Merge remote-tracking branch 'qatar/master'

* qatar/master: (22 commits)
  als: prevent infinite loop in zero_remaining().
  cook: prevent div-by-zero if channels is zero.
  pamenc: switch to encode2().
  svq1enc: switch to encode2().
  dvenc: switch to encode2().
  dpxenc: switch to encode2().
  pngenc: switch to encode2().
  v210enc: switch to encode2().
  xwdenc: switch to encode2().
  ttadec: use branchless unsigned-to-signed unfolding
  avcodec: add a Sun Rasterfile encoder
  sunrast: Move common defines to a new header file.
  cdxl: fix video decoding for some files
  cdxl: fix audio for some samples
  apetag: add proper support for binary tags
  ttadec: remove dead code
  swscale: make access to filter data conditional on filter type.
  swscale: update context offsets after removal of AlpMmxFilter.
  prores: initialise encoder and decoder parts only when needed
  swscale: make monowhite/black RGB-independent.
  ...

Conflicts:
	Changelog
	libavcodec/alsdec.c
	libavcodec/dpxenc.c
	libavcodec/golomb.h
	libavcodec/pamenc.c
	libavcodec/pngenc.c
	libavformat/img2.c
	libswscale/output.c

Merged-by: Michael Niedermayer <michaelni@gmx.at>

1 files changed, 36 insertions, 21 deletions

diff --git a/libswscale/output.c b/libswscale/output.c
index 75d0baad39..cae2c31805 100644
--- a/libswscale/output.c
+++ b/libswscale/output.c
@@ -298,6 +298,9 @@ static void yuv2nv12cX_c(SwsContext *c, const int16_t *chrFilter, int chrFilterS
         }
 }
 
+#define accumulate_bit(acc, val) \
+    acc <<= 1; \
+    acc |= (val) >= (128 + 110)
 #define output_pixel(pos, acc) \
     if (target == PIX_FMT_MONOBLACK) { \
         pos = acc; \
@@ -314,7 +317,6 @@ yuv2mono_X_c_template(SwsContext *c, const int16_t *lumFilter,
                       int y, enum PixelFormat target)
 {
     const uint8_t * const d128=dither_8x8_220[y&7];
-    uint8_t *g = c->table_gU[128 + YUVRGB_TABLE_HEADROOM] + c->table_gV[128 + YUVRGB_TABLE_HEADROOM];
     int i;
     unsigned acc = 0;
 
@@ -333,8 +335,8 @@ yuv2mono_X_c_template(SwsContext *c, const int16_t *lumFilter,
             Y1 = av_clip_uint8(Y1);
             Y2 = av_clip_uint8(Y2);
         }
-        acc += acc + g[Y1 + d128[(i + 0) & 7]];
-        acc += acc + g[Y2 + d128[(i + 1) & 7]];
+        accumulate_bit(acc, Y1 + d128[(i + 0) & 7]);
+        accumulate_bit(acc, Y2 + d128[(i + 1) & 7]);
         if ((i & 7) == 6) {
             output_pixel(*dest++, acc);
         }
@@ -350,19 +352,29 @@ yuv2mono_2_c_template(SwsContext *c, const int16_t *buf[2],
 {
     const int16_t *buf0  = buf[0],  *buf1  = buf[1];
     const uint8_t * const d128 = dither_8x8_220[y & 7];
-    uint8_t *g = c->table_gU[128 + YUVRGB_TABLE_HEADROOM] + c->table_gV[128 + YUVRGB_TABLE_HEADROOM];
     int  yalpha1 = 4095 - yalpha;
     int i;
 
