diff options
Diffstat (limited to 'vp9/encoder/vp9_encodeframe.c')
-rw-r--r-- | vp9/encoder/vp9_encodeframe.c | 876 |
1 files changed, 556 insertions, 320 deletions
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c index a38592240..80abb1fa5 100644 --- a/vp9/encoder/vp9_encodeframe.c +++ b/vp9/encoder/vp9_encodeframe.c @@ -30,6 +30,9 @@ #include "vp9/common/vp9_systemdependent.h" #include "vp9/common/vp9_tile_common.h" +#include "vp9/encoder/vp9_aq_complexity.h" +#include "vp9/encoder/vp9_aq_cyclicrefresh.h" +#include "vp9/encoder/vp9_aq_variance.h" #include "vp9/encoder/vp9_encodeframe.h" #include "vp9/encoder/vp9_encodemb.h" #include "vp9/encoder/vp9_encodemv.h" @@ -38,8 +41,6 @@ #include "vp9/encoder/vp9_rdopt.h" #include "vp9/encoder/vp9_segmentation.h" #include "vp9/encoder/vp9_tokenize.h" -#include "vp9/encoder/vp9_vaq.h" -#include "vp9/encoder/vp9_craq.h" #define GF_ZEROMV_ZBIN_BOOST 0 #define LF_ZEROMV_ZBIN_BOOST 0 @@ -164,13 +165,12 @@ static INLINE void set_modeinfo_offsets(VP9_COMMON *const cm, int mi_col) { const int idx_str = xd->mode_info_stride * mi_row + mi_col; xd->mi_8x8 = cm->mi_grid_visible + idx_str; - xd->prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str; xd->mi_8x8[0] = cm->mi + idx_str; } -static int is_block_in_mb_map(VP9_COMP *cpi, int mi_row, int mi_col, +static int is_block_in_mb_map(const VP9_COMP *cpi, int mi_row, int mi_col, BLOCK_SIZE bsize) { - VP9_COMMON *const cm = &cpi->common; + const VP9_COMMON *const cm = &cpi->common; const int mb_rows = cm->mb_rows; const int mb_cols = cm->mb_cols; const int mb_row = mi_row >> 1; @@ -194,6 +194,16 @@ static int is_block_in_mb_map(VP9_COMP *cpi, int mi_row, int mi_col, return 0; } +static int check_active_map(const VP9_COMP *cpi, const MACROBLOCK *x, + int mi_row, int mi_col, + BLOCK_SIZE bsize) { + if (cpi->active_map_enabled && !x->e_mbd.lossless) { + return is_block_in_mb_map(cpi, mi_row, mi_col, bsize); + } else { + return 1; + } +} + static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile, int mi_row, int mi_col, BLOCK_SIZE bsize) { MACROBLOCK *const x = &cpi->mb; @@ -207,16 +217,11 @@ static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile, const int idx_map = mb_row * cm->mb_cols + mb_col; const struct segmentation *const seg = &cm->seg; - set_skip_context(xd, cpi->above_context, cpi->left_context, mi_row, mi_col); + set_skip_context(xd, xd->above_context, xd->left_context, mi_row, mi_col); // Activity map pointer x->mb_activity_ptr = &cpi->mb_activity_map[idx_map]; - - if (cpi->active_map_enabled && !x->e_mbd.lossless) { - x->in_active_map = is_block_in_mb_map(cpi, mi_row, mi_col, bsize); - } else { - x->in_active_map = 1; - } + x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize); set_modeinfo_offsets(cm, xd, mi_row, mi_col); @@ -276,7 +281,7 @@ static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile, } } -static void duplicate_modeinfo_in_sb(VP9_COMMON * const cm, +static void duplicate_mode_info_in_sb(VP9_COMMON * const cm, MACROBLOCKD *const xd, int mi_row, int mi_col, @@ -300,7 +305,7 @@ static void set_block_size(VP9_COMP * const cpi, MACROBLOCKD *const xd = &cpi->mb.e_mbd; set_modeinfo_offsets(&cpi->common, xd, mi_row, mi_col); xd->mi_8x8[0]->mbmi.sb_type = bsize; - duplicate_modeinfo_in_sb(&cpi->common, xd, mi_row, mi_col, bsize); + duplicate_mode_info_in_sb(&cpi->common, xd, mi_row, mi_col, bsize); } } @@ -829,52 +834,6 @@ static void activity_masking(VP9_COMP *cpi, MACROBLOCK *x) { adjust_act_zbin(cpi, x); } -// Select a segment for the current SB64 -static void select_in_frame_q_segment(VP9_COMP *cpi, - int mi_row, int mi_col, - int output_enabled, int projected_rate) { - VP9_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 = MIN(cm->mi_cols - mi_col, bw); - const int ymis = MIN(cm->mi_rows - mi_row, bh); - int complexity_metric = 64; - int x, y; - - unsigned char segment; - - if (!output_enabled) { - segment = 0; - } 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); - - if (projected_rate < (target_rate / 4)) { - segment = 1; - } else { - segment = 0; - } - - if (target_rate > 0) { - complexity_metric = - clamp((int)((projected_rate * 64) / target_rate), 16, 255); - } - } - - // 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; - cpi->complexity_map[mi_offset + y * cm->mi_cols + x] = - (unsigned char)complexity_metric; - } - } -} - static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col, BLOCK_SIZE bsize, int output_enabled) { @@ -896,21 +855,17 @@ static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx, assert(mi->mbmi.sb_type == bsize); - // For in frame adaptive Q copy over the chosen segment id into the - // mode innfo context for the chosen mode / partition. - if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ || - cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) && + *mi_addr = *mi; + + // For in frame adaptive Q, check for reseting the segment_id and updating + // the cyclic refresh map. + if ((cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) && seg->enabled && output_enabled) { - // Check for reseting segment_id and update cyclic map. - if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && seg->enabled) { - vp9_update_segment_aq(cpi, xd->mi_8x8[0], mi_row, mi_col, bsize, 1); - vp9_init_plane_quantizers(cpi, x); - } - mi->mbmi.segment_id = xd->mi_8x8[0]->mbmi.segment_id; + vp9_cyclic_refresh_update_segment(cpi, &xd->mi_8x8[0]->mbmi, + mi_row, mi_col, bsize, 1); + vp9_init_plane_quantizers(cpi, x); } - *mi_addr = *mi; - max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1; for (i = 0; i < max_plane; ++i) { p[i].coeff = ctx->coeff_pbuf[i][1]; @@ -985,12 +940,7 @@ static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx, #endif if (!frame_is_intra_only(cm)) { if (is_inter_block(mbmi)) { - if (mbmi->sb_type < BLOCK_8X8 || mbmi->mode == NEWMV) { - MV best_mv[2]; - for (i = 0; i < 1 + has_second_ref(mbmi); ++i) - best_mv[i] = mbmi->ref_mvs[mbmi->ref_frame[i]][0].as_mv; - vp9_update_mv_count(cm, xd, best_mv); - } + vp9_update_mv_count(cm, xd); if (cm->interp_filter == SWITCHABLE) { const int ctx = vp9_get_pred_context_switchable_interp(xd); @@ -1111,7 +1061,7 @@ static void rd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile, : cm->last_frame_seg_map; // If segment 1, use rdmult for that segment. if (vp9_get_segment_id(cm, map, bsize, mi_row, mi_col)) - x->rdmult = cpi->cyclic_refresh.rdmult; + x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh); } // Find