Fix Y misscale in GEMV for some cases

[bE554357]

Description

According to the description, $M$ or $N$ could be zero. It is expected that the vector $\mathbf{y}$ will be correctly scaled by scalar for such cases.

For example, let GEMV computes
$\mathbf{y} := 0 \cdot \mathbf{y} + \mathbf{A}\mathbf{x}$
and $\mathbf{x}$ is zero length vector. It is expected, that $\mathbf{y} = 0$ after GEMV, but it is not scaled due to statement in implementation:

- 	 IF ((M.EQ.0) .OR. (N.EQ.0) .OR.
-     +    ((ALPHA.EQ.ZERO).AND. (BETA.EQ.ONE))) RETURN

This pull request is objected to fix the behavior in such cases.

[langou]

Hi,

Note that the code follows the convention as written in the article

J. J. Dongarra, J. Du Croz, S. Hammarling, and R. J. Hanson. An extended set of FORTRAN Basic Linear Algebra Subprograms. ACM Trans. Math. Software, 14:1–32, 399, 1988.

I quote (page 5, section 3 - naming convention)

Note that it is permissible to call the routines with M or N = 0, in which case the routines exit immediately without referencing their vector or matrix arguments.

Cheers,
Julien.