Optimize channelize_poly CUDA kernels and dispatch

[tbensonatl]

Add fused FIR+DFT kernels that keep filtered samples in shared memory or registers, avoiding the intermediate tensor and separate cuFFT launch. Add small-channel leaves for critically sampled M=2..6 (replacing the prior FusedChan kernel) and general leaves for M=8, 10, 16, 20, 32, 40, 64, and 80, plus oversampled M=3..6. The infrastructure is in-place to extend coverage of fusion, but this PR does not do so in order to limit template instantiations. Instead, the aim is for a future PR to add a compile-time property that allows users to opt-in specific channel counts for fusion.

Rework the FIR+cuFFT backends via tuning of tile sizes and launch configuration. Select backends through SelectPlan/ExecutePlan using device attributes (SM count, L2 size, FP64 throughput, memory bus width).

Tests compare every launchable backend, including windowed launches, against the host implementation.

Across a 4,832-shape sweep, the geometric-mean speedup over the previous implementation is 1.89x on L4 and 2.15x on GH200. Some regressions remain with these changes, but the vast majority of cases improve with some cases being more than twice as fast.

[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.

[greptile-apps]

Confidence Score: 5/5

[Medium risk] Optimizes polyphase channelizer kernels and updates tests.

The PR appears safe to merge; no actionable regression was established.

Summary

The PR adds fused CUDA FIR+DFT channelizer kernels and selects among fused and separate backends using shared-memory fit and device characteristics. It also retunes FIR launches, extends backend and windowed-output tests, and documents half-precision limitations.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Validate channelizer arguments] --> B{Fused leaf feasible and preferred?}
  B -- Yes --> C[Fused FIR and DFT]
  B -- No --> D{Critical sampling and Smem suitable?}
  D -- Yes --> E[Whole-channel shared-memory FIR]
  D -- No --> F{Tiled plan fits?}
  F -- Yes --> G[Channel-tiled FIR]
  F -- No --> H[Generic FIR]
  E --> I[cuFFT]
  G --> I
  H --> I
  I --> J{Real filtered samples?}
  J -- Yes --> K[Unpack spectrum]
  J -- No --> L[Output]
  K --> L
  C --> L

Loading

_{Reviews (1) · Last reviewed commit: "Optimize channelize_poly CUDA kernels an..."}

[tbensonatl]

/build

[coveralls]

Coverage is 93.629% — tbenson/channelize-poly-fft-fusion-opts into main. No base build found for main.