Re: [Help-gsl] Why is Negative Stride in GSL not allowed?

[Michael Lehn]

Am 19.11.2012 um 11:52 schrieb Stephan Petzchen:

> But there is no such code in ATLAS. I did look up
> It is not only taking the absolut value of incX also the reference starting 
> point should be the end of the vector.
> 
> 
> In all there testcases they set the incX value to absolute (i.e.  Mabs( INCX 
> )), which simply follow the rule in one of your former messages: *(X + 
> abs(incX)*i) which is not considering the BLAS standard with starting point 
> at the end of the vector: X(1 + (n-i) * abs(INCX)).
> 
> So, my opinion now is: they do not comply with the BLAS standard document for 
> negative stride.
> 
> I hope you or someone other can show me that I am wrong. PLEASE...   ;-)
> 

I just downloaded ATLAS 3.10.  Have a look into 
ATLAS/interfaces/blas/C/src/cblas_daxpy.c.  There
you find


//---- SNIP 
---------------------------------------------------------------------

void cblas_daxpy(const int N, const double alpha, const double *X,
                 const int incX, double *Y, const int incY)
{
   if (N > 0)
   {
      if (incX < 0)
      {
         if (incY < 0) ATL_daxpy(N, alpha, X, -incX, Y, -incY);
         else ATL_daxpy(N, alpha, X+(1-N)*incX, incX, Y, incY);
      }
      else if (incY < 0) ATL_daxpy(N, alpha, X+(N-1)*incX, -incX, Y, -incY);
      else ATL_daxpy(N, alpha, X, incX, Y, incY);
   }
}

//---- SNAP 
---------------------------------------------------------------------


Cheers,

Michael