Normal annotated functions are not working with vikunja::reduce and the CUDA backend

alpaka-group / vikunja

Vikunja is a performance portable algorithm library that defines functions operating on ranges of elements for a variety of purposes . It supports the execution on multi-core CPUs and various GPUs. Vikunja uses alpaka to implement platform-independent primitives such as reduce or transform.

Mozilla Public License 2.0

14 stars 5 forks source link

If I use an annotated function together with vikunja::reduce it is only working with CPU backends but not the CUDA backend (HIP not testet). It compiles, but I get an runtime error.

template<typename TData>
ALPAKA_FN_HOST_ACC TData sum(TData const i, TData const j)
{
    return i + j;
}

//...

P::m_result = vikunja::reduce::deviceReduce<typename P::Acc>(
                        P::devAcc,
                        P::devHost,
                        P::P::queue,
                        P::m_size,
                        alpaka::getPtrNative(P::m_device_mem),
                        sum<Data>);

terminate called after throwing an instance of 'std::runtime_error'
  what():  /home/sehrig/application/miniconda3/envs/vikunja-dev/include/alpaka/mem/buf/BufUniformCudaHipRt.hpp(101) 'cudaFree(reinterpret_cast<void*>(memPtr))' returned error  : 'cudaErrorInvalidPc': 'invalid program counter'!

If I replace the function with a lambda, everything is fine.

alpaka-group / vikunja

Normal annotated functions are not working with vikunja::reduce and the CUDA backend #41