Cuda multi-GPU support: Make some variables device-specific, update Kokkos::fence #6753
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
masterleinad
changed the title
|
Retest this please |
|
Retest this please. |
masterleinad
changed the title
|
The failures in two Cuda build are appear to be unrelated to this pull request and more of a machine issue. |
dalg24
reviewed
Feb 1, 2024
Comment on lines
123
to
127
| inline static std::set<int> cuda_devices = {}; | ||
| inline static std::map<int, unsigned long*> constantMemHostStagingPerDevice = | ||
| {}; | ||
| inline static std::map<int, cudaEvent_t> constantMemReusablePerDevice = {}; | ||
| inline static std::map<int, std::mutex> constantMemMutexPerDevice; |
There was a problem hiding this comment.
Did you put much thought whether to use std::{map,set} versus their unordered counterpart?
There was a problem hiding this comment.
No, I assume these are holding few elements and would never be called in a performance-critical code path.
If we cared about performance, we would rather use a std::vector(or small_vector) here anyway and manually search through it where necessary.
dalg24
reviewed
Feb 1, 2024
dalg24
approved these changes
Feb 1, 2024
Co-authored-by: Damien L-G <dalg24@gmail.com>
tcclevenger
approved these changes
Feb 1, 2024
|
Failures are unrelated. I intend to merge when/if the last CUDA build passes
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Part of #6091. Some of the
staticCudaInternalvariables should be created per device as the comments in the code are already saying. Namely, these areconstantMemHostStaging,constantMemReusable, andconstantMemMutex.This pull request turns each of them into a
static std::mapthat is lazily initialized. To properly deallocate the respective memory inimpl_finalize(), we also have to store the devices thatKokkosis using inside another lazily initializedstd::set.This allows us to also update
cuda_device_synchronizeresp.Kokkos::fencealready to fence all devices that we know of.I decided to
fenceall devices incuda_device_synchronizeinstead of adding adevice idparameter and run through all of them at the call site since it appeared that this matches the implied behavior the best and requires fewer changes.