av1/encoder/mbgraph.c

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <limits.h>

#include "config/av1_rtcd.h"
#include "config/aom_dsp_rtcd.h"

#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/system_state.h"
#include "av1/encoder/segmentation.h"
#include "av1/encoder/mcomp.h"
#include "av1/common/blockd.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"

static unsigned int do_16x16_motion_iteration(AV1_COMP *cpi, const MV *ref_mv,
                                              int mb_row, int mb_col) {
  MACROBLOCK *const x = &cpi->td.mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
  const aom_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];

  const MvLimits tmp_mv_limits = x->mv_limits;
  MV ref_full;
  int cost_list[5];

  // Further step/diamond searches as necessary
  int step_param = mv_sf->reduce_first_step_size;
  step_param = AOMMIN(step_param, MAX_MVSEARCH_STEPS - 2);

  av1_set_mv_search_range(&x->mv_limits, ref_mv);

  ref_full.col = ref_mv->col >> 3;
  ref_full.row = ref_mv->row >> 3;

  /*cpi->sf.search_method == HEX*/
  av1_hex_search(x, &ref_full, step_param, x->errorperbit, 0,
                 cond_cost_list(cpi, cost_list), &v_fn_ptr, 0, ref_mv);

  // Try sub-pixel MC
  // if (bestsme > error_thresh && bestsme < INT_MAX)
  if (cpi->common.cur_frame_force_integer_mv == 1) {
    x->best_mv.as_mv.row *= 8;
    x->best_mv.as_mv.col *= 8;
  } else {
    int distortion;
    unsigned int sse;
    cpi->find_fractional_mv_step(
        x, &cpi->common, mb_row, mb_col, ref_mv,
        cpi->common.allow_high_precision_mv, x->errorperbit, &v_fn_ptr, 0,
        mv_sf->subpel_iters_per_step, cond_cost_list(cpi, cost_list), NULL,
        NULL, &distortion, &sse, NULL, NULL, 0, 0, 0, 0, 0);
  }

  if (has_second_ref(xd->mi[0]))
    xd->mi[0]->mode = NEW_NEWMV;
  else
    xd->mi[0]->mode = NEWMV;

  xd->mi[0]->mv[0] = x->best_mv;
  xd->mi[0]->ref_frame[1] = NONE_FRAME;

  av1_build_inter_predictors_sby(&cpi->common, xd, mb_row, mb_col, NULL,
                                 BLOCK_16X16);

  /* restore UMV window */
  x->mv_limits = tmp_mv_limits;

  return aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
                      xd->plane[0].dst.buf, xd->plane[0].dst.stride);
}

static int do_16x16_motion_search(AV1_COMP *cpi, const MV *ref_mv, int mb_row,
                                  int mb_col) {
  MACROBLOCK *const x = &cpi->td.mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  unsigned int err, tmp_err;
  MV best_mv;

  // Try zero MV first
  // FIXME should really use something like near/nearest MV and/or MV prediction
  err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
  best_mv.col = best_mv.row = 0;

  // Test last reference frame using the previous best mv as the
  // starting point (best reference) for the search
  tmp_err = do_16x16_motion_iteration(cpi, ref_mv, mb_row, mb_col);
  if (tmp_err < err) {
    err = tmp_err;
    best_mv = x->best_mv.as_mv;
  }

  // If the current best reference mv is not centered on 0,0 then do a 0,0
  // based search as well.
  if (ref_mv->row != 0 || ref_mv->col != 0) {
    MV zero_ref_mv = kZeroMv;

    tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, mb_row, mb_col);
    if (tmp_err < err) {
      err = tmp_err;
      best_mv = x->best_mv.as_mv;
    }
  }

  x->best_mv.as_mv = best_mv;
  return err;
}

static int do_16x16_zerozero_search(AV1_COMP *cpi, int_mv *dst_mv) {
  MACROBLOCK *const x = &cpi->td.mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  unsigned int err;

  // Try zero MV first
  // FIXME should really use something like near/nearest MV and/or MV prediction
  err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);

  dst_mv->as_int = 0;

  return err;
}
static int find_best_16x16_intra(AV1_COMP *cpi, PREDICTION_MODE *pbest_mode) {
  const AV1_COMMON *cm = &cpi->common;
  MACROBLOCK *const x = &cpi->td.mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  PREDICTION_MODE best_mode = -1, mode;
  unsigned int best_err = INT_MAX;

