[WIP] Build RMM's Python bindings with PTDS #480
Conversation
|
Please update the changelog in order to start CI tests. View the gpuCI docs here. |
jakirkham
force-pushed
the
bld_w_ptds
branch
2 times, most recently
from
75f826a to
933b1a3
Compare
jakirkham
changed the title
jakirkham
added
3 - Ready for review
Defines `CUDA_API_PER_THREAD_DEFAULT_STREAM` for Setuptools `Extensions` built by `setup.py`. This should make sure the per-thread default stream is used by these Python extensions. https://docs.nvidia.com/cuda/cuda-runtime-api/stream-sync-behavior.html
|
rerun tests |
|
We should be careful to test at least cudf with this before merging as well as doing some basic benchmarking to see if the extra synchronization caused by mixing PTDS and legacy default streams cause slowdowns. |
harrism
changed the title
|
cc @pentschev (for vis) |
| @@ -84,6 +84,10 @@ def get_cuda_version_from_header(cuda_include_dir): | |||
| except Exception: | |||
| nthreads = 0 | |||
|
|
|||
| define_macros = [ | |||
| ("CUDA_API_PER_THREAD_DEFAULT_STREAM", None), | |||
There was a problem hiding this comment.
So does this make this a command line option? How does this work? (Not that familiar with setup.py)
jakirkham
changed the title
jakirkham
added
2 - In Progress
|
Sorry was updating this based on feedback from Peter earlier. Have marked it as a draft. |
|
This PR has been marked stale due to no recent activity in the past 30d. Please close this PR if it is no longer required. Otherwise, please respond with a comment indicating any updates. This PR will be marked rotten if there is no activity in the next 60d. |
|
@jakirkham can we close this? |
|
I think we want to keep this open as a point of reference, but no action should be taken on it. |
|
Ah sorry yeah we can close. I think Peter has PR ( #633 ), which is the successor to this |
|
Neither is a great option. Closing this one. |
|
What do we not like about the other one? Sounds good. |
|
See the discussion starting here. #633 (comment) |
Defines
CUDA_API_PER_THREAD_DEFAULT_STREAMfor SetuptoolsExtensionsbuilt bysetup.py. This should make sure the per-thread default stream is used by these Python extensions.Note: We still need other downstream changes as well ( rapidsai/cudf#5922 ). So we may want to hold on this a bit until those are resolved (hence marked as WIP).Based on discussions with others, it sounds like this change is fine on its own. So marking as ready.https://docs.nvidia.com/cuda/cuda-runtime-api/stream-sync-behavior.html
cc @kkraus14 @shwina @jrhemstad