Update sarbp example default precision and tiling heuristics

[tbensonatl]

Update the sarbp example default precision from mixed to taylor_fast. taylor_fast is faster than mixed in all cases and in most cases will provide sufficient accuracy.

Also adjust the automatic pulse blocking and image tiling heuristics. Previously, the example automatically tuned the pulse block size, but not the image tile size. For large scenes on GPUs with small L2 caches, optimal performance requires tuning image tiling factors in addition to pulse block sizes. The example now automatically tunes both by default. On DGX Spark, this tuning improves performance for the dataset used in the example by ~10%.

[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 changes the SAR backprojection example’s default precision to taylor_fast and introduces joint automatic selection of pulse-block and image-tile sizes based on an estimated L2 working set.

• Adds a standalone cache heuristic supporting automatic and manual overrides.
• Estimates the active range-bin span from scene geometry.
• Updates CLI parsing, diagnostics, defaults, and documentation.
• The heuristic’s boundary behavior currently lacks focused automated coverage.

Confidence Score: 4/5

The PR appears safe to merge, with a non-blocking recommendation to add focused tests for the new heuristic.

The inspected execution paths preserve valid block and tile bounds and do not alter image coverage; the only accepted concern is the maintenance risk from leaving the new boundary-sensitive heuristic untested.

Files Needing Attention: examples/sarbp/sarbp_heuristic.h

Important Files Changed

Filename	Overview
examples/sarbp/sarbp.cu	Integrates automatic image tiling, active-range estimation, the new precision default, and heuristic diagnostics.
examples/sarbp/sarbp_heuristic.h	Implements cache-based joint block and tile selection correctly for inspected paths, but without focused automated tests.
examples/sarbp/README.md	Documents the new defaults, joint tuning behavior, and revised benchmark methodology.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Load SAR input and scene geometry] --> B[Estimate active range-bin span]
    B --> C[Read GPU L2 cache size]
    C --> D{Block size automatic?}
    C --> E{Image tiles automatic?}
    D --> F[Select pulse block size]
    E --> G[Select image tile count]
    F --> H[Process pulse blocks]
    G --> H
    H --> I[Backproject each image tile]
    I --> J[Write focused image]

Loading

_{Reviews (1): Last reviewed commit: "Fix working set image tiling approximati..." | Re-trigger Greptile}

[tbensonatl]

/build

[coveralls]

Coverage is 93.386% — tbenson/sarbp-example-include-image-tiling-in-heuristics into main. No base build found for main.