best coding mode & reconstruct the MB so it is available @@ -1211,21 +1161,21 @@ static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col, int mi_height = num_8x8_blocks_high_lookup[bsize]; for (p = 0; p < MAX_MB_PLANE; p++) { vpx_memcpy( - cpi->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x), + xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x), a + num_4x4_blocks_wide * p, (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >> xd->plane[p].subsampling_x); vpx_memcpy( - cpi->left_context[p] + xd->left_context[p] + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y), l + num_4x4_blocks_high * p, (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >> xd->plane[p].subsampling_y); } - vpx_memcpy(cpi->above_seg_context + mi_col, sa, - sizeof(*cpi->above_seg_context) * mi_width); - vpx_memcpy(cpi->left_seg_context + (mi_row & MI_MASK), sl, - sizeof(cpi->left_seg_context[0]) * mi_height); + vpx_memcpy(xd->above_seg_context + mi_col, sa, + sizeof(*xd->above_seg_context) * mi_width); + vpx_memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl, + sizeof(xd->left_seg_context[0]) * mi_height); } static void save_context(VP9_COMP *cpi, int mi_row, int mi_col, ENTROPY_CONTEXT a[16 * MAX_MB_PLANE], @@ -1244,20 +1194,20 @@ static void save_context(VP9_COMP *cpi, int mi_row, int mi_col, for (p = 0; p < MAX_MB_PLANE; ++p) { vpx_memcpy( a + num_4x4_blocks_wide * p, - cpi->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x), + xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x), (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >> xd->plane[p].subsampling_x); vpx_memcpy( l + num_4x4_blocks_high * p, - cpi->left_context[p] + xd->left_context[p] + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y), (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >> xd->plane[p].subsampling_y); } - vpx_memcpy(sa, cpi->above_seg_context + mi_col, - sizeof(*cpi->above_seg_context) * mi_width); - vpx_memcpy(sl, cpi->left_seg_context + (mi_row & MI_MASK), - sizeof(cpi->left_seg_context[0]) * mi_height); + vpx_memcpy(sa, xd->above_seg_context + mi_col, + sizeof(*xd->above_seg_context) * mi_width); + vpx_memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK), + sizeof(xd->left_seg_context[0]) * mi_height); } static void encode_b(VP9_COMP *cpi, const TileInfo *const tile, @@ -1289,6 +1239,8 @@ static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile, int output_enabled, BLOCK_SIZE bsize) { VP9_COMMON *const cm = &cpi->common; MACROBLOCK *const x = &cpi->mb; + MACROBLOCKD *const xd = &x->e_mbd; + const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4; int ctx; PARTITION_TYPE partition; @@ -1298,8 +1250,7 @@ static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile, return; if (bsize >= BLOCK_8X8) { - ctx = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context, - mi_row, mi_col, bsize); + ctx = partition_plane_context(xd, mi_row, mi_col, bsize); subsize = *get_sb_partitioning(x, bsize); } else { ctx = 0; @@ -1354,8 +1305,7 @@ static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile, } if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8) - update_partition_context(cpi->above_seg_context, cpi->left_seg_context, - mi_row, mi_col, subsize, bsize); + update_partition_context(xd, mi_row, mi_col, subsize, bsize); } // Check to see if the given partition size is allowed for a specified number @@ -1432,6 +1382,7 @@ static void copy_partitioning(VP9_COMMON *cm, MODE_INFO **mi_8x8, for (block_col = 0; block_col < 8; ++block_col) { MODE_INFO *const prev_mi = prev_mi_8x8[block_row * mis + block_col]; const BLOCK_SIZE sb_type = prev_mi ? prev_mi->mbmi.sb_type : 0; + if (prev_mi) { const ptrdiff_t offset = prev_mi - cm->prev_mi; mi_8x8[block_row * mis + block_col] = cm->mi + offset; @@ -1460,45 +1411,39 @@ static int sb_has_motion(const VP9_COMMON *cm, MODE_INFO **prev_mi_8x8) { return 0; } -static void update_state_rt(VP9_COMP *cpi, const PICK_MODE_CONTEXT *ctx, +static void update_state_rt(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col, int bsize) { - int i; VP9_COMMON *const cm = &cpi->common; MACROBLOCK *const x = &cpi->mb; MACROBLOCKD *const xd = &x->e_mbd; MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi; const struct segmentation *const seg = &cm->seg; - // TODO(jingning) We might need PICK_MODE_CONTEXT to buffer coding modes - // associated with variable block sizes. Otherwise, remove this ctx - // from argument list. - (void)ctx; + *(xd->mi_8x8[0]) = ctx->mic; - // Check for reseting segment_id and update cyclic map. - if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && seg->enabled) { - vp9_update_segment_aq(cpi, xd->mi_8x8[0], mi_row, mi_col, bsize, 1); + // For in frame adaptive Q, check for reseting the segment_id and updating + // the cyclic refresh map. + if ((cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) && seg->enabled) { + vp9_cyclic_refresh_update_segment(cpi, &xd->mi_8x8[0]->mbmi, + mi_row, mi_col, bsize, 1); vp9_init_plane_quantizers(cpi, x); } - if (is_inter_block(mbmi)) { - if (mbmi->sb_type < BLOCK_8X8 || mbmi->mode == NEWMV) { - MV best_mv[2]; - for (i = 0; i < 1 + has_second_ref(mbmi); ++i) - best_mv[i] = mbmi->ref_mvs[mbmi->ref_frame[i]][0].as_mv; - vp9_update_mv_count(cm, xd, best_mv); - } + vp9_update_mv_count(cm, xd); if (cm->interp_filter == SWITCHABLE) { const int pred_ctx = vp9_get_pred_context_switchable_interp(xd); ++cm->counts.switchable_interp[pred_ctx][mbmi->interp_filter]; } } + + x->skip = ctx->skip; } static void encode_b_rt(VP9_COMP *cpi, const TileInfo *const tile, - TOKENEXTRA **tp, int mi_row, int mi_col, - int output_enabled, BLOCK_SIZE bsize) { + TOKENEXTRA **tp, int mi_row, int mi_col, + int output_enabled, BLOCK_SIZE bsize) { MACROBLOCK *const x = &cpi->mb; if (bsize < BLOCK_8X8) { @@ -1507,6 +1452,7 @@ static void encode_b_rt(VP9_COMP *cpi, const TileInfo *const tile, if (x->ab_index > 0) return; } + set_offsets(cpi, tile, mi_row, mi_col, bsize); update_state_rt(cpi, get_block_context(x, bsize), mi_row, mi_col, bsize); @@ -1522,6 +1468,8 @@ static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile, int output_enabled, BLOCK_SIZE bsize) { VP9_COMMON *const cm = &cpi->common; MACROBLOCK *const x = &cpi->mb; + MACROBLOCKD *const xd = &x->e_mbd; + const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4; int ctx; PARTITION_TYPE partition; @@ -1534,8 +1482,7 @@ static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile, MACROBLOCKD *const xd = &cpi->mb.e_mbd; const int idx_str = xd->mode_info_stride * mi_row + mi_col; MODE_INFO ** mi_8x8 = cm->mi_grid_visible + idx_str; - ctx = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context, - mi_row, mi_col, bsize); + ctx = partition_plane_context(xd, mi_row, mi_col, bsize); subsize = mi_8x8[0]->mbmi.sb_type; } else { ctx = 0; @@ -1594,8 +1541,7 @@ static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile, } if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8) - update_partition_context(cpi->above_seg_context, cpi->left_seg_context, - mi_row, mi_col, subsize, bsize); + update_partition_context(xd, mi_row, mi_col, subsize, bsize); } static void rd_use_partition(VP9_COMP *cpi, @@ -1606,12 +1552,10 @@ static void rd_use_partition(VP9_COMP *cpi, int do_recon) { VP9_COMMON *const cm = &cpi->common; MACROBLOCK *const x = &cpi->mb; + MACROBLOCKD *const xd = &x->e_mbd; const int mis = cm->mode_info_stride; const int bsl = b_width_log2(bsize); - const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize]; - const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize]; - const int ms = num_4x4_blocks_wide / 2; - const int mh = num_4x4_blocks_high / 2; + const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2; const int bss = (1 << bsl) / 4; int i, pl; PARTITION_TYPE partition = PARTITION_NONE; @@ -1630,10 +1574,14 @@ static void rd_use_partition(VP9_COMP *cpi, BLOCK_SIZE sub_subsize = BLOCK_4X4; int splits_below = 0; BLOCK_SIZE bs_type = mi_8x8[0]->mbmi.sb_type; + int do_partition_search = 1; if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return; + assert(num_4x4_blocks_wide_lookup[bsize] == + num_4x4_blocks_high_lookup[bsize]); + partition = partition_lookup[bsl][bs_type]; subsize = get_subsize(bsize, partition); @@ -1653,9 +1601,22 @@ static void rd_use_partition(VP9_COMP *cpi, if (bsize == BLOCK_16X16) { set_offsets(cpi, tile, mi_row, mi_col, bsize); x->mb_energy = vp9_block_energy(cpi, x, bsize); + } else { + x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize); } - if (cpi->sf.partition_search_type == SEARCH_PARTITION && + if (!x->in_active_map) { + do_partition_search = 0; + if (mi_row + (mi_step >> 1) < cm->mi_rows && + mi_col + (mi_step >> 1) < cm->mi_cols) { + *(get_sb_partitioning(x, bsize)) = bsize; + bs_type = mi_8x8[0]->mbmi.sb_type = bsize; + subsize = bsize; + partition = PARTITION_NONE; + } + } + if (do_partition_search && + cpi->sf.partition_search_type == SEARCH_PARTITION && cpi->sf.adjust_partitioning_from_last_frame) { // Check if any of the sub blocks are further split. if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) { @@ -1673,15 +1634,13 @@ static void rd_use_partition(VP9_COMP *cpi, // If partition is not none try none unless each of the 4 splits are split // even further.. if (partition != PARTITION_NONE && !splits_below && - mi_row + (ms >> 1) < cm->mi_rows && - mi_col + (ms >> 1) < cm->mi_cols) { + mi_row + (mi_step >> 1) < cm->mi_rows && + mi_col + (mi_step >> 1) < cm->mi_cols) { *(get_sb_partitioning(x, bsize)) = bsize; rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &none_rate, &none_dist, bsize, get_block_context(x, bsize), INT64_MAX); - pl = partition_plane_context(cpi->above_seg_context, - cpi->left_seg_context, - mi_row, mi_col, bsize); + pl = partition_plane_context(xd, mi_row, mi_col, bsize); if (none_rate < INT_MAX) { none_rate += x->partition_cost[pl][PARTITION_NONE]; @@ -1706,14 +1665,14 @@ static void rd_use_partition(VP9_COMP *cpi, &last_part_dist, subsize, get_block_context(x, subsize), INT64_MAX); if (last_part_rate != INT_MAX && - bsize >= BLOCK_8X8 && mi_row + (mh >> 1) < cm->mi_rows) { + bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) { int rt = 0; int64_t dt = 0; update_state(cpi, get_block_context(x, subsize), mi_row, mi_col, subsize, 0); encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize); *get_sb_index(x, subsize) = 1; - rd_pick_sb_modes(cpi, tile, mi_row + (ms >> 1), mi_col, &rt, &dt, + rd_pick_sb_modes(cpi, tile, mi_row + (mi_step >> 1), mi_col, &rt, &dt, subsize, get_block_context(x, subsize), INT64_MAX); if (rt == INT_MAX || dt == INT64_MAX) { last_part_rate = INT_MAX; @@ -1731,14 +1690,14 @@ static void rd_use_partition(VP9_COMP *cpi, &last_part_dist, subsize, get_block_context(x, subsize), INT64_MAX); if (last_part_rate != INT_MAX && - bsize >= BLOCK_8X8 && mi_col + (ms >> 1) < cm->mi_cols) { + bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) { int rt = 0; int64_t dt = 0; update_state(cpi, get_block_context(x, subsize), mi_row, mi_col, subsize, 0); encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize); *get_sb_index(x, subsize) = 1; - rd_pick_sb_modes(cpi, tile, mi_row, mi_col + (ms >> 1), &rt, &dt, + rd_pick_sb_modes(cpi, tile, mi_row, mi_col + (mi_step >> 1), &rt, &dt, subsize, get_block_context(x, subsize), INT64_MAX); if (rt == INT_MAX || dt == INT64_MAX) { last_part_rate = INT_MAX; @@ -1754,8 +1713,8 @@ static void rd_use_partition(VP9_COMP *cpi, last_part_rate = 0; last_part_dist = 0; for (i = 0; i < 4; i++) { - int x_idx = (i & 1) * (ms >> 1); - int y_idx = (i >> 1) * (ms >> 1); + int x_idx = (i & 1) * (mi_step >> 1); + int y_idx = (i >> 1) * (mi_step >> 1); int jj = i >> 1, ii = i & 0x01; int rt; int64_t dt; @@ -1781,18 +1740,20 @@ static void rd_use_partition(VP9_COMP *cpi, assert(0); } - pl = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context, - mi_row, mi_col, bsize); + pl = partition_plane_context(xd, mi_row, mi_col, bsize); if (last_part_rate < INT_MAX) { last_part_rate += x->partition_cost[pl][partition]; last_part_rd = RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist); } - if (cpi->sf.adjust_partitioning_from_last_frame + if (do_partition_search + && cpi->sf.adjust_partitioning_from_last_frame && cpi->sf.partition_search_type == SEARCH_PARTITION && partition != PARTITION_SPLIT && bsize > BLOCK_8X8 - && (mi_row + ms < cm->mi_rows || mi_row + (ms >> 1) == cm->mi_rows) - && (mi_col + ms < cm->mi_cols || mi_col + (ms >> 1) == cm->mi_cols)) { + && (mi_row + mi_step < cm->mi_rows || + mi_row + (mi_step >> 1) == cm->mi_rows) + && (mi_col + mi_step < cm->mi_cols || + mi_col + (mi_step >> 1) == cm->mi_cols)) { BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT); chosen_rate = 0; chosen_dist = 0; @@ -1800,8 +1761,8 @@ static void rd_use_partition(VP9_COMP *cpi, // Split partition. for (i = 