Closed jwlawson closed 7 years ago
I'm currently allocating a cl::sycl::buffer
by hand, rather than using the Eigen device. This allows the code to be nicer and easier to use, but does mean that the TF allocator is not used. I'm not sure if this is a problem or not.
Fixes #118
Also the dev/integration
branch is getting too far behind, so the code this patches does not appear there. Can we update that branch?
The TileOP uses an Eigen kernel for Tensors with a small number of dimensions, however it falls back on a custom kernel for larger dimensions. This requires a separate SYCL kernel to handle these larger cases.