michael@0: /*
michael@0:  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS.  All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: #include <limits.h>
michael@0: 
michael@0: #include "vpx_mem/vpx_mem.h"
michael@0: #include "vp9/encoder/vp9_encodeintra.h"
michael@0: #include "vp9/encoder/vp9_rdopt.h"
michael@0: #include "vp9/encoder/vp9_segmentation.h"
michael@0: #include "vp9/encoder/vp9_mcomp.h"
michael@0: #include "vp9/common/vp9_blockd.h"
michael@0: #include "vp9/common/vp9_reconinter.h"
michael@0: #include "vp9/common/vp9_reconintra.h"
michael@0: #include "vp9/common/vp9_systemdependent.h"
michael@0: 
michael@0: 
michael@0: 
michael@0: static unsigned int do_16x16_motion_iteration(VP9_COMP *cpi,
michael@0:                                               int_mv *ref_mv,
michael@0:                                               int_mv *dst_mv,
michael@0:                                               int mb_row,
michael@0:                                               int mb_col) {
michael@0:   MACROBLOCK   *const x  = &cpi->mb;
michael@0:   MACROBLOCKD *const xd = &x->e_mbd;
michael@0:   vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
michael@0:   unsigned int best_err;
michael@0: 
michael@0:   const int tmp_col_min = x->mv_col_min;
michael@0:   const int tmp_col_max = x->mv_col_max;
michael@0:   const int tmp_row_min = x->mv_row_min;
michael@0:   const int tmp_row_max = x->mv_row_max;
michael@0:   int_mv ref_full;
michael@0: 
michael@0:   // Further step/diamond searches as necessary
michael@0:   int step_param = cpi->sf.reduce_first_step_size +
michael@0:       (cpi->speed < 8 ? (cpi->speed > 5 ? 1 : 0) : 2);
michael@0:   step_param = MIN(step_param, (cpi->sf.max_step_search_steps - 2));
michael@0: 
michael@0:   vp9_clamp_mv_min_max(x, &ref_mv->as_mv);
michael@0: 
michael@0:   ref_full.as_mv.col = ref_mv->as_mv.col >> 3;
michael@0:   ref_full.as_mv.row = ref_mv->as_mv.row >> 3;
michael@0: 
michael@0:   /*cpi->sf.search_method == HEX*/
michael@0:   best_err = vp9_hex_search(x, &ref_full.as_mv, step_param, x->errorperbit,
michael@0:                             0, &v_fn_ptr,
michael@0:                             0, &ref_mv->as_mv, &dst_mv->as_mv);
michael@0: 
michael@0:   // Try sub-pixel MC
michael@0:   // if (bestsme > error_thresh && bestsme < INT_MAX)
michael@0:   {
michael@0:     int distortion;
michael@0:     unsigned int sse;
michael@0:     best_err = cpi->find_fractional_mv_step(
michael@0:         x,
michael@0:         &dst_mv->as_mv, &ref_mv->as_mv,
michael@0:         cpi->common.allow_high_precision_mv,
michael@0:         x->errorperbit, &v_fn_ptr,
michael@0:         0, cpi->sf.subpel_iters_per_step, NULL, NULL,
michael@0:         & distortion, &sse);
michael@0:   }
michael@0: 
michael@0:   vp9_set_mbmode_and_mvs(x, NEWMV, dst_mv);
michael@0:   vp9_build_inter_predictors_sby(xd, mb_row, mb_col, BLOCK_16X16);
michael@0:   best_err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
michael@0:                           xd->plane[0].dst.buf, xd->plane[0].dst.stride,
michael@0:                           INT_MAX);
michael@0: 
michael@0:   /* restore UMV window */
michael@0:   x->mv_col_min = tmp_col_min;
michael@0:   x->mv_col_max = tmp_col_max;
michael@0:   x->mv_row_min = tmp_row_min;
michael@0:   x->mv_row_max = tmp_row_max;
michael@0: 
michael@0:   return best_err;
michael@0: }
michael@0: 
michael@0: static int do_16x16_motion_search(VP9_COMP *cpi, int_mv *ref_mv, int_mv *dst_mv,
michael@0:                                   int mb_row, int mb_col) {
michael@0:   MACROBLOCK *const x = &cpi->mb;
michael@0:   MACROBLOCKD *const xd = &x->e_mbd;
michael@0:   unsigned int err, tmp_err;
michael@0:   int_mv tmp_mv;
michael@0: 
michael@0:   // Try zero MV first
michael@0:   // FIXME should really use something like near/nearest MV and/or MV prediction
michael@0:   err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
michael@0:                      xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
michael@0:                      INT_MAX);
michael@0:   dst_mv->as_int = 0;
michael@0: 
michael@0:   // Test last reference frame using the previous best mv as the
michael@0:   // starting point (best reference) for the search
michael@0:   tmp_err = do_16x16_motion_iteration(cpi, ref_mv, &tmp_mv, mb_row, mb_col);
michael@0:   if (tmp_err < err) {
michael@0:     err = tmp_err;
michael@0:     dst_mv->as_int = tmp_mv.as_int;
michael@0:   }
michael@0: 
michael@0:   // If the current best reference mv is not centered on 0,0 then do a 0,0
michael@0:   // based search as well.