0; i < 4; i++) { - int x_idx = (i & 1) * (num_4x4_blocks_wide >> 2); - int y_idx = (i >> 1) * (num_4x4_blocks_wide >> 2); + int x_idx = (i & 1) * (mi_step >> 1); + int y_idx = (i >> 1) * (mi_step >> 1); int rt = 0; int64_t dt = 0; ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE]; @@ -1835,14 +1796,11 @@ static void rd_use_partition(VP9_COMP *cpi, encode_sb(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx, 0, split_subsize); - pl = partition_plane_context(cpi->above_seg_context, - cpi->left_seg_context, - mi_row + y_idx, mi_col + x_idx, + pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx, split_subsize); chosen_rate += x->partition_cost[pl][PARTITION_NONE]; } - pl = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context, - mi_row, mi_col, bsize); + pl = partition_plane_context(xd, mi_row, mi_col, bsize); if (chosen_rate < INT_MAX) { chosen_rate += x->partition_cost[pl][PARTITION_SPLIT]; chosen_rd = RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist); @@ -1880,14 +1838,14 @@ static void rd_use_partition(VP9_COMP *cpi, // and and if necessary apply a Q delta using segmentation to get // closer to the target. if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) { - select_in_frame_q_segment(cpi, mi_row, mi_col, - output_enabled, chosen_rate); - } - if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) { - cpi->cyclic_refresh.projected_rate_sb = chosen_rate; - cpi->cyclic_refresh.projected_dist_sb = chosen_dist; + vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, + output_enabled, chosen_rate); } + if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) + vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh, + chosen_rate, chosen_dist); + encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize); } @@ -1951,77 +1909,71 @@ static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = { // Look at neighboring blocks and set a min and max partition size based on // what they chose. static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile, - int row, int col, + int mi_row, int mi_col, BLOCK_SIZE *min_block_size, BLOCK_SIZE *max_block_size) { - VP9_COMMON * const cm = &cpi->common; + VP9_COMMON *const cm = &cpi->common; MACROBLOCKD *const xd = &cpi->mb.e_mbd; - MODE_INFO ** mi_8x8 = xd->mi_8x8; - MODE_INFO ** prev_mi_8x8 = xd->prev_mi_8x8; - + MODE_INFO **mi_8x8 = xd->mi_8x8; const int left_in_image = xd->left_available && mi_8x8[-1]; const int above_in_image = xd->up_available && mi_8x8[-xd->mode_info_stride]; - MODE_INFO ** above_sb64_mi_8x8; - MODE_INFO ** left_sb64_mi_8x8; + MODE_INFO **above_sb64_mi_8x8; + MODE_INFO **left_sb64_mi_8x8; - int row8x8_remaining = tile->mi_row_end - row; - int col8x8_remaining = tile->mi_col_end - col; + int row8x8_remaining = tile->mi_row_end - mi_row; + int col8x8_remaining = tile->mi_col_end - mi_col; int bh, bw; - + BLOCK_SIZE min_size = BLOCK_4X4; + BLOCK_SIZE max_size = BLOCK_64X64; // Trap case where we do not have a prediction. - if (!left_in_image && !above_in_image && - ((cm->frame_type == KEY_FRAME) || !cm->prev_mi)) { - *min_block_size = BLOCK_4X4; - *max_block_size = BLOCK_64X64; - } else { + if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) { // Default "min to max" and "max to min" - *min_block_size = BLOCK_64X64; - *max_block_size = BLOCK_4X4; + min_size = BLOCK_64X64; + max_size = BLOCK_4X4; // NOTE: each call to get_sb_partition_size_range() uses the previous // passed in values for min and max as a starting point. - // // Find the min and max partition used in previous frame at this location - if (cm->prev_mi && (cm->frame_type != KEY_FRAME)) { - get_sb_partition_size_range(cpi, prev_mi_8x8, - min_block_size, max_block_size); + if (cm->frame_type != KEY_FRAME) { + MODE_INFO **const prev_mi = + &cm->prev_mi_grid_visible[mi_row * xd->mode_info_stride + mi_col]; + get_sb_partition_size_range(cpi, prev_mi, &min_size, &max_size); } - // Find the min and max partition sizes used in the left SB64 if (left_in_image) { left_sb64_mi_8x8 = &mi_8x8[-MI_BLOCK_SIZE]; get_sb_partition_size_range(cpi, left_sb64_mi_8x8, - min_block_size, max_block_size); + &min_size, &max_size); } - // Find the min and max partition sizes used in the above SB64. if (above_in_image) { above_sb64_mi_8x8 = &mi_8x8[-xd->mode_info_stride * MI_BLOCK_SIZE]; get_sb_partition_size_range(cpi, above_sb64_mi_8x8, - min_block_size, max_block_size); + &min_size, &max_size); + } + // adjust observed min and max + if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) { + min_size = min_partition_size[min_size]; + max_size = max_partition_size[max_size]; } } - // adjust observed min and max - if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) { - *min_block_size = min_partition_size[*min_block_size]; - *max_block_size = max_partition_size[*max_block_size]; - } - - // Check border cases where max and min from neighbours may not be legal. - *max_block_size = find_partition_size(*max_block_size, - row8x8_remaining, col8x8_remaining, - &bh, &bw); - *min_block_size = MIN(*min_block_size, *max_block_size); + // Check border cases where max and min from neighbors may not be legal. + max_size = find_partition_size(max_size, + row8x8_remaining, col8x8_remaining, + &bh, &bw); + min_size = MIN(min_size, max_size); // When use_square_partition_only is true, make sure at least one square // partition is allowed by selecting the next smaller square size as // *min_block_size. if (cpi->sf.use_square_partition_only && - next_square_size[*max_block_size] < *min_block_size) { - *min_block_size = next_square_size[*max_block_size]; + next_square_size[max_size] < min_size) { + min_size = next_square_size[max_size]; } + *min_block_size = min_size; + *max_block_size = max_size; } static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) { @@ -2041,7 +1993,8 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, int64_t *dist, int do_recon, int64_t best_rd) { VP9_COMMON *const cm = &cpi->common; MACROBLOCK *const x = &cpi->mb; - const int ms = num_8x8_blocks_wide_lookup[bsize] / 2; + MACROBLOCKD *const xd = &x->e_mbd; + const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2; ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE]; PARTITION_CONTEXT sl[8], sa[8]; TOKENEXTRA *tp_orig = *tp; @@ -2054,8 +2007,8 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, int do_split = bsize >= BLOCK_8X8; int do_rect = 1; // Override skipping rectangular partition operations for edge