As shown, there is a code in pyfr to determine whether to use cuda aware and then determine whether to send the data to the host, is this necessary? When cuda aware can’t send data directly between GPUs, won’t it automatically send the data back to the host?
Not all MPI libraries are CUDA aware. If you pass a CUDA pointer to these libraries PyFR will crash, hence the need to disable the functionality by default.
Regards, Freddie.
Thank you, what I’m puzzled about is that on the Cuda platform alone, we have found that regardless of whether the hardware supports it or not, we can use Cuda ware, which means that pointers can be moved from device to device
It is not a hardware thing but an MPI thing. If it works or not depends on if MPI has been compiled with CUDA awareness.
Regards, Freddie.