Merge pull request #12639 from tomoaki0705:fixFp16CudaFailure

This commit is contained in:
Alexander Alekhin 2018-09-25 13:19:51 +00:00
commit 9faacfbc0d

View File

@ -288,7 +288,7 @@ template <> __device__ __forceinline__ short cast_fp16<float, short>(float v)
{ {
#if __CUDACC_VER_MAJOR__ >= 9 #if __CUDACC_VER_MAJOR__ >= 9
__half h(v); __half h(v);
return *(short*)&v; return *(short*)&h;
#else #else
return (short)__float2half_rn(v); return (short)__float2half_rn(v);
#endif #endif