Improve mGPU partitioning for Gaussian projection operators#547
Merged
matthewdcong merged 2 commits intoopenvdb:mainfrom Mar 16, 2026
Merged
Conversation
Signed-off-by: Matthew Cong <mcong@nvidia.com>
Signed-off-by: Matthew Cong <mcong@nvidia.com>
matthewdcong
changed the title
blackencino
approved these changes
Mar 13, 2026
Contributor
blackencino
left a comment
blackencino
left a comment
There was a problem hiding this comment.
As with #546, I'd love to see this be able to take advantage of some of the nice additions in the dispatch framework, but that's an optimization for a later day. This looks great.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Similar to #546 . Prior to this change, mGPU projection calculations were partitioned across the camera-Gaussian pairs, i.e. C * N. This means that each GPU would need to write to each of the N derivatives associated with the projection outputs which necessitated the use of system-scope atomics across GPUs therefore bottlenecking the interconnect.
Instead, we partition the calculations across Gaussians. Each mGPU processes a distinct segment of Gaussians for all cameras which means that atomics only need to be device scoped (with the exception of the camera pose derivative).
In addition, we load the cameras into shared memory for the backwards pass following the example of the forwards pass. Overall, these changes provide a >5% speedup on 2x RTX 3090s.