Split amplitudes in smaller chunks

[Juergen Reuter]

Large amplitudes can overflow compiler limits. Split into separate modules.

[sschmidt]

just as an example: for the sin-file

alias q = u:U:d:D
alias j = u:U:d:D:g

!process dyuubar =   u, U -> e1, E1, j, j
!process dyuu    =   u, u -> e1, E1, j, j

process dy2 =   q, q -> e1, E1, j, j

compile

if the compiler uses the -O2 flag, the command

libtool: compile:  gfortran -c -I<build>/lib/whizard/mod/models -I<build>/include/omega -I<build>/lib/whizard/mod/whizard-core
  -I<build>/lib/whizard/mod/misc -g -O2 dy2.f90  -fPIC -o .libs/dy2.o

lets the gfortran compiler allocate some hundred MBs of memory (Then I killed it, because it used almost the entire RAM and was still getting bigger). Compilation with -O0 works without problems.

[Juergen Reuter]

OK; this seems to be an urgent must. For sm_ext, the compilation now is a killer. Would rank this even higher than the VAMP stuff (tooth-bitingly).

[ohl]

Update:

• For moderate cases like g g -> g g g g with -O2, there appears to be very little difference when compiling several modules in a single file or in separate files (unless we make use of multiple cpus/cores). That's a big surprise for me.
• The speedup and reduction in memory consumption wrt. the single module case is significant with -O2.
• I have not yet tried multiple functions in a single module.

[ohl]

Suprise: from my first benchmarks (gl gl -> gl gl gl gl) it appears that multiple functions in a single module is significantly faster than the more elaborate approaches. The only explanation that I can come up with is that this approach keeps all variables local to a single module.

Let's see what the longer benchmarks reveal during the night ...

[ohl]

OK, runtime benchmarks show that the penalty for splitting the functions is minimal (O(2%), much smaller than the O(40%) from optimization). Therefore we can safely make it the default.

I will merge the branch, after removing some (a lot) of duplicate code.

[ohl]

Implemented in r2276:

  -target:single_function compute the matrix element(s) in a monolithis function
  -target:split_function split the matrix element(s) into small functions [default, size = 10]
  -target:split_module split the matrix element(s) into small modules
  -target:split_file split the matrix element(s) into small files

[Juergen Reuter]

Milestone v2.0.0final deleted