diff options
Diffstat (limited to 'vp9/encoder')
-rw-r--r-- | vp9/encoder/vp9_bitstream.c | 475 | ||||
-rw-r--r-- | vp9/encoder/vp9_block.h | 9 | ||||
-rw-r--r-- | vp9/encoder/vp9_boolhuff.c | 1 | ||||
-rw-r--r-- | vp9/encoder/vp9_dct.c | 30 | ||||
-rw-r--r-- | vp9/encoder/vp9_encodeframe.c | 424 | ||||
-rw-r--r-- | vp9/encoder/vp9_encodeintra.c | 66 | ||||
-rw-r--r-- | vp9/encoder/vp9_encodeintra.h | 2 | ||||
-rw-r--r-- | vp9/encoder/vp9_encodemb.c | 53 | ||||
-rw-r--r-- | vp9/encoder/vp9_firstpass.c | 2 | ||||
-rw-r--r-- | vp9/encoder/vp9_modecosts.c | 24 | ||||
-rw-r--r-- | vp9/encoder/vp9_onyx_if.c | 174 | ||||
-rw-r--r-- | vp9/encoder/vp9_onyx_int.h | 39 | ||||
-rw-r--r-- | vp9/encoder/vp9_ratectrl.c | 8 | ||||
-rw-r--r-- | vp9/encoder/vp9_rdopt.c | 922 | ||||
-rw-r--r-- | vp9/encoder/vp9_tokenize.c | 15 | ||||
-rw-r--r-- | vp9/encoder/vp9_variance_c.c | 5 | ||||
-rw-r--r-- | vp9/encoder/vp9_write_bit_buffer.h | 42 |
17 files changed, 1303 insertions, 988 deletions
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c index 26f22b4be..a95d7eb46 100644 --- a/vp9/encoder/vp9_bitstream.c +++ b/vp9/encoder/vp9_bitstream.c @@ -12,35 +12,37 @@ #include <stdio.h> #include <limits.h> -#include "vp9/common/vp9_header.h" -#include "vp9/encoder/vp9_encodemv.h" +#include "vpx/vpx_encoder.h" +#include "vpx_mem/vpx_mem.h" + #include "vp9/common/vp9_entropymode.h" #include "vp9/common/vp9_entropymv.h" #include "vp9/common/vp9_findnearmv.h" #include "vp9/common/vp9_tile_common.h" -#include "vp9/encoder/vp9_mcomp.h" -#include "vp9/common/vp9_systemdependent.h" -#include "vp9/common/vp9_pragmas.h" -#include "vpx/vpx_encoder.h" -#include "vpx_mem/vpx_mem.h" -#include "vp9/encoder/vp9_bitstream.h" -#include "vp9/encoder/vp9_segmentation.h" - #include "vp9/common/vp9_seg_common.h" #include "vp9/common/vp9_pred_common.h" #include "vp9/common/vp9_entropy.h" #include "vp9/common/vp9_entropymv.h" #include "vp9/common/vp9_mvref_common.h" #include "vp9/common/vp9_treecoder.h" +#include "vp9/common/vp9_systemdependent.h" +#include "vp9/common/vp9_pragmas.h" + +#include "vp9/encoder/vp9_mcomp.h" +#include "vp9/encoder/vp9_encodemv.h" +#include "vp9/encoder/vp9_bitstream.h" +#include "vp9/encoder/vp9_segmentation.h" +#include "vp9/encoder/vp9_write_bit_buffer.h" + #if defined(SECTIONBITS_OUTPUT) unsigned __int64 Sectionbits[500]; #endif #ifdef ENTROPY_STATS -int intra_mode_stats[VP9_KF_BINTRAMODES] - [VP9_KF_BINTRAMODES] - [VP9_KF_BINTRAMODES]; +int intra_mode_stats[VP9_INTRA_MODES] + [VP9_INTRA_MODES] + [VP9_INTRA_MODES]; vp9_coeff_stats tree_update_hist_4x4[BLOCK_TYPES]; vp9_coeff_stats tree_update_hist_8x8[BLOCK_TYPES]; vp9_coeff_stats tree_update_hist_16x16[BLOCK_TYPES]; @@ -243,15 +245,12 @@ static void update_mbintra_mode_probs(VP9_COMP* const cpi, vp9_writer* const bc) { VP9_COMMON *const cm = &cpi->common; - vp9_prob pnew[VP9_YMODES - 1]; - unsigned int bct[VP9_YMODES - 1][2]; - - update_mode(bc, VP9_YMODES, vp9_ymode_encodings, vp9_ymode_tree, pnew, - cm->fc.ymode_prob, bct, (unsigned int *)cpi->ymode_count); + vp9_prob pnew[VP9_INTRA_MODES - 1]; + unsigned int bct[VP9_INTRA_MODES - 1][2]; - update_mode(bc, VP9_I32X32_MODES, vp9_sb_ymode_encodings, - vp9_sb_ymode_tree, pnew, cm->fc.sb_ymode_prob, bct, - (unsigned int *)cpi->sb_ymode_count); + update_mode(bc, VP9_INTRA_MODES, vp9_intra_mode_encodings, + vp9_intra_mode_tree, pnew, + cm->fc.y_mode_prob, bct, (unsigned int *)cpi->y_mode_count); } void vp9_update_skip_probs(VP9_COMP *cpi) { @@ -322,15 +321,15 @@ static void update_refpred_stats(VP9_COMP *cpi) { // The branch counts table is re-populated during the actual pack stage and in // the decoder to facilitate backwards update of the context. static void update_inter_mode_probs(VP9_COMMON *cm, - int mode_context[INTER_MODE_CONTEXTS][4]) { + int mode_context[INTER_MODE_CONTEXTS][VP9_MVREFS - 1]) { int i, j; - unsigned int (*mv_ref_ct)[4][2] = cm->fc.mv_ref_ct; + unsigned int (*mv_ref_ct)[VP9_MVREFS - 1][2] = cm->fc.mv_ref_ct; vpx_memcpy(mode_context, cm->fc.vp9_mode_contexts, sizeof(cm->fc.vp9_mode_contexts)); for (i = 0; i < INTER_MODE_CONTEXTS; i++) { - for (j = 0; j < 4; j++) { + for (j = 0; j < VP9_MVREFS - 1; j++) { int new_prob, old_cost, new_cost; // Work out cost of coding branches with the old and optimal probability @@ -348,28 +347,8 @@ static void update_inter_mode_probs(VP9_COMMON *cm, } } -static void write_ymode(vp9_writer *bc, int m, const vp9_prob *p) { - write_token(bc, vp9_ymode_tree, p, vp9_ymode_encodings + m); -} - -static void kfwrite_ymode(vp9_writer *bc, int m, const vp9_prob *p) { - write_token(bc, vp9_kf_ymode_tree, p, vp9_kf_ymode_encodings + m); -} - -static void write_sb_ymode(vp9_writer *bc, int m, const vp9_prob *p) { - write_token(bc, vp9_sb_ymode_tree, p, vp9_sb_ymode_encodings + m); -} - -static void sb_kfwrite_ymode(vp9_writer *bc, int m, const vp9_prob *p) { - write_token(bc, vp9_uv_mode_tree, p, vp9_sb_kf_ymode_encodings + m); -} - -static void write_uv_mode(vp9_writer *bc, int m, const vp9_prob *p) { - write_token(bc, vp9_uv_mode_tree, p, vp9_uv_mode_encodings + m); -} - -static void write_kf_bmode(vp9_writer *bc, int m, const vp9_prob *p) { - write_token(bc, vp9_bmode_tree, p, vp9_kf_bmode_encodings + m); +static void write_intra_mode(vp9_writer *bc, int m, const vp9_prob *p) { + write_token(bc, vp9_intra_mode_tree, p, vp9_intra_mode_encodings + m); } static int prob_update_savings(const unsigned int *ct, @@ -471,7 +450,6 @@ static void pack_mb_tokens(vp9_writer* const bc, const vp9_prob *pp; int v = a->value; int n = a->len; - int ncount = n; vp9_prob probs[ENTROPY_NODES]; if (t == EOSB_TOKEN) { @@ -487,18 +465,25 @@ static void pack_mb_tokens(vp9_writer* const bc, assert(pp != 0); /* skip one or two nodes */ +#if !CONFIG_BALANCED_COEFTREE if (p->skip_eob_node) { n -= p->skip_eob_node; i = 2 * p->skip_eob_node; - ncount -= p->skip_eob_node; } +#endif do { const int bb = (v >> --n) & 1; +#if CONFIG_BALANCED_COEFTREE + if (i == 2 && p->skip_eob_node) { + i += 2; + assert(bb == 1); + continue; + } +#endif vp9_write(bc, bb, pp[i >> 1]); i = vp9_coef_tree[i + bb]; - ncount--; - } while (n && ncount); + } while (n); if (b->base_val) { const int e = p->extra, l = b->len; @@ -524,19 +509,10 @@ static void pack_mb_tokens(vp9_writer* const bc, *tp = p; } -static void write_mv_ref(vp9_writer *bc, MB_PREDICTION_MODE m, - const vp9_prob *p) { -#if CONFIG_DEBUG - assert(NEARESTMV <= m && m <= SPLITMV); -#endif - write_token(bc, vp9_mv_ref_tree, p, - vp9_mv_ref_encoding_array - NEARESTMV + m); -} - static void write_sb_mv_ref(vp9_writer *bc, MB_PREDICTION_MODE m, const vp9_prob *p) { #if CONFIG_DEBUG - assert(NEARESTMV <= m && m < SPLITMV); + assert(NEARESTMV <= m && m <= NEWMV); #endif write_token(bc, vp9_sb_mv_ref_tree, p, vp9_sb_mv_ref_encoding_array - NEARESTMV + m); @@ -699,25 +675,37 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, MODE_INFO *m, // Encode the reference frame. encode_ref_frame(bc, pc, xd, segment_id, rf); + if (mi->sb_type >= BLOCK_SIZE_SB8X8 && pc->txfm_mode == TX_MODE_SELECT && + !(rf != INTRA_FRAME && + (skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)))) { + TX_SIZE sz = mi->txfm_size; + // FIXME(rbultje) code ternary symbol once all experiments are merged + vp9_write(bc, sz != TX_4X4, pc->prob_tx[0]); + if (mi->sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) { + vp9_write(bc, sz != TX_8X8, pc->prob_tx[1]); + if (mi->sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8) + vp9_write(bc, sz != TX_16X16, pc->prob_tx[2]); + } + } + if (rf == INTRA_FRAME) { #ifdef ENTROPY_STATS active_section = 6; #endif - if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) - write_sb_ymode(bc, mode, pc->fc.sb_ymode_prob); - - if (m->mbmi.sb_type < BLOCK_SIZE_SB8X8) { + if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) { + write_intra_mode(bc, mode, pc->fc.y_mode_prob); + } else { int idx, idy; int bw = 1 << b_width_log2(mi->sb_type); int bh = 1 << b_height_log2(mi->sb_type); for (idy = 0; idy < 2; idy += bh) for (idx = 0; idx < 2; idx += bw) - write_sb_ymode(bc, m->bmi[idy * 2 + idx].as_mode.first, - pc->fc.sb_ymode_prob); + write_intra_mode(bc, m->bmi[idy * 2 + idx].as_mode.first, + pc->fc.y_mode_prob); } - write_uv_mode(bc, mi->uv_mode, - pc->fc.uv_mode_prob[mode]); + write_intra_mode(bc, mi->uv_mode, + pc->fc.uv_mode_prob[mode]); } else { vp9_prob mv_ref_p[VP9_MVREFS - 1]; @@ -729,21 +717,20 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, MODE_INFO *m, // If segment skip is not enabled code the mode. if (!vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)) { - if (mi->sb_type >= BLOCK_SIZE_SB8X8) + if (mi->sb_type >= BLOCK_SIZE_SB8X8) { write_sb_mv_ref(bc, mode, mv_ref_p); - vp9_accum_mv_refs(&cpi->common, mode, mi->mb_mode_context[rf]); + vp9_accum_mv_refs(&cpi->common, mode, mi->mb_mode_context[rf]); + } } - if (is_inter_mode(mode)) { - if (cpi->common.mcomp_filter_type == SWITCHABLE) { - write_token(bc, vp9_switchable_interp_tree, - vp9_get_pred_probs(&cpi->common, xd, - PRED_SWITCHABLE_INTERP), - vp9_switchable_interp_encodings + - vp9_switchable_interp_map[mi->interp_filter]); - } else { - assert(mi->interp_filter == cpi->common.mcomp_filter_type); - } + if (cpi->common.mcomp_filter_type == SWITCHABLE) { + write_token(bc, vp9_switchable_interp_tree, + vp9_get_pred_probs(&cpi->common, xd, + PRED_SWITCHABLE_INTERP), + vp9_switchable_interp_encodings + + vp9_switchable_interp_map[mi->interp_filter]); + } else { + assert(mi->interp_filter == cpi->common.mcomp_filter_type); } // does the feature use compound prediction or not @@ -753,70 +740,51 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, MODE_INFO *m, vp9_get_pred_prob(pc, xd, PRED_COMP)); } - switch (mode) { /* new, split require MVs */ - case NEWMV: + if (xd->mode_info_context->mbmi.sb_type < BLOCK_SIZE_SB8X8) { + int j; + MB_PREDICTION_MODE blockmode; + int_mv blockmv; + int bwl = b_width_log2(mi->sb_type), bw = 1 << bwl; + int bhl = b_height_log2(mi->sb_type), bh = 1 << bhl; + int idx, idy; + for (idy = 0; idy < 2; idy += bh) { + for (idx = 0; idx < 2; idx += bw) { + j = idy * 2 + idx; + blockmode = cpi->mb.partition_info->bmi[j].mode; + blockmv = cpi->mb.partition_info->bmi[j].mv; + write_sb_mv_ref(bc, blockmode, mv_ref_p); + vp9_accum_mv_refs(&cpi->common, blockmode, mi->mb_mode_context[rf]); + if (blockmode == NEWMV) { #ifdef ENTROPY_STATS - active_section = 5; + active_section = 11; #endif - vp9_encode_mv(bc, - &mi->mv[0].as_mv, &mi->best_mv.as_mv, - nmvc, xd->allow_high_precision_mv); + vp9_encode_mv(bc, &blockmv.as_mv, &mi->best_mv.as_mv, + nmvc, xd->allow_high_precision_mv); - if (mi->second_ref_frame > 0) - vp9_encode_mv(bc, - &mi->mv[1].as_mv, &mi->best_second_mv.as_mv, - nmvc, xd->allow_high_precision_mv); - break; - case SPLITMV: { - int j; - MB_PREDICTION_MODE blockmode; - int_mv blockmv; - int bwl = b_width_log2(mi->sb_type), bw = 1 << bwl; - int bhl = b_height_log2(mi->sb_type), bh = 1 << bhl; - int idx, idy; - for (idy = 0; idy < 2; idy += bh) { - for (idx = 0; idx < 2; idx += bw) { - j = idy * 2 + idx; - blockmode = cpi->mb.partition_info->bmi[j].mode; - blockmv = cpi->mb.partition_info->bmi[j].mv; - write_sb_mv_ref(bc, blockmode, mv_ref_p); - vp9_accum_mv_refs(&cpi->common, blockmode, mi->mb_mode_context[rf]); - if (blockmode == NEWMV) { -#ifdef ENTROPY_STATS - active_section = 11; -#endif - vp9_encode_mv(bc, &blockmv.as_mv, &mi->best_mv.as_mv, + if (mi->second_ref_frame > 0) + vp9_encode_mv(bc, + &cpi->mb.partition_info->bmi[j].second_mv.as_mv, + &mi->best_second_mv.as_mv, nmvc, xd->allow_high_precision_mv); - - if (mi->second_ref_frame > 0) - vp9_encode_mv(bc, - &cpi->mb.partition_info->bmi[j].second_mv.as_mv, - &mi->best_second_mv.as_mv, - nmvc, xd->allow_high_precision_mv); - } } } + } #ifdef MODE_STATS - ++count_mb_seg[mi->partitioning]; + ++count_mb_seg[mi->partitioning]; #endif - break; - } - default: - break; - } - } + } else if (mode == NEWMV) { +#ifdef ENTROPY_STATS + active_section = 5; +#endif + vp9_encode_mv(bc, + &mi->mv[0].as_mv, &mi->best_mv.as_mv, + nmvc, xd->allow_high_precision_mv); - if (mi->sb_type >= BLOCK_SIZE_SB8X8 && pc->txfm_mode == TX_MODE_SELECT && - !(rf != INTRA_FRAME && - (skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)))) { - TX_SIZE sz = mi->txfm_size; - // FIXME(rbultje) code ternary symbol once all experiments are merged - vp9_write(bc, sz != TX_4X4, pc->prob_tx[0]); - if (mi->sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) { - vp9_write(bc, sz != TX_8X8, pc->prob_tx[1]); - if (mi->sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8) - vp9_write(bc, sz != TX_16X16, pc->prob_tx[2]); + if (mi->second_ref_frame > 0) + vp9_encode_mv(bc, + &mi->mv[1].as_mv, &mi->best_second_mv.as_mv, + nmvc, xd->allow_high_precision_mv); } } } @@ -841,14 +809,23 @@ static void write_mb_modes_kf(const VP9_COMP *cpi, vp9_write(bc, skip_coeff, vp9_get_pred_prob(c, xd, PRED_MBSKIP)); } + if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8 && c->txfm_mode == TX_MODE_SELECT) { + TX_SIZE sz = m->mbmi.txfm_size; + // FIXME(rbultje) code ternary symbol once all experiments are merged + vp9_write(bc, sz != TX_4X4, c->prob_tx[0]); + if (m->mbmi.sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) { + vp9_write(bc, sz != TX_8X8, c->prob_tx[1]); + if (m->mbmi.sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8) + vp9_write(bc, sz != TX_16X16, c->prob_tx[2]); + } + } + if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) { const MB_PREDICTION_MODE A = above_block_mode(m, 0, mis); const MB_PREDICTION_MODE L = xd->left_available ? left_block_mode(m, 0) : DC_PRED; - write_kf_bmode(bc, ym, c->kf_bmode_prob[A][L]); - } - - if (m->mbmi.sb_type < BLOCK_SIZE_SB8X8) { + write_intra_mode(bc, ym, c->kf_y_mode_prob[A][L]); + } else { int idx, idy; int bw = 1 << b_width_log2(m->mbmi.sb_type); int bh = 1 << b_height_log2(m->mbmi.sb_type); @@ -858,24 +835,16 @@ static void write_mb_modes_kf(const VP9_COMP *cpi, const MB_PREDICTION_MODE A = above_block_mode(m, i, mis); const MB_PREDICTION_MODE L = (xd->left_available || idx) ? left_block_mode(m, i) : DC_PRED; - write_kf_bmode(bc, m->bmi[i].as_mode.first, - c->kf_bmode_prob[A][L]); + const int bm = m->bmi[i].as_mode.first; +#ifdef ENTROPY_STATS + ++intra_mode_stats[A][L][bm]; +#endif + write_intra_mode(bc, bm, c->kf_y_mode_prob[A][L]); } } } - write_uv_mode(bc, m->mbmi.uv_mode, c->kf_uv_mode_prob[ym]); - - if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8 && c->txfm_mode == TX_MODE_SELECT) { - TX_SIZE sz = m->mbmi.txfm_size; - // FIXME(rbultje) code ternary symbol once all experiments are merged - vp9_write(bc, sz != TX_4X4, c->prob_tx[0]); - if (m->mbmi.sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) { - vp9_write(bc, sz != TX_8X8, c->prob_tx[1]); - if (m->mbmi.sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8) - vp9_write(bc, sz != TX_16X16, c->prob_tx[2]); - } - } + write_intra_mode(bc, m->mbmi.uv_mode, c->kf_uv_mode_prob[ym]); } static void write_modes_b(VP9_COMP *cpi, MODE_INFO *m, vp9_writer *bc, @@ -1045,7 +1014,7 @@ static void print_prob_tree(vp9_coeff_probs *coef_probs, int block_types) { fclose(f); } -static void build_tree_distribution(vp9_coeff_probs *coef_probs, +static void build_tree_distribution(vp9_coeff_probs_model *coef_probs, vp9_coeff_count *coef_counts, unsigned int (*eob_branch_ct)[REF_TYPES] [COEF_BANDS] @@ -1060,6 +1029,7 @@ static void build_tree_distribution(vp9_coeff_probs *coef_probs, #ifdef ENTROPY_STATS int t = 0; #endif + vp9_prob full_probs[ENTROPY_NODES]; for (i = 0; i < block_types; ++i) { for (j = 0; j < REF_TYPES; ++j) { @@ -1068,14 +1038,24 @@ static void build_tree_distribution(vp9_coeff_probs *coef_probs, if (l >= 3 && k == 0) continue; vp9_tree_probs_from_distribution(vp9_coef_tree, - coef_probs[i][j][k][l], + full_probs, coef_branch_ct[i][j][k][l], coef_counts[i][j][k][l], 0); + vpx_memcpy(coef_probs[i][j][k][l], full_probs, + sizeof(vp9_prob) * UNCONSTRAINED_NODES); +#if CONFIG_BALANCED_COEFTREE + coef_branch_ct[i][j][k][l][1][1] = eob_branch_ct[i][j][k][l] - + coef_branch_ct[i][j][k][l][1][0]; + coef_probs[i][j][k][l][1] = + get_binary_prob(coef_branch_ct[i][j][k][l][1][0], + coef_branch_ct[i][j][k][l][1][1]); +#else coef_branch_ct[i][j][k][l][0][1] = eob_branch_ct[i][j][k][l] - coef_branch_ct[i][j][k][l][0][0]; coef_probs[i][j][k][l][0] = get_binary_prob(coef_branch_ct[i][j][k][l][0][0], coef_branch_ct[i][j][k][l][0][1]); +#endif #ifdef ENTROPY_STATS if (!cpi->dummy_packing) { for (t = 0; t < MAX_ENTROPY_TOKENS; ++t) @@ -1127,7 +1107,7 @@ static void update_coef_probs_common( #ifdef ENTROPY_STATS vp9_coeff_stats *tree_update_hist, #endif - vp9_coeff_probs *new_frame_coef_probs, + vp9_coeff_probs_model *new_frame_coef_probs, vp9_coeff_probs_model *old_frame_coef_probs, vp9_coeff_stats *frame_branch_ct, TX_SIZE tx_size) { @@ -1136,7 +1116,6 @@ static void update_coef_probs_common( int savings; const int entropy_nodes_update = UNCONSTRAINED_NODES; - // vp9_prob bestupd = find_coef_update_prob(cpi); const int tstart = 0; /* dry run to see if there is any udpate at all needed */ @@ -1150,7 +1129,7 @@ static void update_coef_probs_common( vp9_prob newp = new_frame_coef_probs[i][j][k][l][t]; const vp9_prob oldp = old_frame_coef_probs[i][j][k][l][t]; const vp9_prob upd = vp9_coef_update_prob[t]; - int s; // = prev_coef_savings[t]; + int s; int u = 0; if (l >= 3 && k == 0) @@ -1192,11 +1171,10 @@ static void update_coef_probs_common( vp9_prob newp = new_frame_coef_probs[i][j][k][l][t]; vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t; const vp9_prob upd = vp9_coef_update_prob[t]; - int s; // = prev_coef_savings[t]; + int s; int u = 0; if (l >= 3 && k == 0) continue; - if (t == PIVOT_NODE) s = prob_diff_update_savings_search_model( frame_branch_ct[i][j][k][l][0], @@ -1278,37 +1256,6 @@ static void update_coef_probs(VP9_COMP* const cpi, vp9_writer* const bc) { } } -#ifdef PACKET_TESTING -FILE *vpxlogc = 0; -#endif - -static void decide_kf_ymode_entropy(VP9_COMP *cpi) { - int mode_cost[MB_MODE_COUNT]; - int bestcost = INT_MAX; - int bestindex = 0; - int i, j; - - for (i = 0; i < 8; i++) { - int cost = 0; - - vp9_cost_tokens(mode_cost, cpi->common.kf_ymode_prob[i], vp9_kf_ymode_tree); - - for (j = 0; j < VP9_YMODES; j++) - cost += mode_cost[j] * cpi->ymode_count[j]; - - vp9_cost_tokens(mode_cost, cpi->common.sb_kf_ymode_prob[i], - vp9_sb_ymode_tree); - for (j = 0; j < VP9_I32X32_MODES; j++) - cost += mode_cost[j] * cpi->sb_ymode_count[j]; - - if (cost < bestcost) { - bestindex = i; - bestcost = cost; - } - } - cpi->common.kf_ymode_probs_index = bestindex; - -} static void segment_reference_frames(VP9_COMP *cpi) { VP9_COMMON *oci = &cpi->common; MODE_INFO *mi = oci->mi; @@ -1417,9 +1364,6 @@ static void encode_segmentation(VP9_COMP *cpi, vp9_writer *w) { // Segmentation map vp9_write_bit(w, xd->update_mb_segmentation_map); -#if CONFIG_IMPLICIT_SEGMENTATION - vp9_write_bit(w, xd->allow_implicit_segment_update); -#endif if (xd->update_mb_segmentation_map) { // Select the coding strategy (temporal or spatial) vp9_choose_segmap_coding_method(cpi); @@ -1482,60 +1426,67 @@ static void encode_segmentation(VP9_COMP *cpi, vp9_writer *w) { } } -void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { - int i; - VP9_HEADER oh; - VP9_COMMON *const pc = &cpi->common; - vp9_writer header_bc, residual_bc; - MACROBLOCKD *const xd = &cpi->mb.e_mbd; - int extra_bytes_packed = 0; - - uint8_t *cx_data = dest; +void write_uncompressed_header(VP9_COMMON *cm, + struct vp9_write_bit_buffer *wb) { + const int scaling_active = cm->width != cm->display_width || + cm->height != cm->display_height; - oh.show_frame = (int) pc->show_frame; - oh.type = (int)pc->frame_type; - oh.version = pc->version; - oh.first_partition_length_in_bytes = 0; + vp9_wb_write_bit(wb, cm->frame_type); + vp9_wb_write_literal(wb, cm->version, 3); + vp9_wb_write_bit(wb, cm->show_frame); + vp9_wb_write_bit(wb, scaling_active); + vp9_wb_write_bit(wb, cm->subsampling_x); + vp9_wb_write_bit(wb, cm->subsampling_y); - cx_data += 3; + if (cm->frame_type == KEY_FRAME) { + vp9_wb_write_literal(wb, SYNC_CODE_0, 8); + vp9_wb_write_literal(wb, SYNC_CODE_1, 8); + vp9_wb_write_literal(wb, SYNC_CODE_2, 8); + } -#if defined(SECTIONBITS_OUTPUT) - Sectionbits[active_section = 1] += sizeof(VP9_HEADER) * 8 * 256; -#endif + if (scaling_active) { + vp9_wb_write_literal(wb, cm->display_width, 16); + vp9_wb_write_literal(wb, cm->display_height, 16); + } - compute_update_table(); + vp9_wb_write_literal(wb, cm->width, 16); + vp9_wb_write_literal(wb, cm->height, 16); - /* every keyframe send startcode, width, height, scale factor, clamp - * and color type. - */ - if (oh.type == KEY_FRAME) { - // Start / synch code - cx_data[0] = 0x49; - cx_data[1] = 0x83; - cx_data[2] = 0x42; - extra_bytes_packed = 3; - cx_data += extra_bytes_packed; + if (!cm->show_frame) { + vp9_wb_write_bit(wb, cm->intra_only); } - if (pc->width != pc->display_width || pc->height != pc->display_height) { - write_le16(cx_data, pc->display_width); - write_le16(cx_data + 2, pc->display_height); - cx_data += 4; - extra_bytes_packed += 4; + vp9_wb_write_literal(wb, cm->frame_context_idx, NUM_FRAME_CONTEXTS_LG2); + vp9_wb_write_bit(wb, cm->clr_type); + + vp9_wb_write_bit(wb, cm->error_resilient_mode); + if (!cm->error_resilient_mode) { + vp9_wb_write_bit(wb, cm->reset_frame_context); + vp9_wb_write_bit(wb, cm->refresh_frame_context); + vp9_wb_write_bit(wb, cm->frame_parallel_decoding_mode); } +} - write_le16(cx_data, pc->width); - write_le16(cx_data + 2, pc->height); - extra_bytes_packed += 4; - cx_data += 4; +void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { + int i, bytes_packed; + VP9_COMMON *const pc = &cpi->common; + vp9_writer header_bc, residual_bc; + MACROBLOCKD *const xd = &cpi->mb.e_mbd; - vp9_start_encode(&header_bc, cx_data); + uint8_t *cx_data = dest; + struct vp9_write_bit_buffer wb = {dest, 0}; + struct vp9_write_bit_buffer first_partition_size_wb; + + write_uncompressed_header(pc, &wb); + first_partition_size_wb = wb; + vp9_wb_write_literal(&wb, 0, 16); // don't know in advance first part. size - // TODO(jkoleszar): remove these two unused bits? - vp9_write_bit(&header_bc, pc->clr_type); + bytes_packed = vp9_rb_bytes_written(&wb); + cx_data += bytes_packed; - // error resilient mode - vp9_write_bit(&header_bc, pc->error_resilient_mode); + compute_update_table(); + + vp9_start_encode(&header_bc, cx_data); encode_loopfilter(pc, xd, &header_bc); @@ -1617,14 +1568,6 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { vp9_write_literal(&header_bc, (pc->mcomp_filter_type), 2); } - if (!pc->error_resilient_mode) { - vp9_write_bit(&header_bc, pc->refresh_frame_context); - vp9_write_bit(&header_bc, pc->frame_parallel_decoding_mode); - } - - vp9_write_literal(&header_bc, pc->frame_context_idx, - NUM_FRAME_CONTEXTS_LG2); - #ifdef ENTROPY_STATS if (pc->frame_type == INTER_FRAME) active_section = 0; @@ -1694,7 +1637,7 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { // changes in the bitstream. if (pc->frame_type != KEY_FRAME) { int i, j; - int new_context[INTER_MODE_CONTEXTS][4]; + int new_context[INTER_MODE_CONTEXTS][VP9_MVREFS - 1]; if (!cpi->dummy_packing) { update_inter_mode_probs(pc, new_context); } else { @@ -1704,7 +1647,7 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { } for (i = 0; i < INTER_MODE_CONTEXTS; i++) { - for (j = 0; j < 4; j++) { + for (j = 0; j < VP9_MVREFS - 1; j++) { if (new_context[i][j] != pc->fc.vp9_mode_contexts[i][j]) { vp9_write(&header_bc, 1, 252); vp9_write_prob(&header_bc, new_context[i][j]); @@ -1731,10 +1674,8 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { vp9_copy(cpi->common.fc.pre_coef_probs_32x32, cpi->common.fc.coef_probs_32x32); - vp9_copy(cpi->common.fc.pre_sb_ymode_prob, cpi->common.fc.sb_ymode_prob); - vp9_copy(cpi->common.fc.pre_ymode_prob, cpi->common.fc.ymode_prob); + vp9_copy(cpi->common.fc.pre_y_mode_prob, cpi->common.fc.y_mode_prob); vp9_copy(cpi->common.fc.pre_uv_mode_prob, cpi->common.fc.uv_mode_prob); - vp9_copy(cpi->common.fc.pre_bmode_prob, cpi->common.fc.bmode_prob); vp9_copy(cpi->common.fc.pre_partition_prob, cpi->common.fc.partition_prob); cpi->common.fc.pre_nmvc = cpi->common.fc.nmvc; vp9_zero(cpi->common.fc.mv_ref_ct); @@ -1750,11 +1691,7 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { vp9_write_prob(&header_bc, pc->mbskip_pred_probs[i]); } - if (pc->frame_type == KEY_FRAME) { - if (!pc->kf_ymode_probs_update) { - vp9_write_literal(&header_bc, pc->kf_ymode_probs_index, 3); - } - } else { + if (pc->frame_type != KEY_FRAME) { // Update the probabilities used to encode reference frame data update_ref_probs(cpi); @@ -1820,35 +1757,11 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) { vp9_stop_encode(&header_bc); - oh.first_partition_length_in_bytes = header_bc.pos; - - /* update frame tag */ - { - int scaling = (pc->width != pc->display_width || - pc->height != pc->display_height); - int v = (oh.first_partition_length_in_bytes << 8) | - (pc->subsampling_y << 7) | - (pc->subsampling_x << 6) | - (scaling << 5) | - (oh.show_frame << 4) | - (oh.version << 1) | - oh.type; - - assert(oh.first_partition_length_in_bytes <= 0xffff); - dest[0] = v; - dest[1] = v >> 8; - dest[2] = v >> 16; - } - - *size = VP9_HEADER_SIZE + extra_bytes_packed + header_bc.pos; - if (pc->frame_type == KEY_FRAME) { - decide_kf_ymode_entropy(cpi); - } else { - /* This is not required if the counts in cpi are consistent with the - * final packing pass */ - // if (!cpi->dummy_packing) vp9_zero(cpi->NMVcount); - } + // first partition size + assert(header_bc.pos <= 0xffff); + vp9_wb_write_literal(&first_partition_size_wb, header_bc.pos, 16); + *size = bytes_packed + header_bc.pos; { int tile_row, tile_col, total_size = 0; diff --git a/vp9/encoder/vp9_block.h b/vp9/encoder/vp9_block.h index 211eca4b4..e6d36cdf8 100644 --- a/vp9/encoder/vp9_block.h +++ b/vp9/encoder/vp9_block.h @@ -114,10 +114,9 @@ struct macroblock { int *nmvsadcost_hp[2]; int **mvsadcost; - int mbmode_cost[2][MB_MODE_COUNT]; + int mbmode_cost[MB_MODE_COUNT]; int intra_uv_mode_cost[2][MB_MODE_COUNT]; - int bmode_costs[VP9_BINTRAMODES][VP9_BINTRAMODES][VP9_BINTRAMODES]; - int inter_bmode_costs[INTRA_MODE_COUNT]; + int y_mode_costs[VP9_INTRA_MODES][VP9_INTRA_MODES][VP9_INTRA_MODES]; int switchable_interp_costs[VP9_SWITCHABLE_FILTERS + 1] [VP9_SWITCHABLE_FILTERS]; @@ -134,7 +133,11 @@ struct macroblock { unsigned char *active_ptr; + // note that token_costs is the cost when eob node is skipped vp9_coeff_count token_costs[TX_SIZE_MAX_SB][BLOCK_TYPES]; +#if CONFIG_BALANCED_COEFTREE + vp9_coeff_count token_costs_noskip[TX_SIZE_MAX_SB][BLOCK_TYPES]; +#endif int optimize; diff --git a/vp9/encoder/vp9_boolhuff.c b/vp9/encoder/vp9_boolhuff.c index 0fcb2579f..86143ca57 100644 --- a/vp9/encoder/vp9_boolhuff.c +++ b/vp9/encoder/vp9_boolhuff.c @@ -10,6 +10,7 @@ #include <assert.h> #include "vp9/encoder/vp9_boolhuff.h" +#include "vp9/common/vp9_entropy.h" #if defined(SECTIONBITS_OUTPUT) unsigned __int64 Sectionbits[500]; diff --git a/vp9/encoder/vp9_dct.c b/vp9/encoder/vp9_dct.c index d22644424..8d4eec139 100644 --- a/vp9/encoder/vp9_dct.c +++ b/vp9/encoder/vp9_dct.c @@ -606,14 +606,13 @@ void vp9_short_walsh4x4_c(short *input, short *output, int pitch) { c1 = ip[2 * pitch_short]; d1 = ip[3 * pitch_short]; - b1 = a1 - b1; - c1 += d1; - e1 = (c1 - b1) >> 1; - a1 += e1; - d1 -= e1; - c1 = a1 - c1; - b1 -= d1; - + a1 += b1; + d1 = d1 - c1; + e1 = (a1 - d1) >> 1; + b1 = e1 - b1; + c1 = e1 - c1; + a1 -= c1; + d1 += b1; op[0] = a1; op[4] = c1; op[8] = d1; @@ -631,14 +630,13 @@ void vp9_short_walsh4x4_c(short *input, short *output, int pitch) { c1 = ip[2]; d1 = ip[3]; - b1 = a1 - b1; - c1 += d1; - e1 = (c1 - b1) >> 1; - a1 += e1; - d1 -= e1; - c1 = a1 - c1; - b1 -= d1; - + a1 += b1; + d1 -= c1; + e1 = (a1 - d1) >> 1; + b1 = e1 - b1; + c1 = e1 - c1; + a1 -= c1; + d1 += b1; op[0] = a1 << WHT_UPSCALE_FACTOR; op[1] = c1 << WHT_UPSCALE_FACTOR; op[2] = d1 << WHT_UPSCALE_FACTOR; diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c index ebee191ad..a38c1ffd3 100644 --- a/vp9/encoder/vp9_encodeframe.c +++ b/vp9/encoder/vp9_encodeframe.c @@ -10,6 +10,7 @@ #include "./vpx_config.h" +#include "./vp9_rtcd.h" #include "vp9/encoder/vp9_encodeframe.h" #include "vp9/encoder/vp9_encodemb.h" #include "vp9/encoder/vp9_encodemv.h" @@ -97,6 +98,8 @@ static unsigned int alt_activity_measure(VP9_COMP *cpi, return vp9_encode_intra(cpi, x, use_dc_pred); } +DECLARE_ALIGNED(16, static const uint8_t, vp9_64x64_zeros[64*64]) = { 0 }; + // Measure the activity of the current macroblock // What we measure here is TBD so abstracted to this function @@ -332,7 +335,9 @@ static void update_state(VP9_COMP *cpi, MACROBLOCKD *const xd = &x->e_mbd; MODE_INFO *mi = &ctx->mic; MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi; - int mb_mode = mi->mbmi.mode; +#if CONFIG_DEBUG || CONFIG_INTERNAL_STATS + MB_PREDICTION_MODE mb_mode = mi->mbmi.mode; +#endif int mb_mode_index = ctx->best_mode_index; const int mis = cpi->common.mode_info_stride; const int bh = 1 << mi_height_log2(bsize), bw = 1 << mi_width_log2(bsize); @@ -362,7 +367,8 @@ static void update_state(VP9_COMP *cpi, ctx->txfm_rd_diff[ALLOW_32X32] = ctx->txfm_rd_diff[ALLOW_16X16]; } - if (mb_mode == SPLITMV) { + if (mbmi->ref_frame != INTRA_FRAME && + mbmi->sb_type < BLOCK_SIZE_SB8X8) { vpx_memcpy(x->partition_info, &ctx->partition_info, sizeof(PARTITION_INFO)); @@ -448,7 +454,8 @@ static void update_state(VP9_COMP *cpi, */ // Note how often each mode chosen as best cpi->mode_chosen_counts[mb_mode_index]++; - if (mbmi->mode == SPLITMV || mbmi->mode == NEWMV) { + if (mbmi->ref_frame != INTRA_FRAME && + (mbmi->sb_type < BLOCK_SIZE_SB8X8 || mbmi->mode == NEWMV)) { int_mv best_mv, best_second_mv; MV_REFERENCE_FRAME rf = mbmi->ref_frame; best_mv.as_int = ctx->best_ref_mv.as_int; @@ -768,6 +775,35 @@ static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col, vpx_memcpy(cm->left_seg_context + (mi_row & MI_MASK), sl, sizeof(PARTITION_CONTEXT) * mh); } +static void save_context(VP9_COMP *cpi, int mi_row, int mi_col, + ENTROPY_CONTEXT a[16 * MAX_MB_PLANE], + ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], + PARTITION_CONTEXT sa[8], + PARTITION_CONTEXT sl[8], + BLOCK_SIZE_TYPE bsize) { + VP9_COMMON *const cm = &cpi->common; + MACROBLOCK *const x = &cpi->mb; + MACROBLOCKD *const xd = &x->e_mbd; + int p; + int bwl = b_width_log2(bsize), bw = 1 << bwl; + int bhl = b_height_log2(bsize), bh = 1 << bhl; + int mwl = mi_width_log2(bsize), mw = 1 << mwl; + int