Add AVX2 SIMD implementation for cdef_dir

cdef_dir_8bpc_c: 629.3
cdef_dir_8bpc_avx2: 82.4

First 1000 frames of Chimera 1080p:
before: 0m23.084s
after: 0m21.860s

1 files changed, 7 insertions, 2 deletions

diff --git a/src/cdef.h b/src/cdef.h
index 9573962..642d125 100644
--- a/src/cdef.h
+++ b/src/cdef.h
@@ -48,8 +48,10 @@ typedef void (*cdef_fn)(pixel *dst, ptrdiff_t stride,
                         /*const*/ pixel *const top[2],
                         int pri_strength, int sec_strength,
                         int dir, int damping, enum CdefEdgeFlags edges);
-typedef int (*cdef_dir_fn)(const pixel *dst, ptrdiff_t stride,
-                           unsigned *var);
+
+#define decl_cdef_dir_fn(name) \
+int (name)(const pixel *dst, ptrdiff_t dst_stride, unsigned *var)
+typedef decl_cdef_dir_fn(*cdef_dir_fn);
 
 typedef struct Dav1dCdefDSPContext {
     cdef_dir_fn dir;
@@ -59,4 +61,7 @@ typedef struct Dav1dCdefDSPContext {
 void dav1d_cdef_dsp_init_8bpc(Dav1dCdefDSPContext *c);
 void dav1d_cdef_dsp_init_10bpc(Dav1dCdefDSPContext *c);
 
+void dav1d_cdef_dsp_init_x86_8bpc(Dav1dCdefDSPContext *c);
+void dav1d_cdef_dsp_init_x86_10bpc(Dav1dCdefDSPContext *c);
+
 #endif /* __DAV1D_SRC_CDEF_H__ */