Fundamentals of GPU programming (3)

CUDA program structure

use CUDA API to query for compatible devices

cudaError_t cudaGetDeviceProperties(cudaDevProp* prop, int dev)

Example: 

int device=1;
cudaDeviceProp props;
cudaGetDeviceProperties(&props, device)

allocate memory on the CUDA device

cudaError_t cudaMalloc(void** devPtr, size_t size)

Example:

size_t E_bytes = N_pts * 3 * sizeof(std::complex<double>);
HANDLE_ERROR(cudaMalloc((void**)&gpu_E, E_bytes));

copy data from main memory to the CUDA device

cudaError_t cudaMemcpy(void* dest, const void* source, size_t size, cudaMemcpyKind type)

Example:

HANDLE_ERROR(cudaMemcpy((void*)gpu_F, (void*)F, F_bytes, cudaMemcpyHostToDevice));

launch the kernel

(To be added in future blog...)

copy data from device to main mamory

Example:

HANDLE_ERROR(cudaMemcpy((void*)E, (void*)gpu_E, E_bytes, cudaMemcpyDeviceToHost));

More explanation is shown below:

Device Properties

Get the number of devices:

cudaError_t cudaGetDeviceCount(int* num) 

Control what devices execute code

cudaError_t cudaGetDevice(int* d)			// get the current device
cudaError_t cudaSetDevice(int d)			// send future commands to this device

Error Checking

Almost all CUDA functions return a cudaError_t value. We need to check it everytime to see if the CUDA operation works or not.

cudaError_t = 
	cudaSuccess
	cudaErrorMemoryAllocation		// not enough memory
	cudaErrorInvalidValue			// a parameter is wrong
	cudaErrorInvalidDeviceProinter	// point wrong, didn't allocate memory maybe

A useful macro

HANDLE_Error is defined in STIM_CUDA_ERROR_H. It allows to display a string describe the error, along with the file and line number where error occured.

#define HANDLE_ERROR( err ) (cuHandleError( err, __FILE__, __LINE__ ))
static void cufftHandleError( cufftResult err, const char*file, int line )
{
    if (err != CUFFT_SUCCESS)
    {
        if(err == CUFFT_INVALID_PLAN)
            std::cout<<"The plan parameter is not a valid handle."<<std::endl;
        else if(err == CUFFT_ALLOC_FAILED)
            std::cout<<"Allocation failed."<<std::endl;
        else if(err == CUFFT_INVALID_VALUE)
            std::cout<<"At least one of the parameters idata, odata, and direction is not valid."<<std::endl;
        else if(err == CUFFT_INTERNAL_ERROR)
            std::cout<<"An internal driver error was detected."<<std::endl;
        else if(err == CUFFT_EXEC_FAILED)
            std::cout<<"CUFFT failed to execute the transform on the GPU."<<std::endl;
        else if(err == CUFFT_SETUP_FAILED)
            std::cout<<"The CUFFT library failed to initialize."<<std::endl;
        else
            std::cout<<"Unknown error: "<<err<<std::endl;

}

}