vp9_pickmode.c 27.28 KiB
/*
 *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */
#include <assert.h>
#include <limits.h>
#include <math.h>
#include <stdio.h>
#include "./vp9_rtcd.h"
#include "vpx_mem/vpx_mem.h"
#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/encoder/vp9_encoder.h"
#include "vp9/encoder/vp9_pickmode.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vp9/encoder/vp9_rdopt.h"
static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
                       const TileInfo *const tile,
                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
                       int_mv *mv_ref_list,
                       int mi_row, int mi_col) {
  const int *ref_sign_bias = cm->ref_frame_sign_bias;
  int i, refmv_count = 0;
  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
  int different_ref_found = 0;
  int context_counter = 0;
  int const_motion = 0;
  // Blank the reference vector list
  vpx_memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
  // The nearest 2 blocks are treated differently
  // if the size < 8x8 we get the mv from the bmi substructure,
  // and we also need to keep a mode count.
  for (i = 0; i < 2; ++i) {
    const POSITION *const mv_ref = &mv_ref_search[i];
    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
                                                   xd->mi_stride];
      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
      // Keep counts for entropy encoding.
      context_counter += mode_2_counter[candidate->mode];
      different_ref_found = 1;
      if (candidate->ref_frame[0] == ref_frame)
        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1));
  const_motion = 1;
  // Check the rest of the neighbors in much the same way
  // as before except we don't need to keep track of sub blocks or
  // mode counts.
  for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
7172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140
const POSITION *const mv_ref = &mv_ref_search[i]; if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) { const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]->mbmi; different_ref_found = 1; if (candidate->ref_frame[0] == ref_frame) ADD_MV_REF_LIST(candidate->mv[0]); } } // Since we couldn't find 2 mvs from the same reference frame // go back through the neighbors and find motion vectors from // different reference frames. if (different_ref_found && !refmv_count) { for (i = 0; i < MVREF_NEIGHBOURS; ++i) { const POSITION *mv_ref = &mv_ref_search[i]; if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) { const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]->mbmi; // If the candidate is INTRA we don't want to consider its mv. IF_DIFF_REF_FRAME_ADD_MV(candidate); } } } Done: mi->mbmi.mode_context[ref_frame] = counter_to_context[context_counter]; // Clamp vectors for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i) clamp_mv_ref(&mv_ref_list[i].as_mv, xd); return const_motion; } static void full_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row, int mi_col, int_mv *tmp_mv, int *rate_mv) { MACROBLOCKD *xd = &x->e_mbd; MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi; struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}}; int step_param; int sadpb = x->sadperbit16; MV mvp_full; int ref = mbmi->ref_frame[0]; const MV ref_mv = mbmi->ref_mvs[ref][0].as_mv; int i; int tmp_col_min = x->mv_col_min; int tmp_col_max = x->mv_col_max; int tmp_row_min = x->mv_row_min; int tmp_row_max = x->mv_row_max; const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, ref); if (scaled_ref_frame) { int i; // Swap out the reference frame for a version that's been scaled to // match the resolution of the current frame, allowing the existing // motion search code to be used without additional modifications. for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0]; vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); } vp9_set_mv_search_range(x, &ref_mv);
141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210
// TODO(jingning) exploiting adaptive motion search control in non-RD // mode decision too. step_param = 6; for (i = LAST_FRAME; i <= LAST_FRAME && cpi->common.show_frame; ++i) { if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) { tmp_mv->as_int = INVALID_MV; if (scaled_ref_frame) { int i; for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i]; } return; } } assert(x->mv_best_ref_index[ref] <= 2); if (x->mv_best_ref_index[ref] < 2) mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv; else mvp_full = x->pred_mv[ref]; mvp_full.col >>= 3; mvp_full.row >>= 3; vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb, &ref_mv, &tmp_mv->as_mv, INT_MAX, 0); x->mv_col_min = tmp_col_min; x->mv_col_max = tmp_col_max; x->mv_row_min = tmp_row_min; x->mv_row_max = tmp_row_max; if (scaled_ref_frame) { int i; for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i]; } // calculate