Fix Issues with CUDA Backend - Githubissues

hpcgarage / spatter

Benchmark for measuring the performance of sparse and irregular memory access.

Other

75 stars 15 forks source link

Fix Issues with CUDA Backend #207

Closed radelja closed 2 months ago

radelja commented 2 months ago

Overview

This PR resolves #201 by updating the report function and addresses an out-of-memory error by moving the device buffers from the ConfigurationBase class to the ClArgs struct.

✨ Change Description/Rationale

Updated report function to remove unnecessary multiplication and subsequent division by nruns
- Now it more closely aligns with the calculation of bytes_moved in Spatter v1.1
Moved dev_sparse, dev_sparse_gather, dev_sparse_scatter, and dev_dense buffers to ClArgs
- Reduces the memory usage of the CUDA backend when multiple configs are provided
Closes https://github.com/hpcgarage/spatter/issues/201

👀 Reviewer Checklist

[ ] All GitHub actions and runners have passed if applicable
[ ] Run Spatter tests on CUDA backend
[ ] Commits are clean and relevant

✅ PR Checklist

[x] Remove or update the template boilerplate text
[x] Commits are relevant and combined where appropriate
[x] Rebase off spatter-devel
[x] Reviewers Requested
[ ] Projects associated
[ ] Commits mention issue and/or PR numbers at the bottom of the message
[x] Relevant issues are linked into the PR
[x] TODOs are completed
[ ] Reviewer checklist is updated

🚀 TODOs

[x] Fix issue freeing device buffers

📌 Future Work

Update Spatter CI workflow to build and run the CUDA backend (currently it only builds)