[ROCm] Triton in XLA for ROCm - ir_emitter_triton related changes.

[zoranjovanovic-ns]

Second commit of the series for enabling Triton in XLA for ROCm.

[ddunl]

I think it'd be good if possible to minimize the number of ifdefs, so maybe ifdefs used everywhere at the top, then

#ifdef GOOGLE_CUDA
<includes> 
#endif
#ifdef TENSORFLOW_USE_ROCM
<includes>
#endif

or similar at the bottom of the includes.

[tdanyluk]

In general I would recommend to reduce the number of ifdefs inside the code, to close to 0.
Could we have some kind of abstraction, like have a
Something class, and a derived CudaSomething and RocmSomething or just use a runtime if based on whether the se::GpuComputeCapability is cuda or rocm?

[tdanyluk]

I talked with my colleagues and a class is not really needed.
We could use something like this:

something.h:

int something();

something_cuda.cc:

int something() {
 // implemented with cuda
}

something_rocm.cc:

int something() {
 // implemented with rocm
}

And then build either something_rocm.cc or something_cuda.cc with a select.
So ir_emitter_triton.cc would remain mostly device vendor independent and just include the shared "something.h'.

[zoranjovanovic-ns]

Thank you for the proposal and clarification. I will change the implementation (based on proposal) and update PR.

[tdanyluk]

thank you!

[zoranjovanovic-ns]

Hi @xla-rotation, I have rebased and pushed, do I need to do something more?

[penpornk]

@zoranjovanovic-ns Thank you for checking and for merging conflicts! We are applying some more fixes to pass internal tests. I'll let you know if we need more help. :)