last Update 10.Feb 2012

http://www.hd-rec.de/Archive/FFTDemo.lha

http://en.wikipedia.org/wiki/Fast_Fourier_transform

thats a Fast Fourier Transformation and a inverse Fast Fourier Transformation in 68k.The 68k instructions can easy replace by PPC opcode by the

JIT.
It cost not much time to run the test, so there are lots values of System i dont own here.

FFT use 4096 steps and execute inner loop 4096 times and then do some few in outer loop.
I think it have a good mix of memread/ memwrite and calculation to come near real world performance in most programs that run too slow
and do Digital Signal Processing as Audio and Image processing.

The float and integer test is Amiblitz 2 or 3 Basic Code.The last test is handoptimized asm code.

the text ( xxxx ms at 500 MHZ) are theory values if the CPU is clock to 500 MHZ.
here you can see how good the performance /clock Ratio is.and can see that
the PPC 604e seem the best PPC ever build in FPU performance /MHZ, sad it do not run on higher clockrates as 375 MHZ.The benchmark currently does not output the values for 500 MHZ.i have done it in calculator

To make it complete there are miss some values.
From a IBM G3 (micro aone OS4 and a Peg G3 MOS)

From a modern AMD System (Phenom or new Athlon with 3 GHZ)

If somebody have such a System, please download the test and Post results.If somebody want do a C native Bench its also intresting to see.

lower values are better.

classic 68060/50

Speed test for FFT + iFFT: (float)
time needed 34257ms for 413696 samples, => .136918902397155x speed
@44100Hz/stereo
------------ (3425ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 30891ms for 413696 samples, => .151838108897209x speed
@44100Hz/stereo
------------ (3089ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 16229ms for 413696 samples, => .289015382528305x speed
@44100Hz/stereo
------------ (1622ms at 500 MHZ)


-------------------------

PPC 604 150 MHZ OS4

Speed test for FFT + iFFT: (float)
time needed 19449ms for 413696 samples, => .241165652871131x speed
@44100Hz/stereo
------------ (5840ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 16485ms for 413696 samples, => .28452718257904x speed
@44100Hz/stereo
------------ (4950ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handopmized 68K ASM)
time needed 7108ms for 413696 samples, => .659880518913269x speed
@44100Hz/stereo
------------ (2134ms at 500 MHZ)

----------------------------------------------------------------------


Blizzard PPC 321Mhz 80.333Mhz bus. OS4.0
Speed test for FFT + iFFT: (float)
time needed 15625ms for 413696 samples, => .300187587738037x speed @44100Hz/stereo
(10080ms at 500MHZ)
Speed test for FFT + iFFT: (integer)
time needed 13146ms for 413696 samples, => .356795281171798x speed @44100Hz/stereo
(8481ms at 500MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 10613ms for 413696 samples, => .44195145368576x speed @44100Hz/stereo
(6847ms at 500MHZ)


Blizzard PPC OS4.O 334Mhz

Speed test for FFT + iFFT: (float)
time needed 15359ms for 413696 samples, => .305386483669281x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer)
time needed 12910ms for 413696 samples, => .363317638635635x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 10423ms for 413696 samples, => .450007766485214x speed @44100Hz/stereo

-------------------------------------------------------------

Cyberstorm PPC @ 366MHz, 66.7MHz bus speed, 70ns ram, OS4:

FFTDemo
Speed test for FFT + iFFT: (float)
time needed 8362ms for 413696 samples, => .560922145843505x speed @44100Hz/stereo
------------ (6148ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 7154ms for 413696 samples, => .655637502670288x speed @44100Hz/stereo
------------ (5260ms at 500 MHZ)

Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 3357ms for 413696 samples, => 1.39720904827117x speed @44100Hz/stereo
------------ (2468ms at 500 MHZ)

Cyberstorm PPC 400MHz, 66.7MHz bus speed, 70ns ram, OS4:

