I am having trouble using zero-copy memory. I read an article talking about using zero-copy memory for a first pass at converting a program to cuda. It sounded like a good idea but I am having a great deal of trouble.
Perhaps someone can comment on my understanding or misunderstanding of zero-copy memory.
My understanding is that after the following code has run, I will have one set of data that can be accessed from either the host or the GPU. The data is accessed on the host by referencing *TF and the data is accessed on the device by referencing *devPtrTF.
Is this correct?
if (cudaHostAlloc( (void**)&TF,sizeof( double ) * Nz * Ny *( Nx+2L),cudaHostAllocMapped))
{ printf("\n CUDA error: cudaHostAlloc TF error", xerror);
return 1;
}
if(cudaHostGetDevicePointer(&devPtrTF,TF,0))
{ printf("\n CUDA error: cudaHostGetDevicePointer TF error", xerror);
return 1;
}
Also, I think the following would be an equivalent call.
TF = (double *)malloc(15L*(Nx+2L)*Ny*Nz*sizeof(double));
if (cudaHostRegister( (void**)&TF,sizeof( double ) * Nz * Ny *( Nx+2L),cudaHostRegisterMapped))
{ printf("\n CUDA error: cudaHostAlloc TF error", xerror);
return 1;
}
if(cudaHostGetDevicePointer(&devPtrTF,TF,0))
{ printf("\n CUDA error: cudaHostGetDevicePointer TF error", xerror);
return 1;
}
Is this also correct?
Thanks for any help with this.