[FFmpeg-cvslog] Bink: clip AC coefficients during dequantization.

Reimar Döffinger git at videolan.org
Wed Sep 7 15:16:22 CEST 2011


ffmpeg | branch: release/0.7 | Reimar Döffinger <Reimar.Doeffinger at gmx.de> | Sun Jul 17 13:03:57 2011 +0200| [a0352d01e95ed084e777d0c259bad1f4cee4f402] | committer: Carl Eugen Hoyos

Bink: clip AC coefficients during dequantization.

Fixes artefacts with Neverwinter Nights WOTCLogo.bik
(http://drmccoy.de/zeugs/WOTCLogo.bik).
Fixes trac ticket #352.

Signed-off-by: Reimar Döffinger <Reimar.Doeffinger at gmx.de>
(cherry picked from commit 47b71eea099b3fe2c7e16644878ad9b7067974e3)

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=a0352d01e95ed084e777d0c259bad1f4cee4f402
---

 libavcodec/bink.c |   20 ++++++++++++++++++--
 1 files changed, 18 insertions(+), 2 deletions(-)

diff --git a/libavcodec/bink.c b/libavcodec/bink.c
index ef07747..4328a43 100644
--- a/libavcodec/bink.c
+++ b/libavcodec/bink.c
@@ -571,6 +571,22 @@ static inline int binkb_get_value(BinkContext *c, int bundle_num)
     return ret;
 }
 
+static inline DCTELEM dequant(DCTELEM in, uint32_t quant, int dc)
+{
+    /* Note: multiplication is unsigned but we want signed shift
+     * otherwise clipping breaks.
+     * TODO: The official decoder does not use clipping at all
+     * but instead uses the full 32-bit result.
+     * However clipping at least gets rid of the case that a
+     * half-black half-white intra block gets black and white swapped
+     * and should cause at most minor differences (except for DC). */
+    int32_t res = in * quant;
+    res >>= 11;
+    if (!dc)
+        res = av_clip_int16(res);
+    return res;
+}
+
 /**
  * Read 8x8 block of DCT coefficients.
  *
@@ -669,10 +685,10 @@ static int read_dct_coeffs(GetBitContext *gb, DCTELEM block[64], const uint8_t *
 
     quant = quant_matrices[quant_idx];
 
-    block[0] = (block[0] * quant[0]) >> 11;
+    block[0] = dequant(block[0], quant[0], 1);
     for (i = 0; i < coef_count; i++) {
         int idx = coef_idx[i];
-        block[scan[idx]] = (block[scan[idx]] * quant[idx]) >> 11;
+        block[scan[idx]] = dequant(block[scan[idx]], quant[idx], 0);
     }
 
     return 0;



More information about the ffmpeg-cvslog mailing list