diff options
Diffstat (limited to 'src/vectors-8-SSE2.h')
-rw-r--r-- | src/vectors-8-SSE2.h | 14 |
1 files changed, 13 insertions, 1 deletions
diff --git a/src/vectors-8-SSE2.h b/src/vectors-8-SSE2.h index bce093c..46a855e 100644 --- a/src/vectors-8-SSE2.h +++ b/src/vectors-8-SSE2.h @@ -8,7 +8,14 @@ #include <assert.h> #include <math.h> -#include <x86intrin.h> +#include <emmintrin.h> +#ifdef __SSE4_1__ +// Intel's SSE 4.1 +# include <smmintrin.h> +#endif +#ifdef __SSE4A__ +// AMD's SSE 4a +# include <ammintrin.h> // Intel compilers don't support SSE 4a. Here is how we can implement // these instructions in assembler instead: @@ -19,6 +26,11 @@ // asm ("movntsd %[x],%[p]" : "=m" (*p) : [p] "m" (*p), [x] "x" (x)); // } +#endif +#ifdef __FMA4__ +# include <fma4intrin.h> +#endif + #ifdef __SSE4_1__ |