Repository navigation
Track cuTile Rust v0.5.0 debugging candidates #1606
Description
Track the cuTile debugging candidates for v0.5.0, including the sub-issues each PR actually covers and the evidence still needed before merging. Merge accepted CUDA Rust changes into main, then cherry-pick the release candidates into release/v0.5.x.
Status updated 2026-10-22. #1602 is merged into main; release integration is in #1613. The other candidates remain open and require their own validation and review.
| Candidate | Issue coverage | Current assessment / next step |
|---|---|---|
#1385 (a450156da) |
#1379 frontend generated-constant fallback; related to #1475 | Arthur addressed the test feedback with exact Line/Full, cross-file/nested caller-chain, serialized-location and cubin checks; reports 16 debug tests passing including live CUDA-GDB. Review the new commit and obtain actual Tile CI results. Backend assertion traps remain separate. |
#1480 (77816f6d0) |
#1475 entry wrapper and wrapper-generated assumptions | Existing fail-before/pass-after, assembler and live CUDA-GDB evidence. Complete the optimized line-info coverage requested in review, then run Tile CI and obtain approval. |
#1528 (f4684741a) |
#1475 short-circuit constants and constant-folded expressions | Still draft and unchanged since the testing feedback. Needs folded-expression, Line/Full, cross-file/caller-chain, serialized-location and const-bool binding coverage, then CI/review. |
#1600 (765cb1f8b) |
#1428 unified debug mode and runtime JIT overrides | Sam added environment-isolated subprocess tests, independent cache-mode comparisons and normal-launcher GPU coverage. Review/validate the update, run CI and finish the v0.5.0 API migration/changelog/versioning treatment. |
#1602 (main merge 9718b4386) |
#1475 arrays, generated reductions, host-relocated/deduplicated check origins; existing device-check/assumption regressions | Merged into main on 2026-10-22 after all 25 applicable CI checks passed, including CodeQL. Local semver, CPU/assembler, exact source-location and live CUDA-GDB/GPU checks pass. Included in release integration PR #1613, pending combined validation and merge. |
#1613 combines the accepted #1602 debug fix with the tested #1516 → #1603 → #1604 host-runtime stack for release/v0.5.x, including the latest read-permission corrections. Its CI also validates PRs targeting release branches. The host-runtime main-branch PRs remain separate; unfinished debugging candidates and external backend/debugger fixes are not claimed by this release integration.
Missing CI on a candidate is not a passing result. In particular, the reported local successes on #1385 and #1600 still need verification and Tile CI; #1480's existing checks do not include the Tile CPU/assembler lanes.
Release integration checklist (check after merge and inclusion in release/v0.5.x):
- fix: correct cuTile source locations for generated constants #1385 generated-constant fallback
- fix(cutile): attribute generated entry setup to the user signature #1480 entry-wrapper source locations
- fix(cutile-compiler): attribute short-circuit boolean constants to expression span #1528 short-circuit/folded-expression locations
- cutile: one CodegenMode for device debug, plus CUTILE_JIT_* run-time overrides #1600 debug-mode/runtime override controls
- fix(cutile): retain source origins for generated code and host checks #1602 arrays/reductions/host-check origins
- Run the combined source-location and debug-mode regressions after integrating the accepted PRs; their changes meet in the same compiler paths.
External dependencies and uncovered scope:
- cuTile: breakpoint on an inlined helper call slides to the next statement #1476: NVIDIA/cuda-gdb#29, reviewed head
45d4455fb. The implementation has synthetic-DWARF and live pending-breakpoint evidence and is a strong candidate for debugger-maintainer review. No approving review or CI result is reported. Confirm the first CUDA-GDB release containing the fix; merging CUDA Rust PRs alone does not deliver it. Testing comment. - [BUG]: cuTile: at a
cuda_tile_assert!trap, cuda-gdb reports the module's first line and drops the inlined helper frames #1379: actual assertion SIGTRAP and reopened GPU-coredump line/frame attribution still needs the Tile IR backend fix and before/after live/core evidence. No backend fix PR is identified in the linked discussions. The frontend constant fix must not close this remaining scope. - cuTile: audit generated source locations (optional follow-ups to #1379) #1475: close checklist items only when their own evidence is complete. Keep backend trap lowering distinct from the frontend audit. Full-debug tileiras 13.3 also coalesces a reduction source-line entry despite correct frontend metadata; fix(cutile): retain source origins for generated code and host checks #1602 does not claim complete backend stepping accuracy.
#1602 local full-suite validation encountered the already documented warmup-cache ordering failure; its 16-test warmup suite passed serially. The semver failure was resolved in 825de3e38; the existing warmup-cache ordering issue is separate.
Activity