mhl = mi_height_log2(bsize), mh = 1 << mhl; + + // buffer the above/left context information of the block in search. + for (p = 0; p < MAX_MB_PLANE; ++p) { + vpx_memcpy(a + bw * p, cm->above_context[p] + + (mi_col * 2 >> xd->plane[p].subsampling_x), + sizeof(ENTROPY_CONTEXT) * bw >> xd->plane[p].subsampling_x); + vpx_memcpy(l + bh * p, cm->left_context[p] + + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y), + sizeof(ENTROPY_CONTEXT) * bh >> xd->plane[p].subsampling_y); + } + vpx_memcpy(sa, cm->above_seg_context + mi_col, + sizeof(PARTITION_CONTEXT) * mw); + vpx_memcpy(sl, cm->left_seg_context + (mi_row & MI_MASK), + sizeof(PARTITION_CONTEXT) * mh); +} static void encode_b(VP9_COMP *cpi, TOKENEXTRA **tp, int mi_row, int mi_col, int output_enabled, @@ -860,6 +896,337 @@ static void encode_sb(VP9_COMP *cpi, TOKENEXTRA **tp, } } +static void set_partitioning(VP9_COMP *cpi, MODE_INFO *m, + BLOCK_SIZE_TYPE bsize) { + VP9_COMMON *const cm = &cpi->common; + const int mis = cm->mode_info_stride; + int bsl = b_width_log2(bsize); + int bs = (1 << bsl) / 2; // + int block_row, block_col; + int row, col; + + // this test function sets the entire macroblock to the same bsize + for (block_row = 0; block_row < 8; block_row += bs) { + for (block_col = 0; block_col < 8; block_col += bs) { + for (row = 0; row < bs; row++) { + for (col = 0; col < bs; col++) { + m[(block_row+row)*mis + block_col+col].mbmi.sb_type = bsize; + } + } + } + } +} + +static void set_block_size(VP9_COMMON *const cm, + MODE_INFO *m, BLOCK_SIZE_TYPE bsize, int mis, + int mi_row, int mi_col) { + int row, col; + int bsl = b_width_log2(bsize); + int bs = (1 << bsl) / 2; // + MODE_INFO *m2 = m + mi_row * mis + mi_col; + for (row = 0; row < bs; row++) { + for (col = 0; col < bs; col++) { + if (mi_row + row >= cm->mi_rows || mi_col + col >= cm->mi_cols) + return; + m2[row*mis+col].mbmi.sb_type = bsize; + } + } +} +typedef struct { + int64_t sum_square_error; + int64_t sum_error; + int count; + int variance; +} var; + +#define VT(TYPE, BLOCKSIZE) \ + typedef struct { \ + var none; \ + var horz[2]; \ + var vert[2]; \ + BLOCKSIZE split[4]; } TYPE; + +VT(v8x8, var) +VT(v16x16, v8x8) +VT(v32x32, v16x16) +VT(v64x64, v32x32) + +typedef enum { + V16X16, + V32X32, + V64X64, +} TREE_LEVEL; + +// Set variance values given sum square error, sum error, count. +static void fill_variance(var *v, int64_t s2, int64_t s, int c) { + v->sum_square_error = s2; + v->sum_error = s; + v->count = c; + v->variance = 256 + * (v->sum_square_error - v->sum_error * v->sum_error / v->count) + / v->count; +} + +// Fills a 16x16 variance tree node by calling get var8x8 var.. +static void fill_16x16_variance(const unsigned char *s, int sp, + const unsigned char *d, int dp, v16x16 *vt) { + unsigned int sse; + int sum; + vp9_get_sse_sum_8x8(s, sp, d, dp, &sse, &sum); + fill_variance(&vt->split[0].none, sse, sum, 64); + vp9_get_sse_sum_8x8(s + 8, sp, d + 8, dp, &sse, &sum); + fill_variance(&vt->split[1].none, sse, sum, 64); + vp9_get_sse_sum_8x8(s + 8 * sp, sp, d + 8 * dp, dp, &sse, &sum); + fill_variance(&vt->split[2].none, sse, sum, 64); + vp9_get_sse_sum_8x8(s + 8 * sp + 8, sp, d + 8 + 8 * dp, dp, &sse, &sum); + fill_variance(&vt->split[3].none, sse, sum, 64); +} + +// Combine 2 variance structures by summing the sum_error, sum_square_error, +// and counts and then calculating the new variance. +void sum_2_variances(var *r, var *a, var*b) { + fill_variance(r, a->sum_square_error + b->sum_square_error, + a->sum_error + b->sum_error, a->count + b->count); +} +// Fill one level of our variance tree, by summing the split sums into each of +// the horizontal, vertical and none from split and recalculating variance. +#define fill_variance_tree(VT) \ + sum_2_variances(VT.horz[0], VT.split[0].none, VT.split[1].none); \ + sum_2_variances(VT.horz[1], VT.split[2].none, VT.split[3].none); \ + sum_2_variances(VT.vert[0], VT.split[0].none, VT.split[2].none); \ + sum_2_variances(VT.vert[1], VT.split[1].none, VT.split[3].none); \ + sum_2_variances(VT.none, VT.vert[0], VT.vert[1]); + +// Set the blocksize in the macroblock info structure if the variance is less +// than our threshold to one of none, horz, vert. +#define set_vt_size(VT, BLOCKSIZE, R, C, ACTION) \ + if (VT.none.variance < threshold) { \ + set_block_size(cm, m, BLOCKSIZE, mis, R, C); \ + ACTION; \ + } \ + if (VT.horz[0].variance < threshold && VT.horz[1].variance < threshold ) { \ + set_block_size(cm, m, get_subsize(BLOCKSIZE, PARTITION_HORZ), mis, R, C); \ + ACTION; \ + } \ + if (VT.vert[0].variance < threshold && VT.vert[1].variance < threshold ) { \ + set_block_size(cm, m, get_subsize(BLOCKSIZE, PARTITION_VERT), mis, R, C); \ + ACTION; \ + } + +static void choose_partitioning(VP9_COMP *cpi, MODE_INFO *m, int mi_row, + int mi_col) { + VP9_COMMON * const cm = &cpi->common; + MACROBLOCK *x = &cpi->mb; + MACROBLOCKD *xd = &cpi->mb.e_mbd; + const int mis = cm->mode_info_stride; + // TODO(JBB): More experimentation or testing of this threshold; + int64_t threshold = 4; + int i, j, k; + v64x64 vt; + unsigned char * s; + int sp; + const unsigned char * d = xd->plane[0].pre->buf; + int dp = xd->plane[0].pre->stride; + + set_offsets(cpi, mi_row, mi_col, BLOCK_SIZE_SB64X64); + s = x->plane[0].src.buf; + sp = x->plane[0].src.stride; + + // TODO(JBB): Clearly the higher the quantizer the fewer partitions we want + // but this needs more experimentation. + threshold = threshold * cpi->common.base_qindex * cpi->common.base_qindex; + + // if ( cm->frame_type == KEY_FRAME ) { + d = vp9_64x64_zeros; + dp = 64; + // } + // Fill in the entire tree of 8x8 variances for splits. + for (i = 0; i < 4; i++) { + const int x32_idx = ((i & 1) << 5); + const int y32_idx = ((i >> 1) << 5); + for (j = 0; j < 4; j++) { + const int x_idx = x32_idx + ((j & 1) << 4); + const int y_idx = y32_idx + ((j >> 1) << 4); + fill_16x16_variance(s + y_idx * sp + x_idx, sp, d + y_idx * dp + x_idx, + dp, &vt.split[i].split[j]); + } + } + // Fill the rest of the variance tree by summing the split partition + // values. + for (i = 0; i < 4; i++) { + for (j = 0; j < 4; j++) { + fill_variance_tree(&vt.split[i].split[j]) + } + fill_variance_tree(&vt.split[i]) + } + fill_variance_tree(&vt) + + // Now go through the entire structure, splitting every blocksize until + // we get to one that's got a variance lower than our threshold, or we + // hit 8x8. + set_vt_size( vt, BLOCK_SIZE_SB64X64, mi_row, mi_col, return); + for (i = 0; i < 4; ++i) { + const int x32_idx = ((i & 1) << 2); + const int y32_idx = ((i >> 1) << 2); + set_vt_size(vt, BLOCK_SIZE_SB32X32, mi_row + y32_idx, mi_col + x32_idx, + continue); + + for (j = 0; j < 4; ++j) { + const int x16_idx = ((j & 1) << 1); + const int y16_idx = ((j >> 1) << 1); + set_vt_size(vt, BLOCK_SIZE_MB16X16, mi_row + y32_idx + y16_idx, + mi_col+x32_idx+x16_idx, continue); + + for (k = 0; k < 4; ++k) { + const int x8_idx = (k & 1); + const int y8_idx = (k >> 1); + set_block_size(cm, m, BLOCK_SIZE_SB8X8, mis, + mi_row + y32_idx + y16_idx + y8_idx, + mi_col + x32_idx + x16_idx + x8_idx); + } + } + } +} +static void rd_use_partition(VP9_COMP *cpi, MODE_INFO *m, TOKENEXTRA **tp, + int mi_row, int mi_col, BLOCK_SIZE_TYPE bsize, + int *rate, int *dist) { + VP9_COMMON * const cm = &cpi->common; + MACROBLOCK * const x = &cpi->mb; + MACROBLOCKD *xd = &cpi->mb.e_mbd; + const int mis = cm->mode_info_stride; + int bwl, bhl; + int bsl = b_width_log2(bsize); + int bs = (1 << bsl); + int bss = (1 << bsl)/4; + int i, pl; + PARTITION_TYPE partition; + BLOCK_SIZE_TYPE subsize; + ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE]; + PARTITION_CONTEXT sl[8], sa[8]; + int r = 0, d = 0; + + if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) + return; + + + bwl = b_width_log2(m->mbmi.sb_type); + bhl = b_height_log2(m->mbmi.sb_type); + + // parse the partition type + if ((bwl == bsl) && (bhl == bsl)) + partition = PARTITION_NONE; + else if ((bwl == bsl) && (bhl < bsl)) + partition = PARTITION_HORZ; + else if ((bwl < bsl) && (bhl == bsl)) + partition = PARTITION_VERT; + else if ((bwl < bsl) && (bhl < bsl)) + partition = PARTITION_SPLIT; + else + assert(0); + + subsize = get_subsize(bsize, partition); + + // TODO(JBB): this restriction is here because pick_sb_modes can return + // r's that are INT_MAX meaning we can't select a mode / mv for this block. + // when the code is made to work for less than sb8x8 we need to come up with + // a solution to this problem. + assert(subsize >= BLOCK_SIZE_SB8X8); + + if (bsize >= BLOCK_SIZE_SB8X8) { + xd->left_seg_context = cm->left_seg_context + (mi_row & MI_MASK); + xd->above_seg_context = cm->above_seg_context + mi_col; + *(get_sb_partitioning(x, bsize)) = subsize; + } + + pl = partition_plane_context(xd, bsize); + save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize); + switch (partition) { + case PARTITION_NONE: + pick_sb_modes(cpi, mi_row, mi_col, tp, &r, &d, bsize, + get_block_context(x, bsize)); + r += x->partition_cost[pl][PARTITION_NONE]; + break; + case PARTITION_HORZ: + *(get_sb_index(xd, subsize)) = 0; + pick_sb_modes(cpi, mi_row, mi_col, tp, &r, &d, subsize, + get_block_context(x, subsize)); + if (mi_row + (bs >> 1) <= cm->mi_rows) { + int rt, dt; + update_state(cpi, get_block_context(x, subsize), subsize, 0); + encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize); + *(get_sb_index(xd, subsize)) = 1; + pick_sb_modes(cpi, mi_row + (bs >> 2), mi_col, tp, &rt, &dt, subsize, + get_block_context(x, subsize)); + r += rt; + d += dt; + } + set_partition_seg_context(cm, xd, mi_row, mi_col); + pl = partition_plane_context(xd, bsize); + r += x->partition_cost[pl][PARTITION_HORZ]; + break; + case PARTITION_VERT: + *(get_sb_index(xd, subsize)) = 0; + pick_sb_modes(cpi, mi_row, mi_col, tp, &r, &d, subsize, + get_block_context(x, subsize)); + if (mi_col + (bs >> 1) <= cm->mi_cols) { + int rt, dt; + update_state(cpi, get_block_context(x, subsize), subsize, 0); + encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize); + *(get_sb_index(xd, subsize)) = 1; + pick_sb_modes(cpi, mi_row, mi_col + (bs >> 2), tp, &rt, &dt, subsize, + get_block_context(x, subsize)); + r += rt; + d += dt; + } + set_partition_seg_context(cm, xd, mi_row, mi_col); + pl = partition_plane_context(xd, bsize); + r += x->partition_cost[pl][PARTITION_VERT]; + restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize); + break; + case PARTITION_SPLIT: + for (i = 0; i < 4; i++) { + int x_idx = (i & 1) * (bs >> 2); + int y_idx = (i >> 1) * (bs >> 2); + int jj = i >> 1, ii = i & 0x01; + int rt, dt; + + if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols)) + continue; + + *(get_sb_index(xd, subsize)) = i; + + rd_use_partition(cpi, m + jj * bss * mis + ii * bss, tp, mi_row + y_idx, + mi_col + x_idx, subsize, &rt, &dt); + r += rt; + d += dt; + } + set_partition_seg_context(cm, xd, mi_row, mi_col); + pl = partition_plane_context(xd, bsize); + r += x->partition_cost[pl][PARTITION_SPLIT]; + break; + default: + assert(0); + } + + // update partition context +#if CONFIG_AB4X4 + if (bsize >= BLOCK_SIZE_SB8X8 && + (bsize == BLOCK_SIZE_SB8X8 || partition != PARTITION_SPLIT)) { +#else + if (bsize > BLOCK_SIZE_SB8X8 + && (bsize == BLOCK_SIZE_MB16X16 || partition != PARTITION_SPLIT)) { +#endif + set_partition_seg_context(cm, xd, mi_row, mi_col); + update_partition_context(xd, subsize, bsize); + } + restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize); + + if (r < INT_MAX && d < INT_MAX) + encode_sb(cpi, tp, mi_row, mi_col, bsize == BLOCK_SIZE_SB64X64, bsize); + *rate = r; + *dist = d; +} + // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are // unlikely to be selected depending on previously rate-distortion optimization @@ -876,7 +1243,7 @@ static void rd_pick_partition(VP9_COMP *cpi, TOKENEXTRA **tp, ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE]; PARTITION_CONTEXT sl[8], sa[8]; TOKENEXTRA *tp_orig = *tp; - int i, p, pl; + int i, pl; BLOCK_SIZE_TYPE subsize; int srate = INT_MAX, sdist = INT_MAX; @@ -888,19 +1255,7 @@ static void rd_pick_partition(VP9_COMP *cpi, TOKENEXTRA **tp, } assert(mi_height_log2(bsize) == mi_width_log2(bsize)); - // buffer the above/left context information of the block in search. - for (p = 0; p < MAX_MB_PLANE; ++p) { - vpx_memcpy(a + bs * p, cm->above_context[p] + - (mi_col * 2 >> xd->plane[p].subsampling_x), - sizeof(ENTROPY_CONTEXT) * bs >> xd->plane[p].subsampling_x); - vpx_memcpy(l + bs * p, cm->left_context[p] + - ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y), - sizeof(ENTROPY_CONTEXT) * bs >> xd->plane[p].subsampling_y); - } - vpx_memcpy(sa, cm->above_seg_context + mi_col, - sizeof(PARTITION_CONTEXT) * ms); - vpx_memcpy(sl, cm->left_seg_context + (mi_row & MI_MASK), - sizeof(PARTITION_CONTEXT) * ms); + save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize); // PARTITION_SPLIT if (bsize >= BLOCK_SIZE_SB8X8) { @@ -1028,6 +1383,8 @@ static void rd_pick_partition(VP9_COMP *cpi, TOKENEXTRA **tp, *rate = srate; *dist = sdist; + restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize); + if (srate < INT_MAX && sdist < INT_MAX) encode_sb(cpi, tp, mi_row, mi_col, bsize == BLOCK_SIZE_SB64X64, bsize); @@ -1053,8 +1410,22 @@ static void encode_sb_row(VP9_COMP *cpi, int mi_row, for (mi_col = cm->cur_tile_mi_col_start; mi_col < cm->cur_tile_mi_col_end; mi_col += 8) { int dummy_rate, dummy_dist; - rd_pick_partition(cpi, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64, - &dummy_rate, &dummy_dist); + // TODO(JBB): remove the border conditions for 64x64 blocks once its fixed + // without this border check choose will fail on the border of every + // non 64x64. + if (cpi->speed < 5 || + mi_col + 8 > cm->cur_tile_mi_col_end || + mi_row + 8 > cm->cur_tile_mi_row_end) { + rd_pick_partition(cpi, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64, + &dummy_rate, &dummy_dist); + } else { + const int idx_str = cm->mode_info_stride * mi_row + mi_col; + MODE_INFO *m = cm->mi + idx_str; + // set_partitioning(cpi, m, BLOCK_SIZE_SB8X8); + choose_partitioning(cpi, cm->mi, mi_row, mi_col); + rd_use_partition(cpi, m, tp, mi_row, mi_col, BLOCK_SIZE_SB64X64, + &dummy_rate, &dummy_dist); + } } } @@ -1093,11 +1464,9 @@ static void init_encode_frame_mb_context(VP9_COMP *cpi) { xd->mode_info_context->mbmi.uv_mode = DC_PRED; vp9_zero(cpi->count_mb_ref_frame_usage) - vp9_zero(cpi->bmode_count) - vp9_zero(cpi->ymode_count) + vp9_zero(cpi->y_mode_count) vp9_zero(cpi->y_uv_mode_count) vp9_zero(cpi->common.fc.mv_ref_ct) - vp9_zero(cpi->sb_ymode_count) vp9_zero(cpi->partition_count); // Note: this memset assumes above_context[0], [1] and [2] @@ -1550,20 +1919,17 @@ static void sum_intra_stats(VP9_COMP *cpi, MACROBLOCK *x) { const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode; const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode; + ++cpi->y_uv_mode_count[m][uvm]; if (xd->mode_info_context->mbmi.sb_type >= BLOCK_SIZE_SB8X8) { - ++cpi->sb_ymode_count[m]; + ++cpi->y_mode_count[m]; } else { - ++cpi->ymode_count[m]; - } - ++cpi->y_uv_mode_count[m][uvm]; - if (m == I4X4_PRED) { int idx, idy; int bw = 1 << b_width_log2(xd->mode_info_context->mbmi.sb_type); int bh = 1 << b_height_log2(xd->mode_info_context->mbmi.sb_type); for (idy = 0; idy < 2; idy += bh) { for (idx = 0; idx < 2; idx += bw) { int m = xd->mode_info_context->bmi[idy * 2 + idx].as_mode.first; - ++cpi->sb_ymode_count[m]; + ++cpi->y_mode_count[m]; } } } @@ -1627,7 +1993,7 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, cpi->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST; else cpi->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST; - } else if (mbmi->mode == SPLITMV) { + } else if (mbmi->sb_type < BLOCK_SIZE_SB8X8) { cpi->zbin_mode_boost = SPLIT_MV_ZBIN_BOOST; } else { cpi->zbin_mode_boost = MV_ZBIN_BOOST; diff --git a/vp9/encoder/vp9_encodeintra.c b/vp9/encoder/vp9_encodeintra.c index 5d7c244f1..57041a90b 100644 --- a/vp9/encoder/vp9_encodeintra.c +++ b/vp9/encoder/vp9_encodeintra.c @@ -16,72 +16,18 @@ #include "vp9/common/vp9_invtrans.h" #include "vp9/encoder/vp9_encodeintra.h" -static void encode_intra4x4block(MACROBLOCK *x, int ib, BLOCK_SIZE_TYPE bs); - int vp9_encode_intra(VP9_COMP *cpi, MACROBLOCK *x, int use_16x16_pred) { MB_MODE_INFO * mbmi = &x->e_mbd.mode_info_context->mbmi; (void) cpi; - + mbmi->mode = DC_PRED; + mbmi->ref_frame = INTRA_FRAME; if (use_16x16_pred) { - mbmi->mode = DC_PRED; - mbmi->uv_mode = DC_PRED; - mbmi->ref_frame = INTRA_FRAME; - - vp9_encode_intra16x16mby(&cpi->common, x); + mbmi->txfm_size = TX_16X16; + vp9_encode_intra_block_y(&cpi->common, x, BLOCK_SIZE_MB16X16); } else { - int i; - - for (i = 0; i < 16; i++) { - encode_intra4x4block(x, i, BLOCK_SIZE_MB16X16); - } + mbmi->txfm_size = TX_4X4; + vp9_encode_intra_block_y(&cpi->common, x, BLOCK_SIZE_MB16X16); } return vp9_get_mb_ss(x->plane[0].src_diff); } - -// This function is used only by the firstpass encoding. -static void encode_intra4x4block(MACROBLOCK *x, int ib, - BLOCK_SIZE_TYPE bsize) { - MACROBLOCKD * const xd = &x->e_mbd; - uint8_t* const src = - raster_block_offset_uint8(xd, bsize, 0, ib, - x->plane[0].src.buf, x->plane[0].src.stride); - uint8_t* const dst = - raster_block_offset_uint8(xd, bsize, 0, ib, - xd->plane[0].dst.buf, xd->plane[0].dst.stride); - int16_t* const src_diff = - raster_block_offset_int16(xd, bsize, 0, ib, - x->plane[0].src_diff); - int16_t* const coeff = BLOCK_OFFSET(x->plane[0].coeff, ib, 16); - const int bwl = b_width_log2(bsize), bhl = b_height_log2(bsize); - - assert(ib < (1 << (bwl + bhl))); - - vp9_intra4x4_predict(&x->e_mbd, ib, bsize, DC_PRED, - dst, xd->plane[0].dst.stride); - vp9_subtract_block(4, 4, src_diff, 4 << bwl, - src, x->plane[0].src.stride, - dst, xd->plane[0].dst.stride); - - x->fwd_txm4x4(src_diff, coeff, 8 << bwl); - x->quantize_b_4x4(x, ib, DCT_DCT, 16); - vp9_inverse_transform_b_4x4_add(&x->e_mbd, xd->plane[0].eobs[ib], - BLOCK_OFFSET(xd->plane[0].dqcoeff, ib, 16), - dst, xd->plane[0].dst.stride); -} - -void vp9_encode_intra16x16mby(VP9_COMMON *const cm, MACROBLOCK *x) { - MACROBLOCKD *xd = &x->e_mbd; - - vp9_build_intra_predictors_sby_s(xd, BLOCK_SIZE_MB16X16); - vp9_encode_sby(cm, x, BLOCK_SIZE_MB16X16); -} - -void vp9_encode_intra16x16mbuv(VP9_COMMON *const cm, MACROBLOCK *x) { - MACROBLOCKD *xd = &x->e_mbd; - - vp9_build_intra_predictors_sbuv_s(xd, BLOCK_SIZE_MB16X16); - vp9_encode_sbuv(cm, x, BLOCK_SIZE_MB16X16); -} - - diff --git a/vp9/encoder/vp9_encodeintra.h b/vp9/encoder/vp9_encodeintra.h index 7da164c6a..14d144b74 100644 --- a/vp9/encoder/vp9_encodeintra.h +++ b/vp9/encoder/vp9_encodeintra.h @@ -14,8 +14,6 @@ #include "vp9/encoder/vp9_onyx_int.h" int vp9_encode_intra(VP9_COMP *cpi, MACROBLOCK *x, int use_16x16_pred); -void vp9_encode_intra16x16mby(VP9_COMMON *const cm, MACROBLOCK *x); -void vp9_encode_intra16x16mbuv(VP9_COMMON *const cm, MACROBLOCK *x); void vp9_encode_intra_block_y(VP9_COMMON *const cm, MACROBLOCK *mb, BLOCK_SIZE_TYPE bs); void vp9_encode_intra_block_uv(VP9_COMMON *const cm, MACROBLOCK *mb, diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c index b7f60b127..98ea98031 100644 --- a/vp9/encoder/vp9_encodemb.c +++ b/vp9/encoder/vp9_encodemb.c @@ -20,6 +20,9 @@ #include "vp9/common/vp9_systemdependent.h" #include "vp9_rtcd.h" +DECLARE_ALIGNED(16, extern const uint8_t, + vp9_pt_energy_class[MAX_ENTROPY_TOKENS]); + void vp9_subtract_block(int rows, int cols, int16_t *diff_ptr, int diff_stride, const uint8_t *src_ptr, int src_stride, @@ -105,7 +108,7 @@ static int trellis_get_coeff_context(const int *scan, uint8_t *token_cache, int pad, int l) { int bak = token_cache[scan[idx]], pt; - token_cache[scan[idx]] = token; + token_cache[scan[idx]] = vp9_pt_energy_class[token]; pt = vp9_get_coef_context(scan, nb, pad, token_cache, idx + 1, l); token_cache[scan[idx]] = bak; return pt; @@ -189,7 +192,8 @@ static void optimize_b(VP9_COMMON *const cm, MACROBLOCK *mb, *(tokens[eob] + 1) = *(tokens[eob] + 0); next = eob; for (i = 0; i < eob; i++) - token_cache[scan[i]] = vp9_dct_value_tokens_ptr[qcoeff_ptr[scan[i]]].token; + token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[ + qcoeff_ptr[scan[i]]].token]; nb = vp9_get_coef_neighbors_handle(scan, &pad); for (i = eob; i-- > i0;) { @@ -211,10 +215,21 @@ static void optimize_b(VP9_COMMON *const cm, MACROBLOCK *mb, band = get_coef_band(band_translate, i + 1); pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache, pad, default_eob); +#if CONFIG_BALANCED_COEFTREE + rate0 += + mb->token_costs_noskip[tx_size][type][ref][band][pt] + [tokens[next][0].token]; + rate1 += + mb->token_costs_noskip[tx_size][type][ref][band][pt] + [tokens[next][1].token]; +#else rate0 += - mb->token_costs[tx_size][type][ref][band][pt][tokens[next][0].token]; + mb->token_costs[tx_size][type][ref][band][pt] + [tokens[next][0].token]; rate1 += - mb->token_costs[tx_size][type][ref][band][pt][tokens[next][1].token]; + mb->token_costs[tx_size][type][ref][band][pt] + [tokens[next][1].token]; +#endif } UPDATE_RD_COST(); /* And pick the best. */ @@ -262,14 +277,32 @@ static void optimize_b(VP9_COMMON *const cm, MACROBLOCK *mb, if (t0 != DCT_EOB_TOKEN) { pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache, pad, default_eob); +#if CONFIG_BALANCED_COEFTREE + if (!x) + rate0 += mb->token_costs[tx_size][type][ref][band][pt][ + tokens[next][0].token]; + else + rate0 += mb->token_costs_noskip[tx_size][type][ref][band][pt][ + tokens[next][0].token]; +#else rate0 += mb->token_costs[tx_size][type][ref][band][pt][ tokens[next][0].token]; +#endif } if (t1 != DCT_EOB_TOKEN) { pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache, pad, default_eob); +#if CONFIG_BALANCED_COEFTREE + if (!x) + rate1 += mb->token_costs[tx_size][type][ref][band][pt][ + tokens[next][1].token]; + else + rate1 += mb->token_costs_noskip[tx_size][type][ref][band][pt][ + tokens[next][1].token]; +#else rate1 += mb->token_costs[tx_size][type][ref][band][pt][ tokens[next][1].token]; +#endif } } @@ -322,8 +355,13 @@ static void optimize_b(VP9_COMMON *const cm, MACROBLOCK *mb, error1 = tokens[next][1].error; t0 = tokens[next][0].token; t1 = tokens[next][1].token; +#if CONFIG_BALANCED_COEFTREE + rate0 += mb->token_costs_noskip[tx_size][type][ref][band][pt][t0]; + rate1 += mb->token_costs_noskip[tx_size][type][ref][band][pt][t1]; +#else rate0 += mb->token_costs[tx_size][type][ref][band][pt][t0]; rate1 += mb->token_costs[tx_size][type][ref][band][pt][t1]; +#endif UPDATE_RD_COST(); best = rd_cost1 < rd_cost0; final_eob = i0 - 1; @@ -610,6 +648,7 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE_TYPE bsize, struct encode_b_args* const args = arg; MACROBLOCK* const x = args->x; MACROBLOCKD* const xd = &x->e_mbd; + MB_MODE_INFO* const mbmi = &xd->mode_info_context->mbmi; const TX_SIZE tx_size = (TX_SIZE)(ss_txfrm_size / 2); const int bw = 4 << (b_width_log2(bsize) - xd->plane[plane].subsampling_x); const int raster_block = txfrm_block_to_raster_block(xd, bsize, plane, @@ -634,9 +673,9 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE_TYPE bsize, TX_TYPE tx_type; int mode, b_mode; - mode = plane == 0? xd->mode_info_context->mbmi.mode: - xd->mode_info_context->mbmi.uv_mode; - if (bsize <= BLOCK_SIZE_SB8X8 && mode == I4X4_PRED && plane == 0) + mode = plane == 0? mbmi->mode: mbmi->uv_mode; + if (mbmi->sb_type < BLOCK_SIZE_SB8X8 && plane == 0 && + mbmi->ref_frame == INTRA_FRAME) b_mode = xd->mode_info_context->bmi[ib].as_mode.first; else b_mode = mode; diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c index 5ec696604..f57c8be6c 100644 --- a/vp9/encoder/vp9_firstpass.c +++ b/vp9/encoder/vp9_firstpass.c @@ -523,6 +523,7 @@ void vp9_first_pass(VP9_COMP *cpi) { xd->left_available = (mb_col != 0); xd->mode_info_context->mbmi.sb_type = BLOCK_SIZE_MB16X16; + xd->mode_info_context->mbmi.ref_frame = INTRA_FRAME; // do intra 16x16 prediction this_error = vp9_encode_intra(cpi, x, use_dc_pred); @@ -619,6 +620,7 @@ void vp9_first_pass(VP9_COMP *cpi) { this_error = motion_error; vp9_set_mbmode_and_mvs(x, NEWMV, &mv); xd->mode_info_context->mbmi.txfm_size = TX_4X4; + xd->mode_info_context->mbmi.ref_frame = LAST_FRAME; vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, BLOCK_SIZE_MB16X16); diff --git a/vp9/encoder/vp9_modecosts.c b/vp9/encoder/vp9_modecosts.c index 171b44bf9..099a04404 100644 --- a/vp9/encoder/vp9_modecosts.c +++ b/vp9/encoder/vp9_modecosts.c @@ -17,29 +17,23 @@ void vp9_init_mode_costs(VP9_COMP *c) { VP9_COMMON *x = &c->common; - const vp9_tree_p T = vp9_bmode_tree; - const vp9_tree_p KT = vp9_bmode_tree; + const vp9_tree_p KT = vp9_intra_mode_tree; int i, j; - for (i = 0; i < VP9_BINTRAMODES; i++) { - for (j = 0; j < VP9_BINTRAMODES; j++) { - vp9_cost_tokens((int *)c->mb.bmode_costs[i][j], - x->kf_bmode_prob[i][j], KT); + for (i = 0; i < VP9_INTRA_MODES; i++) { + for (j = 0; j < VP9_INTRA_MODES; j++) { + vp9_cost_tokens((int *)c->mb.y_mode_costs[i][j], + x->kf_y_mode_prob[i][j], KT); } } - vp9_cost_tokens((int *)c->mb.inter_bmode_costs, x->fc.bmode_prob, T); - // TODO(rbultje) separate tables for superblock costing? - vp9_cost_tokens(c->mb.mbmode_cost[1], x->fc.sb_ymode_prob, - vp9_sb_ymode_tree); - vp9_cost_tokens(c->mb.mbmode_cost[0], - x->sb_kf_ymode_prob[c->common.kf_ymode_probs_index], - vp9_sb_ymode_tree); + vp9_cost_tokens(c->mb.mbmode_cost, x->fc.y_mode_prob, + vp9_intra_mode_tree); vp9_cost_tokens(c->mb.intra_uv_mode_cost[1], - x->fc.uv_mode_prob[VP9_YMODES - 1], vp9_uv_mode_tree); + x->fc.uv_mode_prob[VP9_INTRA_MODES - 1], vp9_intra_mode_tree); vp9_cost_tokens(c->mb.intra_uv_mode_cost[0], - x->kf_uv_mode_prob[VP9_YMODES - 1], vp9_uv_mode_tree); + x->kf_uv_mode_prob[VP9_INTRA_MODES - 1], vp9_intra_mode_tree); for (i = 0; i <= VP9_SWITCHABLE_FILTERS; ++i) vp9_cost_tokens((int *)c->mb.switchable_interp_costs[i], diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c index 4bbb4152b..9c0609ed1 100644 --- a/vp9/encoder/vp9_onyx_if.c +++ b/vp9/encoder/vp9_onyx_if.c @@ -103,9 +103,9 @@ extern int skip_false_count; #ifdef ENTROPY_STATS -extern int intra_mode_stats[VP9_KF_BINTRAMODES] - [VP9_KF_BINTRAMODES] - [VP9_KF_BINTRAMODES]; +extern int intra_mode_stats[VP9_INTRA_MODES] + [VP9_INTRA_MODES] + [VP9_INTRA_MODES]; #endif #ifdef NMV_STATS @@ -258,9 +258,6 @@ void vp9_initialize_enc() { init_done = 1; } } -#ifdef PACKET_TESTING -extern FILE *vpxlogc; -#endif static void setup_features(VP9_COMP *cpi) { MACROBLOCKD *xd = &cpi->mb.e_mbd; @@ -270,9 +267,6 @@ static void setup_features(VP9_COMP *cpi) { xd->update_mb_segmentation_map = 0; xd->update_mb_segmentation_data = 0; -#if CONFIG_IMPLICIT_SEGMENTATION - xd->allow_implicit_segment_update = 0; -#endif vpx_memset(xd->mb_segment_tree_probs, 255, sizeof(xd->mb_segment_tree_probs)); vp9_clearall_segfeatures(xd); @@ -287,7 +281,6 @@ static void setup_features(VP9_COMP *cpi) { set_default_lf_deltas(cpi); } - static void dealloc_compressor_data(VP9_COMP *cpi) { // Delete sementation map vpx_free(cpi->segmentation_map); @@ -358,9 +351,6 @@ static void configure_static_seg_features(VP9_COMP *cpi) { vpx_memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols); xd->update_mb_segmentation_map = 0; xd->update_mb_segmentation_data = 0; -#if CONFIG_IMPLICIT_SEGMENTATION - xd->allow_implicit_segment_update = 0; -#endif cpi->static_mb_pct = 0; // Disable segmentation @@ -374,9 +364,6 @@ static void configure_static_seg_features(VP9_COMP *cpi) { vpx_memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols); xd->update_mb_segmentation_map = 0; xd->update_mb_segmentation_data = 0; -#if CONFIG_IMPLICIT_SEGMENTATION - xd->allow_implicit_segment_update = 0; -#endif cpi->static_mb_pct = 0; // Disable segmentation and individual segment features by default @@ -475,53 +462,6 @@ static void configure_static_seg_features(VP9_COMP *cpi) { } } -#if CONFIG_IMPLICIT_SEGMENTATION -static double implict_seg_q_modifiers[MAX_MB_SEGMENTS] = - {1.0, 0.95, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0}; -static void configure_implicit_segmentation(VP9_COMP *cpi, int frame_qindex) { - VP9_COMMON *cm = &cpi->common; - MACROBLOCKD *xd = &cpi->mb.e_mbd; - int i; - int qi_delta; - double q_baseline = vp9_convert_qindex_to_q(frame_qindex); - - // Set the flags to allow implicit segment update but disallow explicit update - xd->segmentation_enabled = 1; - xd->allow_implicit_segment_update = 1; - xd->update_mb_segmentation_map = 0; - - // For key frames clear down the segment map to a default state. - if (cm->frame_type == KEY_FRAME) { - // Clear down the global segmentation map - vpx_memset(cpi->segmentation_map, 0, (cm->mb_rows * cm->mb_cols)); - - // Clear down the segment features. - vp9_clearall_segfeatures(xd); - - xd->update_mb_segmentation_data = 0; - - // Update the segment data if it is an arf or non overlay gf. - } else if (cpi->refresh_alt_ref_frame || - (cpi->refresh_golden_frame && !cpi->is_src_frame_alt_ref)) { - xd->update_mb_segmentation_data = 1; - - // Enable use of q deltas on segments 1 and up - // Segment 0 is treated as a neutral segment with no changes - for (i = 1; i < MAX_MB_SEGMENTS; ++i) { - qi_delta = compute_qdelta(cpi, q_baseline, - implict_seg_q_modifiers[i] * q_baseline); - vp9_set_segdata(xd, i, SEG_LVL_ALT_Q, qi_delta); - vp9_enable_segfeature(xd, i, SEG_LVL_ALT_Q); - } - - // Where relevant assume segment data is delta data - xd->mb_segment_abs_delta = SEGMENT_DELTADATA; - } else { - xd->update_mb_segmentation_data = 0; - } -} -#endif - #ifdef ENTROPY_STATS void vp9_update_mode_context_stats(VP9_COMP *cpi) { VP9_COMMON *cm = &cpi->common; @@ -541,7 +481,7 @@ void vp9_update_mode_context_stats(VP9_COMP *cpi) { // Add in the values for this frame for (i = 0; i < INTER_MODE_CONTEXTS; i++) { - for (j = 0; j < 4; j++) { + for (j = 0; j < VP9_MVREFS - 1; j++) { mv_ref_stats[i][j][0] += (int64_t)mv_ref_ct[i][j][0]; mv_ref_stats[i][j][1] += (int64_t)mv_ref_ct[i][j][1]; } @@ -558,12 +498,13 @@ void print_mode_context(VP9_COMP *cpi) { int i, j; fprintf(f, "#include \"vp9_entropy.h\"\n"); - fprintf(f, "const int vp9_mode_contexts[INTER_MODE_CONTEXTS][4] ="); + fprintf(f, + "const int vp9_mode_contexts[INTER_MODE_CONTEXTS][VP9_MVREFS - 1] ="); fprintf(f, "{\n"); for (j = 0; j < INTER_MODE_CONTEXTS; j++) { fprintf(f, " {/* %d */ ", j); fprintf(f, " "); - for (i = 0; i < 4; i++) { + for (i = 0; i < VP9_MVREFS - 1; i++) { int this_prob; int64_t count = cpi->mv_ref_stats[j][i][0] + cpi->mv_ref_stats[j][i][1]; if (count) @@ -700,6 +641,25 @@ static void set_rd_speed_thresholds(VP9_COMP *cpi, int mode, int speed) { sf->thresh_mult[THR_COMP_SPLITGA ] += speed_multiplier * 4500; sf->thresh_mult[THR_COMP_SPLITLG ] += speed_multiplier * 4500; + if (speed > 4) { + for (i = 0; i < MAX_MODES; ++i) + sf->thresh_mult[i] = INT_MAX; + + sf->thresh_mult[THR_DC ] = 0; + sf->thresh_mult[THR_TM ] = 0; + sf->thresh_mult[THR_NEWMV ] = 4000; + sf->thresh_mult[THR_NEWG ] = 4000; + sf->thresh_mult[THR_NEWA ] = 4000; + sf->thresh_mult[THR_NEARESTMV] = 0; + sf->thresh_mult[THR_NEARESTG ] = 0; + sf->thresh_mult[THR_NEARESTA ] = 0; + sf->thresh_mult[THR_NEARMV ] = 2000; + sf->thresh_mult[THR_NEARG ] = 2000; + sf->thresh_mult[THR_NEARA ] = 2000; + sf->thresh_mult[THR_COMP_NEARESTLA] = 2000; + sf->recode_loop = 0; + } + /* disable frame modes if flags not set */ if (!