Princeton GPU hackathon 2022 wrap-up

[FlorianReiss]

during the hackathon several issues related to running on GPUs were understood

stack overflow

some functions, in particular the kMatrix, allocate a lot of memory on the stack, which can exceed the default limit. Possible solutions are more memory-friendly implementations, using malloc (usually slower) or increasing the stack size with cuda_error_check(cudaDeviceSetLimit (cudaLimitStackSize, 1024*50));

related issues #287 #242

using RO_CACHE on stack memory

it is not allowed to use RO_CACHE/__ldg on stack memory

related issues #216 #217

see MR #328

to do

• finalize and merge fixes
• document everything learned

[FlorianReiss]

@thboettc @henryiii @JuanBSLeite just a little summary to keep track of the things we fixed or should be able to fix after the hackathon. Feel free to edit