blocks - const int force_horz_split = (mi_row + ms >= cm->mi_rows); - const int force_vert_split = (mi_col + ms >= cm->mi_cols); + const int force_horz_split = (mi_row + mi_step >= cm->mi_rows); + const int force_vert_split = (mi_col + mi_step >= cm->mi_cols); const int xss = x->e_mbd.plane[1].subsampling_x; const int yss = x->e_mbd.plane[1].subsampling_y; @@ -2081,6 +2034,8 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, if (bsize == BLOCK_16X16) { set_offsets(cpi, tile, mi_row, mi_col, bsize); x->mb_energy = vp9_block_energy(cpi, x, bsize); + } else { + x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize); } // Determine partition types in search according to the speed features. @@ -2122,9 +2077,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, ctx, best_rd); if (this_rate != INT_MAX) { if (bsize >= BLOCK_8X8) { - pl = partition_plane_context(cpi->above_seg_context, - cpi->left_seg_context, - mi_row, mi_col, bsize); + pl = partition_plane_context(xd, mi_row, mi_col, bsize); this_rate += x->partition_cost[pl][PARTITION_NONE]; } sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist); @@ -2169,8 +2122,8 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, if (do_split) { subsize = get_subsize(bsize, PARTITION_SPLIT); for (i = 0; i < 4 && sum_rd < best_rd; ++i) { - const int x_idx = (i & 1) * ms; - const int y_idx = (i >> 1) * ms; + const int x_idx = (i & 1) * mi_step; + const int y_idx = (i >> 1) * mi_step; if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols) continue; @@ -2194,9 +2147,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, } } if (sum_rd < best_rd && i == 4) { - pl = partition_plane_context(cpi->above_seg_context, - cpi->left_seg_context, - mi_row, mi_col, bsize); + pl = partition_plane_context(xd, mi_row, mi_col, bsize); sum_rate += x->partition_cost[pl][PARTITION_SPLIT]; sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); if (sum_rd < best_rd) { @@ -2228,7 +2179,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, get_block_context(x, subsize), best_rd); sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); - if (sum_rd < best_rd && mi_row + ms < cm->mi_rows) { + if (sum_rd < best_rd && mi_row + mi_step < cm->mi_rows) { update_state(cpi, get_block_context(x, subsize), mi_row, mi_col, subsize, 0); encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize); @@ -2240,7 +2191,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, partition_none_allowed) get_block_context(x, subsize)->pred_interp_filter = ctx->mic.mbmi.interp_filter; - rd_pick_sb_modes(cpi, tile, mi_row + ms, mi_col, &this_rate, + rd_pick_sb_modes(cpi, tile, mi_row + mi_step, mi_col, &this_rate, &this_dist, subsize, get_block_context(x, subsize), best_rd - sum_rd); if (this_rate == INT_MAX) { @@ -2252,9 +2203,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, } } if (sum_rd < best_rd) { - pl = partition_plane_context(cpi->above_seg_context, - cpi->left_seg_context, - mi_row, mi_col, bsize); + pl = partition_plane_context(xd, mi_row, mi_col, bsize); sum_rate += x->partition_cost[pl][PARTITION_HORZ]; sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); if (sum_rd < best_rd) { @@ -2281,7 +2230,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize, get_block_context(x, subsize), best_rd); sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); - if (sum_rd < best_rd && mi_col + ms < cm->mi_cols) { + if (sum_rd < best_rd && mi_col + mi_step < cm->mi_cols) { update_state(cpi, get_block_context(x, subsize), mi_row, mi_col, subsize, 0); encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize); @@ -2293,7 +2242,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, partition_none_allowed) get_block_context(x, subsize)->pred_interp_filter = ctx->mic.mbmi.interp_filter; - rd_pick_sb_modes(cpi, tile, mi_row, mi_col + ms, &this_rate, + rd_pick_sb_modes(cpi, tile, mi_row, mi_col + mi_step, &this_rate, &this_dist, subsize, get_block_context(x, subsize), best_rd - sum_rd); if (this_rate == INT_MAX) { @@ -2305,9 +2254,7 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, } } if (sum_rd < best_rd) { - pl = partition_plane_context(cpi->above_seg_context, - cpi->left_seg_context, - mi_row, mi_col, bsize); + pl = partition_plane_context(xd, mi_row, mi_col, bsize); sum_rate += x->partition_cost[pl][PARTITION_VERT]; sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); if (sum_rd < best_rd) { @@ -2335,13 +2282,14 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, // and and if necessary apply a Q delta using segmentation to get // closer to the target. if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) { - select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled, best_rate); - } - if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) { - cpi->cyclic_refresh.projected_rate_sb = best_rate; - cpi->cyclic_refresh.projected_dist_sb = best_dist; + vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled, + best_rate); } + if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) + vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh, + best_rate, best_dist); + encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize); } if (bsize == BLOCK_64X64) { @@ -2356,11 +2304,12 @@ static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile, int mi_row, TOKENEXTRA **tp) { VP9_COMMON *const cm = &cpi->common; + MACROBLOCKD *const xd = &cpi->mb.e_mbd; int mi_col; // Initialize the left context for the new SB row - vpx_memset(&cpi->left_context, 0, sizeof(cpi->left_context)); - vpx_memset(cpi->left_seg_context, 0, sizeof(cpi->left_seg_context)); + vpx_memset(&xd->left_context, 0, sizeof(xd->left_context)); + vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context)); // Code each SB in the row for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end; @@ -2486,11 +2435,11 @@ static void init_encode_frame_mb_context(VP9_COMP *cpi) { // Note: this memset assumes above_context[0], [1] and [2] // are allocated as part of the same buffer. - vpx_memset(cpi->above_context[0], 0, - sizeof(*cpi->above_context[0]) * + vpx_memset(xd->above_context[0], 0, + sizeof(*xd->above_context[0]) * 2 * aligned_mi_cols * MAX_MB_PLANE); - vpx_memset(cpi->above_seg_context, 0, - sizeof(*cpi->above_seg_context) * aligned_mi_cols); + vpx_memset(xd->above_seg_context, 0, + sizeof(*xd->above_seg_context) * aligned_mi_cols); } static void switch_lossless_mode(VP9_COMP *cpi, int lossless) { @@ -2533,87 +2482,15 @@ static