Explicit-shape/assumed-size vs. allocatable arrays.

[Janne Blomqvist]

Bart Oldeman wrote:
> Last summer, when introducing allocatable arrays into some research code 
> that used to be all F77, I observed that GFortran generated considerably 
> slower code using allocatable arrays than when passing them to a 
> subroutine that uses explicit shape or assumed size arrays. The 
> descriptor stride gets in the way a little.

It's true that in some cases the descriptor gets in the way. However, 
your benchmark code is flawed in that you should do a warm-up run before 
starting the timing.

When I do that, I get on 1.8 GHz K8 (eyeball-avg of 3 runs):

With -O0 -march=athlon64 -mfpmath=sse -ftree-vectorize

4.1:

  Array syntax:      3.90
  Explicit-shape:    3.42
  Allocatable:       3.86

4.2:

  Array syntax:      3.92
  Explicit-shape:    3.40
  Allocatable:       3.75

trunk:

  Array syntax:      3.90
  Explicit-shape:    3.40
  Allocatable:       3.80

With -O3 -funroll-loops -march=athlon64 -mfpmath=sse -ftree-vectorize 
-ffast-math

4.1:

  Array syntax:      1.80
  Explicit-shape:    1.69
  Allocatable:       1.73

4.2:

  Array syntax:      1.65
  Explicit-shape:    1.70
  Allocatable:       1.64

trunk:

  Array syntax:      1.56
  Explicit-shape:    1.54
  Allocatable:       1.57

So there is a slight advantage to the explicit shape case. Why, I don't 
know. Now, the funny thing is that if I remove the warm-up run (just a 
call to the add subroutine), and first do the loop in the main program, 
the vectorizer can no longer vectorize the loop?!

I also added a third test doing the loop using array syntax; in all 
cases the result was about equal to the other results for all 3 compiler 
versions I tested. So at least for simple loops the scalarizer works 
just fine.

-- 
Janne Blomqvist