[PROTOTYPE] Generalized two pass algorithm and copy_if

[danhoeflinger]

Summary

This PR changes the two pass algorithm to be more generalized for use with other scan-like algorithms like copy_if.

This PR adds copy_if as an example. partition and unique patterns should follow easily. Set algorithms have not been evaluated, but should also be possible.

Structural changes

The "bridge" between blocks now uses a saved carry-out value from the previous block (where it used to depend on the output of the previous block). The output of the previous block only applies to transform_scan patterns, and not scan_copy algorithms. It also required different logic for inclusive and exclusive scans. The change to use the carry unifies this logic between inclusive and exclusive scans.

The following operations have been defined to encode a generalized scan-like pattern:

• _GenReduceInput : a function which accepts the input range and index to generate the data needed by the main output used in the reduction operation (to calculate the global carries)
• _GenScanInput : a function which accepts the input range and index to generate the data needed by the final scan and write operations, for scan patterns
• _ScanPred : a unary function applied to the ouput of _GenScanInput to extract the component used in the scan, but not the part only required for the final write operation
• _ReduceOp : a binary function which is used in the reduction and scan operations
• _FinalOp: A function which accepts output range, index, and output of _GenScanInput applied to the input range.

For transform_scan, these are defined as follows:

• _GenReduceInput : apply the transfom to input at index and return
• _GenScanInput : apply the transfom to input at index and return
• _ScanPred : no_op passthrough
• _ReduceOp : user supplied binary scan operation
• _FinalOp: Simple write value to output range at index

For copy_if these are defined as follows:

• _GenReduceInput : apply user supplied predicate to input at index and return 1 for true, 0 for false
• _GenScanInput : a tuple with 3 elements:
  1) apply user supplied predicate to input at index and return 1 for true, 0 for false (contribution to count)
  2) apply user supplied predicate to input at index (condition for final copy)
  3) original value (value for final copy)
• _ScanPred : std::get<0>(input)
• _ReduceOp : std::plus
• _FinalOp: if the second element of the tuple is true, write the third element to the output range using the index found in the first element of the tuple

Odds and ends

• Makes the two pass scan pattern asynchronous, depending on __result_and_scratch_space for async deallocation

• Changes the single workgroup copy_if to have a smaller threshold, and to use the __result_and_scratch_space struct to match return future type with the copy_if implementation.

• Limits range to only launch non-empty workgroups (removed early exit for empty workgroups)

TODO next:

Replace usages of __parallel_transform_scan_base to route to __parallel_transform_reduce_then_scan instead. Modify the calling code as necessary. This should cover the rest of the scan-like patterns. Prioritize partition, then unique, then finally, set algorithms as time allows.

[adamfidel]

Just some initial comments.

[mmichel11]

First few minor things I see.

[mmichel11]

LGTM

[adamfidel]

LGTM.