Add fast-path TensorAccessor & Optimize SarBp/ChannelizePoly

[tbensonatl]

Adds detail::TensorAccessor (a per-kernel wrapper with an IsUnitStride fast path) and plumbs it through the SAR BP and channelize_poly transforms. Extends the SAR BP shared-mem preamble to float and mixed precisions. Adds a CTILE=32 variant of ChannelizePoly1D_SmemTiled for num_channels <= 32, which replaces the generic ChannelizePoly1D kernel for most small-channel oversampled configs (M=32/D=16, M=40/D=20 etc.) at roughly 3x the throughput. Adds a Full filter smem layout (single copy at smem[p*M + phase]) alongside the existing Rotated layout; dispatcher picks the smaller footprint. Refactors the SmemTiled FIR loops to use a single running filter_idx counter, saving 3 registers/thread. Scopes all channelize_poly internals under matx::detail::cpoly, dropping redundant MATX_CHANNELIZE_POLY1D_ and matxChannelizePoly1DInternal_ prefixes. Adds support for num_channels = 1 to channelize_poly() for completeness, although such cases should use conv1d() for better performance.

The TensorAccessor class improves performance for tensors with unit stride by eliding the load and multiply of the last-dimension stride. Speedups range from 3-12%, depending on the GPU and precision. The float path for the sarbp example improved by > 3x on GPUs with reduced double-precision throughput, but that is due to amortizing the cost of double-to-float conversions.

General polyphase channelizer improvements from the TensorAccessor class range from 0-10%. Oversampled cases with <= 48 channels improved by > 2x due to the addition of a tiled shared memory implementation with a smaller tiling factor.

[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]

Greptile Summary

This PR introduces detail::TensorAccessor — a compile-time-selected wrapper that replaces the per-access stride reload in the SAR BP and channelize_poly hot loops with a single pointer-plus-outer-strides snapshot taken at kernel entry, yielding 3–12% throughput gains. It simultaneously extends the SAR BP float/mixed shared-memory preamble, adds a CTILE=32 ChannelizePoly1D_SmemTiled variant for num_channels ≤ 32 (≈3× speedup over the generic path), and introduces a Full filter smem layout that saves footprint for small-channel oversampled configs.

The previously flagged P1 concern about scalar range_to_mcp has been addressed — the API now requires a MatX operator (enforced by static_assert), and all existing callers already use make_tensor. The rtm_acc int{0} placeholder pattern is fully replaced. Three P2 observations remain (see inline comments): the __CUDA_ARCH__ allowlist, the mutable return from const operator(), and the managed-memory assumption in the new accessor test.

Confidence Score: 5/5

Safe to merge — no P0/P1 findings; the previously flagged scalar range_to_mcp concern is resolved and all P2s are non-blocking.

All P1 concerns from the previous review round have been addressed: the scalar range_to_mcp API now requires a MatX operator (enforced by static_assert), the int{0} rtm_acc placeholder is gone, and all existing callers already use tensor views. The three remaining P2s (CUDA_ARCH allowlist, const accessor mutability, managed-memory test assumption) are quality/style concerns that do not affect correctness of the current implementation.

include/matx/kernels/sar_bp.cuh — CUDA_ARCH allowlist will need updating as new GPU generations are released.

Important Files Changed

Filename	Overview
include/matx/kernels/tensor_accessor.h	New TensorAccessor/BoundAccessor wrapper with correct fast/slow path dispatch; const operator() returns mutable T& by design, matching existing MatX patterns
include/matx/kernels/sar_bp.cuh	Extends shared-mem preamble to Float/Mixed, replaces scalar range_to_mcp with required MatX operator, plumbs TensorAccessor fast-path; CUDA_ARCH whitelist still forward-exclusive
include/matx/transforms/sar_bp.h	Adds unit-stride fast-path detection and dispatches IsUnitStride kernel instantiation; all existing callers already use tensor views for range_to_mcp
include/matx/kernels/channelize_poly.cuh	Adds Full filter smem layout, CTILE=32 SmemTiled variant, running filter_idx counter; TensorAccessor plumbed through all kernel entry points
include/matx/transforms/channelize_poly.h	Renamed to matx::detail::cpoly namespace, adds SmemTiled CTILE=32 dispatcher, Full filter layout selection, and num_channels=1 support; dispatch logic is correct
include/matx/core/tensor.h	Single-line change: adds MATX_DEVICE to Stride() to make it callable from device kernels
test/00_misc/TensorAccessorTests.cu	New host-side unit tests for TensorAccessor fast/slow paths; FastPathWriteThrough does host-side writes assuming managed memory
test/00_transform/ChannelizePoly.cu	Adds num_channels=1 regression test and updates existing tests for cpoly namespace rename
test/CMakeLists.txt	Adds TensorAccessorTests target to the build

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[sar_bp_impl / channelize_poly_impl] --> B{is_tensor_view_v for hot inputs?}
    B -- No --> C[IsUnitStride = false\nlaunch slow-path kernel]
    B -- Yes --> D{runtime: all last-dim strides == 1?}
    D -- No --> C
    D -- Yes --> E[IsUnitStride = true\nlaunch fast-path kernel]

    E --> F[TensorAccessor FastPath=true\ndata_ = op.Data\nouter_strides_ preloaded]
    C --> G[TensorAccessor FastPath=false\nforwards to op operator]

    F --> H["operator()(is...) → data_[offset]\nno stride reload in inner loop"]
    G --> I["operator()(is...) → op(is...)\ngeneric MatX evaluation"]

    subgraph channelize_poly dispatch
      J[num_channels ≤ 6 and D==M and real] --> K[FusedChan kernel]
      L[D==M and smem fits] --> M[Smem kernel]
      N[SmemTiled smem fits] --> O{num_channels ≤ 32?}
      O -- Yes --> P[SmemTiled CTILE=32]
      O -- No --> Q[SmemTiled CTILE=64]
      R[fallback] --> S[Generic kernel]
    end

Loading

_{Reviews (3): Last reviewed commit: "Remove assert() from ChannelizePoly1D_Sm..." | Re-trigger Greptile}

[tbensonatl]

/build

[tbensonatl]

/build

[cliffburdick]

Do we want to use assert in device code? I thought it has a first-time and register penalty.

[tbensonatl]

Good point. It should be removed with NDEBUG defined, but we can rely on the transform dispatch checks to keep this from impacting debug build performance too much.

[coveralls]

Coverage is 91.858% — add-tensor-accessor-optimize-sarbp-cpoly into main. No base build found for main.