diff options
Diffstat (limited to 'media/libaom/src/av1/encoder/encodeframe.c')
| -rw-r--r-- | media/libaom/src/av1/encoder/encodeframe.c | 7310 |
1 files changed, 4023 insertions, 3287 deletions
diff --git a/media/libaom/src/av1/encoder/encodeframe.c b/media/libaom/src/av1/encoder/encodeframe.c index cb226c59e2..53b47d49e9 100644 --- a/media/libaom/src/av1/encoder/encodeframe.c +++ b/media/libaom/src/av1/encoder/encodeframe.c @@ -10,7 +10,9 @@ */ #include <limits.h> +#include <float.h> #include <math.h> +#include <stdbool.h> #include <stdio.h> #include "config/aom_config.h" @@ -45,6 +47,7 @@ #include "av1/encoder/aq_complexity.h" #include "av1/encoder/aq_cyclicrefresh.h" #include "av1/encoder/aq_variance.h" +#include "av1/encoder/corner_detect.h" #include "av1/encoder/global_motion.h" #include "av1/encoder/encodeframe.h" #include "av1/encoder/encodemb.h" @@ -53,23 +56,33 @@ #include "av1/encoder/ethread.h" #include "av1/encoder/extend.h" #include "av1/encoder/ml.h" +#include "av1/encoder/motion_search_facade.h" +#include "av1/encoder/partition_strategy.h" +#if !CONFIG_REALTIME_ONLY #include "av1/encoder/partition_model_weights.h" +#endif #include "av1/encoder/rd.h" #include "av1/encoder/rdopt.h" #include "av1/encoder/reconinter_enc.h" #include "av1/encoder/segmentation.h" #include "av1/encoder/tokenize.h" +#include "av1/encoder/tpl_model.h" +#include "av1/encoder/var_based_part.h" -static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, - ThreadData *td, TOKENEXTRA **t, RUN_TYPE dry_run, - int mi_row, int mi_col, BLOCK_SIZE bsize, - int *rate); +#if CONFIG_TUNE_VMAF +#include "av1/encoder/tune_vmaf.h" +#endif + +static AOM_INLINE void encode_superblock(const AV1_COMP *const cpi, + TileDataEnc *tile_data, ThreadData *td, + TOKENEXTRA **t, RUN_TYPE dry_run, + BLOCK_SIZE bsize, int *rate); // This is used as a reference when computing the source variance for the // purposes of activity masking. // Eventually this should be replaced by custom no-reference routines, // which will be faster. -static const uint8_t AV1_VAR_OFFS[MAX_SB_SIZE] = { +const uint8_t AV1_VAR_OFFS[MAX_SB_SIZE] = { 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, @@ -134,14 +147,42 @@ static const uint16_t AV1_HIGH_VAR_OFFS_12[MAX_SB_SIZE] = { 128 * 16, 128 * 16 }; -#if CONFIG_FP_MB_STATS -static const uint8_t num_16x16_blocks_wide_lookup[BLOCK_SIZES_ALL] = { - 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 4, 8, 8, 1, 1, 1, 2, 2, 4 -}; -static const uint8_t num_16x16_blocks_high_lookup[BLOCK_SIZES_ALL] = { - 1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4, 8, 4, 8, 1, 1, 2, 1, 4, 2 -}; -#endif // CONFIG_FP_MB_STATS +typedef struct { + ENTROPY_CONTEXT a[MAX_MIB_SIZE * MAX_MB_PLANE]; + ENTROPY_CONTEXT l[MAX_MIB_SIZE * MAX_MB_PLANE]; + PARTITION_CONTEXT sa[MAX_MIB_SIZE]; + PARTITION_CONTEXT sl[MAX_MIB_SIZE]; + TXFM_CONTEXT *p_ta; + TXFM_CONTEXT *p_tl; + TXFM_CONTEXT ta[MAX_MIB_SIZE]; + TXFM_CONTEXT tl[MAX_MIB_SIZE]; +} RD_SEARCH_MACROBLOCK_CONTEXT; + +enum { PICK_MODE_RD = 0, PICK_MODE_NONRD }; + +enum { + SB_SINGLE_PASS, // Single pass encoding: all ctxs get updated normally + SB_DRY_PASS, // First pass of multi-pass: does not update the ctxs + SB_WET_PASS // Second pass of multi-pass: finalize and update the ctx +} UENUM1BYTE(SB_MULTI_PASS_MODE); + +// This struct is used to store the statistics used by sb-level multi-pass +// encoding. Currently, this is only used to make a copy of the state before we +// perform the first pass +typedef struct SB_FIRST_PASS_STATS { + RD_SEARCH_MACROBLOCK_CONTEXT x_ctx; + RD_COUNTS rd_count; + + int split_count; + FRAME_COUNTS fc; + InterModeRdModel inter_mode_rd_models[BLOCK_SIZES_ALL]; + int thresh_freq_fact[BLOCK_SIZES_ALL][MAX_MODES]; + int current_qindex; + +#if CONFIG_INTERNAL_STATS + unsigned int mode_chosen_counts[MAX_MODES]; +#endif // CONFIG_INTERNAL_STATS +} SB_FIRST_PASS_STATS; unsigned int av1_get_sby_perpixel_variance(const AV1_COMP *cpi, const struct buf_2d *ref, @@ -156,24 +197,14 @@ unsigned int av1_high_get_sby_perpixel_variance(const AV1_COMP *cpi, const struct buf_2d *ref, BLOCK_SIZE bs, int bd) { unsigned int var, sse; - switch (bd) { - case 10: - var = - cpi->fn_ptr[bs].vf(ref->buf, ref->stride, - CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_10), 0, &sse); - break; - case 12: - var = - cpi->fn_ptr[bs].vf(ref->buf, ref->stride, - CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_12), 0, &sse); - break; - case 8: - default: - var = - cpi->fn_ptr[bs].vf(ref->buf, ref->stride, - CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_8), 0, &sse); - break; - } + assert(bd == 8 || bd == 10 || bd == 12); + const int off_index = (bd - 8) >> 1; + const uint16_t *high_var_offs[3] = { AV1_HIGH_VAR_OFFS_8, + AV1_HIGH_VAR_OFFS_10, + AV1_HIGH_VAR_OFFS_12 }; + var = + cpi->fn_ptr[bs].vf(ref->buf, ref->stride, + CONVERT_TO_BYTEPTR(high_var_offs[off_index]), 0, &sse); return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]); } @@ -183,7 +214,8 @@ static unsigned int get_sby_perpixel_diff_variance(const AV1_COMP *const cpi, BLOCK_SIZE bs) { unsigned int sse, var; uint8_t *last_y; - const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME); + const YV12_BUFFER_CONFIG *last = + get_ref_frame_yv12_buf(&cpi->common, LAST_FRAME); assert(last != NULL); last_y = @@ -206,120 +238,244 @@ static BLOCK_SIZE get_rd_var_based_fixed_partition(AV1_COMP *cpi, MACROBLOCK *x, return BLOCK_8X8; } -// Lighter version of set_offsets that only sets the mode info -// pointers. -static void set_mode_info_offsets(const AV1_COMP *const cpi, - MACROBLOCK *const x, MACROBLOCKD *const xd, - int mi_row, int mi_col) { +static int set_deltaq_rdmult(const AV1_COMP *const cpi, MACROBLOCKD *const xd) { + const AV1_COMMON *const cm = &cpi->common; + const CommonQuantParams *quant_params = &cm->quant_params; + return av1_compute_rd_mult(cpi, quant_params->base_qindex + xd->delta_qindex + + quant_params->y_dc_delta_q); +} + +static AOM_INLINE void set_ssim_rdmult(const AV1_COMP *const cpi, + MACROBLOCK *const x, + const BLOCK_SIZE bsize, const int mi_row, + const int mi_col, int *const rdmult) { + const AV1_COMMON *const cm = &cpi->common; + + const int bsize_base = BLOCK_16X16; + const int num_mi_w = mi_size_wide[bsize_base]; + const int num_mi_h = mi_size_high[bsize_base]; + const int num_cols = (cm->mi_params.mi_cols + num_mi_w - 1) / num_mi_w; + const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h; + const int num_bcols = (mi_size_wide[bsize] + num_mi_w - 1) / num_mi_w; + const int num_brows = (mi_size_high[bsize] + num_mi_h - 1) / num_mi_h; + int row, col; + double num_of_mi = 0.0; + double geom_mean_of_scale = 0.0; + + assert(cpi->oxcf.tuning == AOM_TUNE_SSIM); + + aom_clear_system_state(); + for (row = mi_row / num_mi_w; + row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) { + for (col = mi_col / num_mi_h; + col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) { + const int index = row * num_cols + col; + geom_mean_of_scale += log(cpi->ssim_rdmult_scaling_factors[index]); + num_of_mi += 1.0; + } + } + geom_mean_of_scale = exp(geom_mean_of_scale / num_of_mi); + + *rdmult = (int)((double)(*rdmult) * geom_mean_of_scale + 0.5); + *rdmult = AOMMAX(*rdmult, 0); + set_error_per_bit(x, *rdmult); + aom_clear_system_state(); +} + +static int get_hier_tpl_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x, + const BLOCK_SIZE bsize, const int mi_row, + const int mi_col, int orig_rdmult) { + const AV1_COMMON *const cm = &cpi->common; + assert(IMPLIES(cpi->gf_group.size > 0, + cpi->gf_group.index < cpi->gf_group.size)); + const int tpl_idx = cpi->gf_group.index; + const TplDepFrame *tpl_frame = &cpi->tpl_data.tpl_frame[tpl_idx]; + MACROBLOCKD *const xd = &x->e_mbd; + const int deltaq_rdmult = set_deltaq_rdmult(cpi, xd); + if (tpl_frame->is_valid == 0) return deltaq_rdmult; + if (!is_frame_tpl_eligible((AV1_COMP *)cpi)) return deltaq_rdmult; + if (tpl_idx >= MAX_LAG_BUFFERS) return deltaq_rdmult; + if (cpi->superres_mode != SUPERRES_NONE) return deltaq_rdmult; + if (cpi->oxcf.aq_mode != NO_AQ) return deltaq_rdmult; + + const int bsize_base = BLOCK_16X16; + const int num_mi_w = mi_size_wide[bsize_base]; + const int num_mi_h = mi_size_high[bsize_base]; + const int num_cols = (cm->mi_params.mi_cols + num_mi_w - 1) / num_mi_w; + const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h; + const int num_bcols = (mi_size_wide[bsize] + num_mi_w - 1) / num_mi_w; + const int num_brows = (mi_size_high[bsize] + num_mi_h - 1) / num_mi_h; + int row, col; + double base_block_count = 0.0; + double geom_mean_of_scale = 0.0; + aom_clear_system_state(); + for (row = mi_row / num_mi_w; + row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) { + for (col = mi_col / num_mi_h; + col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) { + const int index = row * num_cols + col; + geom_mean_of_scale += log(cpi->tpl_sb_rdmult_scaling_factors[index]); + base_block_count += 1.0; + } + } + geom_mean_of_scale = exp(geom_mean_of_scale / base_block_count); + int rdmult = (int)((double)orig_rdmult * geom_mean_of_scale + 0.5); + rdmult = AOMMAX(rdmult, 0); + set_error_per_bit(x, rdmult); + aom_clear_system_state(); + if (bsize == cm->seq_params.sb_size) { + const int rdmult_sb = set_deltaq_rdmult(cpi, xd); + assert(rdmult_sb == rdmult); + (void)rdmult_sb; + } + return rdmult; +} + +static int set_segment_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x, + int8_t segment_id) { const AV1_COMMON *const cm = &cpi->common; - const int idx_str = xd->mi_stride * mi_row + mi_col; - xd->mi = cm->mi_grid_visible + idx_str; - xd->mi[0] = cm->mi + idx_str; - x->mbmi_ext = cpi->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col); + av1_init_plane_quantizers(cpi, x, segment_id); + aom_clear_system_state(); + const int segment_qindex = + av1_get_qindex(&cm->seg, segment_id, cm->quant_params.base_qindex); + return av1_compute_rd_mult(cpi, + segment_qindex + cm->quant_params.y_dc_delta_q); } -static void set_offsets_without_segment_id(const AV1_COMP *const cpi, - const TileInfo *const tile, - MACROBLOCK *const x, int mi_row, - int mi_col, BLOCK_SIZE bsize) { +static AOM_INLINE void setup_block_rdmult(const AV1_COMP *const cpi, + MACROBLOCK *const x, int mi_row, + int mi_col, BLOCK_SIZE bsize, + AQ_MODE aq_mode, MB_MODE_INFO *mbmi) { + x->rdmult = cpi->rd.RDMULT; + + if (aq_mode != NO_AQ) { + assert(mbmi != NULL); + if (aq_mode == VARIANCE_AQ) { + if (cpi->vaq_refresh) { + const int energy = bsize <= BLOCK_16X16 + ? x->mb_energy + : av1_log_block_var(cpi, x, bsize); + mbmi->segment_id = energy; + } + x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id); + } else if (aq_mode == COMPLEXITY_AQ) { + x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id); + } else if (aq_mode == CYCLIC_REFRESH_AQ) { + // If segment is boosted, use rdmult for that segment. + if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) + x->rdmult = av1_cyclic_refresh_get_rdmult(cpi->cyclic_refresh); + } + } + + const AV1_COMMON *const cm = &cpi->common; + if (cm->delta_q_info.delta_q_present_flag && + !cpi->sf.rt_sf.use_nonrd_pick_mode) { + x->rdmult = get_hier_tpl_rdmult(cpi, x, bsize, mi_row, mi_col, x->rdmult); + } + + if (cpi->oxcf.tuning == AOM_TUNE_SSIM) { + set_ssim_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult); + } +#if CONFIG_TUNE_VMAF + if (cpi->oxcf.tuning == AOM_TUNE_VMAF_WITHOUT_PREPROCESSING || + cpi->oxcf.tuning == AOM_TUNE_VMAF_MAX_GAIN) { + av1_set_vmaf_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult); + } +#endif +} + +static AOM_INLINE void set_offsets_without_segment_id( + const AV1_COMP *const cpi, const TileInfo *const tile, MACROBLOCK *const x, + int mi_row, int mi_col, BLOCK_SIZE bsize) { const AV1_COMMON *const cm = &cpi->common; const int num_planes = av1_num_planes(cm); MACROBLOCKD *const xd = &x->e_mbd; + assert(bsize < BLOCK_SIZES_ALL); const int mi_width = mi_size_wide[bsize]; const int mi_height = mi_size_high[bsize]; - set_mode_info_offsets(cpi, x, xd, mi_row, mi_col); + set_mode_info_offsets(&cpi->common.mi_params, &cpi->mbmi_ext_info, x, xd, + mi_row, mi_col); - set_skip_context(xd, mi_row, mi_col, num_planes); - xd->above_txfm_context = cm->above_txfm_context[tile->tile_row] + mi_col; + set_entropy_context(xd, mi_row, mi_col, num_planes); + xd->above_txfm_context = cm->above_contexts.txfm[tile->tile_row] + mi_col; xd->left_txfm_context = xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); // Set up destination pointers. - av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row, - mi_col, 0, num_planes); + av1_setup_dst_planes(xd->plane, bsize, &cm->cur_frame->buf, mi_row, mi_col, 0, + num_planes); // Set up limit values for MV components. // Mv beyond the range do not produce new/different prediction block. - x->mv_limits.row_min = - -(((mi_row + mi_height) * MI_SIZE) + AOM_INTERP_EXTEND); - x->mv_limits.col_min = -(((mi_col + mi_width) * MI_SIZE) + AOM_INTERP_EXTEND); - x->mv_limits.row_max = (cm->mi_rows - mi_row) * MI_SIZE + AOM_INTERP_EXTEND; - x->mv_limits.col_max = (cm->mi_cols - mi_col) * MI_SIZE + AOM_INTERP_EXTEND; + av1_set_mv_limits(&cm->mi_params, &x->mv_limits, mi_row, mi_col, mi_height, + mi_width, cpi->oxcf.border_in_pixels); set_plane_n4(xd, mi_width, mi_height, num_planes); // Set up distance of MB to edge of frame in 1/8th pel units. assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1))); - set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, cm->mi_rows, - cm->mi_cols); + set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, + cm->mi_params.mi_rows, cm->mi_params.mi_cols); // Set up source buffers. - av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes); - - // R/D setup. - x->rdmult = cpi->rd.RDMULT; + av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes, bsize); // required by av1_append_sub8x8_mvs_for_idx() and av1_find_best_ref_mvs() xd->tile = *tile; } -static void set_offsets(const AV1_COMP *const cpi, const TileInfo *const tile, - MACROBLOCK *const x, int mi_row, int mi_col, - BLOCK_SIZE bsize) { +static AOM_INLINE void set_offsets(const AV1_COMP *const cpi, + const TileInfo *const tile, + MACROBLOCK *const x, int mi_row, int mi_col, + BLOCK_SIZE bsize) { const AV1_COMMON *const cm = &cpi->common; + const struct segmentation *const seg = &cm->seg; MACROBLOCKD *const xd = &x->e_mbd; MB_MODE_INFO *mbmi; - const struct segmentation *const seg = &cm->seg; set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize); + // Setup segment ID. mbmi = xd->mi[0]; - xd->cfl.mi_row = mi_row; - xd->cfl.mi_col = mi_col; - mbmi->segment_id = 0; - - // Setup segment ID. if (seg->enabled) { if (seg->enabled && !cpi->vaq_refresh) { const uint8_t *const map = - seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map; + seg->update_map ? cpi->enc_seg.map : cm->last_frame_seg_map; mbmi->segment_id = - map ? get_segment_id(cm, map, bsize, mi_row, mi_col) : 0; + map ? get_segment_id(&cm->mi_params, map, bsize, mi_row, mi_col) : 0; } av1_init_plane_quantizers(cpi, x, mbmi->segment_id); } } -static void reset_intmv_filter_type(MB_MODE_INFO *mbmi) { - InterpFilter filters[2]; - - for (int dir = 0; dir < 2; ++dir) { - filters[dir] = av1_extract_interp_filter(mbmi->interp_filters, dir); +static AOM_INLINE void update_filter_type_count(FRAME_COUNTS *counts, + const MACROBLOCKD *xd, + const MB_MODE_INFO *mbmi) { + int dir; + for (dir = 0; dir < 2; ++dir) { + const int ctx = av1_get_pred_context_switchable_interp(xd, dir); + InterpFilter filter = av1_extract_interp_filter(mbmi->interp_filters, dir); + ++counts->switchable_interp[ctx][filter]; } - mbmi->interp_filters = av1_make_interp_filters(filters[0], filters[1]); } -static void update_filter_type_count(uint8_t allow_update_cdf, - FRAME_COUNTS *counts, - const MACROBLOCKD *xd, - const MB_MODE_INFO *mbmi) { +static AOM_INLINE void update_filter_type_cdf(const MACROBLOCKD *xd, + const MB_MODE_INFO *mbmi) { int dir; for (dir = 0; dir < 2; ++dir) { const int ctx = av1_get_pred_context_switchable_interp(xd, dir); InterpFilter filter = av1_extract_interp_filter(mbmi->interp_filters, dir); - ++counts->switchable_interp[ctx][filter]; - if (allow_update_cdf) { - update_cdf(xd->tile_ctx->switchable_interp_cdf[ctx], filter, - SWITCHABLE_FILTERS); - } + update_cdf(xd->tile_ctx->switchable_interp_cdf[ctx], filter, + SWITCHABLE_FILTERS); } } -static void update_global_motion_used(PREDICTION_MODE mode, BLOCK_SIZE bsize, - const MB_MODE_INFO *mbmi, - RD_COUNTS *rdc) { +static AOM_INLINE void update_global_motion_used(PREDICTION_MODE mode, + BLOCK_SIZE bsize, + const MB_MODE_INFO *mbmi, + RD_COUNTS *rdc) { if (mode == GLOBALMV || mode == GLOBAL_GLOBALMV) { const int num_4x4s = mi_size_wide[bsize] * mi_size_high[bsize]; int ref; @@ -329,8 +485,8 @@ static void update_global_motion_used(PREDICTION_MODE mode, BLOCK_SIZE bsize, } } -static void reset_tx_size(MACROBLOCK *x, MB_MODE_INFO *mbmi, - const TX_MODE tx_mode) { +static AOM_INLINE void reset_tx_size(MACROBLOCK *x, MB_MODE_INFO *mbmi, + const TX_MODE tx_mode) { MACROBLOCKD *const xd = &x->e_mbd; if (xd->lossless[mbmi->segment_id]) { mbmi->tx_size = TX_4X4; @@ -344,17 +500,38 @@ static void reset_tx_size(MACROBLOCK *x, MB_MODE_INFO *mbmi, if (is_inter_block(mbmi)) { memset(mbmi->inter_tx_size, mbmi->tx_size, sizeof(mbmi->inter_tx_size)); } - memset(mbmi->txk_type, DCT_DCT, sizeof(mbmi->txk_type[0]) * TXK_TYPE_BUF_LEN); + const int stride = xd->tx_type_map_stride; + const int bw = mi_size_wide[mbmi->sb_type]; + for (int row = 0; row < mi_size_high[mbmi->sb_type]; ++row) { + memset(xd->tx_type_map + row * stride, DCT_DCT, + bw * sizeof(xd->tx_type_map[0])); + } av1_zero(x->blk_skip); - x->skip = 0; + x->force_skip = 0; } -static void update_state(const AV1_COMP *const cpi, - const TileDataEnc *const tile_data, ThreadData *td, - const PICK_MODE_CONTEXT *const ctx, int mi_row, - int mi_col, BLOCK_SIZE bsize, RUN_TYPE dry_run) { +// This function will copy the best reference mode information from +// MB_MODE_INFO_EXT_FRAME to MB_MODE_INFO_EXT. +static INLINE void copy_mbmi_ext_frame_to_mbmi_ext( + MB_MODE_INFO_EXT *mbmi_ext, + const MB_MODE_INFO_EXT_FRAME *const mbmi_ext_best, uint8_t ref_frame_type) { + memcpy(mbmi_ext->ref_mv_stack[ref_frame_type], mbmi_ext_best->ref_mv_stack, + sizeof(mbmi_ext->ref_mv_stack[USABLE_REF_MV_STACK_SIZE])); + memcpy(mbmi_ext->weight[ref_frame_type], mbmi_ext_best->weight, + sizeof(mbmi_ext->weight[USABLE_REF_MV_STACK_SIZE])); + mbmi_ext->mode_context[ref_frame_type] = mbmi_ext_best->mode_context; + mbmi_ext->ref_mv_count[ref_frame_type] = mbmi_ext_best->ref_mv_count; + memcpy(mbmi_ext->global_mvs, mbmi_ext_best->global_mvs, + sizeof(mbmi_ext->global_mvs)); +} + +static AOM_INLINE void update_state(const AV1_COMP *const cpi, ThreadData *td, + const PICK_MODE_CONTEXT *const ctx, + int mi_row, int mi_col, BLOCK_SIZE bsize, + RUN_TYPE dry_run) { int i, x_idx, y; const AV1_COMMON *const cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; const int num_planes = av1_num_planes(cm); RD_COUNTS *const rdc = &td->rd_counts; MACROBLOCK *const x = &td->mb; @@ -366,37 +543,52 @@ static void update_state(const AV1_COMP *const cpi, const struct segmentation *const seg = &cm->seg; const int bw = mi_size_wide[mi->sb_type]; const int bh = mi_size_high[mi->sb_type]; - const int mis = cm->mi_stride; + const int mis = mi_params->mi_stride; const int mi_width = mi_size_wide[bsize]; const int mi_height = mi_size_high[bsize]; assert(mi->sb_type == bsize); *mi_addr = *mi; - *x->mbmi_ext = ctx->mbmi_ext; - - reset_intmv_filter_type(mi_addr); + copy_mbmi_ext_frame_to_mbmi_ext(x->mbmi_ext, &ctx->mbmi_ext_best, + av1_ref_frame_type(ctx->mic.ref_frame)); memcpy(x->blk_skip, ctx->blk_skip, sizeof(x->blk_skip[0]) * ctx->num_4x4_blk); - x->skip = ctx->skip; + x->force_skip = ctx->rd_stats.skip; + + xd->tx_type_map = ctx->tx_type_map; + xd->tx_type_map_stride = mi_size_wide[bsize]; + // If not dry_run, copy the transform type data into the frame level buffer. + // Encoder will fetch tx types when writing bitstream. + if (!dry_run) { + const int grid_idx = get_mi_grid_idx(mi_params, mi_row, mi_col); + uint8_t *const tx_type_map = mi_params->tx_type_map + grid_idx; + const int mi_stride = mi_params->mi_stride; + for (int blk_row = 0; blk_row < bh; ++blk_row) { + av1_copy_array(tx_type_map + blk_row * mi_stride, + xd->tx_type_map + blk_row * xd->tx_type_map_stride, bw); + } + xd->tx_type_map = tx_type_map; + xd->tx_type_map_stride = mi_stride; + } // If segmentation in use if (seg->enabled) { // For in frame complexity AQ copy the segment id from the segment map. if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) { const uint8_t *const map = - seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map; + seg->update_map ? cpi->enc_seg.map : cm->last_frame_seg_map; mi_addr->segment_id = - map ? get_segment_id(cm, map, bsize, mi_row, mi_col) : 0; - reset_tx_size(x, mi_addr, cm->tx_mode); + map ? get_segment_id(mi_params, map, bsize, mi_row, mi_col) : 0; + reset_tx_size(x, mi_addr, x->tx_mode_search_type); } // Else for cyclic refresh mode update the segment map, set the segment id // and then update the quantizer. if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) { av1_cyclic_refresh_update_segment(cpi, mi_addr, mi_row, mi_col, bsize, - ctx->rate, ctx->dist, x->skip); - reset_tx_size(x, mi_addr, cm->tx_mode); + ctx->rd_stats.rate, ctx->rd_stats.dist, + x->force_skip); } if (mi_addr->uv_mode == UV_CFL_PRED && !is_cfl_allowed(xd)) mi_addr->uv_mode = UV_DC_PRED; @@ -412,12 +604,14 @@ static void update_state(const AV1_COMP *const cpi, for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i]; // Restore the coding context of the MB to that that was in place // when the mode was picked for it - for (y = 0; y < mi_height; y++) - for (x_idx = 0; x_idx < mi_width; x_idx++) + for (y = 0; y < mi_height; y++) { + for (x_idx = 0; x_idx < mi_width; x_idx++) { if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) { xd->mi[x_idx + y * mis] = mi_addr; } + } + } if (cpi->oxcf.aq_mode) av1_init_plane_quantizers(cpi, x, mi_addr->segment_id); @@ -457,11 +651,10 @@ static void update_state(const AV1_COMP *const cpi, update_global_motion_used(mi_addr->mode, bsize, mi_addr, rdc); } - if (cm->interp_filter == SWITCHABLE && + if (cm->features.interp_filter == SWITCHABLE && mi_addr->motion_mode != WARPED_CAUSAL && !is_nontrans_global_motion(xd, xd->mi[0])) { - update_filter_type_count(tile_data->allow_update_cdf, td->counts, xd, - mi_addr); + update_filter_type_count(td->counts, xd, mi_addr); } rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff; @@ -469,13 +662,15 @@ static void update_state(const AV1_COMP *const cpi, rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff; } - const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col); - const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row); - av1_copy_frame_mvs(cm, mi, mi_row, mi_col, x_mis, y_mis); + const int x_mis = AOMMIN(bw, mi_params->mi_cols - mi_col); + const int y_mis = AOMMIN(bh, mi_params->mi_rows - mi_row); + if (cm->seq_params.order_hint_info.enable_ref_frame_mvs) + av1_copy_frame_mvs(cm, mi, mi_row, mi_col, x_mis, y_mis); } void av1_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src, - int mi_row, int mi_col, const int num_planes) { + int mi_row, int mi_col, const int num_planes, + BLOCK_SIZE bsize) { // Set current frame pointer. x->e_mbd.cur_buf = src; @@ -483,77 +678,98 @@ void av1_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src, // the static analysis warnings. for (int i = 0; i < AOMMIN(num_planes, MAX_MB_PLANE); i++) { const int is_uv = i > 0; - setup_pred_plane(&x->plane[i].src, x->e_mbd.mi[0]->sb_type, src->buffers[i], - src->crop_widths[is_uv], src->crop_heights[is_uv], - src->strides[is_uv], mi_row, mi_col, NULL, - x->e_mbd.plane[i].subsampling_x, - x->e_mbd.plane[i].subsampling_y); + setup_pred_plane( + &x->plane[i].src, bsize, src->buffers[i], src->crop_widths[is_uv], + src->crop_heights[is_uv], src->strides[is_uv], mi_row, mi_col, NULL, + x->e_mbd.plane[i].subsampling_x, x->e_mbd.plane[i].subsampling_y); } } -static int set_segment_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x, - int8_t segment_id) { - const AV1_COMMON *const cm = &cpi->common; - av1_init_plane_quantizers(cpi, x, segment_id); - aom_clear_system_state(); - int segment_qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex); - return av1_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q); +static EdgeInfo edge_info(const struct buf_2d *ref, const BLOCK_SIZE bsize, + const bool high_bd, const int bd) { + const int width = block_size_wide[bsize]; + const int height = block_size_high[bsize]; + // Implementation requires width to be a multiple of 8. It also requires + // height to be a multiple of 4, but this is always the case. + assert(height % 4 == 0); + if (width % 8 != 0) { + EdgeInfo ei = { .magnitude = 0, .x = 0, .y = 0 }; + return ei; + } + return av1_edge_exists(ref->buf, ref->stride, width, height, high_bd, bd); } -static int set_deltaq_rdmult(const AV1_COMP *const cpi, MACROBLOCKD *const xd) { - const AV1_COMMON *const cm = &cpi->common; +static int use_pb_simple_motion_pred_sse(const AV1_COMP *const cpi) { + // TODO(debargha, yuec): Not in use, need to implement a speed feature + // utilizing this data point, and replace '0' by the corresponding speed + // feature flag. + return 0 && !frame_is_intra_only(&cpi->common); +} - return av1_compute_rd_mult( - cpi, cm->base_qindex + xd->delta_qindex + cm->y_dc_delta_q); +static void hybrid_intra_mode_search(AV1_COMP *cpi, MACROBLOCK *const x, + RD_STATS *rd_cost, BLOCK_SIZE bsize, + PICK_MODE_CONTEXT *ctx) { + // TODO(jianj): Investigate the failure of ScalabilityTest in AOM_Q mode, + // which sets base_qindex to 0 on keyframe. + if (cpi->oxcf.rc_mode != AOM_CBR || !cpi->sf.rt_sf.hybrid_intra_pickmode || + bsize < BLOCK_16X16) + av1_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX); + else + av1_pick_intra_mode(cpi, x, rd_cost, bsize, ctx); } -static void rd_pick_sb_modes(AV1_COMP *const cpi, TileDataEnc *tile_data, - MACROBLOCK *const x, int mi_row, int mi_col, - RD_STATS *rd_cost, PARTITION_TYPE partition, - BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx, - int64_t best_rd) { +static AOM_INLINE void pick_sb_modes(AV1_COMP *const cpi, + TileDataEnc *tile_data, + MACROBLOCK *const x, int mi_row, + int mi_col, RD_STATS *rd_cost, + PARTITION_TYPE partition, BLOCK_SIZE bsize, + PICK_MODE_CONTEXT *ctx, RD_STATS best_rd, + int pick_mode_type) { + if (best_rd.rdcost < 0) { + ctx->rd_stats.rdcost = INT64_MAX; + ctx->rd_stats.skip = 0; + av1_invalid_rd_stats(rd_cost); + return; + } + + set_offsets(cpi, &tile_data->tile_info, x, mi_row, mi_col, bsize); + + if (ctx->rd_mode_is_ready) { + assert(ctx->mic.sb_type == bsize); + assert(ctx->mic.partition == partition); + rd_cost->rate = ctx->rd_stats.rate; + rd_cost->dist = ctx->rd_stats.dist; + rd_cost->rdcost = ctx->rd_stats.rdcost; + return; + } + AV1_COMMON *const cm = &cpi->common; const int num_planes = av1_num_planes(cm); - TileInfo *const tile_info = &tile_data->tile_info; MACROBLOCKD *const xd = &x->e_mbd; MB_MODE_INFO *mbmi; - MB_MODE_INFO *ctx_mbmi = &ctx->mic; struct macroblock_plane *const p = x->plane; struct macroblockd_plane *const pd = xd->plane; const AQ_MODE aq_mode = cpi->oxcf.aq_mode; - const DELTAQ_MODE deltaq_mode = cpi->oxcf.deltaq_mode; - int i, orig_rdmult; + int i; - if (best_rd < 0) { - ctx->rdcost = INT64_MAX; - ctx->skip = 0; - av1_invalid_rd_stats(rd_cost); - return; - } +#if CONFIG_COLLECT_COMPONENT_TIMING + start_timing(cpi, rd_pick_sb_modes_time); +#endif aom_clear_system_state(); - set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize); - mbmi = xd->mi[0]; - - if (ctx->rd_mode_is_ready) { - assert(ctx_mbmi->sb_type == bsize); - assert(ctx_mbmi->partition == partition); - *mbmi = *ctx_mbmi; - rd_cost->rate = ctx->rate; - rd_cost->dist = ctx->dist; - rd_cost->rdcost = ctx->rdcost; - } else { - mbmi->sb_type = bsize; - mbmi->partition = partition; - } + mbmi->sb_type = bsize; + mbmi->partition = partition; #if CONFIG_RD_DEBUG mbmi->mi_row = mi_row; mbmi->mi_col = mi_col; #endif + xd->tx_type_map = x->tx_type_map; + xd->tx_type_map_stride = mi_size_wide[bsize]; + for (i = 0; i < num_planes; ++i) { p[i].coeff = ctx->coeff[i]; p[i].qcoeff = ctx->qcoeff[i]; @@ -564,76 +780,97 @@ static void rd_pick_sb_modes(AV1_COMP *const cpi, TileDataEnc *tile_data, for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i]; - if (!ctx->rd_mode_is_ready) { - ctx->skippable = 0; + ctx->skippable = 0; + // Set to zero to make sure we do not use the previous encoded frame stats + mbmi->skip = 0; + // Reset skip mode flag. + mbmi->skip_mode = 0; - // Set to zero to make sure we do not use the previous encoded frame stats - mbmi->skip = 0; - - // Reset skip mode flag. - mbmi->skip_mode = 0; - } - - x->skip_chroma_rd = - !is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x, - xd->plane[1].subsampling_y); - - if (ctx->rd_mode_is_ready) { - x->skip = ctx->skip; - *x->mbmi_ext = ctx->mbmi_ext; - return; - } - - if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { + if (is_cur_buf_hbd(xd)) { x->source_variance = av1_high_get_sby_perpixel_variance( cpi, &x->plane[0].src, bsize, xd->bd); } else { x->source_variance = av1_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize); } + if (use_pb_simple_motion_pred_sse(cpi)) { + const FULLPEL_MV start_mv = kZeroFullMv; + unsigned int var = 0; + av1_simple_motion_sse_var(cpi, x, mi_row, mi_col, bsize, start_mv, 0, + &x->simple_motion_pred_sse, &var); + } - // Save rdmult before it might be changed, so it can be restored later. - orig_rdmult = x->rdmult; - - if (aq_mode == VARIANCE_AQ) { - if (cpi->vaq_refresh) { - const int energy = bsize <= BLOCK_16X16 - ? x->mb_energy - : av1_log_block_var(cpi, x, bsize); - mbmi->segment_id = energy; - } - x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id); - } else if (aq_mode == COMPLEXITY_AQ) { - x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id); - } else if (aq_mode == CYCLIC_REFRESH_AQ) { - // If segment is boosted, use rdmult for that segment. - if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) - x->rdmult = av1_cyclic_refresh_get_rdmult(cpi->cyclic_refresh); + // If the threshold for disabling wedge search is zero, it means the feature + // should not be used. Use a value that will always succeed in the check. + if (cpi->sf.inter_sf.disable_wedge_search_edge_thresh == 0) { + x->edge_strength = UINT16_MAX; + x->edge_strength_x = UINT16_MAX; + x->edge_strength_y = UINT16_MAX; + } else { + EdgeInfo ei = + edge_info(&x->plane[0].src, bsize, is_cur_buf_hbd(xd), xd->bd); + x->edge_strength = ei.magnitude; + x->edge_strength_x = ei.x; + x->edge_strength_y = ei.y; } - if (deltaq_mode > 0) x->rdmult = set_deltaq_rdmult(cpi, xd); + // Initialize default mode evaluation params + set_mode_eval_params(cpi, x, DEFAULT_EVAL); + + // Save rdmult before it might be changed, so it can be restored later. + const int orig_rdmult = x->rdmult; + setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, aq_mode, mbmi); + // Set error per bit for current rdmult + set_error_per_bit(x, x->rdmult); + av1_rd_cost_update(x->rdmult, &best_rd); // Find best coding mode & reconstruct the MB so it is available // as a predictor for MBs that follow in the SB if (frame_is_intra_only(cm)) { - av1_rd_pick_intra_mode_sb(cpi, x, mi_row, mi_col, rd_cost, bsize, ctx, - best_rd); +#if CONFIG_COLLECT_COMPONENT_TIMING + start_timing(cpi, av1_rd_pick_intra_mode_sb_time); +#endif + switch (pick_mode_type) { + case PICK_MODE_RD: + av1_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd.rdcost); + break; + case PICK_MODE_NONRD: + hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx); + break; + default: assert(0 && "Unknown pick mode type."); + } +#if CONFIG_COLLECT_COMPONENT_TIMING + end_timing(cpi, av1_rd_pick_intra_mode_sb_time); +#endif } else { +#if CONFIG_COLLECT_COMPONENT_TIMING + start_timing(cpi, av1_rd_pick_inter_mode_sb_time); +#endif if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) { av1_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, mi_row, mi_col, - rd_cost, bsize, ctx, best_rd); + rd_cost, bsize, ctx, best_rd.rdcost); } else { - av1_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col, rd_cost, - bsize, ctx, best_rd); + // TODO(kyslov): do the same for pick_inter_mode_sb_seg_skip + switch (pick_mode_type) { + case PICK_MODE_RD: + av1_rd_pick_inter_mode_sb(cpi, tile_data, x, rd_cost, bsize, ctx, + best_rd.rdcost); + break; + case PICK_MODE_NONRD: + av1_nonrd_pick_inter_mode_sb(cpi, tile_data, x, rd_cost, bsize, ctx, + best_rd.rdcost); + break; + default: assert(0 && "Unknown pick mode type."); + } } +#if CONFIG_COLLECT_COMPONENT_TIMING + end_timing(cpi, av1_rd_pick_inter_mode_sb_time); +#endif } // Examine the resulting rate and for AQ mode 2 make a segment choice. - if ((rd_cost->rate != INT_MAX) && (aq_mode == COMPLEXITY_AQ) && - (bsize >= BLOCK_16X16) && - (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame || - cpi->refresh_alt2_ref_frame || - (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) { + if (rd_cost->rate != INT_MAX && aq_mode == COMPLEXITY_AQ && + bsize >= BLOCK_16X16) { av1_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate); } @@ -643,14 +880,19 @@ static void rd_pick_sb_modes(AV1_COMP *const cpi, TileDataEnc *tile_data, // refactored to provide proper exit/return handle. if (rd_cost->rate == INT_MAX) rd_cost->rdcost = INT64_MAX; - ctx->rate = rd_cost->rate; - ctx->dist = rd_cost->dist; - ctx->rdcost = rd_cost->rdcost; + ctx->rd_stats.rate = rd_cost->rate; + ctx->rd_stats.dist = rd_cost->dist; + ctx->rd_stats.rdcost = rd_cost->rdcost; + +#if CONFIG_COLLECT_COMPONENT_TIMING + end_timing(cpi, rd_pick_sb_modes_time); +#endif } -static void update_inter_mode_stats(FRAME_CONTEXT *fc, FRAME_COUNTS *counts, - PREDICTION_MODE mode, int16_t mode_context, - uint8_t allow_update_cdf) { +static AOM_INLINE void update_inter_mode_stats(FRAME_CONTEXT *fc, + FRAME_COUNTS *counts, + PREDICTION_MODE mode, + int16_t mode_context) { (void)counts; int16_t mode_ctx = mode_context & NEWMV_CTX_MASK; @@ -658,38 +900,39 @@ static void update_inter_mode_stats(FRAME_CONTEXT *fc, FRAME_COUNTS *counts, #if CONFIG_ENTROPY_STATS ++counts->newmv_mode[mode_ctx][0]; #endif - if (allow_update_cdf) update_cdf(fc->newmv_cdf[mode_ctx], 0, 2); + update_cdf(fc->newmv_cdf[mode_ctx], 0, 2); return; - } else { + } + #if CONFIG_ENTROPY_STATS - ++counts->newmv_mode[mode_ctx][1]; + ++counts->newmv_mode[mode_ctx][1]; #endif - if (allow_update_cdf) update_cdf(fc->newmv_cdf[mode_ctx], 1, 2); + update_cdf(fc->newmv_cdf[mode_ctx], 1, 2); - mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK; - if (mode == GLOBALMV) { + mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK; + if (mode == GLOBALMV) { #if CONFIG_ENTROPY_STATS - ++counts->zeromv_mode[mode_ctx][0]; + ++counts->zeromv_mode[mode_ctx][0]; #endif - if (allow_update_cdf) update_cdf(fc->zeromv_cdf[mode_ctx], 0, 2); - return; - } else { + update_cdf(fc->zeromv_cdf[mode_ctx], 0, 2); + return; + } + #if CONFIG_ENTROPY_STATS - ++counts->zeromv_mode[mode_ctx][1]; + ++counts->zeromv_mode[mode_ctx][1]; #endif - if (allow_update_cdf) update_cdf(fc->zeromv_cdf[mode_ctx], 1, 2); - mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK; + update_cdf(fc->zeromv_cdf[mode_ctx], 1, 2); + + mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK; #if CONFIG_ENTROPY_STATS - ++counts->refmv_mode[mode_ctx][mode != NEARESTMV]; + ++counts->refmv_mode[mode_ctx][mode != NEARESTMV]; #endif - if (allow_update_cdf) - update_cdf(fc->refmv_cdf[mode_ctx], mode != NEARESTMV, 2); - } - } + update_cdf(fc->refmv_cdf[mode_ctx], mode != NEARESTMV, 2); } -static void update_palette_cdf(MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi, - FRAME_COUNTS *counts, uint8_t allow_update_cdf) { +static AOM_INLINE void update_palette_cdf(MACROBLOCKD *xd, + const MB_MODE_INFO *const mbmi, + FRAME_COUNTS *counts) { FRAME_CONTEXT *fc = xd->tile_ctx; const BLOCK_SIZE bsize = mbmi->sb_type; const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info; @@ -704,17 +947,14 @@ static void update_palette_cdf(MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi, #if CONFIG_ENTROPY_STATS ++counts->palette_y_mode[palette_bsize_ctx][palette_mode_ctx][n > 0]; #endif - if (allow_update_cdf) - update_cdf(fc->palette_y_mode_cdf[palette_bsize_ctx][palette_mode_ctx], - n > 0, 2); + update_cdf(fc->palette_y_mode_cdf[palette_bsize_ctx][palette_mode_ctx], + n > 0, 2); if (n > 0) { #if CONFIG_ENTROPY_STATS ++counts->palette_y_size[palette_bsize_ctx][n - PALETTE_MIN_SIZE]; #endif - if (allow_update_cdf) { - update_cdf(fc->palette_y_size_cdf[palette_bsize_ctx], - n - PALETTE_MIN_SIZE, PALETTE_SIZES); - } + update_cdf(fc->palette_y_size_cdf[palette_bsize_ctx], + n - PALETTE_MIN_SIZE, PALETTE_SIZES); } } @@ -725,30 +965,26 @@ static void update_palette_cdf(MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi, #if CONFIG_ENTROPY_STATS ++counts->palette_uv_mode[palette_uv_mode_ctx][n > 0]; #endif - if (allow_update_cdf) - update_cdf(fc->palette_uv_mode_cdf[palette_uv_mode_ctx], n > 0, 2); + update_cdf(fc->palette_uv_mode_cdf[palette_uv_mode_ctx], n > 0, 2); if (n > 0) { #if CONFIG_ENTROPY_STATS ++counts->palette_uv_size[palette_bsize_ctx][n - PALETTE_MIN_SIZE]; #endif - if (allow_update_cdf) { - update_cdf(fc->palette_uv_size_cdf[palette_bsize_ctx], - n - PALETTE_MIN_SIZE, PALETTE_SIZES); - } + update_cdf(fc->palette_uv_size_cdf[palette_bsize_ctx], + n - PALETTE_MIN_SIZE, PALETTE_SIZES); } } } -static void sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts, - MACROBLOCKD *xd, const MB_MODE_INFO *const mbmi, - const MB_MODE_INFO *above_mi, - const MB_MODE_INFO *left_mi, const int intraonly, - const int mi_row, const int mi_col, - uint8_t allow_update_cdf) { +static AOM_INLINE void sum_intra_stats(const AV1_COMMON *const cm, + FRAME_COUNTS *counts, MACROBLOCKD *xd, + const MB_MODE_INFO *const mbmi, + const MB_MODE_INFO *above_mi, + const MB_MODE_INFO *left_mi, + const int intraonly) { FRAME_CONTEXT *fc = xd->tile_ctx; const PREDICTION_MODE y_mode = mbmi->mode; - const UV_PREDICTION_MODE uv_mode = mbmi->uv_mode; (void)counts; const BLOCK_SIZE bsize = mbmi->sb_type; @@ -760,14 +996,12 @@ static void sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts, const int left_ctx = intra_mode_context[left]; ++counts->kf_y_mode[above_ctx][left_ctx][y_mode]; #endif // CONFIG_ENTROPY_STATS - if (allow_update_cdf) - update_cdf(get_y_mode_cdf(fc, above_mi, left_mi), y_mode, INTRA_MODES); + update_cdf(get_y_mode_cdf(fc, above_mi, left_mi), y_mode, INTRA_MODES); } else { #if CONFIG_ENTROPY_STATS ++counts->y_mode[size_group_lookup[bsize]][y_mode]; #endif // CONFIG_ENTROPY_STATS - if (allow_update_cdf) - update_cdf(fc->y_mode_cdf[size_group_lookup[bsize]], y_mode, INTRA_MODES); + update_cdf(fc->y_mode_cdf[size_group_lookup[bsize]], y_mode, INTRA_MODES); } if (av1_filter_intra_allowed(cm, mbmi)) { @@ -780,14 +1014,11 @@ static void sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts, ->filter_intra_mode[mbmi->filter_intra_mode_info.filter_intra_mode]; } #endif // CONFIG_ENTROPY_STATS - if (allow_update_cdf) { - update_cdf(fc->filter_intra_cdfs[mbmi->sb_type], use_filter_intra_mode, - 2); - if (use_filter_intra_mode) { - update_cdf(fc->filter_intra_mode_cdf, - mbmi->filter_intra_mode_info.filter_intra_mode, - FILTER_INTRA_MODES); - } + update_cdf(fc->filter_intra_cdfs[mbmi->sb_type], use_filter_intra_mode, 2); + if (use_filter_intra_mode) { + update_cdf(fc->filter_intra_mode_cdf, + mbmi->filter_intra_mode_info.filter_intra_mode, + FILTER_INTRA_MODES); } } if (av1_is_directional_mode(mbmi->mode) && av1_use_angle_delta(bsize)) { @@ -795,43 +1026,35 @@ static void sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts, ++counts->angle_delta[mbmi->mode - V_PRED] [mbmi->angle_delta[PLANE_TYPE_Y] + MAX_ANGLE_DELTA]; #endif - if (allow_update_cdf) { - update_cdf(fc->angle_delta_cdf[mbmi->mode - V_PRED], - mbmi->angle_delta[PLANE_TYPE_Y] + MAX_ANGLE_DELTA, - 2 * MAX_ANGLE_DELTA + 1); - } + update_cdf(fc->angle_delta_cdf[mbmi->mode - V_PRED], + mbmi->angle_delta[PLANE_TYPE_Y] + MAX_ANGLE_DELTA, + 2 * MAX_ANGLE_DELTA + 1); } - if (!is_chroma_reference(mi_row, mi_col, bsize, - xd->plane[AOM_PLANE_U].subsampling_x, - xd->plane[AOM_PLANE_U].subsampling_y)) - return; + if (!xd->is_chroma_ref) return; + const UV_PREDICTION_MODE uv_mode = mbmi->uv_mode; + const CFL_ALLOWED_TYPE cfl_allowed = is_cfl_allowed(xd); #if CONFIG_ENTROPY_STATS - ++counts->uv_mode[is_cfl_allowed(xd)][y_mode][uv_mode]; + ++counts->uv_mode[cfl_allowed][y_mode][uv_mode]; #endif // CONFIG_ENTROPY_STATS - if (allow_update_cdf) { - const CFL_ALLOWED_TYPE cfl_allowed = is_cfl_allowed(xd); - update_cdf(fc->uv_mode_cdf[cfl_allowed][y_mode], uv_mode, - UV_INTRA_MODES - !cfl_allowed); - } + update_cdf(fc->uv_mode_cdf[cfl_allowed][y_mode], uv_mode, + UV_INTRA_MODES - !cfl_allowed); if (uv_mode == UV_CFL_PRED) { - const int joint_sign = mbmi->cfl_alpha_signs; - const int idx = mbmi->cfl_alpha_idx; + const int8_t joint_sign = mbmi->cfl_alpha_signs; + const uint8_t idx = mbmi->cfl_alpha_idx; #if CONFIG_ENTROPY_STATS ++counts->cfl_sign[joint_sign]; #endif - if (allow_update_cdf) - update_cdf(fc->cfl_sign_cdf, joint_sign, CFL_JOINT_SIGNS); + update_cdf(fc->cfl_sign_cdf, joint_sign, CFL_JOINT_SIGNS); if (CFL_SIGN_U(joint_sign) != CFL_SIGN_ZERO) { aom_cdf_prob *cdf_u = fc->cfl_alpha_cdf[CFL_CONTEXT_U(joint_sign)]; #if CONFIG_ENTROPY_STATS ++counts->cfl_alpha[CFL_CONTEXT_U(joint_sign)][CFL_IDX_U(idx)]; #endif - if (allow_update_cdf) - update_cdf(cdf_u, CFL_IDX_U(idx), CFL_ALPHABET_SIZE); + update_cdf(cdf_u, CFL_IDX_U(idx), CFL_ALPHABET_SIZE); } if (CFL_SIGN_V(joint_sign) != CFL_SIGN_ZERO) { aom_cdf_prob *cdf_v = fc->cfl_alpha_cdf[CFL_CONTEXT_V(joint_sign)]; @@ -839,8 +1062,7 @@ static void sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts, #if CONFIG_ENTROPY_STATS ++counts->cfl_alpha[CFL_CONTEXT_V(joint_sign)][CFL_IDX_V(idx)]; #endif - if (allow_update_cdf) - update_cdf(cdf_v, CFL_IDX_V(idx), CFL_ALPHABET_SIZE); + update_cdf(cdf_v, CFL_IDX_V(idx), CFL_ALPHABET_SIZE); } } if (av1_is_directional_mode(get_uv_mode(uv_mode)) && @@ -849,526 +1071,449 @@ static void sum_intra_stats(const AV1_COMMON *const cm, FRAME_COUNTS *counts, ++counts->angle_delta[uv_mode - UV_V_PRED] [mbmi->angle_delta[PLANE_TYPE_UV] + MAX_ANGLE_DELTA]; #endif - if (allow_update_cdf) { - update_cdf(fc->angle_delta_cdf[uv_mode - UV_V_PRED], - mbmi->angle_delta[PLANE_TYPE_UV] + MAX_ANGLE_DELTA, - 2 * MAX_ANGLE_DELTA + 1); - } + update_cdf(fc->angle_delta_cdf[uv_mode - UV_V_PRED], + mbmi->angle_delta[PLANE_TYPE_UV] + MAX_ANGLE_DELTA, + 2 * MAX_ANGLE_DELTA + 1); + } + if (av1_allow_palette(cm->features.allow_screen_content_tools, bsize)) { + update_palette_cdf(xd, mbmi, counts); } - if (av1_allow_palette(cm->allow_screen_content_tools, bsize)) - update_palette_cdf(xd, mbmi, counts, allow_update_cdf); } -static void update_stats(const AV1_COMMON *const cm, TileDataEnc *tile_data, - ThreadData *td, int mi_row, int mi_col) { +static AOM_INLINE void update_stats(const AV1_COMMON *const cm, + ThreadData *td) { MACROBLOCK *x = &td->mb; MACROBLOCKD *const xd = &x->e_mbd; const MB_MODE_INFO *const mbmi = xd->mi[0]; const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext; + const CurrentFrame *const current_frame = &cm->current_frame; const BLOCK_SIZE bsize = mbmi->sb_type; FRAME_CONTEXT *fc = xd->tile_ctx; - const uint8_t allow_update_cdf = tile_data->allow_update_cdf; - - // delta quant applies to both intra and inter - const int super_block_upper_left = - ((mi_row & (cm->seq_params.mib_size - 1)) == 0) && - ((mi_col & (cm->seq_params.mib_size - 1)) == 0); - const int seg_ref_active = segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME); - if (cm->skip_mode_flag && !seg_ref_active && is_comp_ref_allowed(bsize)) { + if (current_frame->skip_mode_info.skip_mode_flag && !seg_ref_active && + is_comp_ref_allowed(bsize)) { const int skip_mode_ctx = av1_get_skip_mode_context(xd); #if CONFIG_ENTROPY_STATS td->counts->skip_mode[skip_mode_ctx][mbmi->skip_mode]++; #endif - if (allow_update_cdf) - update_cdf(fc->skip_mode_cdfs[skip_mode_ctx], mbmi->skip_mode, 2); + update_cdf(fc->skip_mode_cdfs[skip_mode_ctx], mbmi->skip_mode, 2); } - if (!mbmi->skip_mode) { - if (!seg_ref_active) { - const int skip_ctx = av1_get_skip_context(xd); + if (!mbmi->skip_mode && !seg_ref_active) { + const int skip_ctx = av1_get_skip_context(xd); #if CONFIG_ENTROPY_STATS - td->counts->skip[skip_ctx][mbmi->skip]++; + td->counts->skip[skip_ctx][mbmi->skip]++; #endif - if (allow_update_cdf) update_cdf(fc->skip_cdfs[skip_ctx], mbmi->skip, 2); - } + update_cdf(fc->skip_cdfs[skip_ctx], mbmi->skip, 2); } - if (cm->delta_q_present_flag && +#if CONFIG_ENTROPY_STATS + // delta quant applies to both intra and inter + const int super_block_upper_left = + ((xd->mi_row & (cm->seq_params.mib_size - 1)) == 0) && + ((xd->mi_col & (cm->seq_params.mib_size - 1)) == 0); + const DeltaQInfo *const delta_q_info = &cm->delta_q_info; + if (delta_q_info->delta_q_present_flag && (bsize != cm->seq_params.sb_size || !mbmi->skip) && super_block_upper_left) { -#if CONFIG_ENTROPY_STATS const int dq = - (mbmi->current_qindex - xd->current_qindex) / cm->delta_q_res; + (mbmi->current_qindex - xd->current_qindex) / delta_q_info->delta_q_res; const int absdq = abs(dq); for (int i = 0; i < AOMMIN(absdq, DELTA_Q_SMALL); ++i) { td->counts->delta_q[i][1]++; } if (absdq < DELTA_Q_SMALL) td->counts->delta_q[absdq][0]++; -#endif - xd->current_qindex = mbmi->current_qindex; - if (cm->delta_lf_present_flag) { - if (cm->delta_lf_multi) { + if (delta_q_info->delta_lf_present_flag) { + if (delta_q_info->delta_lf_multi) { const int frame_lf_count = av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2; for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) { -#if CONFIG_ENTROPY_STATS - const int delta_lf = - (mbmi->delta_lf[lf_id] - xd->delta_lf[lf_id]) / cm->delta_lf_res; + const int delta_lf = (mbmi->delta_lf[lf_id] - xd->delta_lf[lf_id]) / + delta_q_info->delta_lf_res; const int abs_delta_lf = abs(delta_lf); for (int i = 0; i < AOMMIN(abs_delta_lf, DELTA_LF_SMALL); ++i) { td->counts->delta_lf_multi[lf_id][i][1]++; } if (abs_delta_lf < DELTA_LF_SMALL) td->counts->delta_lf_multi[lf_id][abs_delta_lf][0]++; -#endif - xd->delta_lf[lf_id] = mbmi->delta_lf[lf_id]; } } else { -#if CONFIG_ENTROPY_STATS const int delta_lf = (mbmi->delta_lf_from_base - xd->delta_lf_from_base) / - cm->delta_lf_res; + delta_q_info->delta_lf_res; const int abs_delta_lf = abs(delta_lf); for (int i = 0; i < AOMMIN(abs_delta_lf, DELTA_LF_SMALL); ++i) { td->counts->delta_lf[i][1]++; } if (abs_delta_lf < DELTA_LF_SMALL) td->counts->delta_lf[abs_delta_lf][0]++; -#endif - xd->delta_lf_from_base = mbmi->delta_lf_from_base; } } } +#endif if (!is_inter_block(mbmi)) { sum_intra_stats(cm, td->counts, xd, mbmi, xd->above_mbmi, xd->left_mbmi, - frame_is_intra_only(cm), mi_row, mi_col, - tile_data->allow_update_cdf); + frame_is_intra_only(cm)); } if (av1_allow_intrabc(cm)) { - if (allow_update_cdf) - update_cdf(fc->intrabc_cdf, is_intrabc_block(mbmi), 2); + update_cdf(fc->intrabc_cdf, is_intrabc_block(mbmi), 2); #if CONFIG_ENTROPY_STATS ++td->counts->intrabc[is_intrabc_block(mbmi)]; #endif // CONFIG_ENTROPY_STATS } - if (!frame_is_intra_only(cm)) { - RD_COUNTS *rdc = &td->rd_counts; + if (frame_is_intra_only(cm) || mbmi->skip_mode) return; - FRAME_COUNTS *const counts = td->counts; + FRAME_COUNTS *const counts = td->counts; + const int inter_block = is_inter_block(mbmi); - if (mbmi->skip_mode) { - rdc->skip_mode_used_flag = 1; - if (cm->reference_mode == REFERENCE_MODE_SELECT) { - assert(has_second_ref(mbmi)); - rdc->compound_ref_used_flag = 1; - } - set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]); - return; - } - - const int inter_block = is_inter_block(mbmi); - - if (!seg_ref_active) { + if (!seg_ref_active) { #if CONFIG_ENTROPY_STATS - counts->intra_inter[av1_get_intra_inter_context(xd)][inter_block]++; + counts->intra_inter[av1_get_intra_inter_context(xd)][inter_block]++; #endif - if (allow_update_cdf) { - update_cdf(fc->intra_inter_cdf[av1_get_intra_inter_context(xd)], - inter_block, 2); - } - // If the segment reference feature is enabled we have only a single - // reference frame allowed for the segment so exclude it from - // the reference frame counts used to work out probabilities. - if (inter_block) { - const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0]; - const MV_REFERENCE_FRAME ref1 = mbmi->ref_frame[1]; - - av1_collect_neighbors_ref_counts(xd); - - if (cm->reference_mode == REFERENCE_MODE_SELECT) { - if (has_second_ref(mbmi)) - // This flag is also updated for 4x4 blocks - rdc->compound_ref_used_flag = 1; - if (is_comp_ref_allowed(bsize)) { + update_cdf(fc->intra_inter_cdf[av1_get_intra_inter_context(xd)], + inter_block, 2); + // If the segment reference feature is enabled we have only a single + // reference frame allowed for the segment so exclude it from + // the reference frame counts used to work out probabilities. + if (inter_block) { + const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0]; + const MV_REFERENCE_FRAME ref1 = mbmi->ref_frame[1]; + if (current_frame->reference_mode == REFERENCE_MODE_SELECT) { + if (is_comp_ref_allowed(bsize)) { #if CONFIG_ENTROPY_STATS - counts->comp_inter[av1_get_reference_mode_context(xd)] - [has_second_ref(mbmi)]++; + counts->comp_inter[av1_get_reference_mode_context(xd)] + [has_second_ref(mbmi)]++; #endif // CONFIG_ENTROPY_STATS - if (allow_update_cdf) { - update_cdf(av1_get_reference_mode_cdf(xd), has_second_ref(mbmi), - 2); - } - } + update_cdf(av1_get_reference_mode_cdf(xd), has_second_ref(mbmi), 2); } + } - if (has_second_ref(mbmi)) { - const COMP_REFERENCE_TYPE comp_ref_type = has_uni_comp_refs(mbmi) - ? UNIDIR_COMP_REFERENCE - : BIDIR_COMP_REFERENCE; - if (allow_update_cdf) { - update_cdf(av1_get_comp_reference_type_cdf(xd), comp_ref_type, - COMP_REFERENCE_TYPES); - } + if (has_second_ref(mbmi)) { + const COMP_REFERENCE_TYPE comp_ref_type = has_uni_comp_refs(mbmi) + ? UNIDIR_COMP_REFERENCE + : BIDIR_COMP_REFERENCE; + update_cdf(av1_get_comp_reference_type_cdf(xd), comp_ref_type, + COMP_REFERENCE_TYPES); #if CONFIG_ENTROPY_STATS - counts->comp_ref_type[av1_get_comp_reference_type_context(xd)] - [comp_ref_type]++; + counts->comp_ref_type[av1_get_comp_reference_type_context(xd)] + [comp_ref_type]++; #endif // CONFIG_ENTROPY_STATS - if (comp_ref_type == UNIDIR_COMP_REFERENCE) { - const int bit = (ref0 == BWDREF_FRAME); - if (allow_update_cdf) - update_cdf(av1_get_pred_cdf_uni_comp_ref_p(xd), bit, 2); + if (comp_ref_type == UNIDIR_COMP_REFERENCE) { + const int bit = (ref0 == BWDREF_FRAME); + update_cdf(av1_get_pred_cdf_uni_comp_ref_p(xd), bit, 2); #if CONFIG_ENTROPY_STATS - counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p(xd)][0] - [bit]++; + counts + ->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p(xd)][0][bit]++; #endif // CONFIG_ENTROPY_STATS - if (!bit) { - const int bit1 = (ref1 == LAST3_FRAME || ref1 == GOLDEN_FRAME); - if (allow_update_cdf) - update_cdf(av1_get_pred_cdf_uni_comp_ref_p1(xd), bit1, 2); + if (!bit) { + const int bit1 = (ref1 == LAST3_FRAME || ref1 == GOLDEN_FRAME); + update_cdf(av1_get_pred_cdf_uni_comp_ref_p1(xd), bit1, 2); #if CONFIG_ENTROPY_STATS - counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p1(xd)][1] - [bit1]++; + counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p1(xd)][1] + [bit1]++; #endif // CONFIG_ENTROPY_STATS - if (bit1) { - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_uni_comp_ref_p2(xd), - ref1 == GOLDEN_FRAME, 2); - } + if (bit1) { + update_cdf(av1_get_pred_cdf_uni_comp_ref_p2(xd), + ref1 == GOLDEN_FRAME, 2); #if CONFIG_ENTROPY_STATS - counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p2(xd)] - [2][ref1 == GOLDEN_FRAME]++; + counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p2(xd)][2] + [ref1 == GOLDEN_FRAME]++; #endif // CONFIG_ENTROPY_STATS - } } - } else { - const int bit = (ref0 == GOLDEN_FRAME || ref0 == LAST3_FRAME); - if (allow_update_cdf) - update_cdf(av1_get_pred_cdf_comp_ref_p(xd), bit, 2); + } + } else { + const int bit = (ref0 == GOLDEN_FRAME || ref0 == LAST3_FRAME); + update_cdf(av1_get_pred_cdf_comp_ref_p(xd), bit, 2); #if CONFIG_ENTROPY_STATS - counts->comp_ref[av1_get_pred_context_comp_ref_p(xd)][0][bit]++; + counts->comp_ref[av1_get_pred_context_comp_ref_p(xd)][0][bit]++; #endif // CONFIG_ENTROPY_STATS - if (!bit) { - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_comp_ref_p1(xd), - ref0 == LAST2_FRAME, 2); - } + if (!bit) { + update_cdf(av1_get_pred_cdf_comp_ref_p1(xd), ref0 == LAST2_FRAME, + 2); #if CONFIG_ENTROPY_STATS - counts->comp_ref[av1_get_pred_context_comp_ref_p1(xd)][1] - [ref0 == LAST2_FRAME]++; + counts->comp_ref[av1_get_pred_context_comp_ref_p1(xd)][1] + [ref0 == LAST2_FRAME]++; #endif // CONFIG_ENTROPY_STATS - } else { - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_comp_ref_p2(xd), - ref0 == GOLDEN_FRAME, 2); - } + } else { + update_cdf(av1_get_pred_cdf_comp_ref_p2(xd), ref0 == GOLDEN_FRAME, + 2); #if CONFIG_ENTROPY_STATS - counts->comp_ref[av1_get_pred_context_comp_ref_p2(xd)][2] - [ref0 == GOLDEN_FRAME]++; + counts->comp_ref[av1_get_pred_context_comp_ref_p2(xd)][2] + [ref0 == GOLDEN_FRAME]++; #endif // CONFIG_ENTROPY_STATS - } - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_comp_bwdref_p(xd), - ref1 == ALTREF_FRAME, 2); - } + } + update_cdf(av1_get_pred_cdf_comp_bwdref_p(xd), ref1 == ALTREF_FRAME, + 2); #if CONFIG_ENTROPY_STATS - counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p(xd)][0] - [ref1 == ALTREF_FRAME]++; + counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p(xd)][0] + [ref1 == ALTREF_FRAME]++; #endif // CONFIG_ENTROPY_STATS - if (ref1 != ALTREF_FRAME) { - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_comp_bwdref_p1(xd), - ref1 == ALTREF2_FRAME, 2); - } + if (ref1 != ALTREF_FRAME) { + update_cdf(av1_get_pred_cdf_comp_bwdref_p1(xd), + ref1 == ALTREF2_FRAME, 2); #if CONFIG_ENTROPY_STATS - counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p1(xd)][1] - [ref1 == ALTREF2_FRAME]++; + counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p1(xd)][1] + [ref1 == ALTREF2_FRAME]++; #endif // CONFIG_ENTROPY_STATS - } } - } else { - const int bit = (ref0 >= BWDREF_FRAME); - if (allow_update_cdf) - update_cdf(av1_get_pred_cdf_single_ref_p1(xd), bit, 2); + } + } else { + const int bit = (ref0 >= BWDREF_FRAME); + update_cdf(av1_get_pred_cdf_single_ref_p1(xd), bit, 2); #if CONFIG_ENTROPY_STATS - counts->single_ref[av1_get_pred_context_single_ref_p1(xd)][0][bit]++; + counts->single_ref[av1_get_pred_context_single_ref_p1(xd)][0][bit]++; #endif // CONFIG_ENTROPY_STATS - if (bit) { - assert(ref0 <= ALTREF_FRAME); - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_single_ref_p2(xd), - ref0 == ALTREF_FRAME, 2); - } + if (bit) { + assert(ref0 <= ALTREF_FRAME); + update_cdf(av1_get_pred_cdf_single_ref_p2(xd), ref0 == ALTREF_FRAME, + 2); #if CONFIG_ENTROPY_STATS - counts->single_ref[av1_get_pred_context_single_ref_p2(xd)][1] - [ref0 == ALTREF_FRAME]++; + counts->single_ref[av1_get_pred_context_single_ref_p2(xd)][1] + [ref0 == ALTREF_FRAME]++; #endif // CONFIG_ENTROPY_STATS - if (ref0 != ALTREF_FRAME) { - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_single_ref_p6(xd), - ref0 == ALTREF2_FRAME, 2); - } + if (ref0 != ALTREF_FRAME) { + update_cdf(av1_get_pred_cdf_single_ref_p6(xd), + ref0 == ALTREF2_FRAME, 2); #if CONFIG_ENTROPY_STATS - counts->single_ref[av1_get_pred_context_single_ref_p6(xd)][5] - [ref0 == ALTREF2_FRAME]++; + counts->single_ref[av1_get_pred_context_single_ref_p6(xd)][5] + [ref0 == ALTREF2_FRAME]++; #endif // CONFIG_ENTROPY_STATS - } - } else { - const int bit1 = !(ref0 == LAST2_FRAME || ref0 == LAST_FRAME); - if (allow_update_cdf) - update_cdf(av1_get_pred_cdf_single_ref_p3(xd), bit1, 2); + } + } else { + const int bit1 = !(ref0 == LAST2_FRAME || ref0 == LAST_FRAME); + update_cdf(av1_get_pred_cdf_single_ref_p3(xd), bit1, 2); #if CONFIG_ENTROPY_STATS - counts - ->single_ref[av1_get_pred_context_single_ref_p3(xd)][2][bit1]++; + counts->single_ref[av1_get_pred_context_single_ref_p3(xd)][2][bit1]++; #endif // CONFIG_ENTROPY_STATS - if (!bit1) { - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_single_ref_p4(xd), - ref0 != LAST_FRAME, 2); - } + if (!bit1) { + update_cdf(av1_get_pred_cdf_single_ref_p4(xd), ref0 != LAST_FRAME, + 2); #if CONFIG_ENTROPY_STATS - counts->single_ref[av1_get_pred_context_single_ref_p4(xd)][3] - [ref0 != LAST_FRAME]++; + counts->single_ref[av1_get_pred_context_single_ref_p4(xd)][3] + [ref0 != LAST_FRAME]++; #endif // CONFIG_ENTROPY_STATS - } else { - if (allow_update_cdf) { - update_cdf(av1_get_pred_cdf_single_ref_p5(xd), - ref0 != LAST3_FRAME, 2); - } + } else { + update_cdf(av1_get_pred_cdf_single_ref_p5(xd), ref0 != LAST3_FRAME, + 2); #if CONFIG_ENTROPY_STATS - counts->single_ref[av1_get_pred_context_single_ref_p5(xd)][4] - [ref0 != LAST3_FRAME]++; + counts->single_ref[av1_get_pred_context_single_ref_p5(xd)][4] + [ref0 != LAST3_FRAME]++; #endif // CONFIG_ENTROPY_STATS - } } } + } - if (cm->seq_params.enable_interintra_compound && - is_interintra_allowed(mbmi)) { - const int bsize_group = size_group_lookup[bsize]; - if (mbmi->ref_frame[1] == INTRA_FRAME) { + if (cm->seq_params.enable_interintra_compound && + is_interintra_allowed(mbmi)) { + const int bsize_group = size_group_lookup[bsize]; + if (mbmi->ref_frame[1] == INTRA_FRAME) { #if CONFIG_ENTROPY_STATS - counts->interintra[bsize_group][1]++; + counts->interintra[bsize_group][1]++; #endif - if (allow_update_cdf) - update_cdf(fc->interintra_cdf[bsize_group], 1, 2); + update_cdf(fc->interintra_cdf[bsize_group], 1, 2); #if CONFIG_ENTROPY_STATS - counts->interintra_mode[bsize_group][mbmi->interintra_mode]++; + counts->interintra_mode[bsize_group][mbmi->interintra_mode]++; #endif - if (allow_update_cdf) { - update_cdf(fc->interintra_mode_cdf[bsize_group], - mbmi->interintra_mode, INTERINTRA_MODES); - } - if (is_interintra_wedge_used(bsize)) { + update_cdf(fc->interintra_mode_cdf[bsize_group], + mbmi->interintra_mode, INTERINTRA_MODES); + if (av1_is_wedge_used(bsize)) { #if CONFIG_ENTROPY_STATS - counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++; + counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++; #endif - if (allow_update_cdf) { - update_cdf(fc->wedge_interintra_cdf[bsize], - mbmi->use_wedge_interintra, 2); - } - if (mbmi->use_wedge_interintra) { + update_cdf(fc->wedge_interintra_cdf[bsize], + mbmi->use_wedge_interintra, 2); + if (mbmi->use_wedge_interintra) { #if CONFIG_ENTROPY_STATS - counts->wedge_idx[bsize][mbmi->interintra_wedge_index]++; + counts->wedge_idx[bsize][mbmi->interintra_wedge_index]++; #endif - if (allow_update_cdf) { - update_cdf(fc->wedge_idx_cdf[bsize], - mbmi->interintra_wedge_index, 16); - } - } + update_cdf(fc->wedge_idx_cdf[bsize], mbmi->interintra_wedge_index, + 16); } - } else { + } + } else { #if CONFIG_ENTROPY_STATS - counts->interintra[bsize_group][0]++; + counts->interintra[bsize_group][0]++; #endif - if (allow_update_cdf) - update_cdf(fc->interintra_cdf[bsize_group], 0, 2); - } + update_cdf(fc->interintra_cdf[bsize_group], 0, 2); } + } - set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]); - const MOTION_MODE motion_allowed = - cm->switchable_motion_mode - ? motion_mode_allowed(xd->global_motion, xd, mbmi, - cm->allow_warped_motion) - : SIMPLE_TRANSLATION; - if (mbmi->ref_frame[1] != INTRA_FRAME) { - if (motion_allowed == WARPED_CAUSAL) { + const MOTION_MODE motion_allowed = + cm->features.switchable_motion_mode + ? motion_mode_allowed(xd->global_motion, xd, mbmi, + cm->features.allow_warped_motion) + : SIMPLE_TRANSLATION; + if (mbmi->ref_frame[1] != INTRA_FRAME) { + if (motion_allowed == WARPED_CAUSAL) { #if CONFIG_ENTROPY_STATS - counts->motion_mode[bsize][mbmi->motion_mode]++; + counts->motion_mode[bsize][mbmi->motion_mode]++; #endif - if (allow_update_cdf) { - update_cdf(fc->motion_mode_cdf[bsize], mbmi->motion_mode, - MOTION_MODES); - } - } else if (motion_allowed == OBMC_CAUSAL) { + update_cdf(fc->motion_mode_cdf[bsize], mbmi->motion_mode, + MOTION_MODES); + } else if (motion_allowed == OBMC_CAUSAL) { #if CONFIG_ENTROPY_STATS - counts->obmc[bsize][mbmi->motion_mode == OBMC_CAUSAL]++; + counts->obmc[bsize][mbmi->motion_mode == OBMC_CAUSAL]++; #endif - if (allow_update_cdf) { - update_cdf(fc->obmc_cdf[bsize], mbmi->motion_mode == OBMC_CAUSAL, - 2); - } - } + update_cdf(fc->obmc_cdf[bsize], mbmi->motion_mode == OBMC_CAUSAL, 2); } + } - if (has_second_ref(mbmi)) { - assert(cm->reference_mode != SINGLE_REFERENCE && - is_inter_compound_mode(mbmi->mode) && - mbmi->motion_mode == SIMPLE_TRANSLATION); - - const int masked_compound_used = - is_any_masked_compound_used(bsize) && - cm->seq_params.enable_masked_compound; - if (masked_compound_used) { - const int comp_group_idx_ctx = get_comp_group_idx_context(xd); + if (has_second_ref(mbmi)) { + assert(current_frame->reference_mode != SINGLE_REFERENCE && + is_inter_compound_mode(mbmi->mode) && + mbmi->motion_mode == SIMPLE_TRANSLATION); + + const int masked_compound_used = is_any_masked_compound_used(bsize) && + cm->seq_params.enable_masked_compound; + if (masked_compound_used) { + const int comp_group_idx_ctx = get_comp_group_idx_context(xd); #if CONFIG_ENTROPY_STATS - ++counts->comp_group_idx[comp_group_idx_ctx][mbmi->comp_group_idx]; + ++counts->comp_group_idx[comp_group_idx_ctx][mbmi->comp_group_idx]; #endif - if (allow_update_cdf) { - update_cdf(fc->comp_group_idx_cdf[comp_group_idx_ctx], - mbmi->comp_group_idx, 2); - } - } + update_cdf(fc->comp_group_idx_cdf[comp_group_idx_ctx], + mbmi->comp_group_idx, 2); + } - if (mbmi->comp_group_idx == 0) { - const int comp_index_ctx = get_comp_index_context(cm, xd); + if (mbmi->comp_group_idx == 0) { + const int comp_index_ctx = get_comp_index_context(cm, xd); #if CONFIG_ENTROPY_STATS - ++counts->compound_index[comp_index_ctx][mbmi->compound_idx]; + ++counts->compound_index[comp_index_ctx][mbmi->compound_idx]; #endif - if (allow_update_cdf) { - update_cdf(fc->compound_index_cdf[comp_index_ctx], - mbmi->compound_idx, 2); - } - } else { - assert(masked_compound_used); - if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) { + update_cdf(fc->compound_index_cdf[comp_index_ctx], mbmi->compound_idx, + 2); + } else { + assert(masked_compound_used); + if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) { #if CONFIG_ENTROPY_STATS - ++counts->compound_type[bsize][mbmi->interinter_comp.type - 1]; + ++counts->compound_type[bsize][mbmi->interinter_comp.type - + COMPOUND_WEDGE]; #endif - if (allow_update_cdf) { - update_cdf(fc->compound_type_cdf[bsize], - mbmi->interinter_comp.type - 1, COMPOUND_TYPES - 1); - } - } + update_cdf(fc->compound_type_cdf[bsize], + mbmi->interinter_comp.type - COMPOUND_WEDGE, + MASKED_COMPOUND_TYPES); } } - if (mbmi->interinter_comp.type == COMPOUND_WEDGE) { - if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) { + } + if (mbmi->interinter_comp.type == COMPOUND_WEDGE) { + if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) { #if CONFIG_ENTROPY_STATS - counts->wedge_idx[bsize][mbmi->interinter_comp.wedge_index]++; + counts->wedge_idx[bsize][mbmi->interinter_comp.wedge_index]++; #endif - if (allow_update_cdf) { - update_cdf(fc->wedge_idx_cdf[bsize], - mbmi->interinter_comp.wedge_index, 16); - } - } + update_cdf(fc->wedge_idx_cdf[bsize], + mbmi->interinter_comp.wedge_index, 16); } } } + } - if (inter_block && - !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) { - int16_t mode_ctx; - const PREDICTION_MODE mode = mbmi->mode; - - mode_ctx = - av1_mode_context_analyzer(mbmi_ext->mode_context, mbmi->ref_frame); - if (has_second_ref(mbmi)) { + if (inter_block && cm->features.interp_filter == SWITCHABLE && + mbmi->motion_mode != WARPED_CAUSAL && + !is_nontrans_global_motion(xd, mbmi)) { + update_filter_type_cdf(xd, mbmi); + } + if (inter_block && + !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) { + const PREDICTION_MODE mode = mbmi->mode; + const int16_t mode_ctx = + av1_mode_context_analyzer(mbmi_ext->mode_context, mbmi->ref_frame); + if (has_second_ref(mbmi)) { #if CONFIG_ENTROPY_STATS - ++counts->inter_compound_mode[mode_ctx][INTER_COMPOUND_OFFSET(mode)]; + ++counts->inter_compound_mode[mode_ctx][INTER_COMPOUND_OFFSET(mode)]; #endif - if (allow_update_cdf) - update_cdf(fc->inter_compound_mode_cdf[mode_ctx], - INTER_COMPOUND_OFFSET(mode), INTER_COMPOUND_MODES); - } else { - update_inter_mode_stats(fc, counts, mode, mode_ctx, allow_update_cdf); - } - - int mode_allowed = (mbmi->mode == NEWMV); - mode_allowed |= (mbmi->mode == NEW_NEWMV); - if (mode_allowed) { - uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame); - int idx; + update_cdf(fc->inter_compound_mode_cdf[mode_ctx], + INTER_COMPOUND_OFFSET(mode), INTER_COMPOUND_MODES); + } else { + update_inter_mode_stats(fc, counts, mode, mode_ctx); + } - for (idx = 0; idx < 2; ++idx) { - if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) { + const int new_mv = mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV; + if (new_mv) { + const uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame); + for (int idx = 0; idx < 2; ++idx) { + if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) { + const uint8_t drl_ctx = + av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx); + update_cdf(fc->drl_cdf[drl_ctx], mbmi->ref_mv_idx != idx, 2); #if CONFIG_ENTROPY_STATS - uint8_t drl_ctx = - av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx); - ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx]; + ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx]; #endif - - if (mbmi->ref_mv_idx == idx) break; - } + if (mbmi->ref_mv_idx == idx) break; } } + } - if (have_nearmv_in_inter_mode(mbmi->mode)) { - uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame); - int idx; - - for (idx = 1; idx < 3; ++idx) { - if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) { + if (have_nearmv_in_inter_mode(mbmi->mode)) { + const uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame); + for (int idx = 1; idx < 3; ++idx) { + if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) { + const uint8_t drl_ctx = + av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx); + update_cdf(fc->drl_cdf[drl_ctx], mbmi->ref_mv_idx != idx - 1, 2); #if CONFIG_ENTROPY_STATS - uint8_t drl_ctx = - av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx); - ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx - 1]; + ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx - 1]; #endif - - if (mbmi->ref_mv_idx == idx - 1) break; - } + if (mbmi->ref_mv_idx == idx - 1) break; + } + } + } + if (have_newmv_in_inter_mode(mbmi->mode)) { + const int allow_hp = cm->features.cur_frame_force_integer_mv + ? MV_SUBPEL_NONE + : cm->features.allow_high_precision_mv; + if (new_mv) { + for (int ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) { + const int_mv ref_mv = av1_get_ref_mv(x, ref); + av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc, + allow_hp); } + } else if (mbmi->mode == NEAREST_NEWMV || mbmi->mode == NEAR_NEWMV) { + const int ref = 1; + const int_mv ref_mv = av1_get_ref_mv(x, ref); + av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc, + allow_hp); + } else if (mbmi->mode == NEW_NEARESTMV || mbmi->mode == NEW_NEARMV) { + const int ref = 0; + const int_mv ref_mv = av1_get_ref_mv(x, ref); + av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc, + allow_hp); } } } } -typedef struct { - ENTROPY_CONTEXT a[MAX_MIB_SIZE * MAX_MB_PLANE]; - ENTROPY_CONTEXT l[MAX_MIB_SIZE * MAX_MB_PLANE]; - PARTITION_CONTEXT sa[MAX_MIB_SIZE]; - PARTITION_CONTEXT sl[MAX_MIB_SIZE]; - TXFM_CONTEXT *p_ta; - TXFM_CONTEXT *p_tl; - TXFM_CONTEXT ta[MAX_MIB_SIZE]; - TXFM_CONTEXT tl[MAX_MIB_SIZE]; -} RD_SEARCH_MACROBLOCK_CONTEXT; - -static void restore_context(MACROBLOCK *x, - const RD_SEARCH_MACROBLOCK_CONTEXT *ctx, int mi_row, - int mi_col, BLOCK_SIZE bsize, - const int num_planes) { +static AOM_INLINE void restore_context(MACROBLOCK *x, + const RD_SEARCH_MACROBLOCK_CONTEXT *ctx, + int mi_row, int mi_col, BLOCK_SIZE bsize, + const int num_planes) { MACROBLOCKD *xd = &x->e_mbd; int p; - const int num_4x4_blocks_wide = - block_size_wide[bsize] >> tx_size_wide_log2[0]; - const int num_4x4_blocks_high = - block_size_high[bsize] >> tx_size_high_log2[0]; + const int num_4x4_blocks_wide = mi_size_wide[bsize]; + const int num_4x4_blocks_high = mi_size_high[bsize]; int mi_width = mi_size_wide[bsize]; int mi_height = mi_size_high[bsize]; for (p = 0; p < num_planes; p++) { int tx_col = mi_col; int tx_row = mi_row & MAX_MIB_MASK; - memcpy(xd->above_context[p] + (tx_col >> xd->plane[p].subsampling_x), - ctx->a + num_4x4_blocks_wide * p, - (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >> - xd->plane[p].subsampling_x); - memcpy(xd->left_context[p] + (tx_row >> xd->plane[p].subsampling_y), + memcpy( + xd->above_entropy_context[p] + (tx_col >> xd->plane[p].subsampling_x), + ctx->a + num_4x4_blocks_wide * p, + (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >> + xd->plane[p].subsampling_x); + memcpy(xd->left_entropy_context[p] + (tx_row >> xd->plane[p].subsampling_y), ctx->l + num_4x4_blocks_high * p, (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >> xd->plane[p].subsampling_y); } - memcpy(xd->above_seg_context + mi_col, ctx->sa, - sizeof(*xd->above_seg_context) * mi_width); - memcpy(xd->left_seg_context + (mi_row & MAX_MIB_MASK), ctx->sl, - sizeof(xd->left_seg_context[0]) * mi_height); + memcpy(xd->above_partition_context + mi_col, ctx->sa, + sizeof(*xd->above_partition_context) * mi_width); + memcpy(xd->left_partition_context + (mi_row & MAX_MIB_MASK), ctx->sl, + sizeof(xd->left_partition_context[0]) * mi_height); xd->above_txfm_context = ctx->p_ta; xd->left_txfm_context = ctx->p_tl; memcpy(xd->above_txfm_context, ctx->ta, @@ -1377,15 +1522,12 @@ static void restore_context(MACROBLOCK *x, sizeof(*xd->left_txfm_context) * mi_height); } -static void save_context(const MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *ctx, - int mi_row, int mi_col, BLOCK_SIZE bsize, - const int num_planes) { +static AOM_INLINE void save_context(const MACROBLOCK *x, + RD_SEARCH_MACROBLOCK_CONTEXT *ctx, + int mi_row, int mi_col, BLOCK_SIZE bsize, + const int num_planes) { const MACROBLOCKD *xd = &x->e_mbd; int p; - const int num_4x4_blocks_wide = - block_size_wide[bsize] >> tx_size_wide_log2[0]; - const int num_4x4_blocks_high = - block_size_high[bsize] >> tx_size_high_log2[0]; int mi_width = mi_size_wide[bsize]; int mi_height = mi_size_high[bsize]; @@ -1393,19 +1535,18 @@ static void save_context(const MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *ctx, for (p = 0; p < num_planes; ++p) { int tx_col = mi_col; int tx_row = mi_row & MAX_MIB_MASK; - memcpy(ctx->a + num_4x4_blocks_wide * p, - xd->above_context[p] + (tx_col >> xd->plane[p].subsampling_x), - (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >> - xd->plane[p].subsampling_x); - memcpy(ctx->l + num_4x4_blocks_high * p, - xd->left_context[p] + (tx_row >> xd->plane[p].subsampling_y), - (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >> - xd->plane[p].subsampling_y); - } - memcpy(ctx->sa, xd->above_seg_context + mi_col, - sizeof(*xd->above_seg_context) * mi_width); - memcpy(ctx->sl, xd->left_seg_context + (mi_row & MAX_MIB_MASK), - sizeof(xd->left_seg_context[0]) * mi_height); + memcpy( + ctx->a + mi_width * p, + xd->above_entropy_context[p] + (tx_col >> xd->plane[p].subsampling_x), + (sizeof(ENTROPY_CONTEXT) * mi_width) >> xd->plane[p].subsampling_x); + memcpy(ctx->l + mi_height * p, + xd->left_entropy_context[p] + (tx_row >> xd->plane[p].subsampling_y), + (sizeof(ENTROPY_CONTEXT) * mi_height) >> xd->plane[p].subsampling_y); + } + memcpy(ctx->sa, xd->above_partition_context + mi_col, + sizeof(*xd->above_partition_context) * mi_width); + memcpy(ctx->sl, xd->left_partition_context + (mi_row & MAX_MIB_MASK), + sizeof(xd->left_partition_context[0]) * mi_height); memcpy(ctx->ta, xd->above_txfm_context, sizeof(*xd->above_txfm_context) * mi_width); memcpy(ctx->tl, xd->left_txfm_context, @@ -1414,34 +1555,38 @@ static void save_context(const MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *ctx, ctx->p_tl = xd->left_txfm_context; } -static void encode_b(const AV1_COMP *const cpi, TileDataEnc *tile_data, - ThreadData *td, TOKENEXTRA **tp, int mi_row, int mi_col, - RUN_TYPE dry_run, BLOCK_SIZE bsize, - PARTITION_TYPE partition, - const PICK_MODE_CONTEXT *const ctx, int *rate) { +static AOM_INLINE void encode_b(const AV1_COMP *const cpi, + TileDataEnc *tile_data, ThreadData *td, + TOKENEXTRA **tp, int mi_row, int mi_col, + RUN_TYPE dry_run, BLOCK_SIZE bsize, + PARTITION_TYPE partition, + PICK_MODE_CONTEXT *const ctx, int *rate) { TileInfo *const tile = &tile_data->tile_info; MACROBLOCK *const x = &td->mb; MACROBLOCKD *xd = &x->e_mbd; - set_offsets(cpi, tile, x, mi_row, mi_col, bsize); + set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize); + const int origin_mult = x->rdmult; + setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL); MB_MODE_INFO *mbmi = xd->mi[0]; mbmi->partition = partition; - update_state(cpi, tile_data, td, ctx, mi_row, mi_col, bsize, dry_run); + update_state(cpi, td, ctx, mi_row, mi_col, bsize, dry_run); - if (!dry_run) av1_set_coeff_buffer(cpi, x, mi_row, mi_col); - - encode_superblock(cpi, tile_data, td, tp, dry_run, mi_row, mi_col, bsize, - rate); + if (!dry_run) { + x->mbmi_ext_frame->cb_offset = x->cb_offset; + assert(x->cb_offset < + (1 << num_pels_log2_lookup[cpi->common.seq_params.sb_size])); + } - if (dry_run == 0) - x->cb_offset += block_size_wide[bsize] * block_size_high[bsize]; + encode_superblock(cpi, tile_data, td, tp, dry_run, bsize, rate); if (!dry_run) { + const AV1_COMMON *const cm = &cpi->common; + x->cb_offset += block_size_wide[bsize] * block_size_high[bsize]; if (bsize == cpi->common.seq_params.sb_size && mbmi->skip == 1 && - cpi->common.delta_lf_present_flag) { - const int frame_lf_count = av1_num_planes(&cpi->common) > 1 - ? FRAME_LF_COUNT - : FRAME_LF_COUNT - 2; + cm->delta_q_info.delta_lf_present_flag) { + const int frame_lf_count = + av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2; for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) mbmi->delta_lf[lf_id] = xd->delta_lf[lf_id]; mbmi->delta_lf_from_base = xd->delta_lf_from_base; @@ -1453,17 +1598,107 @@ static void encode_b(const AV1_COMP *const cpi, TileDataEnc *tile_data, else mbmi->comp_group_idx = 1; } - update_stats(&cpi->common, tile_data, td, mi_row, mi_col); + + // delta quant applies to both intra and inter + const int super_block_upper_left = + ((mi_row & (cm->seq_params.mib_size - 1)) == 0) && + ((mi_col & (cm->seq_params.mib_size - 1)) == 0); + const DeltaQInfo *const delta_q_info = &cm->delta_q_info; + if (delta_q_info->delta_q_present_flag && + (bsize != cm->seq_params.sb_size || !mbmi->skip) && + super_block_upper_left) { + xd->current_qindex = mbmi->current_qindex; + if (delta_q_info->delta_lf_present_flag) { + if (delta_q_info->delta_lf_multi) { + const int frame_lf_count = + av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2; + for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) { + xd->delta_lf[lf_id] = mbmi->delta_lf[lf_id]; + } + } else { + xd->delta_lf_from_base = mbmi->delta_lf_from_base; + } + } + } + + RD_COUNTS *rdc = &td->rd_counts; + if (mbmi->skip_mode) { + assert(!frame_is_intra_only(cm)); + rdc->skip_mode_used_flag = 1; + if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT) { + assert(has_second_ref(mbmi)); + rdc->compound_ref_used_flag = 1; + } + set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]); + } else { + const int seg_ref_active = + segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME); + if (!seg_ref_active) { + // If the segment reference feature is enabled we have only a single + // reference frame allowed for the segment so exclude it from + // the reference frame counts used to work out probabilities. + if (is_inter_block(mbmi)) { + av1_collect_neighbors_ref_counts(xd); + if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT) { + if (has_second_ref(mbmi)) { + // This flag is also updated for 4x4 blocks + rdc->compound_ref_used_flag = 1; + } + } + set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]); + } + } + } + + if (tile_data->allow_update_cdf) update_stats(&cpi->common, td); + + // Gather obmc and warped motion count to update the probability. + if ((!cpi->sf.inter_sf.disable_obmc && + cpi->sf.inter_sf.prune_obmc_prob_thresh > 0) || + (cm->features.allow_warped_motion && + cpi->sf.inter_sf.prune_warped_prob_thresh > 0)) { + const int inter_block = is_inter_block(mbmi); + const int seg_ref_active = + segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME); + if (!seg_ref_active && inter_block) { + const MOTION_MODE motion_allowed = + cm->features.switchable_motion_mode + ? motion_mode_allowed(xd->global_motion, xd, mbmi, + cm->features.allow_warped_motion) + : SIMPLE_TRANSLATION; + + if (mbmi->ref_frame[1] != INTRA_FRAME) { + if (motion_allowed >= OBMC_CAUSAL) { + td->rd_counts.obmc_used[bsize][mbmi->motion_mode == OBMC_CAUSAL]++; + } + if (motion_allowed == WARPED_CAUSAL) { + td->rd_counts.warped_used[mbmi->motion_mode == WARPED_CAUSAL]++; + } + } + } + } } + // TODO(Ravi/Remya): Move this copy function to a better logical place + // This function will copy the best mode information from block + // level (x->mbmi_ext) to frame level (cpi->mbmi_ext_info.frame_base). This + // frame level buffer (cpi->mbmi_ext_info.frame_base) will be used during + // bitstream preparation. + av1_copy_mbmi_ext_to_mbmi_ext_frame(x->mbmi_ext_frame, x->mbmi_ext, + av1_ref_frame_type(xd->mi[0]->ref_frame)); + x->rdmult = origin_mult; } -static void encode_sb(const AV1_COMP *const cpi, ThreadData *td, - TileDataEnc *tile_data, TOKENEXTRA **tp, int mi_row, - int mi_col, RUN_TYPE dry_run, BLOCK_SIZE bsize, - PC_TREE *pc_tree, int *rate) { +static AOM_INLINE void encode_sb(const AV1_COMP *const cpi, ThreadData *td, + TileDataEnc *tile_data, TOKENEXTRA **tp, + int mi_row, int mi_col, RUN_TYPE dry_run, + BLOCK_SIZE bsize, PC_TREE *pc_tree, + int *rate) { + assert(bsize < BLOCK_SIZES_ALL); const AV1_COMMON *const cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; MACROBLOCK *const x = &td->mb; MACROBLOCKD *const xd = &x->e_mbd; + assert(bsize < BLOCK_SIZES_ALL); const int hbs = mi_size_wide[bsize] / 2; const int is_partition_root = bsize >= BLOCK_8X8; const int ctx = is_partition_root @@ -1475,11 +1710,11 @@ static void encode_sb(const AV1_COMP *const cpi, ThreadData *td, int i; BLOCK_SIZE bsize2 = get_partition_subsize(bsize, PARTITION_SPLIT); - if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return; + if (mi_row >= mi_params->mi_rows || mi_col >= mi_params->mi_cols) return; if (!dry_run && ctx >= 0) { - const int has_rows = (mi_row + hbs) < cm->mi_rows; - const int has_cols = (mi_col + hbs) < cm->mi_cols; + const int has_rows = (mi_row + hbs) < mi_params->mi_rows; + const int has_cols = (mi_col + hbs) < mi_params->mi_cols; if (has_rows && has_cols) { #if CONFIG_ENTROPY_STATS @@ -1502,7 +1737,7 @@ static void encode_sb(const AV1_COMP *const cpi, ThreadData *td, case PARTITION_VERT: encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize, partition, &pc_tree->vertical[0], rate); - if (mi_col + hbs < cm->mi_cols) { + if (mi_col + hbs < mi_params->mi_cols) { encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, subsize, partition, &pc_tree->vertical[1], rate); } @@ -1510,7 +1745,7 @@ static void encode_sb(const AV1_COMP *const cpi, ThreadData *td, case PARTITION_HORZ: encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize, partition, &pc_tree->horizontal[0], rate); - if (mi_row + hbs < cm->mi_rows) { + if (mi_row + hbs < mi_params->mi_rows) { encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, subsize, partition, &pc_tree->horizontal[1], rate); } @@ -1562,7 +1797,7 @@ static void encode_sb(const AV1_COMP *const cpi, ThreadData *td, case PARTITION_HORZ_4: for (i = 0; i < 4; ++i) { int this_mi_row = mi_row + i * quarter_step; - if (i > 0 && this_mi_row >= cm->mi_rows) break; + if (i > 0 && this_mi_row >= mi_params->mi_rows) break; encode_b(cpi, tile_data, td, tp, this_mi_row, mi_col, dry_run, subsize, partition, &pc_tree->horizontal4[i], rate); @@ -1571,8 +1806,7 @@ static void encode_sb(const AV1_COMP *const cpi, ThreadData *td, case PARTITION_VERT_4: for (i = 0; i < 4; ++i) { int this_mi_col = mi_col + i * quarter_step; - if (i > 0 && this_mi_col >= cm->mi_cols) break; - + if (i > 0 && this_mi_col >= mi_params->mi_cols) break; encode_b(cpi, tile_data, td, tp, mi_row, this_mi_col, dry_run, subsize, partition, &pc_tree->vertical4[i], rate); } @@ -1583,38 +1817,19 @@ static void encode_sb(const AV1_COMP *const cpi, ThreadData *td, update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition); } -// Check to see if the given partition size is allowed for a specified number -// of mi block rows and columns remaining in the image. -// If not then return the largest allowed partition size -static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize, int rows_left, - int cols_left, int *bh, int *bw) { - if (rows_left <= 0 || cols_left <= 0) { - return AOMMIN(bsize, BLOCK_8X8); - } else { - for (; bsize > 0; bsize -= 3) { - *bh = mi_size_high[bsize]; - *bw = mi_size_wide[bsize]; - if ((*bh <= rows_left) && (*bw <= cols_left)) { - break; - } - } - } - return bsize; -} - -static void set_partial_sb_partition(const AV1_COMMON *const cm, - MB_MODE_INFO *mi, int bh_in, int bw_in, - int mi_rows_remaining, - int mi_cols_remaining, BLOCK_SIZE bsize, - MB_MODE_INFO **mib) { +static AOM_INLINE void set_partial_sb_partition( + const AV1_COMMON *const cm, MB_MODE_INFO *mi, int bh_in, int bw_in, + int mi_rows_remaining, int mi_cols_remaining, BLOCK_SIZE bsize, + MB_MODE_INFO **mib) { int bh = bh_in; int r, c; for (r = 0; r < cm->seq_params.mib_size; r += bh) { int bw = bw_in; for (c = 0; c < cm->seq_params.mib_size; c += bw) { - const int index = r * cm->mi_stride + c; - mib[index] = mi + index; - mib[index]->sb_type = find_partition_size( + const int grid_index = get_mi_grid_idx(&cm->mi_params, r, c); + const int mi_index = get_alloc_mi_idx(&cm->mi_params, r, c); + mib[grid_index] = mi + mi_index; + mib[grid_index]->sb_type = find_partition_size( bsize, mi_rows_remaining - r, mi_cols_remaining - c, &bh, &bw); } } @@ -1625,28 +1840,34 @@ static void set_partial_sb_partition(const AV1_COMMON *const cm, // However, at the bottom and right borders of the image the requested size // may not be allowed in which case this code attempts to choose the largest // allowable partition. -static void set_fixed_partitioning(AV1_COMP *cpi, const TileInfo *const tile, - MB_MODE_INFO **mib, int mi_row, int mi_col, - BLOCK_SIZE bsize) { +static AOM_INLINE void set_fixed_partitioning(AV1_COMP *cpi, + const TileInfo *const tile, + MB_MODE_INFO **mib, int mi_row, + int mi_col, BLOCK_SIZE bsize) { AV1_COMMON *const cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; const int mi_rows_remaining = tile->mi_row_end - mi_row; const int mi_cols_remaining = tile->mi_col_end - mi_col; - int block_row, block_col; - MB_MODE_INFO *const mi_upper_left = cm->mi + mi_row * cm->mi_stride + mi_col; + MB_MODE_INFO *const mi_upper_left = + mi_params->mi_alloc + get_alloc_mi_idx(mi_params, mi_row, mi_col); int bh = mi_size_high[bsize]; int bw = mi_size_wide[bsize]; + assert(bsize >= mi_params->mi_alloc_bsize && + "Attempted to use bsize < mi_params->mi_alloc_bsize"); assert((mi_rows_remaining > 0) && (mi_cols_remaining > 0)); // Apply the requested partition size to the SB if it is all "in image" if ((mi_cols_remaining >= cm->seq_params.mib_size) && (mi_rows_remaining >= cm->seq_params.mib_size)) { - for (block_row = 0; block_row < cm->seq_params.mib_size; block_row += bh) { - for (block_col = 0; block_col < cm->seq_params.mib_size; + for (int block_row = 0; block_row < cm->seq_params.mib_size; + block_row += bh) { + for (int block_col = 0; block_col < cm->seq_params.mib_size; block_col += bw) { - int index = block_row * cm->mi_stride + block_col; - mib[index] = mi_upper_left + index; - mib[index]->sb_type = bsize; + const int grid_index = get_mi_grid_idx(mi_params, block_row, block_col); + const int mi_index = get_alloc_mi_idx(mi_params, block_row, block_col); + mib[grid_index] = mi_upper_left + mi_index; + mib[grid_index]->sb_type = bsize; } } } else { @@ -1656,12 +1877,12 @@ static void set_fixed_partitioning(AV1_COMP *cpi, const TileInfo *const tile, } } -static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, - TileDataEnc *tile_data, MB_MODE_INFO **mib, - TOKENEXTRA **tp, int mi_row, int mi_col, - BLOCK_SIZE bsize, int *rate, int64_t *dist, - int do_recon, PC_TREE *pc_tree) { +static AOM_INLINE void rd_use_partition( + AV1_COMP *cpi, ThreadData *td, TileDataEnc *tile_data, MB_MODE_INFO **mib, + TOKENEXTRA **tp, int mi_row, int mi_col, BLOCK_SIZE bsize, int *rate, + int64_t *dist, int do_recon, PC_TREE *pc_tree) { AV1_COMMON *const cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; const int num_planes = av1_num_planes(cm); TileInfo *const tile_info = &tile_data->tile_info; MACROBLOCK *const x = &td->mb; @@ -1677,24 +1898,25 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, : PARTITION_NONE; const BLOCK_SIZE subsize = get_partition_subsize(bsize, partition); RD_SEARCH_MACROBLOCK_CONTEXT x_ctx; - RD_STATS last_part_rdc, none_rdc, chosen_rdc; + RD_STATS last_part_rdc, none_rdc, chosen_rdc, invalid_rdc; BLOCK_SIZE sub_subsize = BLOCK_4X4; int splits_below = 0; BLOCK_SIZE bs_type = mib[0]->sb_type; - int do_partition_search = 1; PICK_MODE_CONTEXT *ctx_none = &pc_tree->none; - if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return; + if (mi_row >= mi_params->mi_rows || mi_col >= mi_params->mi_cols) return; assert(mi_size_wide[bsize] == mi_size_high[bsize]); av1_invalid_rd_stats(&last_part_rdc); av1_invalid_rd_stats(&none_rdc); av1_invalid_rd_stats(&chosen_rdc); + av1_invalid_rd_stats(&invalid_rdc); pc_tree->partitioning = partition; - xd->above_txfm_context = cm->above_txfm_context[tile_info->tile_row] + mi_col; + xd->above_txfm_context = + cm->above_contexts.txfm[tile_info->tile_row] + mi_col; xd->left_txfm_context = xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); @@ -1704,16 +1926,22 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, x->mb_energy = av1_log_block_var(cpi, x, bsize); } - if (do_partition_search && - cpi->sf.partition_search_type == SEARCH_PARTITION && - cpi->sf.adjust_partitioning_from_last_frame) { + // Save rdmult before it might be changed, so it can be restored later. + const int orig_rdmult = x->rdmult; + setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL); + + if (cpi->sf.part_sf.partition_search_type == VAR_BASED_PARTITION && + (cpi->sf.part_sf.adjust_var_based_rd_partitioning == 2 || + (cpi->sf.part_sf.adjust_var_based_rd_partitioning == 1 && + cm->quant_params.base_qindex > 190 && bsize <= BLOCK_32X32 && + !frame_is_intra_only(cm)))) { // Check if any of the sub blocks are further split. if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) { sub_subsize = get_partition_subsize(subsize, PARTITION_SPLIT); splits_below = 1; for (i = 0; i < 4; i++) { int jj = i >> 1, ii = i & 0x01; - MB_MODE_INFO *this_mi = mib[jj * hbs * cm->mi_stride + ii * hbs]; + MB_MODE_INFO *this_mi = mib[jj * hbs * mi_params->mi_stride + ii * hbs]; if (this_mi && this_mi->sb_type >= sub_subsize) { splits_below = 0; } @@ -1723,10 +1951,11 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, // If partition is not none try none unless each of the 4 splits are split // even further.. if (partition != PARTITION_NONE && !splits_below && - mi_row + hbs < cm->mi_rows && mi_col + hbs < cm->mi_cols) { + mi_row + hbs < mi_params->mi_rows && + mi_col + hbs < mi_params->mi_cols) { pc_tree->partitioning = PARTITION_NONE; - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, - PARTITION_NONE, bsize, ctx_none, INT64_MAX); + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, + PARTITION_NONE, bsize, ctx_none, invalid_rdc, PICK_MODE_RD); if (none_rdc.rate < INT_MAX) { none_rdc.rate += x->partition_cost[pl][PARTITION_NONE]; @@ -1738,40 +1967,27 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, pc_tree->partitioning = partition; } } - for (int b = 0; b < 2; ++b) { - pc_tree->horizontal[b].skip_ref_frame_mask = 0; - pc_tree->vertical[b].skip_ref_frame_mask = 0; - } - for (int b = 0; b < 3; ++b) { - pc_tree->horizontala[b].skip_ref_frame_mask = 0; - pc_tree->horizontalb[b].skip_ref_frame_mask = 0; - pc_tree->verticala[b].skip_ref_frame_mask = 0; - pc_tree->verticalb[b].skip_ref_frame_mask = 0; - } - for (int b = 0; b < 4; ++b) { - pc_tree->horizontal4[b].skip_ref_frame_mask = 0; - pc_tree->vertical4[b].skip_ref_frame_mask = 0; - } + switch (partition) { case PARTITION_NONE: - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, - PARTITION_NONE, bsize, ctx_none, INT64_MAX); + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, + PARTITION_NONE, bsize, ctx_none, invalid_rdc, PICK_MODE_RD); break; case PARTITION_HORZ: - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, - PARTITION_HORZ, subsize, &pc_tree->horizontal[0], - INT64_MAX); + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, + PARTITION_HORZ, subsize, &pc_tree->horizontal[0], + invalid_rdc, PICK_MODE_RD); if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 && - mi_row + hbs < cm->mi_rows) { + mi_row + hbs < mi_params->mi_rows) { RD_STATS tmp_rdc; const PICK_MODE_CONTEXT *const ctx_h = &pc_tree->horizontal[0]; av1_init_rd_stats(&tmp_rdc); - update_state(cpi, tile_data, td, ctx_h, mi_row, mi_col, subsize, 1); - encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row, - mi_col, subsize, NULL); - rd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &tmp_rdc, - PARTITION_HORZ, subsize, &pc_tree->horizontal[1], - INT64_MAX); + update_state(cpi, td, ctx_h, mi_row, mi_col, subsize, 1); + encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize, + NULL); + pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &tmp_rdc, + PARTITION_HORZ, subsize, &pc_tree->horizontal[1], + invalid_rdc, PICK_MODE_RD); if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) { av1_invalid_rd_stats(&last_part_rdc); break; @@ -1782,20 +1998,21 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, } break; case PARTITION_VERT: - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, - PARTITION_VERT, subsize, &pc_tree->vertical[0], - INT64_MAX); + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, + PARTITION_VERT, subsize, &pc_tree->vertical[0], invalid_rdc, + PICK_MODE_RD); if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 && - mi_col + hbs < cm->mi_cols) { + mi_col + hbs < mi_params->mi_cols) { RD_STATS tmp_rdc; const PICK_MODE_CONTEXT *const ctx_v = &pc_tree->vertical[0]; av1_init_rd_stats(&tmp_rdc); - update_state(cpi, tile_data, td, ctx_v, mi_row, mi_col, subsize, 1); - encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row, - mi_col, subsize, NULL); - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &tmp_rdc, - PARTITION_VERT, subsize, - &pc_tree->vertical[bsize > BLOCK_8X8], INT64_MAX); + update_state(cpi, td, ctx_v, mi_row, mi_col, subsize, 1); + encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize, + NULL); + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &tmp_rdc, + PARTITION_VERT, subsize, + &pc_tree->vertical[bsize > BLOCK_8X8], invalid_rdc, + PICK_MODE_RD); if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) { av1_invalid_rd_stats(&last_part_rdc); break; @@ -1806,6 +2023,11 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, } break; case PARTITION_SPLIT: + if (cpi->sf.part_sf.adjust_var_based_rd_partitioning == 1 && + none_rdc.rate < INT_MAX && none_rdc.skip == 1) { + av1_invalid_rd_stats(&last_part_rdc); + break; + } last_part_rdc.rate = 0; last_part_rdc.dist = 0; last_part_rdc.rdcost = 0; @@ -1814,12 +2036,13 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, int y_idx = (i >> 1) * hbs; int jj = i >> 1, ii = i & 0x01; RD_STATS tmp_rdc; - if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols)) + if ((mi_row + y_idx >= mi_params->mi_rows) || + (mi_col + x_idx >= mi_params->mi_cols)) continue; av1_init_rd_stats(&tmp_rdc); rd_use_partition(cpi, td, tile_data, - mib + jj * hbs * cm->mi_stride + ii * hbs, tp, + mib + jj * hbs * mi_params->mi_stride + ii * hbs, tp, mi_row + y_idx, mi_col + x_idx, subsize, &tmp_rdc.rate, &tmp_rdc.dist, i != 3, pc_tree->split[i]); if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) { @@ -1846,11 +2069,13 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, RDCOST(x->rdmult, last_part_rdc.rate, last_part_rdc.dist); } - if (do_partition_search && cpi->sf.adjust_partitioning_from_last_frame && - cpi->sf.partition_search_type == SEARCH_PARTITION && + if ((cpi->sf.part_sf.partition_search_type == VAR_BASED_PARTITION && + cpi->sf.part_sf.adjust_var_based_rd_partitioning > 2) && partition != PARTITION_SPLIT && bsize > BLOCK_8X8 && - (mi_row + bs < cm->mi_rows || mi_row + hbs == cm->mi_rows) && - (mi_col + bs < cm->mi_cols || mi_col + hbs == cm->mi_cols)) { + (mi_row + bs < mi_params->mi_rows || + mi_row + hbs == mi_params->mi_rows) && + (mi_col + bs < mi_params->mi_cols || + mi_col + hbs == mi_params->mi_cols)) { BLOCK_SIZE split_subsize = get_partition_subsize(bsize, PARTITION_SPLIT); chosen_rdc.rate = 0; chosen_rdc.dist = 0; @@ -1864,14 +2089,15 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, int y_idx = (i >> 1) * hbs; RD_STATS tmp_rdc; - if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols)) + if ((mi_row + y_idx >= mi_params->mi_rows) || + (mi_col + x_idx >= mi_params->mi_cols)) continue; save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); pc_tree->split[i]->partitioning = PARTITION_NONE; - rd_pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx, - &tmp_rdc, PARTITION_SPLIT, split_subsize, - &pc_tree->split[i]->none, INT64_MAX); + pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx, &tmp_rdc, + PARTITION_SPLIT, split_subsize, &pc_tree->split[i]->none, + invalid_rdc, PICK_MODE_RD); restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) { @@ -1930,88 +2156,299 @@ static void rd_use_partition(AV1_COMP *cpi, ThreadData *td, *rate = chosen_rdc.rate; *dist = chosen_rdc.dist; + x->rdmult = orig_rdmult; } -/* clang-format off */ -static const BLOCK_SIZE min_partition_size[BLOCK_SIZES_ALL] = { - BLOCK_4X4, // 4x4 - BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, // 4x8, 8x4, 8x8 - BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 8x16, 16x8, 16x16 - BLOCK_8X8, BLOCK_8X8, BLOCK_16X16, // 16x32, 32x16, 32x32 - BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, // 32x64, 64x32, 64x64 - BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, // 64x128, 128x64, 128x128 - BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 4x16, 16x4, 8x32 - BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, // 32x8, 16x64, 64x16 -}; +static int is_leaf_split_partition(AV1_COMMON *cm, int mi_row, int mi_col, + BLOCK_SIZE bsize) { + const int bs = mi_size_wide[bsize]; + const int hbs = bs / 2; + assert(bsize >= BLOCK_8X8); + const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT); -static const BLOCK_SIZE max_partition_size[BLOCK_SIZES_ALL] = { - BLOCK_8X8, // 4x4 - BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, // 4x8, 8x4, 8x8 - BLOCK_32X32, BLOCK_32X32, BLOCK_32X32, // 8x16, 16x8, 16x16 - BLOCK_64X64, BLOCK_64X64, BLOCK_64X64, // 16x32, 32x16, 32x32 - BLOCK_LARGEST, BLOCK_LARGEST, BLOCK_LARGEST, // 32x64, 64x32, 64x64 - BLOCK_LARGEST, BLOCK_LARGEST, BLOCK_LARGEST, // 64x128, 128x64, 128x128 - BLOCK_16X16, BLOCK_16X16, BLOCK_32X32, // 4x16, 16x4, 8x32 - BLOCK_32X32, BLOCK_LARGEST, BLOCK_LARGEST, // 32x8, 16x64, 64x16 -}; + for (int i = 0; i < 4; i++) { + int x_idx = (i & 1) * hbs; + int y_idx = (i >> 1) * hbs; + if ((mi_row + y_idx >= cm->mi_params.mi_rows) || + (mi_col + x_idx >= cm->mi_params.mi_cols)) + return 0; + if (get_partition(cm, mi_row + y_idx, mi_col + x_idx, subsize) != + PARTITION_NONE && + subsize != BLOCK_8X8) + return 0; + } + return 1; +} -// Next square block size less or equal than current block size. -static const BLOCK_SIZE next_square_size[BLOCK_SIZES_ALL] = { - BLOCK_4X4, // 4x4 - BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 4x8, 8x4, 8x8 - BLOCK_8X8, BLOCK_8X8, BLOCK_16X16, // 8x16, 16x8, 16x16 - BLOCK_16X16, BLOCK_16X16, BLOCK_32X32, // 16x32, 32x16, 32x32 - BLOCK_32X32, BLOCK_32X32, BLOCK_64X64, // 32x64, 64x32, 64x64 - BLOCK_64X64, BLOCK_64X64, BLOCK_128X128, // 64x128, 128x64, 128x128 - BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, // 4x16, 16x4, 8x32 - BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, // 32x8, 16x64, 64x16 -}; -/* clang-format on */ +static AOM_INLINE int do_slipt_check(BLOCK_SIZE bsize) { + return (bsize == BLOCK_16X16 || bsize == BLOCK_32X32); +} -// Look at all the mode_info entries for blocks that are part of this -// partition and find the min and max values for sb_type. -// At the moment this is designed to work on a superblock but could be -// adjusted to use a size parameter. -// -// The min and max are assumed to have been initialized prior to calling this -// function so repeat calls can accumulate a min and max of more than one -// superblock. -static void get_sb_partition_size_range(const AV1_COMMON *const cm, - MACROBLOCKD *xd, MB_MODE_INFO **mib, - BLOCK_SIZE *min_block_size, - BLOCK_SIZE *max_block_size) { - int i, j; - int index = 0; - - // Check the sb_type for each block that belongs to this region. - for (i = 0; i < cm->seq_params.mib_size; ++i) { - for (j = 0; j < cm->seq_params.mib_size; ++j) { - MB_MODE_INFO *mi = mib[index + j]; - BLOCK_SIZE sb_type = mi ? mi->sb_type : BLOCK_4X4; - *min_block_size = AOMMIN(*min_block_size, sb_type); - *max_block_size = AOMMAX(*max_block_size, sb_type); - } - index += xd->mi_stride; +static AOM_INLINE void nonrd_use_partition(AV1_COMP *cpi, ThreadData *td, + TileDataEnc *tile_data, + MB_MODE_INFO **mib, TOKENEXTRA **tp, + int mi_row, int mi_col, + BLOCK_SIZE bsize, PC_TREE *pc_tree) { + AV1_COMMON *const cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; + TileInfo *const tile_info = &tile_data->tile_info; + MACROBLOCK *const x = &td->mb; + MACROBLOCKD *const xd = &x->e_mbd; + // Only square blocks from 8x8 to 128x128 are supported + assert(bsize >= BLOCK_8X8 && bsize <= BLOCK_128X128); + const int bs = mi_size_wide[bsize]; + const int hbs = bs / 2; + const PARTITION_TYPE partition = + (bsize >= BLOCK_8X8) ? get_partition(cm, mi_row, mi_col, bsize) + : PARTITION_NONE; + BLOCK_SIZE subsize = get_partition_subsize(bsize, partition); + assert(subsize <= BLOCK_LARGEST); + const int pl = (bsize >= BLOCK_8X8) + ? partition_plane_context(xd, mi_row, mi_col, bsize) + : 0; + + RD_STATS dummy_cost; + av1_invalid_rd_stats(&dummy_cost); + RD_STATS invalid_rd; + av1_invalid_rd_stats(&invalid_rd); + + if (mi_row >= mi_params->mi_rows || mi_col >= mi_params->mi_cols) return; + + assert(mi_size_wide[bsize] == mi_size_high[bsize]); + + pc_tree->partitioning = partition; + + xd->above_txfm_context = + cm->above_contexts.txfm[tile_info->tile_row] + mi_col; + xd->left_txfm_context = + xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); + + switch (partition) { + case PARTITION_NONE: + if (cpi->sf.rt_sf.nonrd_check_partition_split && do_slipt_check(bsize) && + !frame_is_intra_only(cm)) { + RD_STATS split_rdc, none_rdc, block_rdc; + RD_SEARCH_MACROBLOCK_CONTEXT x_ctx; + + av1_init_rd_stats(&split_rdc); + av1_invalid_rd_stats(&none_rdc); + + save_context(x, &x_ctx, mi_row, mi_col, bsize, 3); + subsize = get_partition_subsize(bsize, PARTITION_SPLIT); + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, + PARTITION_NONE, bsize, &pc_tree->none, invalid_rd, + PICK_MODE_NONRD); + none_rdc.rate += x->partition_cost[pl][PARTITION_NONE]; + none_rdc.rdcost = RDCOST(x->rdmult, none_rdc.rate, none_rdc.dist); + restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3); + + for (int i = 0; i < 4; i++) { + av1_invalid_rd_stats(&block_rdc); + const int x_idx = (i & 1) * hbs; + const int y_idx = (i >> 1) * hbs; + if (mi_row + y_idx >= mi_params->mi_rows || + mi_col + x_idx >= mi_params->mi_cols) + continue; + xd->above_txfm_context = + cm->above_contexts.txfm[tile_info->tile_row] + mi_col + x_idx; + xd->left_txfm_context = + xd->left_txfm_context_buffer + ((mi_row + y_idx) & MAX_MIB_MASK); + pc_tree->split[i]->partitioning = PARTITION_NONE; + pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx, + &block_rdc, PARTITION_NONE, subsize, + &pc_tree->split[i]->none, invalid_rd, PICK_MODE_NONRD); + split_rdc.rate += block_rdc.rate; + split_rdc.dist += block_rdc.dist; + + encode_b(cpi, tile_data, td, tp, mi_row + y_idx, mi_col + x_idx, 1, + subsize, PARTITION_NONE, &pc_tree->split[i]->none, NULL); + } + split_rdc.rate += x->partition_cost[pl][PARTITION_SPLIT]; + split_rdc.rdcost = RDCOST(x->rdmult, split_rdc.rate, split_rdc.dist); + restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3); + + if (none_rdc.rdcost < split_rdc.rdcost) { + mib[0]->sb_type = bsize; + pc_tree->partitioning = PARTITION_NONE; + encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize, partition, + &pc_tree->none, NULL); + } else { + mib[0]->sb_type = subsize; + pc_tree->partitioning = PARTITION_SPLIT; + for (int i = 0; i < 4; i++) { + const int x_idx = (i & 1) * hbs; + const int y_idx = (i >> 1) * hbs; + if (mi_row + y_idx >= mi_params->mi_rows || + mi_col + x_idx >= mi_params->mi_cols) + continue; + + encode_b(cpi, tile_data, td, tp, mi_row + y_idx, mi_col + x_idx, 0, + subsize, PARTITION_NONE, &pc_tree->split[i]->none, NULL); + } + } + + } else { + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &dummy_cost, + PARTITION_NONE, bsize, &pc_tree->none, invalid_rd, + PICK_MODE_NONRD); + encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize, partition, + &pc_tree->none, NULL); + } + break; + case PARTITION_VERT: + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &dummy_cost, + PARTITION_VERT, subsize, &pc_tree->vertical[0], invalid_rd, + PICK_MODE_NONRD); + encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, subsize, + PARTITION_VERT, &pc_tree->vertical[0], NULL); + if (mi_col + hbs < mi_params->mi_cols && bsize > BLOCK_8X8) { + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &dummy_cost, + PARTITION_VERT, subsize, &pc_tree->vertical[1], + invalid_rd, PICK_MODE_NONRD); + encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, 0, subsize, + PARTITION_VERT, &pc_tree->vertical[1], NULL); + } + break; + case PARTITION_HORZ: + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &dummy_cost, + PARTITION_HORZ, subsize, &pc_tree->horizontal[0], + invalid_rd, PICK_MODE_NONRD); + encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, subsize, + PARTITION_HORZ, &pc_tree->horizontal[0], NULL); + + if (mi_row + hbs < mi_params->mi_rows && bsize > BLOCK_8X8) { + pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &dummy_cost, + PARTITION_HORZ, subsize, &pc_tree->horizontal[1], + invalid_rd, PICK_MODE_NONRD); + encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, 0, subsize, + PARTITION_HORZ, &pc_tree->horizontal[1], NULL); + } + break; + case PARTITION_SPLIT: + if (cpi->sf.rt_sf.nonrd_check_partition_merge_mode && + is_leaf_split_partition(cm, mi_row, mi_col, bsize) && + !frame_is_intra_only(cm) && bsize <= BLOCK_32X32) { + RD_SEARCH_MACROBLOCK_CONTEXT x_ctx; + RD_STATS split_rdc, none_rdc; + av1_invalid_rd_stats(&split_rdc); + av1_invalid_rd_stats(&none_rdc); + save_context(x, &x_ctx, mi_row, mi_col, bsize, 3); + xd->above_txfm_context = + cm->above_contexts.txfm[tile_info->tile_row] + mi_col; + xd->left_txfm_context = + xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); + pc_tree->partitioning = PARTITION_NONE; + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, + PARTITION_NONE, bsize, &pc_tree->none, invalid_rd, + PICK_MODE_NONRD); + none_rdc.rate += x->partition_cost[pl][PARTITION_NONE]; + none_rdc.rdcost = RDCOST(x->rdmult, none_rdc.rate, none_rdc.dist); + restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3); + if (cpi->sf.rt_sf.nonrd_check_partition_merge_mode != 2 || + none_rdc.skip != 1 || pc_tree->none.mic.mode == NEWMV) { + av1_init_rd_stats(&split_rdc); + for (int i = 0; i < 4; i++) { + RD_STATS block_rdc; + av1_invalid_rd_stats(&block_rdc); + int x_idx = (i & 1) * hbs; + int y_idx = (i >> 1) * hbs; + if ((mi_row + y_idx >= mi_params->mi_rows) || + (mi_col + x_idx >= mi_params->mi_cols)) + continue; + xd->above_txfm_context = + cm->above_contexts.txfm[tile_info->tile_row] + mi_col + x_idx; + xd->left_txfm_context = xd->left_txfm_context_buffer + + ((mi_row + y_idx) & MAX_MIB_MASK); + pc_tree->split[i]->partitioning = PARTITION_NONE; + pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx, + &block_rdc, PARTITION_NONE, subsize, + &pc_tree->split[i]->none, invalid_rd, + PICK_MODE_NONRD); + split_rdc.rate += block_rdc.rate; + split_rdc.dist += block_rdc.dist; + + encode_b(cpi, tile_data, td, tp, mi_row + y_idx, mi_col + x_idx, 1, + subsize, PARTITION_NONE, &pc_tree->split[i]->none, NULL); + } + restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3); + split_rdc.rate += x->partition_cost[pl][PARTITION_SPLIT]; + split_rdc.rdcost = RDCOST(x->rdmult, split_rdc.rate, split_rdc.dist); + } + if (none_rdc.rdcost < split_rdc.rdcost) { + mib[0]->sb_type = bsize; + pc_tree->partitioning = PARTITION_NONE; + encode_b(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize, partition, + &pc_tree->none, NULL); + } else { + mib[0]->sb_type = subsize; + pc_tree->partitioning = PARTITION_SPLIT; + for (int i = 0; i < 4; i++) { + int x_idx = (i & 1) * hbs; + int y_idx = (i >> 1) * hbs; + if ((mi_row + y_idx >= mi_params->mi_rows) || + (mi_col + x_idx >= mi_params->mi_cols)) + continue; + + encode_b(cpi, tile_data, td, tp, mi_row + y_idx, mi_col + x_idx, 0, + subsize, PARTITION_NONE, &pc_tree->split[i]->none, NULL); + } + } + } else { + for (int i = 0; i < 4; i++) { + int x_idx = (i & 1) * hbs; + int y_idx = (i >> 1) * hbs; + int jj = i >> 1, ii = i & 0x01; + if ((mi_row + y_idx >= mi_params->mi_rows) || + (mi_col + x_idx >= mi_params->mi_cols)) + continue; + nonrd_use_partition(cpi, td, tile_data, + mib + jj * hbs * mi_params->mi_stride + ii * hbs, + tp, mi_row + y_idx, mi_col + x_idx, subsize, + pc_tree->split[i]); + } + } + break; + case PARTITION_VERT_A: + case PARTITION_VERT_B: + case PARTITION_HORZ_A: + case PARTITION_HORZ_B: + case PARTITION_HORZ_4: + case PARTITION_VERT_4: + assert(0 && "Cannot handle extended partition types"); + default: assert(0); break; } } +#if !CONFIG_REALTIME_ONLY +static const FIRSTPASS_STATS *read_one_frame_stats(const TWO_PASS *p, int frm) { + assert(frm >= 0); + if (frm < 0 || + p->stats_buf_ctx->stats_in_start + frm > p->stats_buf_ctx->stats_in_end) { + return NULL; + } + + return &p->stats_buf_ctx->stats_in_start[frm]; +} // Checks to see if a super block is on a horizontal image edge. // In most cases this is the "real" edge unless there are formatting // bars embedded in the stream. static int active_h_edge(const AV1_COMP *cpi, int mi_row, int mi_step) { int top_edge = 0; - int bottom_edge = cpi->common.mi_rows; + int bottom_edge = cpi->common.mi_params.mi_rows; int is_active_h_edge = 0; // For two pass account for any formatting bars detected. - if (cpi->oxcf.pass == 2) { - const TWO_PASS *const twopass = &cpi->twopass; + if (is_stat_consumption_stage_twopass(cpi)) { + const AV1_COMMON *const cm = &cpi->common; + const FIRSTPASS_STATS *const this_frame_stats = read_one_frame_stats( + &cpi->twopass, cm->current_frame.display_order_hint); + if (this_frame_stats == NULL) return AOM_CODEC_ERROR; // The inactive region is specified in MBs not mi units. // The image edge is in the following MB row. - top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2); + top_edge += (int)(this_frame_stats->inactive_zone_rows * 4); - bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2); + bottom_edge -= (int)(this_frame_stats->inactive_zone_rows * 4); bottom_edge = AOMMAX(top_edge, bottom_edge); } @@ -2027,18 +2464,21 @@ static int active_h_edge(const AV1_COMP *cpi, int mi_row, int mi_step) { // bars embedded in the stream. static int active_v_edge(const AV1_COMP *cpi, int mi_col, int mi_step) { int left_edge = 0; - int right_edge = cpi->common.mi_cols; + int right_edge = cpi->common.mi_params.mi_cols; int is_active_v_edge = 0; // For two pass account for any formatting bars detected. - if (cpi->oxcf.pass == 2) { - const TWO_PASS *const twopass = &cpi->twopass; + if (is_stat_consumption_stage_twopass(cpi)) { + const AV1_COMMON *const cm = &cpi->common; + const FIRSTPASS_STATS *const this_frame_stats = read_one_frame_stats( + &cpi->twopass, cm->current_frame.display_order_hint); + if (this_frame_stats == NULL) return AOM_CODEC_ERROR; // The inactive region is specified in MBs not mi units. // The image edge is in the following MB row. - left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2); + left_edge += (int)(this_frame_stats->inactive_zone_cols * 4); - right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2); + right_edge -= (int)(this_frame_stats->inactive_zone_cols * 4); right_edge = AOMMAX(left_edge, right_edge); } @@ -2048,141 +2488,7 @@ static int active_v_edge(const AV1_COMP *cpi, int mi_col, int mi_step) { } return is_active_v_edge; } - -// Checks to see if a super block is at the edge of the active image. -// In most cases this is the "real" edge unless there are formatting -// bars embedded in the stream. -static int active_edge_sb(const AV1_COMP *cpi, int mi_row, int mi_col) { - return active_h_edge(cpi, mi_row, cpi->common.seq_params.mib_size) || - active_v_edge(cpi, mi_col, cpi->common.seq_params.mib_size); -} - -// Look at neighboring blocks and set a min and max partition size based on -// what they chose. -static void rd_auto_partition_range(AV1_COMP *cpi, const TileInfo *const tile, - MACROBLOCKD *const xd, int mi_row, - int mi_col, BLOCK_SIZE *min_block_size, - BLOCK_SIZE *max_block_size) { - AV1_COMMON *const cm = &cpi->common; - MB_MODE_INFO **mi = xd->mi; - const int left_in_image = xd->left_available && mi[-1]; - const int above_in_image = xd->up_available && mi[-xd->mi_stride]; - const int mi_rows_remaining = tile->mi_row_end - mi_row; - const int mi_cols_remaining = tile->mi_col_end - mi_col; - int bh, bw; - BLOCK_SIZE min_size = BLOCK_4X4; - BLOCK_SIZE max_size = BLOCK_LARGEST; - - // Trap case where we do not have a prediction. - if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) { - // Default "min to max" and "max to min" - min_size = BLOCK_LARGEST; - max_size = BLOCK_4X4; - - // NOTE: each call to get_sb_partition_size_range() uses the previous - // passed in values for min and max as a starting point. - // Find the min and max partition used in previous frame at this location - if (cm->frame_type != KEY_FRAME) { - MB_MODE_INFO **prev_mi = - &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col]; - get_sb_partition_size_range(cm, xd, prev_mi, &min_size, &max_size); - } - // Find the min and max partition sizes used in the left superblock - if (left_in_image) { - MB_MODE_INFO **left_sb_mi = &mi[-cm->seq_params.mib_size]; - get_sb_partition_size_range(cm, xd, left_sb_mi, &min_size, &max_size); - } - // Find the min and max partition sizes used in the above suprblock. - if (above_in_image) { - MB_MODE_INFO **above_sb_mi = - &mi[-xd->mi_stride * cm->seq_params.mib_size]; - get_sb_partition_size_range(cm, xd, above_sb_mi, &min_size, &max_size); - } - - // Adjust observed min and max for "relaxed" auto partition case. - if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) { - min_size = min_partition_size[min_size]; - max_size = max_partition_size[max_size]; - } - } - - // Check border cases where max and min from neighbors may not be legal. - max_size = find_partition_size(max_size, mi_rows_remaining, mi_cols_remaining, - &bh, &bw); - min_size = AOMMIN(min_size, max_size); - - // Test for blocks at the edge of the active image. - // This may be the actual edge of the image or where there are formatting - // bars. - if (active_edge_sb(cpi, mi_row, mi_col)) { - min_size = BLOCK_4X4; - } else { - min_size = AOMMIN(cpi->sf.rd_auto_partition_min_limit, min_size); - } - - // When use_square_partition_only is true, make sure at least one square - // partition is allowed by selecting the next smaller square size as - // *min_block_size. - if (min_size >= cpi->sf.use_square_partition_only_threshold) { - min_size = AOMMIN(min_size, next_square_size[max_size]); - } - - *min_block_size = AOMMIN(min_size, cm->seq_params.sb_size); - *max_block_size = AOMMIN(max_size, cm->seq_params.sb_size); -} - -// TODO(jingning) refactor functions setting partition search range -static void set_partition_range(const AV1_COMMON *const cm, - const MACROBLOCKD *const xd, int mi_row, - int mi_col, BLOCK_SIZE bsize, - BLOCK_SIZE *const min_bs, - BLOCK_SIZE *const max_bs) { - const int mi_width = mi_size_wide[bsize]; - const int mi_height = mi_size_high[bsize]; - int idx, idy; - - const int idx_str = cm->mi_stride * mi_row + mi_col; - MB_MODE_INFO **const prev_mi = &cm->prev_mi_grid_visible[idx_str]; - BLOCK_SIZE min_size = cm->seq_params.sb_size; // default values - BLOCK_SIZE max_size = BLOCK_4X4; - - if (prev_mi) { - for (idy = 0; idy < mi_height; ++idy) { - for (idx = 0; idx < mi_width; ++idx) { - const MB_MODE_INFO *const mi = prev_mi[idy * cm->mi_stride + idx]; - const BLOCK_SIZE bs = mi ? mi->sb_type : bsize; - min_size = AOMMIN(min_size, bs); - max_size = AOMMAX(max_size, bs); - } - } - } - - if (xd->left_available) { - for (idy = 0; idy < mi_height; ++idy) { - const MB_MODE_INFO *const mi = xd->mi[idy * cm->mi_stride - 1]; - const BLOCK_SIZE bs = mi ? mi->sb_type : bsize; - min_size = AOMMIN(min_size, bs); - max_size = AOMMAX(max_size, bs); - } - } - - if (xd->up_available) { - for (idx = 0; idx < mi_width; ++idx) { - const MB_MODE_INFO *const mi = xd->mi[idx - cm->mi_stride]; - const BLOCK_SIZE bs = mi ? mi->sb_type : bsize; - min_size = AOMMIN(min_size, bs); - max_size = AOMMAX(max_size, bs); - } - } - - if (min_size == max_size) { - min_size = min_partition_size[min_size]; - max_size = max_partition_size[max_size]; - } - - *min_bs = AOMMIN(min_size, cm->seq_params.sb_size); - *max_bs = AOMMIN(max_size, cm->seq_params.sb_size); -} +#endif // !CONFIG_REALTIME_ONLY static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) { memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv)); @@ -2193,104 +2499,53 @@ static INLINE void load_pred_mv(MACROBLOCK *x, memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv)); } -#if CONFIG_FP_MB_STATS -const int qindex_skip_threshold_lookup[BLOCK_SIZES] = { - 0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120, - // TODO(debargha): What are the correct numbers here? - 130, 130, 150 -}; -const int qindex_split_threshold_lookup[BLOCK_SIZES] = { - 0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120, - // TODO(debargha): What are the correct numbers here? - 160, 160, 240 -}; -const int complexity_16x16_blocks_threshold[BLOCK_SIZES] = { - 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6, - // TODO(debargha): What are the correct numbers here? - 8, 8, 10 -}; - -typedef enum { - MV_ZERO = 0, - MV_LEFT = 1, - MV_UP = 2, - MV_RIGHT = 3, - MV_DOWN = 4, - MV_INVALID -} MOTION_DIRECTION; - -static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) { - if (fp_byte & FPMB_MOTION_ZERO_MASK) { - return MV_ZERO; - } else if (fp_byte & FPMB_MOTION_LEFT_MASK) { - return MV_LEFT; - } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) { - return MV_RIGHT; - } else if (fp_byte & FPMB_MOTION_UP_MASK) { - return MV_UP; - } else { - return MV_DOWN; - } -} - -static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv, - MOTION_DIRECTION that_mv) { - if (this_mv == that_mv) { - return 0; - } else { - return abs(this_mv - that_mv) == 2 ? 2 : 1; - } -} -#endif - +#if !CONFIG_REALTIME_ONLY // Try searching for an encoding for the given subblock. Returns zero if the // rdcost is already too high (to tell the caller not to bother searching for // encodings of further subblocks) static int rd_try_subblock(AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data, TOKENEXTRA **tp, int is_last, int mi_row, int mi_col, BLOCK_SIZE subsize, - RD_STATS *best_rdc, RD_STATS *sum_rdc, - RD_STATS *this_rdc, PARTITION_TYPE partition, + RD_STATS best_rdcost, RD_STATS *sum_rdc, + PARTITION_TYPE partition, PICK_MODE_CONTEXT *prev_ctx, PICK_MODE_CONTEXT *this_ctx) { -#define RTS_X_RATE_NOCOEF_ARG -#define RTS_MAX_RDCOST best_rdc->rdcost - MACROBLOCK *const x = &td->mb; + const int orig_mult = x->rdmult; + setup_block_rdmult(cpi, x, mi_row, mi_col, subsize, NO_AQ, NULL); - if (cpi->sf.adaptive_motion_search) load_pred_mv(x, prev_ctx); + av1_rd_cost_update(x->rdmult, &best_rdcost); + if (cpi->sf.mv_sf.adaptive_motion_search) load_pred_mv(x, prev_ctx); - const int64_t rdcost_remaining = best_rdc->rdcost == INT64_MAX - ? INT64_MAX - : (best_rdc->rdcost - sum_rdc->rdcost); + RD_STATS rdcost_remaining; + av1_rd_stats_subtraction(x->rdmult, &best_rdcost, sum_rdc, &rdcost_remaining); + RD_STATS this_rdc; + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, partition, + subsize, this_ctx, rdcost_remaining, PICK_MODE_RD); - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, this_rdc, - RTS_X_RATE_NOCOEF_ARG partition, subsize, this_ctx, - rdcost_remaining); - - if (this_rdc->rate == INT_MAX) { + if (this_rdc.rate == INT_MAX) { sum_rdc->rdcost = INT64_MAX; } else { - sum_rdc->rate += this_rdc->rate; - sum_rdc->dist += this_rdc->dist; - sum_rdc->rdcost += this_rdc->rdcost; + sum_rdc->rate += this_rdc.rate; + sum_rdc->dist += this_rdc.dist; + av1_rd_cost_update(x->rdmult, sum_rdc); } - if (sum_rdc->rdcost >= RTS_MAX_RDCOST) return 0; + if (sum_rdc->rdcost >= best_rdcost.rdcost) { + x->rdmult = orig_mult; + return 0; + } if (!is_last) { - update_state(cpi, tile_data, td, this_ctx, mi_row, mi_col, subsize, 1); - encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row, mi_col, - subsize, NULL); + update_state(cpi, td, this_ctx, mi_row, mi_col, subsize, 1); + encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize, NULL); } + x->rdmult = orig_mult; return 1; - -#undef RTS_X_RATE_NOCOEF_ARG -#undef RTS_MAX_RDCOST } -static void rd_test_partition3(AV1_COMP *const cpi, ThreadData *td, +static bool rd_test_partition3(AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data, TOKENEXTRA **tp, PC_TREE *pc_tree, RD_STATS *best_rdc, PICK_MODE_CONTEXT ctxs[3], @@ -2299,52 +2554,38 @@ static void rd_test_partition3(AV1_COMP *const cpi, ThreadData *td, int mi_row0, int mi_col0, BLOCK_SIZE subsize0, int mi_row1, int mi_col1, BLOCK_SIZE subsize1, int mi_row2, int mi_col2, BLOCK_SIZE subsize2) { - MACROBLOCK *const x = &td->mb; - MACROBLOCKD *const xd = &x->e_mbd; - RD_STATS sum_rdc, this_rdc; -#define RTP_STX_TRY_ARGS - int pl = partition_plane_context(xd, mi_row, mi_col, bsize); + const MACROBLOCK *const x = &td->mb; + const MACROBLOCKD *const xd = &x->e_mbd; + const int pl = partition_plane_context(xd, mi_row, mi_col, bsize); + RD_STATS sum_rdc; av1_init_rd_stats(&sum_rdc); sum_rdc.rate = x->partition_cost[pl][partition]; sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, 0); if (!rd_try_subblock(cpi, td, tile_data, tp, 0, mi_row0, mi_col0, subsize0, - best_rdc, &sum_rdc, &this_rdc, - RTP_STX_TRY_ARGS partition, ctx, &ctxs[0])) - return; + *best_rdc, &sum_rdc, partition, ctx, &ctxs[0])) + return false; if (!rd_try_subblock(cpi, td, tile_data, tp, 0, mi_row1, mi_col1, subsize1, - best_rdc, &sum_rdc, &this_rdc, - RTP_STX_TRY_ARGS partition, &ctxs[0], &ctxs[1])) - return; - - // With the new layout of mixed partitions for PARTITION_HORZ_B and - // PARTITION_VERT_B, the last subblock might start past halfway through the - // main block, so we might signal it even though the subblock lies strictly - // outside the image. In that case, we won't spend any bits coding it and the - // difference (obviously) doesn't contribute to the error. - const int try_block2 = 1; - if (try_block2 && - !rd_try_subblock(cpi, td, tile_data, tp, 1, mi_row2, mi_col2, subsize2, - best_rdc, &sum_rdc, &this_rdc, - RTP_STX_TRY_ARGS partition, &ctxs[1], &ctxs[2])) - return; + *best_rdc, &sum_rdc, partition, &ctxs[0], &ctxs[1])) + return false; - if (sum_rdc.rdcost >= best_rdc->rdcost) return; + if (!rd_try_subblock(cpi, td, tile_data, tp, 1, mi_row2, mi_col2, subsize2, + *best_rdc, &sum_rdc, partition, &ctxs[1], &ctxs[2])) + return false; + av1_rd_cost_update(x->rdmult, &sum_rdc); + if (sum_rdc.rdcost >= best_rdc->rdcost) return false; sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist); - - if (sum_rdc.rdcost >= best_rdc->rdcost) return; + if (sum_rdc.rdcost >= best_rdc->rdcost) return false; *best_rdc = sum_rdc; pc_tree->partitioning = partition; - -#undef RTP_STX_TRY_ARGS + return true; } -static void reset_partition(PC_TREE *pc_tree, BLOCK_SIZE bsize) { +static AOM_INLINE void reset_partition(PC_TREE *pc_tree, BLOCK_SIZE bsize) { pc_tree->partitioning = PARTITION_NONE; - pc_tree->cb_search_range = SEARCH_FULL_PLANE; - pc_tree->none.skip = 0; + pc_tree->none.rd_stats.skip = 0; if (bsize >= BLOCK_8X8) { BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT); @@ -2353,836 +2594,93 @@ static void reset_partition(PC_TREE *pc_tree, BLOCK_SIZE bsize) { } } -static void rd_pick_sqr_partition(AV1_COMP *const cpi, ThreadData *td, - TileDataEnc *tile_data, TOKENEXTRA **tp, - int mi_row, int mi_col, BLOCK_SIZE bsize, - RD_STATS *rd_cost, int64_t best_rd, - PC_TREE *pc_tree, int64_t *none_rd) { - const AV1_COMMON *const cm = &cpi->common; - TileInfo *const tile_info = &tile_data->tile_info; - MACROBLOCK *const x = &td->mb; - MACROBLOCKD *const xd = &x->e_mbd; - const int mi_step = mi_size_wide[bsize] / 2; - RD_SEARCH_MACROBLOCK_CONTEXT x_ctx; - const TOKENEXTRA *const tp_orig = *tp; - PICK_MODE_CONTEXT *ctx_none = &pc_tree->none; - int tmp_partition_cost[PARTITION_TYPES]; - BLOCK_SIZE subsize; - RD_STATS this_rdc, sum_rdc, best_rdc, pn_rdc; - const int bsize_at_least_8x8 = (bsize >= BLOCK_8X8); - int do_square_split = bsize_at_least_8x8; - const int pl = bsize_at_least_8x8 - ? partition_plane_context(xd, mi_row, mi_col, bsize) - : 0; - const int *partition_cost = - pl >= 0 ? x->partition_cost[pl] : x->partition_cost[0]; - const int num_planes = av1_num_planes(cm); - - int64_t split_rd[4] = { 0, 0, 0, 0 }; - - // Override skipping rectangular partition operations for edge blocks - const int has_rows = (mi_row + mi_step < cm->mi_rows); - const int has_cols = (mi_col + mi_step < cm->mi_cols); - - if (none_rd) *none_rd = 0; - - int partition_none_allowed = has_rows && has_cols; - - (void)*tp_orig; - (void)split_rd; - - if (best_rd < 0) { - pc_tree->none.rdcost = INT64_MAX; - pc_tree->none.skip = 0; - av1_invalid_rd_stats(rd_cost); - return; - } - pc_tree->pc_tree_stats.valid = 1; - - // Override partition costs at the edges of the frame in the same - // way as in read_partition (see decodeframe.c) - if (!