[FFmpeg-devel] [PATCH 01/14] avcodec/vc1: re-implement and expand VC-1 overlap smooting

[Jerome Borsboom]

The existing implementation did overlap smoothing for progressive
frames only. This rewritten version implements overlap smoothing
for all applicable frame types for both progessive and
frame/field-interlace.

Signed-off-by: Jerome Borsboom <jerome.borsboom at carpalis.nl>
---
This patch-set improves the VC-1 software decoder to the point where the fate
checksums are equal to checksums from the Intel hardware decoded image on Haswell.

 libavcodec/vc1.h            |  2 +
 libavcodec/vc1_loopfilter.c | 94 +++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 96 insertions(+)

diff --git a/libavcodec/vc1.h b/libavcodec/vc1.h
index 8fc0729cb8..85504c2f9f 100644
--- a/libavcodec/vc1.h
+++ b/libavcodec/vc1.h
@@ -425,6 +425,8 @@ void ff_vc1_decode_blocks(VC1Context *v);
 void ff_vc1_loop_filter_iblk(VC1Context *v, int pq);
 void ff_vc1_loop_filter_iblk_delayed(VC1Context *v, int pq);
 void ff_vc1_smooth_overlap_filter_iblk(VC1Context *v);
+void ff_vc1_i_overlap_filter(VC1Context *v);
+void ff_vc1_p_overlap_filter(VC1Context *v);
 void ff_vc1_apply_p_loop_filter(VC1Context *v);
 
 void ff_vc1_mc_1mv(VC1Context *v, int dir);
diff --git a/libavcodec/vc1_loopfilter.c b/libavcodec/vc1_loopfilter.c
index 025776bac9..3122b1a258 100644
--- a/libavcodec/vc1_loopfilter.c
+++ b/libavcodec/vc1_loopfilter.c
@@ -208,6 +208,100 @@ void ff_vc1_smooth_overlap_filter_iblk(VC1Context *v)
     }
 }
 
