diff x86/vp8dsp-init.c @ 12204:563339ea87aa libavcodec

Chroma (width=8) inner loopfilter MMX/MMX2/SSE2 for VP8 decoder.

--- a/x86/vp8dsp-init.c	Tue Jul 20 21:54:46 2010 +0000
+++ b/x86/vp8dsp-init.c	Tue Jul 20 22:04:18 2010 +0000
@@ -242,6 +242,19 @@
                                                  int e, int i, int hvt);
 extern void ff_vp8_h_loop_filter16y_inner_sse2  (uint8_t *dst, int stride,
                                                  int e, int i, int hvt);
+
+extern void ff_vp8_v_loop_filter8uv_inner_mmx   (uint8_t *dstU, uint8_t *dstV,
+                                                 int s, int e, int i, int hvt);
+extern void ff_vp8_v_loop_filter8uv_inner_mmxext(uint8_t *dstU, uint8_t *dstV,
+                                                 int s, int e, int i, int hvt);
+extern void ff_vp8_v_loop_filter8uv_inner_sse2  (uint8_t *dstU, uint8_t *dstV,
+                                                 int s, int e, int i, int hvt);
+extern void ff_vp8_h_loop_filter8uv_inner_mmx   (uint8_t *dstU, uint8_t *dstV,
+                                                 int s, int e, int i, int hvt);
+extern void ff_vp8_h_loop_filter8uv_inner_mmxext(uint8_t *dstU, uint8_t *dstV,
+                                                 int s, int e, int i, int hvt);
+extern void ff_vp8_h_loop_filter8uv_inner_sse2  (uint8_t *dstU, uint8_t *dstV,
+                                                 int s, int e, int i, int hvt);
 #endif
 
 #define VP8_LUMA_MC_FUNC(IDX, SIZE, OPT) \
@@ -286,6 +299,8 @@
 
         c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_mmx;
         c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_mmx;
+        c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_mmx;
+        c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_mmx;
     }
 
     /* note that 4-tap width=16 functions are missing because w=16
@@ -304,6 +319,8 @@
 
         c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_mmxext;
         c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_mmxext;
+        c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_mmxext;
+        c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_mmxext;
     }
 
     if (mm_flags & FF_MM_SSE) {
@@ -321,10 +338,12 @@
         c->vp8_h_loop_filter_simple = ff_vp8_h_loop_filter_simple_sse2;
 
         c->vp8_v_loop_filter16y_inner = ff_vp8_v_loop_filter16y_inner_sse2;
+        c->vp8_v_loop_filter8uv_inner = ff_vp8_v_loop_filter8uv_inner_sse2;
     }
 
     if (mm_flags & FF_MM_SSE2) {
         c->vp8_h_loop_filter16y_inner = ff_vp8_h_loop_filter16y_inner_sse2;
+        c->vp8_h_loop_filter8uv_inner = ff_vp8_h_loop_filter8uv_inner_sse2;
     }
 
     if (mm_flags & FF_MM_SSSE3) {