ROCm
Load rocSOLVER with dlopen at runtime
Don't have access to a windows machine so this hasn't been tested on windows yet.
See https://github.com/ROCm/hipBLAS/issues/896 for the request. Similar changes for hipSOLVER have been done at hipSOLVER #302
This change allows hipBLAS to build without rocSOLVER available at buildtime. By default, hipBLAS will try to find rocSOLVER at runtime using dlopen(). If rocSOLVER cannot be found, the rocsolver-based functions will return HIPBLAS_STATUS_NOT_IMPLEMENTED.
The -n flags has been changed to mean "don't build rocSOLVER tests".
The --solver-buildtime flags has been added to find rocSOVLER at build time.
Also added some flags for jenkins so will be testing the various flags on CI.
Is the static build without solver_buildtime the default as that creates an invalid static library?
It's valid. It's just unlikely that you'd want a static hipblas and a dynamic rocsolver.
You may want to adjust the CMake that defines the package dependencies to recommend and/or suggest the rocsolver package to users that install hipblas.
Is the static build without solver_buildtime the default as that creates an invalid static library?
It's valid. It's just unlikely that you'd want a static hipblas and a dynamic rocsolver.
Invalid in the sense that neither a customer nor developer would want it (we don't want failing tests for bad configs), let alone be default. Maybe a conditional option should be used?
Invalid in the sense that neither a customer nor developer would want it (we don't want failing tests for bad configs), let alone be default. Maybe a conditional option should be used?
Keep in mind that we're not distinguishing between 'maybe rocsolver at runtime' and 'no rocsolver'. The configuration of 'static hipblas with no rocsolver' is something that developers may want.
Invalid in the sense that neither a customer nor developer would want it (we don't want failing tests for bad configs), let alone be default. Maybe a conditional option should be used?
Keep in mind that we're not distinguishing between 'maybe rocsolver at runtime' and 'no rocsolver'. The configuration of 'static hipblas with no rocsolver' is something that developers may want.
The flags are confusing me maybe keeping -n --no-solver as never any solver is simpler (both shared and static). What this seems to be is an architeture option to delays .so loading to make it optional. Should discuss an overall roadmap feature request for this change for hip libraries.
Unable to import due to merge conflict please reopen this PR on ROCm/rocm-libraries
