If compiling HiGHs from source, what additional steps are needed to make sure that GPU acceleration for cuPDLP-C works?

[benjaminplong]

We compile HiGHs from the src code and I'm wondering if there are special compilation steps or prerequisites to enable the GPU acceleration. I see an option for it in the CMakeLists file but is providing the option the only thing needed?

Bonus Question: Anyone experienced using the GPU Acceleration while running on an Arm platform? Will this work "normally" or are there additional steps required?

[jajhall]

Have you seen

https://ergo-code.github.io/HiGHS/stable/guide/gpu/#Build-HiGHS-with-GPU-support

We have no Arm GPU support

[benjaminplong]

I was not able to find this when searching. Thank you for linking.

Let me clarify. Not an ARM GPU, but compiling HiGHs for an ARM CPU but expecting to use the GPU acceleration as well.

[galabovaa]

We do have ARM CPU support, and for the CPU code the default settings should work.

We do not really have an arm machine with an nvidia GPU around, but if all works as expected, no additional configuration should be required.

Please let us know whether this works and if you run into any issues!

[benjaminplong]

Follow up question, what would the behavior be if the binary is compiled with GPU support, but it is deployed to a machine without a GPU? Will cuPDLP-c error? Will it run on the CPU instead of the GPU?

[jajhall]

Interesting question. I could speculate, but @galabovaa will have a much better view

[galabovaa]

There will be an error finding the cuda libraries at runtime.

The CPU code is only built if the GPU flags are off.

[samiran-kawtikwar]

In case its still useful, I was able to incorporate the HiGHS PDLP solver in my project by adding this including this HiGHS.cmake module:

set(CUPDLP_GPU       ON CACHE BOOL "Build pdlp with GPU" FORCE)
set(CUPDLP_FIND_CUDA OFF CACHE BOOL "Dont use FindCUDAConf for PDLP" FORCE)
set(BUILD_CXX_EXE    OFF CACHE BOOL "Don't build C++ executable" FORCE)
set(BUILD_TESTING    OFF CACHE BOOL "Don't build tests" FORCE)
set(BUILD_EXAMPLES   OFF CACHE BOOL "Don't build examples" FORCE)
set(CUPDLP_BUILD_EXAMPLES   OFF CACHE BOOL ""    FORCE)
set(CUPDLP_CUDA_BUILD_TESTS OFF CACHE BOOL ""    FORCE)

include(FetchContent)
FetchContent_Declare(
  HiGHS
  GIT_REPOSITORY https://github.com/ERGO-Code/HiGHS.git
  GIT_TAG        latest  # or pin a commit/tag
  EXCLUDE_FROM_ALL
)
FetchContent_MakeAvailable(HiGHS)
get_target_property(_HIGHS_INCLUDES highs::highs INTERFACE_INCLUDE_DIRECTORIES)

In the master CMake, I added:

include(ExternalProject)
include(HiGHS.cmake)
add_executable(main.exe ${SOURCE_FILES})
target_include_directories(main.exe SYSTEM PRIVATE ${_HIGHS_INCLUDES})
target_link_libraries(main.exe PRIVATE
  highs::highs      # from FetchContent
)

configuring and building with this will: Download and install highs into your existing project.
Note, you will need to specify the SOURCE_FILES and change CUPDLP_FIND_CUDA based on your preferences.

[galabovaa]

Great, thank you for sharing!