN(MI_BLOCK_SIZE, cm->mi_cols - MI_BLOCK_SIZE*sbc);
nvb = VPXMIN(MI_BLOCK_SIZE, cm->mi_rows - MI_BLOCK_SIZE*sbr);
for (pli = 0; pli < 3; pli++) {
int16_t dst[MI_BLOCK_SIZE*MI_BLOCK_SIZE*8*8];
int threshold;
#if DERING_REFINEMENT
level = compute_level_from_index(
global_level,
cm->mi_grid_visible[MI_BLOCK_SIZE*sbr*cm->mi_stride +
MI_BLOCK_SIZE*sbc]->mbmi.dering_gain);
#else
level = global_level;
#endif
/* FIXME: This is a temporary hack that uses more conservative
deringing for chroma. */
if (pli) level = (level*5 + 4) >> 3;
if (sb_all_skip(cm, sbr*MI_BLOCK_SIZE, sbc*MI_BLOCK_SIZE)) level = 0;
threshold = level << coeff_shift;
od_dering(
&OD_DERING_VTBL_C,
dst,
MI_BLOCK_SIZE*bsize[pli],
&src[pli][sbr*stride*bsize[pli]*MI_BLOCK_SIZE +
sbc*bsize[pli]*MI_BLOCK_SIZE],
stride, nhb, nvb, sbc, sbr, nhsb, nvsb, dec[pli], dir, pli,
&bskip[MI_BLOCK_SIZE*sbr*cm->mi_cols + MI_BLOCK_SIZE*sbc],
cm->mi_cols, threshold, OD_DERING_NO_CHECK_OVERLAP, coeff_shift);
for (r = 0; r < bsize[pli]*nvb; ++r) {
for (c = 0; c < bsize[pli]*nhb; ++c) {
#if CONFIG_VPX_HIGHBITDEPTH
if (cm->use_highbitdepth) {
CONVERT_TO_SHORTPTR(xd->plane[pli].dst.buf)
[xd->plane[pli].dst.stride*(bsize[pli]*MI_BLOCK_SIZE*sbr + r)
+ sbc*bsize[pli]*MI_BLOCK_SIZE + c] =
dst[r * MI_BLOCK_SIZE * bsize[pli] + c];
} else {
#endif
xd->plane[pli].dst.buf[xd->plane[pli].dst.stride*
(bsize[pli]*MI_BLOCK_SIZE*sbr + r) +
sbc*bsize[pli]*MI_BLOCK_SIZE + c] =
dst[r * MI_BLOCK_SIZE * bsize[pli] + c];
#if CONFIG_VPX_HIGHBITDEPTH
}
#endif
}
}
}
}
}
for (pli = 0; pli < 3; pli++) {
vpx_free(src[pli]);
}
vpx_free(bskip);
}