(has_rows && has_cols)) { - assert(bsize_at_least_8x8 && pl >= 0); - const aom_cdf_prob *partition_cdf = cm->fc->partition_cdf[pl]; - for (int i = 0; i < PARTITION_TYPES; ++i) tmp_partition_cost[i] = INT_MAX; - if (has_cols) { - // At the bottom, the two possibilities are HORZ and SPLIT - aom_cdf_prob bot_cdf[2]; - partition_gather_vert_alike(bot_cdf, partition_cdf, bsize); - static const int bot_inv_map[2] = { PARTITION_HORZ, PARTITION_SPLIT }; - av1_cost_tokens_from_cdf(tmp_partition_cost, bot_cdf, bot_inv_map); - } else if (has_rows) { - // At the right, the two possibilities are VERT and SPLIT - aom_cdf_prob rhs_cdf[2]; - partition_gather_horz_alike(rhs_cdf, partition_cdf, bsize); - static const int rhs_inv_map[2] = { PARTITION_VERT, PARTITION_SPLIT }; - av1_cost_tokens_from_cdf(tmp_partition_cost, rhs_cdf, rhs_inv_map); - } else { - // At the bottom right, we always split - tmp_partition_cost[PARTITION_SPLIT] = 0; - } - - partition_cost = tmp_partition_cost; - } - -#ifndef NDEBUG - // Nothing should rely on the default value of this array (which is just - // leftover from encoding the previous block. Setting it to fixed pattern - // when debugging. - // bit 0, 1, 2 are blk_skip of each plane - // bit 4, 5, 6 are initialization checking of each plane - memset(x->blk_skip, 0x77, sizeof(x->blk_skip)); -#endif // NDEBUG - +// Record the ref frames that have been selected by square partition blocks. +static AOM_INLINE void update_picked_ref_frames_mask(MACROBLOCK *const x, + int ref_type, + BLOCK_SIZE bsize, + int mib_size, int mi_row, + int mi_col) { assert(mi_size_wide[bsize] == mi_size_high[bsize]); - - av1_init_rd_stats(&this_rdc); - av1_init_rd_stats(&sum_rdc); - av1_invalid_rd_stats(&best_rdc); - best_rdc.rdcost = best_rd; - - set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize); - - if (bsize == BLOCK_16X16 && cpi->vaq_refresh) - x->mb_energy = av1_log_block_var(cpi, x, bsize); - - xd->above_txfm_context = cm->above_txfm_context[tile_info->tile_row] + mi_col; - xd->left_txfm_context = - xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); - save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); - -#if CONFIG_DIST_8X8 - if (x->using_dist_8x8) { - if (block_size_high[bsize] <= 8 || block_size_wide[bsize] <= 8) - do_square_split = 0; - } -#endif - - // PARTITION_NONE - if (partition_none_allowed) { - int pt_cost = 0; - if (bsize_at_least_8x8) { - pc_tree->partitioning = PARTITION_NONE; - pt_cost = partition_cost[PARTITION_NONE] < INT_MAX - ? partition_cost[PARTITION_NONE] - : 0; + const int sb_size_mask = mib_size - 1; + const int mi_row_in_sb = mi_row & sb_size_mask; + const int mi_col_in_sb = mi_col & sb_size_mask; + const int mi_size = mi_size_wide[bsize]; + for (int i = mi_row_in_sb; i < mi_row_in_sb + mi_size; ++i) { + for (int j = mi_col_in_sb; j < mi_col_in_sb + mi_size; ++j) { + x->picked_ref_frames_mask[i * 32 + j] |= 1 << ref_type; } - int64_t partition_rd_cost = RDCOST(x->rdmult, pt_cost, 0); - int64_t best_remain_rdcost = best_rdc.rdcost == INT64_MAX - ? INT64_MAX - : (best_rdc.rdcost - partition_rd_cost); - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, - PARTITION_NONE, bsize, ctx_none, best_remain_rdcost); - - pc_tree->pc_tree_stats.rdcost = ctx_none->rdcost; - pc_tree->pc_tree_stats.skip = ctx_none->skip; - - if (none_rd) *none_rd = this_rdc.rdcost; - if (this_rdc.rate != INT_MAX) { - if (bsize_at_least_8x8) { - this_rdc.rate += pt_cost; - this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist); - } - - if (this_rdc.rdcost < best_rdc.rdcost) { - // Adjust dist breakout threshold according to the partition size. - const int64_t dist_breakout_thr = - cpi->sf.partition_search_breakout_dist_thr >> - ((2 * (MAX_SB_SIZE_LOG2 - 2)) - - (mi_size_wide_log2[bsize] + mi_size_high_log2[bsize])); - const int rate_breakout_thr = - cpi->sf.partition_search_breakout_rate_thr * - num_pels_log2_lookup[bsize]; - - best_rdc = this_rdc; - if (bsize_at_least_8x8) pc_tree->partitioning = PARTITION_NONE; - - pc_tree->cb_search_range = SEARCH_FULL_PLANE; - - // If all y, u, v transform blocks in this partition are skippable, and - // the dist & rate are within the thresholds, the partition search is - // terminated for current branch of the partition search tree. - // The dist & rate thresholds are set to 0 at speed 0 to disable the - // early termination at that speed. - if (!x->e_mbd.lossless[xd->mi[0]->segment_id] && - (ctx_none->skippable && best_rdc.dist < dist_breakout_thr && - best_rdc.rate < rate_breakout_thr)) { - do_square_split = 0; - } - } - } - - restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); - } - - // store estimated motion vector - if (cpi->sf.adaptive_motion_search) store_pred_mv(x, ctx_none); - - int64_t temp_best_rdcost = best_rdc.rdcost; - pn_rdc = best_rdc; - - // PARTITION_SPLIT - if (do_square_split) { - int reached_last_index = 0; - subsize = get_partition_subsize(bsize, PARTITION_SPLIT); - int idx; - - for (idx = 0; idx < 4 && sum_rdc.rdcost < temp_best_rdcost; ++idx) { - const int x_idx = (idx & 1) * mi_step; - const int y_idx = (idx >> 1) * mi_step; - - if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols) - continue; - - if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none); - - pc_tree->split[idx]->index = idx; - int64_t *p_split_rd = &split_rd[idx]; - // TODO(Cherma) : Account for partition cost while passing best rd to - // rd_pick_sqr_partition() - rd_pick_sqr_partition(cpi, td, tile_data, tp, mi_row + y_idx, - mi_col + x_idx, subsize, &this_rdc, - temp_best_rdcost - sum_rdc.rdcost, - pc_tree->split[idx], p_split_rd); - - pc_tree->pc_tree_stats.sub_block_rdcost[idx] = this_rdc.rdcost; - pc_tree->pc_tree_stats.sub_block_skip[idx] = - pc_tree->split[idx]->none.skip; - - if (this_rdc.rate == INT_MAX) { - sum_rdc.rdcost = INT64_MAX; - break; - } else { - sum_rdc.rate += this_rdc.rate; - sum_rdc.dist += this_rdc.dist; - sum_rdc.rdcost += this_rdc.rdcost; - } - } - reached_last_index = (idx == 4); - - if (reached_last_index && sum_rdc.rdcost < best_rdc.rdcost) { - sum_rdc.rate += partition_cost[PARTITION_SPLIT]; - sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist); - - if (sum_rdc.rdcost < best_rdc.rdcost) { - best_rdc = sum_rdc; - pc_tree->partitioning = PARTITION_SPLIT; - } - } - - int has_split = 0; - if (pc_tree->partitioning == PARTITION_SPLIT) { - for (int cb_idx = 0; cb_idx <= AOMMIN(idx, 3); ++cb_idx) { - if (pc_tree->split[cb_idx]->partitioning == PARTITION_SPLIT) - ++has_split; - } - - if (has_split >= 3 || sum_rdc.rdcost < (pn_rdc.rdcost >> 1)) { - pc_tree->cb_search_range = SPLIT_PLANE; - } - } - - if (pc_tree->partitioning == PARTITION_NONE) { - pc_tree->cb_search_range = SEARCH_SAME_PLANE; - if (pn_rdc.dist <= sum_rdc.dist) - pc_tree->cb_search_range = NONE_PARTITION_PLANE; - } - - if (pn_rdc.rate == INT_MAX) pc_tree->cb_search_range = NONE_PARTITION_PLANE; - - restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); - } // if (do_split) - - pc_tree->pc_tree_stats.split = pc_tree->partitioning == PARTITION_SPLIT; - if (do_square_split) { - for (int i = 0; i < 4; ++i) { - pc_tree->pc_tree_stats.sub_block_split[i] = - pc_tree->split[i]->partitioning == PARTITION_SPLIT; - } - } - - // TODO(jbb): This code added so that we avoid static analysis - // warning related to the fact that best_rd isn't used after this - // point. This code should be refactored so that the duplicate - // checks occur in some sub function and thus are used... - (void)best_rd; - *rd_cost = best_rdc; - - if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && - pc_tree->index != 3) { - if (bsize == cm->seq_params.sb_size) { - restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); - } else { - encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, DRY_RUN_NORMAL, bsize, - pc_tree, NULL); - } - } - - if (bsize == cm->seq_params.sb_size) { - assert(best_rdc.rate < INT_MAX); - assert(best_rdc.dist < INT64_MAX); - } else { - assert(tp_orig == *tp); } } -#define FEATURE_SIZE 19 -static const float two_pass_split_partition_weights_128[FEATURE_SIZE + 1] = { - 2.683936f, -0.193620f, -4.106470f, -0.141320f, -0.282289f, - 0.125296f, -1.134961f, 0.862757f, -0.418799f, -0.637666f, - 0.016232f, 0.345013f, 0.018823f, -0.393394f, -1.130700f, - 0.695357f, 0.112569f, -0.341975f, -0.513882f, 5.7488966f, -}; - -static const float two_pass_split_partition_weights_64[FEATURE_SIZE + 1] = { - 2.990993f, 0.423273f, -0.926544f, 0.454646f, -0.292698f, - -1.311632f, -0.284432f, 0.717141f, -0.419257f, -0.574760f, - -0.674444f, 0.669047f, -0.374255f, 0.380624f, -0.804036f, - 0.264021f, 0.004163f, 1.896802f, 0.924287f, 0.13490619f, -}; - -static const float two_pass_split_partition_weights_32[FEATURE_SIZE + 1] = { - 2.795181f, -0.136943f, -0.924842f, 0.405330f, -0.463505f, - -0.584076f, -0.831472f, 0.382985f, -0.597544f, -0.138915f, - -1.354350f, 0.466035f, -0.553961f, 0.213202f, -1.166429f, - 0.010776f, -0.096236f, 2.335084f, 1.699857f, -0.58178353f, -}; - -static const float two_pass_split_partition_weights_16[FEATURE_SIZE + 1] = { - 1.987888f, -0.431100f, -1.687703f, 0.262602f, -0.425298f, - -0.463870f, -1.493457f, 0.470917f, -0.528457f, -0.087700f, - -1.815092f, 0.152883f, -0.337908f, 0.093679f, -1.548267f, - -0.042387f, -0.000861f, 2.556746f, 1.619192f, 0.03643292f, -}; - -static const float two_pass_split_partition_weights_8[FEATURE_SIZE + 1] = { - 2.188344f, -0.817528f, -2.119219f, 0.000000f, -0.348167f, - -0.658074f, -1.960362f, 0.000000f, -0.403080f, 0.282699f, - -2.061088f, 0.000000f, -0.431919f, -0.127960f, -1.099550f, - 0.000000f, 0.121622f, 2.017455f, 2.058228f, -0.15475988f, -}; - -static const float two_pass_none_partition_weights_128[FEATURE_SIZE + 1] = { - -1.006689f, 0.777908f, 4.461072f, -0.395782f, -0.014610f, - -0.853863f, 0.729997f, -0.420477f, 0.282429f, -1.194595f, - 3.181220f, -0.511416f, 0.117084f, -1.149348f, 1.507990f, - -0.477212f, 0.202963f, -1.469581f, 0.624461f, -0.89081228f, -}; - -static const float two_pass_none_partition_weights_64[FEATURE_SIZE + 1] = { - -1.241117f, 0.844878f, 5.638803f, -0.489780f, -0.108796f, - -4.576821f, 1.540624f, -0.477519f, 0.227791f, -1.443968f, - 1.586911f, -0.505125f, 0.140764f, -0.464194f, 1.466658f, - -0.641166f, 0.195412f, 1.427905f, 2.080007f, -1.98272777f, -}; - -static const float two_pass_none_partition_weights_32[FEATURE_SIZE + 1] = { - -2.130825f, 0.476023f, 5.907343f, -0.516002f, -0.097471f, - -2.662754f, 0.614858f, -0.576728f, 0.085261f, -0.031901f, - 0.727842f, -0.600034f, 0.079326f, 0.324328f, 0.504502f, - -0.547105f, -0.037670f, 0.304995f, 0.369018f, -2.66299987f, -}; - -static const float two_pass_none_partition_weights_16[FEATURE_SIZE + 1] = { - -1.626410f, 0.872047f, 5.414965f, -0.554781f, -0.084514f, - -3.020550f, 0.467632f, -0.382280f, 0.199568f, 0.426220f, - 0.829426f, -0.467100f, 0.153098f, 0.662994f, 0.327545f, - -0.560106f, -0.141610f, 0.403372f, 0.523991f, -3.02891231f, -}; - -static const float two_pass_none_partition_weights_8[FEATURE_SIZE + 1] = { - -1.463349f, 0.375376f, 4.751430f, 0.000000f, -0.184451f, - -1.655447f, 0.443214f, 0.000000f, 0.127961f, 0.152435f, - 0.083288f, 0.000000f, 0.143105f, 0.438012f, 0.073238f, - 0.000000f, -0.278137f, 0.186134f, 0.073737f, -1.6494962f, -}; - -// split_score indicates confidence of picking split partition; -// none_score indicates confidence of picking none partition; -static int ml_prune_2pass_split_partition(const PC_TREE_STATS *pc_tree_stats, - BLOCK_SIZE bsize, int *split_score, - int *none_score) { - if (!pc_tree_stats->valid) return 0; - const float *split_weights = NULL; - const float *none_weights = NULL; - switch (bsize) { - case BLOCK_4X4: break; - case BLOCK_8X8: - split_weights = two_pass_split_partition_weights_8; - none_weights = two_pass_none_partition_weights_8; - break; - case BLOCK_16X16: - split_weights = two_pass_split_partition_weights_16; - none_weights = two_pass_none_partition_weights_16; - break; - case BLOCK_32X32: - split_weights = two_pass_split_partition_weights_32; - none_weights = two_pass_none_partition_weights_32; - break; - case BLOCK_64X64: - split_weights = two_pass_split_partition_weights_64; - none_weights = two_pass_none_partition_weights_64; - break; - case BLOCK_128X128: - split_weights = two_pass_split_partition_weights_128; - none_weights = two_pass_none_partition_weights_128; - break; - default: assert(0 && "Unexpected bsize."); +// Structure to keep win flags for HORZ and VERT partition evaluations +typedef struct { + bool horz_win; + bool vert_win; +} RD_RECT_PART_WIN_INFO; + +// Decide whether to evaluate the AB partition specified by part_type based on +// split and HORZ/VERT info +int evaluate_ab_partition_based_on_split( + PC_TREE *pc_tree, PARTITION_TYPE rect_part, + RD_RECT_PART_WIN_INFO *rect_part_win_info, int qindex, int split_idx1, + int split_idx2) { + int num_win = 0; + // Threshold for number of winners + // Conservative pruning for high quantizers + const int num_win_thresh = AOMMIN(3 * (2 * (MAXQ - qindex) / MAXQ), 3); + bool sub_part_win = (rect_part_win_info == NULL) + ? (pc_tree->partitioning == rect_part) + : (rect_part == PARTITION_HORZ) + ? rect_part_win_info->horz_win + : rect_part_win_info->vert_win; + num_win += (sub_part_win) ? 1 : 0; + num_win += + (pc_tree->split[split_idx1]->partitioning == PARTITION_NONE) ? 1 : 0; + num_win += + (pc_tree->split[split_idx2]->partitioning == PARTITION_NONE) ? 1 : 0; + if (num_win < num_win_thresh) { + return 0; } - if (!split_weights || !none_weights) return 0; - - aom_clear_system_state(); - - float features[FEATURE_SIZE]; - int feature_index = 0; - features[feature_index++] = (float)pc_tree_stats->split; - features[feature_index++] = (float)pc_tree_stats->skip; - const int rdcost = (int)AOMMIN(INT_MAX, pc_tree_stats->rdcost); - const int rd_valid = rdcost > 0 && rdcost < 1000000000; - features[feature_index++] = (float)rd_valid; - for (int i = 0; i < 4; ++i) { - features[feature_index++] = (float)pc_tree_stats->sub_block_split[i]; - features[feature_index++] = (float)pc_tree_stats->sub_block_skip[i]; - const int sub_rdcost = - (int)AOMMIN(INT_MAX, pc_tree_stats->sub_block_rdcost[i]); - const int sub_rd_valid = sub_rdcost > 0 && sub_rdcost < 1000000000; - features[feature_index++] = (float)sub_rd_valid; - // Ratio between the sub-block RD and the whole-block RD. - float rd_ratio = 1.0f; - if (rd_valid && sub_rd_valid && sub_rdcost < rdcost) - rd_ratio = (float)sub_rdcost / (float)rdcost; - features[feature_index++] = rd_ratio; - } - assert(feature_index == FEATURE_SIZE); - - float score_1 = split_weights[FEATURE_SIZE]; - float score_2 = none_weights[FEATURE_SIZE]; - for (int i = 0; i < FEATURE_SIZE; ++i) { - score_1 += features[i] * split_weights[i]; - score_2 += features[i] * none_weights[i]; - } - *split_score = (int)(score_1 * 100); - *none_score = (int)(score_2 * 100); return 1; } -#undef FEATURE_SIZE - -static void ml_prune_rect_partition(const AV1_COMP *const cpi, - const MACROBLOCK *const x, BLOCK_SIZE bsize, - int64_t best_rd, int64_t none_rd, - int64_t *split_rd, - int *const dst_prune_horz, - int *const dst_prune_vert) { - if (bsize < BLOCK_8X8 || best_rd >= 1000000000) return; - best_rd = AOMMAX(best_rd, 1); - const NN_CONFIG *nn_config = NULL; - const float prob_thresholds[5] = { 0.01f, 0.01f, 0.004f, 0.002f, 0.002f }; - float cur_thresh = 0.0f; - switch (bsize) { - case BLOCK_8X8: - nn_config = &av1_rect_partition_nnconfig_8; - cur_thresh = prob_thresholds[0]; - break; - case BLOCK_16X16: - nn_config = &av1_rect_partition_nnconfig_16; - cur_thresh = prob_thresholds[1]; - break; - case BLOCK_32X32: - nn_config = &av1_rect_partition_nnconfig_32; - cur_thresh = prob_thresholds[2]; - break; - case BLOCK_64X64: - nn_config = &av1_rect_partition_nnconfig_64; - cur_thresh = prob_thresholds[3]; - break; - case BLOCK_128X128: - nn_config = &av1_rect_partition_nnconfig_128; - cur_thresh = prob_thresholds[4]; - break; - default: assert(0 && "Unexpected bsize."); - } - if (!nn_config) return; - aom_clear_system_state(); - - // 1. Compute input features - float features[9]; - - // RD cost ratios - for (int i = 0; i < 5; i++) features[i] = 1.0f; - if (none_rd > 0 && none_rd < 1000000000) - features[0] = (float)none_rd / (float)best_rd; - for (int i = 0; i < 4; i++) { - if (split_rd[i] > 0 && split_rd[i] < 1000000000) - features[1 + i] = (float)split_rd[i] / (float)best_rd; - } - - // Variance ratios - const MACROBLOCKD *const xd = &x->e_mbd; - int whole_block_variance; - if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { - whole_block_variance = av1_high_get_sby_perpixel_variance( - cpi, &x->plane[0].src, bsize, xd->bd); - } else { - whole_block_variance = - av1_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize); - } - whole_block_variance = AOMMAX(whole_block_variance, 1); - - int split_variance[4]; - const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT); - struct buf_2d buf; - buf.stride = x->plane[0].src.stride; - const int bw = block_size_wide[bsize]; - for (int i = 0; i < 4; ++i) { - const int x_idx = (i & 1) * bw / 2; - const int y_idx = (i >> 1) * bw / 2; - buf.buf = x->plane[0].src.buf + x_idx + y_idx * buf.stride; - if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { - split_variance[i] = - av1_high_get_sby_perpixel_variance(cpi, &buf, subsize, xd->bd); - } else { - split_variance[i] = av1_get_sby_perpixel_variance(cpi, &buf, subsize); - } - } - - for (int i = 0; i < 4; i++) - features[5 + i] = (float)split_variance[i] / (float)whole_block_variance; - - // 2. Do the prediction and prune 0-2 partitions based on their probabilities - float raw_scores[3] = { 0.0f }; - av1_nn_predict(features, nn_config, raw_scores); - float probs[3] = { 0.0f }; - av1_nn_softmax(raw_scores, probs, 3); - - // probs[0] is the probability of the fact that both rectangular partitions - // are worse than current best_rd - if (probs[1] <= cur_thresh) (*dst_prune_horz) = 1; - if (probs[2] <= cur_thresh) (*dst_prune_vert) = 1; -} - -// Use a ML model to predict if horz_a, horz_b, vert_a, and vert_b should be -// considered. -static void ml_prune_ab_partition(BLOCK_SIZE bsize, int part_ctx, int var_ctx, - int64_t best_rd, int64_t horz_rd[2], - int64_t vert_rd[2], int64_t split_rd[4], - int *const horza_partition_allowed, - int *const horzb_partition_allowed, - int *const verta_partition_allowed, - int *const vertb_partition_allowed) { - if (bsize < BLOCK_8X8 || best_rd >= 1000000000) return; - const NN_CONFIG *nn_config = NULL; - switch (bsize) { - case BLOCK_8X8: nn_config = NULL; break; - case BLOCK_16X16: nn_config = &av1_ab_partition_nnconfig_16; break; - case BLOCK_32X32: nn_config = &av1_ab_partition_nnconfig_32; break; - case BLOCK_64X64: nn_config = &av1_ab_partition_nnconfig_64; break; - case BLOCK_128X128: nn_config = &av1_ab_partition_nnconfig_128; break; - default: assert(0 && "Unexpected bsize."); - } - if (!nn_config) return; - - aom_clear_system_state(); - - // Generate features. - float features[10]; - int feature_index = 0; - features[feature_index++] = (float)part_ctx; - features[feature_index++] = (float)var_ctx; - const int rdcost = (int)AOMMIN(INT_MAX, best_rd); - int sub_block_rdcost[8] = { 0 }; - int rd_index = 0; - for (int i = 0; i < 2; ++i) { - if (horz_rd[i] > 0 && horz_rd[i] < 1000000000) - sub_block_rdcost[rd_index] = (int)horz_rd[i]; - ++rd_index; - } - for (int i = 0; i < 2; ++i) { - if (vert_rd[i] > 0 && vert_rd[i] < 1000000000) - sub_block_rdcost[rd_index] = (int)vert_rd[i]; - ++rd_index; - } - for (int i = 0; i < 4; ++i) { - if (split_rd[i] > 0 && split_rd[i] < 1000000000) - sub_block_rdcost[rd_index] = (int)split_rd[i]; - ++rd_index; - } - for (int i = 0; i < 8; ++i) { - // Ratio between the sub-block RD and the whole-block RD. - float rd_ratio = 1.0f; - if (sub_block_rdcost[i] > 0 && sub_block_rdcost[i] < rdcost) - rd_ratio = (float)sub_block_rdcost[i] / (float)rdcost; - features[feature_index++] = rd_ratio; - } - assert(feature_index == 10); - - // Calculate scores using the NN model. - float score[16] = { 0.0f }; - av1_nn_predict(features, nn_config, score); - int int_score[16]; - int max_score = -1000; - for (int i = 0; i < 16; ++i) { - int_score[i] = (int)(100 * score[i]); - max_score = AOMMAX(int_score[i], max_score); - } - - // Make decisions based on the model scores. - int thresh = max_score; - switch (bsize) { - case BLOCK_16X16: thresh -= 150; break; - case BLOCK_32X32: thresh -= 100; break; - default: break; - } - *horza_partition_allowed = 0; - *horzb_partition_allowed = 0; - *verta_partition_allowed = 0; - *vertb_partition_allowed = 0; - for (int i = 0; i < 16; ++i) { - if (int_score[i] >= thresh) { - if ((i >> 0) & 1) *horza_partition_allowed = 1; - if ((i >> 1) & 1) *horzb_partition_allowed = 1; - if ((i >> 2) & 1) *verta_partition_allowed = 1; - if ((i >> 3) & 1) *vertb_partition_allowed = 1; - } - } -} - -#define FEATURES 18 -#define LABELS 4 -// Use a ML model to predict if horz4 and vert4 should be considered. -static void ml_prune_4_partition(const AV1_COMP *const cpi, MACROBLOCK *const x, - BLOCK_SIZE bsize, int part_ctx, - int64_t best_rd, int64_t horz_rd[2], - int64_t vert_rd[2], int64_t split_rd[4], - int *const partition_horz4_allowed, - int *const partition_vert4_allowed, - unsigned int pb_source_variance, int mi_row, - int mi_col) { - if (best_rd >= 1000000000) return; - const NN_CONFIG *nn_config = NULL; - switch (bsize) { - case BLOCK_16X16: nn_config = &av1_4_partition_nnconfig_16; break; - case BLOCK_32X32: nn_config = &av1_4_partition_nnconfig_32; break; - case BLOCK_64X64: nn_config = &av1_4_partition_nnconfig_64; break; - default: assert(0 && "Unexpected bsize."); - } - if (!nn_config) return; - - aom_clear_system_state(); - - // Generate features. - float features[FEATURES]; - int feature_index = 0; - features[feature_index++] = (float)part_ctx; - features[feature_index++] = (float)get_unsigned_bits(pb_source_variance); - - const int rdcost = (int)AOMMIN(INT_MAX, best_rd); - int sub_block_rdcost[8] = { 0 }; - int rd_index = 0; - for (int i = 0; i < 2; ++i) { - if (horz_rd[i] > 0 && horz_rd[i] < 1000000000) - sub_block_rdcost[rd_index] = (int)horz_rd[i]; - ++rd_index; - } - for (int i = 0; i < 2; ++i) { - if (vert_rd[i] > 0 && vert_rd[i] < 1000000000) - sub_block_rdcost[rd_index] = (int)vert_rd[i]; - ++rd_index; - } - for (int i = 0; i < 4; ++i) { - if (split_rd[i] > 0 && split_rd[i] < 1000000000) - sub_block_rdcost[rd_index] = (int)split_rd[i]; - ++rd_index; - } - for (int i = 0; i < 8; ++i) { - // Ratio between the sub-block RD and the whole-block RD. - float rd_ratio = 1.0f; - if (sub_block_rdcost[i] > 0 && sub_block_rdcost[i] < rdcost) - rd_ratio = (float)sub_block_rdcost[i] / (float)rdcost; - features[feature_index++] = rd_ratio; - } - - // Get variance of the 1:4 and 4:1 sub-blocks. - unsigned int horz_4_source_var[4] = { 0 }; - unsigned int vert_4_source_var[4] = { 0 }; - { - BLOCK_SIZE horz_4_bs = get_partition_subsize(bsize, PARTITION_HORZ_4); - BLOCK_SIZE vert_4_bs = get_partition_subsize(bsize, PARTITION_VERT_4); - av1_setup_src_planes(x, cpi->source, mi_row, mi_col, - av1_num_planes(&cpi->common)); - const int src_stride = x->plane[0].src.stride; - const uint8_t *src = x->plane[0].src.buf; - const MACROBLOCKD *const xd = &x->e_mbd; - for (int i = 0; i < 4; ++i) { - const uint8_t *horz_src = - src + i * block_size_high[horz_4_bs] * src_stride; - const uint8_t *vert_src = src + i * block_size_wide[vert_4_bs]; - unsigned int horz_var, vert_var, sse; - if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { - switch (xd->bd) { - case 10: - horz_var = cpi->fn_ptr[horz_4_bs].vf( - horz_src, src_stride, CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_10), - 0, &sse); - vert_var = cpi->fn_ptr[vert_4_bs].vf( - vert_src, src_stride, CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_10), - 0, &sse); - break; - case 12: - horz_var = cpi->fn_ptr[horz_4_bs].vf( - horz_src, src_stride, CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_12), - 0, &sse); - vert_var = cpi->fn_ptr[vert_4_bs].vf( - vert_src, src_stride, CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_12), - 0, &sse); - break; - case 8: - default: - horz_var = cpi->fn_ptr[horz_4_bs].vf( - horz_src, src_stride, CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_8), - 0, &sse); - vert_var = cpi->fn_ptr[vert_4_bs].vf( - vert_src, src_stride, CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_8), - 0, &sse); - break; - } - horz_4_source_var[i] = - ROUND_POWER_OF_TWO(horz_var, num_pels_log2_lookup[horz_4_bs]); - vert_4_source_var[i] = - ROUND_POWER_OF_TWO(vert_var, num_pels_log2_lookup[vert_4_bs]); - } else { - horz_var = cpi->fn_ptr[horz_4_bs].vf(horz_src, src_stride, AV1_VAR_OFFS, - 0, &sse); - vert_var = cpi->fn_ptr[vert_4_bs].vf(vert_src, src_stride, AV1_VAR_OFFS, - 0, &sse); - horz_4_source_var[i] = - ROUND_POWER_OF_TWO(horz_var, num_pels_log2_lookup[horz_4_bs]); - vert_4_source_var[i] = - ROUND_POWER_OF_TWO(vert_var, num_pels_log2_lookup[vert_4_bs]); - } - } - } - const float denom = (float)(pb_source_variance + 1); - const float low_b = 0.1f; - const float high_b = 10.0f; - for (int i = 0; i < 4; ++i) { - // Ratio between the 4:1 sub-block variance and the whole-block variance. - float var_ratio = (float)(horz_4_source_var[i] + 1) / denom; - if (var_ratio < low_b) var_ratio = low_b; - if (var_ratio > high_b) var_ratio = high_b; - features[feature_index++] = var_ratio; - } - for (int i = 0; i < 4; ++i) { - // Ratio between the 1:4 sub-block RD and the whole-block RD. - float var_ratio = (float)(vert_4_source_var[i] + 1) / denom; - if (var_ratio < low_b) var_ratio = low_b; - if (var_ratio > high_b) var_ratio = high_b; - features[feature_index++] = var_ratio; - } - assert(feature_index == FEATURES); - - // Calculate scores using the NN model. - float score[LABELS] = { 0.0f }; - av1_nn_predict(features, nn_config, score); - int int_score[LABELS]; - int max_score = -1000; - for (int i = 0; i < LABELS; ++i) { - int_score[i] = (int)(100 * score[i]); - max_score = AOMMAX(int_score[i], max_score); - } - - // Make decisions based on the model scores. - int thresh = max_score; - switch (bsize) { - case BLOCK_16X16: thresh -= 500; break; - case BLOCK_32X32: thresh -= 500; break; - case BLOCK_64X64: thresh -= 200; break; - default: break; - } - *partition_horz4_allowed = 0; - *partition_vert4_allowed = 0; - for (int i = 0; i < LABELS; ++i) { - if (int_score[i] >= thresh) { - if ((i >> 0) & 1) *partition_horz4_allowed = 1; - if ((i >> 1) & 1) *partition_vert4_allowed = 1; - } - } -} -#undef FEATURES -#undef LABELS - -#define FEATURES 4 -// ML-based partition search breakout. -static int ml_predict_breakout(const AV1_COMP *const cpi, BLOCK_SIZE bsize, - const MACROBLOCK *const x, - const RD_STATS *const rd_stats, - unsigned int pb_source_variance) { - const NN_CONFIG *nn_config = NULL; - int thresh = 0; - switch (bsize) { - case BLOCK_8X8: - nn_config = &av1_partition_breakout_nnconfig_8; - thresh = cpi->sf.ml_partition_search_breakout_thresh[0]; - break; - case BLOCK_16X16: - nn_config = &av1_partition_breakout_nnconfig_16; - thresh = cpi->sf.ml_partition_search_breakout_thresh[1]; - break; - case BLOCK_32X32: - nn_config = &av1_partition_breakout_nnconfig_32; - thresh = cpi->sf.ml_partition_search_breakout_thresh[2]; - break; - case BLOCK_64X64: - nn_config = &av1_partition_breakout_nnconfig_64; - thresh = cpi->sf.ml_partition_search_breakout_thresh[3]; - break; - case BLOCK_128X128: - nn_config = &av1_partition_breakout_nnconfig_128; - thresh = cpi->sf.ml_partition_search_breakout_thresh[4]; - break; - default: assert(0 && "Unexpected bsize."); - } - if (!nn_config || thresh < 0) return 0; - - // Generate feature values. - float features[FEATURES]; - int feature_index = 0; - aom_clear_system_state(); - - const int num_pels_log2 = num_pels_log2_lookup[bsize]; - float rate_f = (float)AOMMIN(rd_stats->rate, INT_MAX); - rate_f = ((float)x->rdmult / 128.0f / 512.0f / (float)(1 << num_pels_log2)) * - rate_f; - features[feature_index++] = rate_f; - - const float dist_f = - (float)(AOMMIN(rd_stats->dist, INT_MAX) >> num_pels_log2); - features[feature_index++] = dist_f; - - features[feature_index++] = (float)pb_source_variance; - - const int dc_q = (int)x->plane[0].dequant_QTX[0]; - features[feature_index++] = (float)(dc_q * dc_q) / 256.0f; - assert(feature_index == FEATURES); - - // Calculate score using the NN model. - float score = 0.0f; - av1_nn_predict(features, nn_config, &score); - - // Make decision. - return (int)(score * 100) >= thresh; -} -#undef FEATURES - -// TODO(jingning,jimbankoski,rbultje): properly skip partition types that are -// unlikely to be selected depending on previous rate-distortion optimization -// results, for encoding speed-up. -static void rd_pick_partition(AV1_COMP *const cpi, ThreadData *td, +// Searches for the best partition pattern for a block based on the +// rate-distortion cost, and returns a bool value to indicate whether a valid +// partition pattern is found. The partition can recursively go down to +// the smallest block size. +// +// Inputs: +// cpi: the global compressor setting +// td: thread data +// tile_data: tile data +// tp: the pointer to the start token +// mi_row: row coordinate of the block in a step size of MI_SIZE +// mi_col: column coordinate of the block in a step size of MI_SIZE +// bsize: block size +// max_sq_part: the largest square block size for prediction blocks +// min_sq_part: the smallest square block size for prediction blocks +// rd_cost: the pointer to the final rd cost of the current block +// best_rdc: the upper bound of rd cost for a valid partition +// pc_tree: the pointer to the PC_TREE node storing the picked partitions +// and mode info for the current block +// none_rd: the pointer to the rd cost in the case of not splitting the +// current block +// multi_pass_mode: SB_SINGLE_PASS/SB_DRY_PASS/SB_WET_PASS +// rect_part_win_info: the pointer to a struct storing whether horz/vert +// partition outperforms previously tested partitions +// +// Output: +// a bool value indicating whether a valid partition is found +static bool rd_pick_partition(AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data, TOKENEXTRA **tp, int mi_row, int mi_col, BLOCK_SIZE bsize, - RD_STATS *rd_cost, int64_t best_rd, - PC_TREE *pc_tree, int64_t *none_rd) { + BLOCK_SIZE max_sq_part, BLOCK_SIZE min_sq_part, + RD_STATS *rd_cost, RD_STATS best_rdc, + PC_TREE *pc_tree, int64_t *none_rd, + SB_MULTI_PASS_MODE multi_pass_mode, + RD_RECT_PART_WIN_INFO *rect_part_win_info) { const AV1_COMMON *const cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; const int num_planes = av1_num_planes(cm); TileInfo *const tile_info = &tile_data->tile_info; MACROBLOCK *const x = &td->mb; @@ -3193,62 +2691,82 @@ static void rd_pick_partition(AV1_COMP *const cpi, ThreadData *td, PICK_MODE_CONTEXT *ctx_none = &pc_tree->none; int tmp_partition_cost[PARTITION_TYPES]; BLOCK_SIZE subsize; - RD_STATS this_rdc, sum_rdc, best_rdc; + RD_STATS this_rdc, sum_rdc; const int bsize_at_least_8x8 = (bsize >= BLOCK_8X8); int do_square_split = bsize_at_least_8x8; const int pl = bsize_at_least_8x8 ? partition_plane_context(xd, mi_row, mi_col, bsize) : 0; - const int *partition_cost = - pl >= 0 ? x->partition_cost[pl] : x->partition_cost[0]; + const int *partition_cost = x->partition_cost[pl]; - int do_rectangular_split = 1; + int do_rectangular_split = cpi->oxcf.enable_rect_partitions; int64_t cur_none_rd = 0; int64_t split_rd[4] = { 0, 0, 0, 0 }; int64_t horz_rd[2] = { 0, 0 }; int64_t vert_rd[2] = { 0, 0 }; + int prune_horz = 0; + int prune_vert = 0; + int terminate_partition_search = 0; int split_ctx_is_ready[2] = { 0, 0 }; int horz_ctx_is_ready = 0; int vert_ctx_is_ready = 0; BLOCK_SIZE bsize2 = get_partition_subsize(bsize, PARTITION_SPLIT); + // Initialise HORZ and VERT win flags as true for all split partitions + RD_RECT_PART_WIN_INFO split_part_rect_win[4] = { + { true, true }, { true, true }, { true, true }, { true, true } + }; - if (best_rd < 0) { - pc_tree->none.rdcost = INT64_MAX; - pc_tree->none.skip = 0; + bool found_best_partition = false; + if (best_rdc.rdcost < 0) { av1_invalid_rd_stats(rd_cost); - return; + return found_best_partition; + } + + if (frame_is_intra_only(cm) && bsize == BLOCK_64X64) { + x->quad_tree_idx = 0; + x->cnn_output_valid = 0; } + if (bsize == cm->seq_params.sb_size) x->must_find_valid_partition = 0; // Override skipping rectangular partition operations for edge blocks - const int has_rows = (mi_row + mi_step < cm->mi_rows); - const int has_cols = (mi_col + mi_step < cm->mi_cols); + const int has_rows = (mi_row + mi_step < mi_params->mi_rows); + const int has_cols = (mi_col + mi_step < mi_params->mi_cols); const int xss = x->e_mbd.plane[1].subsampling_x; const int yss = x->e_mbd.plane[1].subsampling_y; - BLOCK_SIZE min_size = x->min_partition_size; - BLOCK_SIZE max_size = x->max_partition_size; - if (none_rd) *none_rd = 0; - -#if CONFIG_FP_MB_STATS - unsigned int src_diff_var = UINT_MAX; - int none_complexity = 0; -#endif - int partition_none_allowed = has_rows && has_cols; - int partition_horz_allowed = has_cols && yss <= xss && bsize_at_least_8x8; - int partition_vert_allowed = has_rows && xss <= yss && bsize_at_least_8x8; + int partition_horz_allowed = + has_cols && bsize_at_least_8x8 && cpi->oxcf.enable_rect_partitions && + get_plane_block_size(get_partition_subsize(bsize, PARTITION_HORZ), xss, + yss) != BLOCK_INVALID; + int partition_vert_allowed = + has_rows && bsize_at_least_8x8 && cpi->oxcf.enable_rect_partitions && + get_plane_block_size(get_partition_subsize(bsize, PARTITION_VERT), xss, + yss) != BLOCK_INVALID; (void)*tp_orig; +#if CONFIG_COLLECT_PARTITION_STATS + int partition_decisions[EXT_PARTITION_TYPES] = { 0 }; + int partition_attempts[EXT_PARTITION_TYPES] = { 0 }; + int64_t partition_times[EXT_PARTITION_TYPES] = { 0 }; + struct aom_usec_timer partition_timer = { 0 }; + int partition_timer_on = 0; +#if CONFIG_COLLECT_PARTITION_STATS == 2 + PartitionStats *part_stats = &cpi->partition_stats; +#endif +#endif + // Override partition costs at the edges of the frame in the same // way as in read_partition (see decodeframe.c) if (!(has_rows && has_cols)) { assert(bsize_at_least_8x8 && pl >= 0); const aom_cdf_prob *partition_cdf = cm->fc->partition_cdf[pl]; - for (int i = 0; i < PARTITION_TYPES; ++i) tmp_partition_cost[i] = INT_MAX; + const int max_cost = av1_cost_symbol(0); + for (int i = 0; i < PARTITION_TYPES; ++i) tmp_partition_cost[i] = max_cost; if (has_cols) { // At the bottom, the two possibilities are HORZ and SPLIT aom_cdf_prob bot_cdf[2]; @@ -3281,229 +2799,209 @@ static void rd_pick_partition(AV1_COMP *const cpi, ThreadData *td, assert(mi_size_wide[bsize] == mi_size_high[bsize]); av1_init_rd_stats(&this_rdc); - av1_invalid_rd_stats(&best_rdc); - best_rdc.rdcost = best_rd; set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize); - if (bsize == BLOCK_16X16 && cpi->vaq_refresh) - x->mb_energy = av1_log_block_var(cpi, x, bsize); - - if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) { - const int cb_partition_search_ctrl = - ((pc_tree->index == 0 || pc_tree->index == 3) + - get_chessboard_index(cm->current_video_frame)) & - 0x1; + // Save rdmult before it might be changed, so it can be restored later. + const int orig_rdmult = x->rdmult; + setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL); - if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size) - set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size); - } + av1_rd_cost_update(x->rdmult, &best_rdc); - // Determine partition types in search according to the speed features. - // The threshold set here has to be of square block size. - if (cpi->sf.auto_min_max_partition_size) { - const int no_partition_allowed = (bsize <= max_size && bsize >= min_size); - // Note: Further partitioning is NOT allowed when bsize == min_size already. - const int partition_allowed = (bsize <= max_size && bsize > min_size); - partition_none_allowed &= no_partition_allowed; - partition_horz_allowed &= partition_allowed || !has_rows; - partition_vert_allowed &= partition_allowed || !has_cols; - do_square_split &= bsize > min_size; - } + if (bsize == BLOCK_16X16 && cpi->vaq_refresh) + x->mb_energy = av1_log_block_var(cpi, x, bsize); - if (bsize > cpi->sf.use_square_partition_only_threshold) { + if (bsize > cpi->sf.part_sf.use_square_partition_only_threshold) { partition_horz_allowed &= !has_rows; partition_vert_allowed &= !has_cols; } - if (bsize > BLOCK_4X4 && x->use_cb_search_range && - cpi->sf.auto_min_max_partition_size == 0) { - int split_score = 0; - int none_score = 0; - const int score_valid = ml_prune_2pass_split_partition( - &pc_tree->pc_tree_stats, bsize, &split_score, &none_score); - if (score_valid) { - { - const int only_split_thresh = 300; - const int no_none_thresh = 250; - const int no_split_thresh = 0; - if (split_score > only_split_thresh) { - partition_none_allowed = 0; - partition_horz_allowed = 0; - partition_vert_allowed = 0; - } else if (split_score > no_none_thresh) { - partition_none_allowed = 0; - } - if (split_score < no_split_thresh) do_square_split = 0; - } - { - const int no_split_thresh = 120; - const int no_none_thresh = -120; - if (none_score > no_split_thresh && partition_none_allowed) - do_square_split = 0; - if (none_score < no_none_thresh) partition_none_allowed = 0; - } - } else { - if (pc_tree->cb_search_range == SPLIT_PLANE) { - partition_none_allowed = 0; - partition_horz_allowed = 0; - partition_vert_allowed = 0; - } - if (pc_tree->cb_search_range == SEARCH_SAME_PLANE) do_square_split = 0; - if (pc_tree->cb_search_range == NONE_PARTITION_PLANE) { - do_square_split = 0; - partition_horz_allowed = 0; - partition_vert_allowed = 0; - } - } - - // Fall back to default values in case all partition modes are rejected. - if (partition_none_allowed == 0 && do_square_split == 0 && - partition_horz_allowed == 0 && partition_vert_allowed == 0) { - do_square_split = bsize_at_least_8x8; - partition_none_allowed = has_rows && has_cols; - partition_horz_allowed = has_cols && yss <= xss && bsize_at_least_8x8; - partition_vert_allowed = has_rows && xss <= yss && bsize_at_least_8x8; - } - } - - xd->above_txfm_context = cm->above_txfm_context[tile_info->tile_row] + mi_col; + xd->above_txfm_context = + cm->above_contexts.txfm[tile_info->tile_row] + mi_col; xd->left_txfm_context = xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); -#if CONFIG_FP_MB_STATS - if (cpi->use_fp_mb_stats) { - set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize); - src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src, mi_row, - mi_col, bsize); - } - - // Decide whether we shall split directly and skip searching NONE by using - // the first pass block statistics - if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_square_split && - partition_none_allowed && src_diff_var > 4 && - cm->base_qindex < qindex_split_threshold_lookup[bsize]) { - int mb_row = mi_row >> 1; - int mb_col = mi_col >> 1; - int mb_row_end = - AOMMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows); - int mb_col_end = - AOMMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols); - int r, c; - - // compute a complexity measure, basically measure inconsistency of motion - // vectors obtained from the first pass in the current block - for (r = mb_row; r < mb_row_end; r++) { - for (c = mb_col; c < mb_col_end; c++) { - const int mb_index = r * cm->mb_cols + c; - - MOTION_DIRECTION this_mv; - MOTION_DIRECTION right_mv; - MOTION_DIRECTION bottom_mv; - - this_mv = - get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]); - - // to its right - if (c != mb_col_end - 1) { - right_mv = get_motion_direction_fp( - cpi->twopass.this_frame_mb_stats[mb_index + 1]); - none_complexity += get_motion_inconsistency(this_mv, right_mv); - } - - // to its bottom - if (r != mb_row_end - 1) { - bottom_mv = get_motion_direction_fp( - cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]); - none_complexity += get_motion_inconsistency(this_mv, bottom_mv); - } - - // do not count its left and top neighbors to avoid double counting - } - } - - if (none_complexity > complexity_16x16_blocks_threshold[bsize]) { - partition_none_allowed = 0; - } + const int try_intra_cnn_split = + !cpi->is_screen_content_type && frame_is_intra_only(cm) && + cpi->sf.part_sf.intra_cnn_split && + cm->seq_params.sb_size >= BLOCK_64X64 && bsize <= BLOCK_64X64 && + bsize >= BLOCK_8X8 && + mi_row + mi_size_high[bsize] <= mi_params->mi_rows && + mi_col + mi_size_wide[bsize] <= mi_params->mi_cols; + + if (try_intra_cnn_split) { + av1_intra_mode_cnn_partition( + &cpi->common, x, bsize, x->quad_tree_idx, &partition_none_allowed, + &partition_horz_allowed, &partition_vert_allowed, &do_rectangular_split, + &do_square_split); + } + + // Use simple_motion_search to prune partitions. This must be done prior to + // PARTITION_SPLIT to propagate the initial mvs to a smaller blocksize. + const int try_split_only = + !cpi->is_screen_content_type && + cpi->sf.part_sf.simple_motion_search_split && do_square_split && + bsize >= BLOCK_8X8 && + mi_row + mi_size_high[bsize] <= mi_params->mi_rows && + mi_col + mi_size_wide[bsize] <= mi_params->mi_cols && + !frame_is_intra_only(cm) && !av1_superres_scaled(cm); + + if (try_split_only) { + av1_simple_motion_search_based_split( + cpi, x, pc_tree, mi_row, mi_col, bsize, &partition_none_allowed, + &partition_horz_allowed, &partition_vert_allowed, &do_rectangular_split, + &do_square_split); + } + + const int try_prune_rect = + !cpi->is_screen_content_type && + cpi->sf.part_sf.simple_motion_search_prune_rect && + !frame_is_intra_only(cm) && do_rectangular_split && + (do_square_split || partition_none_allowed || + (prune_horz && prune_vert)) && + (partition_horz_allowed || partition_vert_allowed) && bsize >= BLOCK_8X8; + + if (try_prune_rect) { + av1_simple_motion_search_prune_rect( + cpi, x, pc_tree, mi_row, mi_col, bsize, &partition_horz_allowed, + &partition_vert_allowed, &prune_horz, &prune_vert); + } + + // Max and min square partition levels are defined as the partition nodes that + // the recursive function rd_pick_partition() can reach. To implement this: + // only PARTITION_NONE is allowed if the current node equals min_sq_part, + // only PARTITION_SPLIT is allowed if the current node exceeds max_sq_part. + assert(block_size_wide[min_sq_part] == block_size_high[min_sq_part]); + assert(block_size_wide[max_sq_part] == block_size_high[max_sq_part]); + assert(min_sq_part <= max_sq_part); + assert(block_size_wide[bsize] == block_size_high[bsize]); + const int max_partition_size = block_size_wide[max_sq_part]; + const int min_partition_size = block_size_wide[min_sq_part]; + const int blksize = block_size_wide[bsize]; + assert(min_partition_size <= max_partition_size); + const int is_le_min_sq_part = blksize <= min_partition_size; + const int is_gt_max_sq_part = blksize > max_partition_size; + if (is_gt_max_sq_part) { + // If current block size is larger than max, only allow split. + partition_none_allowed = 0; + partition_horz_allowed = 0; + partition_vert_allowed = 0; + do_square_split = 1; + } else if (is_le_min_sq_part) { + // If current block size is less or equal to min, only allow none if valid + // block large enough; only allow split otherwise. + partition_horz_allowed = 0; + partition_vert_allowed = 0; + // only disable square split when current block is not at the picture + // boundary. otherwise, inherit the square split flag from previous logic + if (has_rows && has_cols) do_square_split = 0; + partition_none_allowed = !do_square_split; } -#endif - - // Ref frames picked in the [i_th] quarter subblock during square partition - // RD search. It may be used to prune ref frame selection of rect partitions. - int ref_frames_used[4] = { - 0, - }; BEGIN_PARTITION_SEARCH: if (x->must_find_valid_partition) { - partition_none_allowed = has_rows && has_cols; - partition_horz_allowed = has_cols && yss <= xss && bsize_at_least_8x8; - partition_vert_allowed = has_rows && xss <= yss && bsize_at_least_8x8; + do_square_split = bsize_at_least_8x8 && (blksize > min_partition_size); + partition_none_allowed = + has_rows && has_cols && (blksize >= min_partition_size); + partition_horz_allowed = + has_cols && bsize_at_least_8x8 && cpi->oxcf.enable_rect_partitions && + (blksize > min_partition_size) && + get_plane_block_size(get_partition_subsize(bsize, PARTITION_HORZ), xss, + yss) != BLOCK_INVALID; + partition_vert_allowed = + has_rows && bsize_at_least_8x8 && cpi->oxcf.enable_rect_partitions && + (blksize > min_partition_size) && + get_plane_block_size(get_partition_subsize(bsize, PARTITION_VERT), xss, + yss) != BLOCK_INVALID; + terminate_partition_search = 0; } // Partition block source pixel variance. unsigned int pb_source_variance = UINT_MAX; -#if CONFIG_DIST_8X8 - if (x->using_dist_8x8) { - if (block_size_high[bsize] <= 8) partition_horz_allowed = 0; - if (block_size_wide[bsize] <= 8) partition_vert_allowed = 0; - if (block_size_high[bsize] <= 8 || block_size_wide[bsize] <= 8) - do_square_split = 0; - } -#endif + // Partition block sse after simple motion compensation, not in use now, + // but will be used for upcoming speed features + unsigned int pb_simple_motion_pred_sse = UINT_MAX; + (void)pb_simple_motion_pred_sse; // PARTITION_NONE - if (partition_none_allowed) { + if (is_le_min_sq_part && has_rows && has_cols) partition_none_allowed = 1; + assert(terminate_partition_search == 0); + int64_t part_none_rd = INT64_MAX; + if (cpi->is_screen_content_type) + partition_none_allowed = has_rows && has_cols; + if (partition_none_allowed && !is_gt_max_sq_part) { int pt_cost = 0; if (bsize_at_least_8x8) { pt_cost = partition_cost[PARTITION_NONE] < INT_MAX ? partition_cost[PARTITION_NONE] : 0; } - int64_t partition_rd_cost = RDCOST(x->rdmult, pt_cost, 0); - int64_t best_remain_rdcost = (best_rdc.rdcost == INT64_MAX) - ? INT64_MAX - : (best_rdc.rdcost - partition_rd_cost); - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, - PARTITION_NONE, bsize, ctx_none, best_remain_rdcost); + RD_STATS partition_rdcost; + av1_init_rd_stats(&partition_rdcost); + partition_rdcost.rate = pt_cost; + av1_rd_cost_update(x->rdmult, &partition_rdcost); + RD_STATS best_remain_rdcost; + av1_rd_stats_subtraction(x->rdmult, &best_rdc, &partition_rdcost, + &best_remain_rdcost); +#if CONFIG_COLLECT_PARTITION_STATS + if (best_remain_rdcost >= 0) { + partition_attempts[PARTITION_NONE] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } +#endif + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, PARTITION_NONE, + bsize, ctx_none, best_remain_rdcost, PICK_MODE_RD); + av1_rd_cost_update(x->rdmult, &this_rdc); +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_NONE] += time; + partition_timer_on = 0; + } +#endif pb_source_variance = x->source_variance; + pb_simple_motion_pred_sse = x->simple_motion_pred_sse; if (none_rd) *none_rd = this_rdc.rdcost; cur_none_rd = this_rdc.rdcost; if (this_rdc.rate != INT_MAX) { - if (cpi->sf.prune_ref_frame_for_rect_partitions) { + if (cpi->sf.inter_sf.prune_ref_frame_for_rect_partitions) { const int ref_type = av1_ref_frame_type(ctx_none->mic.ref_frame); - for (int i = 0; i < 4; ++i) { - ref_frames_used[i] |= (1 << ref_type); - } + update_picked_ref_frames_mask(x, ref_type, bsize, + cm->seq_params.mib_size, mi_row, mi_col); } if (bsize_at_least_8x8) { this_rdc.rate += pt_cost; this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist); } + part_none_rd = this_rdc.rdcost; if (this_rdc.rdcost < best_rdc.rdcost) { // Adjust dist breakout threshold according to the partition size. const int64_t dist_breakout_thr = - cpi->sf.partition_search_breakout_dist_thr >> + cpi->sf.part_sf.partition_search_breakout_dist_thr >> ((2 * (MAX_SB_SIZE_LOG2 - 2)) - (mi_size_wide_log2[bsize] + mi_size_high_log2[bsize])); const int rate_breakout_thr = - cpi->sf.partition_search_breakout_rate_thr * + cpi->sf.part_sf.partition_search_breakout_rate_thr * num_pels_log2_lookup[bsize]; best_rdc = this_rdc; + found_best_partition = true; if (bsize_at_least_8x8) pc_tree->partitioning = PARTITION_NONE; - if ((do_square_split || do_rectangular_split) && + if (!frame_is_intra_only(cm) && + (do_square_split || do_rectangular_split) && !x->e_mbd.lossless[xd->mi[0]->segment_id] && ctx_none->skippable) { const int use_ml_based_breakout = - bsize <= cpi->sf.use_square_partition_only_threshold && + bsize <= cpi->sf.part_sf.use_square_partition_only_threshold && bsize > BLOCK_4X4 && xd->bd == 8; if (use_ml_based_breakout) { - if (ml_predict_breakout(cpi, bsize, x, &this_rdc, - pb_source_variance)) { + if (av1_ml_predict_breakout(cpi, bsize, x, &this_rdc, + pb_source_variance)) { do_square_split = 0; do_rectangular_split = 0; } @@ -3521,51 +3019,17 @@ BEGIN_PARTITION_SEARCH: } } -#if CONFIG_FP_MB_STATS - // Check if every 16x16 first pass block statistics has zero - // motion and the corresponding first pass residue is small enough. - // If that is the case, check the difference variance between the - // current frame and the last frame. If the variance is small enough, - // stop further splitting in RD optimization - if (cpi->use_fp_mb_stats && do_square_split && - cm->base_qindex > qindex_skip_threshold_lookup[bsize]) { - int mb_row = mi_row >> 1; - int mb_col = mi_col >> 1; - int mb_row_end = - AOMMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows); - int mb_col_end = - AOMMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols); - int r, c; - - int skip = 1; - for (r = mb_row; r < mb_row_end; r++) { - for (c = mb_col; c < mb_col_end; c++) { - const int mb_index = r * cm->mb_cols + c; - if (!