(j3.2006) Mask / Pack / Vector subscript vs. WHERE

[Van Snyder]

I asked a few hours ago whether a method using mask, pack and a vector
subscript, with an allocated array for the vector subscript, would be
faster or slower than a WHERE ... ELSE WHERE ... construct.

The consensus was that the WHERE construct would be faster.

I was surprised that it was not faster using two of three compilers I
commonly use.  At least, it was faster with one of them.

I also created an elemental version, which was fastest by far.  The
other two versions have rank-1 arguments.  Since the elemental version
will accept any rank, it's nice it's also fastest.

I cannot post the code because it's still under review for publication.
If you want to try it yourself, let me know and I'll see what I can do
to let you have the codes.

Here are the results, without attribution:

Compiler A

Faddeyeva_Rev.f90        47.7787
Faddeyeva_Where.f90      63.6593
Faddeyeva_Elemental.f90  27.2149

Compiler A -O

Faddeyeva_Rev.f90        34.5308
Faddeyeva_Where.f90      38.4112
Faddeyeva_Elemental.f90  18.2692

Compiler B

Faddeyeva_Rev.f90        32.0141
Faddeyeva_Where.f90      35.6216
Faddeyeva_Elemental.f90  13.3000

Compiler B -O

Faddeyeva_Rev.f90        32.4971
Faddeyeva_Where.f90      35.6776
Faddeyeva_Elemental.f90  13.2630

Compiler C

Faddeyeva_Rev.f90        45.2341
Faddeyeva_Where.f90      41.8476
Faddeyeva_Elemental.f90  30.4304

Compiler C -O

Faddeyeva_Rev.f90        33.8579
Faddeyeva_Where.f90      30.5444
Faddeyeva_Elemental.f90  17.8903

The program consists of only a main program and two modules.  In every
case, I compiled everything all at once, rather than compiling each file
individually and linking the .o files.  I don't know whether this makes
a difference.