Coded long pulse computational complexity

jan09

Look at the computational complexity for the coded long pulse program. This is done for the echotek system (4K ffts , 5Mhz bw) as well as tammara's machine (mock box with 32 Mhz bw).
Some preliminaries are:

the fft dominates the computation (by about a factor of 10 so ignore accumulation and decoding.
Variables:

K=6 constant for fft complexity
fftlen- length of fft
nheights - number of heights computed each ipp.
ippSec - time for 1 ipp (.01 secs).
nbands- how many bandwidths we do simultaneously (to compare mikes times with mine).
cmps1fft - computations 1 fft
cmps1ipp - computations 1 ipp
cmps1sec - computations 1 second.

Coded long pulse computational breakdown
Step	formula	Ryan's (old) machine	Tammara (new) machine
Cmps1fft	Kfftlenlog2(fftlen)	6.409612.=2.9e-4 GFlps	6.32768.15=2.9e-3 GFlps
cmps1ipp	cmps1Fftnheightsnbands	cmps1fft22502=1.327GFlps	cmps1fft*1591=4.69 GFlps
cmps1sec	cmps1ipp/ippSec	cmps1ipp/.01=132.7 GFlps	cmps1fft/.01=469.g GFlps

Cpus used
cpuName	Description	Software
macNeh	intel 5550 2*4 nehalem 2.67 Ghz	MacOs, accelerate for fft, mikeSulzer asp
aserv11	intel 5550 2.4 nehalem,2.67 ghz	Linux, fftw 3.2,fftwThreaded 3.2 phil, clp1 or clp

Computational speed achieved (2009)
Code	machine	date	time/thread for 10secs	gflopsTotal
mike sulzer, asp, ryans data, 8 copies run	macNeh	2009	320	132.7/(320/10) *8=33 GFlps
phil,Tammara's machine,"clp" multi threaded fftw(nthreads=10), single threads for inp,decode,accum output	aserv11	jan10	180.	469.2/(180/10) *1=26 GFlps
phil,Tamarra's machine clp1 N=8 threads 1 inp thread, N procThreads doing complete processing.	aserv11	feb10	1168	469.2/(1168/10)*8=32 GFlps

--> clp software
home_~phil