[SYSTEMDS-3362] CUDA code gen stream synchronization

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: SystemDS 2.2
• Fix Version/s: SystemDS 3.0
• Component/s: NativeRuntime
• Labels: CUDA GPU code-generation
• Epic Link: spoof-cuda
• Assignee: Unassigned
• Reporter: Mark Dokter
• Created: 28/Apr/22 11:44
• Updated: 10/May/22 09:23
• Resolved: 10/May/22 09:23

Description

The CUDA code generation launcher handles streams per operator at the moment. This is wrong since a read before write can happen on a certain device allocation. Switching to a central stream object for now.

Issue Links

• links to: GitHub Pull Request #1600