OpenCL, Is it possible write directly to local memory?

Question

I am new to OpenCL. Now I trying to understand the local and shared memory. I am doing a sample code, there I am trying to upload a filter of size 4x4 into local memory from CPU, but I am getting the error -52 CL_INVALID_KERNEL_ARGS. The sample code:

void __kernel filter( const __global float* in, __global float* out, const __local float* coeff )
{
    //filtering
}

CPU part:

ret = kernel.setArg(0, input);
ret = kernel.setArg(1, out);
ret = kernel.setArg(2, 16*sizeof(float), &coeff );

the enqueueNDRangeKernel will return -52 error.

If I give NULL instaed of "coeff", no error

ret = kernel.setArg(0, input);
ret = kernel.setArg(1, out);
ret = kernel.setArg(2, 16*sizeof(float), NULL );

Is it possible to write directly to local memory?, I did not find any help in OpenCL books. Please share your experience, if anybody tried this.

Thanks

Answer 1

No, the host CPU does not have the ability to directly access local memory in this manner. All local memory allocations must be initialised inside a kernel running on the device.

In my experience, small filters for image processing are usually better off in constant memory - simply allocate a buffer as you would for global memory but declare the argument with __constant instead of __global . If possible, you may see further performance improvements by embedding your filter into the OpenCL kernel source as a compile-time constant, eg:

__constant float filter[] = {
  -0.000000f, -0.055556f, -0.055556f, 
  -0.055556f, -0.111111f, -0.222222f, 
  -0.222222f, -0.277778f, -0.444444f, 
};