michael@0:   if (ref_mv->as_int) {
michael@0:     unsigned int tmp_err;
michael@0:     int_mv zero_ref_mv, tmp_mv;
michael@0: 
michael@0:     zero_ref_mv.as_int = 0;
michael@0:     tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, &tmp_mv,
michael@0:                                         mb_row, mb_col);
michael@0:     if (tmp_err < err) {
michael@0:       dst_mv->as_int = tmp_mv.as_int;
michael@0:       err = tmp_err;
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   return err;
michael@0: }
michael@0: 
michael@0: static int do_16x16_zerozero_search(VP9_COMP *cpi, int_mv *dst_mv) {
michael@0:   MACROBLOCK *const x = &cpi->mb;
michael@0:   MACROBLOCKD *const xd = &x->e_mbd;
michael@0:   unsigned int err;
michael@0: 
michael@0:   // Try zero MV first
michael@0:   // FIXME should really use something like near/nearest MV and/or MV prediction
michael@0:   err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
michael@0:                      xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
michael@0:                      INT_MAX);
michael@0: 
michael@0:   dst_mv->as_int = 0;
michael@0: 
michael@0:   return err;
michael@0: }
michael@0: static int find_best_16x16_intra(VP9_COMP *cpi,
michael@0:                                  int mb_y_offset,
michael@0:                                  MB_PREDICTION_MODE *pbest_mode) {
michael@0:   MACROBLOCK   *const x  = &cpi->mb;
michael@0:   MACROBLOCKD *const xd = &x->e_mbd;
michael@0:   MB_PREDICTION_MODE best_mode = -1, mode;
michael@0:   unsigned int best_err = INT_MAX;
michael@0: 
michael@0:   // calculate SATD for each intra prediction mode;
michael@0:   // we're intentionally not doing 4x4, we just want a rough estimate
michael@0:   for (mode = DC_PRED; mode <= TM_PRED; mode++) {
michael@0:     unsigned int err;
michael@0: 
michael@0:     xd->mi_8x8[0]->mbmi.mode = mode;
michael@0:     vp9_predict_intra_block(xd, 0, 2, TX_16X16, mode,
michael@0:                             x->plane[0].src.buf, x->plane[0].src.stride,
michael@0:                             xd->plane[0].dst.buf, xd->plane[0].dst.stride);
michael@0:     err = vp9_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
michael@0:                        xd->plane[0].dst.buf, xd->plane[0].dst.stride, best_err);
michael@0: 
michael@0:     // find best
michael@0:     if (err < best_err) {
michael@0:       best_err  = err;
michael@0:       best_mode = mode;
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   if (pbest_mode)
michael@0:     *pbest_mode = best_mode;
michael@0: 
michael@0:   return best_err;
michael@0: }
michael@0: 
michael@0: static void update_mbgraph_mb_stats
michael@0: (
michael@0:   VP9_COMP *cpi,
michael@0:   MBGRAPH_MB_STATS *stats,
michael@0:   YV12_BUFFER_CONFIG *buf,
michael@0:   int mb_y_offset,
michael@0:   YV12_BUFFER_CONFIG *golden_ref,
michael@0:   int_mv *prev_golden_ref_mv,
michael@0:   int gld_y_offset,
michael@0:   YV12_BUFFER_CONFIG *alt_ref,
michael@0:   int_mv *prev_alt_ref_mv,
michael@0:   int arf_y_offset,
michael@0:   int mb_row,
michael@0:   int mb_col
michael@0: ) {
michael@0:   MACROBLOCK *const x = &cpi->mb;
michael@0:   MACROBLOCKD *const xd = &x->e_mbd;
