Support of AMD GPUs

[amontoison]

Adds support for running the first-order PDLP solver on AMD GPUs via HIP/ROCm, alongside the existing NVIDIA/CUDA backend.

The vendor-specific GPU code is refactored behind a thin portability layer so the same sources compile for both, selected at configure time.

• AMD (HIP / ROCm)

cmake -S . -B build -DHIPDLP_HIP=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target highs --parallel

• NVIDIA (CUDA) - unchanged

cmake -S . -B build -DCUPDLP_GPU=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target highs --parallel

@jajhall I got the green light from AMD to help for the maintenance of the HIP / ROCm backend in HiGHS.

I added a workflow to test the compilation with AMD libraries, it passed:
https://github.com/ERGO-Code/HiGHS/actions/runs/27993664916/job/82851309388?pr=3083

[codecov]

Codecov Report

❌ Patch coverage is 77.55102% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 73.26%. Comparing base (f1de972) to head (e26d31f).
⚠️ Report is 928 commits behind head on latest.

Files with missing lines	Patch %	Lines
examples/call_highs_hipdlp.cpp	77.55%	11 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           latest    #3083      +/-   ##
==========================================
+ Coverage   73.16%   73.26%   +0.10%     
==========================================
  Files         430      446      +16     
  Lines      103967   108156    +4189     
  Branches    16719    17329     +610     
==========================================
+ Hits        76067    79245    +3178     
- Misses      27624    28635    +1011     
  Partials      276      276              

☔ 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.

[jajhall]

Many thanks @amontoison!

We'll leave this for v1.16, so we don't introduce too many changes into v1.15. That said, we expect v1.16 to follow fairly soon - once users have had the chance to experiment with the parallel MIP solver (which will not be the default). Hence v1.16 will have parallel MIP by default, and AMD GPU!

[galabovaa]

Hi @amontoison, Would it be possible for you to merge latest in, and give it another quick test on AMD, so we can merge the PR?

[amontoison]

Hello @galabovaa, I can do that next Tuesday (25/08/2026), is it fine for you or do you plan to do the new release before?

[galabovaa]

Tuesday is fine, we don't plan to do a release in the next few weeks so no rush! Whenever you have some spare time. Thank you!

[amontoison]

@galabovaa
I added an example and documentation. 💪
I will test / validate on various AMD architectures next week.

The PR is ready for a review.

Off-topic -- I have these compilation warnings with amdclang++:

/home/montoison/HiGHS/highs/model/HighsHessian.cpp:677:63: warning: variable 'index' is uninitialized when passed as a const pointer argument
      here [-Wuninitialized-const-pointer]
  677 |   return this->call_(kHessianOracleCallTypeProduct,[ 28%] Building CXX object highs/CMakeFiles/highs.dir/pdlp/CupdlpWrapper.cpp.o
 &num_nz, &index, &value,
      |                                                               ^~~~~
/home/montoison/HiGHS/highs/model/HighsHessian.cpp:677:71: warning: variable 'value' is uninitialized when passed as a const pointer argument
      here [-Wuninitialized-const-pointer]
  677 |   return this->call_(kHessianOracleCallTypeProduct, &num_nz, &index, &value,
      |