[SYCL][CUDA] WIP/RFC: cuda symbol table

[hiaselhans]

as pointed out in #1588 cuda fatbins are bundled without the entries table.
In order to have concurrent backends working we will definitely need that.

i was following the advices of @kbobrovs. There is a nvptx *.table and *.sym file now with proper symbols and binary filenames.
I'm asking for a little guidance on how to proceed:

next step would be to construct a llvm-foreach wrapping this:

llvm/clang/lib/Driver/Driver.cpp

Line 3473 in 9e79d31

Action *finalizeNVPTXDependences(Action *Input, const llvm::Triple &TT) {

My approach would be to create a SYCLCompileOffload::ConstructJob in clang/lib/Driver/ToolChains/Clang.cpp which runs the backend and assemble phase returning fatbin filenames.

Is this the right approach or should i be less invasive? As clang-offload-wrapper is handling a single fatbin it would also handle a fatbin batch, right?

Thanks for the help. I think I'm yet too far away..

[Ruyk]

FYI @Naghasan

[Naghasan]

My approach would be to create a SYCLCompileOffload::ConstructJob in clang/lib/Driver/ToolChains/Clang.cpp which runs the backend and assemble phase returning fatbin filenames.

I'm not sure what you mean by this. You want to create a new Tool to handle this ? Then to which ToolChain would it belong to ?

runs the backend and assemble phase a tool doesn't run "phases". There is a binding operation (you can see the result of this operation with -ccc-print-bindings) which then triggers a ConstructJob to build from a file type A to a file type B. It is also this operation that decides the file names.

[hiaselhans]

well, to be honest i'm still trying to find out about the difference of tools / jobs...

but in a way we'd need something like this:
list of bc files -> (tool/job) -> list of fatbins
to feed that to clang-offload-wrapper batch mode.
Then offloadwrapper could enclose the entries names.

I'm not 100% about this but maybe that would also allow to include the gpu_arch in specialization constants?