Eigen support: only tensor module, or arrays and matrices as well?

1

I have a project whose compute code uses Eigen, and I wish to compile it for GPUs using ComputeCpp if possible.
While the Eigen Overview sounds like general support (“Codeplay has added OpenCL™ hardware support to Eigen”), reading a bit further makes it appear as if only Eigen’s tensor module is supported.
However, regarding that topic seems to be quite a while old (3-6 years, and I can’t post the links because of a link number limit as a new contributor).

Is it possible to compile Eigen code for GPUs with ComputeCpp, if the Eigen code uses Eigen::Array and Eigen::Matrix? And if not, is this planned for the (near) future?

2

Here are the links:

and

3

Hi @lens

Let me take the question and pass it on to people who are more familiar with this subject. I will get back to you with an answer.

Thank you.

4

ComputeCpp does not support Eigen anymore. Only DPCPP works with Tensor operation from the unsupported module. The unsupported module means in this context https://eigen.tuxfamily.org/dox/unsupported/index.html. They are modules that are not officially supported by the Eigen core dev team.

If you have any comments or questions please do put them here.

5

Hey illya, thanks for the reply!
On the documentation page you linked, there’s a single link to the “SYCL backend for Eigen”:
https://eigen.tuxfamily.org/dox/unsupported/SYCL_EIGEN.html
Both links on that page point to ComputeCpp-pages, and the tensor benchmark linked in one of them uses it as well:

It looks to me as if ComputeCpp is still required as a SYCL implementation - or the documentation is out of date. Could you clarify?

6

ComputeCpp after this PR ([SYCL-2020 Support] Enabling Intel DPCPP Compiler support to Eigen (!1126) · Merge requests · libeigen / eigen · GitLab), Eigen will not work with ComputeCpp as it requires support for the feature Generic pointer. Once ComputeCpp supports the feature allowing a generic address space, it will work again with Eigen.

Now you can run Eigen with DPC++, using the following:

export CXX=/path/to/intel/dpcpp_compiler/bin/clang++
export CC=/path/to/intel/dpcpp_compiler/bin/clang
export LD_LIBRARY_PATH=/path/to/intel/dpcpp_compiler/lib/
#if you are running on an Intel CPU or GPU
cmake … -DEIGEN_TEST_SYCL=1 -DEIGEN_SYCL_DPCPP=ON -DCMAKE_BUILD_TYPE=Release -DDPCPP_SYCL_TARGET=“spir64” -DSITE=“your site name” -DEIGEN_TEST_BUILD_DOCUMENTATION=1 -DEIGEN_INTERNAL_DOCUMENTATION=1 -GNinja
ctest --build-clean -C Release -D Experimental

The DPCPP target can be changed based on the architecture targeted: -DDPCPP_SYCL_TARGET=“spir64” is for Intel. NVidia has Ptx.

If required, to run only the SYCL tests, you can use the same EIGEN_MAKE_ARGS=“-j4” ./check.sh sycl

7

Sorry, forgot to mention. To use the Intel DPC++, a version LLVM clang compiler for SYCL, use the Intel oneAPI Base Toolkit.