Also run Python example benchmark on CUDA if is available

microsoft / knossos-ksc

Compiler with automatic differentiation

Other

45 stars 10 forks source link

Also run Python example benchmark on CUDA if is available #856

Closed cgravill closed 3 years ago

cgravill commented 3 years ago

Directly use PyTorch reference method on CUDA device for comparison.

AB#19088

cgravill commented 3 years ago

I'm not sure about naming, "manual" and "automatic" aren't very obvious to me what's going on:

PyTorch automatic CUDA PyTorch manual CUDA (with transfer) PyTorch manual CUDA PyTorch

I could just drop the "automatic". From the outside, does anyone have a suggested naming scheme?

Note I've only done the version with tensors on device i.e. no "with transfer".

dcrc2 commented 3 years ago

I'm not sure about naming, "manual" and "automatic" aren't very obvious to me what's going on:

PyTorch automatic CUDA PyTorch manual CUDA (with transfer) PyTorch manual CUDA PyTorch

How about just "PyTorch CUDA" (or "PyTorch GPU") and "Manual CUDA"? There's not really any need to refer to "PyTorch" in the second case because all of our benchmarks are run via PyTorch anyway.

cgravill commented 3 years ago

Thanks, names look clear to me now.