int get_skip_flag(MODE_INFO **mi_8x8, int mis, int ymbs, int xmbs) { return 1; } -static void set_txfm_flag(MODE_INFO **mi_8x8, int mis, int ymbs, int xmbs, - TX_SIZE tx_size) { - int x, y; - - for (y = 0; y < ymbs; y++) { - for (x = 0; x < xmbs; x++) - mi_8x8[y * mis + x]->mbmi.tx_size = tx_size; - } -} - -static void reset_skip_txfm_size_b(const VP9_COMMON *cm, int mis, - TX_SIZE max_tx_size, int bw, int bh, - int mi_row, int mi_col, - MODE_INFO **mi_8x8) { - if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) { - return; - } else { - const MB_MODE_INFO *const mbmi = &mi_8x8[0]->mbmi; - if (mbmi->tx_size > max_tx_size) { - const int ymbs = MIN(bh, cm->mi_rows - mi_row); - const int xmbs = MIN(bw, cm->mi_cols - mi_col); - - assert(vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP) || - get_skip_flag(mi_8x8, mis, ymbs, xmbs)); - set_txfm_flag(mi_8x8, mis, ymbs, xmbs, max_tx_size); - } - } -} - -static void reset_skip_txfm_size_sb(VP9_COMMON *cm, MODE_INFO **mi_8x8, - TX_SIZE max_tx_size, int mi_row, int mi_col, - BLOCK_SIZE bsize) { - const int mis = cm->mode_info_stride; - int bw, bh; - const int bs = num_8x8_blocks_wide_lookup[bsize], hbs = bs / 2; - - if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) - return; - - bw = num_8x8_blocks_wide_lookup[mi_8x8[0]->mbmi.sb_type]; - bh = num_8x8_blocks_high_lookup[mi_8x8[0]->mbmi.sb_type]; - - if (bw == bs && bh == bs) { - reset_skip_txfm_size_b(cm, mis, max_tx_size, bs, bs, mi_row, mi_col, - mi_8x8); - } else if (bw == bs && bh < bs) { - reset_skip_txfm_size_b(cm, mis, max_tx_size, bs, hbs, mi_row, mi_col, - mi_8x8); - reset_skip_txfm_size_b(cm, mis, max_tx_size, bs, hbs, mi_row + hbs, - mi_col, mi_8x8 + hbs * mis); - } else if (bw < bs && bh == bs) { - reset_skip_txfm_size_b(cm, mis, max_tx_size, hbs, bs, mi_row, mi_col, - mi_8x8); - reset_skip_txfm_size_b(cm, mis, max_tx_size, hbs, bs, mi_row, - mi_col + hbs, mi_8x8 + hbs); - } else { - const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize]; - int n; - - assert(bw < bs && bh < bs); - - for (n = 0; n < 4; n++) { - const int mi_dc = hbs * (n & 1); - const int mi_dr = hbs * (n >> 1); - - reset_skip_txfm_size_sb(cm, &mi_8x8[mi_dr * mis + mi_dc], max_tx_size, - mi_row + mi_dr, mi_col + mi_dc, subsize); - } - } -} - static void reset_skip_txfm_size(VP9_COMMON *cm, TX_SIZE txfm_max) { int mi_row, mi_col; const int mis = cm->mode_info_stride; - MODE_INFO **mi_8x8, **mi_ptr = cm->mi_grid_visible; + MODE_INFO **mi_ptr = cm->mi_grid_visible; - for (mi_row = 0; mi_row < cm->mi_rows; mi_row += 8, mi_ptr += 8 * mis) { - mi_8x8 = mi_ptr; - for (mi_col = 0; mi_col < cm->mi_cols; mi_col += 8, mi_8x8 += 8) { - reset_skip_txfm_size_sb(cm, mi_8x8, txfm_max, mi_row, mi_col, - BLOCK_64X64); + for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) { + for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) { + if (mi_ptr[mi_col]->mbmi.tx_size > txfm_max) + mi_ptr[mi_col]->mbmi.tx_size = txfm_max; } } } @@ -2693,6 +2570,7 @@ static void nonrd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile, MACROBLOCKD *const xd = &x->e_mbd; set_offsets(cpi, tile, mi_row, mi_col, bsize); xd->mi_8x8[0]->mbmi.sb_type = bsize; + if (!frame_is_intra_only(cm)) { vp9_pick_inter_mode(cpi, x, tile, mi_row, mi_col, rate, dist, bsize); @@ -2700,7 +2578,338 @@ static void nonrd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile, MB_PREDICTION_MODE intramode = DC_PRED; set_mode_info(&xd->mi_8x8[0]->mbmi, bsize, intramode); } - duplicate_modeinfo_in_sb(cm, xd, mi_row, mi_col, bsize); + duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize); +} + +static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x, + int mi_row, int mi_col, + BLOCK_SIZE bsize, BLOCK_SIZE subsize) { + MACROBLOCKD *xd = &x->e_mbd; + int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4; + PARTITION_TYPE partition = partition_lookup[bsl][subsize]; + + assert(bsize >= BLOCK_8X8); + + if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) + return; + + switch (partition) { + case PARTITION_NONE: + set_modeinfo_offsets(cm, xd, mi_row, mi_col); + *(xd->mi_8x8[0]) = (get_block_context(x, subsize))->mic; + duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize); + break; + case PARTITION_VERT: + *get_sb_index(x, subsize) = 0; + set_modeinfo_offsets(cm, xd, mi_row, mi_col); + *(xd->mi_8x8[0]) = (get_block_context(x, subsize))->mic; + duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize); + + if (mi_col + hbs < cm->mi_cols) { + *get_sb_index(x, subsize) = 1; + set_modeinfo_offsets(cm, xd, mi_row, mi_col + hbs); + *(xd->mi_8x8[0]) = (get_block_context(x, subsize))->mic; + duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, bsize); + } + break; + case PARTITION_HORZ: + *get_sb_index(x, subsize) = 0; + set_modeinfo_offsets(cm, xd, mi_row, mi_col); + *(xd->mi_8x8[0]) = (get_block_context(x, subsize))->mic; + duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize); + if (mi_row + hbs < cm->mi_rows) { + *get_sb_index(x, subsize) = 1; + set_modeinfo_offsets(cm, xd, mi_row + hbs, mi_col); + *(xd->mi_8x8[0]) = (get_block_context(x, subsize))->mic; + duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, bsize); + } + break; + case PARTITION_SPLIT: + *get_sb_index(x, subsize) = 0; + fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, + *(get_sb_partitioning(x, subsize))); + *get_sb_index(x, subsize) = 1; + fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize, + *(get_sb_partitioning(x, subsize))); + *get_sb_index(x, subsize) = 2; + fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize, + *(get_sb_partitioning(x, subsize))); + *get_sb_index(x, subsize) = 3; + fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize, + *(get_sb_partitioning(x, subsize))); + break; + default: + break; + } +} + +static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile, + TOKENEXTRA **tp, int mi_row, + int mi_col, BLOCK_SIZE bsize, int *rate, + int64_t *dist, int do_recon, int64_t best_rd) { + VP9_COMMON *const cm = &cpi->common; + MACROBLOCK *const x = &cpi->mb; + MACROBLOCKD *const xd = &x->e_mbd; + const int ms = num_8x8_blocks_wide_lookup[bsize] / 2; + TOKENEXTRA *tp_orig = *tp; + PICK_MODE_CONTEXT *ctx = get_block_context(x, bsize); + int i; + BLOCK_SIZE subsize; + int this_rate, sum_rate = 0, best_rate = INT_MAX; + int64_t this_dist, sum_dist = 0, best_dist = INT64_MAX; + int64_t sum_rd = 0; + int do_split = bsize >= BLOCK_8X8; + int do_rect = 