Brutex
/
cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

allow atomicAdd for CUDA 10 only 

Signed-off-by: raver119 <raver119@gmail.com>
master
raver119 2019-09-03 13:30:16 +03:00
parent f6f9437a36
commit 9d03bb9425
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
libnd4j/include/templatemath.h
View File

@ -1057,7 +1057,7 @@ inline __device__ uint64_t nd4j_atomicAdd<uint64_t>(uint64_t* address, uint64_t
template <> template <>
inline __device__ float16 nd4j_atomicAdd<float16>(float16* address, float16 val) { inline __device__ float16 nd4j_atomicAdd<float16>(float16* address, float16 val) {
#if __CUDA_ARCH__ >= 700 #if __CUDA_ARCH__ >= 700 && defined(CUDA_10)
atomicAdd(reinterpret_cast<__half*>(address), val.data); atomicAdd(reinterpret_cast<__half*>(address), val.data);
#else #else
auto address_as_ull = (int*) address; auto address_as_ull = (int*) address;