diff options
| -rw-r--r-- | vp9/encoder/vp9_encoder.c | 85 | ||||
| -rw-r--r-- | vp9/encoder/vp9_encoder.h | 8 |
2 files changed, 93 insertions, 0 deletions
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c index ed7f0a110..0953e7734 100644 --- a/vp9/encoder/vp9_encoder.c +++ b/vp9/encoder/vp9_encoder.c @@ -2345,6 +2345,8 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf, cpi->tpl_stats[frame].width = mi_cols; cpi->tpl_stats[frame].height = mi_rows; cpi->tpl_stats[frame].stride = mi_cols; + cpi->tpl_stats[frame].mi_rows = cm->mi_rows; + cpi->tpl_stats[frame].mi_cols = cm->mi_cols; } vp9_set_speed_features_framesize_independent(cpi); @@ -5386,6 +5388,75 @@ uint32_t motion_compensated_prediction(VP9_COMP *cpi, ThreadData *td, return bestsme; } +int get_overlap_area(int grid_pos_row, int grid_pos_col, int ref_pos_row, + int ref_pos_col, int block) { + int overlap_area; + int width = 0, height = 0; + switch (block) { + case 0: + width = grid_pos_col + MI_SIZE - ref_pos_col; + height = grid_pos_row + MI_SIZE - ref_pos_row; + break; + case 1: + width = ref_pos_col + MI_SIZE - grid_pos_col; + height = grid_pos_row + MI_SIZE - ref_pos_row; + break; + case 2: + width = grid_pos_col + MI_SIZE - ref_pos_col; + height = ref_pos_row + MI_SIZE - grid_pos_row; + break; + case 3: + width = ref_pos_col + MI_SIZE - grid_pos_col; + height = ref_pos_row + MI_SIZE - grid_pos_row; + break; + default: assert(0); + } + + return overlap_area = width * height; +} + +void tpl_model_update(TplDepFrame *tpl_frame, TplDepStats *tpl_stats, + int mi_row, int mi_col) { + TplDepFrame *ref_tpl_frame = &tpl_frame[tpl_stats->ref_frame_index]; + TplDepStats *ref_stats = ref_tpl_frame->tpl_stats_ptr; + MV mv = tpl_stats->mv.as_mv; + int mv_row = mv.row >> 3; + int mv_col = mv.col >> 3; + + int ref_pos_row = mi_row * MI_SIZE + mv_row; + int ref_pos_col = mi_col * MI_SIZE + mv_col; + + // top-left on grid block location + int grid_pos_row_base = (ref_pos_row >> MI_SIZE_LOG2) << MI_SIZE_LOG2; + int grid_pos_col_base = (ref_pos_col >> MI_SIZE_LOG2) << MI_SIZE_LOG2; + int block; + + for (block = 0; block < 4; ++block) { + int grid_pos_row = grid_pos_row_base + MI_SIZE * (block >> 1); + int grid_pos_col = grid_pos_col_base + MI_SIZE * (block & 0x01); + + if (grid_pos_row >= 0 && grid_pos_row < ref_tpl_frame->mi_rows * MI_SIZE && + grid_pos_col >= 0 && grid_pos_col < ref_tpl_frame->mi_cols * MI_SIZE) { + int overlap_area = get_overlap_area(grid_pos_row, grid_pos_col, + ref_pos_row, ref_pos_col, block); + int ref_mi_row = grid_pos_row >> MI_SIZE_LOG2; + int ref_mi_col = grid_pos_col >> MI_SIZE_LOG2; + + int64_t mc_flow = tpl_stats->mc_dep_cost - + (tpl_stats->mc_dep_cost * tpl_stats->inter_cost) / + tpl_stats->intra_cost; + + ref_stats[ref_mi_row * ref_tpl_frame->stride + ref_mi_col].mc_flow += + (mc_flow * overlap_area) >> (MI_SIZE_LOG2 * 2); + + ref_stats[ref_mi_row * ref_tpl_frame->stride + ref_mi_col].mc_ref_cost += + ((tpl_stats->intra_cost - tpl_stats->inter_cost) * overlap_area) >> + (MI_SIZE_LOG2 * 2); + assert(overlap_area >= 0); + } + } +} + void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture, int frame_idx) { TplDepFrame *tpl_frame = &cpi->tpl_stats[frame_idx]; YV12_BUFFER_CONFIG *this_frame = gf_picture[frame_idx].frame; @@ -5447,6 +5518,7 @@ void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture, int frame_idx) { set_error_per_bit(&cpi->td.mb, rdmult); vp9_initialize_me_consts(cpi, &cpi->td.mb, ARNR_FILT_QINDEX); + tpl_frame->is_valid = 1; for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row) { // Motion estimation row boundary x->mv_limits.row_min = -((mi_row * MI_SIZE) + (17 - 2 * VP9_INTERP_EXTEND)); @@ -5464,6 +5536,10 @@ void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture, int frame_idx) { int64_t best_intra_cost = INT64_MAX; int64_t intra_cost; PREDICTION_MODE mode; + + TplDepStats *tpl_stats = + &tpl_frame->tpl_stats_ptr[mi_row * tpl_frame->stride + mi_col]; + // Intra prediction search for (mode = DC_PRED; mode <= TM_PRED; ++mode) { uint8_t *src, *dst; @@ -5564,6 +5640,15 @@ void mc_flow_dispenser(VP9_COMP *cpi, GF_PICTURE *gf_picture, int frame_idx) { } // Motion flow dependency dispenser. + best_intra_cost = VPXMAX(best_intra_cost, 1); + best_inter_cost = VPXMIN(best_inter_cost, best_intra_cost); + tpl_stats->inter_cost = best_inter_cost << TPL_DEP_COST_SCALE_LOG2; + tpl_stats->intra_cost = best_intra_cost << TPL_DEP_COST_SCALE_LOG2; + tpl_stats->mc_dep_cost = tpl_stats->intra_cost + tpl_stats->mc_flow; + tpl_stats->ref_frame_index = gf_picture[frame_idx].ref_frame[best_rf_idx]; + tpl_stats->mv.as_int = best_mv.as_int; + + tpl_model_update(cpi->tpl_stats, tpl_stats, mi_row, mi_col); (void)best_mv; (void)best_rf_idx; } diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h index d6d32abf9..0f9bd8dd6 100644 --- a/vp9/encoder/vp9_encoder.h +++ b/vp9/encoder/vp9_encoder.h @@ -283,6 +283,10 @@ typedef struct TplDepStats { uint64_t inter_cost; uint64_t mc_flow; uint64_t mc_dep_cost; + uint64_t mc_ref_cost; + + int ref_frame_index; + int_mv mv; } TplDepStats; typedef struct TplDepFrame { @@ -291,8 +295,12 @@ typedef struct TplDepFrame { int stride; int width; int height; + int mi_rows; + int mi_cols; } TplDepFrame; +#define TPL_DEP_COST_SCALE_LOG2 16 + // TODO(jingning) All spatially adaptive variables should go to TileDataEnc. typedef struct TileDataEnc { TileInfo tile_info; |