1; + // Override skipping rectangular partition operations for edge blocks + const int force_horz_split = (mi_row + ms >= cm->mi_rows); + const int force_vert_split = (mi_col + ms >= cm->mi_cols); + const int xss = x->e_mbd.plane[1].subsampling_x; + const int yss = x->e_mbd.plane[1].subsampling_y; + + int partition_none_allowed = !force_horz_split && !force_vert_split; + int partition_horz_allowed = !force_vert_split && yss <= xss && + bsize >= BLOCK_8X8; + int partition_vert_allowed = !force_horz_split && xss <= yss && + bsize >= BLOCK_8X8; + (void) *tp_orig; + + if (bsize < BLOCK_8X8) { + // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0 + // there is nothing to be done. + if (x->ab_index != 0) { + *rate = 0; + *dist = 0; + return; + } + } + + assert(num_8x8_blocks_wide_lookup[bsize] == + num_8x8_blocks_high_lookup[bsize]); + + x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize); + + // Determine partition types in search according to the speed features. + // The threshold set here has to be of square block size. + if (cpi->sf.auto_min_max_partition_size) { + partition_none_allowed &= (bsize <= cpi->sf.max_partition_size && + bsize >= cpi->sf.min_partition_size); + partition_horz_allowed &= ((bsize <= cpi->sf.max_partition_size && + bsize > cpi->sf.min_partition_size) || + force_horz_split); + partition_vert_allowed &= ((bsize <= cpi->sf.max_partition_size && + bsize > cpi->sf.min_partition_size) || + force_vert_split); + do_split &= bsize > cpi->sf.min_partition_size; + } + if (cpi->sf.use_square_partition_only) { + partition_horz_allowed &= force_horz_split; + partition_vert_allowed &= force_vert_split; + } + + if (!x->in_active_map && (partition_horz_allowed || partition_vert_allowed)) + do_split = 0; + + // PARTITION_NONE + if (partition_none_allowed) { + nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, + &this_rate, &this_dist, bsize); + ctx->mic.mbmi = xd->mi_8x8[0]->mbmi; + + if (this_rate != INT_MAX) { + int pl = partition_plane_context(xd, mi_row, mi_col, bsize); + this_rate += x->partition_cost[pl][PARTITION_NONE]; + sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist); + if (sum_rd < best_rd) { + int64_t stop_thresh = 4096; + int64_t stop_thresh_rd; + + best_rate = this_rate; + best_dist = this_dist; + best_rd = sum_rd; + if (bsize >= BLOCK_8X8) + *(get_sb_partitioning(x, bsize)) = bsize; + + // Adjust threshold according to partition size. + stop_thresh >>= 8 - (b_width_log2_lookup[bsize] + + b_height_log2_lookup[bsize]); + + stop_thresh_rd = RDCOST(x->rdmult, x->rddiv, 0, stop_thresh); + // If obtained distortion is very small, choose current partition + // and stop splitting. + if (!x->e_mbd.lossless && best_rd < stop_thresh_rd) { + do_split = 0; + do_rect = 0; + } + } + } + if (!x->in_active_map) { + do_split = 0; + do_rect = 0; + } + } + + // store estimated motion vector + store_pred_mv(x, ctx); + + // PARTITION_SPLIT + sum_rd = 0; + if (do_split) { + int pl = partition_plane_context(xd, mi_row, mi_col, bsize); + sum_rate += x->partition_cost[pl][PARTITION_SPLIT]; + subsize = get_subsize(bsize, PARTITION_SPLIT); + for (i = 0; i < 4 && sum_rd < best_rd; ++i) { + const int x_idx = (i & 1) * ms; + const int y_idx = (i >> 1) * ms; + + if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols) + continue; + + *get_sb_index(x, subsize) = i; + load_pred_mv(x, ctx); + + nonrd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx, + subsize, &this_rate, &this_dist, 0, + best_rd - sum_rd); + + if (this_rate == INT_MAX) { + sum_rd = INT64_MAX; + } else { + sum_rate += this_rate; + sum_dist += this_dist; + sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); + } + } + + if (sum_rd < best_rd) { + best_rate = sum_rate; + best_dist = sum_dist; + best_rd = sum_rd; + *(get_sb_partitioning(x, bsize)) = subsize; + } else { + // skip rectangular partition test when larger block size + // gives better rd cost + if (cpi->sf.less_rectangular_check) + do_rect &= !partition_none_allowed; + } + } + + // PARTITION_HORZ + if (partition_horz_allowed && do_rect) { + subsize = get_subsize(bsize, PARTITION_HORZ); + *get_sb_index(x, subsize) = 0; + if (cpi->sf.adaptive_motion_search) + load_pred_mv(x, ctx); + + nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, + &this_rate, &this_dist, subsize); + + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; + + sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); + + if (sum_rd < best_rd && mi_row + ms < cm->mi_rows) { + *get_sb_index(x, subsize) = 1; + + load_pred_mv(x, ctx); + + nonrd_pick_sb_modes(cpi, tile, mi_row + ms, mi_col, + &this_rate, &this_dist, subsize); + + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; + + if (this_rate == INT_MAX) { + sum_rd = INT64_MAX; + } else { + int pl = partition_plane_context(xd, mi_row, mi_col, bsize); + this_rate += x->partition_cost[pl][PARTITION_HORZ]; + sum_rate += this_rate; + sum_dist += this_dist; + sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); + } + } + if (sum_rd < best_rd) { + best_rd = sum_rd; + best_rate = sum_rate; + best_dist = sum_dist; + *(get_sb_partitioning(x, bsize)) = subsize; + } + } + + // PARTITION_VERT + if (partition_vert_allowed && do_rect) { + subsize = get_subsize(bsize, PARTITION_VERT); + + *get_sb_index(x, subsize) = 0; + if (cpi->sf.adaptive_motion_search) + load_pred_mv(x, ctx); + + nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, + &this_rate, &this_dist, subsize); + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; + sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); + if (sum_rd < best_rd && mi_col + ms < cm->mi_cols) { + *get_sb_index(x, subsize) = 1; + + load_pred_mv(x, ctx); + + nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col + ms, + &this_rate, &this_dist, subsize); + + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; + + if (this_rate == INT_MAX) { + sum_rd = INT64_MAX; + } else { + int pl = partition_plane_context(xd, mi_row, mi_col, bsize); + this_rate += x->partition_cost[pl][PARTITION_VERT]; + sum_rate += this_rate; + sum_dist += this_dist; + sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist); + } + } + if (sum_rd < best_rd) { + best_rate = sum_rate; + best_dist = sum_dist; + best_rd = sum_rd; + *(get_sb_partitioning(x, bsize)) = subsize; + } + } + + *rate = best_rate; + *dist = best_dist; + + if (best_rate == INT_MAX) + return; + + // update mode info array + fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, + *(get_sb_partitioning(x, bsize))); + + if (best_rate < INT_MAX && best_dist < INT64_MAX && do_recon) { + int output_enabled = (bsize == BLOCK_64X64); + + // Check the projected output rate for this SB against it's target + // and and if necessary apply a Q delta using segmentation to get + // closer to the target. + if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) { + vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled, + best_rate); + } + + if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) + vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh, + best_rate, best_dist); + + encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize); + } + + if (bsize == BLOCK_64X64) { + assert(tp_orig < *tp); + assert(best_rate < INT_MAX); + assert(best_dist < INT64_MAX); + } else { + assert(tp_orig == *tp); + } } static void nonrd_use_partition(VP9_COMP *cpi, @@ -2712,12 +2921,13 @@ static void nonrd_use_partition(VP9_COMP *cpi, int *totrate, int64_t *totdist) { VP9_COMMON *const cm = &cpi->common; MACROBLOCK *const x = &cpi->mb; + MACROBLOCKD *const xd = &x->e_mbd; const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4; const int mis = cm->mode_info_stride; PARTITION_TYPE partition; BLOCK_SIZE subsize; - int rate; - int64_t dist; + int rate = INT_MAX; + int64_t dist = INT64_MAX; if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return; @@ -2733,14 +2943,17 @@ static void nonrd_use_partition(VP9_COMP *cpi, switch (partition) { case PARTITION_NONE: nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist, subsize); + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; break; case PARTITION_VERT: *get_sb_index(x, subsize) = 0; nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist, subsize); + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; if (mi_col + hbs < cm->mi_cols) { *get_sb_index(x, subsize) = 1; nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col + hbs, &rate, &dist, subsize); + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; if (rate != INT_MAX && dist != INT64_MAX && *totrate != INT_MAX && *totdist != INT64_MAX) { *totrate += rate; @@ -2751,10 +2964,12 @@ static void nonrd_use_partition(VP9_COMP *cpi, case PARTITION_HORZ: *get_sb_index(x, subsize) = 0; nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist, subsize); + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; if (mi_row + hbs < cm->mi_rows) { *get_sb_index(x, subsize) = 1; nonrd_pick_sb_modes(cpi, tile, mi_row + hbs, mi_col, &rate, &dist, subsize); + (get_block_context(x, subsize))->mic.mbmi = xd->mi_8x8[0]->mbmi; if (rate != INT_MAX && dist != INT64_MAX && *totrate != INT_MAX && *totdist != INT64_MAX) { *totrate += rate; @@ -2764,7 +2979,6 @@ static void nonrd_use_partition(VP9_COMP *cpi, break; case PARTITION_SPLIT: subsize = get_subsize(bsize, PARTITION_SPLIT); - *get_sb_index(x, subsize) = 0; nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, subsize, output_enabled, totrate, totdist); @@ -2801,10 +3015,9 @@ static void nonrd_use_partition(VP9_COMP *cpi, } if (bsize == BLOCK_64X64 && output_enabled) { - if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) { - cpi->cyclic_refresh.projected_rate_sb = *totrate; - cpi->cyclic_refresh.projected_dist_sb = *totdist; - } + if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) + vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh, + *totrate, *totdist); encode_sb_rt(cpi, tile, tp, mi_row, mi_col, 1, bsize); } } @@ -2812,33 +3025,55 @@ static void nonrd_use_partition(VP9_COMP *cpi, static void encode_nonrd_sb_row(VP9_COMP *cpi, const TileInfo *const tile, int mi_row, TOKENEXTRA **tp) { VP9_COMMON *cm = &cpi->common; + MACROBLOCKD *xd = &cpi->mb.e_mbd; int mi_col; // Initialize the left context for the new SB row - vpx_memset(&cpi->left_context, 0, sizeof(cpi->left_context)); - vpx_memset(cpi->left_seg_context, 0, sizeof(cpi->left_seg_context)); + vpx_memset(&xd->left_context, 0, sizeof(xd->left_context)); + vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context)); // Code each SB in the row for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end; mi_col += MI_BLOCK_SIZE) { - int dummy_rate; - int64_t dummy_dist; + int dummy_rate = 0; + int64_t dummy_dist = 0; const int idx_str = cm->mode_info_stride * mi_row + mi_col; MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str; + MODE_INFO **prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str; + BLOCK_SIZE bsize = cpi->sf.partition_search_type == FIXED_PARTITION ? cpi->sf.always_this_block_size : get_nonrd_var_based_fixed_partition(cpi, mi_row, mi_col); cpi->mb.source_variance = UINT_MAX; + vp9_zero(cpi->mb.pred_mv); // Set the partition type of the 64X64 block - if (cpi->sf.partition_search_type == VAR_BASED_PARTITION) - choose_partitioning(cpi, tile, mi_row, mi_col); - else - set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col, bsize); - - nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64, 1, - &dummy_rate, &dummy_dist); + switch (cpi->sf.partition_search_type) { + case VAR_BASED_PARTITION: + choose_partitioning(cpi, tile, mi_row, mi_col); + nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64, + 1, &dummy_rate, &dummy_dist); + break; + case VAR_BASED_FIXED_PARTITION: + case FIXED_PARTITION: + set_fixed_partitioning(cpi, tile, mi_8x8, mi_row, mi_col, bsize); + nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64, + 1, &dummy_rate, &dummy_dist); + break; + case REFERENCE_PARTITION: + if (cpi->sf.partition_check || sb_has_motion(cm, prev_mi_8x8)) { + nonrd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64, + &dummy_rate, &dummy_dist, 1, INT64_MAX); + } else { + copy_partitioning(cm, mi_8x8, prev_mi_8x8); + nonrd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, + BLOCK_64X64, 1, &dummy_rate, &dummy_dist); + } + break; + default: + assert(0); + } } } // end RTC play code @@ -3192,9 +3427,10 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled, const int mi_height = num_8x8_blocks_high_lookup[bsize]; x->skip_recode = !x->select_txfm_size && mbmi->sb_type >= BLOCK_8X8 && - (cpi->oxcf.aq_mode != COMPLEXITY_AQ && - cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ) && - !cpi->sf.use_nonrd_pick_mode; + cpi->oxcf.aq_mode != COMPLEXITY_AQ && + cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ && + cpi->sf.allow_skip_recode; + x->skip_optimize = ctx->is_coded; ctx->is_coded = 1; x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct; |