michael@0:   int intra_error;
michael@0:   VP9_COMMON *cm = &cpi->common;
michael@0: 
michael@0:   // FIXME in practice we're completely ignoring chroma here
michael@0:   x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
michael@0:   x->plane[0].src.stride = buf->y_stride;
michael@0: 
michael@0:   xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
michael@0:   xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
michael@0: 
michael@0:   // do intra 16x16 prediction
michael@0:   intra_error = find_best_16x16_intra(cpi, mb_y_offset,
michael@0:                                       &stats->ref[INTRA_FRAME].m.mode);
michael@0:   if (intra_error <= 0)
michael@0:     intra_error = 1;
michael@0:   stats->ref[INTRA_FRAME].err = intra_error;
michael@0: 
michael@0:   // Golden frame MV search, if it exists and is different than last frame
michael@0:   if (golden_ref) {
michael@0:     int g_motion_error;
michael@0:     xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
michael@0:     xd->plane[0].pre[0].stride = golden_ref->y_stride;
michael@0:     g_motion_error = do_16x16_motion_search(cpi,
michael@0:                                             prev_golden_ref_mv,
michael@0:                                             &stats->ref[GOLDEN_FRAME].m.mv,
michael@0:                                             mb_row, mb_col);
michael@0:     stats->ref[GOLDEN_FRAME].err = g_motion_error;
michael@0:   } else {
michael@0:     stats->ref[GOLDEN_FRAME].err = INT_MAX;
michael@0:     stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
michael@0:   }
michael@0: 
michael@0:   // Do an Alt-ref frame MV search, if it exists and is different than
michael@0:   // last/golden frame.
michael@0:   if (alt_ref) {
michael@0:     int a_motion_error;
michael@0:     xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
michael@0:     xd->plane[0].pre[0].stride = alt_ref->y_stride;
michael@0:     a_motion_error = do_16x16_zerozero_search(cpi,
michael@0:                                               &stats->ref[ALTREF_FRAME].m.mv);
michael@0: 
michael@0:     stats->ref[ALTREF_FRAME].err = a_motion_error;
michael@0:   } else {
michael@0:     stats->ref[ALTREF_FRAME].err = INT_MAX;
michael@0:     stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
michael@0:   }
michael@0: }
michael@0: 
michael@0: static void update_mbgraph_frame_stats(VP9_COMP *cpi,
michael@0:                                        MBGRAPH_FRAME_STATS *stats,
michael@0:                                        YV12_BUFFER_CONFIG *buf,
michael@0:                                        YV12_BUFFER_CONFIG *golden_ref,
michael@0:                                        YV12_BUFFER_CONFIG *alt_ref) {
michael@0:   MACROBLOCK *const x = &cpi->mb;
michael@0:   MACROBLOCKD *const xd = &x->e_mbd;
michael@0:   VP9_COMMON *const cm = &cpi->common;
michael@0: 
michael@0:   int mb_col, mb_row, offset = 0;
michael@0:   int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
michael@0:   int_mv arf_top_mv, gld_top_mv;
michael@0:   MODE_INFO mi_local = { { 0 } };
michael@0: 
michael@0:   // Set up limit values for motion vectors to prevent them extending outside
michael@0:   // the UMV borders.
