CUDA programming before the introduction of Unified Memory

For completeness sake, I like to discuss how to write CUDA programs without using Unified Memory
Recall: the host memory and device memory are separate memory systems:
CUDA programs before the introduction of Unified Memory:
Uses the host function cudaMemcopy( ) to transfer data between CPU (host) memory and GPU (device) memory

Allocating host (CPU) memory and device (GPU) memory

Before we can perform computations:
a program must first allocate memory to store its data in variables

How to allocate variable variables in the host (CPU) memory:

dataType *x = (dataType *) malloc( NBytes ); // CS255 material Example: allocate an array of 1000 doubles in CPU memory: double *x = (double *) malloc( 1000*sizeof(double) );

How to allocate variable variables in the device (GPU) memory:

dataType *d_x; // Device variables are prefixed with d_ cudaMalloc(&d_x, NBytes ) ; Example: allocate an array of 1000 doubles in GPU memory: double *d_x; cudaMalloc( &d_x, 1000*sizeof(double) );

Transfer data between host (CPU) memory and device (GPU) memory

The cudaMemcopy( ) function:

cudaMemcpy(to, from, Nbytes, MODE): if MODE = cudaMemcpyHostToDevice: transfers Nbytes from host memory to device memory if MODE = cudaMemcpyDeviceToHost: transfers Nbytes from device memory to host memory

How to use cudaMemcopy( ) with GPU kernel functions:

(1) Transfer input data from host memory to device memory: cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice); (2) Execute the kernel function on the device data: vectorAdd<<<B, T>>>(d_x, d_y, d_z, N); (3) Transfer output data from device memory to host memory: cudaMemcpy(z, d_z, N*sizeof(float), cudaMemcpyDeviceToHost);