Feature/add oversampled polyphase channelizer

[tbensonatl]

Add support for oversampling to the polyphase channelizer.

This update adds support for decimation factors (D) that are lower than the number of channels (M) with the polyphase channelizer. For all cases, the channelizer generates M outputs for each D inputs with any remaining partial set of inputs being zero-padded to D elements. Cases of D == M correspond to the maximally decimated, or critically sampled, case, which was previously supported. With maximal decimation, the channel frequency bands partition the frequency space, so the channel frequency support is adjacent but not overlapping. The oversampled cases D < M maintain the same channel center frequencies, but the channels have some overlap.

The per-channel length of the output tensor is (input_len + D - 1) / D. Thus, for example, with M=20 D=10, there is a 2x oversampling factor and the output tensor will have twice as many samples per channel as with M=20 D=20.

[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.

[tbensonatl]

/build

[greptile-apps]

Greptile Summary

This PR adds oversampling support to the polyphase channelizer, enabling decimation factors D < M. It introduces a new tiled shared-memory kernel (ChannelizePoly1D_SmemTiled) supporting both the maximally-decimated and oversampled paths, a fused complex MAC helper (channelize_cmac), comprehensive tests covering integer/rational oversampling, large-M tiling, and batch modes, and a Python reference generator for cross-validation. Previously flagged P0/P1 issues (division-by-zero, alignment, loop-invariant s, misleading comment) are all addressed in the current HEAD.

Confidence Score: 5/5

Safe to merge; all previously flagged P0/P1 issues are resolved and remaining findings are P2 style/comment improvements

All prior P0/P1 concerns (division-by-zero, smem alignment, loop-invariant hoisting, misleading inequality comment) are addressed. The new kernel logic, phase-rotation math, and dispatch hierarchy are sound. Remaining findings are a silently-ignored size parameter in the Python test generator and a misleading test comment — neither affects correctness of the production code.

test/test_vectors/generators/00_transforms.py — harris2003_oversampled_operators::channelize() ignores self.size

Important Files Changed

Filename	Overview
include/matx/kernels/channelize_poly.cuh	Adds MaximallyDecimated template parameter to ChannelizePoly1D, new ChannelizePoly1D_SmemTiled kernel (D==M and D<M paths), and channelize_cmac FMA helper; logic appears correct with proper circular-buffer and phase-rotation handling
include/matx/transforms/channelize_poly.h	Dispatch logic updated to route oversampled inputs through SmemTiled or generic kernels; fused-DFT and Smem kernels correctly gated to D==M only
include/matx/operators/channelize_poly.h	Output dimension formula correctly updated from num_channels to decimation_factor; documentation substantially expanded with Harris convention explanation and reference
test/00_transform/ChannelizePoly.cu	Thorough new test suite (identity filter, integer/rational oversampling, large-M tiling, batched, complex filter, fallback path); one test comment incorrectly describes what triggers the fallback
test/test_vectors/generators/00_transforms.py	Adds channelize_oversampled reference and harris2003_oversampled_operators; harris class hardcodes M/D/input_len constants and silently ignores the size parameter passed from C++
examples/channelize_poly_bench.cu	Refactored to accept explicit -M/-D flags with proper validation guards; atol() used for parsing without detecting invalid string inputs

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["channelize_poly_impl(in, filter, M, D)"] --> B{"real input\n& real filter?"}
    B -- yes --> C{"D==M and M<=6?"}
    B -- no --> D{"D==M and M<=6?"}
    C -- yes --> E["FusedChan kernel"]
    C -- no --> F{"D==M and Smem fits?"}
    F -- yes --> G["Smem kernel"]
    F -- no --> H{"SmemTiled eligible?"}
    H -- yes --> I["SmemTiled kernel"]
    H -- no --> J["Generic ChannelizePoly1D"]
    D -- yes --> E
    D -- no --> K{"D==M and Smem fits?"}
    K -- yes --> G
    K -- no --> L{"SmemTiled eligible?"}
    L -- yes --> I
    L -- no --> J
    I --> M{"MaximallyDecimated\nD==M?"}
    M -- yes --> N["Fixed phase per channel\nincremental buf_row advance"]
    M -- no --> O["phase = (c + t*D) % M\nK-rotation filter cache"]

Loading

_{Reviews (5): Last reviewed commit: "Make ::detail constants visible for host..." | Re-trigger Greptile}

[tbensonatl]

/build

[tbensonatl]

/build

[tbensonatl]

/build

[coveralls]

Coverage is 91.829% — feature/add-oversampled-polyphase-channelizer into main. No base build found for main.

[cliffburdick]

Awesome work!