GPU crunching question

[Vyper]

Tried it on Vista X64 and a 8800GTX with newest beta driver.

It seems to crash after 30 minutes of running, the console text stopps and nothing happens, the program is running in background though.

Stopped after 2 hours of non progress.

I will try it later when i come home instead of remoting my home computer trying it out

Kind Reg Vyper

[Devaster]

send here a stderr.txt please ....

[Vyper]

Hehe chill    Im gonna upload a stderr.txt within 5 hours or so..

Kind Reg. Vyper

[Devaster]

there is some issue with rapidmind and remote gpus: https://developer.rapidmind.net/documentation/kb/gpus-remotely

thats may be your case .

new .exe - small optimization in fft calc and more speedy method on data upload ,other method to show progress :
1st number - fft size
2nd number - fft chunk calculated and analyzed
3rd number - total count of fft chunks.

[attachment deleted by admin]

[Vyper]

Ok here is a stderr.txt from the first client, the progress seemed slow so i aborted it..

Check if it's ok even if i aborted it.

Going to try your second compile now.

Kind Reg Vyper

P.S  How long does it take and how far shall the counter go? for the 2nd value that just climbs climbs and climbs?? D.S

[attachment deleted by admin]

[pepperammi]

I quickly set that second one up for a run over night. Seemed to be flying quite nicely when I left it but I just had a look and seems to have stuck at "after powerspectrum 8 131071 131072". Looks like stopped at the last chunk or something. stderr.txt attached.
Still using the cpu at max by the way.

This is on a Vista x64 system with a PD945 @3.62 and 7950GT OC edition with 512MB and using current latest beta drivers. Bit strapped for time but will try and get it on a second machine and/or this one with the latest 'official' drivers instead of beta ones.

[attachment deleted by admin]

[Devaster]

today i will check it what is the problem ...

app is maximally unoptimized and without any exception trap 

[Freddy]

Running your client on 8800GTS/winXP, everithing seems nice so far...
I'll upload the txt file later.
One core is used at about 80% while exe is running. (X2 4400+@2,5Ghz)

Stoped Running at powerspectrum 8 131071 131072 CPU    is used by 100% now.


[attachment deleted by admin]

[Devaster]

there is generated a exception but there is any trap so application stop ...

[Maxxx]

@Devaster

Running  client on a ATI 1950XT -> winXP

core 1 is at  70-90%  (Core2Duo@3333MHZ)

Stoped  at powerspectrum 8 131071 131072 CPU   

Lg Maxxx
good luck

[Arnulf]

I ran the client on my computer with dual Opterons and a nVidia 7800GS card ( AGP ).

Processor usage - one core at 60 to 70%.

Attached stderror.txt

A.

[attachment deleted by admin]

[Devaster]

for now i am trying in cooperation with rapidmind guys to maximally parallelize the fft code ...

for now whole analyze is running in one cycle in serial my idea is this (thanx to vyper for basic idea )   :

separate  the whole analyze process to  three stages:
1. calculate part ffts on whole datas at one time
2. calculate whole powerspectrum at one time
3. check for the spikes/Gaussians/pulses at one time

usage of the rapidmind can parallelize code efficiently ...

1. trying encapsulate the walkthrough code over data array into shaders code then may be in separate shader unit run many ffts side by side ... 

[Vyper]

Heyy i got a thanks.. Me is soo  now ...

Good luck dude.. Keep in touch with us, and let us know further hicupps so you can get another angle to solve issues..

Kind Regards Vyper

[Devaster]

i have found in my code a big error : seti uses fftlen as is but raoidmind uses fftlen as 2^fftlen  so it calculates ....

[Vyper]

debug...debug...debug... and yet again debug

When tweaking is done the appropriate result will show ..

I remember in the good old days when i programmed O/S friendly assembly with library calls on the Amiga, i was so disturbed with that the library routine to plot pixels was soo slow, then i discovered that with a simple library patch i managed to increase the performance with 1100 - 1200 % when i wrote the pixel routine in pure assembly instead, and boy have i used the trace register function alot in these days.. 15% of time is writing the code 85% of it is pure tracing too see that the program does what i've intended in my brain  ..
Boy how fun the code could be sometimes but work pays off and one day (voila) the code does what it was intended to do .. Many nights!! Late hours and alot of caffeinepills and just when the code is done u notice that the clock is about 05.00 at night and the school started at 08.00 .

Why the hell sleep   , keep it up devie..

I had to write some nostalgia here to trigger you further 

Kind Regads Vyper