diff options
Diffstat (limited to 'vp9')
-rw-r--r-- | vp9/common/vp9_blockd.h | 4 | ||||
-rw-r--r-- | vp9/common/vp9_thread_common.c | 6 | ||||
-rw-r--r-- | vp9/decoder/vp9_decodeframe.c | 382 | ||||
-rw-r--r-- | vp9/encoder/vp9_denoiser.c | 33 | ||||
-rw-r--r-- | vp9/encoder/vp9_denoiser.h | 5 | ||||
-rw-r--r-- | vp9/encoder/vp9_encoder.c | 4 | ||||
-rw-r--r-- | vp9/encoder/vp9_noise_estimate.c | 4 | ||||
-rw-r--r-- | vp9/encoder/vp9_pickmode.c | 8 | ||||
-rw-r--r-- | vp9/encoder/vp9_ratectrl.c | 92 | ||||
-rw-r--r-- | vp9/encoder/vp9_rdopt.c | 2 | ||||
-rw-r--r-- | vp9/encoder/vp9_svc_layercontext.c | 10 | ||||
-rw-r--r-- | vp9/encoder/vp9_svc_layercontext.h | 4 |
12 files changed, 483 insertions, 71 deletions
diff --git a/vp9/common/vp9_blockd.h b/vp9/common/vp9_blockd.h index 65db065a6..f0887157e 100644 --- a/vp9/common/vp9_blockd.h +++ b/vp9/common/vp9_blockd.h @@ -131,6 +131,8 @@ struct macroblockd_plane { // encoder const int16_t *dequant; + + int *eob; }; #define BLOCK_OFFSET(x, i) ((x) + (i)*16) @@ -194,6 +196,8 @@ typedef struct macroblockd { int corrupted; struct vpx_internal_error_info *error_info; + + PARTITION_TYPE *partition; } MACROBLOCKD; static INLINE PLANE_TYPE get_plane_type(int plane) { diff --git a/vp9/common/vp9_thread_common.c b/vp9/common/vp9_thread_common.c index 36530fae6..ba9aa69d0 100644 --- a/vp9/common/vp9_thread_common.c +++ b/vp9/common/vp9_thread_common.c @@ -8,6 +8,7 @@ * be found in the AUTHORS file in the root of the source tree. */ +#include <limits.h> #include "./vpx_config.h" #include "vpx_dsp/vpx_dsp_common.h" #include "vpx_mem/vpx_mem.h" @@ -402,6 +403,11 @@ static int get_next_row(VP9_COMMON *cm, VP9LfSync *lf_sync) { pthread_mutex_unlock(&lf_sync->recon_done_mutex[cur_row]); pthread_mutex_lock(&lf_sync->lf_mutex); if (lf_sync->corrupted) { + int row = return_val >> MI_BLOCK_SIZE_LOG2; + pthread_mutex_lock(&lf_sync->mutex[row]); + lf_sync->cur_sb_col[row] = INT_MAX; + pthread_cond_signal(&lf_sync->cond[row]); + pthread_mutex_unlock(&lf_sync->mutex[row]); return_val = -1; } pthread_mutex_unlock(&lf_sync->lf_mutex); diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c index bc0fc6197..c9c85053d 100644 --- a/vp9/decoder/vp9_decodeframe.c +++ b/vp9/decoder/vp9_decodeframe.c @@ -45,6 +45,12 @@ #define MAX_VP9_HEADER_SIZE 80 +typedef int (*predict_recon_func)(TileWorkerData *twd, MODE_INFO *const mi, + int plane, int row, int col, TX_SIZE tx_size); + +typedef void (*intra_recon_func)(TileWorkerData *twd, MODE_INFO *const mi, + int plane, int row, int col, TX_SIZE tx_size); + static int read_is_valid(const uint8_t *start, size_t len, const uint8_t *end) { return len != 0 && len <= (size_t)(end - start); } @@ -325,6 +331,59 @@ static void predict_and_reconstruct_intra_block(TileWorkerData *twd, } } +static void parse_intra_block_row_mt(TileWorkerData *twd, MODE_INFO *const mi, + int plane, int row, int col, + TX_SIZE tx_size) { + MACROBLOCKD *const xd = &twd->xd; + PREDICTION_MODE mode = (plane == 0) ? mi->mode : mi->uv_mode; + + if (mi->sb_type < BLOCK_8X8) + if (plane == 0) mode = xd->mi[0]->bmi[(row << 1) + col].as_mode; + + if (!mi->skip) { + struct macroblockd_plane *const pd = &xd->plane[plane]; + const TX_TYPE tx_type = + (plane || xd->lossless) ? DCT_DCT : intra_mode_to_tx_type_lookup[mode]; + const scan_order *sc = (plane || xd->lossless) + ? &vp9_default_scan_orders[tx_size] + : &vp9_scan_orders[tx_size][tx_type]; + *pd->eob = vp9_decode_block_tokens(twd, plane, sc, col, row, tx_size, + mi->segment_id); + /* Keep the alignment to 16 */ + pd->dqcoeff += (16 << (tx_size << 1)); + pd->eob++; + } +} + +static void predict_and_reconstruct_intra_block_row_mt(TileWorkerData *twd, + MODE_INFO *const mi, + int plane, int