Monday, December 3, 2012

Ternary Operator ?

Suppose you had a variable x that has a value depends on the variables a and b. The most obvious choice is to write an if-else statement to determine what x is:
if(a > b) then
   x = a
else
   x = b
endif

But this seems a bit cumbersome, particularly when you look at a more compressed form using the ternary operator ? as in C (among other languages):
x = a > b ? a : b;

I used to be jealous of C because they had the inline-if and I believed that Fortran did not. Fortunately for us, Fortran added this inline-if into the 1995 standard. Unfortunately, they chose the most random of intrinsic function names I have ever heard of: merge. It is used quite like the above:
x = merge(a,b,a>b)

More explicitly,
variable = merge(value if true, value if false, condition)

The following is a snippet from a simulation I wrote that involved a Monte Carlo sampling of positions in cylindrical coordinates. R,P,Z are the rho-phi-zed coordinate-arrays.
do i=1,numPoints
   call random_number(rTemp)
   call random_number(pTemp)
   call random_number(zTemp)
   call random_number(cTemp)
   R(i) = sqrt(rTemp)*rScale
   P(i) = 2.*pi*pTemp
   Z(i) = merge(zTemp*zScale,-zTemp*zScale, cTemp > 0.5)
enddo

A couple notes on the above:
(1) Fortran uses 1-index, rather than C's 0-index--which leads me to the aside: quickly count to 10. Did you start at 0? Fortran can actually start at 0, if you want. You would have to define the array to start there by using
real,dimension(0:numPoints)::myArray

(2) Fortran uses curved braces for the array index, whereas C uses the square braces for the same thing.
(3) The variable pi is not an intrinsic (why not!?) and needs to be declared beforehand; I find the easiest option is using pi=acos(-1.0), though others may suggest pi=4*atan(1.0) (which is fine, but it seems to me that one operation is better/faster than two).
(4) The intrinsic subroutine random_number(variable) was used because it is superior to the intrinsic function rand(). The former has a period of 2123 (~1037) while the latter has a period of (I think) 232 (~109).


Comments always welcome.

6 comments:

  1. pi=acos(-1.0) is superior to pi=4*atan(1.0) because the latter loses two bits of precision.

    ReplyDelete
  2. Shouldn't line 3 read
    "else if (a < b) then" ?
    For equality, the merge condition evaluates to false

    ReplyDelete
  3. +MartinDiehl: Actually, I think it should just be "else" in line 3 because we're testing for strictly greater than with merge and the C ternary operator. Using the "else if" there would eliminate equality (which should evaluate to false in the three cases, if they are to be equivalent).

    ReplyDelete
  4. It should be noted that in contrast to C, both arguments of merge function are calculated (or can be calculated, depending on a compiler), and you'll probably get a problem if one of them throws an exception even if this argument is not used according to the mask. So, the following call: merge(array(index), array(-index), index > 0) being innocent at first sight, can crash your program.

    ReplyDelete
  5. This blog awesome and i learn a lot about programming from here.The best thing about this blog is that you doing from beginning to experts level.

    Love from

    ReplyDelete
  6. This is a pretty good post. I think merge was really intended as a way to combine two arrays into a third, e.g. c=merge(a,b,a>0) means take the elements of a where they are greater than 0 and otherwise take the elements of b. For this, the name "merge" makes sense.

    It's kind of similar to the where() construct, e.g. where(a<=0)a=0, so the last argument is really a "mask", i.e. an array of logicals that.

    PS:
    Counting Model (1-based) vs. Location Model (0-based)
    I was a Fortran programmer for a long time and have recently added C/C++ to the repertoire. To help you come to terms with C's zero-indexing, think about this: would you start a number line axis at 1? No. Zero is the most natural "origin". The C array model is really just a memory layout model. You can get the address of a scalar 'a' in memory as '&a'. How do you refer to the memory block to the left/right of that? C chose that &a[1] is next door to the right, &a[-1] is next door to the left and &a[0] is the same as &a. When 'a' is an array, the same logic applies: &a[0] is the origin, &a[1] is the next element, etc. I have heard people say [] in C is a memory "offset" operator which is also used for array indexing. Did you know in most places in Europe, the ground floor is 0? It's kind of cool because then the below ground floor numbers are standardized: -1, -2, etc. instead of the "B1", "B2", "G", etc.

    ReplyDelete