FFTDemo
Speed test for FFT + iFFT: (float)
time needed 7478ms
Speed test for FFT + iFFT: (integer)
time needed 6419ms
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 3045ms


----------------------

Sam 667 MHZ OS4

Speed test for FFT + iFFT: (float)
time needed 6136ms for 413696 samples, => .764411807060241x speed
@44100Hz/stereo
------------ (8185ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 4735ms for 413696 samples, => .990587294101715x speed
@44100Hz/stereo
------------ (6316 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 2510ms for 413696 samples, => 1.86869752407073x speed
@44100Hz/stereo
------------ (3348ms at 500 MHZ)

--------------------------------------------------

Sam 440-EP Flex 800Mhz OS4.1U2:

FFTDemo
Speed test for FFT + iFFT: (float)
time needed 5614ms for 413696 samples, => .835488200187683x speed @44100Hz/stereo
------------ (8982ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 4205ms for 413696 samples, => 1.11544132232666x speed @44100Hz/stereo
------------ (6728ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 2308ms for 413696 samples, => 2.03224897384643x speed @44100Hz/stereo
------------ (3692ms at 500 MHZ)

A1-SE@733MHz, OS4.1.2

Speed test for FFT + iFFT: (float)
time needed 3933ms for 413696 samples, => 1.19258344173431x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer)
time needed 2899ms for 413696 samples, => 1.61794781684875x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 1339ms for 413696 samples, => 3.50293564796447x speed @44100Hz/stereo

--------------------------------------------

aone 800 MHZ G4 OS4

Speed test for FFT + iFFT: (float)
time needed 4171ms for 413696 samples, => 1.12453389167785x speed
@44100Hz/stereo
------------ (6673 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 2839ms for 413696 samples, => 1.65214192867279x speed
@44100Hz/stereo
------------ (4542 ms at 500MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 1225ms for 413696 samples, => 3.82892322540283x speed
@44100Hz/stereo
------------ (1960 ms at 500 MHZ)

---------------------------------

OS4: Peg2 1GHZ G4

Speed test for FFT + iFFT: (float)
time needed 3183ms for 413696 samples, => 1.47358810901641x speed
@44100Hz/stereo
----------------(6362 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 2235ms for 413696 samples, => 2.09862685203552x speed
@44100Hz/stereo
----------------(4470 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 961ms for 413696 samples, => 4.88078117370605x speed
@44100Hz/stereo
----------------(1922 ms at 500 MHZ)

------------------------------

MorphOS: Peg2 1 GHZ G4

time needed 3022ms for 413696 samples, => 1.55209481716156x speed
@44100Hz/stereo
----------------(6044 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 1626ms for 413696 samples, => 2.88464379310607x speed
@44100Hz/stereo
----------------(3252 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 619ms for 413696 samples, => 7.57743263244628x speed
@44100Hz/stereo
----------------(1238 ms at 500 MHZ)

-----------------------------

------------------------------

X1000 1.8 GHZ

Speed test for FFT + iFFT: (float)
time needed 1720ms for 413696 samples, => 2.72699475288391x speed @44100Hz/stereo
-------6192 ms @500 MHZ
Speed test for FFT + iFFT: (integer)
time needed 1276ms for 413696 samples, => 3.6758861541748x speed @44100Hz/stereo
-------4593 ms @500 MHZ
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 617ms for 413696 samples, => 7.60199499130249x speed @44100Hz/stereo
-------2221 ms @500 MHZ

-----------------

Mac Mini 1,5 GHZ MOS 2.4

Speed test for FFT + iFFT: (float)
time needed 2020ms for 413696 samples, => 2.32199525833129x speed
@44100Hz/stereo
------------ (6060 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 1071ms for 413696 samples, => 4.37948703765869x speed
@44100Hz/stereo
------------ (3213 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 407ms for 413696 samples, => 11.5243997573852x speed
@44100Hz/stereo
------------ (1221 ms at 500 MHZ)

------

Next are the benchmarks on some Systems run on winuae.