(cpi->twopass.this_frame_mb_stats[mb_index] & - FPMB_MOTION_ZERO_MASK) || - !(cpi->twopass.this_frame_mb_stats[mb_index] & - FPMB_ERROR_SMALL_MASK)) { - skip = 0; - break; - } - } - if (skip == 0) { - break; - } - } - if (skip) { - if (src_diff_var == UINT_MAX) { - set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize); - src_diff_var = get_sby_perpixel_diff_variance( - cpi, &x->plane[0].src, mi_row, mi_col, bsize); - } - if (src_diff_var < 8) { - do_square_split = 0; - do_rectangular_split = 0; - } - } + if (cpi->sf.part_sf.simple_motion_search_early_term_none && + cm->show_frame && !frame_is_intra_only(cm) && + bsize >= BLOCK_16X16 && mi_row + mi_step < mi_params->mi_rows && + mi_col + mi_step < mi_params->mi_cols && + this_rdc.rdcost < INT64_MAX && this_rdc.rdcost >= 0 && + this_rdc.rate < INT_MAX && this_rdc.rate >= 0 && + (do_square_split || do_rectangular_split)) { + av1_simple_motion_search_early_term_none(cpi, x, pc_tree, mi_row, + mi_col, bsize, &this_rdc, + &terminate_partition_search); } -#endif } } @@ -3573,130 +3037,152 @@ BEGIN_PARTITION_SEARCH: } // store estimated motion vector - if (cpi->sf.adaptive_motion_search) store_pred_mv(x, ctx_none); + if (cpi->sf.mv_sf.adaptive_motion_search) store_pred_mv(x, ctx_none); // PARTITION_SPLIT - if (do_square_split) { + int64_t part_split_rd = INT64_MAX; + if ((!terminate_partition_search && do_square_split) || is_gt_max_sq_part) { av1_init_rd_stats(&sum_rdc); subsize = get_partition_subsize(bsize, PARTITION_SPLIT); sum_rdc.rate = partition_cost[PARTITION_SPLIT]; sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, 0); int idx; +#if CONFIG_COLLECT_PARTITION_STATS + if (best_rdc.rdcost - sum_rdc.rdcost >= 0) { + partition_attempts[PARTITION_SPLIT] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } +#endif for (idx = 0; idx < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++idx) { const int x_idx = (idx & 1) * mi_step; const int y_idx = (idx >> 1) * mi_step; - if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols) + if (mi_row + y_idx >= mi_params->mi_rows || + mi_col + x_idx >= mi_params->mi_cols) continue; - if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none); + if (cpi->sf.mv_sf.adaptive_motion_search) load_pred_mv(x, ctx_none); pc_tree->split[idx]->index = idx; int64_t *p_split_rd = &split_rd[idx]; - int64_t best_remain_rdcost = best_rdc.rdcost == INT64_MAX - ? INT64_MAX - : (best_rdc.rdcost - sum_rdc.rdcost); - if (cpi->sf.prune_ref_frame_for_rect_partitions) - pc_tree->split[idx]->none.rate = INT_MAX; - rd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx, mi_col + x_idx, - subsize, &this_rdc, best_remain_rdcost, - pc_tree->split[idx], p_split_rd); - if (this_rdc.rate == INT_MAX) { - sum_rdc.rdcost = INT64_MAX; + RD_STATS best_remain_rdcost; + av1_rd_stats_subtraction(x->rdmult, &best_rdc, &sum_rdc, + &best_remain_rdcost); + + int curr_quad_tree_idx = 0; + if (frame_is_intra_only(cm) && bsize <= BLOCK_64X64) { + curr_quad_tree_idx = x->quad_tree_idx; + x->quad_tree_idx = 4 * curr_quad_tree_idx + idx + 1; + } + if (!rd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx, + mi_col + x_idx, subsize, max_sq_part, min_sq_part, + &this_rdc, best_remain_rdcost, pc_tree->split[idx], + p_split_rd, multi_pass_mode, + &split_part_rect_win[idx])) { + av1_invalid_rd_stats(&sum_rdc); break; - } else { - sum_rdc.rate += this_rdc.rate; - sum_rdc.dist += this_rdc.dist; - sum_rdc.rdcost += this_rdc.rdcost; - if (cpi->sf.prune_ref_frame_for_rect_partitions && - pc_tree->split[idx]->none.rate != INT_MAX) { - const int ref_type = - av1_ref_frame_type(pc_tree->split[idx]->none.mic.ref_frame); - ref_frames_used[idx] |= (1 << ref_type); - } - if (idx <= 1 && (bsize <= BLOCK_8X8 || - pc_tree->split[idx]->partitioning == PARTITION_NONE)) { - const MB_MODE_INFO *const mbmi = &pc_tree->split[idx]->none.mic; - const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info; - // Neither palette mode nor cfl predicted - if (pmi->palette_size[0] == 0 && pmi->palette_size[1] == 0) { - if (mbmi->uv_mode != UV_CFL_PRED) split_ctx_is_ready[idx] = 1; - } + } + if (frame_is_intra_only(cm) && bsize <= BLOCK_64X64) { + x->quad_tree_idx = curr_quad_tree_idx; + } + + sum_rdc.rate += this_rdc.rate; + sum_rdc.dist += this_rdc.dist; + av1_rd_cost_update(x->rdmult, &sum_rdc); + if (idx <= 1 && (bsize <= BLOCK_8X8 || + pc_tree->split[idx]->partitioning == PARTITION_NONE)) { + const MB_MODE_INFO *const mbmi = &pc_tree->split[idx]->none.mic; + const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info; + // Neither palette mode nor cfl predicted + if (pmi->palette_size[0] == 0 && pmi->palette_size[1] == 0) { + if (mbmi->uv_mode != UV_CFL_PRED) split_ctx_is_ready[idx] = 1; } } } +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_SPLIT] += time; + partition_timer_on = 0; + } +#endif const int reached_last_index = (idx == 4); + part_split_rd = sum_rdc.rdcost; if (reached_last_index && sum_rdc.rdcost < best_rdc.rdcost) { sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist); - if (sum_rdc.rdcost < best_rdc.rdcost) { best_rdc = sum_rdc; + found_best_partition = true; pc_tree->partitioning = PARTITION_SPLIT; } - } else if (cpi->sf.less_rectangular_check_level > 0) { - // skip rectangular partition test when larger block size - // gives better rd cost - if (cpi->sf.less_rectangular_check_level == 2 || idx <= 2) - do_rectangular_split &= !partition_none_allowed; + } else if (cpi->sf.part_sf.less_rectangular_check_level > 0) { + // Skip rectangular partition test when partition type none gives better + // rd than partition type split. + if (cpi->sf.part_sf.less_rectangular_check_level == 2 || idx <= 2) { + const int partition_none_valid = cur_none_rd > 0; + const int partition_none_better = cur_none_rd < sum_rdc.rdcost; + do_rectangular_split &= + !(partition_none_valid && partition_none_better); + } } restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } // if (do_split) - pc_tree->horizontal[0].skip_ref_frame_mask = 0; - pc_tree->horizontal[1].skip_ref_frame_mask = 0; - pc_tree->vertical[0].skip_ref_frame_mask = 0; - pc_tree->vertical[1].skip_ref_frame_mask = 0; - if (cpi->sf.prune_ref_frame_for_rect_partitions) { - int used_frames; - used_frames = ref_frames_used[0] | ref_frames_used[1]; - if (used_frames) pc_tree->horizontal[0].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[2] | ref_frames_used[3]; - if (used_frames) pc_tree->horizontal[1].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[0] | ref_frames_used[2]; - if (used_frames) pc_tree->vertical[0].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[1] | ref_frames_used[3]; - if (used_frames) pc_tree->vertical[1].skip_ref_frame_mask = ~used_frames; + if (cpi->sf.part_sf.ml_early_term_after_part_split_level && + !frame_is_intra_only(cm) && !terminate_partition_search && + do_rectangular_split && + (partition_horz_allowed || partition_vert_allowed)) { + av1_ml_early_term_after_split(cpi, x, pc_tree, bsize, best_rdc.rdcost, + part_none_rd, part_split_rd, split_rd, mi_row, + mi_col, &terminate_partition_search); } - int prune_horz = 0; - int prune_vert = 0; - if (cpi->sf.ml_prune_rect_partition && !frame_is_intra_only(cm) && - (partition_horz_allowed || partition_vert_allowed)) { - av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes); - ml_prune_rect_partition(cpi, x, bsize, best_rdc.rdcost, cur_none_rd, - split_rd, &prune_horz, &prune_vert); + if (!cpi->sf.part_sf.ml_early_term_after_part_split_level && + cpi->sf.part_sf.ml_prune_rect_partition && !frame_is_intra_only(cm) && + (partition_horz_allowed || partition_vert_allowed) && + !(prune_horz || prune_vert) && !terminate_partition_search) { + av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes, bsize); + av1_ml_prune_rect_partition(cpi, x, bsize, best_rdc.rdcost, cur_none_rd, + split_rd, &prune_horz, &prune_vert); } // PARTITION_HORZ - if (partition_horz_allowed && !prune_horz && - (do_rectangular_split || active_h_edge(cpi, mi_row, mi_step))) { + assert(IMPLIES(!cpi->oxcf.enable_rect_partitions, !partition_horz_allowed)); + if (!terminate_partition_search && partition_horz_allowed && !prune_horz && + (do_rectangular_split || active_h_edge(cpi, mi_row, mi_step)) && + !is_gt_max_sq_part) { av1_init_rd_stats(&sum_rdc); subsize = get_partition_subsize(bsize, PARTITION_HORZ); - if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none); - if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 && - partition_none_allowed) { - pc_tree->horizontal[0].pred_interp_filter = - av1_extract_interp_filter(ctx_none->mic.interp_filters, 0); - } - int64_t best_remain_rdcost = best_rdc.rdcost == INT64_MAX - ? INT64_MAX - : (best_rdc.rdcost - sum_rdc.rdcost); + if (cpi->sf.mv_sf.adaptive_motion_search) load_pred_mv(x, ctx_none); sum_rdc.rate = partition_cost[PARTITION_HORZ]; sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, 0); - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, - PARTITION_HORZ, subsize, &pc_tree->horizontal[0], - best_remain_rdcost); + RD_STATS best_remain_rdcost; + av1_rd_stats_subtraction(x->rdmult, &best_rdc, &sum_rdc, + &best_remain_rdcost); +#if CONFIG_COLLECT_PARTITION_STATS + if (best_remain_rdcost >= 0) { + partition_attempts[PARTITION_HORZ] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } +#endif + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, PARTITION_HORZ, + subsize, &pc_tree->horizontal[0], best_remain_rdcost, + PICK_MODE_RD); + av1_rd_cost_update(x->rdmult, &this_rdc); if (this_rdc.rate == INT_MAX) { sum_rdc.rdcost = INT64_MAX; } else { sum_rdc.rate += this_rdc.rate; sum_rdc.dist += this_rdc.dist; - sum_rdc.rdcost += this_rdc.rdcost; + av1_rd_cost_update(x->rdmult, &sum_rdc); } horz_rd[0] = this_rdc.rdcost; @@ -3708,20 +3194,18 @@ BEGIN_PARTITION_SEARCH: if (pmi->palette_size[0] == 0 && pmi->palette_size[1] == 0) { if (mbmi->uv_mode != UV_CFL_PRED) horz_ctx_is_ready = 1; } - update_state(cpi, tile_data, td, ctx_h, mi_row, mi_col, subsize, 1); - encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row, mi_col, - subsize, NULL); + update_state(cpi, td, ctx_h, mi_row, mi_col, subsize, 1); + encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize, NULL); - if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_h); + if (cpi->sf.mv_sf.adaptive_motion_search) load_pred_mv(x, ctx_h); - if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 && - partition_none_allowed) { - pc_tree->horizontal[1].pred_interp_filter = - av1_extract_interp_filter(ctx_h->mic.interp_filters, 0); - } - rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col, &this_rdc, - PARTITION_HORZ, subsize, &pc_tree->horizontal[1], - best_rdc.rdcost - sum_rdc.rdcost); + av1_rd_stats_subtraction(x->rdmult, &best_rdc, &sum_rdc, + &best_remain_rdcost); + + pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col, &this_rdc, + PARTITION_HORZ, subsize, &pc_tree->horizontal[1], + best_remain_rdcost, PICK_MODE_RD); + av1_rd_cost_update(x->rdmult, &this_rdc); horz_rd[1] = this_rdc.rdcost; if (this_rdc.rate == INT_MAX) { @@ -3729,74 +3213,88 @@ BEGIN_PARTITION_SEARCH: } else { sum_rdc.rate += this_rdc.rate; sum_rdc.dist += this_rdc.dist; - sum_rdc.rdcost += this_rdc.rdcost; + av1_rd_cost_update(x->rdmult, &sum_rdc); } } +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_HORZ] += time; + partition_timer_on = 0; + } +#endif if (sum_rdc.rdcost < best_rdc.rdcost) { sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist); if (sum_rdc.rdcost < best_rdc.rdcost) { best_rdc = sum_rdc; + found_best_partition = true; pc_tree->partitioning = PARTITION_HORZ; } + } else { + // Update HORZ win flag + if (rect_part_win_info != NULL) { + rect_part_win_info->horz_win = false; + } } restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } // PARTITION_VERT - if (partition_vert_allowed && !prune_vert && - (do_rectangular_split || active_v_edge(cpi, mi_col, mi_step))) { + assert(IMPLIES(!cpi->oxcf.enable_rect_partitions, !partition_vert_allowed)); + if (!terminate_partition_search && partition_vert_allowed && !prune_vert && + (do_rectangular_split || active_v_edge(cpi, mi_col, mi_step)) && + !is_gt_max_sq_part) { av1_init_rd_stats(&sum_rdc); subsize = get_partition_subsize(bsize, PARTITION_VERT); - if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none); + if (cpi->sf.mv_sf.adaptive_motion_search) load_pred_mv(x, ctx_none); - if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 && - partition_none_allowed) { - pc_tree->vertical[0].pred_interp_filter = - av1_extract_interp_filter(ctx_none->mic.interp_filters, 0); - } sum_rdc.rate = partition_cost[PARTITION_VERT]; sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, 0); - int64_t best_remain_rdcost = best_rdc.rdcost == INT64_MAX - ? INT64_MAX - : (best_rdc.rdcost - sum_rdc.rdcost); - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, - PARTITION_VERT, subsize, &pc_tree->vertical[0], - best_remain_rdcost); + RD_STATS best_remain_rdcost; + av1_rd_stats_subtraction(x->rdmult, &best_rdc, &sum_rdc, + &best_remain_rdcost); +#if CONFIG_COLLECT_PARTITION_STATS + if (best_remain_rdcost >= 0) { + partition_attempts[PARTITION_VERT] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } +#endif + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, PARTITION_VERT, + subsize, &pc_tree->vertical[0], best_remain_rdcost, + PICK_MODE_RD); + av1_rd_cost_update(x->rdmult, &this_rdc); if (this_rdc.rate == INT_MAX) { sum_rdc.rdcost = INT64_MAX; } else { sum_rdc.rate += this_rdc.rate; sum_rdc.dist += this_rdc.dist; - sum_rdc.rdcost += this_rdc.rdcost; + av1_rd_cost_update(x->rdmult, &sum_rdc); } vert_rd[0] = this_rdc.rdcost; - const int64_t vert_max_rdcost = best_rdc.rdcost; - if (sum_rdc.rdcost < vert_max_rdcost && has_cols) { + if (sum_rdc.rdcost < best_rdc.rdcost && has_cols) { const MB_MODE_INFO *const mbmi = &pc_tree->vertical[0].mic; const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info; // Neither palette mode nor cfl predicted if (pmi->palette_size[0] == 0 && pmi->palette_size[1] == 0) { if (mbmi->uv_mode != UV_CFL_PRED) vert_ctx_is_ready = 1; } - update_state(cpi, tile_data, td, &pc_tree->vertical[0], mi_row, mi_col, - subsize, 1); - encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, mi_row, mi_col, - subsize, NULL); + update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 1); + encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize, NULL); - if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none); + if (cpi->sf.mv_sf.adaptive_motion_search) load_pred_mv(x, ctx_none); - if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 && - partition_none_allowed) { - pc_tree->vertical[1].pred_interp_filter = - av1_extract_interp_filter(ctx_none->mic.interp_filters, 0); - } - rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step, &this_rdc, - PARTITION_VERT, subsize, &pc_tree->vertical[1], - best_rdc.rdcost - sum_rdc.rdcost); + av1_rd_stats_subtraction(x->rdmult, &best_rdc, &sum_rdc, + &best_remain_rdcost); + pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step, &this_rdc, + PARTITION_VERT, subsize, &pc_tree->vertical[1], + best_remain_rdcost, PICK_MODE_RD); + av1_rd_cost_update(x->rdmult, &this_rdc); vert_rd[1] = this_rdc.rdcost; if (this_rdc.rate == INT_MAX) { @@ -3804,15 +3302,27 @@ BEGIN_PARTITION_SEARCH: } else { sum_rdc.rate += this_rdc.rate; sum_rdc.dist += this_rdc.dist; - sum_rdc.rdcost += this_rdc.rdcost; + av1_rd_cost_update(x->rdmult, &sum_rdc); } } +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_VERT] += time; + partition_timer_on = 0; + } +#endif + av1_rd_cost_update(x->rdmult, &sum_rdc); if (sum_rdc.rdcost < best_rdc.rdcost) { - sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist); - if (sum_rdc.rdcost < best_rdc.rdcost) { - best_rdc = sum_rdc; - pc_tree->partitioning = PARTITION_VERT; + best_rdc = sum_rdc; + found_best_partition = true; + pc_tree->partitioning = PARTITION_VERT; + } else { + // Update VERT win flag + if (rect_part_win_info != NULL) { + rect_part_win_info->vert_win = false; } } @@ -3820,8 +3330,8 @@ BEGIN_PARTITION_SEARCH: } if (pb_source_variance == UINT_MAX) { - av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes); - if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { + av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes, bsize); + if (is_cur_buf_hbd(xd)) { pb_source_variance = av1_high_get_sby_perpixel_variance( cpi, &x->plane[0].src, bsize, xd->bd); } else { @@ -3830,25 +3340,30 @@ BEGIN_PARTITION_SEARCH: } } + if (use_pb_simple_motion_pred_sse(cpi) && + pb_simple_motion_pred_sse == UINT_MAX) { + const FULLPEL_MV start_mv = kZeroFullMv; + unsigned int var = 0; + + av1_simple_motion_sse_var(cpi, x, mi_row, mi_col, bsize, start_mv, 0, + &pb_simple_motion_pred_sse, &var); + } + + assert(IMPLIES(!cpi->oxcf.enable_rect_partitions, !do_rectangular_split)); + const int ext_partition_allowed = - do_rectangular_split && bsize > BLOCK_8X8 && partition_none_allowed; + do_rectangular_split && + bsize > cpi->sf.part_sf.ext_partition_eval_thresh && has_rows && has_cols; // The standard AB partitions are allowed whenever ext-partition-types are // allowed - int horzab_partition_allowed = ext_partition_allowed; - int vertab_partition_allowed = ext_partition_allowed; - -#if CONFIG_DIST_8X8 - if (x->using_dist_8x8) { - if (block_size_high[bsize] <= 8 || block_size_wide[bsize] <= 8) { - horzab_partition_allowed = 0; - vertab_partition_allowed = 0; - } - } -#endif + int horzab_partition_allowed = + ext_partition_allowed & cpi->oxcf.enable_ab_partitions; + int vertab_partition_allowed = + ext_partition_allowed & cpi->oxcf.enable_ab_partitions; - if (cpi->sf.prune_ext_partition_types_search_level) { - if (cpi->sf.prune_ext_partition_types_search_level == 1) { + if (cpi->sf.part_sf.prune_ext_partition_types_search_level) { + if (cpi->sf.part_sf.prune_ext_partition_types_search_level == 1) { // TODO(debargha,huisu@google.com): may need to tune the threshold for // pb_source_variance. horzab_partition_allowed &= (pc_tree->partitioning == PARTITION_HORZ || @@ -3876,10 +3391,10 @@ BEGIN_PARTITION_SEARCH: } int horza_partition_allowed = horzab_partition_allowed; int horzb_partition_allowed = horzab_partition_allowed; - if (cpi->sf.prune_ext_partition_types_search_level) { + if (cpi->sf.part_sf.prune_ext_partition_types_search_level) { const int64_t horz_a_rd = horz_rd[1] + split_rd[0] + split_rd[1]; const int64_t horz_b_rd = horz_rd[0] + split_rd[2] + split_rd[3]; - switch (cpi->sf.prune_ext_partition_types_search_level) { + switch (cpi->sf.part_sf.prune_ext_partition_types_search_level) { case 1: horza_partition_allowed &= (horz_a_rd / 16 * 14 < best_rdc.rdcost); horzb_partition_allowed &= (horz_b_rd / 16 * 14 < best_rdc.rdcost); @@ -3894,10 +3409,10 @@ BEGIN_PARTITION_SEARCH: int verta_partition_allowed = vertab_partition_allowed; int vertb_partition_allowed = vertab_partition_allowed; - if (cpi->sf.prune_ext_partition_types_search_level) { + if (cpi->sf.part_sf.prune_ext_partition_types_search_level) { const int64_t vert_a_rd = vert_rd[1] + split_rd[0] + split_rd[2]; const int64_t vert_b_rd = vert_rd[0] + split_rd[1] + split_rd[3]; - switch (cpi->sf.prune_ext_partition_types_search_level) { + switch (cpi->sf.part_sf.prune_ext_partition_types_search_level) { case 1: verta_partition_allowed &= (vert_a_rd / 16 * 14 < best_rdc.rdcost); vertb_partition_allowed &= (vert_b_rd / 16 * 14 < best_rdc.rdcost); @@ -3910,20 +3425,32 @@ BEGIN_PARTITION_SEARCH: } } - if (cpi->sf.ml_prune_ab_partition && ext_partition_allowed && + if (cpi->sf.part_sf.ml_prune_ab_partition && ext_partition_allowed && partition_horz_allowed && partition_vert_allowed) { - // TODO(huisu@google.com): x->source_variance may not be the current block's - // variance. The correct one to use is pb_source_variance. - // Need to re-train the model to fix it. - ml_prune_ab_partition(bsize, pc_tree->partitioning, - get_unsigned_bits(x->source_variance), - best_rdc.rdcost, horz_rd, vert_rd, split_rd, - &horza_partition_allowed, &horzb_partition_allowed, - &verta_partition_allowed, &vertb_partition_allowed); + // TODO(huisu@google.com): x->source_variance may not be the current + // block's variance. The correct one to use is pb_source_variance. Need to + // re-train the model to fix it. + av1_ml_prune_ab_partition( + bsize, pc_tree->partitioning, get_unsigned_bits(x->source_variance), + best_rdc.rdcost, horz_rd, vert_rd, split_rd, &horza_partition_allowed, + &horzb_partition_allowed, &verta_partition_allowed, + &vertb_partition_allowed); + } + + horza_partition_allowed &= cpi->oxcf.enable_ab_partitions; + horzb_partition_allowed &= cpi->oxcf.enable_ab_partitions; + verta_partition_allowed &= cpi->oxcf.enable_ab_partitions; + vertb_partition_allowed &= cpi->oxcf.enable_ab_partitions; + + if (cpi->sf.part_sf.prune_ab_partition_using_split_info && + horza_partition_allowed) { + horza_partition_allowed &= evaluate_ab_partition_based_on_split( + pc_tree, PARTITION_HORZ, rect_part_win_info, x->qindex, 0, 1); } // PARTITION_HORZ_A - if (partition_horz_allowed && horza_partition_allowed) { + if (!terminate_partition_search && partition_horz_allowed && + horza_partition_allowed && !is_gt_max_sq_part) { subsize = get_partition_subsize(bsize, PARTITION_HORZ_A); pc_tree->horizontala[0].rd_mode_is_ready = 0; pc_tree->horizontala[1].rd_mode_is_ready = 0; @@ -3939,30 +3466,44 @@ BEGIN_PARTITION_SEARCH: pc_tree->horizontala[1].rd_mode_is_ready = 1; } } - pc_tree->horizontala[0].skip_ref_frame_mask = 0; - pc_tree->horizontala[1].skip_ref_frame_mask = 0; - pc_tree->horizontala[2].skip_ref_frame_mask = 0; - if (cpi->sf.prune_ref_frame_for_rect_partitions) { - int used_frames; - used_frames = ref_frames_used[0]; - if (used_frames) - pc_tree->horizontala[0].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[1]; - if (used_frames) - pc_tree->horizontala[1].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[2] | ref_frames_used[3]; - if (used_frames) - pc_tree->horizontala[2].skip_ref_frame_mask = ~used_frames; - } - rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc, - pc_tree->horizontala, ctx_none, mi_row, mi_col, bsize, - PARTITION_HORZ_A, mi_row, mi_col, bsize2, mi_row, - mi_col + mi_step, bsize2, mi_row + mi_step, mi_col, - subsize); +#if CONFIG_COLLECT_PARTITION_STATS + { + RD_STATS tmp_sum_rdc; + av1_init_rd_stats(&tmp_sum_rdc); + tmp_sum_rdc.rate = x->partition_cost[pl][PARTITION_HORZ_A]; + tmp_sum_rdc.rdcost = RDCOST(x->rdmult, tmp_sum_rdc.rate, 0); + if (best_rdc.rdcost - tmp_sum_rdc.rdcost >= 0) { + partition_attempts[PARTITION_HORZ_A] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } + } +#endif + found_best_partition |= rd_test_partition3( + cpi, td, tile_data, tp, pc_tree, &best_rdc, pc_tree->horizontala, + ctx_none, mi_row, mi_col, bsize, PARTITION_HORZ_A, mi_row, mi_col, + bsize2, mi_row, mi_col + mi_step, bsize2, mi_row + mi_step, mi_col, + subsize); +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_HORZ_A] += time; + partition_timer_on = 0; + } +#endif restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } + + if (cpi->sf.part_sf.prune_ab_partition_using_split_info && + horzb_partition_allowed) { + horzb_partition_allowed &= evaluate_ab_partition_based_on_split( + pc_tree, PARTITION_HORZ, rect_part_win_info, x->qindex, 2, 3); + } + // PARTITION_HORZ_B - if (partition_horz_allowed && horzb_partition_allowed) { + if (!terminate_partition_search && partition_horz_allowed && + horzb_partition_allowed && !is_gt_max_sq_part) { subsize = get_partition_subsize(bsize, PARTITION_HORZ_B); pc_tree->horizontalb[0].rd_mode_is_ready = 0; pc_tree->horizontalb[1].rd_mode_is_ready = 0; @@ -3972,31 +3513,45 @@ BEGIN_PARTITION_SEARCH: pc_tree->horizontalb[0].mic.partition = PARTITION_HORZ_B; pc_tree->horizontalb[0].rd_mode_is_ready = 1; } - pc_tree->horizontalb[0].skip_ref_frame_mask = 0; - pc_tree->horizontalb[1].skip_ref_frame_mask = 0; - pc_tree->horizontalb[2].skip_ref_frame_mask = 0; - if (cpi->sf.prune_ref_frame_for_rect_partitions) { - int used_frames; - used_frames = ref_frames_used[0] | ref_frames_used[1]; - if (used_frames) - pc_tree->horizontalb[0].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[2]; - if (used_frames) - pc_tree->horizontalb[1].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[3]; - if (used_frames) - pc_tree->horizontalb[2].skip_ref_frame_mask = ~used_frames; - } - rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc, - pc_tree->horizontalb, ctx_none, mi_row, mi_col, bsize, - PARTITION_HORZ_B, mi_row, mi_col, subsize, - mi_row + mi_step, mi_col, bsize2, mi_row + mi_step, - mi_col + mi_step, bsize2); +#if CONFIG_COLLECT_PARTITION_STATS + { + RD_STATS tmp_sum_rdc; + av1_init_rd_stats(&tmp_sum_rdc); + tmp_sum_rdc.rate = x->partition_cost[pl][PARTITION_HORZ_B]; + tmp_sum_rdc.rdcost = RDCOST(x->rdmult, tmp_sum_rdc.rate, 0); + if (best_rdc.rdcost - tmp_sum_rdc.rdcost >= 0) { + partition_attempts[PARTITION_HORZ_B] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } + } +#endif + found_best_partition |= rd_test_partition3( + cpi, td, tile_data, tp, pc_tree, &best_rdc, pc_tree->horizontalb, + ctx_none, mi_row, mi_col, bsize, PARTITION_HORZ_B, mi_row, mi_col, + subsize, mi_row + mi_step, mi_col, bsize2, mi_row + mi_step, + mi_col + mi_step, bsize2); + +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_HORZ_B] += time; + partition_timer_on = 0; + } +#endif restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } + if (cpi->sf.part_sf.prune_ab_partition_using_split_info && + verta_partition_allowed) { + verta_partition_allowed &= evaluate_ab_partition_based_on_split( + pc_tree, PARTITION_VERT, rect_part_win_info, x->qindex, 0, 2); + } + // PARTITION_VERT_A - if (partition_vert_allowed && verta_partition_allowed) { + if (!terminate_partition_search && partition_vert_allowed && + verta_partition_allowed && !is_gt_max_sq_part) { subsize = get_partition_subsize(bsize, PARTITION_VERT_A); pc_tree->verticala[0].rd_mode_is_ready = 0; pc_tree->verticala[1].rd_mode_is_ready = 0; @@ -4006,27 +3561,44 @@ BEGIN_PARTITION_SEARCH: pc_tree->verticala[0].mic.partition = PARTITION_VERT_A; pc_tree->verticala[0].rd_mode_is_ready = 1; } - pc_tree->verticala[0].skip_ref_frame_mask = 0; - pc_tree->verticala[1].skip_ref_frame_mask = 0; - pc_tree->verticala[2].skip_ref_frame_mask = 0; - if (cpi->sf.prune_ref_frame_for_rect_partitions) { - int used_frames; - used_frames = ref_frames_used[0]; - if (used_frames) pc_tree->verticala[0].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[2]; - if (used_frames) pc_tree->verticala[1].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[1] | ref_frames_used[3]; - if (used_frames) pc_tree->verticala[2].skip_ref_frame_mask = ~used_frames; - } - rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc, - pc_tree->verticala, ctx_none, mi_row, mi_col, bsize, - PARTITION_VERT_A, mi_row, mi_col, bsize2, - mi_row + mi_step, mi_col, bsize2, mi_row, - mi_col + mi_step, subsize); +#if CONFIG_COLLECT_PARTITION_STATS + { + RD_STATS tmp_sum_rdc; + av1_init_rd_stats(&tmp_sum_rdc); + tmp_sum_rdc.rate = x->partition_cost[pl][PARTITION_VERT_A]; + tmp_sum_rdc.rdcost = RDCOST(x->rdmult, tmp_sum_rdc.rate, 0); + if (best_rdc.rdcost - tmp_sum_rdc.rdcost >= 0) { + partition_attempts[PARTITION_VERT_A] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } + } +#endif + found_best_partition |= rd_test_partition3( + cpi, td, tile_data, tp, pc_tree, &best_rdc, pc_tree->verticala, + ctx_none, mi_row, mi_col, bsize, PARTITION_VERT_A, mi_row, mi_col, + bsize2, mi_row + mi_step, mi_col, bsize2, mi_row, mi_col + mi_step, + subsize); +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_VERT_A] += time; + partition_timer_on = 0; + } +#endif restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } + + if (cpi->sf.part_sf.prune_ab_partition_using_split_info && + vertb_partition_allowed) { + vertb_partition_allowed &= evaluate_ab_partition_based_on_split( + pc_tree, PARTITION_VERT, rect_part_win_info, x->qindex, 1, 3); + } + // PARTITION_VERT_B - if (partition_vert_allowed && vertb_partition_allowed) { + if (!terminate_partition_search && partition_vert_allowed && + vertb_partition_allowed && !is_gt_max_sq_part) { subsize = get_partition_subsize(bsize, PARTITION_VERT_B); pc_tree->verticalb[0].rd_mode_is_ready = 0; pc_tree->verticalb[1].rd_mode_is_ready = 0; @@ -4036,35 +3608,53 @@ BEGIN_PARTITION_SEARCH: pc_tree->verticalb[0].mic.partition = PARTITION_VERT_B; pc_tree->verticalb[0].rd_mode_is_ready = 1; } - pc_tree->verticalb[0].skip_ref_frame_mask = 0; - pc_tree->verticalb[1].skip_ref_frame_mask = 0; - pc_tree->verticalb[2].skip_ref_frame_mask = 0; - if (cpi->sf.prune_ref_frame_for_rect_partitions) { - int used_frames; - used_frames = ref_frames_used[0] | ref_frames_used[2]; - if (used_frames) pc_tree->verticalb[0].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[1]; - if (used_frames) pc_tree->verticalb[1].skip_ref_frame_mask = ~used_frames; - used_frames = ref_frames_used[3]; - if (used_frames) pc_tree->verticalb[2].skip_ref_frame_mask = ~used_frames; - } - rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc, - pc_tree->verticalb, ctx_none, mi_row, mi_col, bsize, - PARTITION_VERT_B, mi_row, mi_col, subsize, mi_row, - mi_col + mi_step, bsize2, mi_row + mi_step, - mi_col + mi_step, bsize2); +#if CONFIG_COLLECT_PARTITION_STATS + { + RD_STATS tmp_sum_rdc; + av1_init_rd_stats(&tmp_sum_rdc); + tmp_sum_rdc.rate = x->partition_cost[pl][PARTITION_VERT_B]; + tmp_sum_rdc.rdcost = RDCOST(x->rdmult, tmp_sum_rdc.rate, 0); + if (!frame_is_intra_only(cm) && + best_rdc.rdcost - tmp_sum_rdc.rdcost >= 0) { + partition_attempts[PARTITION_VERT_B] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } + } +#endif + found_best_partition |= rd_test_partition3( + cpi, td, tile_data, tp, pc_tree, &best_rdc, pc_tree->verticalb, + ctx_none, mi_row, mi_col, bsize, PARTITION_VERT_B, mi_row, mi_col, + subsize, mi_row, mi_col + mi_step, bsize2, mi_row + mi_step, + mi_col + mi_step, bsize2); +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_VERT_B] += time; + partition_timer_on = 0; + } +#endif restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } // partition4_allowed is 1 if we can use a PARTITION_HORZ_4 or // PARTITION_VERT_4 for this block. This is almost the same as - // ext_partition_allowed, except that we don't allow 128x32 or 32x128 blocks, - // so we require that bsize is not BLOCK_128X128. - const int partition4_allowed = - ext_partition_allowed && bsize != BLOCK_128X128; - int partition_horz4_allowed = partition4_allowed && partition_horz_allowed; - int partition_vert4_allowed = partition4_allowed && partition_vert_allowed; - if (cpi->sf.prune_ext_partition_types_search_level == 2) { + // ext_partition_allowed, except that we don't allow 128x32 or 32x128 + // blocks, so we require that bsize is not BLOCK_128X128. + const int partition4_allowed = cpi->oxcf.enable_1to4_partitions && + ext_partition_allowed && + bsize != BLOCK_128X128; + + int partition_horz4_allowed = + partition4_allowed && partition_horz_allowed && + get_plane_block_size(get_partition_subsize(bsize, PARTITION_HORZ_4), xss, + yss) != BLOCK_INVALID; + int partition_vert4_allowed = + partition4_allowed && partition_vert_allowed && + get_plane_block_size(get_partition_subsize(bsize, PARTITION_VERT_4), xss, + yss) != BLOCK_INVALID; + if (cpi->sf.part_sf.prune_ext_partition_types_search_level == 2) { partition_horz4_allowed &= (pc_tree->partitioning == PARTITION_HORZ || pc_tree->partitioning == PARTITION_HORZ_A || pc_tree->partitioning == PARTITION_HORZ_B || @@ -4076,26 +3666,45 @@ BEGIN_PARTITION_SEARCH: pc_tree->partitioning == PARTITION_SPLIT || pc_tree->partitioning == PARTITION_NONE); } - if (cpi->sf.ml_prune_4_partition && partition4_allowed && + if (cpi->sf.part_sf.ml_prune_4_partition && partition4_allowed && partition_horz_allowed && partition_vert_allowed) { - ml_prune_4_partition(cpi, x, bsize, pc_tree->partitioning, best_rdc.rdcost, - horz_rd, vert_rd, split_rd, &partition_horz4_allowed, - &partition_vert4_allowed, pb_source_variance, mi_row, - mi_col); + av1_ml_prune_4_partition(cpi, x, bsize, pc_tree->partitioning, + best_rdc.rdcost, horz_rd, vert_rd, split_rd, + &partition_horz4_allowed, &partition_vert4_allowed, + pb_source_variance, mi_row, mi_col); + } + + if (blksize < (min_partition_size << 2)) { + partition_horz4_allowed = 0; + partition_vert4_allowed = 0; } -#if CONFIG_DIST_8X8 - if (x->using_dist_8x8) { - if (block_size_high[bsize] <= 16 || block_size_wide[bsize] <= 16) { + if (cpi->sf.part_sf.prune_4_partition_using_split_info && + (partition_horz4_allowed || partition_vert4_allowed)) { + // Count of child blocks in which HORZ or VERT partition has won + int num_child_horz_win = 0, num_child_vert_win = 0; + for (int idx = 0; idx < 4; idx++) { + num_child_horz_win += (split_part_rect_win[idx].horz_win) ? 1 : 0; + num_child_vert_win += (split_part_rect_win[idx].vert_win) ? 