Patch for parallel option "offload"

[PokhodenkoSA]

Introduced by @reazulhoque.
Possibly numba should provide a way to extend parallel options.

Field gen_spirv corresponds to offload option. gen_spirv is used only in numba.dppl.

[DrTodd13]

This will certainly remove gen_spirv as a baked in Numba parallel option. If you had access at the right time in the compile process then you could monkey-patch this option back in from outside of default Numba. If these PRs are going to create a hook at the appropriate time then monkey patching would work.

[PokhodenkoSA]

Investigation results:

1. gen_spirv variable is not necessary and could be removed, but value.pop('offload', False) make values empty and eliminates exception.
2. dppl uses offload flag in examples and tests.

• the final goal is to remove offload option for njit
  • examples and tests should be updated to use with context.
  • and dppl sources should also be changed to not use offload option
• if it not be removed then we need to upstream it or monkey-patch as @DrTodd13 proposed.

Final conclusion: We need to modify tests and examples first and then remove the option.

/CC @diptorupd

[DrTodd13]

What is the proposal for how to convey offload mode to the compilation pipeline when compiling a function called within a with context? @PokhodenkoSA

[PokhodenkoSA]

What is the proposal for how to convey offload mode to the compilation pipeline when compiling a function called within a with context?

Sorry, I didn't understand the question. Could you please provide an example? It is possible that we missed something.

[DrTodd13]

What is the proposal for how to convey offload mode to the compilation pipeline when compiling a function called within a with context?

Sorry, I didn't understand the question. Could you please provide an example? It is possible that we missed something.

Previously, compilation would look at the parallel flags to see if it should go through the GPU offload code. If that flag is removed from the parallel flags object then what will compilation look at to determine whether to go through the GPU offload code or not?

[reazulhoque]

Hi @DrTodd13 I think right now we check if we are in dpctl device_context (https://github.com/IntelPython/numba/blob/pydppl/numba/dppl/target_dispatcher.py#L63).

[PokhodenkoSA]

@diptorupd @reazulhoque
Is it enough to have only with device_context to manipulate offloading? Is it really no need anymore for njit flag for offloading?
Possibly, I am not familiar well with last discussions with customers and our new commitments.

[diptorupd]

Yes, the semantics that we finalized is to not add any new njit flag and only use with device_context to trigger offloading. The njit can be removed IMO.

[PokhodenkoSA]

Parallel option "offload" was removed.