[starpu-devel] Port StarPU onto Graphcore IPUs

[Changjiang Gou <gouchangjiang@gmail.com>]

Dear developers of StarPU,

First of all, thank you for developing this amazing runtime that makes running applications on heterogenous platforms far easier.

I would like to run PaStiX on Graphcore IPUs, which is an accelerator. The first thing I need to do is to port StarPU onto IPUs. I would like to have your suggestions to see if I am on the right track. Many thanks in advance.

According to one of your paper Exploiting the Cell/BE architecture with the StarPUunified runtime system, I need to do mainly three things:

• Find the runtime library of IPUs
• design/write a driver
• implement some codelets

For these three steps:

• Find the runtime library of IPUs: Poplar is the library, like CUDA from Nvidia. Many predefined functions such as linear algebra operations, element-wise tensor operations, non-linearities and reductions are implemented. It is possible to write my own code that runs directly on the IPU as a codelet in C++ or assembly code. Particularly, it already provides these linear algebra functions: Cholesky, ConvParams, ConvPreplan, ConvUtil, Convolution, FullyConnected, MatMul, MeshGrid, MultiConvolution, Norms, TriangularSolve.
• design/write a driver: I am planning to do it as the driver for Nvidia GPUs.
• implement some codelets: I am planning to do it when I need it.

My first step is to implement a simple matrix multiplication through StarPU, like what you did in the paper. Did I miss something for the first step? 

Could you please provide me a INRIA gitlab account? so I can fork a branch. I had an account before, GOU Changjiang · GitLab (inria.fr). Since I graduated from ENS de Lyon, I lost my email. I cannot login into it anymore. It would be really nice if I can use this account again.

Best regards,

Changjiang