From d642294b1c57a5adacb1038ff45766c38bae8a6d Mon Sep 17 00:00:00 2001 From: Jingning Han Date: Thu, 11 Feb 2016 12:36:49 -0800 Subject: [PATCH] Fix tsan error in VP9 sub8x8 intra mode search This commit fixes issue 1141. The issue was triggered in multi-tile encoding. The change properly saves and restores the block context information in the real-time mode selection process. It removes several redundant memcpy operations in sub8x8 intra block mode search. Change-Id: I35c9ad197f4bd500ec39b5fc833f052f19eee010 --- vp9/encoder/vp9_encodeframe.c | 22 ++++++++++++++++++++++ vp9/encoder/vp9_rdopt.c | 29 ++++++++++++++--------------- 2 files changed, 36 insertions(+), 15 deletions(-) diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c index 4159fe5e7..4109c19a9 100644 --- a/vp9/encoder/vp9_encodeframe.c +++ b/vp9/encoder/vp9_encodeframe.c @@ -3031,10 +3031,24 @@ static void nonrd_pick_sb_modes(VP9_COMP *cpi, TileInfo *const tile_info = &tile_data->tile_info; MACROBLOCKD *const xd = &x->e_mbd; MODE_INFO *mi; + ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE]; + BLOCK_SIZE bs = VPXMAX(bsize, BLOCK_8X8); // processing unit block size + const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bs]; + const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bs]; + int plane; + set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize); mi = xd->mi[0]; mi->sb_type = bsize; + for (plane = 0; plane < MAX_MB_PLANE; ++plane) { + struct macroblockd_plane *pd = &xd->plane[plane]; + memcpy(a + num_4x4_blocks_wide * plane, pd->above_context, + (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x); + memcpy(l + num_4x4_blocks_high * plane, pd->left_context, + (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y); + } + if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) if (cyclic_refresh_segment_id_boosted(mi->segment_id)) x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh); @@ -3052,6 +3066,14 @@ static void nonrd_pick_sb_modes(VP9_COMP *cpi, duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize); + for (plane = 0; plane < MAX_MB_PLANE; ++plane) { + struct macroblockd_plane *pd = &xd->plane[plane]; + memcpy(pd->above_context, a + num_4x4_blocks_wide * plane, + (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x); + memcpy(pd->left_context, l + num_4x4_blocks_high * plane, + (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y); + } + if (rd_cost->rate == INT_MAX) vp9_rd_cost_reset(rd_cost); diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c index 1480ea418..193c9d33c 100644 --- a/vp9/encoder/vp9_rdopt.c +++ b/vp9/encoder/vp9_rdopt.c @@ -787,9 +787,9 @@ static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, #if CONFIG_VP9_HIGHBITDEPTH uint16_t best_dst16[8 * 8]; #endif + memcpy(ta, a, num_4x4_blocks_wide * sizeof(a[0])); + memcpy(tl, l, num_4x4_blocks_high * sizeof(l[0])); - memcpy(ta, a, sizeof(ta)); - memcpy(tl, l, sizeof(tl)); xd->mi[0]->tx_size = TX_4X4; #if CONFIG_VP9_HIGHBITDEPTH @@ -810,8 +810,8 @@ static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, continue; } - memcpy(tempa, ta, sizeof(ta)); - memcpy(templ, tl, sizeof(tl)); + memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0])); + memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0])); for (idy = 0; idy < num_4x4_blocks_high; ++idy) { for (idx = 0; idx < num_4x4_blocks_wide; ++idx) { @@ -874,8 +874,8 @@ static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, *bestdistortion = distortion; best_rd = this_rd; *best_mode = mode; - memcpy(a, tempa, sizeof(tempa)); - memcpy(l, templ, sizeof(templ)); + memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0])); + memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0])); for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) { memcpy(best_dst16 + idy * 8, CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride), @@ -914,8 +914,8 @@ static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, continue; } - memcpy(tempa, ta, sizeof(ta)); - memcpy(templ, tl, sizeof(tl)); + memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0])); + memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0])); for (idy = 0; idy < num_4x4_blocks_high; ++idy) { for (idx = 0; idx < num_4x4_blocks_wide; ++idx) { @@ -976,8 +976,8 @@ static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, *bestdistortion = distortion; best_rd = this_rd; *best_mode = mode; - memcpy(a, tempa, sizeof(tempa)); - memcpy(l, templ, sizeof(templ)); + memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0])); + memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0])); for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) memcpy(best_dst + idy * 8, dst_init + idy * dst_stride, num_4x4_blocks_wide * 4); @@ -1013,12 +1013,8 @@ static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb, int64_t total_distortion = 0; int tot_rate_y = 0; int64_t total_rd = 0; - ENTROPY_CONTEXT t_above[4], t_left[4]; const int *bmode_costs = cpi->mbmode_cost; - memcpy(t_above, xd->plane[0].above_context, sizeof(t_above)); - memcpy(t_left, xd->plane[0].left_context, sizeof(t_left)); - // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block. for (idy = 0; idy < 2; idy += num_4x4_blocks_high) { for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) { @@ -1034,8 +1030,11 @@ static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb, } this_rd = rd_pick_intra4x4block(cpi, mb, idy, idx, &best_mode, - bmode_costs, t_above + idx, t_left + idy, + bmode_costs, + xd->plane[0].above_context + idx, + xd->plane[0].left_context + idy, &r, &ry, &d, bsize, best_rd - total_rd); + if (this_rd >= best_rd - total_rd) return INT64_MAX; -- 2.40.0