The changes to the ScatterNd op in 63ac86a placed the SYCl kernel registration inside GOOGLE_CUDA guards, which means that they weren't registered except when compiled with both CUDA and SYCL support. Here I ensure that the SYCL registrations are outside of the GOOGLE_CUDA scope.
The changes to the ScatterNd op in 63ac86a placed the SYCl kernel registration inside
GOOGLE_CUDAguards, which means that they weren't registered except when compiled with both CUDA and SYCL support. Here I ensure that the SYCL registrations are outside of theGOOGLE_CUDAscope.