Question

in my C OpenCL code I use clSetKernelArg to create ‘variable size’ __local memory for use in my kernels, which is not available in OpenCL per se. See my example:

clSetKernelArg(clKernel, ArgCounter++, sizeof(cl_mem), (void *)&d_B);
...
clSetKernelArg(clKernel, ArgCounter++, sizeof(float)*block_size*block_size, NULL);
...
kernel="    
    matrixMul(__global float* C,
                   ...
              __local float* A_temp,
                  ...
             )"
{...

My question is now, how to do the same in pyopencl?

I looked through the examples that come with pyopencl, but the only thing I could find was an approach using templates, which seems as to me as I understood it like an overkill. See example.

kernel = """
        __kernel void matrixMul(__global float* C,...){
                     ... 
            __local float A_temp[ %(mem_size) ];
                     ...
        }

What do you recommend?

Answer 1

It is similar to C. You pass it a fixed size array as a local. Here is an example from Enja’s radix sort. Notice the last argument is a local memory array.

 def naive_scan(self, num):
    nhist = num/2/self.cta_size*16
    global_size = (nhist,)
    local_size = (nhist,)
    extra_space = nhist / 16 #NUM_BANKS defined as 16 in RadixSort.cpp
    shared_mem_size = self.uintsz * (nhist + extra_space)
    scan_args = (   self.mCountersSum,
                    self.mCounters,
                    np.uint32(nhist),
                    cl.LocalMemory(2*shared_mem_size)
                )
    self.radix_prg.scanNaive(self.queue, global_size, local_size, *(scan_args)).wait()