This directory contains:

Readme:

This file

gambc-v4_6_0.tgz:

A gzipped tarfile of a modified version of Gambit-C suitable for benchmarking
the double-precision, complex, direct fft code in bignum multiplication.

TO BUILD GAMBIT-C:

tar zxf gambc-v4_6_0.tgz
cd gambc-v4_6_0
./configure CC=gcc --enable-single-host

Please, remember the --enable-single-host.

The CC variable contains the compiler version.  The default optimization
options are 

-O1 -fschedule-insns2

For some versions of gcc,

-fno-move-loop-invariants

is added to the list of optimizations because move-loop-invariants takes too
much compilation time.

You can change the compiler and the options at configure time by, e.g.,

./configure CC='gcc-4.5 -O2 -fschedule-insns' --enable-single-host

Optimizations in the CC variable are added to the front of the options
list.  I find that on powerpc (32 and 64 bit) and on x86-64

-fschedule-insns

offers the greatest benefit.

TO BENCHMARK THE FFT CODE

The following code is an accurate test only on 64-bit machines, unfortunately.
(On 32-bit machines, a different algorithm is used internally.)

In the Gambit-C build directory give the command

gsi/gsi -e '(define a (expt 3 10000000))(set! *bench-bignum-fft* #t)(define b (* a a))'

The output will look like:

(time (direct-fft-recursive-4 a table))
    174 ms real time
    180 ms cpu time (180 user, 0 system)
    no collections
    16 bytes allocated
    no minor faults
    no major faults

so it's easy to strip out the result.

If that's too little CPU time to measure accurately, then try

gsi/gsi -e '(define a (expt 3 100000000))(set! *bench-bignum-fft* #t)(define b (* a a))'
(time (direct-fft-recursive-4 a table))
    3388 ms real time
    3380 ms cpu time (3380 user, 0 system)
    no collections
    16 bytes allocated
    no minor faults
    no major faults