|
These
benchmarks were run on an Apple
Power Macintosh G4 with dual 800 MHz PowerPCs. [Only
one of the PowerPCs is actually used in the benchmark.]
All but one were run on Yellow Dog Linux 2.3 with the stock
SMP kernel (2.4.19-4asmp) and gcc-3.3-20030303. The one
maked "Jaguar" was run on MacOS X 10.2.4.
vBigDSP is the fastest of these FFTs over most of
the benchmarks. It is run with two different compilers and
operating systems to illustrate that this can make a
significant performance difference. Although Jaguar appears
to be generally faster for vBigDSP, this was not the case
for all of the FFT implementations.
moto_fft does well for very small FFTs, but gives
a bit-reversed output. When the output is descrambled (using
my admittedly unoptimized function), the performance is
poor.
cfft2 is very slow, and included only for another
data point. We also included FFTW
2.1.5 (which does not use AltiVec) for comparison with the
others. Please note that fftw-3.0-fma
is faster than any of the codes provided here.
[Note: vBigDSP does not compile on CodeWarrior
because it lacks valloc(). You can substitue malloc(), which
gives vector-aligned data on MacOS X. The AltiVec PIM says
we should use vec_malloc(), but that is missing from both
platforms!]
All benchmarks are on one-dimensional complex radix-2
FFTs with interleaved data. FLOPS are computed as 5 N
log2 N, divided by execution time. The execution
time is the mean for a single transform, computed over 0.25
second. MFLOPS are 10^6 FLOPS. Here is the
code and the Matlab
script that generated this
graph.
|
