Under OSX with the OSX BLAS (Accelerate Framework). The following program
real*8 :: x(2)
real*8 :: zero=0.0d0
x(1)=1.0d0
x(2)=1/zero
print*, x(2)*0.0d0
call dscal(2,0.d0,x,1)
print*, x
end
gives me the following output
NaN
0.0000000000000000 0.0000000000000000
and the following output on OSX+MKL BLAS
NaN
0.0000000000000000 NaN
i.e. it seems that the OSX BLAS DSCAL makes a short-circuit in the multiplication by 0 (we got X(2)=0 although 0*Inf=NaN should be obtained)
Is it a known bug/feature ?
This is neither a bug nor a feature. It is in the discretion of the implementer what the result of 0*NaN
0*Inf
is. What is the purpose of the above anyway?