gst/deinterlace2/tvtime/greedy.c: Fix the C implementation to produce correct results and optimize the

Original commit message from CVS:
* gst/deinterlace2/tvtime/greedy.c:
(deinterlace_greedy_packed422_scanline_c),
(deinterlace_greedy_packed422_scanline_mmxext),
(deinterlace_greedy_packed422_scanline):
Fix the C implementation to produce correct results and optimize the
MMXEXT implementation.
Handle odd widths and don't read over array boundaries in the MMXEXT
implementation.
* gst/deinterlace2/tvtime/vfir.c: (deinterlace_line_c),
(deinterlace_line_mmx), (deinterlace_scanline_vfir):
Fix a small rounding bug in the MMX implementation, the MMX
implementation doesn't actually need MMXEXT instructions so don't mark
it as such.
Handle odd widths in both implementations.

1 files changed, 41 insertions, 36 deletions

diff --git a/gst/deinterlace2/tvtime/vfir.c b/gst/deinterlace2/tvtime/vfir.c
index f32be654..479ee440 100644
--- a/gst/deinterlace2/tvtime/vfir.c
+++ b/gst/deinterlace2/tvtime/vfir.c
@@ -49,10 +49,36 @@
  * filter taps here are: [-1 4 2 4 -1].
  */
 
+/**
+  * C implementation.
+  */
+static inline void
+deinterlace_line_c (uint8_t * dst, uint8_t * lum_m4,
+    uint8_t * lum_m3, uint8_t * lum_m2,
+    uint8_t * lum_m1, uint8_t * lum, int size)
+{
+  int sum;
+
+  for (; size > 0; size--) {
+    sum = -lum_m4[0];
+    sum += lum_m3[0] << 2;
+    sum += lum_m2[0] << 1;
+    sum += lum_m1[0] << 2;
+    sum += -lum[0];
+    dst[0] = (sum + 4) >> 3;    // This needs to be clipped at 0 and 255: cm[(sum + 4) >> 3];
+    lum_m4++;
+    lum_m3++;
+    lum_m2++;
+    lum_m1++;
+    lum++;
+    dst++;
+  }
+}
+
 #ifdef HAVE_CPU_I386
 #include "mmx.h"
 static void
-deinterlace_line_mmxext (uint8_t * dst, uint8_t * lum_m4,
+deinterlace_line_mmx (uint8_t * dst, uint8_t * lum_m4,
     uint8_t * lum_m3, uint8_t * lum_m2,
     uint8_t * lum_m1, uint8_t * lum, int size)
 {
@@ -63,14 +89,15 @@ deinterlace_line_mmxext (uint8_t * dst, uint8_t * lum_m4,
   rounder.uw[2] = 4;
   rounder.uw[3] = 4;
   pxor_r2r (mm7, mm7);
-  movq_m2r (rounder, mm6);
+  movd_m2r (rounder, mm6);
+  punpcklbw_r2r (mm7, mm6);
 
   for (; size > 3; size -= 4) {
-    movd_m2r (lum_m4[0], mm0);
-    movd_m2r (lum_m3[0], mm1);
-    movd_m2r (lum_m2[0], mm2);
-    movd_m2r (lum_m1[0], mm3);
-    movd_m2r (lum[0], mm4);
+    movd_m2r (*lum_m4, mm0);
+    movd_m2r (*lum_m3, mm1);
+    movd_m2r (*lum_m2, mm2);
+    movd_m2r (*lum_m1, mm3);
+    movd_m2r (*lum, mm4);
     punpcklbw_r2r (mm7, mm0);
     punpcklbw_r2r (mm7, mm1);
     punpcklbw_r2r (mm7, mm2);
@@ -85,7 +112,7 @@ deinterlace_line_mmxext (uint8_t * dst, uint8_t * lum_m4,
     psubusw_r2r (mm0, mm1);
     psrlw_i2r (3, mm1);         // 3
     packuswb_r2r (mm7, mm1);
-    movd_r2m (mm1, dst[0]);
+    movd_r2m (mm1, *dst);
     lum_m4 += 4;
     lum_m3 += 4;
     lum_m2 += 4;
@@ -94,34 +121,12 @@ deinterlace_line_mmxext (uint8_t * dst, uint8_t * lum_m4,
     dst += 4;
   }
   emms ();
-}
-#endif
 
-/**
-  * C implementation.
-  */
-static void
-deinterlace_line_c (uint8_t * dst, uint8_t * lum_m4,
-    uint8_t * lum_m3, uint8_t * lum_m2,
-    uint8_t * lum_m1, uint8_t * lum, int size)
-{
-  int sum;
-
-  for (; size > 0; size--) {
-    sum = -lum_m4[0];
-    sum += lum_m3[0] << 2;
-    sum += lum_m2[0] << 1;
-    sum += lum_m1[0] << 2;
-    sum += -lum[0];
-    dst[0] = (sum + 4) >> 3;    // This needs to be clipped at 0 and 255: cm[(sum + 4) >> 3];
-    lum_m4++;
-    lum_m3++;
-    lum_m2++;
-    lum_m1++;
-    lum++;
-    dst++;
-  }
+  /* Handle odd widths */
+  if (size > 0)
+    deinterlace_line_c (dst, lum_m4, lum_m3, lum_m2, lum_m1, lum, size);
 }
+#endif
 
 /*
  * The commented-out method below that uses the bottom_field member is more
@@ -134,8 +139,8 @@ deinterlace_scanline_vfir (GstDeinterlace2 * object,
     deinterlace_scanline_data_t * data, uint8_t * output)
 {
 #ifdef HAVE_CPU_I386
-  if (object->cpu_feature_flags & OIL_IMPL_FLAG_MMXEXT) {
-    deinterlace_line_mmxext (output, data->tt1, data->t0, data->m1, data->b0,
+  if (object->cpu_feature_flags & OIL_IMPL_FLAG_MMX) {
+    deinterlace_line_mmx (output, data->tt1, data->t0, data->m1, data->b0,
         data->bb1, object->frame_width * 2);
   } else {
     deinterlace_line_c (output, data->tt1, data->t0, data->m1, data->b0,