P2322R6 accumulator types for reduce

[gevtushenko]

This PR addresses the following issue. I also found that we use copy assignment operator on uninitialized memory, which might lead to issues for not primitive types.

⚠️ Breaking changes:

1. AgentReduce:
   1. now has extra template parameter representing the accumulator type.
   2. ConsumeTile uses accumulator type for thread aggregate instead of output iterator value type.
   3. ConsumeTile returns accumulator type instead of output iterator value type.
2. DeviceReduceKernel doesn't accept output iterator as a template parameter. Apart from that, it now accepts accumulator type.
3. DeviceReduceSingleTileKernel now accepts accumulator type.
4. DeviceSegmentedReduceKernel now accepts accumulator type.
5. DeviceReducePolicy now accepts accumulator type instead of input iterator value type. It also doesn't accept output iterator value type now.
6. DispatchReduce:
   1. now accepts accumulator type as last parameter.
   2. now accepts initializer type instead of output iterator value type.
   3. doesn't work with device extended lambdas now.
   4. Constructor now accepts init as initial type instead of output iterator value type.
7. DispatchSegmentedReduce:
   1. accepts accumulator type as last parameter.
   2. accepts initializer type instead of output iterator value type.
8. Thread operators now accept parameters using different types: Equality, Inequality, InequalityWrapper, Sum, Difference, Division, Max, ArgMax, Min, ArgMin.
9. ThreadReduce now accepts accumulator type and use different type for prefix
10. Default accumulator type is now selected by introspecting return type of the reduce operator.

[gevtushenko]

This PR should be merged after libcu++ supports device lambdas in invoke result: NVIDIA/libcudacxx#284

[alliepiper]

Can you add the release: breaking change label and put a description of these changes to the Dispatch interface in the PR description? That way I'll make sure to call these changes out in the release notes.

Same goes for all of the accumulator-type changes in behavior -- I check that label when I'm building relnotes from the list of PRs.

[alliepiper]

This will need a couple of things before merging:

• Rebase for debug_synchronous/DebugSyncStream/CDP changes
• Add summary of breaking changes to PR description.

[miscco]

I am not sure if I am yet qualified to really approve it but I can complain

[gevtushenko]

⚠️ Breaking changes:

1. AgentReduce:
   1. now has extra template parameter representing the accumulator type.
   2. ConsumeTile uses accumulator type for thread aggregate instead of output iterator value type.
   3. ConsumeTile returns accumulator type instead of output iterator value type.
2. DeviceReduceKernel doesn't accept output iterator as a template parameter. Apart from that, it now accepts accumulator type.
3. DeviceReduceSingleTileKernel now accepts accumulator type.
4. DeviceSegmentedReduceKernel now accepts accumulator type.
5. DeviceReducePolicy now accepts accumulator type instead of input iterator value type. It also doesn't accept output iterator value type now.
6. DispatchReduce:
   1. now accepts accumulator type as last parameter.
   2. now accepts initializer type instead of output iterator value type.
   3. doesn't work with device extended lambdas now.
   4. Constructor now accepts init as initial type instead of output iterator value type.
7. DispatchSegmentedReduce:
   1. accepts accumulator type as last parameter.
   2. accepts initializer type instead of output iterator value type.
8. Thread operators now accept parameters using different types: Equality, Inequality, InequalityWrapper, Sum, Difference, Division, Max, ArgMax, Min, ArgMin.
9. ThreadReduce now accepts accumulator type and use different type for prefix
10. Default accumulator type is now selected by introspecting return type of the reduce operator.

[alliepiper]

Breaking changes

Thanks for writing this up! Can you edit the PR description/first comment and add this to it? I'm less likely to overlook it that way :)

[gevtushenko]

Breaking changes

Thanks for writing this up! Can you edit the PR description/first comment and add this to it? I'm less likely to overlook it that way :)

Sorry, didn't think about this aspect 😄 I'll definitely update the description.