diff options
Diffstat (limited to 'vp9/encoder')
-rw-r--r-- | vp9/encoder/vp9_encoder.h | 4 | ||||
-rw-r--r-- | vp9/encoder/vp9_firstpass.c | 26 | ||||
-rw-r--r-- | vp9/encoder/vp9_pickmode.c | 63 | ||||
-rw-r--r-- | vp9/encoder/vp9_ratectrl.c | 40 | ||||
-rw-r--r-- | vp9/encoder/vp9_ratectrl.h | 2 | ||||
-rw-r--r-- | vp9/encoder/vp9_speed_features.c | 4 |
6 files changed, 110 insertions, 29 deletions
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h index f448f961c..ddba3275e 100644 --- a/vp9/encoder/vp9_encoder.h +++ b/vp9/encoder/vp9_encoder.h @@ -119,8 +119,8 @@ typedef enum { COMPLEXITY_AQ = 2, CYCLIC_REFRESH_AQ = 3, EQUATOR360_AQ = 4, - PSNR_AQ = 5, - PERCEPTUAL_AQ = 6, + PERCEPTUAL_AQ = 5, + PSNR_AQ = 6, // AQ based on lookahead temporal // variance (only valid for altref frames) LOOKAHEAD_AQ = 7, diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c index 665b4c36e..392d1088e 100644 --- a/vp9/encoder/vp9_firstpass.c +++ b/vp9/encoder/vp9_firstpass.c @@ -2642,6 +2642,17 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { rc->source_alt_ref_pending = 0; } +#define LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR 0.2 + rc->arf_active_best_quality_adjustment_factor = 1.0; + if (rc->source_alt_ref_pending && !is_lossless_requested(&cpi->oxcf) && + rc->frames_to_key <= rc->arf_active_best_quality_adjustment_window) { + rc->arf_active_best_quality_adjustment_factor = + LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR + + (1.0 - LAST_ALR_ACTIVE_BEST_QUALITY_ADJUSTMENT_FACTOR) * + (rc->frames_to_key - i) / + VPXMAX(1, (rc->arf_active_best_quality_adjustment_window - i)); + } + #ifdef AGGRESSIVE_VBR // Limit maximum boost based on interval length. rc->gfu_boost = VPXMIN((int)rc->gfu_boost, i * 140); @@ -3190,6 +3201,11 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) { // Default to normal-sized frame on keyframes. cpi->rc.next_frame_size_selector = UNSCALED; } +#define ARF_ACTIVE_BEST_QUALITY_ADJUSTMENT_WINDOW_SIZE 64 + // TODO(ravi.chaudhary@ittiam.com): Experiment without the below min + // condition. This might be helpful for small key frame intervals. + rc->arf_active_best_quality_adjustment_window = + VPXMIN(ARF_ACTIVE_BEST_QUALITY_ADJUSTMENT_WINDOW_SIZE, rc->frames_to_key); } static int is_skippable_frame(const VP9_COMP *cpi) { @@ -3225,6 +3241,11 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) { if (gf_group->update_type[gf_group->index] == ARF_UPDATE) { int target_rate; + vp9_zero(this_frame); + this_frame = + cpi->twopass.stats_in_start[cm->current_video_frame + + gf_group->arf_src_offset[gf_group->index]]; + vp9_configure_buffer_updates(cpi, gf_group->index); target_rate = gf_group->bit_allocation[gf_group->index]; @@ -3240,6 +3261,11 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) { cpi->partition_search_skippable_frame = is_skippable_frame(cpi); } + // The multiplication by 256 reverses a scaling factor of (>> 8) + // applied when combining MB error values for the frame. + twopass->mb_av_energy = log((this_frame.intra_error * 256.0) + 1.0); + twopass->mb_smooth_pct = this_frame.intra_smooth_pct; + return; } diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c index a431e4ca6..d6052fade 100644 --- a/vp9/encoder/vp9_pickmode.c +++ b/vp9/encoder/vp9_pickmode.c @@ -376,10 +376,52 @@ static TX_SIZE calculate_tx_size(VP9_COMP *const cpi, BLOCK_SIZE bsize, tx_size = VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); } - return tx_size; } +static void compute_intra_yprediction(PREDICTION_MODE mode, BLOCK_SIZE bsize, + MACROBLOCK *x, MACROBLOCKD *xd) { + struct macroblockd_plane *const pd = &xd->plane[0]; + struct macroblock_plane *const p = &x->plane[0]; + uint8_t *const src_buf_base = p->src.buf; + uint8_t *const dst_buf_base = pd->dst.buf; + const int src_stride = p->src.stride; + const int dst_stride = pd->dst.stride; + // block and transform sizes, in number of 4x4 blocks log 2 ("*_b") + // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8 + const TX_SIZE tx_size = max_txsize_lookup[bsize]; + const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize]; + const int num_4x4_h = num_4x4_blocks_high_lookup[bsize]; + int row, col; + // If mb_to_right_edge is < 0 we are in a situation in which + // the current block size extends into the UMV and we won't + // visit the sub blocks that are wholly within the UMV. + 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)); + + // Keep track of the row and column of the blocks we use so that we know + // if we are in the unrestricted motion border. + for (row = 0; row < max_blocks_high; row += (1 << tx_size)) { + // Skip visiting the sub blocks that are wholly within the UMV. + for (col = 0; col < max_blocks_wide; col += (1 << tx_size)) { + p->src.buf = &src_buf_base[4 * (row * (int64_t)src_stride + col)]; + pd->dst.buf = &dst_buf_base[4 * (row * (int64_t)dst_stride + col)]; + vp9_predict_intra_block(xd, b_width_log2_lookup[bsize], tx_size, mode, + x->skip_encode ? p->src.buf : pd->dst.buf, + x->skip_encode ? src_stride : dst_stride, + pd->dst.buf, dst_stride, col, row, 0); + } + } + p->src.buf = src_buf_base; + pd->dst.buf = dst_buf_base; +} + static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x, MACROBLOCKD *xd, int *out_rate_sum, int64_t *out_dist_sum, @@ -1013,8 +1055,8 @@ static void estimate_block_intra(int plane, int block, int row, int col, (void)block; - p->src.buf = &src_buf_base[4 * (row * src_stride + col)]; - pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)]; + p->src.buf = &src_buf_base[4 * (row * (int64_t)src_stride + col)]; + pd->dst.buf = &dst_buf_base[4 * (row * (int64_t)dst_stride + col)]; // Use source buffer as an approximation for the fully reconstructed buffer. vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size, args->mode, x->skip_encode ? p->src.buf : pd->dst.buf, @@ -1023,13 +1065,12 @@ static void estimate_block_intra(int plane, int block, int row, int col, if (plane == 0) { int64_t this_sse = INT64_MAX; - // TODO(jingning): This needs further refactoring. block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx, VPXMIN(tx_size, TX_16X16), 0, 1); } else { unsigned int var = 0; unsigned int sse = 0; - model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane, + model_rd_for_sb_uv(cpi, bsize_tx, x, xd, &this_rdc, &var, &sse, plane, plane); } @@ -2472,13 +2513,12 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data, bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad && !x->lowvar_highsumdiff)) { struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 }; + int64_t this_sse = INT64_MAX; int i; PRED_BUFFER *const best_pred = best_pickmode.best_pred; TX_SIZE intra_tx_size = VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); - if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16) - intra_tx_size = TX_16X16; if (reuse_inter_pred && best_pred != NULL) { if (best_pred->data == orig_dst.buf) { @@ -2539,8 +2579,13 @@ void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data, args.skippable = 1; args.rdc = &this_rdc; mi->tx_size = intra_tx_size; - vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra, - &args); + + compute_intra_yprediction(this_mode, bsize, x, xd); + model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist, + &var_y, &sse_y, 1); + block_yrd(cpi, x, &this_rdc, &args.skippable, &this_sse, bsize, + VPXMIN(mi->tx_size, TX_16X16), 1, 1); + // Check skip cost here since skippable is not set for for uv, this // mirrors the behavior used by inter if (args.skippable) { diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c index b51f6797b..e45030ab8 100644 --- a/vp9/encoder/vp9_ratectrl.c +++ b/vp9/encoder/vp9_ratectrl.c @@ -211,17 +211,15 @@ int vp9_rc_clamp_pframe_target_size(const VP9_COMP *const cpi, int target) { const RATE_CONTROL *rc = &cpi->rc; const VP9EncoderConfig *oxcf = &cpi->oxcf; - if (cpi->oxcf.pass != 2) { - const int min_frame_target = - VPXMAX(rc->min_frame_bandwidth, rc->avg_frame_bandwidth >> 5); - if (target < min_frame_target) target = min_frame_target; - if (cpi->refresh_golden_frame && rc->is_src_frame_alt_ref) { - // If there is an active ARF at this location use the minimum - // bits on this frame even if it is a constructed arf. - // The active maximum quantizer insures that an appropriate - // number of bits will be spent if needed for constructed ARFs. - target = min_frame_target; - } + const int min_frame_target = + VPXMAX(rc->min_frame_bandwidth, rc->avg_frame_bandwidth >> 5); + if (target < min_frame_target) target = min_frame_target; + if (cpi->refresh_golden_frame && rc->is_src_frame_alt_ref) { + // If there is an active ARF at this location use the minimum + // bits on this frame even if it is a constructed arf. + // The active maximum quantizer insures that an appropriate + // number of bits will be spent if needed for constructed ARFs. + target = min_frame_target; } // Clip the frame target to the maximum allowed value. @@ -437,6 +435,7 @@ void vp9_rc_init(const VP9EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) { rc->last_post_encode_dropped_scene_change = 0; rc->use_post_encode_drop = 0; rc->ext_use_post_encode_drop = 0; + rc->arf_active_best_quality_adjustment_factor = 1.0; } static int check_buffer_above_thresh(VP9_COMP *cpi, int drop_mark) { @@ -1417,6 +1416,8 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi, int *bottom_index, int active_worst_quality = cpi->twopass.active_worst_quality; int q; int *inter_minq; + int arf_active_best_quality_adjustment, arf_active_best_quality_max; + int *arfgf_high_motion_minq; const int boost_frame = !rc->is_src_frame_alt_ref && (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame); @@ -1441,14 +1442,17 @@ static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi, int *bottom_index, // For constrained quality dont allow Q less than the cq level if (oxcf->rc_mode == VPX_CQ) { if (q < cq_level) q = cq_level; + } + active_best_quality = get_gf_active_quality(cpi, q, cm->bit_depth); - active_best_quality = get_gf_active_quality(cpi, q, cm->bit_depth); + ASSIGN_MINQ_TABLE(cm->bit_depth, arfgf_high_motion_minq); + arf_active_best_quality_max = arfgf_high_motion_minq[q]; + arf_active_best_quality_adjustment = + arf_active_best_quality_max - active_best_quality; + active_best_quality = arf_active_best_quality_max - + (int)(arf_active_best_quality_adjustment * + rc->arf_active_best_quality_adjustment_factor); - // Constrained quality use slightly lower active best. - active_best_quality = active_best_quality * 15 / 16; - } 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) { @@ -3049,7 +3053,7 @@ int vp9_encodedframe_overshoot(VP9_COMP *cpi, int frame_size, int *q) { // Lower thresh_qp for video (more overshoot at lower Q) to be // more conservative for video. if (cpi->oxcf.content != VP9E_CONTENT_SCREEN) - thresh_qp = rc->worst_quality >> 1; + thresh_qp = 3 * (rc->worst_quality >> 2); // If this decision is not based on an encoded frame size but just on // scene/slide change detection (i.e., re_encode_overshoot_cbr_rt == // FAST_DETECTION_MAXQ), for now skip the (frame_size > thresh_rate) diff --git a/vp9/encoder/vp9_ratectrl.h b/vp9/encoder/vp9_ratectrl.h index a5c1f4cf0..2c2048edc 100644 --- a/vp9/encoder/vp9_ratectrl.h +++ b/vp9/encoder/vp9_ratectrl.h @@ -197,6 +197,8 @@ typedef struct { int ext_use_post_encode_drop; int damped_adjustment[RATE_FACTOR_LEVELS]; + double arf_active_best_quality_adjustment_factor; + int arf_active_best_quality_adjustment_window; } RATE_CONTROL; struct VP9_COMP; diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c index 18130dded..df2afceda 100644 --- a/vp9/encoder/vp9_speed_features.c +++ b/vp9/encoder/vp9_speed_features.c @@ -725,6 +725,10 @@ static void set_rt_speed_feature_framesize_independent( if (!frame_is_intra_only(cm) && cm->width * cm->height <= 352 * 288) { sf->nonrd_use_ml_partition = 1; } +#if CONFIG_VP9_HIGHBITDEPTH + if (cpi->Source->flags & YV12_FLAG_HIGHBITDEPTH) + sf->nonrd_use_ml_partition = 0; +#endif if (content == VP9E_CONTENT_SCREEN) sf->mv.subpel_force_stop = HALF_PEL; // Only keep INTRA_DC mode for speed 8. if (!is_keyframe) { |