row, + int col, + TX_SIZE tx_size) { + MACROBLOCKD *const xd = &twd->xd; + struct macroblockd_plane *const pd = &xd->plane[plane]; + PREDICTION_MODE mode = (plane == 0) ? mi->mode : mi->uv_mode; + uint8_t *dst = &pd->dst.buf[4 * row * pd->dst.stride + 4 * col]; + + if (mi->sb_type < BLOCK_8X8) + if (plane == 0) mode = xd->mi[0]->bmi[(row << 1) + col].as_mode; + + vp9_predict_intra_block(xd, pd->n4_wl, tx_size, mode, dst, pd->dst.stride, + dst, pd->dst.stride, col, row, plane); + + if (!mi->skip) { + const TX_TYPE tx_type = + (plane || xd->lossless) ? DCT_DCT : intra_mode_to_tx_type_lookup[mode]; + if (*pd->eob > 0) { + inverse_transform_block_intra(xd, plane, tx_type, tx_size, dst, + pd->dst.stride, *pd->eob); + } + /* Keep the alignment to 16 */ + pd->dqcoeff += (16 << (tx_size << 1)); + pd->eob++; + } +} + static int reconstruct_inter_block(TileWorkerData *twd, MODE_INFO *const mi, int plane, int row, int col, TX_SIZE tx_size) { @@ -342,6 +401,41 @@ static int reconstruct_inter_block(TileWorkerData *twd, MODE_INFO *const mi, return eob; } +static int parse_inter_block_row_mt(TileWorkerData *twd, MODE_INFO *const mi, + int plane, int row, int col, + TX_SIZE tx_size) { + MACROBLOCKD *const xd = &twd->xd; + struct macroblockd_plane *const pd = &xd->plane[plane]; + const scan_order *sc = &vp9_default_scan_orders[tx_size]; + const int eob = vp9_decode_block_tokens(twd, plane, sc, col, row, tx_size, + mi->segment_id); + + *pd->eob = eob; + pd->dqcoeff += (16 << (tx_size << 1)); + pd->eob++; + + return eob; +} + +static int reconstruct_inter_block_row_mt(TileWorkerData *twd, + MODE_INFO *const mi, int plane, + int row, int col, TX_SIZE tx_size) { + MACROBLOCKD *const xd = &twd->xd; + struct macroblockd_plane *const pd = &xd->plane[plane]; + const int eob = *pd->eob; + + (void)mi; + if (eob > 0) { + inverse_transform_block_inter( + xd, plane, tx_size, &pd->dst.buf[4 * row * pd->dst.stride + 4 * col], + pd->dst.stride, eob); + } + pd->dqcoeff += (16 << (tx_size << 1)); + pd->eob++; + + return eob; +} + static void build_mc_border(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int x, int y, int b_w, int b_h, int w, int h) { @@ -689,6 +783,25 @@ static void set_plane_n4(MACROBLOCKD *const xd, int bw, int bh, int bwl, } } +static MODE_INFO *set_offsets_recon(VP9_COMMON *const cm, MACROBLOCKD *const xd, + int mi_row, int mi_col, int bw, int bh, + int bwl, int bhl) { + const int offset = mi_row * cm->mi_stride + mi_col; + const TileInfo *const tile = &xd->tile; + xd->mi = cm->mi_grid_visible + offset; + + set_plane_n4(xd, bw, bh, bwl, bhl); + + set_skip_context(xd, mi_row, mi_col); + + // Distance of Mb to the various image edges. These are specified to 8th pel + // as they are always compared to values that are in 1/8th pel units + set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols); + + vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col); + return xd->mi[0]; +} + static MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd, BLOCK_SIZE bsize, int mi_row, int mi_col, int bw, int bh, int x_mis, int y_mis, int bwl, int bhl) { @@ -718,6 +831,66 @@ static MODE_INFO *set_offsets(VP9_COMMON *const cm, MACROBLOCKD *const xd, return xd->mi[0]; } +static INLINE int predict_recon_inter(MACROBLOCKD *xd, MODE_INFO *mi, + TileWorkerData *twd, + predict_recon_func func) { + int eobtotal = 0; + int plane; + for (plane = 0; plane < MAX_MB_PLANE; ++plane) { + const struct macroblockd_plane *const pd = &xd->plane[plane]; + const TX_SIZE tx_size = plane ? get_uv_tx_size(mi, pd) : mi->tx_size; + const int num_4x4_w = pd->n4_w; + const int num_4x4_h = pd->n4_h; + const int step = (1 << tx_size); + int row, col; + const int max_blocks_wide = + num_4x4_w + (xd->mb_to_right_edge >= 0 + ? 