Note:
In theory a PPC should be lots faster to emulate 68k.
The X86 need byteswap lots data, need emulate Chipset and also X86 have fewer registers as 68k.
but seem because of the better performance /clock value
in FPU the newer X86 are faster at same clockrate as PPC.But the Core I7 rule in all values too when compare performance /Clock

--------------------------------------------------------------

this are CPU that have very worse performance /clock.P3 cant buy, but a Atom 270 oh no its really slow, but of course due to 1,6 GHZ faster as a SAM.also it have second level cache

winuae PIII 650 MHZ

Speed test for FFT + iFFT: (float)
time needed 9852ms for 413696 samples, => .618871986865997x speed @44100Hz/stereo
------- (12807 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 7628ms for 413696 samples, => .614896535873413x speed @44100Hz/stereo
------- (9916 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 3651ms for 413696 samples, => 1.2846975326538x speed @44100Hz/stereo
------- (4746 ms at 500 MHZ)

--------------------------------

Atom N270 1,6 GHZ Netbook speedmode/ fastest = shorttest time of accu

Speed test for FFT + iFFT: (float)
time needed 3837ms for 413696 samples, => 1.22242140769958x speed @44100Hz/stereo
------- (12278 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 3123ms for 413696 samples, => 1.50189912319183x speed @44100Hz/stereo
-------(9993 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 1654ms for 413696 samples, => 2.83581066131591x speed @44100Hz/stereo
-------(5292 ms at 500 MHZ)

Atom N270 1,6 GHZ Netbook slowest = longest run of Accu.

Speed test for FFT + iFFT: (float)
time needed 9063ms for 413696 samples, => .517536222934722x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer)
time needed 7137ms for 413696 samples, => .657199203968048x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 3701ms for 413696 samples, => 1.26734149456024x speed @44100Hz/stereo


--------------------------------

AMD Sempron 1,8 GHZ Amithlon

Speed test for FFT + iFFT: (float)
time needed 1234ms for 413696 samples, => 3.80099749565124x speed @44100Hz/stereo
--------- (4442 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 826ms for 413696 samples, => 5.67848777770996x speed @44100Hz/stereo
--------- (2973 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 373ms for 413696 samples, => 12.5748815536499x speed @44100Hz/stereo
--------- (1342 ms at 500 MHZ)

-------------------------------


winuae Core Duo 1,8 GHZ Notebook

Speed test for FFT + iFFT: (float)
time needed 1149ms for 413696 samples, => 4.0821852684021x speed
@44100Hz/stereo
------------ (4136 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 1275ms for 413696 samples, => 3.67876935005188x speed
@44100Hz/stereo
------------ (4590 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 604ms for 413696 samples, => 7.76561403274536x speed
@44100Hz/stereo
------------ (2174 ms at 500 MHZ)

------------------------------------------

winuae AMD64 3000+ (real 1,8 GHZ)

17.h0:wbstartup> "Ram Disk:FFTDemo"
Speed test for FFT + iFFT: (float)
time needed 1092ms for 413696 samples, => 4.29526615142822x speed
@44100Hz/stereo
------------ (3931 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 1098ms for 413696 samples, => 4.27179479598999x speed
@44100Hz/stereo
------------ (3952 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 509ms for 413696 samples, => 9.21499156951904x speed
@44100Hz/stereo
------------ (1832 ms at 500 MHZ)


-------------------------------

Intel I5 760 3.33 GHZ(2.8 GHZ clock labeled but default Intel turboboost work here because of 1 thread)

15.h0:wbstartup> "Ram Disk:FFTDemo"
Speed test for FFT + iFFT: (float)
time needed 440ms for 413696 samples, => 10.6600704193115x speed @44100Hz/stereo
-------2930 ms @500 MHZ
Speed test for FFT + iFFT: (integer)
time needed 484ms for 413696 samples, => 9.69097328186035x speed @44100Hz/stereo
-------3223 ms @500 MHZ
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 196ms for 413696 samples, => 23.9307689666748x speed @44100Hz/stereo
-------1305 ms @500 MHZ


--------------------------------

winuae 3.9GHz Core i7 920.