1 : 0; + } + + // Prune HORZ4/VERT4 partitions based on number of HORZ/VERT winners of + // split partiitons. + // Conservative pruning for high quantizers + const int num_win_thresh = AOMMIN(3 * (MAXQ - x->qindex) / MAXQ + 1, 3); + if (num_child_horz_win < num_win_thresh) { partition_horz4_allowed = 0; + } + if (num_child_vert_win < num_win_thresh) { partition_vert4_allowed = 0; } } -#endif // PARTITION_HORZ_4 - if (partition_horz4_allowed && has_rows && - (do_rectangular_split || active_h_edge(cpi, mi_row, mi_step))) { + assert(IMPLIES(!cpi->oxcf.enable_rect_partitions, !partition_horz4_allowed)); + if (!terminate_partition_search && partition_horz4_allowed && has_rows && + (do_rectangular_split || active_h_edge(cpi, mi_row, mi_step)) && + !is_gt_max_sq_part) { av1_init_rd_stats(&sum_rdc); const int quarter_step = mi_size_high[bsize] / 4; PICK_MODE_CONTEXT *ctx_prev = ctx_none; @@ -4104,42 +3713,54 @@ BEGIN_PARTITION_SEARCH: sum_rdc.rate = partition_cost[PARTITION_HORZ_4]; sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, 0); +#if CONFIG_COLLECT_PARTITION_STATS + if (best_rdc.rdcost - sum_rdc.rdcost >= 0) { + partition_attempts[PARTITION_HORZ_4] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } +#endif for (int i = 0; i < 4; ++i) { const int this_mi_row = mi_row + i * quarter_step; - if (i > 0 && this_mi_row >= cm->mi_rows) break; + if (i > 0 && this_mi_row >= mi_params->mi_rows) break; PICK_MODE_CONTEXT *ctx_this = &pc_tree->horizontal4[i]; ctx_this->rd_mode_is_ready = 0; - ctx_this->skip_ref_frame_mask = 0; - if (cpi->sf.prune_ref_frame_for_rect_partitions) { - const int used_frames = i <= 1 - ? (ref_frames_used[0] | ref_frames_used[1]) - : (ref_frames_used[2] | ref_frames_used[3]); - if (used_frames) ctx_this->skip_ref_frame_mask = ~used_frames; - } if (!rd_try_subblock(cpi, td, tile_data, tp, (i == 3), this_mi_row, - mi_col, subsize, &best_rdc, &sum_rdc, &this_rdc, - PARTITION_HORZ_4, ctx_prev, ctx_this)) + mi_col, subsize, best_rdc, &sum_rdc, + PARTITION_HORZ_4, ctx_prev, ctx_this)) { + av1_invalid_rd_stats(&sum_rdc); break; + } ctx_prev = ctx_this; } + av1_rd_cost_update(x->rdmult, &sum_rdc); if (sum_rdc.rdcost < best_rdc.rdcost) { - sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist); - if (sum_rdc.rdcost < best_rdc.rdcost) { - best_rdc = sum_rdc; - pc_tree->partitioning = PARTITION_HORZ_4; - } + best_rdc = sum_rdc; + found_best_partition = true; + pc_tree->partitioning = PARTITION_HORZ_4; + } + +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_HORZ_4] += time; + partition_timer_on = 0; } +#endif restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } // PARTITION_VERT_4 - if (partition_vert4_allowed && has_cols && - (do_rectangular_split || active_v_edge(cpi, mi_row, mi_step))) { + assert(IMPLIES(!cpi->oxcf.enable_rect_partitions, !partition_vert4_allowed)); + if (!terminate_partition_search && partition_vert4_allowed && has_cols && + (do_rectangular_split || active_v_edge(cpi, mi_row, mi_step)) && + !is_gt_max_sq_part) { av1_init_rd_stats(&sum_rdc); const int quarter_step = mi_size_wide[bsize] / 4; PICK_MODE_CONTEXT *ctx_prev = ctx_none; @@ -4148,58 +3769,105 @@ BEGIN_PARTITION_SEARCH: sum_rdc.rate = partition_cost[PARTITION_VERT_4]; sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, 0); +#if CONFIG_COLLECT_PARTITION_STATS + if (best_rdc.rdcost - sum_rdc.rdcost >= 0) { + partition_attempts[PARTITION_VERT_4] += 1; + aom_usec_timer_start(&partition_timer); + partition_timer_on = 1; + } +#endif for (int i = 0; i < 4; ++i) { const int this_mi_col = mi_col + i * quarter_step; - if (i > 0 && this_mi_col >= cm->mi_cols) break; + if (i > 0 && this_mi_col >= mi_params->mi_cols) break; PICK_MODE_CONTEXT *ctx_this = &pc_tree->vertical4[i]; ctx_this->rd_mode_is_ready = 0; - ctx_this->skip_ref_frame_mask = 0; - if (cpi->sf.prune_ref_frame_for_rect_partitions) { - const int used_frames = i <= 1 - ? (ref_frames_used[0] | ref_frames_used[2]) - : (ref_frames_used[1] | ref_frames_used[3]); - if (used_frames) ctx_this->skip_ref_frame_mask = ~used_frames; - } if (!rd_try_subblock(cpi, td, tile_data, tp, (i == 3), mi_row, - this_mi_col, subsize, &best_rdc, &sum_rdc, &this_rdc, - PARTITION_VERT_4, ctx_prev, ctx_this)) + this_mi_col, subsize, best_rdc, &sum_rdc, + PARTITION_VERT_4, ctx_prev, ctx_this)) { + av1_invalid_rd_stats(&sum_rdc); break; + } ctx_prev = ctx_this; } + av1_rd_cost_update(x->rdmult, &sum_rdc); if (sum_rdc.rdcost < best_rdc.rdcost) { - sum_rdc.rdcost = RDCOST(x->rdmult, sum_rdc.rate, sum_rdc.dist); - if (sum_rdc.rdcost < best_rdc.rdcost) { - best_rdc = sum_rdc; - pc_tree->partitioning = PARTITION_VERT_4; - } + best_rdc = sum_rdc; + found_best_partition = true; + pc_tree->partitioning = PARTITION_VERT_4; + } +#if CONFIG_COLLECT_PARTITION_STATS + if (partition_timer_on) { + aom_usec_timer_mark(&partition_timer); + int64_t time = aom_usec_timer_elapsed(&partition_timer); + partition_times[PARTITION_VERT_4] += time; + partition_timer_on = 0; } +#endif restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes); } - if (bsize == cm->seq_params.sb_size && best_rdc.rate == INT_MAX) { + if (bsize == cm->seq_params.sb_size && !found_best_partition) { // Did not find a valid partition, go back and search again, with less // constraint on which partition types to search. x->must_find_valid_partition = 1; +#if CONFIG_COLLECT_PARTITION_STATS == 2 + part_stats->partition_redo += 1; +#endif goto BEGIN_PARTITION_SEARCH; } - // TODO(jbb): This code added so that we avoid static analysis - // warning related to the fact that best_rd isn't used after this - // point. This code should be refactored so that the duplicate - // checks occur in some sub function and thus are used... - (void)best_rd; *rd_cost = best_rdc; - if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && - pc_tree->index != 3) { +#if CONFIG_COLLECT_PARTITION_STATS + if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX) { + partition_decisions[pc_tree->partitioning] += 1; + } +#endif + +#if CONFIG_COLLECT_PARTITION_STATS == 1 + // If CONFIG_COLLECT_PARTITION_STATS is 1, then print out the stats for each + // prediction block + FILE *f = fopen("data.csv", "a"); + fprintf(f, "%d,%d,%d,", bsize, cm->show_frame, frame_is_intra_only(cm)); + for (int idx = 0; idx < EXT_PARTITION_TYPES; idx++) { + fprintf(f, "%d,", partition_decisions[idx]); + } + for (int idx = 0; idx < EXT_PARTITION_TYPES; idx++) { + fprintf(f, "%d,", partition_attempts[idx]); + } + for (int idx = 0; idx < EXT_PARTITION_TYPES; idx++) { + fprintf(f, "%ld,", partition_times[idx]); + } + fprintf(f, "\n"); + fclose(f); +#endif + +#if CONFIG_COLLECT_PARTITION_STATS == 2 + // If CONFIG_COLLECTION_PARTITION_STATS is 2, then we print out the stats for + // the whole clip. So we need to pass the information upstream to the encoder + const int bsize_idx = av1_get_bsize_idx_for_part_stats(bsize); + int *agg_attempts = part_stats->partition_attempts[bsize_idx]; + int *agg_decisions = part_stats->partition_decisions[bsize_idx]; + int64_t *agg_times = part_stats->partition_times[bsize_idx]; + for (int idx = 0; idx < EXT_PARTITION_TYPES; idx++) { + agg_attempts[idx] += partition_attempts[idx]; + agg_decisions[idx] += partition_decisions[idx]; + agg_times[idx] += partition_times[idx]; + } +#endif + + if (found_best_partition && pc_tree->index != 3) { if (bsize == cm->seq_params.sb_size) { + const int emit_output = multi_pass_mode != SB_DRY_PASS; + const RUN_TYPE run_type = emit_output ? OUTPUT_ENABLED : DRY_RUN_NORMAL; + x->cb_offset = 0; - encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, OUTPUT_ENABLED, bsize, + encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, run_type, bsize, pc_tree, NULL); } else { encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, DRY_RUN_NORMAL, bsize, @@ -4213,331 +3881,1134 @@ BEGIN_PARTITION_SEARCH: } else { assert(tp_orig == *tp); } + + x->rdmult = orig_rdmult; + return found_best_partition; +} +#endif // !CONFIG_REALTIME_ONLY +#undef NUM_SIMPLE_MOTION_FEATURES + +#if !CONFIG_REALTIME_ONLY + +static int get_rdmult_delta(AV1_COMP *cpi, BLOCK_SIZE bsize, int analysis_type, + int mi_row, int mi_col, int orig_rdmult) { + AV1_COMMON *const cm = &cpi->common; + assert(IMPLIES(cpi->gf_group.size > 0, + cpi->gf_group.index < cpi->gf_group.size)); + const int tpl_idx = cpi->gf_group.index; + TplParams *const tpl_data = &cpi->tpl_data; + TplDepFrame *tpl_frame = &tpl_data->tpl_frame[tpl_idx]; + TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr; + const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2; + int tpl_stride = tpl_frame->stride; + int64_t intra_cost = 0; + int64_t mc_dep_cost = 0; + const int mi_wide = mi_size_wide[bsize]; + const int mi_high = mi_size_high[bsize]; + + if (tpl_frame->is_valid == 0) return orig_rdmult; + + if (!is_frame_tpl_eligible(cpi)) return orig_rdmult; + + if (cpi->gf_group.index >= MAX_LAG_BUFFERS) return orig_rdmult; + + int64_t mc_count = 0, mc_saved = 0; + int mi_count = 0; + const int mi_col_sr = + coded_to_superres_mi(mi_col, cm->superres_scale_denominator); + const int mi_col_end_sr = + coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator); + const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width); + const int step = 1 << block_mis_log2; + for (int row = mi_row; row < mi_row + mi_high; row += step) { + for (int col = mi_col_sr; col < mi_col_end_sr; col += step) { + if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) continue; + TplDepStats *this_stats = + &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)]; + int64_t mc_dep_delta = + RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate, + this_stats->mc_dep_dist); + intra_cost += this_stats->recrf_dist << RDDIV_BITS; + mc_dep_cost += (this_stats->recrf_dist << RDDIV_BITS) + mc_dep_delta; + mc_count += this_stats->mc_count; + mc_saved += this_stats->mc_saved; + mi_count++; + } + } + + aom_clear_system_state(); + + double beta = 1.0; + if (analysis_type == 0) { + if (mc_dep_cost > 0 && intra_cost > 0) { + const double r0 = cpi->rd.r0; + const double rk = (double)intra_cost / mc_dep_cost; + beta = (r0 / rk); + } + } else if (analysis_type == 1) { + const double mc_count_base = (mi_count * cpi->rd.mc_count_base); + beta = (mc_count + 1.0) / (mc_count_base + 1.0); + beta = pow(beta, 0.5); + } else if (analysis_type == 2) { + const double mc_saved_base = (mi_count * cpi->rd.mc_saved_base); + beta = (mc_saved + 1.0) / (mc_saved_base + 1.0); + beta = pow(beta, 0.5); + } + + int rdmult = av1_get_adaptive_rdmult(cpi, beta); + + aom_clear_system_state(); + + rdmult = AOMMIN(rdmult, orig_rdmult * 3 / 2); + rdmult = AOMMAX(rdmult, orig_rdmult * 1 / 2); + + rdmult = AOMMAX(1, rdmult); + + return rdmult; +} + +static int get_tpl_stats_b(AV1_COMP *cpi, BLOCK_SIZE bsize, int mi_row, + int mi_col, int64_t *intra_cost_b, + int64_t *inter_cost_b, + int_mv mv_b[][INTER_REFS_PER_FRAME], int *stride) { + if (!cpi->oxcf.enable_tpl_model) return 0; + if (cpi->superres_mode != SUPERRES_NONE) return 0; + if (cpi->common.current_frame.frame_type == KEY_FRAME) return 0; + const FRAME_UPDATE_TYPE update_type = get_frame_update_type(&cpi->gf_group); + if (update_type == INTNL_OVERLAY_UPDATE || update_type == OVERLAY_UPDATE) + return 0; + assert(IMPLIES(cpi->gf_group.size > 0, + cpi->gf_group.index < cpi->gf_group.size)); + + AV1_COMMON *const cm = &cpi->common; + const int gf_group_index = cpi->gf_group.index; + TplParams *const tpl_data = &cpi->tpl_data; + TplDepFrame *tpl_frame = &tpl_data->tpl_frame[gf_group_index]; + TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr; + int tpl_stride = tpl_frame->stride; + const int mi_wide = mi_size_wide[bsize]; + const int mi_high = mi_size_high[bsize]; + + if (tpl_frame->is_valid == 0) return 0; + if (gf_group_index >= MAX_LAG_BUFFERS) return 0; + + int mi_count = 0; + int count = 0; + const int mi_col_sr = + coded_to_superres_mi(mi_col, cm->superres_scale_denominator); + const int mi_col_end_sr = + coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator); + // mi_cols_sr is mi_cols at superres case. + const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width); + + // TPL store unit size is not the same as the motion estimation unit size. + // Here always use motion estimation size to avoid getting repetitive inter/ + // intra cost. + const BLOCK_SIZE tpl_bsize = convert_length_to_bsize(MC_FLOW_BSIZE_1D); + const int step = mi_size_wide[tpl_bsize]; + assert(mi_size_wide[tpl_bsize] == mi_size_high[tpl_bsize]); + + // Stride is only based on SB size, and we fill in values for every 16x16 + // block in a SB. + *stride = (mi_col_end_sr - mi_col_sr) / step; + + for (int row = mi_row; row < mi_row + mi_high; row += step) { + for (int col = mi_col_sr; col < mi_col_end_sr; col += step) { + // Handle partial SB, so that no invalid values are used later. + if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) { + inter_cost_b[count] = INT64_MAX; + intra_cost_b[count] = INT64_MAX; + for (int i = 0; i < INTER_REFS_PER_FRAME; ++i) { + mv_b[count][i].as_int = INVALID_MV; + } + count++; + continue; + } + + TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos( + row, col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)]; + inter_cost_b[count] = this_stats->inter_cost; + intra_cost_b[count] = this_stats->intra_cost; + memcpy(mv_b[count], this_stats->mv, sizeof(this_stats->mv)); + mi_count++; + count++; + } + } + + return mi_count; +} + +// analysis_type 0: Use mc_dep_cost and intra_cost +// analysis_type 1: Use count of best inter predictor chosen +// analysis_type 2: Use cost reduction from intra to inter for best inter +// predictor chosen +static int get_q_for_deltaq_objective(AV1_COMP *const cpi, BLOCK_SIZE bsize, + int mi_row, int mi_col) { + AV1_COMMON *const cm = &cpi->common; + assert(IMPLIES(cpi->gf_group.size > 0, + cpi->gf_group.index < cpi->gf_group.size)); + const int tpl_idx = cpi->gf_group.index; + TplParams *const tpl_data = &cpi->tpl_data; + TplDepFrame *tpl_frame = &tpl_data->tpl_frame[tpl_idx]; + TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr; + const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2; + int tpl_stride = tpl_frame->stride; + int64_t intra_cost = 0; + int64_t mc_dep_cost = 0; + const int mi_wide = mi_size_wide[bsize]; + const int mi_high = mi_size_high[bsize]; + const int base_qindex = cm->quant_params.base_qindex; + + if (tpl_frame->is_valid == 0) return base_qindex; + + if (!is_frame_tpl_eligible(cpi)) return base_qindex; + + if (cpi->gf_group.index >= MAX_LAG_BUFFERS) return base_qindex; + + int64_t mc_count = 0, mc_saved = 0; + int mi_count = 0; + const int mi_col_sr = + coded_to_superres_mi(mi_col, cm->superres_scale_denominator); + const int mi_col_end_sr = + coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator); + const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width); + const int step = 1 << block_mis_log2; + for (int row = mi_row; row < mi_row + mi_high; row += step) { + for (int col = mi_col_sr; col < mi_col_end_sr; col += step) { + if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) continue; + TplDepStats *this_stats = + &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)]; + int64_t mc_dep_delta = + RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate, + this_stats->mc_dep_dist); + intra_cost += this_stats->recrf_dist << RDDIV_BITS; + mc_dep_cost += (this_stats->recrf_dist << RDDIV_BITS) + mc_dep_delta; + mc_count += this_stats->mc_count; + mc_saved += this_stats->mc_saved; + mi_count++; + } + } + + aom_clear_system_state(); + + int offset = 0; + double beta = 1.0; + if (mc_dep_cost > 0 && intra_cost > 0) { + const double r0 = cpi->rd.r0; + const double rk = (double)intra_cost / mc_dep_cost; + beta = (r0 / rk); + assert(beta > 0.0); + } + offset = av1_get_deltaq_offset(cpi, base_qindex, beta); + aom_clear_system_state(); + + const DeltaQInfo *const delta_q_info = &cm->delta_q_info; + offset = AOMMIN(offset, delta_q_info->delta_q_res * 9 - 1); + offset = AOMMAX(offset, -delta_q_info->delta_q_res * 9 + 1); + int qindex = cm->quant_params.base_qindex + offset; + qindex = AOMMIN(qindex, MAXQ); + qindex = AOMMAX(qindex, MINQ); + + return qindex; +} + +static AOM_INLINE void setup_delta_q(AV1_COMP *const cpi, ThreadData *td, + MACROBLOCK *const x, + const TileInfo *const tile_info, + int mi_row, int mi_col, int num_planes) { + AV1_COMMON *const cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; + const DeltaQInfo *const delta_q_info = &cm->delta_q_info; + assert(delta_q_info->delta_q_present_flag); + + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + // Delta-q modulation based on variance + av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes, sb_size); + + int current_qindex = cm->quant_params.base_qindex; + if (cpi->oxcf.deltaq_mode == DELTA_Q_PERCEPTUAL) { + if (DELTA_Q_PERCEPTUAL_MODULATION == 1) { + const int block_wavelet_energy_level = + av1_block_wavelet_energy_level(cpi, x, sb_size); + x->sb_energy_level = block_wavelet_energy_level; + current_qindex = av1_compute_q_from_energy_level_deltaq_mode( + cpi, block_wavelet_energy_level); + } else { + const int block_var_level = av1_log_block_var(cpi, x, sb_size); + x->sb_energy_level = block_var_level; + current_qindex = + av1_compute_q_from_energy_level_deltaq_mode(cpi, block_var_level); + } + } else if (cpi->oxcf.deltaq_mode == DELTA_Q_OBJECTIVE && + cpi->oxcf.enable_tpl_model) { + // Setup deltaq based on tpl stats + current_qindex = get_q_for_deltaq_objective(cpi, sb_size, mi_row, mi_col); + } + + const int delta_q_res = delta_q_info->delta_q_res; + // Right now aq only works with tpl model. So if tpl is disabled, we set the + // current_qindex to base_qindex. + if (cpi->oxcf.enable_tpl_model && cpi->oxcf.deltaq_mode != NO_DELTA_Q) { + current_qindex = + clamp(current_qindex, delta_q_res, 256 - delta_q_info->delta_q_res); + } else { + current_qindex = cm->quant_params.base_qindex; + } + + MACROBLOCKD *const xd = &x->e_mbd; + const int sign_deltaq_index = + current_qindex - xd->current_qindex >= 0 ? 1 : -1; + const int deltaq_deadzone = delta_q_res / 4; + const int qmask = ~(delta_q_res - 1); + int abs_deltaq_index = abs(current_qindex - xd->current_qindex); + abs_deltaq_index = (abs_deltaq_index + deltaq_deadzone) & qmask; + current_qindex = xd->current_qindex + sign_deltaq_index * abs_deltaq_index; + current_qindex = AOMMAX(current_qindex, MINQ + 1); + assert(current_qindex > 0); + + xd->delta_qindex = current_qindex - cm->quant_params.base_qindex; + set_offsets(cpi, tile_info, x, mi_row, mi_col, sb_size); + xd->mi[0]->current_qindex = current_qindex; + av1_init_plane_quantizers(cpi, x, xd->mi[0]->segment_id); + + // keep track of any non-zero delta-q used + td->deltaq_used |= (xd->delta_qindex != 0); + + if (cpi->oxcf.deltalf_mode) { + const int delta_lf_res = delta_q_info->delta_lf_res; + const int lfmask = ~(delta_lf_res - 1); + const int delta_lf_from_base = + ((xd->delta_qindex / 2 + delta_lf_res / 2) & lfmask); + const int8_t delta_lf = + (int8_t)clamp(delta_lf_from_base, -MAX_LOOP_FILTER, MAX_LOOP_FILTER); + const int frame_lf_count = + av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2; + const int mib_size = cm->seq_params.mib_size; + + // pre-set the delta lf for loop filter. Note that this value is set + // before mi is assigned for each block in current superblock + for (int j = 0; j < AOMMIN(mib_size, mi_params->mi_rows - mi_row); j++) { + for (int k = 0; k < AOMMIN(mib_size, mi_params->mi_cols - mi_col); k++) { + const int grid_idx = get_mi_grid_idx(mi_params, mi_row + j, mi_col + k); + mi_params->mi_grid_base[grid_idx]->delta_lf_from_base = delta_lf; + for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) { + mi_params->mi_grid_base[grid_idx]->delta_lf[lf_id] = delta_lf; + } + } + } + } +} +#endif // !CONFIG_REALTIME_ONLY + +#define AVG_CDF_WEIGHT_LEFT 3 +#define AVG_CDF_WEIGHT_TOP_RIGHT 1 + +static AOM_INLINE void avg_cdf_symbol(aom_cdf_prob *cdf_ptr_left, + aom_cdf_prob *cdf_ptr_tr, int num_cdfs, + int cdf_stride, int nsymbs, int wt_left, + int wt_tr) { + for (int i = 0; i < num_cdfs; i++) { + for (int j = 0; j <= nsymbs; j++) { + cdf_ptr_left[i * cdf_stride + j] = + (aom_cdf_prob)(((int)cdf_ptr_left[i * cdf_stride + j] * wt_left + + (int)cdf_ptr_tr[i * cdf_stride + j] * wt_tr + + ((wt_left + wt_tr) / 2)) / + (wt_left + wt_tr)); + assert(cdf_ptr_left[i * cdf_stride + j] >= 0 && + cdf_ptr_left[i * cdf_stride + j] < CDF_PROB_TOP); + } + } +} + +#define AVERAGE_CDF(cname_left, cname_tr, nsymbs) \ + AVG_CDF_STRIDE(cname_left, cname_tr, nsymbs, CDF_SIZE(nsymbs)) + +#define AVG_CDF_STRIDE(cname_left, cname_tr, nsymbs, cdf_stride) \ + do { \ + aom_cdf_prob *cdf_ptr_left = (aom_cdf_prob *)cname_left; \ + aom_cdf_prob *cdf_ptr_tr = (aom_cdf_prob *)cname_tr; \ + int array_size = (int)sizeof(cname_left) / sizeof(aom_cdf_prob); \ + int num_cdfs = array_size / cdf_stride; \ + avg_cdf_symbol(cdf_ptr_left, cdf_ptr_tr, num_cdfs, cdf_stride, nsymbs, \ + wt_left, wt_tr); \ + } while (0) + +static AOM_INLINE void avg_nmv(nmv_context *nmv_left, nmv_context *nmv_tr, + int wt_left, int wt_tr) { + AVERAGE_CDF(nmv_left->joints_cdf, nmv_tr->joints_cdf, 4); + for (int i = 0; i < 2; i++) { + AVERAGE_CDF(nmv_left->comps[i].classes_cdf, nmv_tr->comps[i].classes_cdf, + MV_CLASSES); + AVERAGE_CDF(nmv_left->comps[i].class0_fp_cdf, + nmv_tr->comps[i].class0_fp_cdf, MV_FP_SIZE); + AVERAGE_CDF(nmv_left->comps[i].fp_cdf, nmv_tr->comps[i].fp_cdf, MV_FP_SIZE); + AVERAGE_CDF(nmv_left->comps[i].sign_cdf, nmv_tr->comps[i].sign_cdf, 2); + AVERAGE_CDF(nmv_left->comps[i].class0_hp_cdf, + nmv_tr->comps[i].class0_hp_cdf, 2); + AVERAGE_CDF(nmv_left->comps[i].hp_cdf, nmv_tr->comps[i].hp_cdf, 2); + AVERAGE_CDF(nmv_left->comps[i].class0_cdf, nmv_tr->comps[i].class0_cdf, + CLASS0_SIZE); + AVERAGE_CDF(nmv_left->comps[i].bits_cdf, nmv_tr->comps[i].bits_cdf, 2); + } +} + +// In case of row-based multi-threading of encoder, since we always +// keep a top - right sync, we can average the top - right SB's CDFs and +// the left SB's CDFs and use the same for current SB's encoding to +// improve the performance. This function facilitates the averaging +// of CDF and used only when row-mt is enabled in encoder. +static AOM_INLINE void avg_cdf_symbols(FRAME_CONTEXT *ctx_left, + FRAME_CONTEXT *ctx_tr, int wt_left, + int wt_tr) { + AVERAGE_CDF(ctx_left->txb_skip_cdf, ctx_tr->txb_skip_cdf, 2); + AVERAGE_CDF(ctx_left->eob_extra_cdf, ctx_tr->eob_extra_cdf, 2); + AVERAGE_CDF(ctx_left->dc_sign_cdf, ctx_tr->dc_sign_cdf, 2); + AVERAGE_CDF(ctx_left->eob_flag_cdf16, ctx_tr->eob_flag_cdf16, 5); + AVERAGE_CDF(ctx_left->eob_flag_cdf32, ctx_tr->eob_flag_cdf32, 6); + AVERAGE_CDF(ctx_left->eob_flag_cdf64, ctx_tr->eob_flag_cdf64, 7); + AVERAGE_CDF(ctx_left->eob_flag_cdf128, ctx_tr->eob_flag_cdf128, 8); + AVERAGE_CDF(ctx_left->eob_flag_cdf256, ctx_tr->eob_flag_cdf256, 9); + AVERAGE_CDF(ctx_left->eob_flag_cdf512, ctx_tr->eob_flag_cdf512, 10); + AVERAGE_CDF(ctx_left->eob_flag_cdf1024, ctx_tr->eob_flag_cdf1024, 11); + AVERAGE_CDF(ctx_left->coeff_base_eob_cdf, ctx_tr->coeff_base_eob_cdf, 3); + AVERAGE_CDF(ctx_left->coeff_base_cdf, ctx_tr->coeff_base_cdf, 4); + AVERAGE_CDF(ctx_left->coeff_br_cdf, ctx_tr->coeff_br_cdf, BR_CDF_SIZE); + AVERAGE_CDF(ctx_left->newmv_cdf, ctx_tr->newmv_cdf, 2); + AVERAGE_CDF(ctx_left->zeromv_cdf, ctx_tr->zeromv_cdf, 2); + AVERAGE_CDF(ctx_left->refmv_cdf, ctx_tr->refmv_cdf, 2); + AVERAGE_CDF(ctx_left->drl_cdf, ctx_tr->drl_cdf, 2); + AVERAGE_CDF(ctx_left->inter_compound_mode_cdf, + ctx_tr->inter_compound_mode_cdf, INTER_COMPOUND_MODES); + AVERAGE_CDF(ctx_left->compound_type_cdf, ctx_tr->compound_type_cdf, + MASKED_COMPOUND_TYPES); + AVERAGE_CDF(ctx_left->wedge_idx_cdf, ctx_tr->wedge_idx_cdf, 16); + AVERAGE_CDF(ctx_left->interintra_cdf, ctx_tr->interintra_cdf, 2); + AVERAGE_CDF(ctx_left->wedge_interintra_cdf, ctx_tr->wedge_interintra_cdf, 2); + AVERAGE_CDF(ctx_left->interintra_mode_cdf, ctx_tr->interintra_mode_cdf, + INTERINTRA_MODES); + AVERAGE_CDF(ctx_left->motion_mode_cdf, ctx_tr->motion_mode_cdf, MOTION_MODES); + AVERAGE_CDF(ctx_left->obmc_cdf, ctx_tr->obmc_cdf, 2); + AVERAGE_CDF(ctx_left->palette_y_size_cdf, ctx_tr->palette_y_size_cdf, + PALETTE_SIZES); + AVERAGE_CDF(ctx_left->palette_uv_size_cdf, ctx_tr->palette_uv_size_cdf, + PALETTE_SIZES); + for (int j = 0; j < PALETTE_SIZES; j++) { + int nsymbs = j + PALETTE_MIN_SIZE; + AVG_CDF_STRIDE(ctx_left->palette_y_color_index_cdf[j], + ctx_tr->palette_y_color_index_cdf[j], nsymbs, + CDF_SIZE(PALETTE_COLORS)); + AVG_CDF_STRIDE(ctx_left->palette_uv_color_index_cdf[j], + ctx_tr->palette_uv_color_index_cdf[j], nsymbs, + CDF_SIZE(PALETTE_COLORS)); + } + AVERAGE_CDF(ctx_left->palette_y_mode_cdf, ctx_tr->palette_y_mode_cdf, 2); + AVERAGE_CDF(ctx_left->palette_uv_mode_cdf, ctx_tr->palette_uv_mode_cdf, 2); + AVERAGE_CDF(ctx_left->comp_inter_cdf, ctx_tr->comp_inter_cdf, 2); + AVERAGE_CDF(ctx_left->single_ref_cdf, ctx_tr->single_ref_cdf, 2); + AVERAGE_CDF(ctx_left->comp_ref_type_cdf, ctx_tr->comp_ref_type_cdf, 2); + AVERAGE_CDF(ctx_left->uni_comp_ref_cdf, ctx_tr->uni_comp_ref_cdf, 2); + AVERAGE_CDF(ctx_left->comp_ref_cdf, ctx_tr->comp_ref_cdf, 2); + AVERAGE_CDF(ctx_left->comp_bwdref_cdf, ctx_tr->comp_bwdref_cdf, 2); + AVERAGE_CDF(ctx_left->txfm_partition_cdf, ctx_tr->txfm_partition_cdf, 2); + AVERAGE_CDF(ctx_left->compound_index_cdf, ctx_tr->compound_index_cdf, 2); + AVERAGE_CDF(ctx_left->comp_group_idx_cdf, ctx_tr->comp_group_idx_cdf, 2); + AVERAGE_CDF(ctx_left->skip_mode_cdfs, ctx_tr->skip_mode_cdfs, 2); + AVERAGE_CDF(ctx_left->skip_cdfs, ctx_tr->skip_cdfs, 2); + AVERAGE_CDF(ctx_left->intra_inter_cdf, ctx_tr->intra_inter_cdf, 2); + avg_nmv(&ctx_left->nmvc, &ctx_tr->nmvc, wt_left, wt_tr); + avg_nmv(&ctx_left->ndvc, &ctx_tr->ndvc, wt_left, wt_tr); + AVERAGE_CDF(ctx_left->intrabc_cdf, ctx_tr->intrabc_cdf, 2); + AVERAGE_CDF(ctx_left->seg.tree_cdf, ctx_tr->seg.tree_cdf, MAX_SEGMENTS); + AVERAGE_CDF(ctx_left->seg.pred_cdf, ctx_tr->seg.pred_cdf, 2); + AVERAGE_CDF(ctx_left->seg.spatial_pred_seg_cdf, + ctx_tr->seg.spatial_pred_seg_cdf, MAX_SEGMENTS); + AVERAGE_CDF(ctx_left->filter_intra_cdfs, ctx_tr->filter_intra_cdfs, 2); + AVERAGE_CDF(ctx_left->filter_intra_mode_cdf, ctx_tr->filter_intra_mode_cdf, + FILTER_INTRA_MODES); + AVERAGE_CDF(ctx_left->switchable_restore_cdf, ctx_tr->switchable_restore_cdf, + RESTORE_SWITCHABLE_TYPES); + AVERAGE_CDF(ctx_left->wiener_restore_cdf, ctx_tr->wiener_restore_cdf, 2); + AVERAGE_CDF(ctx_left->sgrproj_restore_cdf, ctx_tr->sgrproj_restore_cdf, 2); + AVERAGE_CDF(ctx_left->y_mode_cdf, ctx_tr->y_mode_cdf, INTRA_MODES); + AVG_CDF_STRIDE(ctx_left->uv_mode_cdf[0], ctx_tr->uv_mode_cdf[0], + UV_INTRA_MODES - 1, CDF_SIZE(UV_INTRA_MODES)); + AVERAGE_CDF(ctx_left->uv_mode_cdf[1], ctx_tr->uv_mode_cdf[1], UV_INTRA_MODES); + for (int i = 0; i < PARTITION_CONTEXTS; i++) { + if (i < 4) { + AVG_CDF_STRIDE(ctx_left->partition_cdf[i], ctx_tr->partition_cdf[i], 4, + CDF_SIZE(10)); + } else if (i < 16) { + AVERAGE_CDF(ctx_left->partition_cdf[i], ctx_tr->partition_cdf[i], 10); + } else { + AVG_CDF_STRIDE(ctx_left->partition_cdf[i], ctx_tr->partition_cdf[i], 8, + CDF_SIZE(10)); + } + } + AVERAGE_CDF(ctx_left->switchable_interp_cdf, ctx_tr->switchable_interp_cdf, + SWITCHABLE_FILTERS); + AVERAGE_CDF(ctx_left->kf_y_cdf, ctx_tr->kf_y_cdf, INTRA_MODES); + AVERAGE_CDF(ctx_left->angle_delta_cdf, ctx_tr->angle_delta_cdf, + 2 * MAX_ANGLE_DELTA + 1); + AVG_CDF_STRIDE(ctx_left->tx_size_cdf[0], ctx_tr->tx_size_cdf[0], MAX_TX_DEPTH, + CDF_SIZE(MAX_TX_DEPTH + 1)); + AVERAGE_CDF(ctx_left->tx_size_cdf[1], ctx_tr->tx_size_cdf[1], + MAX_TX_DEPTH + 1); + AVERAGE_CDF(ctx_left->tx_size_cdf[2], ctx_tr->tx_size_cdf[2], + MAX_TX_DEPTH + 1); + AVERAGE_CDF(ctx_left->tx_size_cdf[3], ctx_tr->tx_size_cdf[3], + MAX_TX_DEPTH + 1); + AVERAGE_CDF(ctx_left->delta_q_cdf, ctx_tr->delta_q_cdf, DELTA_Q_PROBS + 1); + AVERAGE_CDF(ctx_left->delta_lf_cdf, ctx_tr->delta_lf_cdf, DELTA_LF_PROBS + 1); + for (int i = 0; i < FRAME_LF_COUNT; i++) { + AVERAGE_CDF(ctx_left->delta_lf_multi_cdf[i], ctx_tr->delta_lf_multi_cdf[i], + DELTA_LF_PROBS + 1); + } + AVG_CDF_STRIDE(ctx_left->intra_ext_tx_cdf[1], ctx_tr->intra_ext_tx_cdf[1], 7, + CDF_SIZE(TX_TYPES)); + AVG_CDF_STRIDE(ctx_left->intra_ext_tx_cdf[2], ctx_tr->intra_ext_tx_cdf[2], 5, + CDF_SIZE(TX_TYPES)); + AVG_CDF_STRIDE(ctx_left->inter_ext_tx_cdf[1], ctx_tr->inter_ext_tx_cdf[1], 16, + CDF_SIZE(TX_TYPES)); + AVG_CDF_STRIDE(ctx_left->inter_ext_tx_cdf[2], ctx_tr->inter_ext_tx_cdf[2], 12, + CDF_SIZE(TX_TYPES)); + AVG_CDF_STRIDE(ctx_left->inter_ext_tx_cdf[3], ctx_tr->inter_ext_tx_cdf[3], 2, + CDF_SIZE(TX_TYPES)); + AVERAGE_CDF(ctx_left->cfl_sign_cdf, ctx_tr->cfl_sign_cdf, CFL_JOINT_SIGNS); + AVERAGE_CDF(ctx_left->cfl_alpha_cdf, ctx_tr->cfl_alpha_cdf, + CFL_ALPHABET_SIZE); +} + +#if !CONFIG_REALTIME_ONLY +static AOM_INLINE void adjust_rdmult_tpl_model(AV1_COMP *cpi, MACROBLOCK *x, + int mi_row, int mi_col) { + const BLOCK_SIZE sb_size = cpi->common.seq_params.sb_size; + const int orig_rdmult = cpi->rd.RDMULT; + + assert(IMPLIES(cpi->gf_group.size > 0, + cpi->gf_group.index < cpi->gf_group.size)); + const int gf_group_index = cpi->gf_group.index; + if (cpi->oxcf.enable_tpl_model && cpi->oxcf.aq_mode == NO_AQ && + cpi->oxcf.deltaq_mode == NO_DELTA_Q && gf_group_index > 0 && + cpi->gf_group.update_type[gf_group_index] == ARF_UPDATE) { + const int dr = + get_rdmult_delta(cpi, sb_size, 0, mi_row, mi_col, orig_rdmult); + x->rdmult = dr; + } +} +#endif + +static void source_content_sb(AV1_COMP *cpi, MACROBLOCK *x, int shift) { + unsigned int tmp_sse; + unsigned int tmp_variance; + const BLOCK_SIZE bsize = BLOCK_64X64; + uint8_t *src_y = cpi->source->y_buffer; + int src_ystride = cpi->source->y_stride; + uint8_t *last_src_y = cpi->last_source->y_buffer; + int last_src_ystride = cpi->last_source->y_stride; + uint64_t avg_source_sse_threshold = 100000; // ~5*5*(64*64) + uint64_t avg_source_sse_threshold_high = 1000000; // ~15*15*(64*64) + uint64_t sum_sq_thresh = 10000; // sum = sqrt(thresh / 64*64)) ~1.5 +#if CONFIG_AV1_HIGHBITDEPTH + MACROBLOCKD *xd = &x->e_mbd; + if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) return; +#endif + src_y += shift; + last_src_y += shift; + tmp_variance = cpi->fn_ptr[bsize].vf(src_y, src_ystride, last_src_y, + last_src_ystride, &tmp_sse); + // Note: tmp_sse - tmp_variance = ((sum * sum) >> 12) + // Detect large lighting change. + if (tmp_variance < (tmp_sse >> 1) && (tmp_sse - tmp_variance) > sum_sq_thresh) + x->content_state_sb = kLowVarHighSumdiff; + else if (tmp_sse < avg_source_sse_threshold) + x->content_state_sb = kLowSad; + else if (tmp_sse > avg_source_sse_threshold_high) + x->content_state_sb = kHighSad; +} + +static AOM_INLINE void encode_nonrd_sb(AV1_COMP *cpi, ThreadData *td, + TileDataEnc *tile_data, + PC_TREE *const pc_root, TOKENEXTRA **tp, + const int mi_row, const int mi_col, + const int seg_skip) { + AV1_COMMON *const cm = &cpi->common; + MACROBLOCK *const x = &td->mb; + const SPEED_FEATURES *const sf = &cpi->sf; + const TileInfo *const tile_info = &tile_data->tile_info; + MB_MODE_INFO **mi = cm->mi_params.mi_grid_base + + get_mi_grid_idx(&cm->mi_params, mi_row, mi_col); + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + if (sf->rt_sf.source_metrics_sb_nonrd && sb_size == BLOCK_64X64 && + cpi->svc.number_spatial_layers <= 1 && + cm->current_frame.frame_type != KEY_FRAME) { + int shift = cpi->source->y_stride * (mi_row << 2) + (mi_col << 2); + source_content_sb(cpi, x, shift); + } + if (sf->part_sf.partition_search_type == FIXED_PARTITION || seg_skip) { + set_offsets(cpi, tile_info, x, mi_row, mi_col, sb_size); + const BLOCK_SIZE bsize = + seg_skip ? sb_size : sf->part_sf.always_this_block_size; + set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize); + } else if (cpi->partition_search_skippable_frame) { + set_offsets(cpi, tile_info, x, mi_row, mi_col, sb_size); + const BLOCK_SIZE bsize = + get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col); + set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize); + } else if (sf->part_sf.partition_search_type == VAR_BASED_PARTITION) { + set_offsets_without_segment_id(cpi, tile_info, x, mi_row, mi_col, sb_size); + av1_choose_var_based_partitioning(cpi, tile_info, td, x, mi_row, mi_col); + } + assert(sf->part_sf.partition_search_type == FIXED_PARTITION || seg_skip || + cpi->partition_search_skippable_frame || + sf->part_sf.partition_search_type == VAR_BASED_PARTITION); + td->mb.cb_offset = 0; + nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, sb_size, + pc_root); +} + +// Memset the mbmis at the current superblock to 0 +static INLINE void reset_mbmi(CommonModeInfoParams *const mi_params, + BLOCK_SIZE sb_size, int mi_row, int mi_col) { + // size of sb in unit of mi (BLOCK_4X4) + const int sb_size_mi = mi_size_wide[sb_size]; + const int mi_alloc_size_1d = mi_size_wide[mi_params->mi_alloc_bsize]; + // size of sb in unit of allocated mi size + const int sb_size_alloc_mi = mi_size_wide[sb_size] / mi_alloc_size_1d; + assert(mi_params->mi_alloc_stride % sb_size_alloc_mi == 0 && + "mi is not allocated as a multiple of sb!"); + assert(mi_params->mi_stride % sb_size_mi == 0 && + "mi_grid_base is not allocated as a multiple of sb!"); + + const int mi_rows = mi_size_high[sb_size]; + for (int cur_mi_row = 0; cur_mi_row < mi_rows; cur_mi_row++) { + assert(get_mi_grid_idx(mi_params, 0, mi_col + mi_alloc_size_1d) < + mi_params->mi_stride); + const int mi_grid_idx = + get_mi_grid_idx(mi_params, mi_row + cur_mi_row, mi_col); + const int alloc_mi_idx = + get_alloc_mi_idx(mi_params, mi_row + cur_mi_row, mi_col); + memset(&mi_params->mi_grid_base[mi_grid_idx], 0, + sb_size_mi * sizeof(*mi_params->mi_grid_base)); + memset(&mi_params->tx_type_map[mi_grid_idx], 0, + sb_size_mi * sizeof(*mi_params->tx_type_map)); + if (cur_mi_row % mi_alloc_size_1d == 0) { + memset(&mi_params->mi_alloc[alloc_mi_idx], 0, + sb_size_alloc_mi * sizeof(*mi_params->mi_alloc)); + } + } } -// Set all the counters as max. -static void init_first_partition_pass_stats_tables( - FIRST_PARTITION_PASS_STATS *stats) { - for (int i = 0; i < FIRST_PARTITION_PASS_STATS_TABLES; ++i) { - memset(stats[i].ref0_counts, 0xff, sizeof(stats[i].ref0_counts)); - memset(stats[i].ref1_counts, 0xff, sizeof(stats[i].ref1_counts)); - stats[i].sample_counts = INT_MAX; +static INLINE void backup_sb_state(SB_FIRST_PASS_STATS *sb_fp_stats, + const AV1_COMP *cpi, ThreadData *td, + const TileDataEnc *tile_data, int mi_row, + int mi_col) { + MACROBLOCK *x = &td->mb; + MACROBLOCKD *xd = &x->e_mbd; + const TileInfo *tile_info = &tile_data->tile_info; + + const AV1_COMMON *cm = &cpi->common; + const int num_planes = av1_num_planes(cm); + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + + xd->above_txfm_context = + cm->above_contexts.txfm[tile_info->tile_row] + mi_col; + xd->left_txfm_context = + xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); + save_context(x, &sb_fp_stats->x_ctx, mi_row, mi_col, sb_size, num_planes); + + sb_fp_stats->rd_count = cpi->td.rd_counts; + sb_fp_stats->split_count = cpi->td.mb.txb_split_count; + + sb_fp_stats->fc = *td->counts; + + memcpy(sb_fp_stats->inter_mode_rd_models, tile_data->inter_mode_rd_models, + sizeof(sb_fp_stats->inter_mode_rd_models)); + + memcpy(sb_fp_stats->thresh_freq_fact, x->thresh_freq_fact, + sizeof(sb_fp_stats->thresh_freq_fact)); + + const int alloc_mi_idx = get_alloc_mi_idx(&cm->mi_params, mi_row, mi_col); + sb_fp_stats->current_qindex = + cm->mi_params.mi_alloc[alloc_mi_idx].current_qindex; + +#if CONFIG_INTERNAL_STATS + memcpy(sb_fp_stats->mode_chosen_counts, cpi->mode_chosen_counts, + sizeof(sb_fp_stats->mode_chosen_counts)); +#endif // CONFIG_INTERNAL_STATS +} + +static INLINE void restore_sb_state(const SB_FIRST_PASS_STATS *sb_fp_stats, + AV1_COMP *cpi, ThreadData *td, + TileDataEnc *tile_data, int mi_row, + int mi_col) { + MACROBLOCK *x = &td->mb; + + const AV1_COMMON *cm = &cpi->common; + const int num_planes = av1_num_planes(cm); + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + + restore_context(x, &sb_fp_stats->x_ctx, mi_row, mi_col, sb_size, num_planes); + + cpi->td.rd_counts = sb_fp_stats->rd_count; + cpi->td.mb.txb_split_count = sb_fp_stats->split_count; + + *td->counts = sb_fp_stats->fc; + + memcpy(tile_data->inter_mode_rd_models, sb_fp_stats->inter_mode_rd_models, + sizeof(sb_fp_stats->inter_mode_rd_models)); + memcpy(x->thresh_freq_fact, sb_fp_stats->thresh_freq_fact, + sizeof(sb_fp_stats->thresh_freq_fact)); + + const int alloc_mi_idx = get_alloc_mi_idx(&cm->mi_params, mi_row, mi_col); + cm->mi_params.mi_alloc[alloc_mi_idx].current_qindex = + sb_fp_stats->current_qindex; + +#if CONFIG_INTERNAL_STATS + memcpy(cpi->mode_chosen_counts, sb_fp_stats->mode_chosen_counts, + sizeof(sb_fp_stats->mode_chosen_counts)); +#endif // CONFIG_INTERNAL_STATS +} + +#if !CONFIG_REALTIME_ONLY +static void init_ref_frame_space(AV1_COMP *cpi, ThreadData *td, int mi_row, + int mi_col) { + const AV1_COMMON *cm = &cpi->common; + const CommonModeInfoParams *const mi_params = &cm->mi_params; + MACROBLOCK *x = &td->mb; + const int frame_idx = cpi->gf_group.index; + TplParams *const tpl_data = &cpi->tpl_data; + TplDepFrame *tpl_frame = &tpl_data->tpl_frame[frame_idx]; + const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2; + + av1_zero(x->search_ref_frame); + + if (tpl_frame->is_valid == 0) return; + if (!is_frame_tpl_eligible(cpi)) return; + if (frame_idx >= MAX_LAG_BUFFERS) return; + if (cpi->superres_mode != SUPERRES_NONE) return; + if (cpi->oxcf.aq_mode != NO_AQ) return; + + const int is_overlay = cpi->gf_group.update_type[frame_idx] == OVERLAY_UPDATE; + if (is_overlay) { + memset(x->search_ref_frame, 1, sizeof(x->search_ref_frame)); + return; + } + + TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr; + const int tpl_stride = tpl_frame->stride; + int64_t inter_cost[INTER_REFS_PER_FRAME] = { 0 }; + const int step = 1 << block_mis_log2; + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + const int mi_row_end = + AOMMIN(mi_size_high[sb_size] + mi_row, mi_params->mi_rows); + const int mi_col_end = + AOMMIN(mi_size_wide[sb_size] + mi_col, mi_params->mi_cols); + + for (int row = mi_row; row < mi_row_end; row += step) { + for (int col = mi_col; col < mi_col_end; col += step) { + const TplDepStats *this_stats = + &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)]; + int64_t tpl_pred_error[INTER_REFS_PER_FRAME] = { 0 }; + // Find the winner ref frame idx for the current block + int64_t best_inter_cost = this_stats->pred_error[0]; + int best_rf_idx = 0; + for (int idx = 1; idx < INTER_REFS_PER_FRAME; ++idx) { + if ((this_stats->pred_error[idx] < best_inter_cost) && + (this_stats->pred_error[idx] != 0)) { + best_inter_cost = this_stats->pred_error[idx]; + best_rf_idx = idx; + } + } + // tpl_pred_error is the pred_error reduction of best_ref w.r.t. + // LAST_FRAME. + tpl_pred_error[best_rf_idx] = this_stats->pred_error[best_rf_idx] - + this_stats->pred_error[LAST_FRAME - 1]; + + for (int rf_idx = 1; rf_idx < INTER_REFS_PER_FRAME; ++rf_idx) + inter_cost[rf_idx] += tpl_pred_error[rf_idx]; + } + } + + int rank_index[INTER_REFS_PER_FRAME - 1]; + for (int idx = 0; idx < INTER_REFS_PER_FRAME - 1; ++idx) { + rank_index[idx] = idx + 1; + for (int i = idx; i > 0; --i) { + if (inter_cost[rank_index[i - 1]] > inter_cost[rank_index[i]]) { + const int tmp = rank_index[i - 1]; + rank_index[i - 1] = rank_index[i]; + rank_index[i] = tmp; + } + } + } + + x->search_ref_frame[INTRA_FRAME] = 1; + x->search_ref_frame[LAST_FRAME] = 1; + + int cutoff_ref = 0; + for (int idx = 0; idx < INTER_REFS_PER_FRAME - 1; ++idx) { + x->search_ref_frame[rank_index[idx] + LAST_FRAME] = 1; + if (idx > 2) { + if (!cutoff_ref) { + // If the predictive coding gains are smaller than the previous more + // relevant frame over certain amount, discard this frame and all the + // frames afterwards. + if (llabs(inter_cost[rank_index[idx]]) < + llabs(inter_cost[rank_index[idx - 1]]) / 8 || + inter_cost[rank_index[idx]] == 0) + cutoff_ref = 1; + } + + if (cutoff_ref) x->search_ref_frame[rank_index[idx] + LAST_FRAME] = 0; + } } } +#endif // !CONFIG_REALTIME_ONLY + +// This function initializes the stats for encode_rd_sb. +static INLINE void init_encode_rd_sb(AV1_COMP *cpi, ThreadData *td, + const TileDataEnc *tile_data, + PC_TREE *pc_root, RD_STATS *rd_cost, + int mi_row, int mi_col, + int gather_tpl_data) { + const AV1_COMMON *cm = &cpi->common; + const TileInfo *tile_info = &tile_data->tile_info; + MACROBLOCK *x = &td->mb; -// clear pc_tree_stats -static INLINE void clear_pc_tree_stats(PC_TREE *pt) { - if (pt == NULL) return; - pt->pc_tree_stats.valid = 0; - for (int i = 0; i < 4; ++i) { - clear_pc_tree_stats(pt->split[i]); + const SPEED_FEATURES *sf = &cpi->sf; + const int use_simple_motion_search = + (sf->part_sf.simple_motion_search_split || + sf->part_sf.simple_motion_search_prune_rect || + sf->part_sf.simple_motion_search_early_term_none || + sf->part_sf.ml_early_term_after_part_split_level) && + !frame_is_intra_only(cm); + if (use_simple_motion_search) { + init_simple_motion_search_mvs(pc_root); } + +#if !CONFIG_REALTIME_ONLY + init_ref_frame_space(cpi, td, mi_row, mi_col); + x->sb_energy_level = 0; + x->cnn_output_valid = 0; + if (gather_tpl_data) { + if (cm->delta_q_info.delta_q_present_flag) { + const int num_planes = av1_num_planes(cm); + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + setup_delta_q(cpi, td, x, tile_info, mi_row, mi_col, num_planes); + av1_tpl_rdmult_setup_sb(cpi, x, sb_size, mi_row, mi_col); + } + if (cpi->oxcf.enable_tpl_model) { + adjust_rdmult_tpl_model(cpi, x, mi_row, mi_col); + } + } +#else + (void)tile_info; + (void)mi_row; + (void)mi_col; + (void)gather_tpl_data; +#endif + + // Reset hash state for transform/mode rd hash information + reset_hash_records(x, cpi->sf.tx_sf.use_inter_txb_hash); + av1_zero(x->picked_ref_frames_mask); + av1_zero(x->pred_mv); + av1_invalid_rd_stats(rd_cost); } -// Minimum number of samples to trigger the -// mode_pruning_based_on_two_pass_partition_search feature. -#define FIRST_PARTITION_PASS_MIN_SAMPLES 16 +static AOM_INLINE void encode_rd_sb(AV1_COMP *cpi, ThreadData *td, + TileDataEnc *tile_data, + PC_TREE *const pc_root, TOKENEXTRA **tp, + const int mi_row, const int mi_col, + const int seg_skip) { + AV1_COMMON *const cm = &cpi->common; + MACROBLOCK *const x = &td->mb; + const SPEED_FEATURES *const sf = &cpi->sf; + const TileInfo *const tile_info = &tile_data->tile_info; + MB_MODE_INFO **mi = cm->mi_params.mi_grid_base + + get_mi_grid_idx(&cm->mi_params, mi_row, mi_col); + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + int dummy_rate; + int64_t dummy_dist; + RD_STATS dummy_rdc; + +#if CONFIG_REALTIME_ONLY + (void)seg_skip; +#endif // CONFIG_REALTIME_ONLY + + init_encode_rd_sb(cpi, td, tile_data, pc_root, &dummy_rdc, mi_row, mi_col, 1); + + if (sf->part_sf.partition_search_type == VAR_BASED_PARTITION) { + set_offsets_without_segment_id(cpi, tile_info, x, mi_row, mi_col, sb_size); + av1_choose_var_based_partitioning(cpi, tile_info, td, x, mi_row, mi_col); + rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, sb_size, + &dummy_rate, &dummy_dist, 1, pc_root); + } +#if !CONFIG_REALTIME_ONLY + else if (sf->part_sf.partition_search_type == FIXED_PARTITION || seg_skip) { + set_offsets(cpi, tile_info, x, mi_row, mi_col, sb_size); + const BLOCK_SIZE bsize = + seg_skip ? sb_size : sf->part_sf.always_this_block_size; + set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize); + rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, sb_size, + &dummy_rate, &dummy_dist, 1, pc_root); + } else if (cpi->partition_search_skippable_frame) { + set_offsets(cpi, tile_info, x, mi_row, mi_col, sb_size); + const BLOCK_SIZE bsize = + get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col); + set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize); + rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, sb_size, + &dummy_rate, &dummy_dist, 1, pc_root); + } else { + // No stats for overlay frames. Exclude key frame. + x->valid_cost_b = + get_tpl_stats_b(cpi, sb_size, mi_row, mi_col, x->intra_cost_b, + x->inter_cost_b, x->mv_b, &x->cost_stride); + + reset_partition(pc_root, sb_size); + +#if CONFIG_COLLECT_COMPONENT_TIMING + start_timing(cpi, rd_pick_partition_time); +#endif + BLOCK_SIZE max_sq_size = x->max_partition_size; + BLOCK_SIZE min_sq_size = x->min_partition_size; + + if (use_auto_max_partition(cpi, sb_size, mi_row, mi_col)) { + float features[FEATURE_SIZE_MAX_MIN_PART_PRED] = { 0.0f }; + + av1_get_max_min_partition_features(cpi, x, mi_row, mi_col, features); + max_sq_size = AOMMAX( + AOMMIN(av1_predict_max_partition(cpi, x, features), max_sq_size), + min_sq_size); + } + + const int num_passes = cpi->oxcf.sb_multipass_unit_test ? 