提交 9bc1a1a1 编写于 作者: D dangqingqing

fix cuda_helper.h

上级 31f59d2b
......@@ -34,8 +34,6 @@ USE_CUDA_ATOMIC(Add, float);
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600
USE_CUDA_ATOMIC(Add, double);
#else
// Custom implementation of atomicAdd for double.
// This implementation is copied from CUDA manual.
CUDA_ATOMIC_WRAPPER(Add, double) {
unsigned long long int* address_as_ull =
reinterpret_cast<unsigned long long int*>(address);
......@@ -50,8 +48,8 @@ CUDA_ATOMIC_WRAPPER(Add, double) {
} while (assumed != old);
return __longlong_as_double(old);
#endif
}
#endif
} // namespace platform
} // namespace paddle
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册