Asynchronous copy/compute support for GPU about monolish

[t-hishinuma]

Thanks for the reply.

for the near term EU exascale computers

I'm sorry, If you need MPI communication, MPI support in monolish will be next year.

The first goal of monolish is to develop a complete numerical library for any hardware on one node.

Asynchronous communication is important, but we're not going to get around to it for now.
We look forward to your pull requests.

When you have an API design for asynchronous communication, please share it with us.

regards.

[v1j4y]

Thanks for the detailed comments.

I am quite interested in this project due to the lack of standard libraries which are hardware agnostic and are written in C/C++. We're also working on a similar project albeit for Qauantum Monte Carlo applications. Here the goal is also to have a backend agnostic QMC kernel library for use by physicists for development of new methods for the near term EU exascale computers.

In our efforts, we have realized asynchronous copy/compute is essential for optimal saturation of the GPUs. We're are currently developing interfaces (for Fortran and C/C++) which includes the async copy/compute features along with the ability to simultaneously use CPU/GPU for the given workload in order to fully saturate the compute node.

I can have a look at your interfaces for GPU based Linear Algebra functions in details and can eventually send a pull request so that we can perhaps discuss in details what I mean by a unified interface.

[t-hishinuma]

Hi, Thanks for looking at this library.

There are no plans to support asynchrony in the near future.
This is because we don't have an idea for a unified and intuitive interface.
If you have an idea for an interface, we would be glad to hear from you.

AMD support is only a matter of time.
We have chosen OpenMP Offloading to support AMD Radeon and Intel Xe.
However, in the current version, many functions call cuBLAS, cuSPARSE, or cuSOLVER.
We will support AMD Radeon in the near future, either by implementing Rocm calls, or by implementing Reference BLAS for devices using OpenMP Offloading.

Regards,
T.Hishinuma