CUDA - typical program structure

Global variables declaration

• __host__

• __device__

• __global__

• __constant__

• texture

Function prototypes

• __global__ void kernelOne(...)

• __device__ / __host__ float handyFunction(...)

Main()

• allocate memory space on the device - cudaMalloc(&d_GlobalVarPtr, bytes)

• transfer data from host to device - cudaMemCpy(d_GlobalVarPtr, h_GlobalVa...)

• execution configuration setup

• kernel call - kernelOne<<<execution configuration>>>(args ...);

• transfer results from device to host - cudaMemCpy(h_GlobalVarPtr, d_Global...)

• free memory space on device - cudaFree(d_GlobalVarPtr);

Kernel - void kernelOne(type args,...)

• variables declaration:

  • __shared__

  • automatic variables transparently assigned to registers or local memory

• __syncthreads()...

__global__ void cudaKernel(float *a, float *b, float *c)
{
    int tID = some_mean_to_identify_thread_index;
    c[tID] = a[tID] + b[tID];
}



int main(int argc, char* argv[]) {
    /* Allocate and initialize vector a,b and c on both CPU and GPU */
    /* Data transfer for copying input vectors a and b on GPU. */
    cudaKernel<<<some_environment_info>>>(a, b, c);
    /* Data transfer for copying output vector c on CPU */
    /* Free memory. */
    return 0;
}