michael@0:   arf_top_mv.as_int = 0;
michael@0:   gld_top_mv.as_int = 0;
michael@0:   x->mv_row_min     = -BORDER_MV_PIXELS_B16;
michael@0:   x->mv_row_max     = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
michael@0:   xd->up_available  = 0;
michael@0:   xd->plane[0].dst.stride  = buf->y_stride;
michael@0:   xd->plane[0].pre[0].stride  = buf->y_stride;
michael@0:   xd->plane[1].dst.stride = buf->uv_stride;
michael@0:   xd->mi_8x8[0] = &mi_local;
michael@0:   mi_local.mbmi.sb_type = BLOCK_16X16;
michael@0:   mi_local.mbmi.ref_frame[0] = LAST_FRAME;
michael@0:   mi_local.mbmi.ref_frame[1] = NONE;
michael@0: 
michael@0:   for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
michael@0:     int_mv arf_left_mv, gld_left_mv;
michael@0:     int mb_y_in_offset  = mb_y_offset;
michael@0:     int arf_y_in_offset = arf_y_offset;
michael@0:     int gld_y_in_offset = gld_y_offset;
michael@0: 
michael@0:     // Set up limit values for motion vectors to prevent them extending outside
michael@0:     // the UMV borders.
michael@0:     arf_left_mv.as_int = arf_top_mv.as_int;
michael@0:     gld_left_mv.as_int = gld_top_mv.as_int;
michael@0:     x->mv_col_min      = -BORDER_MV_PIXELS_B16;
michael@0:     x->mv_col_max      = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
michael@0:     xd->left_available = 0;
michael@0: 
michael@0:     for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
michael@0:       MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
michael@0: 
michael@0:       update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset,
michael@0:                               golden_ref, &gld_left_mv, gld_y_in_offset,
michael@0:                               alt_ref,    &arf_left_mv, arf_y_in_offset,
michael@0:                               mb_row, mb_col);
michael@0:       arf_left_mv.as_int = mb_stats->ref[ALTREF_FRAME].m.mv.as_int;
michael@0:       gld_left_mv.as_int = mb_stats->ref[GOLDEN_FRAME].m.mv.as_int;
michael@0:       if (mb_col == 0) {
michael@0:         arf_top_mv.as_int = arf_left_mv.as_int;
michael@0:         gld_top_mv.as_int = gld_left_mv.as_int;
michael@0:       }
michael@0:       xd->left_available = 1;
michael@0:       mb_y_in_offset    += 16;
michael@0:       gld_y_in_offset   += 16;
michael@0:       arf_y_in_offset   += 16;
michael@0:       x->mv_col_min     -= 16;
michael@0:       x->mv_col_max     -= 16;
michael@0:     }
michael@0:     xd->up_available = 1;
michael@0:     mb_y_offset     += buf->y_stride * 16;
michael@0:     gld_y_offset    += golden_ref->y_stride * 16;
michael@0:     if (alt_ref)
michael@0:       arf_y_offset    += alt_ref->y_stride * 16;
michael@0:     x->mv_row_min   -= 16;
michael@0:     x->mv_row_max   -= 16;
michael@0:     offset          += cm->mb_cols;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // void separate_arf_mbs_byzz
michael@0: static void separate_arf_mbs(VP9_COMP *cpi) {
michael@0:   VP9_COMMON *const cm = &cpi->common;
michael@0:   int mb_col, mb_row, offset, i;
michael@0:   int mi_row, mi_col;
michael@0:   int ncnt[4] = { 0 };
michael@0:   int n_frames = cpi->mbgraph_n_frames;
michael@0: 
michael@0:   int *arf_not_zz;
michael@0: 
michael@0:   CHECK_MEM_ERROR(cm, arf_not_zz,
michael@0:                   vpx_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz),
michael@0:                              1));
michael@0: 
michael@0:   // We are not interested in results beyond the alt ref itself.
michael@0:   if (n_frames > cpi->frames_till_gf_update_due)
michael@0:     n_frames = cpi->frames_till_gf_update_due;
michael@0: 
michael@0:   // defer cost to reference frames
michael@0:   for (i = n_frames - 1; i >= 0; i--) {
michael@0:     MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
michael@0: 
michael@0:     for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
michael@0:          offset += cm->mb_cols, mb_row++) {
michael@0:       for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
michael@0:         MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
michael@0: 
michael@0:         int altref_err = mb_stats->ref[ALTREF_FRAME].err;
michael@0:         int intra_err  = mb_stats->ref[INTRA_FRAME ].err;
michael@0:         int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
michael@0: 
michael@0:         // Test for altref vs intra and gf and that its mv was 0,0.