the bit cost on motion vector mvp_full.row = tmp_mv->as_mv.row * 8; mvp_full.col = tmp_mv->as_mv.col * 8; *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); } static void sub_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row, int mi_col, MV *tmp_mv) { MACROBLOCKD *xd = &x->e_mbd; MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi; struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}}; int ref = mbmi->ref_frame[0]; MV ref_mv = mbmi->ref_mvs[ref][0].as_mv; int dis; const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, ref); if (scaled_ref_frame) { int i; // Swap out the reference frame for a version that's been scaled to // match the resolution of the current frame, allowing the existing // motion search code to be used without additional modifications. for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0]; vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); }
211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280
cpi->find_fractional_mv_step(x, tmp_mv, &ref_mv, cpi->common.allow_high_precision_mv, x->errorperbit, &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop, cpi->sf.mv.subpel_iters_per_step, x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref]); if (scaled_ref_frame) { int i; for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i]; } x->pred_mv[ref] = *tmp_mv; } static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x, MACROBLOCKD *xd, int *out_rate_sum, int64_t *out_dist_sum, unsigned int *var_y, unsigned int *sse_y) { // Note our transform coeffs are 8 times an orthogonal transform. // Hence quantizer step is also 8 times. To get effective quantizer // we need to divide by 8 before sending to modeling function. unsigned int sse; int rate; int64_t dist; struct macroblock_plane *const p = &x->plane[0]; struct macroblockd_plane *const pd = &xd->plane[0]; const uint32_t dc_quant = pd->dequant[0]; const uint32_t ac_quant = pd->dequant[1]; unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse); *var_y = var; *sse_y = sse; if (sse < dc_quant * dc_quant >> 6) x->skip_txfm = 1; else if (var < ac_quant * ac_quant >> 6) x->skip_txfm = 2; else x->skip_txfm = 0; if (cpi->common.tx_mode == TX_MODE_SELECT) { if (sse > (var << 2)) xd->mi[0]->mbmi.tx_size = MIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); else xd->mi[0]->mbmi.tx_size = TX_8X8; } else { xd->mi[0]->mbmi.tx_size = MIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); } vp9_model_rd_from_var_lapndz(sse - var, 1 << num_pels_log2_lookup[bsize], dc_quant >> 3, &rate, &dist); *out_rate_sum = rate >> 1; *out_dist_sum = dist << 3; vp9_model_rd_from_var_lapndz(var, 1 << num_pels_log2_lookup[bsize], ac_quant >> 3, &rate, &dist); *out_rate_sum += rate; *out_dist_sum += dist << 4; } static int get_pred_buffer(PRED_BUFFER *p, int len) { int i; for (i = 0; i < len; i++) {
281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350
if (!p[i].in_use) { p[i].in_use = 1; return i; } } return -1; } static void free_pred_buffer(PRED_BUFFER *p) { p->in_use = 0; } static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row, int mi_col, MV_REFERENCE_FRAME ref_frame, PREDICTION_MODE this_mode, unsigned int var_y, unsigned int sse_y, struct buf_2d yv12_mb[][MAX_MB_PLANE], int *rate, int64_t *dist) { MACROBLOCKD *xd = &x->e_mbd; MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi; const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]); unsigned int var = var_y, sse = sse_y; // Skipping threshold for ac. unsigned int thresh_ac; // Skipping threshold for dc. unsigned int thresh_dc; if (x->encode_breakout > 0) { // Set a maximum for threshold to avoid big PSNR loss in low bit rate // case. Use extreme low threshold for static frames to limit // skipping. const unsigned int max_thresh = 36000; // The encode_breakout input const unsigned int min_thresh = MIN(((unsigned int)x->encode_breakout << 4), max_thresh); // Calculate threshold according to dequant value. thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9; thresh_ac = clamp(thresh_ac, min_thresh, max_thresh); // Adjust ac threshold according to partition size. thresh_ac >>= 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]); thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6); } else { thresh_ac = 0; thresh_dc = 0; } // Y skipping condition checking for ac and dc. if (var <= thresh_ac && (sse - var) <= thresh_dc) { unsigned int sse_u, sse_v; unsigned int var_u, var_v; // Skip UV prediction unless breakout is zero (lossless) to save // computation with low impact on the result if (x->encode_breakout == 0) { xd->plane[1].pre[0] = yv12_mb[ref_frame][1]; xd->plane[2].pre[0] = yv12_mb[ref_frame][2]; vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize); } var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride, xd->plane[1].dst.buf, xd->plane[1].dst.stride, &sse_u); // U skipping condition checking