src2 = vec_pack(s2, sA);
src3 = vec_pack(s3, sB);
+#if HAVE_BIGENDIAN
p0 = vec_lvsl (0, dest);
p1 = vec_lvsl (stride, dest);
p = vec_splat_u8 (-1);
perm0 = vec_mergeh (p, p0);
perm1 = vec_mergeh (p, p1);
+#define GET_TMP2(dst, p) \
+ tmp = vec_ld (0, dest); \
+ tmp2 = (vector signed short)vec_perm (tmp, vec_splat_u8(0), p);
+#else
+#define GET_TMP2(dst,p) \
+ tmp = vec_vsx_ld (0, dst); \
+ tmp2 = (vector signed short)vec_mergeh (tmp, vec_splat_u8(0));
+#endif
#define ADD(dest,src,perm) \
- /* *(uint64_t *)&tmp = *(uint64_t *)dest; */ \
- tmp = vec_ld (0, dest); \
- tmp2 = (vector signed short)vec_perm (tmp, vec_splat_u8(0), perm); \
+ GET_TMP2(dest, perm); \
tmp3 = vec_adds (tmp2, src); \
tmp = vec_packsu (tmp3, tmp3); \
vec_ste ((vector unsigned int)tmp, 0, (unsigned int *)dest); \