3.harddrive0:d> fftdemo
Speed test for FFT + iFFT: (float)
time needed 312ms for 413696 samples, => 15.0334320068359x speed
@44100Hz/stereo
------------ (2433 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer)
time needed 340ms for 413696 samples, => 13.7953844070434x speed
@44100Hz/stereo
------------ (2652 ms at 500 MHZ)
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 150ms for 413696 samples, => 31.2695388793945x speed
@44100Hz/stereo
------------ (1170 ms at 500 MHZ)
3.harddrive0:d>

Last edited by bernd_afa on 10-Feb-2012 at 03:46 PM.
Last edited by bernd_afa on 10-Feb-2012 at 03:45 PM.
Last edited by bernd_afa on 05-Feb-2012 at 10:15 AM.
Last edited by bernd_afa on 16-Sep-2010 at 03:30 PM.
Last edited by bernd_afa on 05-Sep-2010 at 11:08 AM.
Last edited by bernd_afa on 28-Aug-2010 at 04:01 PM.
Last edited by bernd_afa on 26-Aug-2010 at 07:15 PM.
Last edited by bernd_afa on 26-Aug-2010 at 07:13 PM.
Last edited by bernd_afa on 26-Aug-2010 at 07:13 PM.
Last edited by bernd_afa on 26-Aug-2010 at 07:07 PM.
Last edited by bernd_afa on 26-Aug-2010 at 07:01 PM.
Last edited by bernd_afa on 26-Aug-2010 at 06:51 PM.
Last edited by bernd_afa on 13-Dec-2009 at 02:43 PM.
Last edited by bernd_afa on 15-Nov-2009 at 09:32 AM.
Last edited by bernd_afa on 15-Nov-2009 at 09:32 AM.
Last edited by bernd_afa on 14-Nov-2009 at 03:17 PM.
Last edited by bernd_afa on 10-Nov-2009 at 03:11 PM.
Last edited by bernd_afa on 07-Nov-2009 at 01:01 PM.
Last edited by bernd_afa on 07-Nov-2009 at 12:58 PM.

@bernd_afa

Quote:

The X86 need byteswap lots data, need emulate Chipset and also X86 have fewer registers as 68k.

What if you run it on Amithlon?

I don't know if registers matter; do we even know if the PPC JITs make use of its own registers by properly placing 68k registers into PPC registers? Regardless of this x86 JIT is fast and x86 chips are made so that non-register and memory data accesses are quick anyway.

@Hypex

ah yes, values of Amithlon are also usefull.if somebody have one he can try.

But Amithlon have a performance Problem get with last 2 patches and FPU.
when use patch 1 fpu jit work fastest.

>I don't know if registers matter; do we even know if the PPC JITs make use of its >own registers by properly placing 68k registers into PPC registers?

the register problem you can see more, when you use the asm optimized code.this use more 68k registers as amiblitz code and as you can see the asm optimized code get more speedup on the mac mini aone and 604 in compare to winuae.

Last edited by bernd_afa on 07-Nov-2009 at 02:02 PM.
Last edited by bernd_afa on 07-Nov-2009 at 01:51 PM.
Last edited by bernd_afa on 07-Nov-2009 at 01:51 PM.
Last edited by bernd_afa on 07-Nov-2009 at 01:47 PM.

Um, Since I dont really have any reasonably powerful hardware to try things on, or set up etc, what sort of algorithm is being used for this FFT function?.. You say inner loop, and outer loop, which sounds like the brute force version with sine and cosine calculated each time?.. or is it one of the more optimised, Butterfly, algorithms thats all additions and few static multiplications (Twiddle Factors) and just shows how much processing power is really needed for modern code?

