[eigen] Use of SYCL h/w acceleration in current version of Eigen

["Peterka, Thomas" <tpeterka@xxxxxxxxxxx>]

Hi,

I understand that the current version of Eigen has had some amount of GPU acceleration via the SYCL standard. I am using the latest Eigen version 3.3.7. I am using an alpha version of SYCL from Intel, part of their DPMC++ package, on an Intel Xeon with integrated Intel Gen9 CPU.

When I run a program w/ Eigen that was compiled with SYCL on a GPU-enabled machine, and I profile the execution, I do not see any Eigen kernels being offloaded to the GPU. I confirmed that other SYCL kernels that I wrote in my code are being offloaded to the GPU correctly, but nothing from Eigen.

Can someone explain what kernels in Eigen have been accelerated, and how to get them to run offloaded on a GPU? Eg., do I need to #define EIGEN_USE_SYCL?

I am using Eigen Vector and dense Matrix operations in my code. I am not using any of the Tensor module at this time. Is only the tensor module accelerated, and is this perhaps why I don’t see any offloading?

I could not find any documentation about this. I learned about the SYCL acceleration in Eigen from the Codeplay website:

https://developer.codeplay.com/products/computecpp/ce/guides/eigen-guide

and from this paper:

https://dl.acm.org/citation.cfm?id=3078160

Thanks in advance,

Tom

-----------------

Tom Peterka
Mathematics and Computer Science Division
Argonne National Laboratory
9700 S. Cass Ave.
Lemont, IL 60439
Bldg 240, Rm 2148
e-mail: tpeterka@xxxxxxxxxxx
web: www.mcs.anl.gov/~tpeterka
phone: 630-252-7198