with the 3.8.4.x releases I'm having some problems in Octave, described in the thread here: https://mailman.cae.wisc.edu/pipermail/octave-maintainers/2011-April/023614.html. In few words, dgemm can read uninitialized values when BETA=0. The problem disappears if I revert the fix (w.r.t. 3.8.3) described here: http://math-atlas.sourceforge.net/errata.html#scal0. I'm sorry I cannot provide a small fortran or C file to show the issue.
Reference dgemm initializes C to zero when BETA=0 (and not multiplies C by BETA). Does ATLAS 3.8.4.x do the same?