2D Pooling on PyCudaHandler fails for big inputs

[Qwlouse]

The error:

File ".../brainstorm/layers/pooling_layer_2d.py", line 89, in forward_pass
    flat_argmax)
  File ".../brainstorm/handlers/pycuda_handler.py", line 314, in maxpool2d_forward_batch
    grid=(NUM_CUDA_THREADS, 1, 1))
  File ".../lib/python3.4/site-packages/pycuda/driver.py", line 379, in function_call
    func._set_block_shape(*block)
pycuda._driver.LogicError: cuFuncSetBlockShape failed: invalid argument

How to reproduce: In the CIFAR-10 example change the number of maps in the first convolution to something bigger than 46 and uncomment the line with the handler to run it on GPU.