Publish passing benchmark results - #6
Conversation
The native vector width can be two doubles, which drops the z coordinate in nbody and breaks Mandelbrot lane access. Use fixed 256-bit vectors so both programs retain all four lanes.
Calling upstream generators exhausts the stack at large inputs. Yield before each request so the coroutine chain resumes without recursing through every stage on the same call stack.
plajjan
changed the title
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
One failed language blocks every result from reaching Pages. Publish verified languages and show which ones are missing. Keep failed jobs visible and reject empty or mixed-run results.
e8491da to
4bbfc1c
Compare
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e8491daacd
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".
|
|
||
| def collect_results(): | ||
| results = BENCH / 'build/_results' | ||
| available = sorted(p.name for p in results.iterdir() if p.is_dir()) if results.exists() else [] |
There was a problem hiding this comment.
When all jobs are re-run, artifacts from earlier attempts remain under the same workflow run, and a language overwrites its artifact only if it reaches the upload step. Because available accepts every downloaded result directory while verification checks the SHA and run ID but not the githubRunAttempt recorded in bench/tool/Program.cs:776, a language that fails before uploading in the current attempt can be silently published using measurements from a prior attempt. Namespace or filter artifacts by attempt, or otherwise verify that each selected artifact belongs to an acceptable attempt before deriving available.
Useful? React with 👍 / 👎.
GCC can use AVX-512 fused arithmetic even with -mno-fma. Disable expression contraction to keep the expected Mandelbrot output on these CPUs.
1 participant
A failed language currently blocks every result from reaching Pages. Publish the languages that pass and list missing languages on the site.
Fix container image lookup, C and C# CPU assumptions, and the TypeScript sieve stack overflow on the dedicated runner.