(cpi->ref_frame_flags & VP9_LAST_FLAG)) { sf->thresh_mult[THR_NEWMV ] = INT_MAX; @@ -778,16 +738,12 @@ void vp9_set_speed_features(VP9_COMP *cpi) { sf->optimize_coefficients = !cpi->oxcf.lossless; sf->first_step = 0; sf->max_step_search_steps = MAX_MVSEARCH_STEPS; - sf->comp_inter_joint_serach = 1; + sf->comp_inter_joint_search = 1; #if CONFIG_MULTIPLE_ARF // Switch segmentation off. sf->static_segmentation = 0; #else -#if CONFIG_IMPLICIT_SEGMENTATION sf->static_segmentation = 0; -#else - sf->static_segmentation = 0; -#endif #endif sf->mb16_breakout = 0; @@ -801,55 +757,16 @@ void vp9_set_speed_features(VP9_COMP *cpi) { // Switch segmentation off. sf->static_segmentation = 0; #else -#if CONFIG_IMPLICIT_SEGMENTATION - sf->static_segmentation = 0; -#else sf->static_segmentation = 0; #endif -#endif sf->mb16_breakout = 0; if (speed > 0) { - /* Disable coefficient optimization above speed 0 */ sf->optimize_coefficients = 0; sf->no_skip_block4x4_search = 0; - sf->comp_inter_joint_serach = 0; - + sf->comp_inter_joint_search = 0; sf->first_step = 1; - - cpi->mode_check_freq[THR_SPLITG] = 2; - cpi->mode_check_freq[THR_SPLITA] = 2; - cpi->mode_check_freq[THR_SPLITMV] = 0; - - cpi->mode_check_freq[THR_COMP_SPLITGA] = 2; - cpi->mode_check_freq[THR_COMP_SPLITLG] = 2; - cpi->mode_check_freq[THR_COMP_SPLITLA] = 0; } - - if (speed > 1) { - cpi->mode_check_freq[THR_SPLITG] = 4; - cpi->mode_check_freq[THR_SPLITA] = 4; - cpi->mode_check_freq[THR_SPLITMV] = 2; - - cpi->mode_check_freq[THR_COMP_SPLITGA] = 4; - cpi->mode_check_freq[THR_COMP_SPLITLG] = 4; - cpi->mode_check_freq[THR_COMP_SPLITLA] = 2; - } - - if (speed > 2) { - cpi->mode_check_freq[THR_SPLITG] = 15; - cpi->mode_check_freq[THR_SPLITA] = 15; - cpi->mode_check_freq[THR_SPLITMV] = 7; - - cpi->mode_check_freq[THR_COMP_SPLITGA] = 15; - cpi->mode_check_freq[THR_COMP_SPLITLG] = 15; - cpi->mode_check_freq[THR_COMP_SPLITLA] = 7; - - // Only do recode loop on key frames, golden frames and - // alt ref frames - sf->recode_loop = 2; - } - break; }; /* switch */ @@ -1194,6 +1111,7 @@ void vp9_change_config(VP9_PTR ptr, VP9_CONFIG *oxcf) { cpi->refresh_golden_frame = 0; cpi->refresh_last_frame = 1; cm->refresh_frame_context = 1; + cm->reset_frame_context = 0; setup_features(cpi); cpi->mb.e_mbd.allow_high_precision_mv = 0; // Default mv precision adaptation @@ -1756,18 +1674,18 @@ void vp9_remove_compressor(VP9_PTR *ptr) { fprintf(fmode, "\n#include \"vp9_entropymode.h\"\n\n"); fprintf(fmode, "const unsigned int vp9_kf_default_bmode_counts "); - fprintf(fmode, "[VP9_KF_BINTRAMODES][VP9_KF_BINTRAMODES]" - "[VP9_KF_BINTRAMODES] =\n{\n"); + fprintf(fmode, "[VP9_INTRA_MODES][VP9_INTRA_MODES]" + "[VP9_INTRA_MODES] =\n{\n"); - for (i = 0; i < VP9_KF_BINTRAMODES; i++) { + for (i = 0; i < VP9_INTRA_MODES; i++) { fprintf(fmode, " { // Above Mode : %d\n", i); - for (j = 0; j < VP9_KF_BINTRAMODES; j++) { + for (j = 0; j < VP9_INTRA_MODES; j++) { fprintf(fmode, " {"); - for (k = 0; k < VP9_KF_BINTRAMODES; k++) { + for (k = 0; k < VP9_INTRA_MODES; k++) { if (!intra_mode_stats[i][j][k]) fprintf(fmode, " %5d, ", 1); else @@ -2621,6 +2539,7 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, (cpi->oxcf.frame_parallel_decoding_mode != 0); if (cm->error_resilient_mode) { cm->frame_parallel_decoding_mode = 1; + cm->reset_frame_context = 0; cm->refresh_frame_context = 0; } } @@ -2931,16 +2850,7 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, } } -#if CONFIG_IMPLICIT_SEGMENTATION - if (!cm->error_resilient_mode && !cpi->sf.static_segmentation) { - configure_implicit_segmentation(cpi, q); - } -#endif - // transform / motion compensation build reconstruction frame - if (cm->frame_type == KEY_FRAME) { - vp9_default_coef_probs(cm); - } vp9_encode_frame(cpi); @@ -3183,15 +3093,7 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, cpi->dummy_packing = 0; vp9_pack_bitstream(cpi, dest, size); -#if CONFIG_IMPLICIT_SEGMENTATION - // Should we allow implicit update of the segment map. - if (xd->allow_implicit_segment_update && !cm->error_resilient_mode) { - vp9_implicit_segment_map_update(cm); - // or has there been an explicit update - } else if (xd->update_mb_segmentation_map) { -#else if (xd->update_mb_segmentation_map) { -#endif update_reference_segmentation_map(cpi); } @@ -3212,10 +3114,8 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, } if (cpi->common.frame_type != KEY_FRAME) { - vp9_copy(cpi->common.fc.sb_ymode_counts, cpi->sb_ymode_count); - vp9_copy(cpi->common.fc.ymode_counts, cpi->ymode_count); + vp9_copy(cpi->common.fc.y_mode_counts, cpi->y_mode_count); vp9_copy(cpi->common.fc.uv_mode_counts, cpi->y_uv_mode_count); - vp9_copy(cpi->common.fc.bmode_counts, cpi->bmode_count); vp9_copy(cpi->common.fc.partition_counts, cpi->partition_count); cpi->common.fc.NMVcount = cpi->NMVcount; if (!cpi->common.error_resilient_mode && @@ -3680,6 +3580,7 @@ int vp9_get_compressed_data(VP9_PTR ptr, unsigned int *frame_flags, } cm->show_frame = 0; + cm->intra_only = 0; cpi->refresh_alt_ref_frame = 1; cpi->refresh_golden_frame = 0; cpi->refresh_last_frame = 0; @@ -3884,6 +3785,7 @@ int vp9_get_compressed_data(VP9_PTR ptr, unsigned int *frame_flags, cpi->droppable = !frame_is_reference(cpi); // return to normal state + cm->reset_frame_context = 0; cm->refresh_frame_context = 1; cpi->refresh_alt_ref_frame = 0; cpi->refresh_golden_frame = 0; diff --git a/vp9/encoder/vp9_onyx_int.h b/vp9/encoder/vp9_onyx_int.h index 15f9571bb..9e259762d 100644 --- a/vp9/encoder/vp9_onyx_int.h +++ b/vp9/encoder/vp9_onyx_int.h @@ -81,17 +81,15 @@ typedef struct { vp9_coeff_probs_model coef_probs_16x16[BLOCK_TYPES]; vp9_coeff_probs_model coef_probs_32x32[BLOCK_TYPES]; - vp9_prob sb_ymode_prob[VP9_I32X32_MODES - 1]; - vp9_prob ymode_prob[VP9_YMODES - 1]; /* interframe intra mode probs */ - vp9_prob uv_mode_prob[VP9_YMODES][VP9_UV_MODES - 1]; - vp9_prob bmode_prob[VP9_BINTRAMODES - 1]; + vp9_prob y_mode_prob[VP9_INTRA_MODES - 1]; /* interframe intra mode probs */ + vp9_prob uv_mode_prob[VP9_INTRA_MODES][VP9_INTRA_MODES - 1]; vp9_prob partition_prob[NUM_PARTITION_CONTEXTS][PARTITION_TYPES - 1]; vp9_prob switchable_interp_prob[VP9_SWITCHABLE_FILTERS + 1] [VP9_SWITCHABLE_FILTERS - 1]; - int mv_ref_ct[INTER_MODE_CONTEXTS][4][2]; - int vp9_mode_contexts[INTER_MODE_CONTEXTS][4]; + int mv_ref_ct[INTER_MODE_CONTEXTS][VP9_MVREFS - 1][2]; + int vp9_mode_contexts[INTER_MODE_CONTEXTS][VP9_MVREFS - 1]; } CODING_CONTEXT; @@ -115,8 +113,7 @@ typedef struct { double new_mv_count; double duration; double count; -} -FIRSTPASS_STATS; +} FIRSTPASS_STATS; typedef struct { int frames_so_far; @@ -128,7 +125,6 @@ typedef struct { double frame_mvr_abs; double frame_mvc; double frame_mvc_abs; - } ONEPASS_FRAMESTATS; typedef struct { @@ -200,8 +196,7 @@ typedef enum { THR_COMP_SPLITLG, THR_COMP_SPLITLA, THR_COMP_SPLITGA, -} -THR_MODES; +} THR_MODES; typedef enum { DIAMOND = 0, @@ -225,7 +220,7 @@ typedef struct { int search_best_filter; int mb16_breakout; int static_segmentation; - int comp_inter_joint_serach; + int comp_inter_joint_search; } SPEED_FEATURES; enum BlockSize { @@ -336,8 +331,8 @@ typedef struct VP9_COMP { int single_pred_count[COMP_PRED_CONTEXTS]; // FIXME contextualize int txfm_count_32x32p[TX_SIZE_MAX_SB]; - int txfm_count_16x16p[TX_SIZE_MAX_MB]; - int txfm_count_8x8p[TX_SIZE_MAX_MB - 1]; + int txfm_count_16x16p[TX_SIZE_MAX_SB - 1]; + int txfm_count_8x8p[TX_SIZE_MAX_SB - 2]; int64_t rd_tx_select_diff[NB_TXFM_MODES]; int rd_tx_select_threshes[4][NB_TXFM_MODES]; @@ -413,28 +408,26 @@ typedef struct VP9_COMP { int cq_target_quality; - int sb_ymode_count [VP9_I32X32_MODES]; - int ymode_count[VP9_YMODES]; /* intra MB type cts this frame */ - int bmode_count[VP9_BINTRAMODES]; - int y_uv_mode_count[VP9_YMODES][VP9_UV_MODES]; + int y_mode_count[VP9_INTRA_MODES]; + int y_uv_mode_count[VP9_INTRA_MODES][VP9_INTRA_MODES]; unsigned int partition_count[NUM_PARTITION_CONTEXTS][PARTITION_TYPES]; nmv_context_counts NMVcount; vp9_coeff_count coef_counts_4x4[BLOCK_TYPES]; - vp9_coeff_probs frame_coef_probs_4x4[BLOCK_TYPES]; + vp9_coeff_probs_model frame_coef_probs_4x4[BLOCK_TYPES]; vp9_coeff_stats frame_branch_ct_4x4[BLOCK_TYPES]; vp9_coeff_count coef_counts_8x8[BLOCK_TYPES]; - vp9_coeff_probs frame_coef_probs_8x8[BLOCK_TYPES]; + vp9_coeff_probs_model frame_coef_probs_8x8[BLOCK_TYPES]; vp9_coeff_stats frame_branch_ct_8x8[BLOCK_TYPES]; vp9_coeff_count coef_counts_16x16[BLOCK_TYPES]; - vp9_coeff_probs frame_coef_probs_16x16[BLOCK_TYPES]; + vp9_coeff_probs_model frame_coef_probs_16x16[BLOCK_TYPES]; vp9_coeff_stats frame_branch_ct_16x16[BLOCK_TYPES]; vp9_coeff_count coef_counts_32x32[BLOCK_TYPES]; - vp9_coeff_probs frame_coef_probs_32x32[BLOCK_TYPES]; + vp9_coeff_probs_model frame_coef_probs_32x32[BLOCK_TYPES]; vp9_coeff_stats frame_branch_ct_32x32[BLOCK_TYPES]; int gfu_boost; @@ -619,7 +612,7 @@ typedef struct VP9_COMP { #endif #ifdef ENTROPY_STATS - int64_t mv_ref_stats[INTER_MODE_CONTEXTS][4][2]; + int64_t mv_ref_stats[INTER_MODE_CONTEXTS][VP9_MVREFS - 1][2]; #endif } VP9_COMP; diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c index cf1132e3a..f4426adaa 100644 --- a/vp9/encoder/vp9_ratectrl.c +++ b/vp9/encoder/vp9_ratectrl.c @@ -124,9 +124,7 @@ void vp9_save_coding_context(VP9_COMP *cpi) { vp9_copy(cc->vp9_mode_contexts, cm->fc.vp9_mode_contexts); - vp9_copy(cc->ymode_prob, cm->fc.ymode_prob); - vp9_copy(cc->sb_ymode_prob, cm->fc.sb_ymode_prob); - vp9_copy(cc->bmode_prob, cm->fc.bmode_prob); + vp9_copy(cc->y_mode_prob, cm->fc.y_mode_prob); vp9_copy(cc->uv_mode_prob, cm->fc.uv_mode_prob); vp9_copy(cc->partition_prob, cm->fc.partition_prob); @@ -163,9 +161,7 @@ void vp9_restore_coding_context(VP9_COMP *cpi) { vp9_copy(cm->fc.vp9_mode_contexts, cc->vp9_mode_contexts); - vp9_copy(cm->fc.ymode_prob, cc->ymode_prob); - vp9_copy(cm->fc.sb_ymode_prob, cc->sb_ymode_prob); - vp9_copy(cm->fc.bmode_prob, cc->bmode_prob); + vp9_copy(cm->fc.y_mode_prob, cc->y_mode_prob); vp9_copy(cm->fc.uv_mode_prob, cc->uv_mode_prob); vp9_copy(cm->fc.partition_prob, cc->partition_prob); diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c index aa0557735..862e72f24 100644 --- a/vp9/encoder/vp9_rdopt.c +++ b/vp9/encoder/vp9_rdopt.c @@ -46,6 +46,12 @@ /* Factor to weigh the rate for switchable interp filters */ #define SWITCHABLE_INTERP_RATE_FACTOR 1 +DECLARE_ALIGNED(16, extern const uint8_t, + vp9_pt_energy_class[MAX_ENTROPY_TOKENS]); + +#define I4X4_PRED 0x8000 +#define SPLITMV 0x10000 + const MODE_DEFINITION vp9_mode_order[MAX_MODES] = { {ZEROMV, LAST_FRAME, NONE}, {DC_PRED, INTRA_FRAME, NONE}, @@ -81,7 +87,7 @@ const MODE_DEFINITION vp9_mode_order[MAX_MODES] = { {SPLITMV, GOLDEN_FRAME, NONE}, {SPLITMV, ALTREF_FRAME, NONE}, - {I4X4_PRED, INTRA_FRAME, NONE}, + {I4X4_PRED, INTRA_FRAME, NONE}, /* compound prediction modes */ {ZEROMV, LAST_FRAME, GOLDEN_FRAME}, @@ -105,11 +111,31 @@ const MODE_DEFINITION vp9_mode_order[MAX_MODES] = { {SPLITMV, GOLDEN_FRAME, ALTREF_FRAME}, }; +#if CONFIG_BALANCED_COEFTREE +static void fill_token_costs(vp9_coeff_count *c, + vp9_coeff_count *cnoskip, + vp9_coeff_probs_model *p, + TX_SIZE tx_size) { + int i, j, k, l; + for (i = 0; i < BLOCK_TYPES; i++) + for (j = 0; j < REF_TYPES; j++) + for (k = 0; k < COEF_BANDS; k++) + for (l = 0; l < PREV_COEF_CONTEXTS; l++) { + vp9_prob probs[ENTROPY_NODES]; + vp9_model_to_full_probs(p[i][j][k][l], probs); + vp9_cost_tokens((int *)cnoskip[i][j][k][l], probs, + vp9_coef_tree); + // Replace the eob node prob with a very small value so that the + // cost approximately equals the cost without the eob node + probs[1] = 1; + vp9_cost_tokens((int *)c[i][j][k][l], probs, vp9_coef_tree); + } +} +#else static void fill_token_costs(vp9_coeff_count *c, vp9_coeff_probs_model *p, TX_SIZE tx_size) { int i, j, k, l; - for (i = 0; i < BLOCK_TYPES; i++) for (j = 0; j < REF_TYPES; j++) for (k = 0; k < COEF_BANDS; k++) @@ -120,6 +146,7 @@ static void fill_token_costs(vp9_coeff_count *c, vp9_coef_tree); } } +#endif static int rd_iifactor[32] = { 4, 4, 3, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, @@ -210,6 +237,20 @@ void vp9_initialize_rd_consts(VP9_COMP *cpi, int qindex) { } } +#if CONFIG_BALANCED_COEFTREE + fill_token_costs(cpi->mb.token_costs[TX_4X4], + cpi->mb.token_costs_noskip[TX_4X4], + cpi->common.fc.coef_probs_4x4, TX_4X4); + fill_token_costs(cpi->mb.token_costs[TX_8X8], + cpi->mb.token_costs_noskip[TX_8X8], + cpi->common.fc.coef_probs_8x8, TX_8X8); + fill_token_costs(cpi->mb.token_costs[TX_16X16], + cpi->mb.token_costs_noskip[TX_16X16], + cpi->common.fc.coef_probs_16x16, TX_16X16); + fill_token_costs(cpi->mb.token_costs[TX_32X32], + cpi->mb.token_costs_noskip[TX_32X32], + cpi->common.fc.coef_probs_32x32, TX_32X32); +#else fill_token_costs(cpi->mb.token_costs[TX_4X4], cpi->common.fc.coef_probs_4x4, TX_4X4); fill_token_costs(cpi->mb.token_costs[TX_8X8], @@ -218,6 +259,7 @@ void vp9_initialize_rd_consts(VP9_COMP *cpi, int qindex) { cpi->common.fc.coef_probs_16x16, TX_16X16); fill_token_costs(cpi->mb.token_costs[TX_32X32], cpi->common.fc.coef_probs_32x32, TX_32X32); +#endif for (i = 0; i < NUM_PARTITION_CONTEXTS; i++) vp9_cost_tokens(cpi->mb.partition_cost[i], @@ -225,7 +267,6 @@ void vp9_initialize_rd_consts(VP9_COMP *cpi, int qindex) { vp9_partition_tree); /*rough estimate for costing*/ - cpi->common.kf_ymode_probs_index = cpi->common.base_qindex >> 4; vp9_init_mode_costs(cpi); if (cpi->common.frame_type != KEY_FRAME) { @@ -271,7 +312,13 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb, TX_TYPE tx_type = DCT_DCT; const int segment_id = xd->mode_info_context->mbmi.segment_id; +#if CONFIG_BALANCED_COEFTREE + unsigned int (*token_costs_noskip)[PREV_COEF_CONTEXTS][MAX_ENTROPY_TOKENS] = + mb->token_costs_noskip[tx_size][type][ref]; +#else vp9_prob coef_probs[COEF_BANDS][PREV_COEF_CONTEXTS][ENTROPY_NODES]; +#endif + int seg_eob, default_eob; uint8_t token_cache[1024]; const uint8_t * band_translate; @@ -291,8 +338,10 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb, get_tx_type_4x4(xd, block) : DCT_DCT; above_ec = A[0] != 0; left_ec = L[0] != 0; +#if !CONFIG_BALANCED_COEFTREE vp9_model_to_full_probs_sb(cm->fc.coef_probs_4x4[type][ref], coef_probs); +#endif seg_eob = 16; scan = get_scan_4x4(tx_type); band_translate = vp9_coefband_trans_4x4; @@ -307,8 +356,10 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb, above_ec = (A[0] + A[1]) != 0; left_ec = (L[0] + L[1]) != 0; scan = get_scan_8x8(tx_type); +#if !CONFIG_BALANCED_COEFTREE vp9_model_to_full_probs_sb(cm->fc.coef_probs_8x8[type][ref], coef_probs); +#endif seg_eob = 64; band_translate = vp9_coefband_trans_8x8plus; break; @@ -320,8 +371,10 