  // calculate SATD for each intra prediction mode;
  // we're intentionally not doing 4x4, we just want a rough estimate
  for (mode = DC_PRED; mode <= PAETH_PRED; mode++) {
    unsigned int err;

    xd->mi[0]->mode = mode;
    av1_predict_intra_block(cm, xd, 16, 16, TX_16X16, mode, 0, 0,
                            FILTER_INTRA_MODES, x->plane[0].src.buf,
                            x->plane[0].src.stride, xd->plane[0].dst.buf,
                            xd->plane[0].dst.stride, 0, 0, 0);
    err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
                       xd->plane[0].dst.buf, xd->plane[0].dst.stride);

    // find best
    if (err < best_err) {
      best_err = err;
      best_mode = mode;
    }
  }

  if (pbest_mode) *pbest_mode = best_mode;

  return best_err;
}

static void update_mbgraph_mb_stats(AV1_COMP *cpi, MBGRAPH_MB_STATS *stats,
                                    YV12_BUFFER_CONFIG *buf, int mb_y_offset,
                                    YV12_BUFFER_CONFIG *golden_ref,
                                    const MV *prev_golden_ref_mv,
                                    YV12_BUFFER_CONFIG *alt_ref, int mb_row,
                                    int mb_col) {
  MACROBLOCK *const x = &cpi->td.mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int intra_error;
  AV1_COMMON *cm = &cpi->common;

  // FIXME in practice we're completely ignoring chroma here
  x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
  x->plane[0].src.stride = buf->y_stride;

  xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
  xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;

  // do intra 16x16 prediction
  intra_error = find_best_16x16_intra(cpi, &stats->ref[INTRA_FRAME].m.mode);
  if (intra_error <= 0) intra_error = 1;
  stats->ref[INTRA_FRAME].err = intra_error;

  // Golden frame MV search, if it exists and is different than last frame
  if (golden_ref) {
    int g_motion_error;
    xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
    xd->plane[0].pre[0].stride = golden_ref->y_stride;
    g_motion_error =
        do_16x16_motion_search(cpi, prev_golden_ref_mv, mb_row, mb_col);
    stats->ref[GOLDEN_FRAME].m.mv = x->best_mv;
    stats->ref[GOLDEN_FRAME].err = g_motion_error;
  } else {
    stats->ref[GOLDEN_FRAME].err = INT_MAX;
    stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
  }

  // Do an Alt-ref frame MV search, if it exists and is different than
  // last/golden frame.
  if (alt_ref) {
    int a_motion_error;
    xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
    xd->plane[0].pre[0].stride = alt_ref->y_stride;
    a_motion_error =
        do_16x16_zerozero_search(cpi, &stats->ref[ALTREF_FRAME].m.mv);

    stats->ref[ALTREF_FRAME].err = a_motion_error;
  } else {
    stats->ref[ALTREF_FRAME].err = INT_MAX;
    stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
  }
}

static void update_mbgraph_frame_stats(AV1_COMP *cpi,
                                       MBGRAPH_FRAME_STATS *stats,
                                       YV12_BUFFER_CONFIG *buf,
                                       YV12_BUFFER_CONFIG *golden_ref,
                                       YV12_BUFFER_CONFIG *alt_ref) {
  MACROBLOCK *const x = &cpi->td.mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  AV1_COMMON *const cm = &cpi->common;

  int mb_col, mb_row, offset = 0;
  int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
  MV gld_top_mv = kZeroMv;
  MB_MODE_INFO mi_local;

  av1_zero(mi_local);
  // Set up limit values for motion vectors to prevent them extending outside
  // the UMV borders.
  x->mv_limits.row_min = -BORDER_MV_PIXELS_B16;
  x->mv_limits.row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
  xd->up_available = 0;
  xd->plane[0].dst.stride = buf->y_stride;
  xd->plane[0].pre[0].stride = buf->y_stride;
  xd->plane[1].dst.stride = buf->uv_stride;
  xd->mi[0] = &mi_local;
  mi_local.sb_type = BLOCK_16X16;
  mi_local.ref_frame[0] = LAST_FRAME;
  mi_local.ref_frame[1] = NONE_FRAME;

  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
    MV gld_left_mv = gld_top_mv;
    int mb_y_in_offset = mb_y_offset;
    int arf_y_in_offset = arf_y_offset;
    int gld_y_in_offset = gld_y_offset;

