Partial Comparison:
DAXPY in CUDA versus OpenCL

// DAXPY in CUDA (from file: daxpy.cu)
void doDAXPY(double a, const double* xHost, const double* yHost,
             int n, double* zHost)
{
    … (including a series of calls to create and initialize GPU buffers)
    daxpy<<<nblocks, threadsPerBlock>>>(a, xGPU, yGPU, n, zGPU);
    …
}

// DAXPY in CUDA (from file: daxpy.cu)
__global__
void daxpy(double a, const double* x, const double* y,
           int n, double* z)
{
    int i = blockIdx.x*blockDim.x + threadIdx.x;
    if (i < n)
        z[i] = a*x[i] + y[i];
}

// DAXPY in OpenCL (from file: daxpy.c++)
void doDAXPY(double a, const double* xHost, const double* yHost,
             int n, double* zHost)
{
    … (including a series of calls to create and initialize GPU buffers)
    … (including a series of "clSetKernelArg" calls to bind parameters), e.g.:
    clSetKernelArg(kernel, 0, sizeof(double), &a);
    clSetKernelArg(kernel, 1, sizeof(cl_mem), &bufferX);
    // ... remainder of the clSetKernelArg calls ...
    clEnqueueNDRangeKernel(cmdQueue,
        kernel, // object created by clCreateKernel
        workDimension, // here it is 1
        globalWorkOffset, // []; nullptr ==> none
        globalWorkSize, // []; { ceil(n/localWorkSize)*localWorkSize }
        localWorkSize, // []; here {threadsPerBlock}; can be nullptr
        numEvents, // here 0 ==> none
        eventWaitList, // nullptr==> none
        event); // nullptr==> do not generate one)
    …
}

// DAXPY in OpenCL (from file: daxpy.cl)
__kernel
void daxpy(double a, __global const double* x, __global const double* y,
           int n, __global double* z)
{
    int i = get_global_id(0);
    if (i < n)
        z[i] = a*x[i] + y[i];
}