Alternatives instead of "if ... else" to speed up code

[Wee_Beng_T_]

Hi,

In my codes there's a lot of "if..else" statements. I'm told that they tend to slow down codes.

So I'm thinking of other alternatives, like the one below:

The 1st subroutineif_elseuses the "if...else" statement but the 2nd if_else2 does some arithimetics to avoid that.

I tested the runtime but they are almost the same, both in debug and optimized mode. Does this mean it makes no difference whether "if..else" is used?

Thanks!

program if_else_test

!test whether if_else can be avoided

implicit none

integer,parameter :: loop_no=10000,size=10000,call_time=20000

integer :: i

real :: array(size)

do i=1,call_time

call if_else2(loop_no,size,array) !or call if_else(loop_no,size,array)

end do

contains

subroutine if_else(loop_no,size,array)

!use if...else

integer, intent(in) :: loop_no,size

integer :: w(size)

real :: u(size),v(size)

real, intent(inout) :: array(size)

do i=1,loop_no

call random_number(u(i)); call random_number(v(i))

w(i)=(u(i)-v(i))/abs(u(i)-v(i))

if (w(i)==1) then

array(i)=2.*u(i)

else if (w(i)==-1) then

array(i)=2.*v(i)

end if

end do

end subroutine if_else

subroutine if_else2(loop_no,size,array)

!use alternative

integer, intent(in) :: loop_no,size

integer :: w(size)

real :: u(size),v(size)

real, intent(inout) :: array(size)

do i=1,loop_no

call random_number(u(i)); call random_number(v(i))

w(i)=(u(i)-v(i))/abs(u(i)-v(i))

array(i)= (w(i)+1)*u(i) - (w(i)-1)*v(i)

end do

end subroutine if_else2

end program if_else_test

[tom_p]

I would suggest something like this (untested):

[fortran]subroutine if_else(loop_no, size, array)
  integer, intent(in) :: loop_no, size
  real, intent(out) :: array(size)
  integer :: i
  real :: u, v

  do i = 1, loop_no
    call random_number(u)
    call random_number(v)
    array(i) = 2*max(u, v)
  end do
end subroutine if_else
[/fortran]

I think this is easier to understand and does not suffer from potential floating point rounding errors as the original code does (there is no guarantee w(i) will be exactly equal to 1 or -1, and abs(u(i)-v(i)) could be zero).

However, I would not expect a large difference in speed. The biggest impact could, I guess, be achieved by using a different random number generator that is adapted to your problem (fast, but still "random" enough).

[TimP]

Performance impact of if..else is more significant when it affects auto-vectorization. As you call a serial random number generator, the impact of the if..else would be relatively small here.
As the preceding response hinted, a selection by max() (or by an if which the compiler recognizes as equivalent) would be vectorizable in the absense of the subroutine call.

[Wee_Beng_T_]

Thanks Tom_p and TimP!

Well, I'm actually giving an example. In my e.g., I am trying to use :

array(i) = (w(i)+1)*u(i) - (w(i)-1)*v(i)

instead of

if (w(i)==1) then

array(i)=2.*u(i)

else if (w(i)==-1) then

array(i)=2.*v(i)

end if

The 1st case uses more operations but it avoids the use of "if..else".

Hence I wonder if the 1st case is better? Especially when optimizations are enabled.

[Steven_L_Intel1]

Don't forget that RANDOM_NUMBER can take an array argument to get many at once.