Preducer

A precision reducer for Fortran77 files, from double precision to real.

Why?

We needed a way to downgrade the precision in F77 files, to observe the differences in roundoff. Note that using preducer does not necessarily improve performance, as the created code contains various cast operations. Preducer is only really useful for studying roundoff effects, before modifying the precision by other means if the reduced precision is still sufficient. You wouldn't have used Preducer in a production code anyways, would you?

How?

Preducer is a Python script that parses F77 files, finds all subroutines and functions contained therein, and changes all their variables from double to single precision. The new routines/functions will have the original name with _sp appended. For example, subroutine foo becomes subroutine foo_sp.

To make it easier to call these new routines, a wrapper will be created that has the original name (in this case foo) and all arguments still have the original precision. This wrapper will cast all inputs to single precision, then call the low precision routine (foo_sp) using single precision inputs, and finally casts the results to double precision before returning. This way, the files created by Preducer can be used as drop-in replacements in existing software.

Installation

Preducer imports the following packages, (pip) install them as necessary:

fparser, sys, re, textwrap

Usage

The script can be directly called from the command line as

./preducer FILENAME.f
./preducer FILENAME.f SUBROUTINENAME
./preducer -verbose FILENAME.f
./preducer -verbose FILENAME.f SUBROUTINENAME

CUTEST

We used preducer as part of a Julia wrapper for CUTEst optimization problems. The Julia wrapper uses SIFDECODER to create a F77 file, after which we call Preducer to modify the precision. Preducer is tested on the files that SIFDECODER creates, and may break in various ways on other input files. The Julia code was modified to call Preducer and pass verbosity flags appropriately, the result is also included in this repository.

CUTEST installation

If you plan to use Preducer with the Julia CUTEST package, it is additionally necessary to define an environment variable called "PREDUCER" holding the Preducer directory. In the CUTEST.jl file, the follwing lines must be added directly below if isfile("ELFUN.f"):

      if "-preduce" in args
        preducenv = ENV["PREDUCER"]
        preducer = "$preducenv/preducer.py"
        if verbose
          preducer = ["$preducenv/preducer.py","-verbose"]
        end
        run(`$preducer ELFUN.f`)
        run(`mv ELFUN_preduced.f ELFUN.f`)
        run(`$preducer EXTER.f`)
        run(`mv EXTER_preduced.f EXTER.f`)
        run(`$preducer GROUP.f`)
        run(`mv GROUP_preduced.f GROUP.f`)
        run(`$preducer RANGE.f`)
        run(`mv RANGE_preduced.f RANGE.f`)
      end

CUTEST usage

Use Preducer from Julia's CUTEST as follows:

nlp = CUTEstModel("ROSENBR", "-preduce")
nlp = CUTEstModel("ROSENBR", verbose=true, "-preduce")