add sse instructions support for double on MAC OS

paddle/cuda/include/hl_matrix_type.cuh

@@ -27,6 +27,8 @@ typedef float4 vecType;
 typedef double2 vecType;
 #endif
 #else
+#include <mmintrin.h>
+#include <xmmintrin.h>
 #include <emmintrin.h>
 #ifndef HPPL_TYPE_DOUBLE
 typedef __m128  vecType;

paddle/cuda/include/hl_sse_matrix_kernel.cuh

@@ -25,6 +25,9 @@ limitations under the License. */
 #define     VECTOR_LEN      4
 #define     VECTOR_SET      _mm_set_ps1
 #else
+#if   defined(__APPLE__) || defined(__OSX__)
+#define     _mm_set_pd1     _mm_set1_pd
+#endif
 /* number of double in vector */
 #define     VECTOR_LEN      2
 #define     VECTOR_SET      _mm_set_pd1