Do not perform copy for the phase space point passed from MadSpace and reshape of the underlaying array. Use the passed pointer and unify the layout.
We are already able to hide the latency with cudaMallocAsync so memory allocator test on the MadMatrix side did not improve anything. Also verified that same address is reused on cuda. The gain would come purely from not storing the data twice.
Do not perform copy for the phase space point passed from MadSpace and reshape of the underlaying array. Use the passed pointer and unify the layout.
We are already able to hide the latency with
cudaMallocAsyncso memory allocator test on the MadMatrix side did not improve anything. Also verified that same address is reused on cuda. The gain would come purely from not storing the data twice.