Mark Harris comments

Results 363 comments of


                                            Mark Harris

[FEA] Add a argument or flag to control whether the constructor and operator= should be async

> I think the original issue brings up valid points. There's a lot of things I don't like about `device_buffer` that I fixed in `device_uvector`: > > 1. There is...

[FEA] Add a argument or flag to control whether the constructor and operator= should be async

> @jrhemstad are you suggesting we remove this constructor? @jrhemstad can you comment? I'd like to make progress on closing or implementing this FEA.

[FEA] Add a argument or flag to control whether the constructor and operator= should be async

Wouldn't an sync factory need to call an async constructor?

[DO NOT MERGE] Experiment with using cuda::memory_resource interface

Thanks @robertmaynard ! #883 is the more relevant PR for this. Mentioning it here so this shows up on that PR.

[WIP] Support enabling PTDS via CUDA_PTDS environment variable

> When you say "this is the right place", are you referring to RMM itself, to the Cython binding or to the `CudaStreamView`? I mean the constructor of CudaStreamView. I...

[WIP] Support enabling PTDS via CUDA_PTDS environment variable

> Are you saying that you want a way to explicitly request cudaStreamLegacy even when we have the environment variable explicitly passed by the user? I mean that it already...

[WIP] Support enabling PTDS via CUDA_PTDS environment variable

> it gets very annoying to type all that I generally type env variables into scripts, not at the command line, so the length doesn't matter much. PTDS is not...

[WIP] Support enabling PTDS via CUDA_PTDS environment variable

I'm firmly against PTDS in the name of the environment variable. It's inconsistent with the rest of RMM and with CUDA.

[WIP] Support enabling PTDS via CUDA_PTDS environment variable

Please retarget to `branch-0.18` when ready. Adding to 0.18 project for tracking.

[WIP] Support enabling PTDS via CUDA_PTDS environment variable

One thing to consider here is that this environment variable currently only affects the Python side. If the app is using RMM from both Python and C++ they may see...