From 8a379fce1651640f36bbc979653c62aeedcc1c9c Mon Sep 17 00:00:00 2001
From: liaogang <liaogang@baidu.com>
Date: Wed, 14 Sep 2016 14:52:33 +0800
Subject: [PATCH] add sse instructions support for double on MAC OS

---
 paddle/cuda/include/hl_matrix_type.cuh       | 2 ++
 paddle/cuda/include/hl_sse_matrix_kernel.cuh | 3 +++
 2 files changed, 5 insertions(+)

diff --git a/paddle/cuda/include/hl_matrix_type.cuh b/paddle/cuda/include/hl_matrix_type.cuh
index 85b60cc313f..6917f362901 100644
--- a/paddle/cuda/include/hl_matrix_type.cuh
+++ b/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;
diff --git a/paddle/cuda/include/hl_sse_matrix_kernel.cuh b/paddle/cuda/include/hl_sse_matrix_kernel.cuh
index d774150c21e..c90d49e4ade 100644
--- a/paddle/cuda/include/hl_sse_matrix_kernel.cuh
+++ b/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
-- 
GitLab