    // Set up limit values for motion vectors to prevent them extending outside
    // the UMV borders.
    x->mv_limits.col_min = -BORDER_MV_PIXELS_B16;
    x->mv_limits.col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
    xd->left_available = 0;

    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
      MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];

      update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset, golden_ref,
                              &gld_left_mv, alt_ref, mb_row, mb_col);
      gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
      if (mb_col == 0) {
        gld_top_mv = gld_left_mv;
      }
      xd->left_available = 1;
      mb_y_in_offset += 16;
      gld_y_in_offset += 16;
      arf_y_in_offset += 16;
      x->mv_limits.col_min -= 16;
      x->mv_limits.col_max -= 16;
    }
    xd->up_available = 1;
    mb_y_offset += buf->y_stride * 16;
    gld_y_offset += golden_ref->y_stride * 16;
    if (alt_ref) arf_y_offset += alt_ref->y_stride * 16;
    x->mv_limits.row_min -= 16;
    x->mv_limits.row_max -= 16;
    offset += cm->mb_cols;
  }
}

// void separate_arf_mbs_byzz
static void separate_arf_mbs(AV1_COMP *cpi) {
  AV1_COMMON *const cm = &cpi->common;
  int mb_col, mb_row, offset, i;
  int mi_row, mi_col;
  int ncnt[4] = { 0 };
  int n_frames = cpi->mbgraph_n_frames;

  int *arf_not_zz;

  CHECK_MEM_ERROR(
      cm, arf_not_zz,
      aom_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz), 1));

  // We are not interested in results beyond the alt ref itself.
  if (n_frames > cpi->rc.frames_till_gf_update_due)
    n_frames = cpi->rc.frames_till_gf_update_due;

  // defer cost to reference frames
  for (i = n_frames - 1; i >= 0; i--) {
    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];

    for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
         offset += cm->mb_cols, mb_row++) {
      for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
        MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];

        int altref_err = mb_stats->ref[ALTREF_FRAME].err;
        int intra_err = mb_stats->ref[INTRA_FRAME].err;
        int golden_err = mb_stats->ref[GOLDEN_FRAME].err;

        // Test for altref vs intra and gf and that its mv was 0,0.
        if (altref_err > 1000 || altref_err > intra_err ||
            altref_err > golden_err) {
          arf_not_zz[offset + mb_col]++;
        }
      }
    }
  }

  // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
  // of bound access in segmentation_map
  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
      // If any of the blocks in the sequence failed then the MB
      // goes in segment 0
      if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
        ncnt[0]++;
        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
      } else {
        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
        ncnt[1]++;
      }
    }
  }

  // Only bother with segmentation if over 10% of the MBs in static segment
  // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
  if (1) {
    // Note % of blocks that are marked as static
    if (cm->MBs)
      cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);

    // This error case should not be reachable as this function should
    // never be called with the common data structure uninitialized.
    else
      cpi->static_mb_pct = 0;

    av1_enable_segmentation(&cm->seg);
  } else {
    cpi->static_mb_pct = 0;
    av1_disable_segmentation(&cm->seg);
  }

  // Free localy allocated storage
  aom_free(arf_not_zz);
}

void av1_update_mbgraph_stats(AV1_COMP *cpi) {
  AV1_COMMON *const cm = &cpi->common;
  int i, n_frames = av1_lookahead_depth(cpi->lookahead);
  YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);

  assert(golden_ref != NULL);

  // we need to look ahead beyond where the ARF transitions into
  // being a GF - so exit if we don't look ahead beyond that
  if (n_frames <= cpi->rc.frames_till_gf_update_due) return;

  if (n_frames > MAX_LAG_BUFFERS) n_frames = MAX_LAG_BUFFERS;

  cpi->mbgraph_n_frames = n_frames;
  for (i = 0; i < n_frames; i++) {
    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
    memset(frame_stats->mb_stats, 0,
           cm->mb_rows * cm->mb_cols * sizeof(*cpi->mbgraph_stats[i].mb_stats));
  }

  // do motion search to find contribution of each reference to data
  // later on in this GF group
  // FIXME really, the GF/last MC search should be done forward, and
  // the ARF MC search backwards, to get optimal results for MV caching
  for (i = 0; i < n_frames; i++) {
    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
    struct lookahead_entry *q_cur = av1_lookahead_peek(cpi->lookahead, i);

    assert(q_cur != NULL);

    update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img, golden_ref,
                               cpi->source);
  }

  aom_clear_system_state();

  separate_arf_mbs(cpi);
}