JuliaGPU/CUDAnative.jl

Cthulhu integration

Closed this issue · 3 comments

@device_code_warntype interactive=true app()

@vchuravy are you ready to commit on an API (mainly just descend_code_typed and descend_code_warntype) and declare a 1.0? Makes it easier to manage the package as a dependency.

What you don't like my current integration? https://github.com/JuliaDebug/Cthulhu.jl/blob/14b2e068f41742e90f116284e288c24ea72ad849/src/reflection.jl#L82-L84 ;)

For me the question is backend and frontend integration. E.g. how does CUDAnative tell Cthulhu which about the method-instance hidden by cufunction, similar problems exist for @async and how does a frontend package interface with Cthulhu.

For the frontend the descend_code_* methods should be stable enough. Albeit you should take a look that they are currently in the form you want them to be. Happy to tag a 1.0 if that is the case.

For me the question is backend and frontend integration. E.g. how does CUDAnative tell Cthulhu which about the method-instance hidden by cufunction, similar problems exist for @async and how does a frontend package interface with Cthulhu.

Right now CUDAnative doesn't rewrite the function, so (f,tt) as captured by the compiler hook identifies the code executed? And even if we'd contextualize, wouldn't (contextualize(f), tt) cover that?