[gonidelis]

fixes #5606

Boilerplate code for extracting types information (stream, mr, tuning_t etc.) is too big and repetitive across the new device environment based interfaces we introduced. This PR extracts the code into a separate function and re-uses it in the existing environment based device APIs that we have (DeviceScan and DeviceReduce).

Some consideration about the design for the reviewers:

• Each device primitive has its own quirks regarding which deterministm_t is supported. For example DeviceReduce::Reduce can support both gpu_to_gpu and run_to_run determinism, while DeviceReduce::ArgMax/Min or DeviceScan only support run_to_run at the moment. That means the determinism heuristics cannot be incorporated into the boilerplate code. Future environment-based APIs must individually evaluate each algorithm to determine and support the appropriate deterministic types.
• The existing boilerplate code uses a lambda callable to pass the specific deterministic algorithm implementation by packing the arguments.

      auto reduce_callable = [&](auto tuning, void* storage, size_t& bytes, auto... args) {
        using tuning_t = decltype(tuning);
        return reduce_impl<tuning_t>(storage, bytes, args...);
      };

      // Dispatch with environment - handles all boilerplate
      return detail::dispatch_with_env(
        env, determinism_t{}, reduce_callable, d_in, d_out, num_items, reduction_op, ::cuda::std::identity{}, init);
    }

I need some feedback on whether this interface on the dispatch_with_env() looks sane.

[gonidelis]

I need to figure out why this bug was not caught from the tests

[gonidelis]

LLVMs 15-18 are not friendly with complex C++20 aggregate expressions codegen hence this PR breaks CCCL for these compilers/C++ version.

Investigating what's the best course of action. I see three options:

a) simplify the code in this pr
b) file a bug to llvm
c) reject pr

digging deeper to see what's the exact issue.

[gonidelis]

i would promise we tried the same thing with @miscco

[gonidelis]

Once green this should be good to go

[bernhardmgruber]

Suggested change

	//! @param algorithm_callable Callable that invokes the algorithm implementation with determinism specified
	//! @param algorithm_callable Callable that invokes the algorithm

[bernhardmgruber]

Why do we need this here?

[bernhardmgruber]

Also, I think this is unused.

[gonidelis]

will patch it up

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 14h 56m: Pass: 100%/95 | Total: 5d 01h | Max: 5h 51m | Hits: 69%/91590

See results here.