0 + : xd->mb_to_right_edge >> (5 + pd->subsampling_x)); + const int max_blocks_high = + num_4x4_h + (xd->mb_to_bottom_edge >= 0 + ? 0 + : xd->mb_to_bottom_edge >> (5 + pd->subsampling_y)); + + xd->max_blocks_wide = xd->mb_to_right_edge >= 0 ? 0 : max_blocks_wide; + xd->max_blocks_high = xd->mb_to_bottom_edge >= 0 ? 0 : max_blocks_high; + + for (row = 0; row < max_blocks_high; row += step) + for (col = 0; col < max_blocks_wide; col += step) + eobtotal += func(twd, mi, plane, row, col, tx_size); + } + return eobtotal; +} + +static INLINE void predict_recon_intra(MACROBLOCKD *xd, MODE_INFO *mi, + TileWorkerData *twd, + intra_recon_func func) { + int plane; + for (plane = 0; plane < MAX_MB_PLANE; ++plane) { + const struct macroblockd_plane *const pd = &xd->plane[plane]; + const TX_SIZE tx_size = plane ? get_uv_tx_size(mi, pd) : mi->tx_size; + const int num_4x4_w = pd->n4_w; + const int num_4x4_h = pd->n4_h; + const int step = (1 << tx_size); + int row, col; + const int max_blocks_wide = + num_4x4_w + (xd->mb_to_right_edge >= 0 + ? 0 + : xd->mb_to_right_edge >> (5 + pd->subsampling_x)); + const int max_blocks_high = + num_4x4_h + (xd->mb_to_bottom_edge >= 0 + ? 0 + : xd->mb_to_bottom_edge >> (5 + pd->subsampling_y)); + + xd->max_blocks_wide = xd->mb_to_right_edge >= 0 ? 0 : max_blocks_wide; + xd->max_blocks_high = xd->mb_to_bottom_edge >= 0 ? 0 : max_blocks_high; + + for (row = 0; row < max_blocks_high; row += step) + for (col = 0; col < max_blocks_wide; col += step) + func(twd, mi, plane, row, col, tx_size); + } +} + static void decode_block(TileWorkerData *twd, VP9Decoder *const pbi, int mi_row, int mi_col, BLOCK_SIZE bsize, int bwl, int bhl) { VP9_COMMON *const cm = &pbi->common; @@ -818,6 +991,81 @@ static void decode_block(TileWorkerData *twd, VP9Decoder *const pbi, int mi_row, } } +static void recon_block(TileWorkerData *twd, VP9Decoder *const pbi, int mi_row, + int mi_col, BLOCK_SIZE bsize, int bwl, int bhl) { + VP9_COMMON *const cm = &pbi->common; + const int bw = 1 << (bwl - 1); + const int bh = 1 << (bhl - 1); + MACROBLOCKD *const xd = &twd->xd; + + MODE_INFO *mi = set_offsets_recon(cm, xd, mi_row, mi_col, bw, bh, bwl, bhl); + + if (bsize >= BLOCK_8X8 && (cm->subsampling_x || cm->subsampling_y)) { + const BLOCK_SIZE uv_subsize = + ss_size_lookup[bsize][cm->subsampling_x][cm->subsampling_y]; + if (uv_subsize == BLOCK_INVALID) + vpx_internal_error(xd->error_info, VPX_CODEC_CORRUPT_FRAME, + "Invalid block size."); + } + + if (!is_inter_block(mi)) { + predict_recon_intra(xd, mi, twd, + predict_and_reconstruct_intra_block_row_mt); + } else { + // Prediction + dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col); + + // Reconstruction + if (!mi->skip) { + predict_recon_inter(xd, mi, twd, reconstruct_inter_block_row_mt); + } + } + + vp9_build_mask(cm, mi, mi_row, mi_col, bw, bh); +} + +static void parse_block(TileWorkerData *twd, VP9Decoder *const pbi, int mi_row, + int mi_col, BLOCK_SIZE bsize, int bwl, int bhl) { + VP9_COMMON *const cm = &pbi->common; + const int less8x8 = bsize < BLOCK_8X8; + const int bw = 1 << (bwl - 1); + const int bh = 1 << (bhl - 1); + const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col); + const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row); + vpx_reader *r = &twd->bit_reader; + MACROBLOCKD *const xd = &twd->xd; + + MODE_INFO *mi = set_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, + y_mis, bwl, bhl); + + if (bsize >= BLOCK_8X8 && (cm->subsampling_x || cm->subsampling_y)) { + const BLOCK_SIZE uv_subsize = + ss_size_lookup[bsize][cm->subsampling_x][cm->subsampling_y]; + if (uv_subsize == BLOCK_INVALID) + vpx_internal_error(xd->error_info, VPX_CODEC_CORRUPT_FRAME, + "Invalid block