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);
}
// 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;
}
}
}
void vp9_encode_fp_row_mt(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const int tile_rows = 1 << cm->log2_tile_rows;
MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
TileDataEnc *first_tile_col;
int num_workers = VPXMAX(cpi->oxcf.max_threads, 1);
int i;
if (multi_thread_ctxt->allocated_tile_cols < tile_cols ||
multi_thread_ctxt->allocated_tile_rows < tile_rows ||
multi_thread_ctxt->allocated_vert_unit_rows < cm->mb_rows) {
vp9_row_mt_mem_dealloc(cpi);
vp9_init_tile_data(cpi);
vp9_row_mt_mem_alloc(cpi);
} else {
vp9_init_tile_data(cpi);
}
create_enc_workers(cpi, num_workers);
vp9_assign_tile_to_thread(multi_thread_ctxt, tile_cols, cpi->num_workers);
vp9_prepare_job_queue(cpi, FIRST_PASS_JOB);
vp9_multi_thread_tile_init(cpi);
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;
}
}
launch_enc_workers(cpi, first_pass_worker_hook, multi_thread_ctxt,
num_workers);
first_tile_col = &cpi->tile_data[0];
for (i = 1; i < tile_cols; i++) {
TileDataEnc *this_tile = &cpi->tile_data[i];
accumulate_fp_tile_stat(first_tile_col, this_tile);
}
}
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);
}
static int get_max_tile_cols(VP9_COMP *cpi) {
const int aligned_width = ALIGN_POWER_OF_TWO(cpi->oxcf.width, MI_SIZE_LOG2);
int mi_cols = aligned_width >> MI_SIZE_LOG2;
int min_log2_tile_cols, max_log2_tile_cols;
int log2_tile_cols;
vp9_get_tile_n_bits(mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
log2_tile_cols =
clamp(cpi->oxcf.tile_columns, min_log2_tile_cols, max_log2_tile_cols);
if (cpi->oxcf.target_level == LEVEL_AUTO) {
const int level_tile_cols =
log_tile_cols_from_picsize_level(cpi->common.width, cpi->common.height);
if (log2_tile_cols > level_tile_cols) {
log2_tile_cols = VPXMAX(level_tile_cols, min_log2_tile_cols);
}
}
return (1 << log2_tile_cols);
}
static int compute_rd_thresh_factor(int qindex, vpx_bit_depth_t bit_depth) {
double q;
#if CONFIG_VP9_HIGHBITDEPTH
switch (bit_depth) {
case VPX_BITS_8:
q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
break;
case VPX_BITS_10:
q = vp9_dc_quant(qindex, 0, VPX_BITS_10) / 16.0;
break;
case VPX_BITS_12:
q = vp9_dc_quant(qindex, 0, VPX_BITS_12) / 64.0;
break;
default:
assert(0 && "bit_depth should be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
return -1;
}
#else
(void) bit_depth;
q = vp9_dc_quant(qindex, 0, VPX_BITS_8) / 4.0;
#endif // CONFIG_VP9_HIGHBITDEPTH
// TODO(debargha): Adjust the function below.
return VPXMAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
}
void vp9_encode_tiles_row_mt(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const int tile_rows = 1 << cm->log2_tile_rows;
MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
int num_workers = VPXMAX(cpi->oxcf.max_threads, 1);
int i;
if (multi_thread_ctxt->allocated_tile_cols < tile_cols ||
multi_thread_ctxt->allocated_tile_rows < tile_rows ||
multi_thread_ctxt->allocated_vert_unit_rows < cm->mb_rows) {
vp9_row_mt_mem_dealloc(cpi);
vp9_init_tile_data(cpi);
vp9_row_mt_mem_alloc(cpi);
} else {
vp9_init_tile_data(cpi);
}
create_enc_workers(cpi, num_workers);
vp9_assign_tile_to_thread(multi_thread_ctxt, tile_cols, cpi->num_workers);
vp9_prepare_job_queue(cpi, ENCODE_JOB);
vp9_multi_thread_tile_init(cpi);
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_row_mt_worker_hook, multi_thread_ctxt,
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);
}
}
}
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);
}
