[galabovaa]

Description

Follow-up to #3293, incorporating the review discussion there.

What the code here introduces, largely following the previous PR:

• Check that OpenBLAS actually runs in serial

Before solving, HiPO sets the number of OpenBLAS threads to 1. Previously this was a silent no-op if the build was misconfigured: if HiGHS was linked against OpenBLAS but HIPO_USES_OPENBLAS was not defined, the call did nothing and OpenBLAS ran multi-threaded, which slows HiPO down considerably.

Now highs_openblas_set_num_threads returns the number of threads OpenBLAS reports after the call (-1 if HiGHS was built without OpenBLAS support). Before solving, HiPO checks this against the BLAS provider reported in the build metadata: if the provider is OpenBLAS but the thread count is not 1, HiPO logs an error and returns kSolveError instead of running slowly without warning. Calling it with 0 or less just returns the current number of threads without changing it (used by the tests).

This check is skipped when hipo_serial_blas is false (see below).

• Detect OpenBLAS from BLAS_LIBRARIES

If the library name contains "openblas", the BLAS metadata is set to OpenBLAS, so HIPO_USES_OPENBLAS is defined.

• Opt-out

New option hipo_serial_blas (default true). The number of OpenBLAS threads is process-wide, so users running other OpenBLAS work in the same process can set it to false. HiPO then neither changes the number of OpenBLAS threads nor checks it.

• Thread safety

Calls are serialised with a mutex, and openblas_set_num_threads is only called when the thread count actually needs to change, so concurrent HiGHS instances don't reconfigure OpenBLAS while another is running BLAS routines.

• Provider checks

Added a portable, case-insensitive is_substring_case to stringutil.h, used for the OpenBLAS and Apple Accelerate checks.

• Docs

New "BLAS threads" section in [parallel.md](http://parallel.md/), the option definition, and notes in the installation, Python and Julia pages.

• libblastrampoline

HiGHS.jl uses libblastrampoline, so HiPO does not change the number of BLAS threads there. We should probably document on the JuMP side that users should call BLAS.set_num_threads(1) to get the best performance out of HiPO, noting that this applies to the whole Julia session, so it may slow down other code that relies on parallel BLAS. cc @odow

• Tests

Unit tests for is_substring_case and the OpenBLAS thread wrapper, and a HiPO test checking that OpenBLAS runs in serial by default and is left alone when hipo_serial_blas is false.

Checklist

• I have read the contributing guidelines
• This PR targets the latest branch
• Tests are passing
• Documentation was updated where relevant
• This PR is not primarily AI-generated (per the AI contributions policy in CONTRIBUTING.md)

[codecov]

Codecov Report

❌ Patch coverage is 53.06122% with 23 lines in your changes missing coverage. Please review.
✅ Project coverage is 75.19%. Comparing base (b0f37ed) to head (d76287b).
⚠️ Report is 1 commits behind head on latest.

Files with missing lines	Patch %	Lines
highs/ipm/IpxWrapper.cpp	0.00%	10 Missing ⚠️
check/TestHighsExternalDeps.cpp	59.09%	9 Missing ⚠️
extern/HighsExtrasExternalDeps.h	0.00%	2 Missing ⚠️
highs/ipm/hipo/ipm/FactorHighsSolver.cpp	0.00%	1 Missing ⚠️
highs/lp_data/HighsOptions.h	75.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           latest    #3402      +/-   ##
==========================================
- Coverage   75.19%   75.19%   -0.01%     
==========================================
  Files         460      460              
  Lines      112421   112464      +43     
  Branches    17826    17828       +2     
==========================================
+ Hits        84540    84569      +29     
- Misses      27599    27613      +14     
  Partials      282      282              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[filikat]

Thanks Ivet, this looks great.
The only thing I would notice is that we are claiming multiple times in the documentation that HiPO runs better when BLAS is serial. This is definitely true for OpenBLAS, but I don't know if it's true for other BLAS libraries, like MKL.

[galabovaa]

@filikat I updates the sections to specify OpenBLAS. Would you please have another look? Thank you!