Hands-On GPU Computing with Python

Unified Memory Access (UMA)

This feature allows data pointers to allocate memory on the GPU device that can be accessed by both the CPU and the GPU. With this shared memory access, the need for CPU allocation for host variables is eliminated. The following declaration example differentiates UMA from conventional memory allocation:

Conventional CPU memory allocation	Unified memory allocation
char *data;data = (char *)malloc(N)	char *datacudaMallocManaged(&data, N)

 

By using cudaMallocManaged, we can allocate memory on the GPU that can be migrated between both the CPU and the GPU. So, with unified memory access, we can pass that data pointer to a CPU function, as well as a GPU kernel!