Enable fixed-size reductions for expressions

[cliffburdick]

Fixes #1242

CUB reductions typically get transformed into an iterator for generic types. However, when we know the reduction is a basic pointer-style reduction on an inner dimension we can avoid materializing operators and iterators on input. This improves reductions up to about 170x, depending on size:

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[cliffburdick]

/build

[greptile-apps]

Greptile Summary

Enables CUB fixed-size segmented reductions for compatible random-access expressions, avoiding intermediate materialization.

• Adds fixed-size input/output reduction adapters and permutation compatibility selection.
• Applies the optimized path to generic, sum, minimum, and maximum reductions while retaining the offset-based fallback.
• Skips segmented CUB calls for empty outputs and adds expression-reduction coverage and documentation.

Confidence Score: 5/5

The PR appears safe to merge with no concrete correctness, security, or build failures identified.

The optimized path preserves the fallback path’s collapsed logical iteration and equal segment boundaries, while incompatible top-level permutations continue to use explicit offsets.

Important Files Changed

Filename	Overview
include/matx/core/reduce_utils.h	Adds compatibility detection and fixed-size adapters that preserve logical iterator ordering and output conversion behavior.
include/matx/transforms/cub.h	Selects fixed-size CUB segmented reductions for compatible non-scalar inputs, retains the prior fallback, and avoids calls for empty outputs.
test/00_operators/ReductionTests.cu	Adds numerical coverage for fixed-size sum, minimum, maximum, and product reductions over expressions.
docs_input/api/manipulation/selecting/reduce.rst	Documents intermediate-free fixed-size reductions for compatible CUDA expressions.
docs_input/api/math/sumprod/sum.rst	Documents the optimized fixed-size sum path for compatible CUDA expressions.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Reduction expression] --> B{Output empty?}
  B -->|Yes| C[Return]
  B -->|No| D{Input and output contiguous tensor views?}
  D -->|Yes| E[Direct CUB reduction]
  D -->|No| F{Fixed-size compatible?}
  F -->|Yes| G[Collapse expression]
  G --> H[Fixed-size segmented CUB reduction]
  F -->|No| I[Offset-based segmented CUB reduction]

Loading

_{Reviews (1): Last reviewed commit: "Enable fixed-size reductions for express..." | Re-trigger Greptile}

[coveralls]

Coverage is 93.4% — cburdick/issue-1242-fixed-size-expression-reductions into main. No base build found for main.