Weight updates on the GPU about somoclu

[peterwittek]

Commits 0ff6a92 and 31d41b5 implement in-place codebook updates for the dense and sparse CPU kernels when compiled without MPI.

A similar mechanism was implemented for the GPU (commit b3b4440), but the update does not happen on the GPU. There are two ways of doing it, and neither of them would be efficient:

1. Formulate the in-place update as W = W + H X, where W is the codebook, X is the data, and H is the matrix of the update weights (it includes the learning rate). H could be calculated efficiently on the GPU, and W+HX is a standard BLAS call. The problem is that H would have to be in the GPU memory, and its size is #rows x #columns x #data_points, where the rows and columns refer to the map. So calculating H would severely restrict the size of the problems that can be solved on the GPU.
2. The second option is writing a kernel that calculates the W update directly. The problem here is the memory access patterns: each entry of the X matrix would have to be fetched #rows x #columns times, which simply cannot be efficient.