Re: [gfortran] patch for PR 15149 - problem computing PI

[Tobias Schlüter <tobias dot schlueter at physik dot uni-muenchen dot de>]

Bud Davis wrote:
> test suite file
>
> ! pr 15149
> ! note this is not a test of the random
> ! number generator, just that "something"
> ! exists and returns values not always 0
> TYPE point_type
>  SEQUENCE
>  REAL :: pos(2)
> END TYPE
> INTEGER, PARAMETER :: N=1000000
> TYPE(point_type)     :: points(N)
> REAL, DIMENSION(2*N) :: scratch
> EQUIVALENCE(points,scratch)
> CALL RANDOM_NUMBER(scratch)
> scratch=2.0*scratch-1.0
> count=0
> DO i=1,N
>   IF (SUM(points(i)%pos**2)<1.0) THEN
>      count=count+1
>   ENDIF
> ENDDO
> IF (4.0*REAL(count)/REAL(N).LT.3.14) THEN
>     CALL ABORT
> ENDIF
> END

I believe you should reduce N in  the testcase and allow a bigger error 
margin. For me this test takes 4/10 seconds. Given that every test is 
run 8 times this would mean a slowdown of the testsuite by two seconds. 
Not that this is dramatical, but it seems unnecessary.

With N=1000 I still get results which are well above 3, and the program 
runtime is dominated by startup time, i.e. they don't get shorter if I 
reduce N further.

- Tobi

- Tobi