diff options
Diffstat (limited to 'tensorflow/core/kernels/pack_op.cc')
| -rw-r--r-- | tensorflow/core/kernels/pack_op.cc | 25 |
1 files changed, 25 insertions, 0 deletions
diff --git a/tensorflow/core/kernels/pack_op.cc b/tensorflow/core/kernels/pack_op.cc index e072eb36b3..4977ad1d7c 100644 --- a/tensorflow/core/kernels/pack_op.cc +++ b/tensorflow/core/kernels/pack_op.cc @@ -34,6 +34,9 @@ typedef Eigen::ThreadPoolDevice CPUDevice; #if GOOGLE_CUDA typedef Eigen::GpuDevice GPUDevice; #endif // GOOGLE_CUDA +#ifdef TENSORFLOW_USE_SYCL +typedef Eigen::SyclDevice SYCLDevice; +#endif // TENSORFLOW_USE_SYCL // -------------------------------------------------------------------------- template <typename Device, typename T> @@ -156,4 +159,26 @@ REGISTER_KERNEL_BUILDER(Name("Pack") #endif // GOOGLE_CUDA +#ifdef TENSORFLOW_USE_SYCL + +#define REGISTER_SYCL(type) \ + REGISTER_KERNEL_BUILDER( \ + Name("Pack").Device(DEVICE_SYCL).TypeConstraint<type>("T"), \ + PackOp<SYCLDevice, type>) + +REGISTER_SYCL(float); +#undef REGISTER_SYCL + +// A special GPU kernel for int32. +// TODO(b/25387198): Also enable int32 in device memory. This kernel +// registration requires all int32 inputs and outputs to be in host memory. +REGISTER_KERNEL_BUILDER(Name("Pack") + .Device(DEVICE_SYCL) + .HostMemory("values") + .HostMemory("output") + .TypeConstraint<int32>("T"), + PackOp<CPUDevice, int32>); + +#endif // TENSORFLOW_USE_SYCL + } // namespace tensorflow |