diff --git a/video/decode/gpu_memcpy_sse4.h b/video/decode/gpu_memcpy_sse4.h
index c441ff38fe..160209bdc5 100644
--- a/video/decode/gpu_memcpy_sse4.h
+++ b/video/decode/gpu_memcpy_sse4.h
@@ -19,14 +19,18 @@
  *  Taken from the QuickSync decoder by Eric Gur
  */
 
-#include <emmintrin.h>
+#ifndef GPU_MEMCPY_SSE4_H_
+#define GPU_MEMCPY_SSE4_H_
+
+#pragma GCC push_options
+#pragma GCC target("sse4.1")
+#include <smmintrin.h>
 
 // gpu_memcpy is a memcpy style function that copied data very fast from a
 // GPU tiled memory (write back)
 // Performance tip: page offset (12 lsb) of both addresses should be different
 //  optimally use a 2K offset between them.
-__attribute__((target("sse4"))) static inline void
-*gpu_memcpy(void *restrict d, const void *restrict s, size_t size)
+static inline void *gpu_memcpy(void *restrict d, const void *restrict s, size_t size)
 {
     static const size_t regsInLoop = sizeof(size_t) * 2; // 8 or 16
 
@@ -127,3 +131,6 @@ __attribute__((target("sse4"))) static inline void
 
     return d;
 }
+
+#pragma GCC pop_options
+#endif