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb, TX_TYPE tx_type = (type == PLANE_TYPE_Y_WITH_DC) ? get_tx_type_16x16(xd, y + (x >> 2)) : DCT_DCT; scan = get_scan_16x16(tx_type); +#if !CONFIG_BALANCED_COEFTREE vp9_model_to_full_probs_sb(cm->fc.coef_probs_16x16[type][ref], coef_probs); +#endif seg_eob = 256; above_ec = (A[0] + A[1] + A[2] + A[3]) != 0; left_ec = (L[0] + L[1] + L[2] + L[3]) != 0; @@ -330,8 +383,10 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb, } case TX_32X32: scan = vp9_default_scan_32x32; +#if !CONFIG_BALANCED_COEFTREE vp9_model_to_full_probs_sb(cm->fc.coef_probs_32x32[type][ref], coef_probs); +#endif seg_eob = 1024; above_ec = (A[0] + A[1] + A[2] + A[3] + A[4] + A[5] + A[6] + A[7]) != 0; left_ec = (L[0] + L[1] + L[2] + L[3] + L[4] + L[5] + L[6] + L[7]) != 0; @@ -362,18 +417,30 @@ static INLINE int cost_coeffs(VP9_COMMON *const cm, MACROBLOCK *mb, if (c) pt = vp9_get_coef_context(scan, nb, pad, token_cache, c, default_eob); +#if CONFIG_BALANCED_COEFTREE + if (!c || token_cache[scan[c - 1]]) // do not skip eob + cost += token_costs_noskip[band][pt][t] + vp9_dct_value_cost_ptr[v]; + else + cost += token_costs[band][pt][t] + vp9_dct_value_cost_ptr[v]; +#else cost += token_costs[band][pt][t] + vp9_dct_value_cost_ptr[v]; - if (!c || token_cache[scan[c - 1]]) cost += vp9_cost_bit(coef_probs[band][pt][0], 1); - token_cache[scan[c]] = t; +#endif + token_cache[scan[c]] = vp9_pt_energy_class[t]; } if (c < seg_eob) { if (c) pt = vp9_get_coef_context(scan, nb, pad, token_cache, c, default_eob); +#if CONFIG_BALANCED_COEFTREE + cost += mb->token_costs_noskip[tx_size][type][ref] + [get_coef_band(band_translate, c)] + [pt][DCT_EOB_TOKEN]; +#else cost += mb->token_costs[tx_size][type][ref] [get_coef_band(band_translate, c)] [pt][DCT_EOB_TOKEN]; +#endif } } @@ -556,9 +623,25 @@ static void super_block_yrd(VP9_COMP *cpi, int64_t txfm_cache[NB_TXFM_MODES]) { VP9_COMMON *const cm = &cpi->common; int r[TX_SIZE_MAX_SB][2], d[TX_SIZE_MAX_SB], s[TX_SIZE_MAX_SB]; + MACROBLOCKD *xd = &x->e_mbd; + MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi; vp9_subtract_sby(x, bs); + if (cpi->speed > 4) { + if (bs >= BLOCK_SIZE_SB32X32) { + mbmi->txfm_size = TX_32X32; + } else if (bs >= BLOCK_SIZE_MB16X16) { + mbmi->txfm_size = TX_16X16; + } else if (bs >= BLOCK_SIZE_SB8X8) { + mbmi->txfm_size = TX_8X8; + } else { + mbmi->txfm_size = TX_4X4; + } + super_block_yrd_for_txfm(cm, x, rate, distortion, skip, bs, + mbmi->txfm_size); + return; + } if (bs >= BLOCK_SIZE_SB32X32) super_block_yrd_for_txfm(cm, x, &r[TX_32X32][0], &d[TX_32X32], &s[TX_32X32], bs, TX_32X32); @@ -611,11 +694,7 @@ static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib, int64_t this_rd; int ratey = 0; - xd->mode_info_context->bmi[ib].as_mode.first = mode; - if (cm->frame_type == KEY_FRAME) - rate = bmode_costs[mode]; - else - rate = x->mbmode_cost[cm->frame_type][mode]; + rate = bmode_costs[mode]; distortion = 0; vpx_memcpy(tempa, ta, sizeof(ta)); @@ -653,9 +732,6 @@ static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib, distortion += vp9_block_error(coeff, BLOCK_OFFSET(xd->plane[0].dqcoeff, block, 16), 16) >> 2; - vp9_intra4x4_predict(xd, block, BLOCK_SIZE_SB8X8, mode, - dst, xd->plane[0].dst.stride); - if (best_tx_type != DCT_DCT) vp9_short_iht4x4_add(BLOCK_OFFSET(xd->plane[0].dqcoeff, block, 16), dst, xd->plane[0].dst.stride, best_tx_type); @@ -726,16 +802,15 @@ static int64_t rd_pick_intra4x4mby_modes(VP9_COMP *cpi, MACROBLOCK *mb, int64_t total_rd = 0; ENTROPY_CONTEXT t_above[4], t_left[4]; int *bmode_costs; + MODE_INFO *const mic = xd->mode_info_context; vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above)); vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left)); - xd->mode_info_context->mbmi.mode = I4X4_PRED; - bmode_costs = mb->inter_bmode_costs; + bmode_costs = mb->mbmode_cost; for (idy = 0; idy < 2; idy += bh) { for (idx = 0; idx < 2; idx += bw) { - MODE_INFO *const mic = xd->mode_info_context; const int mis = xd->mode_info_stride; MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode); int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(ry); @@ -747,7 +822,7 @@ static int64_t rd_pick_intra4x4mby_modes(VP9_COMP *cpi, MACROBLOCK *mb, const MB_PREDICTION_MODE L = (xd->left_available || idx) ? left_block_mode(mic, i) : DC_PRED; - bmode_costs = mb->bmode_costs[A][L]; + bmode_costs = mb->y_mode_costs[A][L]; } total_rd += rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs, @@ -774,6 +849,7 @@ static int64_t rd_pick_intra4x4mby_modes(VP9_COMP *cpi, MACROBLOCK *mb, *Rate = cost; *rate_y = tot_rate_y; *Distortion = distortion; + xd->mode_info_context->mbmi.mode = mic->bmi[3].as_mode.first; return RDCOST(mb->rdmult, mb->rddiv, cost, distortion); } @@ -785,12 +861,13 @@ static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x, int64_t txfm_cache[NB_TXFM_MODES]) { MB_PREDICTION_MODE mode; MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected); - MACROBLOCKD *xd = &x->e_mbd; + MACROBLOCKD *const xd = &x->e_mbd; int this_rate, this_rate_tokenonly; int this_distortion, s; int64_t best_rd = INT64_MAX, this_rd; TX_SIZE UNINITIALIZED_IS_SAFE(best_tx); int i; + int *bmode_costs = x->mbmode_cost; if (bsize < BLOCK_SIZE_SB8X8) { x->e_mbd.mode_info_context->mbmi.txfm_size = TX_4X4; @@ -805,17 +882,19 @@ static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x, int64_t local_txfm_cache[NB_TXFM_MODES]; MODE_INFO *const mic = xd->mode_info_context; const int mis = xd->mode_info_stride; - const MB_PREDICTION_MODE A = above_block_mode(mic, 0, mis); - const MB_PREDICTION_MODE L = xd->left_available ? - left_block_mode(mic, 0) : DC_PRED; - - int *bmode_costs = x->bmode_costs[A][L]; + if (cpi->common.frame_type == KEY_FRAME) { + const MB_PREDICTION_MODE A = above_block_mode(mic, 0, mis); + const MB_PREDICTION_MODE L = xd->left_available ? + left_block_mode(mic, 0) : DC_PRED; + bmode_costs = x->y_mode_costs[A][L]; + } x->e_mbd.mode_info_context->mbmi.mode = mode; vp9_build_intra_predictors_sby_s(&x->e_mbd, bsize); super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, bsize, local_txfm_cache); + this_rate = this_rate_tokenonly + bmode_costs[mode]; this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion); @@ -925,10 +1004,10 @@ int vp9_cost_mv_ref(VP9_COMP *cpi, VP9_COMMON *pc = &cpi->common; vp9_prob p [VP9_MVREFS - 1]; - assert(NEARESTMV <= m && m <= SPLITMV); + assert(NEARESTMV <= m && m <= NEWMV); vp9_mv_ref_probs(pc, p, mode_context); - return cost_token(vp9_mv_ref_tree, p, - vp9_mv_ref_encoding_array - NEARESTMV + m); + return cost_token(vp9_sb_mv_ref_tree, p, + vp9_sb_mv_ref_encoding_array - NEARESTMV + m); } else return 0; } @@ -938,19 +1017,18 @@ void vp9_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, int_mv *mv) { x->e_mbd.mode_info_context->mbmi.mv[0].as_int = mv->as_int; } -static int labels2mode(MACROBLOCK *x, - int const *labelings, int which_label, +static int labels2mode(MACROBLOCK *x, int i, MB_PREDICTION_MODE this_mode, int_mv *this_mv, int_mv *this_second_mv, int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], - int_mv seg_mvs[MAX_REF_FRAMES - 1], + int_mv seg_mvs[MAX_REF_FRAMES], int_mv *best_ref_mv, int_mv *second_best_ref_mv, int *mvjcost, int *mvcost[2], VP9_COMP *cpi) { MACROBLOCKD *const xd = &x->e_mbd; MODE_INFO *const mic = xd->mode_info_context; MB_MODE_INFO * mbmi = &mic->mbmi; - int i, cost = 0, thismvcost = 0; + int cost = 0, thismvcost = 0; int idx, idy; int bw = 1 << b_width_log2(mbmi->sb_type); int bh = 1 << b_height_log2(mbmi->sb_type); @@ -958,72 +1036,61 @@ static int labels2mode(MACROBLOCK *x, /* We have to be careful retrieving previously-encoded motion vectors. Ones from this macroblock have to be pulled from the BLOCKD array as they have not yet made it to the bmi array in our MB_MODE_INFO. */ - for (i = 0; i < 4; ++i) { - MB_PREDICTION_MODE m; - - if (labelings[i] != which_label) - continue; + MB_PREDICTION_MODE m; - { - // the only time we should do costing for new motion vector or mode - // is when we are on a new label (jbb May 08, 2007) - switch (m = this_mode) { - case NEWMV: - if (mbmi->second_ref_frame > 0) { - this_mv->as_int = seg_mvs[mbmi->ref_frame - 1].as_int; - this_second_mv->as_int = - seg_mvs[mbmi->second_ref_frame - 1].as_int; - } - - thismvcost = vp9_mv_bit_cost(this_mv, best_ref_mv, mvjcost, mvcost, - 102, xd->allow_high_precision_mv); - if (mbmi->second_ref_frame > 0) { - thismvcost += vp9_mv_bit_cost(this_second_mv, second_best_ref_mv, - mvjcost, mvcost, 102, - xd->allow_high_precision_mv); - } - break; - case NEARESTMV: - this_mv->as_int = frame_mv[NEARESTMV][mbmi->ref_frame].as_int; - if (mbmi->second_ref_frame > 0) - this_second_mv->as_int = - frame_mv[NEARESTMV][mbmi->second_ref_frame].as_int; - break; - case NEARMV: - this_mv->as_int = frame_mv[NEARMV][mbmi->ref_frame].as_int; - if (mbmi->second_ref_frame > 0) - this_second_mv->as_int = - frame_mv[NEARMV][mbmi->second_ref_frame].as_int; - break; - case ZEROMV: - this_mv->as_int = 0; - if (mbmi->second_ref_frame > 0) - this_second_mv->as_int = 0; - break; - default: - break; + // the only time we should do costing for new motion vector or mode + // is when we are on a new label (jbb May 08, 2007) + switch (m = this_mode) { + case NEWMV: + this_mv->as_int = seg_mvs[mbmi->ref_frame].as_int; + thismvcost = vp9_mv_bit_cost(this_mv, best_ref_mv, mvjcost, mvcost, + 102, xd->allow_high_precision_mv); + if (mbmi->second_ref_frame > 0) { + this_second_mv->as_int = seg_mvs[mbmi->second_ref_frame].as_int; + thismvcost += vp9_mv_bit_cost(this_second_mv, second_best_ref_mv, + mvjcost, mvcost, 102, + xd->allow_high_precision_mv); } + break; + case NEARESTMV: + this_mv->as_int = frame_mv[NEARESTMV][mbmi->ref_frame].as_int; + if (mbmi->second_ref_frame > 0) + this_second_mv->as_int = + frame_mv[NEARESTMV][mbmi->second_ref_frame].as_int; + break; + case NEARMV: + this_mv->as_int = frame_mv[NEARMV][mbmi->ref_frame].as_int; + if (mbmi->second_ref_frame > 0) + this_second_mv->as_int = + frame_mv[NEARMV][mbmi->second_ref_frame].as_int; + break; + case ZEROMV: + this_mv->as_int = 0; + if (mbmi->second_ref_frame > 0) + this_second_mv->as_int = 0; + break; + default: + break; + } - cost = vp9_cost_mv_ref(cpi, this_mode, - mbmi->mb_mode_context[mbmi->ref_frame]); - } + cost = vp9_cost_mv_ref(cpi, this_mode, + mbmi->mb_mode_context[mbmi->ref_frame]); - mic->bmi[i].as_mv[0].as_int = this_mv->as_int; - if (mbmi->second_ref_frame > 0) - mic->bmi[i].as_mv[1].as_int = this_second_mv->as_int; + mic->bmi[i].as_mv[0].as_int = this_mv->as_int; + if (mbmi->second_ref_frame > 0) + mic->bmi[i].as_mv[1].as_int = this_second_mv->as_int; - x->partition_info->bmi[i].mode = m; - x->partition_info->bmi[i].mv.as_int = this_mv->as_int; - if (mbmi->second_ref_frame > 0) - x->partition_info->bmi[i].second_mv.as_int = this_second_mv->as_int; - for (idy = 0; idy < bh; ++idy) { - for (idx = 0; idx < bw; ++idx) { - vpx_memcpy(&mic->bmi[i + idy * 2 + idx], - &mic->bmi[i], sizeof(mic->bmi[i])); - vpx_memcpy(&x->partition_info->bmi[i + idy * 2 + idx], - &x->partition_info->bmi[i], - sizeof(x->partition_info->bmi[i])); - } + x->partition_info->bmi[i].mode = m; + x->partition_info->bmi[i].mv.as_int = this_mv->as_int; + if (mbmi->second_ref_frame > 0) + x->partition_info->bmi[i].second_mv.as_int = this_second_mv->as_int; + for (idy = 0; idy < bh; ++idy) { + for (idx = 0; idx < bw; ++idx) { + vpx_memcpy(&mic->bmi[i + idy * 2 + idx], + &mic->bmi[i], sizeof(mic->bmi[i])); + vpx_memcpy(&x->partition_info->bmi[i + idy * 2 + idx], + &x->partition_info->bmi[i], + sizeof(x->partition_info->bmi[i])); } } @@ -1033,90 +1100,86 @@ static int labels2mode(MACROBLOCK *x, static int64_t encode_inter_mb_segment(VP9_COMMON *const cm, MACROBLOCK *x, - int const *labels, - int which_label, + int i, int *labelyrate, int *distortion, ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl) { - int i, k; + int k; MACROBLOCKD *xd = &x->e_mbd; BLOCK_SIZE_TYPE bsize = xd->mode_info_context->mbmi.sb_type; int bwl = b_width_log2(bsize), bw = 1 << bwl; int bhl = b_height_log2(bsize), bh = 1 << bhl; int idx, idy; + const int src_stride = x->plane[0].src.stride; + uint8_t* const src = + raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, + x->plane[0].src.buf, src_stride); + int16_t* src_diff = + raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, i, + x->plane[0].src_diff); + int16_t* coeff = BLOCK_OFFSET(x->plane[0].coeff, 16, i); + uint8_t* const pre = + raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, + xd->plane[0].pre[0].buf, + xd->plane[0].pre[0].stride); + uint8_t* const dst = + raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, + xd->plane[0].dst.buf, + xd->plane[0].dst.stride); + int thisdistortion = 0; + int thisrate = 0; *labelyrate = 0; *distortion = 0; - for (i = 0; i < 4; i++) { - if (labels[i] == which_label) { - const int src_stride = x->plane[0].src.stride; - uint8_t* const src = - raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, - x->plane[0].src.buf, src_stride); - int16_t* src_diff = - raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, i, - x->plane[0].src_diff); - int16_t* coeff = BLOCK_OFFSET(x->plane[0].coeff, 16, i); - uint8_t* const pre = - raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, - xd->plane[0].pre[0].buf, - xd->plane[0].pre[0].stride); - uint8_t* const dst = - raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, - xd->plane[0].dst.buf, - xd->plane[0].dst.stride); - int thisdistortion = 0; - int thisrate = 0; - - vp9_build_inter_predictor(pre, - xd->plane[0].pre[0].stride, - dst, - xd->plane[0].dst.stride, - &xd->mode_info_context->bmi[i].as_mv[0], - &xd->scale_factor[0], - 4 * bw, 4 * bh, 0 /* no avg */, &xd->subpix); - - // TODO(debargha): Make this work properly with the - // implicit-compoundinter-weight experiment when implicit - // weighting for splitmv modes is turned on. - if (xd->mode_info_context->mbmi.second_ref_frame > 0) { - uint8_t* const second_pre = - raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, - xd->plane[0].pre[1].buf, - xd->plane[0].pre[1].stride); - vp9_build_inter_predictor(second_pre, xd->plane[0].pre[1].stride, - dst, xd->plane[0].dst.stride, - &xd->mode_info_context->bmi[i].as_mv[1], - &xd->scale_factor[1], 4 * bw, 4 * bh, 1, - &xd->subpix); - } - vp9_subtract_block(4 * bh, 4 * bw, src_diff, 8, - src, src_stride, - dst, xd->plane[0].dst.stride); + vp9_build_inter_predictor(pre, + xd->plane[0].pre[0].stride, + dst, + xd->plane[0].dst.stride, + &xd->mode_info_context->bmi[i].as_mv[0], + &xd->scale_factor[0], + 4 * bw, 4 * bh, 0 /* no avg */, &xd->subpix); + + // TODO(debargha): Make this work properly with the + // implicit-compoundinter-weight experiment when implicit + // weighting for splitmv modes is turned on. + if (xd->mode_info_context->mbmi.second_ref_frame > 0) { + uint8_t* const second_pre = + raster_block_offset_uint8(xd, BLOCK_SIZE_SB8X8, 0, i, + xd->plane[0].pre[1].buf, + xd->plane[0].pre[1].stride); + vp9_build_inter_predictor(second_pre, xd->plane[0].pre[1].stride, + dst, xd->plane[0].dst.stride, + &xd->mode_info_context->bmi[i].as_mv[1], + &xd->scale_factor[1], 4 * bw, 4 * bh, 1, + &xd->subpix); + } - k = i; - for (idy = 0; idy < bh; ++idy) { - for (idx = 0; idx < bw; ++idx) { - k += (idy * 2 + idx); - src_diff = raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, k, - x->plane[0].src_diff); - coeff = BLOCK_OFFSET(x->plane[0].coeff, 16, k); - x->fwd_txm4x4(src_diff, coeff, 16); - x->quantize_b_4x4(x, k, DCT_DCT, 16); - thisdistortion += vp9_block_error(coeff, - BLOCK_OFFSET(xd->plane[0].dqcoeff, - k, 16), 16); - thisrate += cost_coeffs(cm, x, 0, k, PLANE_TYPE_Y_WITH_DC, - ta + (k & 1), - tl + (k >> 1), TX_4X4, 16); - } - } - *distortion += thisdistortion; - *labelyrate += thisrate; + vp9_subtract_block(4 * bh, 4 * bw, src_diff, 8, + src, src_stride, + dst, xd->plane[0].dst.stride); + + k = i; + for (idy = 0; idy < bh; ++idy) { + for (idx = 0; idx < bw; ++idx) { + k += (idy * 2 + idx); + src_diff = raster_block_offset_int16(xd, BLOCK_SIZE_SB8X8, 0, k, + x->plane[0].src_diff); + coeff = BLOCK_OFFSET(x->plane[0].coeff, 16, k); + x->fwd_txm4x4(src_diff, coeff, 16); + x->quantize_b_4x4(x, k, DCT_DCT, 16); + thisdistortion += vp9_block_error(coeff, + BLOCK_OFFSET(xd->plane[0].dqcoeff, + k, 16), 16); + thisrate += cost_coeffs(cm, x, 0, k, PLANE_TYPE_Y_WITH_DC, + ta + (k & 1), + tl + (k >> 1), TX_4X4, 16); } } + *distortion += thisdistortion; + *labelyrate += thisrate; + *distortion >>= 2; return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion); } @@ -1188,11 +1251,45 @@ static enum BlockSize get_block_size(int bw, int bh) { return -1; } +static INLINE void mi_buf_shift(MACROBLOCK *x, int i) { + MB_MODE_INFO *mbmi = &x->e_mbd.mode_info_context->mbmi; + x->plane[0].src.buf = + raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, i, + x->plane[0].src.buf, + x->plane[0].src.stride); + assert(((intptr_t)x->e_mbd.plane[0].pre[0].buf & 0x7) == 0); + x->e_mbd.plane[0].pre[0].buf = + raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, i, + x->e_mbd.plane[0].pre[0].buf, + x->e_mbd.plane[0].pre[0].stride); + if (mbmi->second_ref_frame) + x->e_mbd.plane[0].pre[1].buf = + raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, i, + x->e_mbd.plane[0].pre[1].buf, + x->e_mbd.plane[0].pre[1].stride); +} + +static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src, + struct buf_2d orig_pre[2]) { + MB_MODE_INFO *mbmi = &x->e_mbd.mode_info_context->mbmi; + x->plane[0].src = orig_src; + x->e_mbd.plane[0].pre[0] = orig_pre[0]; + if (mbmi->second_ref_frame) + x->e_mbd.plane[0].pre[1] = orig_pre[1]; +} + +static void iterative_motion_search(VP9_COMP *cpi, MACROBLOCK *x, + BLOCK_SIZE_TYPE bsize, + int_mv *frame_mv, + YV12_BUFFER_CONFIG **scaled_ref_frame, + int mi_row, int mi_col, + int_mv single_newmv[MAX_REF_FRAMES]); + static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, BEST_SEG_INFO *bsi, - int_mv seg_mvs[4][MAX_REF_FRAMES - 1]) { + int_mv seg_mvs[4][MAX_REF_FRAMES], + int mi_row, int mi_col) { int i, j; - static const int labels[4] = { 0, 1, 2, 3 }; int br = 0, bd = 0; MB_PREDICTION_MODE this_mode; MB_MODE_INFO * mbmi = &x->e_mbd.mode_info_context->mbmi; @@ -1208,7 +1305,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, int bhl = b_height_log2(bsize), bh = 1 << bhl; int idx, idy; vp9_variance_fn_ptr_t *v_fn_ptr; - + YV12_BUFFER_CONFIG *scaled_ref_frame[2] = {NULL, NULL}; ENTROPY_CONTEXT t_above[4], t_left[4]; ENTROPY_CONTEXT t_above_b[4], t_left_b[4]; @@ -1255,18 +1352,21 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, int distortion; int labelyrate; ENTROPY_CONTEXT t_above_s[4], t_left_s[4]; + const struct buf_2d orig_src = x->plane[0].src; + struct buf_2d orig_pre[2]; + + vpx_memcpy(orig_pre, x->e_mbd.plane[0].pre, sizeof(orig_pre)); vpx_memcpy(t_above_s, t_above, sizeof(t_above_s)); vpx_memcpy(t_left_s, t_left, sizeof(t_left_s)); // motion search for newmv (single predictor case only) if (mbmi->second_ref_frame <= 0 && this_mode == NEWMV) { - int sseshift, n; int step_param = 0; int further_steps; int thissme, bestsme = INT_MAX; - const struct buf_2d orig_src = x->plane[0].src; - const struct buf_2d orig_pre = x->e_mbd.plane[0].pre[0]; + int sadpb = x->sadperbit4; + int_mv mvp_full; /* Is the best so far sufficiently good that we cant justify doing * and new motion search. */ @@ -1287,55 +1387,35 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param; - { - int sadpb = x->sadperbit4; - int_mv mvp_full; - - mvp_full.as_mv.row = bsi->mvp.as_mv.row >> 3; - mvp_full.as_mv.col = bsi->mvp.as_mv.col >> 3; - - // find first label - n = i; - - // adjust src pointer for this segment - x->plane[0].src.buf = - raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, n, - x->plane[0].src.buf, - x->plane[0].src.stride); - assert(((intptr_t)x->e_mbd.plane[0].pre[0].buf & 0x7) == 0); - x->e_mbd.plane[0].pre[0].buf = - raster_block_offset_uint8(&x->e_mbd, BLOCK_SIZE_SB8X8, 0, n, - x->e_mbd.plane[0].pre[0].buf, - x->e_mbd.plane[0].pre[0].stride); - - bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param, - sadpb, further_steps, 0, v_fn_ptr, - bsi->ref_mv, &mode_mv[NEWMV]); - - sseshift = 0; - - // Should we do a full search (best quality only) - if ((cpi->compressor_speed == 0) && (bestsme >> sseshift) > 4000) { - /* Check if mvp_full is within the range. */ - clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max, - x->mv_row_min, x->mv_row_max); - - thissme = cpi->full_search_sad(x, &mvp_full, - sadpb, 16, v_fn_ptr, - x->nmvjointcost, x->mvcost, - bsi->ref_mv, - n); - - if (thissme < bestsme) { - bestsme = thissme; - mode_mv[NEWMV].as_int = - x->e_mbd.mode_info_context->bmi[n].as_mv[0].as_int; - } else { - /* The full search result is actually worse so re-instate the - * previous best vector */ - x->e_mbd.mode_info_context->bmi[n].as_mv[0].as_int = - mode_mv[NEWMV].as_int; - } + mvp_full.as_mv.row = bsi->mvp.as_mv.row >> 3; + mvp_full.as_mv.col = bsi->mvp.as_mv.col >> 3; + + // adjust src pointer for this block + mi_buf_shift(x, i); + bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param, + sadpb, further_steps, 0, v_fn_ptr, + bsi->ref_mv, &mode_mv[NEWMV]); + + // Should we do a full search (best quality only) + if (cpi->compressor_speed == 0) { + /* Check if mvp_full is within the range. */ + clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max, + x->mv_row_min, x->mv_row_max); + + thissme = cpi->full_search_sad(x, &mvp_full, + sadpb, 16, v_fn_ptr, + x->nmvjointcost, x->mvcost, + bsi->ref_mv, i); + + if (thissme < bestsme) { + bestsme = thissme; + mode_mv[NEWMV].as_int = + x->e_mbd.mode_info_context->bmi[i].as_mv[0].as_int; + } else { + /* The full search result is actually worse so re-instate the + * previous best vector */ + x->e_mbd.mode_info_context->bmi[i].as_mv[0].as_int = + mode_mv[NEWMV].as_int; } } @@ -1348,23 +1428,32 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, &distortion, &sse); // safe motion search result for use in compound prediction - seg_mvs[i][mbmi->ref_frame - 1].as_int = mode_mv[NEWMV].as_int; + seg_mvs[i][mbmi->ref_frame].as_int = mode_mv[NEWMV].as_int; } // restore src pointers - x->plane[0].src = orig_src; - x->e_mbd.plane[0].pre[0] = orig_pre; + mi_buf_restore(x, orig_src, orig_pre); } else if (mbmi->second_ref_frame > 0 && this_mode == NEWMV) { - /* NEW4X4 */ - /* motion search not completed? Then skip newmv for this block with - * comppred */ - if (seg_mvs[i][mbmi->second_ref_frame - 1].as_int == INVALID_MV || - seg_mvs[i][mbmi->ref_frame - 1].as_int == INVALID_MV) { + if (seg_mvs[i][mbmi->second_ref_frame].as_int == INVALID_MV || + seg_mvs[i][mbmi->ref_frame ].as_int == INVALID_MV) continue; + + // adjust src pointers + mi_buf_shift(x, i); + if (cpi->sf.comp_inter_joint_search) { + iterative_motion_search(cpi, x, bsize, frame_mv[this_mode], + scaled_ref_frame, + mi_row, mi_col, seg_mvs[i]); + seg_mvs[i][mbmi->ref_frame].as_int = + frame_mv[this_mode][mbmi->ref_frame].as_int; + seg_mvs[i][mbmi->second_ref_frame].as_int = + frame_mv[this_mode][mbmi->second_ref_frame].as_int; } + // restore src pointers + mi_buf_restore(x, orig_src, orig_pre); } - rate = labels2mode(x, labels, i, this_mode, &mode_mv[this_mode], + rate = labels2mode(x, i, this_mode, &mode_mv[this_mode], &second_mode_mv[this_mode], frame_mv, seg_mvs[i], bsi->ref_mv, bsi->second_ref_mv, x->nmvjointcost, x->mvcost, cpi); @@ -1381,7 +1470,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, continue; this_rd = encode_inter_mb_segment(&cpi->common, - x, labels, i, &labelyrate, + x, i, &labelyrate, &distortion, t_above_s, t_left_s); this_rd += RDCOST(x->rdmult, x->rddiv, rate, 0); rate += labelyrate; @@ -1392,10 +1481,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, bestlabelyrate = labelyrate; mode_selected = this_mode; best_label_rd = this_rd; - for (j = 0; j < 4; j++) - if (labels[j] == i) - best_eobs[j] = x->e_mbd.plane[0].eobs[j]; - + best_eobs[i] = x->e_mbd.plane[0].eobs[i]; vpx_memcpy(t_above_b, t_above_s, sizeof(t_above_s)); vpx_memcpy(t_left_b, t_left_s, sizeof(t_left_s)); } @@ -1404,7 +1490,7 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, vpx_memcpy(t_above, t_above_b, sizeof(t_above)); vpx_memcpy(t_left, t_left_b, sizeof(t_left)); - labels2mode(x, labels, i, mode_selected, &mode_mv[mode_selected], + labels2mode(x, i, mode_selected, &mode_mv[mode_selected], &second_mode_mv[mode_selected], frame_mv, seg_mvs[i], bsi->ref_mv, bsi->second_ref_mv, x->nmvjointcost, x->mvcost, cpi); @@ -1443,12 +1529,6 @@ static void rd_check_segment_txsize(VP9_COMP *cpi, MACROBLOCK *x, } } -static void rd_check_segment(VP9_COMP *cpi, MACROBLOCK *x, - BEST_SEG_INFO *bsi, - int_mv seg_mvs[4][MAX_REF_FRAMES - 1]) { - rd_check_segment_txsize(cpi, x, bsi, seg_mvs); -} - static int rd_pick_best_mbsegmentation(VP9_COMP *cpi, MACROBLOCK *x, int_mv *best_ref_mv, int_mv *second_best_ref_mv, @@ -1457,7 +1537,8 @@ static int rd_pick_best_mbsegmentation(VP9_COMP *cpi, MACROBLOCK *x, int *returnyrate, int *returndistortion, int *skippable, int mvthresh, - int_mv seg_mvs[4][MAX_REF_FRAMES - 1]) { + int_mv seg_mvs[4][MAX_REF_FRAMES], + int mi_row, int mi_col) { int i; BEST_SEG_INFO bsi; MB_MODE_INFO * mbmi = &x->e_mbd.mode_info_context->mbmi; @@ -1473,7 +1554,7 @@ static int rd_pick_best_mbsegmentation(VP9_COMP *cpi, MACROBLOCK *x, for (i = 0; i < 4; i++) bsi.modes[i] = ZEROMV; - rd_check_segment(cpi, x, &bsi, seg_mvs); + rd_check_segment_txsize(cpi, x, &bsi, seg_mvs, mi_row, mi_col); /* set it to the best */ for (i = 0; i < 4; i++) { @@ -1504,6 +1585,7 @@ static int rd_pick_best_mbsegmentation(VP9_COMP *cpi, MACROBLOCK *x, *returndistortion = bsi.d; *returnyrate = bsi.segment_yrate; *skippable = vp9_sby_is_skippable(&x->e_mbd, BLOCK_SIZE_SB8X8); + mbmi->mode = bsi.modes[3]; return (int)(bsi.segment_rd); } @@ -1878,6 +1960,154 @@ static INLINE int get_switchable_rate(VP9_COMMON *cm, MACROBLOCK *x) { return SWITCHABLE_INTERP_RATE_FACTOR * x->switchable_interp_costs[c][m]; } +static void iterative_motion_search(VP9_COMP *cpi, MACROBLOCK *x, + BLOCK_SIZE_TYPE bsize, + int_mv *frame_mv, + YV12_BUFFER_CONFIG **scaled_ref_frame, + int mi_row, int mi_col, + int_mv single_newmv[MAX_REF_FRAMES]) { + int pw = 4 << b_width_log2(bsize), ph = 4 << b_height_log2(bsize); + MACROBLOCKD *xd = &x->e_mbd; + MB_MODE_INFO *mbmi = &xd->mode_info_context->mbmi; + int refs[2] = { mbmi->ref_frame, + (mbmi->second_ref_frame < 0 ? 0 : mbmi->second_ref_frame) }; + int_mv ref_mv[2]; + const enum BlockSize block_size = get_plane_block_size(bsize, &xd->plane[0]); + int ite; + // Prediction buffer from second frame. + uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t)); + + // Do joint motion search in compound mode to get more accurate mv. + struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}}; + struct buf_2d backup_second_yv12[MAX_MB_PLANE] = {{0}}; + struct buf_2d scaled_first_yv12; + int last_besterr[2] = {INT_MAX, INT_MAX}; + + ref_mv[0] = mbmi->ref_mvs[refs[0]][0]; + ref_mv[1] = mbmi->ref_mvs[refs[1]][0]; + + if (scaled_ref_frame[0]) { + int i; + // Swap out the reference frame for a version that's been scaled to + // match the resolution of the current frame, allowing the existing + // motion search code to be used without additional modifications. + for (i = 0; i < MAX_MB_PLANE; i++) + backup_yv12[i] = xd->plane[i].pre[0]; + setup_pre_planes(xd, scaled_ref_frame[0], NULL, mi_row, mi_col, + NULL, NULL); + } + + if (scaled_ref_frame[1]) { + int i; + for (i = 0; i < MAX_MB_PLANE; i++) + backup_second_yv12[i] = xd->plane[i].pre[1]; + + setup_pre_planes(xd, scaled_ref_frame[1], NULL, mi_row, mi_col, + NULL, NULL); + } + + xd->scale_factor[0].set_scaled_offsets(&xd->scale_factor[0], + mi_row, mi_col); + xd->scale_factor[1].set_scaled_offsets(&xd->scale_factor[1], + mi_row, mi_col); + scaled_first_yv12 = xd->plane[0].pre[0]; + + // Initialize mv using single prediction mode result. + frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int; + frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int; + + // Allow joint search multiple times iteratively for each ref frame + // and break out the search loop if it couldn't find better mv. + for (ite = 0; ite < 4; ite++) { + struct buf_2d ref_yv12[2]; + int bestsme = INT_MAX; + int sadpb = x->sadperbit16; + int_mv tmp_mv; + int search_range = 3; + + int tmp_col_min = x->mv_col_min; + int tmp_col_max = x->mv_col_max; + int tmp_row_min = x->mv_row_min; + int tmp_row_max = x->mv_row_max; + int id = ite % 2; + + // Initialized here because of compiler problem in Visual Studio. + ref_yv12[0] = xd->plane[0].pre[0]; + ref_yv12[1] = xd->plane[0].pre[1]; + + // Get pred block from second frame. + vp9_build_inter_predictor(ref_yv12[!id].buf, + ref_yv12[!id].stride, + second_pred, pw, + &frame_mv[refs[!id]], + &xd->scale_factor[!id], + pw, ph, 0, + &xd->subpix); + + // Compound motion search on first ref frame. + if (id) + xd->plane[0].pre[0] = ref_yv12[id]; + vp9_clamp_mv_min_max(x, &ref_mv[id]); + + // Use mv result from single mode as mvp. + tmp_mv.as_int = frame_mv[refs[id]].as_int; + + tmp_mv.as_mv.col >>= 3; + tmp_mv.as_mv.row >>= 3; + + // Small-range full-pixel motion search + bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, + search_range, + &cpi->fn_ptr[block_size], + x->nmvjointcost, x->mvcost, + &ref_mv[id], second_pred, + pw, ph); + + x->mv_col_min = tmp_col_min; + x->mv_col_max = tmp_col_max; + x->mv_row_min = tmp_row_min; + x->mv_row_max = tmp_row_max; + + if (bestsme < INT_MAX) { + int dis; /* TODO: use dis in distortion calculation later. */ + unsigned int sse; + + bestsme = vp9_find_best_sub_pixel_comp(x, &tmp_mv, + &ref_mv[id], + x->errorperbit, + &cpi->fn_ptr[block_size], + x->nmvjointcost, x->mvcost, + &dis, &sse, second_pred, + pw, ph); + } + + if (id) + xd->plane[0].pre[0] = scaled_first_yv12; + + if (bestsme < last_besterr[id]) { + frame_mv[refs[id]].as_int = tmp_mv.as_int; + last_besterr[id] = bestsme; + } else { + break; + } + } + + // restore the predictor + if (scaled_ref_frame[0]) { + int i; + for (i = 0; i < MAX_MB_PLANE; i++) + xd->plane[i].pre[0] = backup_yv12[i]; + } + + if (scaled_ref_frame[1]) { + int i; + for (i = 0; i < MAX_MB_PLANE; i++) + xd->plane[i].pre[1] = backup_second_yv12[i]; + } + + vpx_free(second_pred); +} + static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE_TYPE bsize, int64_t txfm_cache[], @@ -1920,145 +2150,13 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, ref_mv[1] = mbmi->ref_mvs[refs[1]][0]; if (is_comp_pred) { - if (cpi->sf.comp_inter_joint_serach) { - int pw = 4 << b_width_log2(bsize), ph = 4 << b_height_log2(bsize); - int ite; - // Prediction buffer from second frame. - uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t)); - - // Do joint motion search in compound mode to get more accurate mv. - struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}}; - struct buf_2d backup_second_yv12[MAX_MB_PLANE] = {{0}}; - struct buf_2d scaled_first_yv12; - int last_besterr[2] = {INT_MAX, INT_MAX}; - - if (scaled_ref_frame[0]) { - int i; - - // Swap out the reference frame for a version that's been scaled to - // match the resolution of the current frame, allowing the existing - // motion search code to be used without additional modifications. - for (i = 0; i < MAX_MB_PLANE; i++) - backup_yv12[i] = xd->plane[i].pre[0]; - - setup_pre_planes(xd, scaled_ref_frame[0], NULL, mi_row, mi_col, - NULL, NULL); - } - - if (scaled_ref_frame[1]) { - int i; - - for (i = 0; i < MAX_MB_PLANE; i++) - backup_second_yv12[i] = xd->plane[i].pre[1]; - - setup_pre_planes(xd, scaled_ref_frame[1], NULL, mi_row, mi_col, - NULL, NULL); - } - xd->scale_factor[0].set_scaled_offsets(&xd->scale_factor[0], - mi_row, mi_col); - xd->scale_factor[1].set_scaled_offsets(&xd->scale_factor[1], - mi_row, mi_col); - - scaled_first_yv12 = xd->plane[0].pre[0]; - - // Initialize mv using single prediction mode result. - frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int; - frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int; - - // Allow joint search multiple times iteratively for each ref frame - // and break out the search loop if it couldn't find better mv. - for (ite = 0; ite < 4; ite++) { - struct buf_2d ref_yv12[2]; - int bestsme = INT_MAX; - int sadpb = x->sadperbit16; - int_mv tmp_mv; - int search_range = 3; - - int tmp_col_min = x->mv_col_min; - int tmp_col_max = x->mv_col_max; - int tmp_row_min = x->mv_row_min; - int tmp_row_max = x->mv_row_max; - int id = ite % 2; - - // Initialized here because of compiler problem in Visual Studio. - ref_yv12[0] = xd->plane[0].pre[0]; - ref_yv12[1] = xd->plane[0].pre[1]; - - // Get pred block from second frame. - vp9_build_inter_predictor(ref_yv12[!id].buf, - ref_yv12[!id].stride, - second_pred, pw, - &frame_mv[refs[!id]], - &xd->scale_factor[!id], - pw, ph, 0, - &xd->subpix); - - // Compound motion search on first ref frame. - if (id) - xd->plane[0].pre[0] = ref_yv12[id]; - vp9_clamp_mv_min_max(x, &ref_mv[id]); - - // Use mv result from single mode as mvp. - tmp_mv.as_int = frame_mv[refs[id]].as_int; - - tmp_mv.as_mv.col >>= 3; - tmp_mv.as_mv.row >>= 3; - - // Small-range full-pixel motion search - bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, - search_range, - &cpi->fn_ptr[block_size], - x->nmvjointcost, x->mvcost, - &ref_mv[id], second_pred, - pw, ph); - - x->mv_col_min = tmp_col_min; - x->mv_col_max = tmp_col_max; - x->mv_row_min = tmp_row_min; - x->mv_row_max = tmp_row_max; - - if (bestsme < INT_MAX) { - int dis; /* TODO: use dis in distortion calculation later. */ - unsigned int sse; - - bestsme = vp9_find_best_sub_pixel_comp(x, &tmp_mv, - &ref_mv[id], - x->errorperbit, - &cpi->fn_ptr[block_size], - x->nmvjointcost, x->mvcost, - &dis, &sse, second_pred, - pw, ph); - } - - if (id) - xd->plane[0].pre[0] = scaled_first_yv12; - - if (bestsme < last_besterr[id]) { - frame_mv[refs[id]].as_int = - xd->mode_info_context->bmi[0].as_mv[1].as_int = tmp_mv.as_int; - last_besterr[id] = bestsme; - } else { - break; - } - } - - // restore the predictor - if (scaled_ref_frame[0]) { - int i; - - for (i = 0; i < MAX_MB_PLANE; i++) - xd->plane[i].pre[0] = backup_yv12[i]; - } + // Initialize mv using single prediction mode result. + frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int; + frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int; - if (scaled_ref_frame[1]) { - int i; - - for (i = 0; i < MAX_MB_PLANE; i++) - xd->plane[i].pre[1] = backup_second_yv12[i]; - } - - vpx_free(second_pred); - } + if (cpi->sf.comp_inter_joint_search) + iterative_motion_search(cpi, x, bsize, frame_mv, scaled_ref_frame, + mi_row, mi_col, single_newmv); if (frame_mv[refs[0]].as_int == INVALID_MV || frame_mv[refs[1]].as_int == INVALID_MV) @@ -2134,8 +2232,7 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, x->nmvjointcost, x->mvcost, &dis, &sse); } - frame_mv[refs[0]].as_int = - xd->mode_info_context->bmi[0].as_mv[0].as_int = tmp_mv.as_int; + frame_mv[refs[0]].as_int = tmp_mv.as_int; single_newmv[refs[0]].as_int = tmp_mv.as_int; // Add the new motion vector cost to our rolling cost variable @@ -2191,7 +2288,9 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, (mbmi->mv[1].as_mv.col & 15) == 0; // Search for best switchable filter by checking the variance of // pred error irrespective of whether the filter will be used - if (1) { + if (cpi->speed > 4) { + *best_filter = EIGHTTAP; + } else { int i, newbest; int tmp_rate_sum = 0, tmp_dist_sum = 0; for (i = 0; i < VP9_SWITCHABLE_FILTERS; ++i) { @@ -2328,6 +2427,7 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, // Y cost and distortion super_block_yrd(cpi, x, rate_y, distortion_y, &skippable_y, bsize, txfm_cache); + *rate2 += *rate_y; *distortion += *distortion_y; @@ -2393,16 +2493,14 @@ void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, *returnrate = rate4x4_y + rate_uv + vp9_cost_bit(vp9_get_pred_prob(cm, xd, PRED_MBSKIP), 0); *returndist = dist4x4_y + (dist_uv >> 2); - for (i = 0; i < NB_TXFM_MODES; i++) { - ctx->txfm_rd_diff[i] = MIN(err4x4, err - txfm_cache[i]); - } + vpx_memset(ctx->txfm_rd_diff, 0, sizeof(ctx->txfm_rd_diff)); xd->mode_info_context->mbmi.txfm_size = TX_4X4; } else { *returnrate = rate_y + rate_uv + vp9_cost_bit(vp9_get_pred_prob(cm, xd, PRED_MBSKIP), 0); *returndist = dist_y + (dist_uv >> 2); for (i = 0; i < NB_TXFM_MODES; i++) { - ctx->txfm_rd_diff[i] = MIN(err4x4, err - txfm_cache[i]); + ctx->txfm_rd_diff[i] = txfm_cache[i] - txfm_cache[cm->txfm_mode]; } xd->mode_info_context->mbmi.txfm_size = txfm_size; xd->mode_info_context->mbmi.mode = mode; @@ -2457,14 +2555,14 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, int64_t frame_distortions[MAX_REF_FRAMES] = {-1}; int intra_cost_penalty = 20 * vp9_dc_quant(cpi->common.base_qindex, cpi->common.y_dc_delta_q); - int_mv seg_mvs[4][MAX_REF_FRAMES - 1]; + int_mv seg_mvs[4][MAX_REF_FRAMES]; union b_mode_info best_bmodes[4]; PARTITION_INFO best_partition; for (i = 0; i < 4; i++) { int j; - for (j = 0; j < MAX_REF_FRAMES - 1; j++) + for (j = 0; j < MAX_REF_FRAMES; j++) seg_mvs[i][j].as_int = INVALID_MV; } // Everywhere the flag is set the error is much higher than its neighbors. @@ -2563,11 +2661,6 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, this_mode = vp9_mode_order[mode_index].mode; ref_frame = vp9_mode_order[mode_index].ref_frame; - if (!(ref_frame == INTRA_FRAME - || (cpi->ref_frame_flags & flag_list[ref_frame]))) { - continue; - } - if (cpi->speed > 0 && bsize >= BLOCK_SIZE_SB8X8) { if (!(ref_frame_mask & (1 << ref_frame))) { continue; @@ -2585,6 +2678,15 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, mbmi->ref_frame = ref_frame; mbmi->second_ref_frame = vp9_mode_order[mode_index].second_ref_frame; + if (!(ref_frame == INTRA_FRAME + || (cpi->ref_frame_flags & flag_list[ref_frame]))) { + continue; + } + if (!(mbmi->second_ref_frame == NONE + || (cpi->ref_frame_flags & flag_list[mbmi->second_ref_frame]))) { + continue; + } + // TODO(jingning, jkoleszar): scaling reference frame not supported for // SPLITMV. if (mbmi->ref_frame > 0 && @@ -2680,8 +2782,6 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, if (this_mode == I4X4_PRED) { int rate; - // Note the rate value returned here includes the cost of coding - // the I4X4_PRED mode : x->mbmode_cost[xd->frame_type][I4X4_PRED]; mbmi->txfm_size = TX_4X4; rd_pick_intra4x4mby_modes(cpi, x, &rate, &rate_y, &distortion_y, INT64_MAX); @@ -2716,7 +2816,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, skippable = skippable && skip_uv[uv_tx]; mbmi->uv_mode = mode_uv[uv_tx]; - rate2 = rate_y + x->mbmode_cost[cm->frame_type][mbmi->mode] + rate_uv; + rate2 = rate_y + x->mbmode_cost[mbmi->mode] + rate_uv; if (mbmi->mode != DC_PRED && mbmi->mode != TM_PRED) rate2 += intra_cost_penalty; distortion2 = distortion_y + distortion_uv; @@ -2755,7 +2855,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, second_ref, INT64_MAX, &rate, &rate_y, &distortion, &skippable, - (int)this_rd_thresh, seg_mvs); + (int)this_rd_thresh, seg_mvs, + mi_row, mi_col); if (cpi->common.mcomp_filter_type == SWITCHABLE) { const int rs = get_switchable_rate(cm, x); tmp_rd += RDCOST(x->rdmult, x->rddiv, rs, 0); @@ -2794,7 +2895,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, second_ref, INT64_MAX, &rate, &rate_y, &distortion, &skippable, - (int)this_rd_thresh, seg_mvs); + (int)this_rd_thresh, seg_mvs, + mi_row, mi_col); } else { if (cpi->common.mcomp_filter_type == SWITCHABLE) { int rs = get_switchable_rate(cm, x); @@ -2843,7 +2945,6 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, compmode_cost = vp9_cost_bit(vp9_get_pred_prob(cm, xd, PRED_COMP), is_comp_pred); - mbmi->mode = this_mode; } else { YV12_BUFFER_CONFIG *scaled_ref_frame[2] = {NULL, NULL}; int fb = get_ref_frame_idx(cpi, mbmi->ref_frame); @@ -2938,14 +3039,16 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, best_mode = this_mode; } - // Store the respective mode distortions for later use. - if (mode_distortions[this_mode] == -1 - || distortion2 < mode_distortions[this_mode]) { - mode_distortions[this_mode] = distortion2; - } - if (frame_distortions[mbmi->ref_frame] == -1 - || distortion2 < frame_distortions[mbmi->ref_frame]) { - frame_distortions[mbmi->ref_frame] = distortion2; + if (this_mode != I4X4_PRED && this_mode != SPLITMV) { + // Store the respective mode distortions for later use. + if (mode_distortions[this_mode] == -1 + || distortion2 < mode_distortions[this_mode]) { + mode_distortions[this_mode] = distortion2; + } + if (frame_distortions[mbmi->ref_frame] == -1 + || distortion2 < frame_distortions[mbmi->ref_frame]) { + frame_distortions[mbmi->ref_frame] = distortion2; + } } // Did this mode help.. i.e. is it the new best mode @@ -2954,7 +3057,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, // Note index of best mode so far best_mode_index = mode_index; - if (this_mode <= I4X4_PRED) { + if (ref_frame == INTRA_FRAME) { /* required for left and above block mv */ mbmi->mv[0].as_int = 0; } @@ -3052,8 +3155,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, // Flag all modes that have a distortion thats > 2x the best we found at // this level. for (mode_index = 0; mode_index < MB_MODE_COUNT; ++mode_index) { - if (mode_index == NEARESTMV || mode_index == NEARMV || mode_index == NEWMV - || mode_index == SPLITMV) + if (mode_index == NEARESTMV || mode_index == NEARMV || mode_index == NEWMV) continue; if (mode_distortions[mode_index] > 2 * *returndistortion) { @@ -3077,7 +3179,7 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, assert((cm->mcomp_filter_type == SWITCHABLE) || (cm->mcomp_filter_type == best_mbmode.interp_filter) || - (best_mbmode.mode <= I4X4_PRED)); + (best_mbmode.ref_frame == INTRA_FRAME)); // Accumulate filter usage stats // TODO(agrange): Use RD criteria to select interpolation filter mode. @@ -3129,13 +3231,15 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, // macroblock modes vpx_memcpy(mbmi, &best_mbmode, sizeof(MB_MODE_INFO)); - if (best_mbmode.mode == I4X4_PRED) { + if (best_mbmode.ref_frame == INTRA_FRAME && + best_mbmode.sb_type < BLOCK_SIZE_SB8X8) { for (i = 0; i < 4; i++) { xd->mode_info_context->bmi[i].as_mode = best_bmodes[i].as_mode; } } - if (best_mbmode.mode == SPLITMV) { + if (best_mbmode.ref_frame != INTRA_FRAME && + best_mbmode.sb_type < BLOCK_SIZE_SB8X8) { for (i = 0; i < 4; i++) xd->mode_info_context->bmi[i].as_mv[0].as_int = best_bmodes[i].as_mv[0].as_int; diff --git a/vp9/encoder/vp9_tokenize.c b/vp9/encoder/vp9_tokenize.c index 08efc84d4..79f72bb4b 100644 --- a/vp9/encoder/vp9_tokenize.c +++ b/vp9/encoder/vp9_tokenize.c @@ -36,6 +36,9 @@ extern vp9_coeff_stats tree_update_hist_16x16[BLOCK_TYPES]; extern vp9_coeff_stats tree_update_hist_32x32[BLOCK_TYPES]; #endif /* ENTROPY_STATS */ +DECLARE_ALIGNED(16, extern const uint8_t, + vp9_pt_energy_class[MAX_ENTROPY_TOKENS]); + static TOKENVALUE dct_value_tokens[DCT_MAX_VALUE * 2]; const TOKENVALUE *vp9_dct_value_tokens_ptr; static int dct_value_cost[DCT_MAX_VALUE * 2]; @@ -221,14 +224,24 @@ static void tokenize_b(int plane, int block, BLOCK_SIZE_TYPE bsize, t->token = token; t->context_tree = coef_probs[type][ref][band][pt]; t->skip_eob_node = (c > 0) && (token_cache[scan[c - 1]] == 0); + +#if CONFIG_BALANCED_COEFTREE + assert(token <= ZERO_TOKEN || + vp9_coef_encodings[t->token].len - t->skip_eob_node > 0); +#else assert(vp9_coef_encodings[t->token].len - t->skip_eob_node > 0); +#endif if (!dry_run) { ++counts[type][ref][band][pt][token]; +#if CONFIG_BALANCED_COEFTREE + if (!t->skip_eob_node && token > ZERO_TOKEN) +#else if (!t->skip_eob_node) +#endif ++cpi->common.fc.eob_branch_counts[tx_size][type][ref][band][pt]; } - token_cache[scan[c]] = token; + token_cache[scan[c]] = vp9_pt_energy_class[token]; ++t; } while (c < eob && ++c < seg_eob); diff --git a/vp9/encoder/vp9_variance_c.c b/vp9/encoder/vp9_variance_c.c index 7231dcf22..e9b36f356 100644 --- a/vp9/encoder/vp9_variance_c.c +++ b/vp9/encoder/vp9_variance_c.c @@ -318,6 +318,11 @@ unsigned int vp9_variance16x8_c(const uint8_t *src_ptr, return (var - (((unsigned int)avg * avg) >> 7)); } +void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, + const uint8_t *ref_ptr, int ref_stride, + unsigned int *sse, int *sum) { + variance(src_ptr, source_stride, ref_ptr, ref_stride, 8, 8, sse, sum); +} unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, diff --git a/vp9/encoder/vp9_write_bit_buffer.h b/vp9/encoder/vp9_write_bit_buffer.h new file mode 100644 index 000000000..18cf40366 --- /dev/null +++ b/vp9/encoder/vp9_write_bit_buffer.h @@ -0,0 +1,42 @@ +/* + * Copyright (c) 2013 The WebM project authors. All Rights Reserved. + * + * Use of this source code is governed by a BSD-style license + * that can be found in the LICENSE file in the root of the source + * tree. An additional intellectual property rights grant can be found + * in the file PATENTS. All contributing project authors may + * be found in the AUTHORS file in the root of the source tree. + */ + +#ifndef VP9_BIT_WRITE_BUFFER_H_ +#define VP9_BIT_WRITE_BUFFER_H_ + +#include "vpx/vpx_integer.h" + +struct vp9_write_bit_buffer { + uint8_t *bit_buffer; + size_t bit_offset; +}; + +static size_t vp9_rb_bytes_written(struct vp9_write_bit_buffer *wb) { + return wb->bit_offset / CHAR_BIT + (wb->bit_offset % CHAR_BIT > 0); +} + +static void vp9_wb_write_bit(struct vp9_write_bit_buffer *wb, int bit) { + const int off = wb->bit_offset; + const int p = off / CHAR_BIT; + const int q = CHAR_BIT - 1 - off % CHAR_BIT; + wb->bit_buffer[p] &= ~(1 << q); + wb->bit_buffer[p] |= bit << q; + wb->bit_offset = off + 1; +} + +static void vp9_wb_write_literal(struct vp9_write_bit_buffer *wb, + int data, int bits) { + int bit; + for (bit = bits - 1; bit >= 0; bit--) + vp9_wb_write_bit(wb, (data >> bit) & 1); +} + + +#endif // VP9_BIT_WRITE_BUFFER_H_ |