Calling cuda/rocm libraries from openmp

[lucaparisi91]

• [x] call cuda kernel using an openmp mapped variable
• [x] use openmp with a device variable
• [ ] call cuda/rocm library from an openmp context
• [ ] fortran examples

[lucaparisi91]

Pointers Translation

OpenMP to HIP/CUDA

You can get the device address of a host variable mapped to the device. An example is

double * x = ..... // allocate x on host
#pragma omp target data map(to:x[0:count]) 
{
    #pragma omp target data use_device_addr(x[:,0])
   { 
       // launch hip/cuda kernel here
       launch_my_hip_kernel(x)    
   }
}

HIP/Cuda to OpenMP

double * x_device;
allocate_and_copy_from_host_to_device() // cuda/hip code

#pragma omp target is_device_ptr(x_device)
{
  // code using x_device here
}