Rapid Transpose Dot Product of 2 CSR Matrices

Sorry to answer myself, but the fastest solution was to rewrite the simple XW.T code in cython (remember, these are sparse vectors not matrices). Here y is used to represent the W weights. Here the sparse vector is passed as 2 ndarrays, one of indices, the other of corresponding values. This is a CSRMatrix's native format if you're dealing with an nx1 matrix (aka vector), so you can easily pass CSRMatrices to this code. In a program where XW.T is getting done ~10k times, this went from taking 33s with a standard numpy XW.T to 0.76 seconds, so this is really* faster, and the long runs are more on the order of 1e6 to 1e7 calls, so you can imagine the time saved :-P

import numpy as np
cimport numpy as np
import cython

@cython.boundscheck(False)
def t_dot( np.ndarray[int, ndim=1] x_indices, 
           np.ndarray[np.float64_t, ndim=1] x_data,
           np.ndarray[np.float64_t, ndim=1] y):   
    """
    Dot product of the transpose.
    """
    #Using this type is important, this gives us exactly the same values as it
    #will give exact same values as numpy's equivalent X*W.T
    cdef np.float64_t sum
    cdef int idx, ptr
    sum = 0.0
    for idx, ptr in enumerate(x_indices):
        sum += x_data[idx] * y[ptr]    
    return sum