2 : 1; -static void encode_rd_sb_row(AV1_COMP *cpi, ThreadData *td, - TileDataEnc *tile_data, int mi_row, - TOKENEXTRA **tp) { + if (num_passes == 1) { + rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, sb_size, + max_sq_size, min_sq_size, &dummy_rdc, dummy_rdc, + pc_root, NULL, SB_SINGLE_PASS, NULL); + } else { + // First pass + SB_FIRST_PASS_STATS sb_fp_stats; + backup_sb_state(&sb_fp_stats, cpi, td, tile_data, mi_row, mi_col); + rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, sb_size, + max_sq_size, min_sq_size, &dummy_rdc, dummy_rdc, + pc_root, NULL, SB_DRY_PASS, NULL); + + // Second pass + init_encode_rd_sb(cpi, td, tile_data, pc_root, &dummy_rdc, mi_row, mi_col, + 0); + reset_mbmi(&cm->mi_params, sb_size, mi_row, mi_col); + reset_partition(pc_root, sb_size); + + restore_sb_state(&sb_fp_stats, cpi, td, tile_data, mi_row, mi_col); + + rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, sb_size, + max_sq_size, min_sq_size, &dummy_rdc, dummy_rdc, + pc_root, NULL, SB_WET_PASS, NULL); + } + // Reset to 0 so that it wouldn't be used elsewhere mistakenly. + x->valid_cost_b = 0; +#if CONFIG_COLLECT_COMPONENT_TIMING + end_timing(cpi, rd_pick_partition_time); +#endif + } +#endif // !CONFIG_REALTIME_ONLY + + // TODO(angiebird): Let inter_mode_rd_model_estimation support multi-tile. + if (cpi->sf.inter_sf.inter_mode_rd_model_estimation == 1 && + cm->tiles.cols == 1 && cm->tiles.rows == 1) { + av1_inter_mode_data_fit(tile_data, x->rdmult); + } +} + +static AOM_INLINE void set_cost_upd_freq(AV1_COMP *cpi, ThreadData *td, + const TileInfo *const tile_info, + const int mi_row, const int mi_col) { AV1_COMMON *const cm = &cpi->common; const int num_planes = av1_num_planes(cm); + MACROBLOCK *const x = &td->mb; + MACROBLOCKD *const xd = &x->e_mbd; + + switch (cpi->oxcf.coeff_cost_upd_freq) { + case COST_UPD_TILE: // Tile level + if (mi_row != tile_info->mi_row_start) break; + AOM_FALLTHROUGH_INTENDED; + case COST_UPD_SBROW: // SB row level in tile + if (mi_col != tile_info->mi_col_start) break; + AOM_FALLTHROUGH_INTENDED; + case COST_UPD_SB: // SB level + if (cpi->sf.inter_sf.disable_sb_level_coeff_cost_upd && + mi_col != tile_info->mi_col_start) + break; + av1_fill_coeff_costs(&td->mb, xd->tile_ctx, num_planes); + break; + default: assert(0); + } + + switch (cpi->oxcf.mode_cost_upd_freq) { + case COST_UPD_TILE: // Tile level + if (mi_row != tile_info->mi_row_start) break; + AOM_FALLTHROUGH_INTENDED; + case COST_UPD_SBROW: // SB row level in tile + if (mi_col != tile_info->mi_col_start) break; + AOM_FALLTHROUGH_INTENDED; + case COST_UPD_SB: // SB level + av1_fill_mode_rates(cm, x, xd->tile_ctx); + break; + default: assert(0); + } + switch (cpi->oxcf.mv_cost_upd_freq) { + case COST_UPD_OFF: break; + case COST_UPD_TILE: // Tile level + if (mi_row != tile_info->mi_row_start) break; + AOM_FALLTHROUGH_INTENDED; + case COST_UPD_SBROW: // SB row level in tile + if (mi_col != tile_info->mi_col_start) break; + AOM_FALLTHROUGH_INTENDED; + case COST_UPD_SB: // SB level + if (cpi->sf.inter_sf.disable_sb_level_mv_cost_upd && + mi_col != tile_info->mi_col_start) + break; + av1_fill_mv_costs(xd->tile_ctx, cm->features.cur_frame_force_integer_mv, + cm->features.allow_high_precision_mv, x); + break; + default: assert(0); + } +} + +static AOM_INLINE void encode_sb_row(AV1_COMP *cpi, ThreadData *td, + TileDataEnc *tile_data, int mi_row, + TOKENEXTRA **tp) { + AV1_COMMON *const cm = &cpi->common; const TileInfo *const tile_info = &tile_data->tile_info; MACROBLOCK *const x = &td->mb; MACROBLOCKD *const xd = &x->e_mbd; - SPEED_FEATURES *const sf = &cpi->sf; - const int leaf_nodes = 256; + const int sb_cols_in_tile = av1_get_sb_cols_in_tile(cm, tile_data->tile_info); + const BLOCK_SIZE sb_size = cm->seq_params.sb_size; + const int mib_size = cm->seq_params.mib_size; + const int mib_size_log2 = cm->seq_params.mib_size_log2; + const int sb_row = (mi_row - tile_info->mi_row_start) >> mib_size_log2; + const int use_nonrd_mode = cpi->sf.rt_sf.use_nonrd_pick_mode; + +#if CONFIG_COLLECT_COMPONENT_TIMING + start_timing(cpi, encode_sb_time); +#endif // Initialize the left context for the new SB row av1_zero_left_context(xd); // Reset delta for every tile - if (mi_row == tile_info->mi_row_start) { - if (cm->delta_q_present_flag) xd->current_qindex = cm->base_qindex; - if (cm->delta_lf_present_flag) { + if (mi_row == tile_info->mi_row_start || cpi->row_mt) { + if (cm->delta_q_info.delta_q_present_flag) + xd->current_qindex = cm->quant_params.base_qindex; + if (cm->delta_q_info.delta_lf_present_flag) { av1_reset_loop_filter_delta(xd, av1_num_planes(cm)); } } + reset_thresh_freq_fact(x); - PC_TREE *const pc_root = - td->pc_root[cm->seq_params.mib_size_log2 - MIN_MIB_SIZE_LOG2]; // Code each SB in the row - for (int mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end; - mi_col += cm->seq_params.mib_size) { - av1_fill_coeff_costs(&td->mb, xd->tile_ctx, num_planes); - av1_fill_mode_rates(cm, x, xd->tile_ctx); - - if (sf->adaptive_pred_interp_filter) { - for (int i = 0; i < leaf_nodes; ++i) { - td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE; - td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE; - td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE; - td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE; + for (int mi_col = tile_info->mi_col_start, sb_col_in_tile = 0; + mi_col < tile_info->mi_col_end; mi_col += mib_size, sb_col_in_tile++) { + (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row, + sb_col_in_tile); + if (tile_data->allow_update_cdf && (cpi->row_mt == 1) && + (tile_info->mi_row_start != mi_row)) { + if ((tile_info->mi_col_start == mi_col)) { + // restore frame context of 1st column sb + memcpy(xd->tile_ctx, x->row_ctx, sizeof(*xd->tile_ctx)); + } else { + int wt_left = AVG_CDF_WEIGHT_LEFT; + int wt_tr = AVG_CDF_WEIGHT_TOP_RIGHT; + if (tile_info->mi_col_end > (mi_col + mib_size)) + avg_cdf_symbols(xd->tile_ctx, x->row_ctx + sb_col_in_tile, wt_left, + wt_tr); + else + avg_cdf_symbols(xd->tile_ctx, x->row_ctx + sb_col_in_tile - 1, + wt_left, wt_tr); } } - x->mb_rd_record.num = x->mb_rd_record.index_start = 0; + set_cost_upd_freq(cpi, td, tile_info, mi_row, mi_col); - av1_zero(x->txb_rd_record_8X8); - av1_zero(x->txb_rd_record_16X16); - av1_zero(x->txb_rd_record_32X32); - av1_zero(x->txb_rd_record_64X64); - av1_zero(x->txb_rd_record_intra); + x->color_sensitivity[0] = 0; + x->color_sensitivity[1] = 0; + x->content_state_sb = 0; - av1_zero(x->pred_mv); + PC_TREE *const pc_root = td->pc_root; pc_root->index = 0; + xd->cur_frame_force_integer_mv = cm->features.cur_frame_force_integer_mv; + td->mb.cb_coef_buff = av1_get_cb_coeff_buffer(cpi, mi_row, mi_col); + x->source_variance = UINT_MAX; + x->simple_motion_pred_sse = UINT_MAX; + const struct segmentation *const seg = &cm->seg; int seg_skip = 0; if (seg->enabled) { const uint8_t *const map = - seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map; + seg->update_map ? cpi->enc_seg.map : cm->last_frame_seg_map; const int segment_id = - map ? get_segment_id(cm, map, cm->seq_params.sb_size, mi_row, mi_col) + map ? get_segment_id(&cm->mi_params, map, sb_size, mi_row, mi_col) : 0; seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP); } - xd->cur_frame_force_integer_mv = cm->cur_frame_force_integer_mv; - x->sb_energy_level = 0; - if (cm->delta_q_present_flag) { - // Delta-q modulation based on variance - av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes); - - int offset_qindex; - if (DELTAQ_MODULATION == 1) { - const int block_wavelet_energy_level = - av1_block_wavelet_energy_level(cpi, x, cm->seq_params.sb_size); - x->sb_energy_level = block_wavelet_energy_level; - offset_qindex = av1_compute_deltaq_from_energy_level( - cpi, block_wavelet_energy_level); - } else { - const int block_var_level = - av1_log_block_var(cpi, x, cm->seq_params.sb_size); - x->sb_energy_level = block_var_level; - offset_qindex = - av1_compute_deltaq_from_energy_level(cpi, block_var_level); - } - const int qmask = ~(cm->delta_q_res - 1); - int current_qindex = clamp(cm->base_qindex + offset_qindex, - cm->delta_q_res, 256 - cm->delta_q_res); - current_qindex = - ((current_qindex - cm->base_qindex + cm->delta_q_res / 2) & qmask) + - cm->base_qindex; - assert(current_qindex > 0); - - xd->delta_qindex = current_qindex - cm->base_qindex; - set_offsets(cpi, tile_info, x, mi_row, mi_col, cm->seq_params.sb_size); - xd->mi[0]->current_qindex = current_qindex; - av1_init_plane_quantizers(cpi, x, xd->mi[0]->segment_id); - if (cpi->oxcf.deltaq_mode == DELTA_Q_LF) { - const int lfmask = ~(cm->delta_lf_res - 1); - const int delta_lf_from_base = - ((offset_qindex / 2 + cm->delta_lf_res / 2) & lfmask); - - // pre-set the delta lf for loop filter. Note that this value is set - // before mi is assigned for each block in current superblock - for (int j = 0; - j < AOMMIN(cm->seq_params.mib_size, cm->mi_rows - mi_row); j++) { - for (int k = 0; - k < AOMMIN(cm->seq_params.mib_size, cm->mi_cols - mi_col); k++) { - cm->mi[(mi_row + j) * cm->mi_stride + (mi_col + k)] - .delta_lf_from_base = - clamp(delta_lf_from_base, -MAX_LOOP_FILTER, MAX_LOOP_FILTER); - const int frame_lf_count = - av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2; - for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) { - cm->mi[(mi_row + j) * cm->mi_stride + (mi_col + k)] - .delta_lf[lf_id] = - clamp(delta_lf_from_base, -MAX_LOOP_FILTER, MAX_LOOP_FILTER); - } - } - } - } - } - - int dummy_rate; - int64_t dummy_dist; - RD_STATS dummy_rdc; - const int idx_str = cm->mi_stride * mi_row + mi_col; - MB_MODE_INFO **mi = cm->mi_grid_visible + idx_str; - x->source_variance = UINT_MAX; - if (sf->partition_search_type == FIXED_PARTITION || seg_skip) { - set_offsets(cpi, tile_info, x, mi_row, mi_col, cm->seq_params.sb_size); - const BLOCK_SIZE bsize = - seg_skip ? cm->seq_params.sb_size : sf->always_this_block_size; - set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize); - rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, - cm->seq_params.sb_size, &dummy_rate, &dummy_dist, 1, - pc_root); - } else if (cpi->partition_search_skippable_frame) { - set_offsets(cpi, tile_info, x, mi_row, mi_col, cm->seq_params.sb_size); - const BLOCK_SIZE bsize = - get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col); - set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize); - rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, - cm->seq_params.sb_size, &dummy_rate, &dummy_dist, 1, - pc_root); + if (use_nonrd_mode) { + encode_nonrd_sb(cpi, td, tile_data, pc_root, tp, mi_row, mi_col, + seg_skip); } else { - // If required set upper and lower partition size limits - if (sf->auto_min_max_partition_size) { - set_offsets(cpi, tile_info, x, mi_row, mi_col, cm->seq_params.sb_size); - rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col, - &x->min_partition_size, &x->max_partition_size); - } - - reset_partition(pc_root, cm->seq_params.sb_size); - x->use_cb_search_range = 0; - init_first_partition_pass_stats_tables(x->first_partition_pass_stats); - // Do the first pass if we need two pass partition search - if (cpi->sf.two_pass_partition_search && - cpi->sf.use_square_partition_only_threshold > BLOCK_4X4 && - mi_row + mi_size_high[cm->seq_params.sb_size] < cm->mi_rows && - mi_col + mi_size_wide[cm->seq_params.sb_size] < cm->mi_cols && - cm->frame_type != KEY_FRAME) { - x->cb_partition_scan = 1; - // Reset the stats tables. - if (sf->mode_pruning_based_on_two_pass_partition_search) - av1_zero(x->first_partition_pass_stats); - clear_pc_tree_stats(pc_root); - rd_pick_sqr_partition(cpi, td, tile_data, tp, mi_row, mi_col, - cm->seq_params.sb_size, &dummy_rdc, INT64_MAX, - pc_root, NULL); - x->cb_partition_scan = 0; - - x->source_variance = UINT_MAX; - if (sf->adaptive_pred_interp_filter) { - for (int i = 0; i < leaf_nodes; ++i) { - td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE; - td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE; - td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE; - td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE; - } - } - - x->mb_rd_record.num = x->mb_rd_record.index_start = 0; - av1_zero(x->txb_rd_record_8X8); - av1_zero(x->txb_rd_record_16X16); - av1_zero(x->txb_rd_record_32X32); - av1_zero(x->txb_rd_record_64X64); - av1_zero(x->txb_rd_record_intra); - av1_zero(x->pred_mv); - pc_root->index = 0; - - for (int idy = 0; idy < mi_size_high[cm->seq_params.sb_size]; ++idy) { - for (int idx = 0; idx < mi_size_wide[cm->seq_params.sb_size]; ++idx) { - const int offset = cm->mi_stride * (mi_row + idy) + (mi_col + idx); - cm->mi_grid_visible[offset] = 0; - } - } - - x->use_cb_search_range = 1; - - if (sf->mode_pruning_based_on_two_pass_partition_search) { - for (int i = 0; i < FIRST_PARTITION_PASS_STATS_TABLES; ++i) { - FIRST_PARTITION_PASS_STATS *const stat = - &x->first_partition_pass_stats[i]; - if (stat->sample_counts < FIRST_PARTITION_PASS_MIN_SAMPLES) { - // If there are not enough samples collected, make all available. - memset(stat->ref0_counts, 0xff, sizeof(stat->ref0_counts)); - memset(stat->ref1_counts, 0xff, sizeof(stat->ref1_counts)); - } else if (sf->selective_ref_frame < 2) { - // ALTREF2_FRAME and BWDREF_FRAME may be skipped during the - // initial partition scan, so we don't eliminate them. - stat->ref0_counts[ALTREF2_FRAME] = 0xff; - stat->ref1_counts[ALTREF2_FRAME] = 0xff; - stat->ref0_counts[BWDREF_FRAME] = 0xff; - stat->ref1_counts[BWDREF_FRAME] = 0xff; - } - } - } - } - - rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, - cm->seq_params.sb_size, &dummy_rdc, INT64_MAX, pc_root, - NULL); + encode_rd_sb(cpi, td, tile_data, pc_root, tp, mi_row, mi_col, seg_skip); } -#if CONFIG_COLLECT_INTER_MODE_RD_STATS - // TODO(angiebird): Let inter_mode_rd_model_estimation support multi-tile. - if (cpi->sf.inter_mode_rd_model_estimation && cm->tile_cols == 1 && - cm->tile_rows == 1) { - av1_inter_mode_data_fit(tile_data, x->rdmult); + + if (tile_data->allow_update_cdf && (cpi->row_mt == 1) && + (tile_info->mi_row_end > (mi_row + mib_size))) { + if (sb_cols_in_tile == 1) + memcpy(x->row_ctx, xd->tile_ctx, sizeof(*xd->tile_ctx)); + else if (sb_col_in_tile >= 1) + memcpy(x->row_ctx + sb_col_in_tile - 1, xd->tile_ctx, + sizeof(*xd->tile_ctx)); } -#endif + (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, sb_row, + sb_col_in_tile, sb_cols_in_tile); } +#if CONFIG_COLLECT_COMPONENT_TIMING + end_timing(cpi, encode_sb_time); +#endif } -static void init_encode_frame_mb_context(AV1_COMP *cpi) { +static AOM_INLINE void init_encode_frame_mb_context(AV1_COMP *cpi) { AV1_COMMON *const cm = &cpi->common; const int num_planes = av1_num_planes(cm); MACROBLOCK *const x = &cpi->td.mb; MACROBLOCKD *const xd = &x->e_mbd; // Copy data over into macro block data structures. - av1_setup_src_planes(x, cpi->source, 0, 0, num_planes); + av1_setup_src_planes(x, cpi->source, 0, 0, num_planes, + cm->seq_params.sb_size); av1_setup_block_planes(xd, cm->seq_params.subsampling_x, cm->seq_params.subsampling_y, num_planes); } -static MV_REFERENCE_FRAME get_frame_type(const AV1_COMP *cpi) { - if (frame_is_intra_only(&cpi->common)) return INTRA_FRAME; - // We will not update the golden frame with an internal overlay frame - else if ((cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame) || - cpi->rc.is_src_frame_ext_arf) - return ALTREF_FRAME; - else if (cpi->refresh_golden_frame || cpi->refresh_alt2_ref_frame || - cpi->refresh_alt_ref_frame) - return GOLDEN_FRAME; - else - // TODO(zoeliu): To investigate whether a frame_type other than - // INTRA/ALTREF/GOLDEN/LAST needs to be specified seperately. - return LAST_FRAME; -} - -static TX_MODE select_tx_mode(const AV1_COMP *cpi) { - if (cpi->common.coded_lossless) return ONLY_4X4; - if (cpi->sf.tx_size_search_method == USE_LARGESTALL) - return TX_MODE_LARGEST; - else if (cpi->sf.tx_size_search_method == USE_FULL_RD || - cpi->sf.tx_size_search_method == USE_FAST_RD) - return TX_MODE_SELECT; - else - return cpi->common.tx_mode; -} - void av1_alloc_tile_data(AV1_COMP *cpi) { AV1_COMMON *const cm = &cpi->common; - const int tile_cols = cm->tile_cols; - const int tile_rows = cm->tile_rows; - int tile_col, tile_row; + const int tile_cols = cm->tiles.cols; + const int tile_rows = cm->tiles.rows; if (cpi->tile_data != NULL) aom_free(cpi->tile_data); CHECK_MEM_ERROR( cm, cpi->tile_data, aom_memalign(32, tile_cols * tile_rows * sizeof(*cpi->tile_data))); - cpi->allocated_tiles = tile_cols * tile_rows; - for (tile_row = 0; tile_row < tile_rows; ++tile_row) - for (tile_col = 0; tile_col < tile_cols; ++tile_col) { - TileDataEnc *const tile_data = - &cpi->tile_data[tile_row * tile_cols + tile_col]; - int i, j; - for (i = 0; i < BLOCK_SIZES_ALL; ++i) { - for (j = 0; j < MAX_MODES; ++j) { - tile_data->thresh_freq_fact[i][j] = 32; - tile_data->mode_map[i][j] = j; - } - } - } + cpi->allocated_tiles = tile_cols * tile_rows; } void av1_init_tile_data(AV1_COMP *cpi) { AV1_COMMON *const cm = &cpi->common; const int num_planes = av1_num_planes(cm); - const int tile_cols = cm->tile_cols; - const int tile_rows = cm->tile_rows; + const int tile_cols = cm->tiles.cols; + const int tile_rows = cm->tiles.rows; int tile_col, tile_row; TOKENEXTRA *pre_tok = cpi->tile_tok[0][0]; TOKENLIST *tplist = cpi->tplist[0][0]; @@ -4558,9 +5029,10 @@ void av1_init_tile_data(AV1_COMP *cpi) { cpi->tplist[tile_row][tile_col] = tplist + tplist_count; tplist = cpi->tplist[tile_row][tile_col]; tplist_count = av1_get_sb_rows_in_tile(cm, tile_data->tile_info); - tile_data->allow_update_cdf = !cm->large_scale_tile; + tile_data->allow_update_cdf = !cm->tiles.large_scale; tile_data->allow_update_cdf = - tile_data->allow_update_cdf && !cm->disable_cdf_update; + tile_data->allow_update_cdf && !cm->features.disable_cdf_update; + tile_data->tctx = *cm->fc; } } } @@ -4569,24 +5041,22 @@ void av1_encode_sb_row(AV1_COMP *cpi, ThreadData *td, int tile_row, int tile_col, int mi_row) { AV1_COMMON *const cm = &cpi->common; const int num_planes = av1_num_planes(cm); - const int tile_cols = cm->tile_cols; + const int tile_cols = cm->tiles.cols; TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col]; const TileInfo *const tile_info = &this_tile->tile_info; TOKENEXTRA *tok = NULL; - int sb_row_in_tile; - int tile_mb_cols = (tile_info->mi_col_end - tile_info->mi_col_start + 2) >> 2; - - int num_mb_rows_in_sb = - ((1 << (cm->seq_params.mib_size_log2 + MI_SIZE_LOG2)) + 8) >> 4; - - sb_row_in_tile = + const int sb_row_in_tile = (mi_row - tile_info->mi_row_start) >> cm->seq_params.mib_size_log2; + const int tile_mb_cols = + (tile_info->mi_col_end - tile_info->mi_col_start + 2) >> 2; + const int num_mb_rows_in_sb = + ((1 << (cm->seq_params.mib_size_log2 + MI_SIZE_LOG2)) + 8) >> 4; get_start_tok(cpi, tile_row, tile_col, mi_row, &tok, cm->seq_params.mib_size_log2 + MI_SIZE_LOG2, num_planes); cpi->tplist[tile_row][tile_col][sb_row_in_tile].start = tok; - encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok); + encode_sb_row(cpi, td, this_tile, mi_row, &tok); cpi->tplist[tile_row][tile_col][sb_row_in_tile].stop = tok; cpi->tplist[tile_row][tile_col][sb_row_in_tile].count = @@ -4607,42 +5077,30 @@ void av1_encode_tile(AV1_COMP *cpi, ThreadData *td, int tile_row, int tile_col) { AV1_COMMON *const cm = &cpi->common; TileDataEnc *const this_tile = - &cpi->tile_data[tile_row * cm->tile_cols + tile_col]; + &cpi->tile_data[tile_row * cm->tiles.cols + tile_col]; const TileInfo *const tile_info = &this_tile->tile_info; - int mi_row; -#if CONFIG_COLLECT_INTER_MODE_RD_STATS - av1_inter_mode_data_init(this_tile); -#endif + if (!cpi->sf.rt_sf.use_nonrd_pick_mode) av1_inter_mode_data_init(this_tile); av1_zero_above_context(cm, &td->mb.e_mbd, tile_info->mi_col_start, tile_info->mi_col_end, tile_row); - av1_init_above_context(cm, &td->mb.e_mbd, tile_row); + av1_init_above_context(&cm->above_contexts, av1_num_planes(cm), tile_row, + &td->mb.e_mbd); - // Set up pointers to per thread motion search counters. - this_tile->m_search_count = 0; // Count of motion search hits. - this_tile->ex_search_count = 0; // Exhaustive mesh search hits. - td->mb.m_search_count_ptr = &this_tile->m_search_count; - td->mb.ex_search_count_ptr = &this_tile->ex_search_count; - this_tile->tctx = *cm->fc; - td->mb.e_mbd.tile_ctx = &this_tile->tctx; - - cfl_init(&td->mb.e_mbd.cfl, &cm->seq_params); + if (cpi->oxcf.enable_cfl_intra) cfl_init(&td->mb.e_mbd.cfl, &cm->seq_params); av1_crc32c_calculator_init(&td->mb.mb_rd_record.crc_calculator); - td->intrabc_used_this_tile = 0; - - for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end; + for (int mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end; mi_row += cm->seq_params.mib_size) { av1_encode_sb_row(cpi, td, tile_row, tile_col, mi_row); } } -static void encode_tiles(AV1_COMP *cpi) { +static AOM_INLINE void encode_tiles(AV1_COMP *cpi) { AV1_COMMON *const cm = &cpi->common; - const int tile_cols = cm->tile_cols; - const int tile_rows = cm->tile_rows; + const int tile_cols = cm->tiles.cols; + const int tile_rows = cm->tiles.rows; int tile_col, tile_row; if (cpi->tile_data == NULL || cpi->allocated_tiles < tile_cols * tile_rows) @@ -4652,26 +5110,19 @@ static void encode_tiles(AV1_COMP *cpi) { for (tile_row = 0; tile_row < tile_rows; ++tile_row) { for (tile_col = 0; tile_col < tile_cols; ++tile_col) { + TileDataEnc *const this_tile = + &cpi->tile_data[tile_row * cm->tiles.cols + tile_col]; + cpi->td.intrabc_used = 0; + cpi->td.deltaq_used = 0; + cpi->td.mb.e_mbd.tile_ctx = &this_tile->tctx; + cpi->td.mb.tile_pb_ctx = &this_tile->tctx; av1_encode_tile(cpi, &cpi->td, tile_row, tile_col); - cpi->intrabc_used |= cpi->td.intrabc_used_this_tile; + cpi->intrabc_used |= cpi->td.intrabc_used; + cpi->deltaq_used |= cpi->td.deltaq_used; } } } -#if CONFIG_FP_MB_STATS -static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats, - AV1_COMMON *cm, uint8_t **this_frame_mb_stats) { - uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start + - cm->current_video_frame * cm->MBs * sizeof(uint8_t); - - if (mb_stats_in > firstpass_mb_stats->mb_stats_end) return EOF; - - *this_frame_mb_stats = mb_stats_in; - - return 1; -} -#endif - #define GLOBAL_TRANS_TYPES_ENC 3 // highest motion model to search static int gm_get_params_cost(const WarpedMotionParams *gm, const WarpedMotionParams *ref_gm, int allow_hp) { @@ -4722,156 +5173,63 @@ static int gm_get_params_cost(const WarpedMotionParams *gm, return (params_cost << AV1_PROB_COST_SHIFT); } -static int do_gm_search_logic(SPEED_FEATURES *const sf, int num_refs_using_gm, - int frame) { - (void)num_refs_using_gm; +static int do_gm_search_logic(SPEED_FEATURES *const sf, int frame) { (void)frame; - switch (sf->gm_search_type) { + switch (sf->gm_sf.gm_search_type) { case GM_FULL_SEARCH: return 1; - case GM_REDUCED_REF_SEARCH: + case GM_REDUCED_REF_SEARCH_SKIP_L2_L3: return !(frame == LAST2_FRAME || frame == LAST3_FRAME); + case GM_REDUCED_REF_SEARCH_SKIP_L2_L3_ARF2: + return !(frame == LAST2_FRAME || frame == LAST3_FRAME || + (frame == ALTREF2_FRAME)); case GM_DISABLE_SEARCH: return 0; default: assert(0); } return 1; } -// Estimate if the source frame is screen content, based on the portion of -// blocks that have no more than 4 (experimentally selected) luma colors. -static int is_screen_content(const uint8_t *src, int use_hbd, int bd, - int stride, int width, int height) { - assert(src != NULL); - int counts = 0; - const int blk_w = 16; - const int blk_h = 16; - const int limit = 4; - for (int r = 0; r + blk_h <= height; r += blk_h) { - for (int c = 0; c + blk_w <= width; c += blk_w) { - int count_buf[1 << 12]; // Maximum (1 << 12) color levels. - const int n_colors = - use_hbd ? av1_count_colors_highbd(src + r * stride + c, stride, blk_w, - blk_h, bd, count_buf) - : av1_count_colors(src + r * stride + c, stride, blk_w, blk_h, - count_buf); - if (n_colors > 1 && n_colors <= limit) counts++; - } - } - // The threshold is 10%. - return counts * blk_h * blk_w * 10 > width * height; -} - -static const uint8_t ref_frame_flag_list[REF_FRAMES] = { 0, - AOM_LAST_FLAG, - AOM_LAST2_FLAG, - AOM_LAST3_FLAG, - AOM_GOLD_FLAG, - AOM_BWD_FLAG, - AOM_ALT2_FLAG, - AOM_ALT_FLAG }; - -// Enforce the number of references for each arbitrary frame limited to -// (INTER_REFS_PER_FRAME - 1) -static void enforce_max_ref_frames(AV1_COMP *cpi) { - AV1_COMMON *const cm = &cpi->common; +// Set the relative distance of a reference frame w.r.t. current frame +static AOM_INLINE void set_rel_frame_dist(AV1_COMP *cpi) { + const AV1_COMMON *const cm = &cpi->common; + const OrderHintInfo *const order_hint_info = &cm->seq_params.order_hint_info; MV_REFERENCE_FRAME ref_frame; - int total_valid_refs = 0; + int min_past_dist = INT32_MAX, min_future_dist = INT32_MAX; + cpi->nearest_past_ref = NONE_FRAME; + cpi->nearest_future_ref = NONE_FRAME; for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) { - if (cpi->ref_frame_flags & ref_frame_flag_list[ref_frame]) - total_valid_refs++; - } - - // NOTE(zoeliu): When all the possible reference frames are availble, we - // reduce the number of reference frames by 1, following the rules of: - // (1) Retain GOLDEN_FARME/ALTEF_FRAME; - // (2) Check the earliest 2 remaining reference frames, and remove the one - // with the lower quality factor, otherwise if both have been coded at - // the same quality level, remove the earliest reference frame. - - if (total_valid_refs == INTER_REFS_PER_FRAME) { - unsigned int min_ref_offset = UINT_MAX; - unsigned int second_min_ref_offset = UINT_MAX; - MV_REFERENCE_FRAME earliest_ref_frames[2] = { LAST3_FRAME, LAST2_FRAME }; - int earliest_buf_idxes[2] = { 0 }; - - // Locate the earliest two reference frames except GOLDEN/ALTREF. - for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) { - // Retain GOLDEN/ALTERF - if (ref_frame == GOLDEN_FRAME || ref_frame == ALTREF_FRAME) continue; - - const int buf_idx = cm->frame_refs[ref_frame - LAST_FRAME].idx; - if (buf_idx >= 0) { - const unsigned int ref_offset = - cm->buffer_pool->frame_bufs[buf_idx].cur_frame_offset; - - if (min_ref_offset == UINT_MAX) { - min_ref_offset = ref_offset; - earliest_ref_frames[0] = ref_frame; - earliest_buf_idxes[0] = buf_idx; - } else { - if (get_relative_dist(cm, ref_offset, min_ref_offset) < 0) { - second_min_ref_offset = min_ref_offset; - earliest_ref_frames[1] = earliest_ref_frames[0]; - earliest_buf_idxes[1] = earliest_buf_idxes[0]; - - min_ref_offset = ref_offset; - earliest_ref_frames[0] = ref_frame; - earliest_buf_idxes[0] = buf_idx; - } else if (second_min_ref_offset == UINT_MAX || - get_relative_dist(cm, ref_offset, second_min_ref_offset) < - 0) { - second_min_ref_offset = ref_offset; - earliest_ref_frames[1] = ref_frame; - earliest_buf_idxes[1] = buf_idx; - } - } + cpi->ref_relative_dist[ref_frame - LAST_FRAME] = 0; + if (cpi->ref_frame_flags & av1_ref_frame_flag_list[ref_frame]) { + int dist = av1_encoder_get_relative_dist( + order_hint_info, + cm->cur_frame->ref_display_order_hint[ref_frame - LAST_FRAME], + cm->current_frame.display_order_hint); + cpi->ref_relative_dist[ref_frame - LAST_FRAME] = dist; + // Get the nearest ref_frame in the past + if (abs(dist) < min_past_dist && dist < 0) { + cpi->nearest_past_ref = ref_frame; + min_past_dist = abs(dist); + } + // Get the nearest ref_frame in the future + if (dist < min_future_dist && dist > 0) { + cpi->nearest_future_ref = ref_frame; + min_future_dist = dist; } - } - // Check the coding quality factors of the two earliest reference frames. - RATE_FACTOR_LEVEL ref_rf_level[2]; - double ref_rf_deltas[2]; - for (int i = 0; i < 2; ++i) { - ref_rf_level[i] = cpi->frame_rf_level[earliest_buf_idxes[i]]; - ref_rf_deltas[i] = rate_factor_deltas[ref_rf_level[i]]; - } - (void)ref_rf_level; - (void)ref_rf_deltas; - -#define USE_RF_LEVEL_TO_ENFORCE 1 -#if USE_RF_LEVEL_TO_ENFORCE - // If both earliest two reference frames are coded using the same rate- - // factor, disable the earliest reference frame; Otherwise disable the - // reference frame that uses a lower rate-factor delta. - const MV_REFERENCE_FRAME ref_frame_to_disable = - (ref_rf_deltas[0] <= ref_rf_deltas[1]) ? earliest_ref_frames[0] - : earliest_ref_frames[1]; -#else - // Always disable the earliest reference frame - const MV_REFERENCE_FRAME ref_frame_to_disable = earliest_ref_frames[0]; -#endif // USE_RF_LEVEL_TO_ENFORCE -#undef USE_RF_LEVEL_TO_ENFORCE - - switch (ref_frame_to_disable) { - case LAST_FRAME: cpi->ref_frame_flags &= ~AOM_LAST_FLAG; break; - case LAST2_FRAME: cpi->ref_frame_flags &= ~AOM_LAST2_FLAG; break; - case LAST3_FRAME: cpi->ref_frame_flags &= ~AOM_LAST3_FLAG; break; - case BWDREF_FRAME: cpi->ref_frame_flags &= ~AOM_BWD_FLAG; break; - case ALTREF2_FRAME: cpi->ref_frame_flags &= ~AOM_ALT2_FLAG; break; - default: break; } } } -static INLINE int av1_refs_are_one_sided(const AV1_COMMON *cm) { +static INLINE int refs_are_one_sided(const AV1_COMMON *cm) { assert(!frame_is_intra_only(cm)); int one_sided_refs = 1; - for (int ref = 0; ref < INTER_REFS_PER_FRAME; ++ref) { - const int buf_idx = cm->frame_refs[ref].idx; - if (buf_idx == INVALID_IDX) continue; - - const int ref_offset = - cm->buffer_pool->frame_bufs[buf_idx].cur_frame_offset; - if (get_relative_dist(cm, ref_offset, (int)cm->frame_offset) > 0) { + for (int ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref) { + const RefCntBuffer *const buf = get_ref_frame_buf(cm, ref); + if (buf == NULL) continue; + + const int ref_display_order_hint = buf->display_order_hint; + if (av1_encoder_get_relative_dist( + &cm->seq_params.order_hint_info, ref_display_order_hint, + (int)cm->current_frame.display_order_hint) > 0) { one_sided_refs = 0; // bwd reference break; } @@ -4880,31 +5238,36 @@ static INLINE int av1_refs_are_one_sided(const AV1_COMMON *cm) { } static INLINE void get_skip_mode_ref_offsets(const AV1_COMMON *cm, - int ref_offset[2]) { - ref_offset[0] = ref_offset[1] = 0; - if (!cm->is_skip_mode_allowed) return; - - const int buf_idx_0 = cm->frame_refs[cm->ref_frame_idx_0].idx; - const int buf_idx_1 = cm->frame_refs[cm->ref_frame_idx_1].idx; - assert(buf_idx_0 != INVALID_IDX && buf_idx_1 != INVALID_IDX); - - ref_offset[0] = cm->buffer_pool->frame_bufs[buf_idx_0].cur_frame_offset; - ref_offset[1] = cm->buffer_pool->frame_bufs[buf_idx_1].cur_frame_offset; + int ref_order_hint[2]) { + const SkipModeInfo *const skip_mode_info = &cm->current_frame.skip_mode_info; + ref_order_hint[0] = ref_order_hint[1] = 0; + if (!skip_mode_info->skip_mode_allowed) return; + + const RefCntBuffer *const buf_0 = + get_ref_frame_buf(cm, LAST_FRAME + skip_mode_info->ref_frame_idx_0); + const RefCntBuffer *const buf_1 = + get_ref_frame_buf(cm, LAST_FRAME + skip_mode_info->ref_frame_idx_1); + assert(buf_0 != NULL && buf_1 != NULL); + + ref_order_hint[0] = buf_0->order_hint; + ref_order_hint[1] = buf_1->order_hint; } static int check_skip_mode_enabled(AV1_COMP *const cpi) { AV1_COMMON *const cm = &cpi->common; av1_setup_skip_mode_allowed(cm); - if (!cm->is_skip_mode_allowed) return 0; + if (!cm->current_frame.skip_mode_info.skip_mode_allowed) return 0; // Turn off skip mode if the temporal distances of the reference pair to the // current frame are different by more than 1 frame. - const int cur_offset = (int)cm->frame_offset; + const int cur_offset = (int)cm->current_frame.order_hint; int ref_offset[2]; get_skip_mode_ref_offsets(cm, ref_offset); - const int cur_to_ref0 = get_relative_dist(cm, cur_offset, ref_offset[0]); - const int cur_to_ref1 = abs(get_relative_dist(cm, cur_offset, ref_offset[1])); + const int cur_to_ref0 = get_relative_dist(&cm->seq_params.order_hint_info, + cur_offset, ref_offset[0]); + const int cur_to_ref1 = abs(get_relative_dist(&cm->seq_params.order_hint_info, + cur_offset, ref_offset[1])); if (abs(cur_to_ref0 - cur_to_ref1) > 1) return 0; // High Latency: Turn off skip mode if all refs are fwd. @@ -4918,8 +5281,10 @@ static int check_skip_mode_enabled(AV1_COMP *const cpi) { AOM_BWD_FLAG, AOM_ALT2_FLAG, AOM_ALT_FLAG }; - const int ref_frame[2] = { cm->ref_frame_idx_0 + LAST_FRAME, - cm->ref_frame_idx_1 + LAST_FRAME }; + const int ref_frame[2] = { + cm->current_frame.skip_mode_info.ref_frame_idx_0 + LAST_FRAME, + cm->current_frame.skip_mode_info.ref_frame_idx_1 + LAST_FRAME + }; if (!(cpi->ref_frame_flags & flag_list[ref_frame[0]]) || !(cpi->ref_frame_flags & flag_list[ref_frame[1]])) return 0; @@ -4933,65 +5298,380 @@ static INLINE int skip_gm_frame(AV1_COMMON *const cm, int ref_frame) { if ((ref_frame == LAST3_FRAME || ref_frame == LAST2_FRAME) && cm->global_motion[GOLDEN_FRAME].wmtype != IDENTITY) { return get_relative_dist( - cm, cm->cur_frame->ref_frame_offset[ref_frame - LAST_FRAME], - cm->cur_frame->ref_frame_offset[GOLDEN_FRAME - LAST_FRAME]) <= 0; + &cm->seq_params.order_hint_info, + cm->cur_frame->ref_order_hints[ref_frame - LAST_FRAME], + cm->cur_frame->ref_order_hints[GOLDEN_FRAME - LAST_FRAME]) <= 0; } return 0; } -static void set_default_interp_skip_flags(AV1_COMP *cpi) { - const int num_planes = av1_num_planes(&cpi->common); - cpi->default_interp_skip_flags = (num_planes == 1) - ? DEFAULT_LUMA_INTERP_SKIP_FLAG - : DEFAULT_INTERP_SKIP_FLAG; +static AOM_INLINE void set_default_interp_skip_flags( + const AV1_COMMON *cm, InterpSearchFlags *interp_search_flags) { + const int num_planes = av1_num_planes(cm); + interp_search_flags->default_interp_skip_flags = + (num_planes == 1) ? INTERP_SKIP_LUMA_EVAL_CHROMA + : INTERP_SKIP_LUMA_SKIP_CHROMA; +} + +// TODO(Remya): Can include erroradv_prod_tr[] for threshold calculation +static INLINE int64_t calc_erroradv_threshold(AV1_COMP *cpi, + int64_t ref_frame_error) { + if (!cpi->sf.gm_sf.disable_adaptive_warp_error_thresh) + return (int64_t)( + ref_frame_error * erroradv_tr[cpi->sf.gm_sf.gm_erroradv_type] + 0.5); + else + return INT64_MAX; } -static void encode_frame_internal(AV1_COMP *cpi) { +static void compute_global_motion_for_ref_frame( + AV1_COMP *cpi, YV12_BUFFER_CONFIG *ref_buf[REF_FRAMES], int frame, + int *num_frm_corners, int *frm_corners, unsigned char *frm_buffer, + MotionModel *params_by_motion, uint8_t *segment_map, + const int segment_map_w, const int segment_map_h, + const WarpedMotionParams *ref_params) { ThreadData *const td = &cpi->td; MACROBLOCK *const x = &td->mb; AV1_COMMON *const cm = &cpi->common; MACROBLOCKD *const xd = &x->e_mbd; - RD_COUNTS *const rdc = &cpi->td.rd_counts; int i; + // clang-format off + static const double kIdentityParams[MAX_PARAMDIM - 1] = { + 0.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0 + }; + // clang-format on + WarpedMotionParams tmp_wm_params; + const double *params_this_motion; + int inliers_by_motion[RANSAC_NUM_MOTIONS]; + assert(ref_buf[frame] != NULL); + if (*num_frm_corners < 0) { + // compute interest points using FAST features + *num_frm_corners = av1_fast_corner_detect( + frm_buffer, cpi->source->y_width, cpi->source->y_height, + cpi->source->y_stride, frm_corners, MAX_CORNERS); + } + TransformationType model; - x->min_partition_size = AOMMIN(x->min_partition_size, cm->seq_params.sb_size); - x->max_partition_size = AOMMIN(x->max_partition_size, cm->seq_params.sb_size); -#if CONFIG_DIST_8X8 - x->using_dist_8x8 = cpi->oxcf.using_dist_8x8; - x->tune_metric = cpi->oxcf.tuning; -#endif - cm->setup_mi(cm); + aom_clear_system_state(); - xd->mi = cm->mi_grid_visible; - xd->mi[0] = cm->mi; + // TODO(sarahparker, debargha): Explore do_adaptive_gm_estimation = 1 + const int do_adaptive_gm_estimation = 0; + + const int ref_frame_dist = get_relative_dist( + &cm->seq_params.order_hint_info, cm->current_frame.order_hint, + cm->cur_frame->ref_order_hints[frame - LAST_FRAME]); + const GlobalMotionEstimationType gm_estimation_type = + cm->seq_params.order_hint_info.enable_order_hint && + abs(ref_frame_dist) <= 2 && do_adaptive_gm_estimation + ? GLOBAL_MOTION_DISFLOW_BASED + : GLOBAL_MOTION_FEATURE_BASED; + for (model = ROTZOOM; model < GLOBAL_TRANS_TYPES_ENC; ++model) { + int64_t best_warp_error = INT64_MAX; + // Initially set all params to identity. + for (i = 0; i < RANSAC_NUM_MOTIONS; ++i) { + memcpy(params_by_motion[i].params, kIdentityParams, + (MAX_PARAMDIM - 1) * sizeof(*(params_by_motion[i].params))); + params_by_motion[i].num_inliers = 0; + } + + av1_compute_global_motion( + model, frm_buffer, cpi->source->y_width, cpi->source->y_height, + cpi->source->y_stride, frm_corners, *num_frm_corners, ref_buf[frame], + cpi->common.seq_params.bit_depth, gm_estimation_type, inliers_by_motion, + params_by_motion, RANSAC_NUM_MOTIONS); + int64_t ref_frame_error = 0; + for (i = 0; i < RANSAC_NUM_MOTIONS; ++i) { + if (inliers_by_motion[i] == 0) continue; + + params_this_motion = params_by_motion[i].params; + av1_convert_model_to_params(params_this_motion, &tmp_wm_params); + + if (tmp_wm_params.wmtype != IDENTITY) { + av1_compute_feature_segmentation_map( + segment_map, segment_map_w, segment_map_h, + params_by_motion[i].inliers, params_by_motion[i].num_inliers); + + ref_frame_error = av1_segmented_frame_error( + is_cur_buf_hbd(xd), xd->bd, ref_buf[frame]->y_buffer, + ref_buf[frame]->y_stride, cpi->source->y_buffer, + cpi->source->y_width, cpi->source->y_height, cpi->source->y_stride, + segment_map, segment_map_w); + + int64_t erroradv_threshold = + calc_erroradv_threshold(cpi, ref_frame_error); + + const int64_t warp_error = av1_refine_integerized_param( + &tmp_wm_params, tmp_wm_params.wmtype, is_cur_buf_hbd(xd), xd->bd, + ref_buf[frame]->y_buffer, ref_buf[frame]->y_width, + ref_buf[frame]->y_height, ref_buf[frame]->y_stride, + cpi->source->y_buffer, cpi->source->y_width, cpi->source->y_height, + cpi->source->y_stride, GM_REFINEMENT_COUNT, best_warp_error, + segment_map, segment_map_w, erroradv_threshold); + + if (warp_error < best_warp_error) { + best_warp_error = warp_error; + // Save the wm_params modified by + // av1_refine_integerized_param() rather than motion index to + // avoid rerunning refine() below. + memcpy(&(cm->global_motion[frame]), &tmp_wm_params, + sizeof(WarpedMotionParams)); + } + } + } + if (cm->global_motion[frame].wmtype <= AFFINE) + if (!av1_get_shear_params(&cm->global_motion[frame])) + cm->global_motion[frame] = default_warp_params; - av1_zero(*td->counts); - av1_zero(rdc->comp_pred_diff); + if (cm->global_motion[frame].wmtype == TRANSLATION) { + cm->global_motion[frame].wmmat[0] = + convert_to_trans_prec(cm->features.allow_high_precision_mv, + cm->global_motion[frame].wmmat[0]) * + GM_TRANS_ONLY_DECODE_FACTOR; + cm->global_motion[frame].wmmat[1] = + convert_to_trans_prec(cm->features.allow_high_precision_mv, + cm->global_motion[frame].wmmat[1]) * + GM_TRANS_ONLY_DECODE_FACTOR; + } - if (frame_is_intra_only(cm)) { - if (cm->seq_params.force_screen_content_tools == 2) { - cm->allow_screen_content_tools = - cpi->oxcf.content == AOM_CONTENT_SCREEN || - is_screen_content(cpi->source->y_buffer, - cpi->source->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd, - cpi->source->y_stride, cpi->source->y_width, - cpi->source->y_height); + if (cm->global_motion[frame].wmtype == IDENTITY) continue; + + if (ref_frame_error == 0) continue; + + // If the best error advantage found doesn't meet the threshold for + // this motion type, revert to IDENTITY. + if (!av1_is_enough_erroradvantage( + (double)best_warp_error / ref_frame_error, + gm_get_params_cost(&cm->global_motion[frame], ref_params, + cm->features.allow_high_precision_mv), + cpi->sf.gm_sf.gm_erroradv_type)) { + cm->global_motion[frame] = default_warp_params; + } + + if (cm->global_motion[frame].wmtype != IDENTITY) break; + } + + aom_clear_system_state(); +} + +typedef struct { + int distance; + MV_REFERENCE_FRAME frame; +} FrameDistPair; + +static INLINE void update_valid_ref_frames_for_gm( + AV1_COMP *cpi, YV12_BUFFER_CONFIG *ref_buf[REF_FRAMES], + FrameDistPair *past_ref_frame, FrameDistPair *future_ref_frame, + int *num_past_ref_frames, int *num_future_ref_frames) { + AV1_COMMON *const cm = &cpi->common; + const OrderHintInfo *const order_hint_info = &cm->seq_params.order_hint_info; + for (int frame = ALTREF_FRAME; frame >= LAST_FRAME; --frame) { + const MV_REFERENCE_FRAME ref_frame[2] = { frame, NONE_FRAME }; + RefCntBuffer *buf = get_ref_frame_buf(cm, frame); + const int ref_disabled = + !(cpi->ref_frame_flags & av1_ref_frame_flag_list[frame]); + ref_buf[frame] = NULL; + cm->global_motion[frame] = default_warp_params; + // Skip global motion estimation for invalid ref frames + if (buf == NULL || + (ref_disabled && cpi->sf.hl_sf.recode_loop != DISALLOW_RECODE)) { + cpi->gm_info.params_cost[frame] = 0; + continue; } else { - cm->allow_screen_content_tools = - cm->seq_params.force_screen_content_tools; + ref_buf[frame] = &buf->buf; + } + + if (ref_buf[frame]->y_crop_width == cpi->source->y_crop_width && + ref_buf[frame]->y_crop_height == cpi->source->y_crop_height && + do_gm_search_logic(&cpi->sf, frame) && + !prune_ref_by_selective_ref_frame( + cpi, NULL, ref_frame, cm->cur_frame->ref_display_order_hint) && + !(cpi->sf.gm_sf.selective_ref_gm && skip_gm_frame(cm, frame))) { + assert(ref_buf[frame] != NULL); + int relative_frame_dist = av1_encoder_get_relative_dist( + order_hint_info, buf->display_order_hint, + cm->cur_frame->display_order_hint); + // Populate past and future ref frames + if (relative_frame_dist <= 0) { + past_ref_frame[*num_past_ref_frames].distance = + abs(relative_frame_dist); + past_ref_frame[*num_past_ref_frames].frame = frame; + (*num_past_ref_frames)++; + } else { + future_ref_frame[*num_future_ref_frames].distance = + abs(relative_frame_dist); + future_ref_frame[*num_future_ref_frames].frame = frame; + (*num_future_ref_frames)++; + } + } + } +} + +static INLINE void compute_gm_for_valid_ref_frames( + AV1_COMP *cpi, YV12_BUFFER_CONFIG *ref_buf[REF_FRAMES], int frame, + int *num_frm_corners, int *frm_corners, unsigned char *frm_buffer, + MotionModel *params_by_motion, uint8_t *segment_map, + const int segment_map_w, const int segment_map_h) { + AV1_COMMON *const cm = &cpi->common; + GlobalMotionInfo *const gm_info = &cpi->gm_info; + const WarpedMotionParams *ref_params = + cm->prev_frame ? &cm->prev_frame->global_motion[frame] + : &default_warp_params; + + compute_global_motion_for_ref_frame( + cpi, ref_buf, frame, num_frm_corners, frm_corners, frm_buffer, + params_by_motion, segment_map, segment_map_w, segment_map_h, ref_params); + + gm_info->params_cost[frame] = + gm_get_params_cost(&cm->global_motion[frame], ref_params, + cm->features.allow_high_precision_mv) + + gm_info->type_cost[cm->global_motion[frame].wmtype] - + gm_info->type_cost[IDENTITY]; +} + +static int compare_distance(const void *a, const void *b) { + const int diff = + ((FrameDistPair *)a)->distance - ((FrameDistPair *)b)->distance; + if (diff > 0) + return 1; + else if (diff < 0) + return -1; + return 0; +} + +static INLINE void compute_global_motion_for_references( + AV1_COMP *cpi, YV12_BUFFER_CONFIG *ref_buf[REF_FRAMES], + FrameDistPair reference_frame[REF_FRAMES - 1], int num_ref_frames, + int *num_frm_corners, int *frm_corners, unsigned char *frm_buffer, + MotionModel *params_by_motion, uint8_t *segment_map, + const int segment_map_w, const int segment_map_h) { + AV1_COMMON *const cm = &cpi->common; + // Compute global motion w.r.t. reference frames starting from the nearest ref + // frame in a given direction + for (int frame = 0; frame < num_ref_frames; frame++) { + int ref_frame = reference_frame[frame].frame; + compute_gm_for_valid_ref_frames(cpi, ref_buf, ref_frame, num_frm_corners, + frm_corners, frm_buffer, params_by_motion, + segment_map, segment_map_w, segment_map_h); + // If global motion w.r.t. current ref frame is + // INVALID/TRANSLATION/IDENTITY, skip the evaluation of global motion w.r.t + // the remaining ref frames in that direction. The below exit is disabled + // when ref frame distance w.r.t. current frame is zero. E.g.: + // source_alt_ref_frame w.r.t. ARF frames + if (cpi->sf.gm_sf.prune_ref_frame_for_gm_search && + reference_frame[frame].distance != 0 && + cm->global_motion[ref_frame].wmtype != ROTZOOM) + break; + } +} + +static AOM_INLINE void setup_prune_ref_frame_mask(AV1_COMP *cpi) { + if (!cpi->sf.rt_sf.use_nonrd_pick_mode && + cpi->sf.inter_sf.selective_ref_frame >= 2) { + AV1_COMMON *const cm = &cpi->common; + const OrderHintInfo *const order_hint_info = + &cm->seq_params.order_hint_info; + const int cur_frame_display_order_hint = + cm->current_frame.display_order_hint; + unsigned int *ref_display_order_hint = + cm->cur_frame->ref_display_order_hint; + const int arf2_dist = av1_encoder_get_relative_dist( + order_hint_info, ref_display_order_hint[ALTREF2_FRAME - LAST_FRAME], + cur_frame_display_order_hint); + const int bwd_dist = av1_encoder_get_relative_dist( + order_hint_info, ref_display_order_hint[BWDREF_FRAME - LAST_FRAME], + cur_frame_display_order_hint); + + for (int ref_idx = REF_FRAMES; ref_idx < MODE_CTX_REF_FRAMES; ++ref_idx) { + MV_REFERENCE_FRAME rf[2]; + av1_set_ref_frame(rf, ref_idx); + if (!(cpi->ref_frame_flags & av1_ref_frame_flag_list[rf[0]]) || + !