     for (i = 0; i < dstW - 7; i += 8) {
-        int acc =    g[((buf0[i    ] * yalpha1 + buf1[i    ] * yalpha) >> 19) + d128[0]];
-        acc += acc + g[((buf0[i + 1] * yalpha1 + buf1[i + 1] * yalpha) >> 19) + d128[1]];
-        acc += acc + g[((buf0[i + 2] * yalpha1 + buf1[i + 2] * yalpha) >> 19) + d128[2]];
-        acc += acc + g[((buf0[i + 3] * yalpha1 + buf1[i + 3] * yalpha) >> 19) + d128[3]];
-        acc += acc + g[((buf0[i + 4] * yalpha1 + buf1[i + 4] * yalpha) >> 19) + d128[4]];
-        acc += acc + g[((buf0[i + 5] * yalpha1 + buf1[i + 5] * yalpha) >> 19) + d128[5]];
-        acc += acc + g[((buf0[i + 6] * yalpha1 + buf1[i + 6] * yalpha) >> 19) + d128[6]];
-        acc += acc + g[((buf0[i + 7] * yalpha1 + buf1[i + 7] * yalpha) >> 19) + d128[7]];
+        int Y, acc = 0;
+
+        Y = (buf0[i + 0] * yalpha1 + buf1[i + 0] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[0]);
+        Y = (buf0[i + 1] * yalpha1 + buf1[i + 1] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[1]);
+        Y = (buf0[i + 2] * yalpha1 + buf1[i + 2] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[2]);
+        Y = (buf0[i + 3] * yalpha1 + buf1[i + 3] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[3]);
+        Y = (buf0[i + 4] * yalpha1 + buf1[i + 4] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[4]);
+        Y = (buf0[i + 5] * yalpha1 + buf1[i + 5] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[5]);
+        Y = (buf0[i + 6] * yalpha1 + buf1[i + 6] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[6]);
+        Y = (buf0[i + 7] * yalpha1 + buf1[i + 7] * yalpha) >> 19;
+        accumulate_bit(acc, Y + d128[7]);
+
         output_pixel(*dest++, acc);
     }
 }
@@ -374,23 +386,26 @@ yuv2mono_1_c_template(SwsContext *c, const int16_t *buf0,
                       int uvalpha, int y, enum PixelFormat target)
 {
     const uint8_t * const d128 = dither_8x8_220[y & 7];
-    uint8_t *g = c->table_gU[128 + YUVRGB_TABLE_HEADROOM] + c->table_gV[128 + YUVRGB_TABLE_HEADROOM];
     int i;
 
     for (i = 0; i < dstW - 7; i += 8) {
-        int acc =    g[((buf0[i    ] + 64) >> 7) + d128[0]];
-        acc += acc + g[((buf0[i + 1] + 64) >> 7) + d128[1]];
-        acc += acc + g[((buf0[i + 2] + 64) >> 7) + d128[2]];
-        acc += acc + g[((buf0[i + 3] + 64) >> 7) + d128[3]];
-        acc += acc + g[((buf0[i + 4] + 64) >> 7) + d128[4]];
-        acc += acc + g[((buf0[i + 5] + 64) >> 7) + d128[5]];
-        acc += acc + g[((buf0[i + 6] + 64) >> 7) + d128[6]];
-        acc += acc + g[((buf0[i + 7] + 64) >> 7) + d128[7]];
+        int acc = 0;
+
+        accumulate_bit(acc, ((buf0[i + 0] + 64) >> 7) + d128[0]);
+        accumulate_bit(acc, ((buf0[i + 1] + 64) >> 7) + d128[1]);
+        accumulate_bit(acc, ((buf0[i + 2] + 64) >> 7) + d128[2]);
+        accumulate_bit(acc, ((buf0[i + 3] + 64) >> 7) + d128[3]);
+        accumulate_bit(acc, ((buf0[i + 4] + 64) >> 7) + d128[4]);
+        accumulate_bit(acc, ((buf0[i + 5] + 64) >> 7) + d128[5]);
+        accumulate_bit(acc, ((buf0[i + 6] + 64) >> 7) + d128[6]);
+        accumulate_bit(acc, ((buf0[i + 7] + 64) >> 7) + d128[7]);
+
         output_pixel(*dest++, acc);
     }
 }
 
 #undef output_pixel
+#undef accumulate_bit
 
 #define YUV2PACKEDWRAPPER(name, base, ext, fmt) \
 static void name ## ext ## _X_c(SwsContext *c, const int16_t *lumFilter, \