That is, origional FFT code needed n*n calculations for comparing all samples with all samples, but optimised FFT only needs n*ln(n) calculations or so? That is, 4096 samples is 2^12, so you would have 12 processing steps, so the bigger the sample, the more efficient?

Sorry if Ive got things completely confused 8(.. Its nice to see the speeds though, very compatible.

_________________
The older and more respected a scientist is, the longer it takes to prove him wrong.

@RodTerl

see in source for all and better indent, its a optimized version, that use a sin table.The sin is also only calculate in the 3.rd loop.

The code also not speedup lots with a 2. Level cache, see the theoretical 604 performance at 500 MHZ, its really fast, also 604 have very slow EDO Ram.SO you can see the X86 does not help the fast 2 Channel ram or 3 Channel RAM of I 920

but all in all, when the same code run on several Hardware it doesnt matter for speedcompare lots if the FFT is optimize or not.

For k = 0 To \order-1
le = le2-1
le2 LSL 1
le8 LSL 1
le16 LSL 1
ur.f = 1.0
ui.f = 0.0
wr.f = \cosT\f[k]
wi.f = \sinT\f[k] * df
For j = 0 To le
p1r.l = buffer_r + (j LSL 3)
p2r.l = p1r+le8
For i = j To npoints1 Step le2
x.f = Peek.f(p2r)
y.f = Peek.f(p2r+4)
tr.f = x*ur - y*ui
ti.f = x*ui + y*ur

x.f = Peek.f(p1r)
Poke.f p2r, x-tr
Poke.f p1r, x+tr

x.f = Peek.f(p1r+4)
Poke.f p2r+4, x-ti
Poke.f p1r+4, x+ti

p1r + le16
p2r + le16
Next
t.f = (ur*wr - ui*wi)
ui = (ur*wi + ui*wr)
ur = t
Next
Next

Last edited by bernd_afa on 07-Nov-2009 at 02:13 PM.
Last edited by bernd_afa on 07-Nov-2009 at 02:11 PM.
Last edited by bernd_afa on 07-Nov-2009 at 02:09 PM.

@bernd_afa

68k JIT in OS4/MOS must be really good - two times faster than JIT in WinUAE (according to your results and estimated CPU performance of Core i7 and 7445).

@pavlor

i dont see how you come to this result.
the benchmark is not that highewr values are better.lower values are better.

ms mean the time need in ms to run this Test

when you see the FPU values of mac mini at 500 MHZ

Mac Mini 1,5 GHZ MOS 2.4

Speed test for FFT + iFFT: (float)
time needed 2020ms for 413696 samples, => 2.32199525833129x speed
@44100Hz/stereo
(6060 ms at 500 MHZ)

and

winuae 3.9GHz Core i7 920.

3.harddrive0:d> fftdemo
Speed test for FFT + iFFT: (float)
time needed 312ms for 413696 samples, => 15.0334320068359x speed
@44100Hz/stereo
(2433 ms at 500 MHZ)

then you see that the I7 run with 500 MHZ on winuae is more than 2* faster as a mac mini that run on 500 MHZ

2433 ms is faster than 6060 ms

Last edited by bernd_afa on 07-Nov-2009 at 04:01 PM.
Last edited by bernd_afa on 07-Nov-2009 at 02:58 PM.
Last edited by bernd_afa on 07-Nov-2009 at 02:57 PM.

Im wondering that the x86 is still faster per clock because of the multithreading, and the massive code reuse in FFT transforms..

I was trying to find the actual name for what I thought you are trying to do, and found this page,

http://www.ibm.com/developerworks/power/cell/open_source.html

There are things like Cooly Turkey? N Log N Fourier transforms, and even an AI Operating system, as well as many other data handlign functions. Given the PPC core of the Cell, and other things, I was hoping that this might be of someuse, intrest, or even known to various people here, especially with the magic words, Open Source 8)

Im intrested in FFT, because wavelet is FFT with Sinc(x), instead of normal Hamming windowing etc, but I was looking at the method of doing a Hilbert scan of MPEG Macroblocks first, turning the 2D block into a 1D fractal array, then applying N log N Fourier algorithm to it, so that the bigger the block, the faster, relative, your algorithm becomes, until your processing the whole frame as a single scan.