(cpi->ref_frame_flags & av1_ref_frame_flag_list[rf[1]])) { + continue; + } + + if (!cpi->all_one_sided_refs) { + int ref_dist[2]; + for (int i = 0; i < 2; ++i) { + ref_dist[i] = av1_encoder_get_relative_dist( + order_hint_info, ref_display_order_hint[rf[i] - LAST_FRAME], + cur_frame_display_order_hint); + } + + // One-sided compound is used only when all reference frames are + // one-sided. + if ((ref_dist[0] > 0) == (ref_dist[1] > 0)) { + cpi->prune_ref_frame_mask |= 1 << ref_idx; + } + } + + if (cpi->sf.inter_sf.selective_ref_frame >= 4 && + (rf[0] == ALTREF2_FRAME || rf[1] == ALTREF2_FRAME) && + (cpi->ref_frame_flags & av1_ref_frame_flag_list[BWDREF_FRAME])) { + // Check if both ALTREF2_FRAME and BWDREF_FRAME are future references. + if (arf2_dist > 0 && bwd_dist > 0 && bwd_dist <= arf2_dist) { + // Drop ALTREF2_FRAME as a reference if BWDREF_FRAME is a closer + // reference to the current frame than ALTREF2_FRAME + cpi->prune_ref_frame_mask |= 1 << ref_idx; + } + } } } +} + +#define CHECK_PRECOMPUTED_REF_FRAME_MAP 0 + +static AOM_INLINE void encode_frame_internal(AV1_COMP *cpi) { + ThreadData *const td = &cpi->td; + MACROBLOCK *const x = &td->mb; + AV1_COMMON *const cm = &cpi->common; + CommonModeInfoParams *const mi_params = &cm->mi_params; + FeatureFlags *const features = &cm->features; + MACROBLOCKD *const xd = &x->e_mbd; + RD_COUNTS *const rdc = &cpi->td.rd_counts; + GlobalMotionInfo *const gm_info = &cpi->gm_info; + FrameProbInfo *const frame_probs = &cpi->frame_probs; + IntraBCHashInfo *const intrabc_hash_info = &x->intrabc_hash_info; + int i; + + if (!cpi->sf.rt_sf.use_nonrd_pick_mode) { + mi_params->setup_mi(mi_params); + } + + set_mi_offsets(mi_params, xd, 0, 0); + +#if CONFIG_AV1_HIGHBITDEPTH + x->fwd_txfm4x4 = aom_fdct4x4; +#else + x->fwd_txfm4x4 = aom_fdct4x4_lp; +#endif + + av1_zero(*td->counts); + av1_zero(rdc->comp_pred_diff); + av1_zero(rdc->tx_type_used); + av1_zero(rdc->obmc_used); + av1_zero(rdc->warped_used); - // Allow intrabc when screen content tools are enabled. - cm->allow_intrabc = cm->allow_screen_content_tools; // Reset the flag. cpi->intrabc_used = 0; // Need to disable intrabc when superres is selected if (av1_superres_scaled(cm)) { - cm->allow_intrabc = 0; + features->allow_intrabc = 0; } - if (cpi->oxcf.pass != 1 && av1_use_hash_me(cm)) { + features->allow_intrabc &= (cpi->oxcf.enable_intrabc); + + if (features->allow_warped_motion && + cpi->sf.inter_sf.prune_warped_prob_thresh > 0) { + const FRAME_UPDATE_TYPE update_type = get_frame_update_type(&cpi->gf_group); + if (frame_probs->warped_probs[update_type] < + cpi->sf.inter_sf.prune_warped_prob_thresh) + features->allow_warped_motion = 0; + } + + int hash_table_created = 0; + if (!is_stat_generation_stage(cpi) && av1_use_hash_me(cpi) && + !cpi->sf.rt_sf.use_nonrd_pick_mode) { + // TODO(any): move this outside of the recoding loop to avoid recalculating + // the hash table. // add to hash table const int pic_width = cpi->source->y_crop_width; const int pic_height = cpi->source->y_crop_height; @@ -5011,46 +5691,28 @@ static void encode_frame_internal(AV1_COMP *cpi) { } } - av1_hash_table_create(&cm->cur_frame->hash_table); - av1_generate_block_2x2_hash_value(cpi->source, block_hash_values[0], - is_block_same[0], &cpi->td.mb); - av1_generate_block_hash_value(cpi->source, 4, block_hash_values[0], - block_hash_values[1], is_block_same[0], - is_block_same[1], &cpi->td.mb); - av1_add_to_hash_map_by_row_with_precal_data( - &cm->cur_frame->hash_table, block_hash_values[1], is_block_same[1][2], - pic_width, pic_height, 4); - av1_generate_block_hash_value(cpi->source, 8, block_hash_values[1], - block_hash_values[0], is_block_same[1], - is_block_same[0], &cpi->td.mb); - av1_add_to_hash_map_by_row_with_precal_data( - &cm->cur_frame->hash_table, block_hash_values[0], is_block_same[0][2], - pic_width, pic_height, 8); - av1_generate_block_hash_value(cpi->source, 16, block_hash_values[0], - block_hash_values[1], is_block_same[0], - is_block_same[1], &cpi->td.mb); - av1_add_to_hash_map_by_row_with_precal_data( - &cm->cur_frame->hash_table, block_hash_values[1], is_block_same[1][2], - pic_width, pic_height, 16); - av1_generate_block_hash_value(cpi->source, 32, block_hash_values[1], - block_hash_values[0], is_block_same[1], - is_block_same[0], &cpi->td.mb); - av1_add_to_hash_map_by_row_with_precal_data( - &cm->cur_frame->hash_table, block_hash_values[0], is_block_same[0][2], - pic_width, pic_height, 32); - av1_generate_block_hash_value(cpi->source, 64, block_hash_values[0], - block_hash_values[1], is_block_same[0], - is_block_same[1], &cpi->td.mb); - av1_add_to_hash_map_by_row_with_precal_data( - &cm->cur_frame->hash_table, block_hash_values[1], is_block_same[1][2], - pic_width, pic_height, 64); - - av1_generate_block_hash_value(cpi->source, 128, block_hash_values[1], - block_hash_values[0], is_block_same[1], - is_block_same[0], &cpi->td.mb); - av1_add_to_hash_map_by_row_with_precal_data( - &cm->cur_frame->hash_table, block_hash_values[0], is_block_same[0][2], - pic_width, pic_height, 128); + av1_hash_table_init(intrabc_hash_info); + av1_hash_table_create(&intrabc_hash_info->intrabc_hash_table); + hash_table_created = 1; + av1_generate_block_2x2_hash_value(intrabc_hash_info, cpi->source, + block_hash_values[0], is_block_same[0]); + // Hash data generated for screen contents is used for intraBC ME + const int min_alloc_size = block_size_wide[mi_params->mi_alloc_bsize]; + const int max_sb_size = + (1 << (cm->seq_params.mib_size_log2 + MI_SIZE_LOG2)); + int src_idx = 0; + for (int size = 4; size <= max_sb_size; size *= 2, src_idx = !src_idx) { + const int dst_idx = !src_idx; + av1_generate_block_hash_value( + intrabc_hash_info, cpi->source, size, block_hash_values[src_idx], + block_hash_values[dst_idx], is_block_same[src_idx], + is_block_same[dst_idx]); + if (size >= min_alloc_size) { + av1_add_to_hash_map_by_row_with_precal_data( + &intrabc_hash_info->intrabc_hash_table, block_hash_values[dst_idx], + is_block_same[dst_idx][2], pic_width, pic_height, size); + } + } for (k = 0; k < 2; k++) { for (j = 0; j < 2; j++) { @@ -5063,49 +5725,69 @@ static void encode_frame_internal(AV1_COMP *cpi) { } } + const CommonQuantParams *quant_params = &cm->quant_params; for (i = 0; i < MAX_SEGMENTS; ++i) { - const int qindex = cm->seg.enabled - ? av1_get_qindex(&cm->seg, i, cm->base_qindex) - : cm->base_qindex; - xd->lossless[i] = qindex == 0 && cm->y_dc_delta_q == 0 && - cm->u_dc_delta_q == 0 && cm->u_ac_delta_q == 0 && - cm->v_dc_delta_q == 0 && cm->v_ac_delta_q == 0; - if (xd->lossless[i]) cpi->has_lossless_segment = 1; + const int qindex = + cm->seg.enabled ? av1_get_qindex(&cm->seg, i, quant_params->base_qindex) + : quant_params->base_qindex; + xd->lossless[i] = + qindex == 0 && quant_params->y_dc_delta_q == 0 && + quant_params->u_dc_delta_q == 0 && quant_params->u_ac_delta_q == 0 && + quant_params->v_dc_delta_q == 0 && quant_params->v_ac_delta_q == 0; + if (xd->lossless[i]) cpi->enc_seg.has_lossless_segment = 1; xd->qindex[i] = qindex; if (xd->lossless[i]) { - cpi->optimize_seg_arr[i] = 0; + cpi->optimize_seg_arr[i] = NO_TRELLIS_OPT; } else { - cpi->optimize_seg_arr[i] = cpi->optimize_speed_feature; + cpi->optimize_seg_arr[i] = cpi->sf.rd_sf.optimize_coefficients; } } - cm->coded_lossless = is_coded_lossless(cm, xd); - cm->all_lossless = cm->coded_lossless && !av1_superres_scaled(cm); - - cm->tx_mode = select_tx_mode(cpi); + features->coded_lossless = is_coded_lossless(cm, xd); + features->all_lossless = features->coded_lossless && !av1_superres_scaled(cm); // Fix delta q resolution for the moment - cm->delta_q_res = DEFAULT_DELTA_Q_RES; + cm->delta_q_info.delta_q_res = 0; + if (cpi->oxcf.deltaq_mode == DELTA_Q_OBJECTIVE) + cm->delta_q_info.delta_q_res = DEFAULT_DELTA_Q_RES_OBJECTIVE; + else if (cpi->oxcf.deltaq_mode == DELTA_Q_PERCEPTUAL) + cm->delta_q_info.delta_q_res = DEFAULT_DELTA_Q_RES_PERCEPTUAL; // Set delta_q_present_flag before it is used for the first time - cm->delta_lf_res = DEFAULT_DELTA_LF_RES; - cm->delta_q_present_flag = cpi->oxcf.deltaq_mode != NO_DELTA_Q; - cm->delta_lf_present_flag = cpi->oxcf.deltaq_mode == DELTA_Q_LF; - cm->delta_lf_multi = DEFAULT_DELTA_LF_MULTI; - // update delta_q_present_flag and delta_lf_present_flag based on base_qindex - cm->delta_q_present_flag &= cm->base_qindex > 0; - cm->delta_lf_present_flag &= cm->base_qindex > 0; + cm->delta_q_info.delta_lf_res = DEFAULT_DELTA_LF_RES; + cm->delta_q_info.delta_q_present_flag = cpi->oxcf.deltaq_mode != NO_DELTA_Q; - av1_frame_init_quantizer(cpi); + // Turn off cm->delta_q_info.delta_q_present_flag if objective delta_q is used + // for ineligible frames. That effectively will turn off row_mt usage. + // Note objective delta_q and tpl eligible frames are only altref frames + // currently. + if (cm->delta_q_info.delta_q_present_flag) { + if (cpi->oxcf.deltaq_mode == DELTA_Q_OBJECTIVE && + !is_frame_tpl_eligible(cpi)) + cm->delta_q_info.delta_q_present_flag = 0; + } + + // Reset delta_q_used flag + cpi->deltaq_used = 0; + + cm->delta_q_info.delta_lf_present_flag = + cm->delta_q_info.delta_q_present_flag && cpi->oxcf.deltalf_mode; + cm->delta_q_info.delta_lf_multi = DEFAULT_DELTA_LF_MULTI; + // update delta_q_present_flag and delta_lf_present_flag based on + // base_qindex + cm->delta_q_info.delta_q_present_flag &= quant_params->base_qindex > 0; + cm->delta_q_info.delta_lf_present_flag &= quant_params->base_qindex > 0; + + av1_frame_init_quantizer(cpi); av1_initialize_rd_consts(cpi); - av1_initialize_me_consts(cpi, x, cm->base_qindex); + av1_initialize_me_consts(cpi, x, quant_params->base_qindex); + init_encode_frame_mb_context(cpi); - set_default_interp_skip_flags(cpi); - if (cm->prev_frame) + set_default_interp_skip_flags(cm, &cpi->interp_search_flags); + if (cm->prev_frame && cm->prev_frame->seg.enabled) cm->last_frame_seg_map = cm->prev_frame->seg_map; else cm->last_frame_seg_map = NULL; - cm->current_frame_seg_map = cm->cur_frame->seg_map; - if (cm->allow_intrabc || cm->coded_lossless) { + if (features->allow_intrabc || features->coded_lossless) { av1_set_default_ref_deltas(cm->lf.ref_deltas); av1_set_default_mode_deltas(cm->lf.mode_deltas); } else if (cm->prev_frame) { @@ -5115,205 +5797,251 @@ static void encode_frame_internal(AV1_COMP *cpi) { memcpy(cm->cur_frame->ref_deltas, cm->lf.ref_deltas, REF_FRAMES); memcpy(cm->cur_frame->mode_deltas, cm->lf.mode_deltas, MAX_MODE_LF_DELTAS); - // Special case: set prev_mi to NULL when the previous mode info - // context cannot be used. - cm->prev_mi = cm->allow_ref_frame_mvs ? cm->prev_mip : NULL; + cpi->all_one_sided_refs = + frame_is_intra_only(cm) ? 0 : refs_are_one_sided(cm); + + cpi->prune_ref_frame_mask = 0; + // Figure out which ref frames can be skipped at frame level. + setup_prune_ref_frame_mask(cpi); x->txb_split_count = 0; +#if CONFIG_SPEED_STATS + x->tx_search_count = 0; +#endif // CONFIG_SPEED_STATS +#if CONFIG_COLLECT_COMPONENT_TIMING + start_timing(cpi, av1_compute_global_motion_time); +#endif av1_zero(rdc->global_motion_used); - av1_zero(cpi->gmparams_cost); -#if !CONFIG_GLOBAL_MOTION_SEARCH - cpi->global_motion_search_done = 1; -#endif // !CONFIG_GLOBAL_MOTION_SEARCH - if (cpi->common.frame_type == INTER_FRAME && cpi->source && - !cpi->global_motion_search_done) { + av1_zero(gm_info->params_cost); + if (cpi->common.current_frame.frame_type == INTER_FRAME && cpi->source && + cpi->oxcf.enable_global_motion && !gm_info->search_done) { YV12_BUFFER_CONFIG *ref_buf[REF_FRAMES]; - int frame; - double params_by_motion[RANSAC_NUM_MOTIONS * (MAX_PARAMDIM - 1)]; - const double *params_this_motion; - int inliers_by_motion[RANSAC_NUM_MOTIONS]; - WarpedMotionParams tmp_wm_params; - static const double kIdentityParams[MAX_PARAMDIM - 1] = { - 0.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0 + MotionModel params_by_motion[RANSAC_NUM_MOTIONS]; + for (int m = 0; m < RANSAC_NUM_MOTIONS; m++) { + memset(¶ms_by_motion[m], 0, sizeof(params_by_motion[m])); + params_by_motion[m].inliers = + aom_malloc(sizeof(*(params_by_motion[m].inliers)) * 2 * MAX_CORNERS); + } + + int num_frm_corners = -1; + int frm_corners[2 * MAX_CORNERS]; + unsigned char *frm_buffer = cpi->source->y_buffer; + if (cpi->source->flags & YV12_FLAG_HIGHBITDEPTH) { + // The frame buffer is 16-bit, so we need to convert to 8 bits for the + // following code. We cache the result until the frame is released. + frm_buffer = + av1_downconvert_frame(cpi->source, cpi->common.seq_params.bit_depth); + } + const int segment_map_w = + (cpi->source->y_width + WARP_ERROR_BLOCK) >> WARP_ERROR_BLOCK_LOG; + const int segment_map_h = + (cpi->source->y_height + WARP_ERROR_BLOCK) >> WARP_ERROR_BLOCK_LOG; + + uint8_t *segment_map = + aom_malloc(sizeof(*segment_map) * segment_map_w * segment_map_h); + memset(segment_map, 0, + sizeof(*segment_map) * segment_map_w * segment_map_h); + + FrameDistPair future_ref_frame[REF_FRAMES - 1] = { + { -1, NONE_FRAME }, { -1, NONE_FRAME }, { -1, NONE_FRAME }, + { -1, NONE_FRAME }, { -1, NONE_FRAME }, { -1, NONE_FRAME }, + { -1, NONE_FRAME } + }; + FrameDistPair past_ref_frame[REF_FRAMES - 1] = { + { -1, NONE_FRAME }, { -1, NONE_FRAME }, { -1, NONE_FRAME }, + { -1, NONE_FRAME }, { -1, NONE_FRAME }, { -1, NONE_FRAME }, + { -1, NONE_FRAME } }; - int num_refs_using_gm = 0; + int num_past_ref_frames = 0; + int num_future_ref_frames = 0; + // Populate ref_buf for valid ref frames in global motion + update_valid_ref_frames_for_gm(cpi, ref_buf, past_ref_frame, + future_ref_frame, &num_past_ref_frames, + &num_future_ref_frames); + + // Sort the ref frames in the ascending order of their distance from the + // current frame + qsort(past_ref_frame, num_past_ref_frames, sizeof(past_ref_frame[0]), + compare_distance); + qsort(future_ref_frame, num_future_ref_frames, sizeof(future_ref_frame[0]), + compare_distance); + + // Compute global motion w.r.t. past reference frames + if (num_past_ref_frames > 0) + compute_global_motion_for_references( + cpi, ref_buf, past_ref_frame, num_past_ref_frames, &num_frm_corners, + frm_corners, frm_buffer, params_by_motion, segment_map, segment_map_w, + segment_map_h); + + // Compute global motion w.r.t. future reference frames + if (num_future_ref_frames > 0) + compute_global_motion_for_references( + cpi, ref_buf, future_ref_frame, num_future_ref_frames, + &num_frm_corners, frm_corners, frm_buffer, params_by_motion, + segment_map, segment_map_w, segment_map_h); + + aom_free(segment_map); + + gm_info->search_done = 1; + for (int m = 0; m < RANSAC_NUM_MOTIONS; m++) { + aom_free(params_by_motion[m].inliers); + } + } + memcpy(cm->cur_frame->global_motion, cm->global_motion, + REF_FRAMES * sizeof(WarpedMotionParams)); +#if CONFIG_COLLECT_COMPONENT_TIMING + end_timing(cpi, av1_compute_global_motion_time); +#endif - for (frame = ALTREF_FRAME; frame >= LAST_FRAME; --frame) { - ref_buf[frame] = get_ref_frame_buffer(cpi, frame); - int pframe; - cm->global_motion[frame] = default_warp_params; - const WarpedMotionParams *ref_params = - cm->prev_frame ? &cm->prev_frame->global_motion[frame] - : &default_warp_params; - // check for duplicate buffer - for (pframe = ALTREF_FRAME; pframe > frame; --pframe) { - if (ref_buf[frame] == ref_buf[pframe]) break; - } - if (pframe > frame) { - memcpy(&cm->global_motion[frame], &cm->global_motion[pframe], - sizeof(WarpedMotionParams)); - } else if (ref_buf[frame] && - ref_buf[frame]->y_crop_width == cpi->source->y_crop_width && - ref_buf[frame]->y_crop_height == cpi->source->y_crop_height && - do_gm_search_logic(&cpi->sf, num_refs_using_gm, frame) && - !(cpi->sf.selective_ref_gm && skip_gm_frame(cm, frame))) { - TransformationType model; - const int64_t ref_frame_error = - av1_frame_error(xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd, - ref_buf[frame]->y_buffer, ref_buf[frame]->y_stride, - cpi->source->y_buffer, cpi->source->y_width, - cpi->source->y_height, cpi->source->y_stride); - - if (ref_frame_error == 0) continue; - - aom_clear_system_state(); - for (model = ROTZOOM; model < GLOBAL_TRANS_TYPES_ENC; ++model) { - int64_t best_warp_error = INT64_MAX; - // Initially set all params to identity. - for (i = 0; i < RANSAC_NUM_MOTIONS; ++i) { - memcpy(params_by_motion + (MAX_PARAMDIM - 1) * i, kIdentityParams, - (MAX_PARAMDIM - 1) * sizeof(*params_by_motion)); - } +#if CONFIG_COLLECT_COMPONENT_TIMING + start_timing(cpi, av1_setup_motion_field_time); +#endif + if (features->allow_ref_frame_mvs) av1_setup_motion_field(cm); +#if CONFIG_COLLECT_COMPONENT_TIMING + end_timing(cpi, av1_setup_motion_field_time); +#endif - compute_global_motion_feature_based( - model, cpi->source, ref_buf[frame], - cpi->common.seq_params.bit_depth, inliers_by_motion, - params_by_motion, RANSAC_NUM_MOTIONS); - - for (i = 0; i < RANSAC_NUM_MOTIONS; ++i) { - if (inliers_by_motion[i] == 0) continue; - - params_this_motion = params_by_motion + (MAX_PARAMDIM - 1) * i; - convert_model_to_params(params_this_motion, &tmp_wm_params); - - if (tmp_wm_params.wmtype != IDENTITY) { - const int64_t warp_error = refine_integerized_param( - &tmp_wm_params, tmp_wm_params.wmtype, - xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd, - ref_buf[frame]->y_buffer, ref_buf[frame]->y_width, - ref_buf[frame]->y_height, ref_buf[frame]->y_stride, - cpi->source->y_buffer, cpi->source->y_width, - cpi->source->y_height, cpi->source->y_stride, 5, - best_warp_error); - if (warp_error < best_warp_error) { - best_warp_error = warp_error; - // Save the wm_params modified by refine_integerized_param() - // rather than motion index to avoid rerunning refine() below. - memcpy(&(cm->global_motion[frame]), &tmp_wm_params, - sizeof(WarpedMotionParams)); - } - } - } - if (cm->global_motion[frame].wmtype <= AFFINE) - if (!get_shear_params(&cm->global_motion[frame])) - cm->global_motion[frame] = default_warp_params; - - if (cm->global_motion[frame].wmtype == TRANSLATION) { - cm->global_motion[frame].wmmat[0] = - convert_to_trans_prec(cm->allow_high_precision_mv, - cm->global_motion[frame].wmmat[0]) * - GM_TRANS_ONLY_DECODE_FACTOR; - cm->global_motion[frame].wmmat[1] = - convert_to_trans_prec(cm->allow_high_precision_mv, - cm->global_motion[frame].wmmat[1]) * - GM_TRANS_ONLY_DECODE_FACTOR; - } + cm->current_frame.skip_mode_info.skip_mode_flag = + check_skip_mode_enabled(cpi); - // If the best error advantage found doesn't meet the threshold for - // this motion type, revert to IDENTITY. - if (!is_enough_erroradvantage( - (double)best_warp_error / ref_frame_error, - gm_get_params_cost(&cm->global_motion[frame], ref_params, - cm->allow_high_precision_mv), - cpi->sf.gm_erroradv_type)) { - cm->global_motion[frame] = default_warp_params; - } - if (cm->global_motion[frame].wmtype != IDENTITY) break; - } - aom_clear_system_state(); - } - if (cm->global_motion[frame].wmtype != IDENTITY) num_refs_using_gm++; - cpi->gmparams_cost[frame] = - gm_get_params_cost(&cm->global_motion[frame], ref_params, - cm->allow_high_precision_mv) + - cpi->gmtype_cost[cm->global_motion[frame].wmtype] - - cpi->gmtype_cost[IDENTITY]; - } - // clear disabled ref_frames - for (frame = LAST_FRAME; frame <= ALTREF_FRAME; ++frame) { - const int ref_disabled = - !(cpi->ref_frame_flags & ref_frame_flag_list[frame]); - if (ref_disabled && cpi->sf.recode_loop != DISALLOW_RECODE) { - cpi->gmparams_cost[frame] = 0; - cm->global_motion[frame] = default_warp_params; + cpi->row_mt_sync_read_ptr = av1_row_mt_sync_read_dummy; + cpi->row_mt_sync_write_ptr = av1_row_mt_sync_write_dummy; + cpi->row_mt = 0; + + if (cpi->oxcf.row_mt && (cpi->oxcf.max_threads > 1)) { + cpi->row_mt = 1; + cpi->row_mt_sync_read_ptr = av1_row_mt_sync_read; + cpi->row_mt_sync_write_ptr = av1_row_mt_sync_write; + av1_encode_tiles_row_mt(cpi); + } else { + if (AOMMIN(cpi->oxcf.max_threads, cm->tiles.cols * cm->tiles.rows) > 1) + av1_encode_tiles_mt(cpi); + else + encode_tiles(cpi); + } + + // If intrabc is allowed but never selected, reset the allow_intrabc flag. + if (features->allow_intrabc && !cpi->intrabc_used) { + features->allow_intrabc = 0; + } + if (features->allow_intrabc) { + cm->delta_q_info.delta_lf_present_flag = 0; + } + + if (cm->delta_q_info.delta_q_present_flag && cpi->deltaq_used == 0) { + cm->delta_q_info.delta_q_present_flag = 0; + } + + // Set the transform size appropriately before bitstream creation + const MODE_EVAL_TYPE eval_type = + cpi->sf.winner_mode_sf.enable_winner_mode_for_tx_size_srch + ? WINNER_MODE_EVAL + : DEFAULT_EVAL; + const TX_SIZE_SEARCH_METHOD tx_search_type = + cpi->winner_mode_params.tx_size_search_methods[eval_type]; + assert(cpi->oxcf.enable_tx64 || tx_search_type != USE_LARGESTALL); + features->tx_mode = select_tx_mode(cm, tx_search_type); + + if (cpi->sf.tx_sf.tx_type_search.prune_tx_type_using_stats) { + const FRAME_UPDATE_TYPE update_type = get_frame_update_type(&cpi->gf_group); + + for (i = 0; i < TX_SIZES_ALL; i++) { + int sum = 0; + int j; + int left = 1024; + + for (j = 0; j < TX_TYPES; j++) + sum += cpi->td.rd_counts.tx_type_used[i][j]; + + for (j = TX_TYPES - 1; j >= 0; j--) { + const int new_prob = + sum ? 1024 * cpi->td.rd_counts.tx_type_used[i][j] / sum + : (j ? 0 : 1024); + int prob = + (frame_probs->tx_type_probs[update_type][i][j] + new_prob) >> 1; + left -= prob; + if (j == 0) prob += left; + frame_probs->tx_type_probs[update_type][i][j] = prob; } } - cpi->global_motion_search_done = 1; } - memcpy(cm->cur_frame->global_motion, cm->global_motion, - REF_FRAMES * sizeof(WarpedMotionParams)); - av1_setup_motion_field(cm); + if (!cpi->sf.inter_sf.disable_obmc && + cpi->sf.inter_sf.prune_obmc_prob_thresh > 0) { + const FRAME_UPDATE_TYPE update_type = get_frame_update_type(&cpi->gf_group); - cpi->all_one_sided_refs = - frame_is_intra_only(cm) ? 0 : av1_refs_are_one_sided(cm); + for (i = 0; i < BLOCK_SIZES_ALL; i++) { + int sum = 0; + for (int j = 0; j < 2; j++) sum += cpi->td.rd_counts.obmc_used[i][j]; - cm->skip_mode_flag = check_skip_mode_enabled(cpi); + const int new_prob = + sum ? 128 * cpi->td.rd_counts.obmc_used[i][1] / sum : 0; + frame_probs->obmc_probs[update_type][i] = + (frame_probs->obmc_probs[update_type][i] + new_prob) >> 1; + } + } - { - struct aom_usec_timer emr_timer; - aom_usec_timer_start(&emr_timer); + if (features->allow_warped_motion && + cpi->sf.inter_sf.prune_warped_prob_thresh > 0) { + const FRAME_UPDATE_TYPE update_type = get_frame_update_type(&cpi->gf_group); + int sum = 0; + for (i = 0; i < 2; i++) sum += cpi->td.rd_counts.warped_used[i]; + const int new_prob = sum ? 128 * cpi->td.rd_counts.warped_used[1] / sum : 0; + frame_probs->warped_probs[update_type] = + (frame_probs->warped_probs[update_type] + new_prob) >> 1; + } -#if CONFIG_FP_MB_STATS - if (cpi->use_fp_mb_stats) { - input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm, - &cpi->twopass.this_frame_mb_stats); - } -#endif + if (cm->current_frame.frame_type != KEY_FRAME && + cpi->sf.interp_sf.adaptive_interp_filter_search == 2 && + features->interp_filter == SWITCHABLE) { + const FRAME_UPDATE_TYPE update_type = get_frame_update_type(&cpi->gf_group); - if (cpi->row_mt && (cpi->oxcf.max_threads > 1)) - av1_encode_tiles_mt(cpi); - else if (AOMMIN(cpi->oxcf.max_threads, cm->tile_cols * cm->tile_rows) > 1) - av1_encode_tiles_mt(cpi); - else - encode_tiles(cpi); + for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) { + int sum = 0; + int j; + int left = 1536; - aom_usec_timer_mark(&emr_timer); - cpi->time_encode_sb_row += aom_usec_timer_elapsed(&emr_timer); + for (j = 0; j < SWITCHABLE_FILTERS; j++) { + sum += cpi->td.counts->switchable_interp[i][j]; + } + + for (j = SWITCHABLE_FILTERS - 1; j >= 0; j--) { + const int new_prob = + sum ? 1536 * cpi->td.counts->switchable_interp[i][j] / sum + : (j ? 0 : 1536); + int prob = (frame_probs->switchable_interp_probs[update_type][i][j] + + new_prob) >> + 1; + left -= prob; + if (j == 0) prob += left; + frame_probs->switchable_interp_probs[update_type][i][j] = prob; + } + } } - // If intrabc is allowed but never selected, reset the allow_intrabc flag. - if (cm->allow_intrabc && !cpi->intrabc_used) cm->allow_intrabc = 0; - if (cm->allow_intrabc) cm->delta_lf_present_flag = 0; + if ((!is_stat_generation_stage(cpi) && av1_use_hash_me(cpi) && + !cpi->sf.rt_sf.use_nonrd_pick_mode) || + hash_table_created) { + av1_hash_table_destroy(&intrabc_hash_info->intrabc_hash_table); + } } void av1_encode_frame(AV1_COMP *cpi) { AV1_COMMON *const cm = &cpi->common; + CurrentFrame *const current_frame = &cm->current_frame; + FeatureFlags *const features = &cm->features; const int num_planes = av1_num_planes(cm); // Indicates whether or not to use a default reduced set for ext-tx // rather than the potential full set of 16 transforms - cm->reduced_tx_set_used = 0; - - if (cm->show_frame == 0) { - int arf_offset = AOMMIN( - (MAX_GF_INTERVAL - 1), - cpi->twopass.gf_group.arf_src_offset[cpi->twopass.gf_group.index]); - int brf_offset = - cpi->twopass.gf_group.brf_src_offset[cpi->twopass.gf_group.index]; - arf_offset = AOMMIN((MAX_GF_INTERVAL - 1), arf_offset + brf_offset); - cm->frame_offset = cm->current_video_frame + arf_offset; - } else { - cm->frame_offset = cm->current_video_frame; - } - cm->frame_offset %= (1 << (cm->seq_params.order_hint_bits_minus_1 + 1)); + features->reduced_tx_set_used = cpi->oxcf.reduced_tx_type_set; // Make sure segment_id is no larger than last_active_segid. if (cm->seg.enabled && cm->seg.update_map) { - const int mi_rows = cm->mi_rows; - const int mi_cols = cm->mi_cols; + const int mi_rows = cm->mi_params.mi_rows; + const int mi_cols = cm->mi_params.mi_cols; const int last_active_segid = cm->seg.last_active_segid; - uint8_t *map = cpi->segmentation_map; + uint8_t *map = cpi->enc_seg.map; for (int mi_row = 0; mi_row < mi_rows; ++mi_row) { for (int mi_col = 0; mi_col < mi_cols; ++mi_col) { map[mi_col] = AOMMIN(map[mi_col], last_active_segid); @@ -5323,86 +6051,105 @@ void av1_encode_frame(AV1_COMP *cpi) { } av1_setup_frame_buf_refs(cm); - if (cpi->sf.selective_ref_frame >= 2) enforce_max_ref_frames(cpi); + enforce_max_ref_frames(cpi, &cpi->ref_frame_flags); + set_rel_frame_dist(cpi); av1_setup_frame_sign_bias(cm); +#if CHECK_PRECOMPUTED_REF_FRAME_MAP + GF_GROUP *gf_group = &cpi->gf_group; + // TODO(yuec): The check is disabled on OVERLAY frames for now, because info + // in cpi->gf_group has been refreshed for the next GOP when the check is + // performed for OVERLAY frames. Since we have not support inter-GOP ref + // frame map computation, the precomputed ref map for an OVERLAY frame is all + // -1 at this point (although it is meaning before gf_group is refreshed). + if (!frame_is_intra_only(cm) && gf_group->index != 0) { + const RefCntBuffer *const golden_buf = get_ref_frame_buf(cm, GOLDEN_FRAME); + + if (golden_buf) { + const int golden_order_hint = golden_buf->order_hint; + + for (int ref = LAST_FRAME; ref < EXTREF_FRAME; ++ref) { + const RefCntBuffer *const buf = get_ref_frame_buf(cm, ref); + const int ref_disp_idx_precomputed = + gf_group->ref_frame_disp_idx[gf_group->index][ref - LAST_FRAME]; + + (void)ref_disp_idx_precomputed; + + if (buf != NULL) { + const int ref_disp_idx = + get_relative_dist(&cm->seq_params.order_hint_info, + buf->order_hint, golden_order_hint); + + if (ref_disp_idx >= 0) + assert(ref_disp_idx == ref_disp_idx_precomputed); + else + assert(ref_disp_idx_precomputed == -1); + } else { + assert(ref_disp_idx_precomputed == -1); + } + } + } + } +#endif + #if CONFIG_MISMATCH_DEBUG mismatch_reset_frame(num_planes); #else (void)num_planes; #endif - cpi->allow_comp_inter_inter = !frame_is_intra_only(cm); - - if (cpi->sf.frame_parameter_update) { - int i; - RD_OPT *const rd_opt = &cpi->rd; + if (cpi->sf.hl_sf.frame_parameter_update) { RD_COUNTS *const rdc = &cpi->td.rd_counts; - // This code does a single RD pass over the whole frame assuming - // either compound, single or hybrid prediction as per whatever has - // worked best for that type of frame in the past. - // It also predicts whether another coding mode would have worked - // better than this coding mode. If that is the case, it remembers - // that for subsequent frames. - // It does the same analysis for transform size selection also. - // - // TODO(zoeliu): To investigate whether a frame_type other than - // INTRA/ALTREF/GOLDEN/LAST needs to be specified seperately. - const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi); - int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type]; - const int is_alt_ref = frame_type == ALTREF_FRAME; - - /* prediction (compound, single or hybrid) mode selection */ - // NOTE: "is_alt_ref" is true only for OVERLAY/INTNL_OVERLAY frames - if (is_alt_ref || !cpi->allow_comp_inter_inter) - cm->reference_mode = SINGLE_REFERENCE; + if (frame_is_intra_only(cm)) + current_frame->reference_mode = SINGLE_REFERENCE; else - cm->reference_mode = REFERENCE_MODE_SELECT; + current_frame->reference_mode = REFERENCE_MODE_SELECT; - cm->interp_filter = SWITCHABLE; - if (cm->large_scale_tile) cm->interp_filter = EIGHTTAP_REGULAR; + features->interp_filter = SWITCHABLE; + if (cm->tiles.large_scale) features->interp_filter = EIGHTTAP_REGULAR; - cm->switchable_motion_mode = 1; + features->switchable_motion_mode = 1; rdc->compound_ref_used_flag = 0; rdc->skip_mode_used_flag = 0; encode_frame_internal(cpi); - for (i = 0; i < REFERENCE_MODES; ++i) - mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2; - - if (cm->reference_mode == REFERENCE_MODE_SELECT) { + if (current_frame->reference_mode == REFERENCE_MODE_SELECT) { // Use a flag that includes 4x4 blocks if (rdc->compound_ref_used_flag == 0) { - cm->reference_mode = SINGLE_REFERENCE; + current_frame->reference_mode = SINGLE_REFERENCE; #if CONFIG_ENTROPY_STATS av1_zero(cpi->td.counts->comp_inter); #endif // CONFIG_ENTROPY_STATS } } - // Re-check on the skip mode status as reference mode may have been changed. - if (frame_is_intra_only(cm) || cm->reference_mode == SINGLE_REFERENCE) { - cm->is_skip_mode_allowed = 0; - cm->skip_mode_flag = 0; + // Re-check on the skip mode status as reference mode may have been + // changed. + SkipModeInfo *const skip_mode_info = ¤t_frame->skip_mode_info; + if (frame_is_intra_only(cm) || + current_frame->reference_mode == SINGLE_REFERENCE) { + skip_mode_info->skip_mode_allowed = 0; + skip_mode_info->skip_mode_flag = 0; } - if (cm->skip_mode_flag && rdc->skip_mode_used_flag == 0) - cm->skip_mode_flag = 0; + if (skip_mode_info->skip_mode_flag && rdc->skip_mode_used_flag == 0) + skip_mode_info->skip_mode_flag = 0; - if (!cm->large_scale_tile) { - if (cm->tx_mode == TX_MODE_SELECT && cpi->td.mb.txb_split_count == 0) - cm->tx_mode = TX_MODE_LARGEST; + if (!cm->tiles.large_scale) { + if (features->tx_mode == TX_MODE_SELECT && + cpi->td.mb.txb_split_count == 0) + features->tx_mode = TX_MODE_LARGEST; } } else { encode_frame_internal(cpi); } } -static void update_txfm_count(MACROBLOCK *x, MACROBLOCKD *xd, - FRAME_COUNTS *counts, TX_SIZE tx_size, int depth, - int blk_row, int blk_col, - uint8_t allow_update_cdf) { +static AOM_INLINE void update_txfm_count(MACROBLOCK *x, MACROBLOCKD *xd, + FRAME_COUNTS *counts, TX_SIZE tx_size, + int depth, int blk_row, int blk_col, + uint8_t allow_update_cdf) { MB_MODE_INFO *mbmi = xd->mi[0]; const BLOCK_SIZE bsize = mbmi->sb_type; const int max_blocks_high = max_block_high(xd, bsize, 0); @@ -5465,30 +6212,33 @@ static void update_txfm_count(MACROBLOCK *x, MACROBLOCKD *xd, } } -static void tx_partition_count_update(const AV1_COMMON *const cm, MACROBLOCK *x, - BLOCK_SIZE plane_bsize, int mi_row, - int mi_col, FRAME_COUNTS *td_counts, - uint8_t allow_update_cdf) { +static AOM_INLINE void tx_partition_count_update(const AV1_COMMON *const cm, + MACROBLOCK *x, + BLOCK_SIZE plane_bsize, + FRAME_COUNTS *td_counts, + uint8_t allow_update_cdf) { MACROBLOCKD *xd = &x->e_mbd; - const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0]; - const int mi_height = block_size_high[plane_bsize] >> tx_size_high_log2[0]; + const int mi_width = mi_size_wide[plane_bsize]; + const int mi_height = mi_size_high[plane_bsize]; const TX_SIZE max_tx_size = get_vartx_max_txsize(xd, plane_bsize, 0); const int bh = tx_size_high_unit[max_tx_size]; const int bw = tx_size_wide_unit[max_tx_size]; - int idx, idy; - xd->above_txfm_context = cm->above_txfm_context[xd->tile.tile_row] + mi_col; + xd->above_txfm_context = + cm->above_contexts.txfm[xd->tile.tile_row] + xd->mi_col; xd->left_txfm_context = - xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); + xd->left_txfm_context_buffer + (xd->mi_row & MAX_MIB_MASK); - for (idy = 0; idy < mi_height; idy += bh) - for (idx = 0; idx < mi_width; idx += bw) + for (int idy = 0; idy < mi_height; idy += bh) { + for (int idx = 0; idx < mi_width; idx += bw) { update_txfm_count(x, xd, td_counts, max_tx_size, 0, idy, idx, allow_update_cdf); + } + } } -static void set_txfm_context(MACROBLOCKD *xd, TX_SIZE tx_size, int blk_row, - int blk_col) { +static AOM_INLINE void set_txfm_context(MACROBLOCKD *xd, TX_SIZE tx_size, + int blk_row, int blk_col) { MB_MODE_INFO *mbmi = xd->mi[0]; const BLOCK_SIZE bsize = mbmi->sb_type; const int max_blocks_high = max_block_high(xd, bsize, 0); @@ -5525,29 +6275,31 @@ static void set_txfm_context(MACROBLOCKD *xd, TX_SIZE tx_size, int blk_row, } } -static void tx_partition_set_contexts(const AV1_COMMON *const cm, - MACROBLOCKD *xd, BLOCK_SIZE plane_bsize, - int mi_row, int mi_col) { - const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0]; - const int mi_height = block_size_high[plane_bsize] >> tx_size_high_log2[0]; +static AOM_INLINE void tx_partition_set_contexts(const AV1_COMMON *const cm, + MACROBLOCKD *xd, + BLOCK_SIZE plane_bsize) { + const int mi_width = mi_size_wide[plane_bsize]; + const int mi_height = mi_size_high[plane_bsize]; const TX_SIZE max_tx_size = get_vartx_max_txsize(xd, plane_bsize, 0); const int bh = tx_size_high_unit[max_tx_size]; const int bw = tx_size_wide_unit[max_tx_size]; - int idx, idy; - xd->above_txfm_context = cm->above_txfm_context[xd->tile.tile_row] + mi_col; + xd->above_txfm_context = + cm->above_contexts.txfm[xd->tile.tile_row] + xd->mi_col; xd->left_txfm_context = - xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK); + xd->left_txfm_context_buffer + (xd->mi_row & MAX_MIB_MASK); - for (idy = 0; idy < mi_height; idy += bh) - for (idx = 0; idx < mi_width; idx += bw) + for (int idy = 0; idy < mi_height; idy += bh) { + for (int idx = 0; idx < mi_width; idx += bw) { set_txfm_context(xd, max_tx_size, idy, idx); + } + } } -static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, - ThreadData *td, TOKENEXTRA **t, RUN_TYPE dry_run, - int mi_row, int mi_col, BLOCK_SIZE bsize, - int *rate) { +static AOM_INLINE void encode_superblock(const AV1_COMP *const cpi, + TileDataEnc *tile_data, ThreadData *td, + TOKENEXTRA **t, RUN_TYPE dry_run, + BLOCK_SIZE bsize, int *rate) { const AV1_COMMON *const cm = &cpi->common; const int num_planes = av1_num_planes(cm); MACROBLOCK *const x = &td->mb; @@ -5556,53 +6308,35 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, MB_MODE_INFO *mbmi = mi_4x4[0]; const int seg_skip = segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP); - const int mis = cm->mi_stride; + const int mis = cm->mi_params.mi_stride; const int mi_width = mi_size_wide[bsize]; const int mi_height = mi_size_high[bsize]; const int is_inter = is_inter_block(mbmi); - if (cpi->sf.mode_pruning_based_on_two_pass_partition_search && - x->cb_partition_scan) { - for (int row = mi_row; row < mi_row + mi_width; - row += FIRST_PARTITION_PASS_SAMPLE_REGION) { - for (int col = mi_col; col < mi_col + mi_height; - col += FIRST_PARTITION_PASS_SAMPLE_REGION) { - const int index = av1_first_partition_pass_stats_index(row, col); - FIRST_PARTITION_PASS_STATS *const stats = - &x->first_partition_pass_stats[index]; - // Increase the counter of data samples. - ++stats->sample_counts; - // Increase the counter for ref_frame[0] and ref_frame[1]. - if (stats->ref0_counts[mbmi->ref_frame[0]] < 255) - ++stats->ref0_counts[mbmi->ref_frame[0]]; - if (mbmi->ref_frame[1] >= 0 && - stats->ref1_counts[mbmi->ref_frame[0]] < 255) - ++stats->ref1_counts[mbmi->ref_frame[1]]; - } - } - } + // Initialize tx_mode and tx_size_search_method + set_tx_size_search_method( + cm, &cpi->winner_mode_params, x, + cpi->sf.winner_mode_sf.enable_winner_mode_for_tx_size_srch, 1); + const int mi_row = xd->mi_row; + const int mi_col = xd->mi_col; if (!is_inter) { - xd->cfl.is_chroma_reference = - is_chroma_reference(mi_row, mi_col, bsize, cm->seq_params.subsampling_x, - cm->seq_params.subsampling_y); xd->cfl.store_y = store_cfl_required(cm, xd); mbmi->skip = 1; for (int plane = 0; plane < num_planes; ++plane) { - av1_encode_intra_block_plane(cpi, x, bsize, plane, - cpi->optimize_seg_arr[mbmi->segment_id], - mi_row, mi_col); + av1_encode_intra_block_plane(cpi, x, bsize, plane, dry_run, + cpi->optimize_seg_arr[mbmi->segment_id]); } // If there is at least one lossless segment, force the skip for intra // block to be 0, in order to avoid the segment_id to be changed by in // write_segment_id(). if (!cpi->common.seg.segid_preskip && cpi->common.seg.update_map && - cpi->has_lossless_segment) + cpi->enc_seg.has_lossless_segment) mbmi->skip = 0; xd->cfl.store_y = 0; - if (av1_allow_palette(cm->allow_screen_content_tools, bsize)) { + if (av1_allow_palette(cm->features.allow_screen_content_tools, bsize)) { for (int plane = 0; plane < AOMMIN(2, num_planes); ++plane) { if (mbmi->palette_mode_info.palette_size[plane] > 0) { if (!dry_run) { @@ -5617,7 +6351,7 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, } } - av1_update_txb_context(cpi, td, dry_run, bsize, rate, mi_row, mi_col, + av1_update_txb_context(cpi, td, dry_run, bsize, tile_data->allow_update_cdf); } else { int ref; @@ -5625,15 +6359,19 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]); for (ref = 0; ref < 1 + is_compound; ++ref) { - YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, mbmi->ref_frame[ref]); + const YV12_BUFFER_CONFIG *cfg = + get_ref_frame_yv12_buf(cm, mbmi->ref_frame[ref]); assert(IMPLIES(!is_intrabc_block(mbmi), cfg)); av1_setup_pre_planes(xd, ref, cfg, mi_row, mi_col, - &xd->block_refs[ref]->sf, num_planes); + xd->block_ref_scale_factors[ref], num_planes); + } + int start_plane = (cpi->sf.rt_sf.reuse_inter_pred_nonrd) ? 1 : 0; + av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, + start_plane, av1_num_planes(cm) - 1); + if (mbmi->motion_mode == OBMC_CAUSAL) { + assert(cpi->oxcf.enable_obmc == 1); + av1_build_obmc_inter_predictors_sb(cm, xd); } - - av1_build_inter_predictors_sb(cm, xd, mi_row, mi_col, NULL, bsize); - if (mbmi->motion_mode == OBMC_CAUSAL) - av1_build_obmc_inter_predictors_sb(cm, xd, mi_row, mi_col); #if CONFIG_MISMATCH_DEBUG if (dry_run == OUTPUT_ENABLED) { @@ -5645,9 +6383,9 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x, pd->subsampling_y)) continue; - mismatch_record_block_pre(pd->dst.buf, pd->dst.stride, cm->frame_offset, - plane, pixel_c, pixel_r, pd->width, - pd->height, + mismatch_record_block_pre(pd->dst.buf, pd->dst.stride, + cm->current_frame.order_hint, plane, pixel_c, + pixel_r, pd->width, pd->height, xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH); } } @@ -5655,18 +6393,18 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, (void)num_planes; #endif - av1_encode_sb(cpi, x, bsize, mi_row, mi_col, dry_run); - av1_tokenize_sb_vartx(cpi, td, t, dry_run, mi_row, mi_col, bsize, rate, + av1_encode_sb(cpi, x, bsize, dry_run); + av1_tokenize_sb_vartx(cpi, td, dry_run, bsize, rate, tile_data->allow_update_cdf); } if (!dry_run) { - if (av1_allow_intrabc(cm) && is_intrabc_block(mbmi)) - td->intrabc_used_this_tile = 1; - if (cm->tx_mode == TX_MODE_SELECT && !xd->lossless[mbmi->segment_id] && - mbmi->sb_type > BLOCK_4X4 && !(is_inter && (mbmi->skip || seg_skip))) { + if (av1_allow_intrabc(cm) && is_intrabc_block(mbmi)) td->intrabc_used = 1; + if (x->tx_mode_search_type == TX_MODE_SELECT && + !xd->lossless[mbmi->segment_id] && mbmi->sb_type > BLOCK_4X4 && + !(is_inter && (mbmi->skip || seg_skip))) { if (is_inter) { - tx_partition_count_update(cm, x, bsize, mi_row, mi_col, td->counts, + tx_partition_count_update(cm, x, bsize, td->counts, tile_data->allow_update_cdf); } else { if (mbmi->tx_size != max_txsize_rect_lookup[bsize]) @@ -5694,7 +6432,7 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, if (xd->lossless[mbmi->segment_id]) { intra_tx_size = TX_4X4; } else { - intra_tx_size = tx_size_from_tx_mode(bsize, cm->tx_mode); + intra_tx_size = tx_size_from_tx_mode(bsize, x->tx_mode_search_type); } } else { intra_tx_size = mbmi->tx_size; @@ -5702,17 +6440,18 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, for (j = 0; j < mi_height; j++) for (i = 0; i < mi_width; i++) - if (mi_col + i < cm->mi_cols && mi_row + j < cm->mi_rows) + if (mi_col + i < cm->mi_params.mi_cols && + mi_row + j < cm->mi_params.mi_rows) mi_4x4[mis * j + i]->tx_size = intra_tx_size; if (intra_tx_size != max_txsize_rect_lookup[bsize]) ++x->txb_split_count; } } - if (cm->tx_mode == TX_MODE_SELECT && block_signals_txsize(mbmi->sb_type) && - is_inter && !(mbmi->skip || seg_skip) && - !xd->lossless[mbmi->segment_id]) { - if (dry_run) tx_partition_set_contexts(cm, xd, bsize, mi_row, mi_col); + if (x->tx_mode_search_type == TX_MODE_SELECT && + block_signals_txsize(mbmi->sb_type) && is_inter && + !(mbmi->skip || seg_skip) && !xd->lossless[mbmi->segment_id]) { + if (dry_run) tx_partition_set_contexts(cm, xd, bsize); } else { TX_SIZE tx_size = mbmi->tx_size; // The new intra coding scheme requires no change of transform size @@ -5720,20 +6459,17 @@ static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data, if (xd->lossless[mbmi->segment_id]) { tx_size = TX_4X4; } else { - tx_size = tx_size_from_tx_mode(bsize, cm->tx_mode); + tx_size = tx_size_from_tx_mode(bsize, x->tx_mode_search_type); } } else { tx_size = (bsize > BLOCK_4X4) ? tx_size : TX_4X4; } mbmi->tx_size = tx_size; - set_txfm_ctxs(tx_size, xd->n4_w, xd->n4_h, + set_txfm_ctxs(tx_size, xd->width, xd->height, (mbmi->skip || seg_skip) && is_inter_block(mbmi), xd); } - CFL_CTX *const cfl = &xd->cfl; - if (is_inter_block(mbmi) && - !is_chroma_reference(mi_row, mi_col, bsize, cfl->subsampling_x, - cfl->subsampling_y) && - is_cfl_allowed(xd)) { + + if (is_inter_block(mbmi) && !xd->is_chroma_ref && is_cfl_allowed(xd)) { cfl_store_block(xd, mbmi->sb_type, mbmi->tx_size); } } |