experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Avoid using cuda specific GPU function to unbreak ROCM build.

Browse Source 
Indirect them through core/util/gpu_functions.h

PiperOrigin-RevId: 314622940
Change-Id: I0e5a349f759d0af6ff13acc43b34080a5104c9cc
This commit is contained in:
Gunhan Gulsoy 2020-06-03 16:05:26 -07:00 committed by TensorFlower Gardener
parent cfc9a6852e
commit a1e26e4298
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/core/kernels/scatter_nd_op_gpu.cu.cc
View File

@ -58,14 +58,14 @@ struct LeftUpdate<T, scatter_nd_op::UpdateOp::SUB> {
template <typename T> template <typename T>
struct LeftUpdate<T, scatter_nd_op::UpdateOp::MAX> { struct LeftUpdate<T, scatter_nd_op::UpdateOp::MAX> {
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void operator()(T* out, const T& val) { EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void operator()(T* out, const T& val) {
CudaAtomicMax(out, val); GpuAtomicMax(out, val);
} }
}; };
template <typename T> template <typename T>
struct LeftUpdate<T, scatter_nd_op::UpdateOp::MIN> { struct LeftUpdate<T, scatter_nd_op::UpdateOp::MIN> {
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void operator()(T* out, const T& val) { EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC void operator()(T* out, const T& val) {
CudaAtomicMin(out, val); GpuAtomicMin(out, val);
} }
}; };