Project

General

Profile

Task #2608

Remove code duplication between OpenCL and CUDA

Added by Roland Schulz 3 months ago. Updated about 2 months ago.

Status:
New
Priority:
Normal
Assignee:
-
Category:
-
Target version:
-
Difficulty:
uncategorized
Close

Description

Find agreement on how to express code that must compile with CUDA and equivalent OpenCL code. Some of the higher level code is identical if only we wrap the API calls.


Related issues

Related to GROMACS - Task #2530: consider using CUDA Driver APINew

History

#1 Updated by Roland Schulz 3 months ago

  • Related to Task #2530: consider using CUDA Driver API added

#2 Updated by Roland Schulz 3 months ago

Could we use https://github.com/CNugteren/CLCudaAPI for the wrapping or do we need to create our own?

#3 Updated by Mark Abraham 3 months ago

It looks great at a glance, except that we do want to use CUDA textures, and perhaps need to see about support for multiple streams

#4 Updated by Szilárd Páll about 2 months ago

An API like this would be great, but a concern is that it suffers from issues that third-party compatibility generally APIs do, especially if it's maintained by parties that can't afford being very up-to-date (like HIP): it easily gets out of data, especially given that CUDA features are added rather frequently.

Also available in: Atom PDF