Skip to content

Reduce ordinary numerical test runtime and Windows LDC coverage memory - #305

Merged
jmh530 merged 3 commits into
libmir:masterfrom
jmh530:codex/discrete-reference-ci
Nov 18, 2026
Merged

jmh530 merged 3 commits into
libmir:masterfrom
jmh530:codex/discrete-reference-ci

Conversation

@jmh530

@jmh530 jmh530 commented Nov 17, 2026 •

Copy link
Copy Markdown
Contributor

Reduce ordinary CI work by moving broader extended-precision log-probability reference cases to the optimized slow-numerics suite, and reduce Windows LDC coverage build memory by compiling descriptive and remaining tests separately.

Numerical reference tests

Retain representative precision and boundary checks in ordinary CI while moving broader binomial-log Fp!128, Poisson-log Fp!128/Fp!256, and four costly negative-binomial Fp!256/Fp!512 reference cases to slow numerics. Update the generators to reproduce the partition without removing reference cases.

Windows LDC coverage

Match the existing Linux descriptive/remaining grouping with two sequential package builds. Preserve relative coverage paths, reuse dependency selections, and retain resource logging for each group. Propagate build failures, reject missing, empty, or duplicate reports, and clean up temporary build directories.

Validation

  • Validated the numerical changes with focused DMD/LDC tests, optimized slow numerics, 32-bit DMD checks, and local coverage comparisons.
  • The local Windows split preserved all 111 coverage reports and every previously covered line. All 90 test modules passed their assertions; both baseline and split runs subsequently encountered the known local coverage shutdown failure.
  • The local split experiment reduced elapsed time from 237 to 182 seconds and peak compiler working memory from 18.1 to 12.0 GiB.
  • CI passed after the split, with 29 descriptive and 61 remaining test modules passing and all 111 reports collected.
  • Windows LDC coverage time fell from approximately 410 to 255 seconds between CI runs. Highest sampled compiler private memory fell from 18.1 to 10.9 GiB. These runs used different runner CPU models and master revisions, so the comparison is indicative rather than a controlled benchmark; memory measurements were sampled every 30 seconds.

Related to #157.

AI-Assisted-By: Codex (GPT-6 Astra Medium)

Retain representative precision and boundary checks in ordinary CI
while moving the remaining binomial-log Fp!128 and Poisson-log
Fp!128 and Fp!256 reference cases to the slow-numerics suite.

Update the generators to reproduce the partition and preserve all
454 existing reference rows. Local validation confirms that the
optimized slow suite executes the moved cases and ordinary Linux
coverage retains the covered production lines.

AI-Assisted-By: Codex (GPT-6 Astra Medium)
@codecov

codecov Bot commented Nov 17, 2026

Copy link
Copy Markdown

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

Run the Fp!256 and Fp!512 reference cases for 57 failures before
19 successes and 7 failures before 12 successes in the slow-numerics
suite. Retain the same inputs at lower precisions and the remaining
boundary cases at all four precisions in ordinary CI.

Update the generator to preserve all 36 reference rows across the
32 ordinary and four slow cases. Validate both suites on 32-bit DMD,
the optimized slow-numerics DUB job on LDC, and ordinary Linux
coverage after the move.

AI-Assisted-By: Codex (GPT-6 Astra Medium)
@jmh530
jmh530 marked this pull request as ready for review November 17, 2026 22:46
Compile descriptive tests and the remaining tests in sequential
package builds, matching the Linux coverage groups. Preserve
relative report paths, dependency selections, and resource logging.

Collect reports for the existing upload steps, reject missing,
empty, or duplicate reports, propagate build failures, and clean
up temporary build directories.

The local experiment preserved all 111 reports and covered lines,
while reducing peak compiler memory from 18.1 to 12.0 GiB.
All 90 test modules passed before the known local coverage
shutdown failure. Validate the complete workflow in CI.

AI-Assisted-By: Codex (GPT-6 Astra Medium)
@jmh530 jmh530 changed the title Reduce ordinary CI time for extended-precision log reference tests Reduce ordinary numerical test runtime and Windows LDC coverage memory Nov 18, 2026
@jmh530
jmh530 merged commit b5382cf into libmir:master Nov 18, 2026
23 checks passed
@jmh530
jmh530 deleted the codex/discrete-reference-ci branch November 18, 2026 03:29
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant