cuda: build.rs add trigger#8708
Conversation
Signed-off-by: mag1cian <mag1cian@icloud.com>
|
@mag1c1an1 Good catch, thank you! |
Merging this PR will not alter performance
|
| Mode | Benchmark | BASE |
HEAD |
Efficiency | |
|---|---|---|---|---|---|
| ❌ | Simulation | rebuild_naive |
91.4 µs | 101.6 µs | -10.04% |
| ⚡ | Simulation | bitwise_not_vortex_buffer_mut[128] |
273.6 ns | 244.4 ns | +11.93% |
Tip
Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.
Comparing mag1c1an1:cuda (71eeeed) with develop (1bcaf3f)
Footnotes
-
4 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩
rebuild_naive flipped between ~91us and ~105us (reported as -10% to -16% regressions or +15% improvements depending on the baseline) on about ten recently merged PRs that never touched vortex-zstd or ListView rebuild, including a docs-only Spark PR (#8728) and a build.rs-only CUDA PR (#8708). The benchmark cannot be stabilized: its trace is dominated by zstd's internal window/output copies (glibc memcpy resolved per runner image via ifunc), not by Vortex code. It is also of little value: the fixture is degenerate - a 4-string dictionary with all-zero offsets - and the rebuild paths it nominally covers are benchmarked properly across element types and list shapes in vortex-array/benches/listview_rebuild.rs. Remove it rather than keep a noise source; the divan dev-dependency goes with it since this was the crate's only benchmark. Signed-off-by: Claude <noreply@anthropic.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T6PPcdrqcNeUkfi1EGd4oC
rebuild_naive flipped between ~91us and ~105us (reported as -10% to -16% regressions or +15% improvements depending on the baseline) on about ten recently merged PRs that never touched vortex-zstd or ListView rebuild, including a docs-only Spark PR (#8728) and a build.rs-only CUDA PR (#8708). The benchmark cannot be stabilized: its trace is dominated by zstd's internal window/output copies (glibc memcpy resolved per runner image via ifunc), not by Vortex code. It is also of little value: the fixture is degenerate - a 4-string dictionary with all-zero offsets - and the rebuild paths it nominally covers are benchmarked properly across element types and list shapes in vortex-array/benches/listview_rebuild.rs. Remove it rather than keep a noise source; the divan dev-dependency goes with it since this was the crate's only benchmark. Signed-off-by: Claude <noreply@anthropic.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T6PPcdrqcNeUkfi1EGd4oC
build.rscaches stale PTX when nvcc availability changes (norerun-if-env-changed=PATH) #8707