+static av_always_inline void vc1_h_overlap_filter(VC1Context *v, int16_t (*left_block)[64],
+                                                  int16_t (*right_block)[64], int block_num)
+{
+    if (left_block != right_block || (block_num & 5) == 1) {
+        if (block_num > 3)
+            v->vc1dsp.vc1_h_s_overlap(left_block[block_num], right_block[block_num]);
+        else if (block_num & 1)
+            v->vc1dsp.vc1_h_s_overlap(right_block[block_num - 1], right_block[block_num]);
+        else
+            v->vc1dsp.vc1_h_s_overlap(left_block[block_num + 1], right_block[block_num]);
+    }
+}
+
+static av_always_inline void vc1_v_overlap_filter(VC1Context *v, int16_t (*top_block)[64],
+                                                  int16_t (*bottom_block)[64], int block_num)
+{
+    if (top_block != bottom_block || block_num & 2) {
+        if (block_num > 3)
+            v->vc1dsp.vc1_v_s_overlap(top_block[block_num], bottom_block[block_num]);
+        else if (block_num & 2)
+            v->vc1dsp.vc1_v_s_overlap(bottom_block[block_num - 2], bottom_block[block_num]);
+        else
+            v->vc1dsp.vc1_v_s_overlap(top_block[block_num + 2], bottom_block[block_num]);
+    }
+}
+
+void ff_vc1_i_overlap_filter(VC1Context *v)
+{
+    MpegEncContext *s = &v->s;
+    int16_t (*topleft_blk)[64], (*top_blk)[64], (*left_blk)[64], (*cur_blk)[64];
+    int block_count = CONFIG_GRAY && (s->avctx->flags & AV_CODEC_FLAG_GRAY) ? 4 : 6;
+    int mb_pos = s->mb_x + s->mb_y * s->mb_stride;
+    int i;
+
+    topleft_blk = v->block[v->topleft_blk_idx];
+    top_blk = v->block[v->top_blk_idx];
+    left_blk = v->block[v->left_blk_idx];
+    cur_blk = v->block[v->cur_blk_idx];
+
+    /* Within a MB, the horizontal overlap always runs before the vertical.
+     * To accomplish that, we run the H on the left and internal vertical
+     * borders of the currently decoded MB. Then, we wait for the next overlap
+     * iteration to do H overlap on the right edge of this MB, before moving
+     * over and running the V overlap on the top and internal horizontal
+     * borders. Therefore, the H overlap trails by one MB col and the
+     * V overlap trails by one MB row. This is reflected in the time at which
+     * we run the put_pixels loop, i.e. delayed by one row and one column. */
+    for (i = 0; i < block_count; i++)
+        if (v->pq >= 9 || v->condover == CONDOVER_ALL ||
+            (v->over_flags_plane[mb_pos] && ((i & 5) == 1 || v->over_flags_plane[mb_pos - 1])))
+            vc1_h_overlap_filter(v, s->mb_x ? left_blk : cur_blk, cur_blk, i);
+
+    if (v->fcm != ILACE_FRAME)
+        for (i = 0; i < block_count; i++) {
+            if (s->mb_x && (v->pq >= 9 || v->condover == CONDOVER_ALL ||
+                (v->over_flags_plane[mb_pos - 1] &&
+                 ((i & 2) || v->over_flags_plane[mb_pos - 1 - s->mb_stride]))))
+                vc1_v_overlap_filter(v, s->first_slice_line ? left_blk : topleft_blk, left_blk, i);
+            if (s->mb_x == s->mb_width - 1)
+                if (v->pq >= 9 || v->condover == CONDOVER_ALL ||
+                    (v->over_flags_plane[mb_pos] &&
+                     ((i & 2) || v->over_flags_plane[mb_pos - s->mb_stride])))
+                    vc1_v_overlap_filter(v, s->first_slice_line ? cur_blk : top_blk, cur_blk, i);
+        }
+}
+
+void ff_vc1_p_overlap_filter(VC1Context *v)
+{
+    MpegEncContext *s = &v->s;
+    int16_t (*topleft_blk)[64], (*top_blk)[64], (*left_blk)[64], (*cur_blk)[64];
+    int block_count = CONFIG_GRAY && (s->avctx->flags & AV_CODEC_FLAG_GRAY) ? 4 : 6;
+    int i;
+
+    topleft_blk = v->block[v->topleft_blk_idx];
+    top_blk = v->block[v->top_blk_idx];
+    left_blk = v->block[v->left_blk_idx];
+    cur_blk = v->block[v->cur_blk_idx];
+
+    for (i = 0; i < block_count; i++)
+        if (v->mb_type[0][s->block_index[i]] && (s->mb_x == 0 || v->mb_type[0][s->block_index[i] - 1]))
+            vc1_h_overlap_filter(v, s->mb_x ? left_blk : cur_blk, cur_blk, i);
+
+    if (v->fcm != ILACE_FRAME)
+        for (i = 0; i < block_count; i++) {
+            if (s->mb_x && v->mb_type[0][s->block_index[i] - 1] &&
+                (s->first_slice_line || v->mb_type[0][s->block_index[i] - s->block_wrap[i] - 1]))
+                vc1_v_overlap_filter(v, s->first_slice_line ? left_blk : topleft_blk, left_blk, i);
+            if (s->mb_x == s->mb_width - 1)
+                if (v->mb_type[0][s->block_index[i]] &&
+                    (s->first_slice_line || v->mb_type[0][s->block_index[i] - s->block_wrap[i]]))
+                    vc1_v_overlap_filter(v, s->first_slice_line ? cur_blk : top_blk, cur_blk, i);
+        }
+}
+
 static av_always_inline void vc1_apply_p_v_loop_filter(VC1Context *v, int block_num)
 {
     MpegEncContext *s  = &v->s;
-- 
2.13.6