size."); + } + + vp9_read_mode_info(twd, pbi, mi_row, mi_col, x_mis, y_mis); + + if (mi->skip) { + dec_reset_skip_context(xd); + } + + if (!is_inter_block(mi)) { + predict_recon_intra(xd, mi, twd, parse_intra_block_row_mt); + } else { + if (!mi->skip) { + const int eobtotal = + predict_recon_inter(xd, mi, twd, parse_inter_block_row_mt); + + if (!less8x8 && eobtotal == 0) mi->skip = 1; // skip loopfilter + } + } + + xd->corrupted |= vpx_reader_has_error(r); +} + static INLINE int dec_partition_plane_context(TileWorkerData *twd, int mi_row, int mi_col, int bsl) { const PARTITION_CONTEXT *above_ctx = twd->xd.above_seg_context + mi_col; @@ -924,6 +1172,118 @@ static void decode_partition(TileWorkerData *twd, VP9Decoder *const pbi, dec_update_partition_context(twd, mi_row, mi_col, subsize, num_8x8_wh); } +static void recon_partition(TileWorkerData *twd, VP9Decoder *const pbi, + int mi_row, int mi_col, BLOCK_SIZE bsize, + int n4x4_l2) { + VP9_COMMON *const cm = &pbi->common; + const int n8x8_l2 = n4x4_l2 - 1; + const int num_8x8_wh = 1 << n8x8_l2; + const int hbs = num_8x8_wh >> 1; + PARTITION_TYPE partition; + BLOCK_SIZE subsize; + const int has_rows = (mi_row + hbs) < cm->mi_rows; + const int has_cols = (mi_col + hbs) < cm->mi_cols; + MACROBLOCKD *const xd = &twd->xd; + + if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return; + + partition = *xd->partition; + xd->partition++; + + subsize = get_subsize(bsize, partition); + if (!hbs) { + // calculate bmode block dimensions (log 2) + xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT); + xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ); + recon_block(twd, pbi, mi_row, mi_col, subsize, 1, 1); + } else { + switch (partition) { + case PARTITION_NONE: + recon_block(twd, pbi, mi_row, mi_col, subsize, n4x4_l2, n4x4_l2); + break; + case PARTITION_HORZ: + recon_block(twd, pbi, mi_row, mi_col, subsize, n4x4_l2, n8x8_l2); + if (has_rows) + recon_block(twd, pbi, mi_row + hbs, mi_col, subsize, n4x4_l2, + n8x8_l2); + break; + case PARTITION_VERT: + recon_block(twd, pbi, mi_row, mi_col, subsize, n8x8_l2, n4x4_l2); + if (has_cols) + recon_block(twd, pbi, mi_row, mi_col + hbs, subsize, n8x8_l2, + n4x4_l2); + break; + case PARTITION_SPLIT: + recon_partition(twd, pbi, mi_row, mi_col, subsize, n8x8_l2); + recon_partition(twd, pbi, mi_row, mi_col + hbs, subsize, n8x8_l2); + recon_partition(twd, pbi, mi_row + hbs, mi_col, subsize, n8x8_l2); + recon_partition(twd, pbi, mi_row + hbs, mi_col + hbs, subsize, n8x8_l2); + break; + default: assert(0 && "Invalid partition type"); + } + } +} + +static void parse_partition(TileWorkerData *twd, VP9Decoder *const pbi, + int mi_row, int mi_col, BLOCK_SIZE bsize, + int n4x4_l2) { + VP9_COMMON *const cm = &pbi->common; + const int n8x8_l2 = n4x4_l2 - 1; + const int num_8x8_wh = 1 << n8x8_l2; + const int hbs = num_8x8_wh >> 1; + PARTITION_TYPE partition; + BLOCK_SIZE subsize; + const int has_rows = (mi_row + hbs) < cm->mi_rows; + const int has_cols = (mi_col + hbs) < cm->mi_cols; + MACROBLOCKD *const xd = &twd->xd; + + if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return; + + *xd->partition = + read_partition(twd, mi_row, mi_col, has_rows, has_cols, n8x8_l2); + + partition = *xd->partition; + xd->partition++; + + subsize = get_subsize(bsize, partition); + if (!hbs) { + // calculate bmode block dimensions (log 2) + xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT); + xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ); + parse_block(twd, pbi, mi_row, mi_col, subsize, 1, 1); + } else { + switch (partition) { + case PARTITION_NONE: + parse_block(twd, pbi, mi_row, mi_col, subsize, n4x4_l2, n4x4_l2); + break; + case PARTITION_HORZ: + parse_block(twd, pbi, mi_row, mi_col, subsize, n4x4_l2, n8x8_l2); + if (has_rows) + parse_block(twd, pbi, mi_row + hbs, mi_col, subsize, n4x4_l2, + n8x8_l2); + break; + case PARTITION_VERT: + parse_block(twd, pbi, mi_row, mi_col, subsize, n8x8_l2, n4x4_l2); + if (has_cols) + parse_block(twd, pbi, mi_row, mi_col + hbs, subsize, n8x8_l2, + n4x4_l2); + break; + case PARTITION_SPLIT: + parse_partition(twd, pbi, mi_row, mi_col, subsize, n8x8_l2); + parse_partition(twd, pbi, mi_row, mi_col + hbs, subsize, n8x8_l2); + parse_partition(twd, pbi, mi_row + hbs, mi_col, subsize, n8x8_l2); + parse_partition(twd, pbi, mi_row + hbs, mi_col + hbs, subsize, n8x8_l2); + break; + default: assert(0 && "Invalid partition type"); + } + } + + // update partition context + if (bsize >= BLOCK_8X8 && + (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT)) + dec_update_partition_context(twd, mi_row, mi_col, subsize, num_8x8_wh); +} + static void setup_token_decoder(const uint8_t *data, const uint8_t *data_end, size_t read_size, struct vpx_internal_error_info *error_info, @@ -1406,7 +1766,27 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi, const uint8_t *data, vp9_zero(tile_data->xd.left_seg_context); for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end; mi_col += MI_BLOCK_SIZE) { - decode_partition(tile_data, pbi, mi_row, mi_col, BLOCK_64X64, 4); + if (pbi->row_mt == 1) { + int plane; + RowMTWorkerData *const row_mt_worker_data = pbi->row_mt_worker_data; + for (plane = 0; plane < MAX_MB_PLANE; ++plane) { + tile_data->xd.plane[plane].eob = row_mt_worker_data->eob[plane]; + tile_data->xd.plane[plane].dqcoeff = + row_mt_worker_data->dqcoeff[plane]; + } + tile_data->xd.partition = row_mt_worker_data->partition; + parse_partition(tile_data, pbi, mi_row, mi_col, BLOCK_64X64, 4); + + for (plane = 0; plane < MAX_MB_PLANE; ++plane) { + tile_data->xd.plane[plane].eob = row_mt_worker_data->eob[plane]; + tile_data->xd.plane[plane].dqcoeff = + row_mt_worker_data->dqcoeff[plane]; + } + tile_data->xd.partition = row_mt_worker_data->partition; + recon_partition(tile_data, pbi, mi_row, mi_col, BLOCK_64X64, 4); + } else { + decode_partition(tile_data, pbi, mi_row, mi_col, BLOCK_64X64, 4); + } } pbi->mb.corrupted |= tile_data->xd.corrupted; if (pbi->mb.corrupted) diff --git a/vp9/encoder/vp9_denoiser.c b/vp9/encoder/vp9_denoiser.c index b70890e68..2820b71b4 100644 --- a/vp9/encoder/vp9_denoiser.c +++ b/vp9/encoder/vp9_denoiser.c @@ -692,6 +692,7 @@ int vp9_denoiser_alloc(VP9_COMMON *cm, struct SVC *svc, VP9_DENOISER *denoiser, denoiser->denoising_level = kDenLow; denoiser->prev_denoising_level = kDenLow; denoiser->reset = 0; + denoiser->current_denoiser_frame = 0; return 0; } @@ -716,13 +717,29 @@ void vp9_denoiser_free(VP9_DENOISER *denoiser) { vpx_free_frame_buffer(&denoiser->last_source); } -void vp9_denoiser_set_noise_level(VP9_DENOISER *denoiser, int noise_level) { +static void force_refresh_longterm_ref(VP9_COMP *const cpi) { + SVC *const svc = &cpi->svc; + // If long term reference is used, force refresh of that slot, so + // denoiser buffer for long term reference stays in sync. + if (svc->use_gf_temporal_ref_current_layer) { + int index = svc->spatial_layer_id; + if (svc->number_spatial_layers == 3) index = svc->spatial_layer_id - 1; + assert(index >= 0); + cpi->alt_fb_idx = svc->buffer_gf_temporal_ref[index].idx; + cpi->refresh_alt_ref_frame = 1; + } +} + +void vp9_denoiser_set_noise_level(VP9_COMP *const cpi, int noise_level) { + VP9_DENOISER *const denoiser = &cpi->denoiser; denoiser->denoising_level = noise_level; if (denoiser->denoising_level > kDenLowLow && - denoiser->prev_denoising_level == kDenLowLow) + denoiser->prev_denoising_level == kDenLowLow) { denoiser->reset = 1; - else + force_refresh_longterm_ref(cpi); + } else { denoiser->reset = 0; + } denoiser->prev_denoising_level = denoiser->denoising_level; } @@ -754,14 +771,24 @@ int64_t vp9_scale_acskip_thresh(int64_t threshold, return threshold; } +void vp9_denoiser_reset_on_first_frame(VP9_COMP *const cpi) { + if (vp9_denoise_svc_non_key(cpi) && + cpi->denoiser.current_denoiser_frame == 0) { + cpi->denoiser.reset = 1; + force_refresh_longterm_ref(cpi); + } +} + void vp9_denoiser_update_ref_frame(VP9_COMP *const cpi) { VP9_COMMON *const cm = &cpi->common; SVC *const svc = &cpi->svc; + if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) && cpi->denoiser.denoising_level > kDenLowLow) { int svc_refresh_denoiser_buffers = 0; int denoise_svc_second_layer = 0; FRAME_TYPE frame_type = cm->intra_only ? KEY_FRAME : cm->frame_type; + cpi->denoiser.current_denoiser_frame++; if (cpi->use_svc) { const int svc_buf_shift = svc->number_spatial_layers - svc->spatial_layer_id == 2 diff --git a/vp9/encoder/vp9_denoiser.h b/vp9/encoder/vp9_denoiser.h index 2362c4f50..1973e9898 100644 --- a/vp9/encoder/vp9_denoiser.h +++ b/vp9/encoder/vp9_denoiser.h @@ -50,6 +50,7 @@ typedef struct vp9_denoiser { int reset; int num_ref_frames; int num_layers; + unsigned int current_denoiser_frame; VP9_DENOISER_LEVEL denoising_level; VP9_DENOISER_LEVEL prev_denoising_level; } VP9_DENOISER; @@ -111,7 +112,9 @@ static INLINE int total_adj_strong_thresh(BLOCK_SIZE bs, void vp9_denoiser_free(VP9_DENOISER *denoiser); -void vp9_denoiser_set_noise_level(VP9_DENOISER *denoiser, int noise_level); +void vp9_denoiser_set_noise_level(struct VP9_COMP *const cpi, int noise_level); + +void vp9_denoiser_reset_on_first_frame(struct VP9_COMP *const cpi); int64_t vp9_scale_part_thresh(int64_t threshold, VP9_DENOISER_LEVEL noise_level, int content_state, int temporal_layer_id); diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c index 746f234e2..a73185623 100644 --- a/vp9/encoder/vp9_encoder.c +++ b/vp9/encoder/vp9_encoder.c @@ -3815,6 +3815,10 @@ static int encode_without_recode_loop(VP9_COMP *cpi, size_t *size, cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv)); } +#if CONFIG_VP9_TEMPORAL_DENOISING + if (cpi->oxcf.noise_sensitivity > 0 && cpi->use_svc) + vp9_denoiser_reset_on_first_frame(cpi); +#endif vp9_update_noise_estimate(cpi); // Scene detection is always used for VBR mode or screen-content case. diff --git a/vp9/encoder/vp9_noise_estimate.c b/vp9/encoder/vp9_noise_estimate.c index 8c9a40f55..fc189dbb1 100644 --- a/vp9/encoder/vp9_noise_estimate.c +++ b/vp9/encoder/vp9_noise_estimate.c @@ -159,7 +159,7 @@ void vp9_update_noise_estimate(VP9_COMP *const cpi) { #if CONFIG_VP9_TEMPORAL_DENOISING if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi) && cpi->svc.current_superframe > 1) { - vp9_denoiser_set_noise_level(&cpi->denoiser, ne->level); + vp9_denoiser_set_noise_level(cpi, ne->level); copy_frame(&cpi->denoiser.last_source, cpi->Source); } #endif @@ -269,7 +269,7 @@ void vp9_update_noise_estimate(VP9_COMP *const cpi) { ne->level = vp9_noise_estimate_extract_level(ne); #if CONFIG_VP9_TEMPORAL_DENOISING if (cpi->oxcf.noise_sensitivity > 0 && noise_est_svc(cpi)) - vp9_denoiser_set_noise_level(&cpi->denoiser, ne->level); + vp9_denoiser_set_noise_level(cpi, ne->level); #endif } } diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c index 1324b5bc8..fe8f24444 100644 --- a/vp9/encoder/vp9_pickmode.c +++ b/vp9/encoder/vp9_pickmode.c @@ -1815,13 +1815,7 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data, #if CONFIG_VP9_TEMPORAL_DENOISING if (cpi->oxcf.noise_sensitivity > 0) { - if (cpi->use_svc) { - int layer = - LAYER_IDS_TO_IDX(svc->spatial_layer_id, svc->temporal_layer_id, - svc->number_temporal_layers); - LAYER_CONTEXT *lc = &svc->layer_context[layer]; - denoise_svc_pickmode = denoise_svc(cpi) && !lc->is_key_frame; - } + if (cpi->use_svc) denoise_svc_pickmode = vp9_denoise_svc_non_key(cpi); if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode) vp9_denoiser_reset_frame_stats(ctx); } diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c index ac8fda496..b5c002aea 100644 --- a/vp9/encoder/vp9_ratectrl.c +++ b/vp9/encoder/vp9_ratectrl.c @@ -570,10 +570,25 @@ int post_encode_drop_cbr(VP9_COMP *cpi, size_t *size) { cpi->last_frame_dropped = 1; cpi->ext_refresh_frame_flags_pending = 0; if (cpi->use_svc) { - cpi->svc.last_layer_dropped[cpi->svc.spatial_layer_id] = 1; - cpi->svc.drop_spatial_layer[cpi->svc.spatial_layer_id] = 1; - cpi->svc.drop_count[cpi->svc.spatial_layer_id]++; - cpi->svc.skip_enhancement_layer = 1; + SVC *svc = &cpi->svc; + int sl = 0; + int tl = 0; + svc->last_layer_dropped[svc->spatial_layer_id] = 1; + svc->drop_spatial_layer[svc->spatial_layer_id] = 1; + svc->drop_count[svc->spatial_layer_id]++; + svc->skip_enhancement_layer = 1; + // Postencode drop is only checked on base spatial layer, + // for now if max-q is set on base we force it on all layers. + for (sl = 0; sl < svc->number_spatial_layers; ++sl) { + for (tl = 0; tl < svc->number_temporal_layers; ++tl) { + const int layer = + LAYER_IDS_TO_IDX(sl, tl, svc->number_temporal_layers); + LAYER_CONTEXT *lc = &svc->layer_context[layer]; + RATE_CONTROL *lrc = &lc->rc; + lrc->force_max_q = 1; + lrc->avg_frame_qindex[INTER_FRAME] = cpi->rc.worst_quality; + } + } } return 1; } @@ -1394,16 +1409,7 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi, int *bottom_index, return rc_constant_q(cpi, bottom_index, top_index, gf_group_index); if (frame_is_intra_only(cm)) { - if (rc->frames_to_key == 1 && oxcf->rc_mode == VPX_Q) { - // If the next frame is also a key frame or the current frame is the - // only frame in the sequence in AOM_Q mode, just use the cq_level - // as q. - active_best_quality = cq_level; - active_worst_quality = cq_level; - } else { - pick_kf_q_bound_two_pass(cpi, &active_best_quality, - &active_worst_quality); - } + pick_kf_q_bound_two_pass(cpi, &active_best_quality, &active_worst_quality); } else if (!rc->is_src_frame_alt_ref && (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) { // Use the lower of active_worst_quality and recent @@ -1434,54 +1440,31 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi, int *bottom_index, ((layer_depth - 1) * q + active_best_quality + layer_depth / 2) / layer_depth; } - } else if (oxcf->rc_mode == VPX_Q) { - if (!cpi->refresh_alt_ref_frame) { - active_best_quality = cq_level; - } else { - active_best_quality = get_gf_active_quality(cpi, q, cm->bit_depth); - - // Modify best quality for second level arfs. For mode VPX_Q this - // becomes the baseline frame q. - if (gf_group->rf_level[gf_group_index] == GF_ARF_LOW) { - const int layer_depth = gf_group->layer_depth[gf_group_index]; - // linearly fit the frame q depending on the layer depth index from - // the base layer ARF. - active_best_quality = ((layer_depth - 1) * cq_level + - active_best_quality + layer_depth / 2) / - layer_depth; - } - } } else { active_best_quality = get_gf_active_quality(cpi, q, cm->bit_depth); } } else { - if (oxcf->rc_mode == VPX_Q) { - active_best_quality = cq_level; - } else { - active_best_quality = inter_minq[active_worst_quality]; + active_best_quality = inter_minq[active_worst_quality]; - // For the constrained quality mode we don't want - // q to fall below the cq level. - if ((oxcf->rc_mode == VPX_CQ) && (active_best_quality < cq_level)) { - active_best_quality = cq_level; - } + // For the constrained quality mode we don't want + // q to fall below the cq level. + if ((oxcf->rc_mode == VPX_CQ) && (active_best_quality < cq_level)) { + active_best_quality = cq_level; } } // Extension to max or min Q if undershoot or overshoot is outside // the permitted range. - if (cpi->oxcf.rc_mode != VPX_Q) { - if (frame_is_intra_only(cm) || - (!rc->is_src_frame_alt_ref && - (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))) { - active_best_quality -= - (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast); - active_worst_quality += (cpi->twopass.extend_maxq / 2); - } else { - active_best_quality -= - (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast) / 2; - active_worst_quality += cpi->twopass.extend_maxq; - } + if (frame_is_intra_only(cm) || + (!rc->is_src_frame_alt_ref && + (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))) { + active_best_quality -= + (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast); + active_worst_quality += (cpi->twopass.extend_maxq / 2); + } else { + active_best_quality -= + (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast) / 2; + active_worst_quality += cpi->twopass.extend_maxq; } // For normal frames do not allow an active minq lower than the q used for @@ -1517,10 +1500,7 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi, int *bottom_index, active_worst_quality = clamp(active_worst_quality, active_best_quality, rc->worst_quality); - if (oxcf->rc_mode == VPX_Q) { - q = active_best_quality; - // Special case code to try and match quality with forced key frames. - } else if (frame_is_intra_only(cm) && rc->this_key_frame_forced) { + if (frame_is_intra_only(cm) && rc->this_key_frame_forced) { // If static since last kf use better of last boosted and last kf q. if (cpi->twopass.last_kfgroup_zeromotion_pct >= STATIC_MOTION_THRESH) { q = VPXMIN(rc->last_kf_qindex, rc->last_boosted_qindex); diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c index 2e1aa1d30..b55e2ddb4 100644 --- a/vp9/encoder/vp9_rdopt.c +++ b/vp9/encoder/vp9_rdopt.c @@ -2461,7 +2461,7 @@ static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, #if CONFIG_NON_GREEDY_MV this_me = vp9_full_pixel_diamond_new( cpi, x, &mvp_full, VPXMAX(step_param, MAX_MVSEARCH_STEPS - step), - lambda, 1, &cpi->fn_ptr[bsize], nb_full_mvs, &tmp_mv->as_mv, &mv_dist, + lambda, 1, &cpi->fn_ptr[bsize], nb_full_mvs, &this_mv, &mv_dist, &mv_cost); #else // CONFIG_NON_GREEDY_MV this_me = vp9_full_pixel_search( diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c index 21b920f11..510087580 100644 --- a/vp9/encoder/vp9_svc_layercontext.c +++ b/vp9/encoder/vp9_svc_layercontext.c @@ -1106,6 +1106,16 @@ void vp9_svc_assert_constraints_pattern(VP9_COMP *const cpi) { } } +#if CONFIG_VP9_TEMPORAL_DENOISING +int vp9_denoise_svc_non_key(VP9_COMP *const cpi) { + int layer = + LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id, cpi->svc.temporal_layer_id, + cpi->svc.number_temporal_layers); + LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer]; + return denoise_svc(cpi) && !lc->is_key_frame; +} +#endif + void vp9_svc_check_spatial_layer_sync(VP9_COMP *const cpi) { SVC *const svc = &cpi->svc; // Only for superframes whose base is not key, as those are diff --git a/vp9/encoder/vp9_svc_layercontext.h b/vp9/encoder/vp9_svc_layercontext.h index 945312044..f1f2457b2 100644 --- a/vp9/encoder/vp9_svc_layercontext.h +++ b/vp9/encoder/vp9_svc_layercontext.h @@ -235,6 +235,10 @@ struct lookahead_entry *vp9_svc_lookahead_pop(struct VP9_COMP *const cpi, // Start a frame and initialize svc parameters int vp9_svc_start_frame(struct VP9_COMP *const cpi); +#if CONFIG_VP9_TEMPORAL_DENOISING +int vp9_denoise_svc_non_key(struct VP9_COMP *const cpi); +#endif + void vp9_copy_flags_ref_update_idx(struct VP9_COMP *const cpi); int vp9_one_pass_cbr_svc_start_layer(struct VP9_COMP *const cpi); |