Ah.. sorry.. Ive been trying to code this stuff for years, and only ever managed to get one crude basic loop running, yet never got it working right. I cant even rememebr which language or which system it was on, and that was only in teh last year, but Its still useful, to me, as a comparison of what to expect as the code is optmised and used on various machines. Just need to get it working first. 8)

_________________
The older and more respected a scientist is, the longer it takes to prove him wrong.

@RodTerl

>Im wondering that the x86 is still faster per clock because of the multithreading, >and the massive code reuse in FFT transforms..

real world apps for FFT are also here, so speedup of code is welcome
Thilo do with that code a denoiser plugin for hd-rec

http://www.hd-rec.de/Archive/fx_denoiser.lha

its also possible to do pitchshifter, or vocoder fft based

on winuae 68k emu there is only 1 thread, this thread do the calculation and Chipset emu.

that winuae is use only 1 thread most time thats also the reason, that the I7 CPU run on 3,9 GHZ.thats the so called Turbo Boost of I7

The I7 increase the clock rate intern , when only 1 thread run.
now there are newer I7 8xx or i5 xxx out that are cheaper, need less power and run with cheaper mainboards as I 920 and allow more turboboost offical.

Last edited by bernd_afa on 07-Nov-2009 at 04:45 PM.
Last edited by bernd_afa on 07-Nov-2009 at 04:43 PM.
Last edited by bernd_afa on 07-Nov-2009 at 04:43 PM.
Last edited by bernd_afa on 07-Nov-2009 at 04:42 PM.

@bernd_afa

I wrote that OS4/MOS JIT is faster than JIT in WinUAE, not that hardware with OS4/68k JIT is faster than WinUAE on x86...

Try your test on G3/G4 equivalent hardware - eg. Pentium III 800 MHz, I estimate computer with OS4 (68k JIT, G4 800 MHz) will be 2 times faster than WinUAE on this hardware (PIII).

@pavlor

>Try your test on G3/G4 equivalent hardware - eg. Pentium III 800 MHz, I estimate >computer with OS4 (68k JIT, G4 800 MHz) will be 2 times faster than WinUAE on >this hardware (PIII).

yes sure, if a X86 is only as slow as a PPC at same clockrate, the MOS or OS4 JIT must be faster, because PPC have more register and need no bytswap and need no chipset emulate.

But when OS4 or MOS do not run on faster CPU this mean nothing, because P3 or P4 is sell since many years not.an old used Athlon CPU from 2002 with 1,3 GHZ is also fast.only the P4 was a flop.

intresting is when MOS or OS4 run on PPC970 and do the benchmark here.PPC970 is lots faster than this AMCC and freescale or IBM G3 PPC at same clockrate

Last edited by bernd_afa on 07-Nov-2009 at 07:16 PM.
Last edited by bernd_afa on 07-Nov-2009 at 07:16 PM.
Last edited by bernd_afa on 07-Nov-2009 at 07:15 PM.

OK, here are some pics


http://img689.imageshack.us/img689/4558/test11.png

http://img21.imageshack.us/img21/600/test12.png

http://img145.imageshack.us/img145/2816/winuae.png
http://img267.imageshack.us/img267/1838/winuae2.png

http://img200.imageshack.us/img200/2961/timesc.png

http://img32.imageshack.us/img32/8669/all1o.png
http://img145.imageshack.us/img145/9130/all2.png

_________________
Download 499.26 Mbps, 659.94 Mbps Upload :)

@asymetrix

Nice charts but it show not so good the value, because the step is always 500 ms

maybe its possible to do the Chart in log so every double speed is ---

for example value in ms

64000 ----

32000 -----

12000-----

6400------

3200-----

1600----

800-----

400------

200------

0 -------

Last edited by bernd_afa on 09-Nov-2009 at 10:28 AM.

