Re: OT: negative zeros with ATLAS 3.8.4.x

[John W. Eaton]

On 12-Apr-2011, Marco Caliari wrote:

| In reference BLAS dgemm I see that if beta is zero, than C is set to zero 
| (which is exactly what we want). So, the problem should be in ATLAS.

I'm not sure reference BLAS gets this right, since it will not properly
generate NaN for cases where C includes Inf values and beta happens to
be zero (instead of having zero always mean that the operation should
be skipped).

So I think there really should be a separate function that doesn't do
the +beta*C part if you don't need it.

| P.S. Why in my example was used the "general GEMM operation" and not the 
| "column vector by row vector -> matrix operations" in dMatrix.cc?

Do you mean

  // column vector by row vector -> matrix operations

  Matrix
  operator * (const ColumnVector& v, const RowVector& a)

?

Looking at the body of that function, it just calls DGEMM internally.

And the other reason is that in Octave, your row and column vectors
are actually stored as Matrix objects not RowVector and ColumnVector
objects.

jwe