perf(autoware_tensorrt_plugins): remove Thrust from sort kernels

[mojomex]

Stack

This PR is stacked on #12561, which adds the reference kernel tests and the minimal pre-existing unique counts fix those tests expose. Review #12561 first; this PR contains the no-thrust implementation and cleanup on top.

Summary

Removes Thrust from the TensorRT plugin sort kernels and keeps the mutable unique temp-storage follow-up used in the benchmarked variant.

This draft PR corresponds to the benchmarked variant ptv3-t18-no-thrust-c8f76ed-20260506.
All PRs in this cohort target main; each later PR contains the changes benchmarked in the earlier ones.

Cohort

• perf(autoware_tensorrt_plugins): remove Thrust from sort kernels (ptv3-t18-no-thrust-c8f76ed-20260506)
• perf(autoware_tensorrt_plugins): keep SegmentCSR allocation-free (ptv3-t18-no-thrust-no-alloc-e9515b790-20260506)
• perf(autoware_tensorrt_plugins): keep plugin outputs stream-ordered (ptv3-t18-no-thrust-no-alloc-no-sync-13f3672a0-20260506)
• perf(autoware_tensorrt_plugins): cache max sort workspace bounds (ptv3-t18-no-thrust-no-alloc-no-sync-maxnumel-47bf5656f-20260506)
• fix(ptv3): set TensorRT max auxiliary streams to 1 (ptv3-t18-no-thrust-no-alloc-no-sync-maxnumel-maxauxstreams1)

Benchmarks

Source report: reports/2026-05-07_22-20-12/report.md

Total Latency Summary

Variant	Measurement	CPU mean (ms)	CPU p95 (ms)	CPU faster vs baseline	GPU mean (ms)	GPU p95 (ms)	GPU faster vs baseline	Mean voxels
ptv3-t18	series (7 x 50)	29.117	31.755	+0.0%	28.020	30.598	+0.0%	120572
ptv3-t18-no-thrust-c8f76ed-20260506	series (7 x 50)	27.559	28.502	+5.7%	26.416	27.306	+6.1%	120572
ptv3-t18-no-thrust-no-alloc-e9515b790-20260506	series (7 x 50)	26.874	28.863	+8.3%	25.741	27.665	+8.9%	120572
ptv3-t18-no-thrust-no-alloc-no-sync-13f3672a0-20260506	series (7 x 50)	26.278	27.471	+10.8%	25.140	26.288	+11.5%	120572
ptv3-t18-no-thrust-no-alloc-no-sync-maxnumel-47bf5656f-20260506	series (7 x 50)	26.214	26.642	+11.1%	25.084	25.488	+11.7%	120572
ptv3-t18-no-thrust-no-alloc-no-sync-maxnumel-maxauxstreams1	series (7 x 50)	25.378	25.984	+14.7%	24.230	24.792	+15.6%	120572
ptv3-t18-no-thrust-no-alloc-no-sync-maxnumel-maxauxstreams3	series (7 x 50)	26.809	28.083	+8.6%	25.614	26.849	+9.4%	120572

Relative Performance

[github-actions]

Thank you for contributing to the Autoware project!

🚧 If your pull request is in progress, switch it to draft mode.

Please ensure:

• You've checked our contribution guidelines.
• Your PR follows our pull request guidelines.
• All required CI checks pass before marking the PR ready for review.

[mojomex]

@codex review. Is the implementation memory safe and has no race conditions? Do the unit tests guarantee correct behavior? Is the implementation as concise and readable as possible?

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Can't wait for the next one!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[mojomex]

@codex Code isn't documented and behavior is not immediately obvious to the reader, e.g. write_unique_offset_sentinel. Document function contracts/pre/post conditions concisely, annotate argsort and unique flows so user can easily follow. For design choices, e.g. where to put sentinels etc., explain the "why".

[chatgpt-codex-connector]

To use Codex here, create an environment for this repo.