michael@0:         if (altref_err > 1000 ||
michael@0:             altref_err > intra_err ||
michael@0:             altref_err > golden_err) {
michael@0:           arf_not_zz[offset + mb_col]++;
michael@0:         }
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
michael@0:   // of bound access in segmentation_map
michael@0:   for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
michael@0:     for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
michael@0:       // If any of the blocks in the sequence failed then the MB
michael@0:       // goes in segment 0
michael@0:       if (arf_not_zz[mi_row/2*cm->mb_cols + mi_col/2]) {
michael@0:         ncnt[0]++;
michael@0:         cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
michael@0:       } else {
michael@0:         cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
michael@0:         ncnt[1]++;
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   // Only bother with segmentation if over 10% of the MBs in static segment
michael@0:   // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
michael@0:   if (1) {
michael@0:     // Note % of blocks that are marked as static
michael@0:     if (cm->MBs)
michael@0:       cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
michael@0: 
michael@0:     // This error case should not be reachable as this function should
michael@0:     // never be called with the common data structure uninitialized.
michael@0:     else
michael@0:       cpi->static_mb_pct = 0;
michael@0: 
michael@0:     cpi->seg0_cnt = ncnt[0];
michael@0:     vp9_enable_segmentation((VP9_PTR)cpi);
michael@0:   } else {
michael@0:     cpi->static_mb_pct = 0;
michael@0:     vp9_disable_segmentation((VP9_PTR)cpi);
michael@0:   }
michael@0: 
michael@0:   // Free localy allocated storage
michael@0:   vpx_free(arf_not_zz);
michael@0: }
michael@0: 
michael@0: void vp9_update_mbgraph_stats(VP9_COMP *cpi) {
michael@0:   VP9_COMMON *const cm = &cpi->common;
michael@0:   int i, n_frames = vp9_lookahead_depth(cpi->lookahead);
michael@0:   YV12_BUFFER_CONFIG *golden_ref =
michael@0:       &cm->yv12_fb[cm->ref_frame_map[cpi->gld_fb_idx]];
michael@0: 
michael@0:   // we need to look ahead beyond where the ARF transitions into
michael@0:   // being a GF - so exit if we don't look ahead beyond that
michael@0:   if (n_frames <= cpi->frames_till_gf_update_due)
michael@0:     return;
michael@0:   if (n_frames > (int)cpi->frames_till_alt_ref_frame)
michael@0:     n_frames = cpi->frames_till_alt_ref_frame;
michael@0:   if (n_frames > MAX_LAG_BUFFERS)
michael@0:     n_frames = MAX_LAG_BUFFERS;
michael@0: 
michael@0:   cpi->mbgraph_n_frames = n_frames;
michael@0:   for (i = 0; i < n_frames; i++) {
michael@0:     MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
michael@0:     vpx_memset(frame_stats->mb_stats, 0,
michael@0:                cm->mb_rows * cm->mb_cols *
michael@0:                sizeof(*cpi->mbgraph_stats[i].mb_stats));
michael@0:   }
michael@0: 
michael@0:   // do motion search to find contribution of each reference to data
michael@0:   // later on in this GF group
michael@0:   // FIXME really, the GF/last MC search should be done forward, and
michael@0:   // the ARF MC search backwards, to get optimal results for MV caching
michael@0:   for (i = 0; i < n_frames; i++) {
michael@0:     MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
michael@0:     struct lookahead_entry *q_cur = vp9_lookahead_peek(cpi->lookahead, i);
michael@0: 
michael@0:     assert(q_cur != NULL);
michael@0: 
michael@0:     update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img,
michael@0:                                golden_ref, cpi->Source);
michael@0:   }
michael@0: 
michael@0:   vp9_clear_system_state();  // __asm emms;
michael@0: 
michael@0:   separate_arf_mbs(cpi);
michael@0: }