Matrix Multiplication Example from the "Next Chapter" 2023 Webinar

[rokelley]

I watched the "Next Chapter" webinar from yesterday (2/1/2023) and I really like the simple example of matrix multiplication in Fortran with OpenMP offload. When I offload to CPU, it works great. However, when I attempt to force GPU offload with

OMP_TARGET_OFFLOAD=mandatory

LIBOMPTARGET_DEVICETYPE=gpu

I get

matrix_multiply.f90:5:26: Libomptarget fatal error 1: failure of target construct while offloading is mandatory

(among many other messages of course)

My only question at this point is: should this work if the only device I have available is the Iris Xe integrated GPU from an i7-12700H?

[rokelley · Accepted solution]

Just in case anyone reads this later: yes, it apparently should work. In this case my input matrices were apparently too large (??), when I reduced the sizes to 2000x2000 it ran fine on CPU and GPU offload with OpenMP. I've been pushing the code to GitLab as I'm working on it here: https://gitlab.com/ryan.okelley/oneapi-gpu-offload-sample