diff options
Diffstat (limited to 'tensorflow/core/kernels/shape_ops.cc')
| -rw-r--r-- | tensorflow/core/kernels/shape_ops.cc | 37 |
1 files changed, 37 insertions, 0 deletions
diff --git a/tensorflow/core/kernels/shape_ops.cc b/tensorflow/core/kernels/shape_ops.cc index 7ff812cf27..496865de02 100644 --- a/tensorflow/core/kernels/shape_ops.cc +++ b/tensorflow/core/kernels/shape_ops.cc @@ -210,6 +210,43 @@ REGISTER_KERNEL_BUILDER(Name("ShapeN") ShapeNOp<int64>); #endif +#if TENSORFLOW_USE_SYCL +#define REGISTER_SYCL_KERNEL(type) \ + REGISTER_KERNEL_BUILDER(Name("ShapeN") \ + .Device(DEVICE_SYCL) \ + .HostMemory("output") \ + .TypeConstraint<int32>("out_type") \ + .TypeConstraint<type>("T"), \ + ShapeNOp<int32>); \ + REGISTER_KERNEL_BUILDER(Name("ShapeN") \ + .Device(DEVICE_SYCL) \ + .HostMemory("output") \ + .TypeConstraint<int64>("out_type") \ + .TypeConstraint<type>("T"), \ + ShapeNOp<int64>) + +TF_CALL_NUMBER_TYPES_NO_INT32(REGISTER_SYCL_KERNEL); +#undef REGISTER_SYCL_KERNEL + +// 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("ShapeN") + .Device(DEVICE_SYCL) + .HostMemory("input") + .HostMemory("output") + .TypeConstraint<int32>("T") + .TypeConstraint<int32>("out_type"), + ShapeNOp<int32>); +REGISTER_KERNEL_BUILDER(Name("ShapeN") + .Device(DEVICE_SYCL) + .HostMemory("input") + .HostMemory("output") + .TypeConstraint<int32>("T") + .TypeConstraint<int64>("out_type"), + ShapeNOp<int64>); +#endif // TENSORFLOW_USE_SYCL + class RankOp : public OpKernel { public: explicit RankOp(OpKernelConstruction* ctx) : OpKernel(ctx) {} |