[CUDA] Define __USE_FAST_MATH__ when __FAST_MATH__ is defined.

[jlebar]

Summary

The CUDA headers look for USE_FAST_MATH.

[jlebar]

Actually, after talking offline with Chandler, I need something more complicated than this. I will send a new patch.

Sorry for the noise.

lib/Headers/__clang_cuda_runtime_wrapper.h

  * this file during every CUDA compilation.
  */
 
 #ifndef __CLANG_CUDA_RUNTIME_WRAPPER_H__
 #define __CLANG_CUDA_RUNTIME_WRAPPER_H__
 
 #if defined(__CUDA__) && defined(__clang__)
 
+// The CUDA headers use the __USE_FAST_MATH__ macro to determine whether we get
+// the slow-but-accurate or fast-but-inaccurate versions of functions like sin
+// and exp, but -ffast-math defines only __FAST_MATH__.
+#if defined(__FAST_MATH__) && !defined(__USE_FAST_MATH__)
+#define __USE_FAST_MATH__
+#endif
+
 // Include some forward declares that must come before cmath.
 #include <__clang_cuda_math_forward_declares.h>
 
 // Include some standard headers to avoid CUDA headers including them
 // while some required macros (like __THROW) are in a weird state.
 #include <cmath>
 #include <cstdlib>
 #include <stdlib.h>