Regarding the performance of this CUDA version Trixi

[yiyuef]

Hey @huiyuxie!

It is an amazing repo that makes the original Trixi more versitile! Throughout my initial trial with this package, I found that on my laptop, with Nvidia hardware and CUDA installed, the examples are running slower than the running with the original Trixi.

For instance, I tried this case in TrixiCUDA.jl and this case in Trixi.jl, and the code difference only appears at solver = DGSEMGPU(polydeg = 3, surface_flux = flux_lax_friedrichs, volume_integral = VolumeIntegralWeakForm()),
semi = SemidiscretizationHyperbolicGPU(mesh, equations, initial_condition, solver, source_terms = source_terms_convergence_test)
and
ode = semidiscretizeGPU(semi, tspan).

According to the output log they both shows that the case was single thread. See screenshot below. (They both reached same results).

I am curious about: How can I test or know that the process is indeed running on CUDA, not my CPU?

Thanks!

[huiyuxie]

Thanks for your feedback!

According to the output log they both shows that the case was single thread.

It shows the number of CPU threads and we default to single thread on CPU. If you can run the example directly from our examples (as you mentioned) without any errors, then it is indeed running on the GPU.

the examples are running slower than when using the original Trixi.

Yes, that is true for most examples due to unresolved issues like #113 and #118, which severely harm performance. However, fixing these issues requires lots of changes to the code structure, so it will take some time. If you try running a single kernel (or a few kernels together), you will find that the GPU runs faster than the CPU, but overall the performance is not good enough.

I will try to resolve them as soon as I can.