Output with empty matrix produces invalid data

[projekter]

Before carrying out the multiplication, _sparse_dot_dense and _sparse_dot_sparse check whether the result is zero and take a shortcut. However, if the out parameter is specified, this shortcut will just return the output matrix unchanged instead of setting it to zero.
Edit: Basically, the out_scalar parameter is completely ignored.
Second edit: Actually, there's another issue due to the inconsistent way in which output arrays are handled. spmm does not accept an out parameter, so correctly an exception is thrown if one is provided. spmmd does allow for one; however, it always acts as if out_scalar=0 (so the default None corresponds to 0 in this case). sparse_dot just ignores the out_scalar parameter. sparse_?_mm and sparse_?_mv also allows for the output, and this case the out_scalar is converted using _mkl_scalar, i.e., the default None corresponds to 1 in this case. I think there are only two ways to resolve this inconsistency in the general dot_product_mkl interface:

• The backwards-compatible one is to check for out_scalar different from None or 0 in the spmmd case and throw an exception, plus add this to the documentation.
• The consistency-oriented one is to change the default value of out_scalar from None to 0 and add the exception conditioned on it not being 0.

import sparse_dot_mkl, scipy.sparse, numpy

output = numpy.random.rand(3, 3)
sparse_dot_mkl._sparse_dense._sparse_dot_dense(
    numpy.array([[1., 2., 3.], [4., 5., 6.], [7., 8., 9.]]),
    scipy.sparse.csr_matrix((3, 3)),
    out=output,
    out_scalar=0
)
# now out is still the random matrix
sparse_dot_mkl._sparse_sparse._sparse_dot_sparse(
    scipy.sparse.csr_matrix((3, 3)),
    scipy.sparse.csr_matrix((3, 3)),
    out=output,
    dense=True,
    out_scalar=0
)
# and also now