Hi,

I run benchmark on Amithlon with Sempron 1.8GHz . Here are results

Speed test for FFT + iFFT: (float)
time needed 1234ms for 413696 samples, => 3.80099749565124x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer)
time needed 826ms for 413696 samples, => 5.67848777770996x speed @44100Hz/stereo
Speed test for FFT + iFFT: (integer handoptimized 68K ASM)
time needed 373ms for 413696 samples, => 12.5748815536499x speed @44100Hz/stereo

Bye

I update the intro post by adding P3 650 values and a Atom N270 Netbook run at 1,6 GHZ.

seem very old X86 CPU are really slow in compare to PPC.also the N270 is very slow in clock/performance, but of course have 1,6 GHZ and overall faster as a SAM

funny that the N270 reach near same Performance /clock rate as the P3.
Maybe intel resurrect the P3 design.

Last edited by bernd_afa on 14-Nov-2009 at 03:21 PM.

@bernd_afa

Quote:

seem very old X86 CPU are really slow in compare to PPC.

More precise: WinUAE JIT seems to be slower than OS4/MOS JIT, exactly as I wrote.

@pavlor

in compare with a G4 P3 and Atom is really much slower at same clockrate, but this test also show how worse PPC CPU is and that modern X86 CPU are lots faster at same clockrate due to clever out of order execution, now and they reach higher clockrate since 2003.

but all theory doesnt help to run software faster on PPC when X86 have lots faster CPU als with performance /clockrate

but its also possible that this old X86 CPU are slower as PPC on native code too .in the past was told PPC is faster at same clockrate.this seem true until Athlon and core duo come.

now there is miss benchmark of efika.its intresting to see what performance the e300 core get in compare to 440ep

Please can a efika user do the test and post values ?

Last edited by bernd_afa on 14-Nov-2009 at 05:15 PM.
Last edited by bernd_afa on 14-Nov-2009 at 05:14 PM.

@all

Is it so that G3 and G4 FPUs were developed from 603 FPU, not from 604 FPU?

_________________
- KimmoK
// For freedom, for honor, for AMIGA
//
// Thing that I should find more time for: CC64 - 64bit Community Computer?

@bernd_afa

Quote:

but its also possible that this old X86 CPU are slower as PPC on native code too .

I think your benchmark is good for comparison of different 68k JIT/emulators (eg. OS4, MOS, Amithlon, WinUAE), but it can´t directly compare real speed of CPU architectures (eg. PPC vs x86).

Amithlon results will be much better than WinUAE results on same hardware. Again, G3/G4 is as fast as PIII on the same clockrate.

Quote:

now there is miss benchmark of efika.its intresting to see what performance the e300 core get in compare to 440ep

Look at EEMBC benchmarks:
Consumer benchmark:
440EP 666 MHz - 60.6
603e 300 MHz - 21.5
750GX 1 GHz - 124.0
74447A 1.4 GHz - 197.2

Automotive benchmark:
440EP 666 MHz - 575.0
603e 300 MHz - 196.2
750GX 1 GHz - 1155.8
7447A 1.4 GHz - 1564.1

See www.eembc.com for more details.

@bernd

Thanks for hotlinking to private links.

The FFT algorithm used here is a Radix-2 FFT. It has a complexity of O(n*log(n)).
It uses a windows size of 4096 samples, and does FFT and inverse FFT.
The result ... x speed means, the speed of processing an audio signal in realtime.
So 2x speed means, you could run it twice (if there is no other overhead). The higher the faster is the machine. It is the inverse of the Realtimefacter.

_________________
--
Author of
HD-Rec, Sweeper, Samplemanager, ArTKanoid, Monkeyscript, Toadies, AsteroidsTR, TuiTED, PosTED, TKPlayer, AudioConverter, ScreenCam, PerlinFX, MapEdit, AB3 Includes and many more...
Homepage: http://www.hd-rec.de