c-style error format #304
c-style error format #304
Conversation
There was a problem hiding this comment.
I'm on board with using printf for kernel debugging (there really shouldn't be any other use case to print from a kernel). However I'm a little concerned about having to switch all of our error logging to printf just because of it. Just think about the amount of changes in GEOS this will cause. It's also quite cumbersome to have to specify an additional format string "%s" everywhere while already using advanced formatting via fmt.
I think it's time we had separate macros for in-kernel vs out-of-kernel logging. The rule should be that any function/lambda marked as DEVICE or HOST_DEVICE is only allowed to use the former. Enforced (hopefully) via a CUDA CI build with debugging enabled.
| "***** Block: [%u, %u, %u]\n" \ | ||
| "***** Thread: [%u, %u, %u]\n" \ | ||
| "***** Controlling expression (should be false): " STRINGIZE( EXP ) "\n", \ | ||
| "***** MSG: " STRINGIZE( MSG )"\n\n" \ |
There was a problem hiding this comment.
If MSG is a quoted format string now, do you still need to STRINGIZE it?
|
@klevzoff, thanks for the comments. This PR was a shot in the dark regarding debugging CUDA kernels: I think it just doesn't work and requires a cleaner rework from scratch based on #303. I think I'll close this as :
|
Fix #303.
The rationale is that only
printfis supported onCUDAkernels, so we switch toc-styleformat.