Multithreaded Quaternion Julia Sets renderer

Index > Linux > Multithreaded Quaternion Julia Sets renderer

Goto page Previous 1, 2, 3 Next

Author

Thread

comrade

Joined: 16 Jun 2003
Posts: 1150
Location: Russian Federation

comrade 24 Mar 2013, 08:44

randall wrote:

Yes, but it requires SSE 4 support. I have old Core2 Duo at home so only SSSE3.

An interesting project (somebody was talking about compos on this board) is emulating extended instruction sets on older CPUs. There are several approaches to try out performance wise (similar to VMs):

1) Dynamic code rewriting. Some of these extended instructions are so long that you can easily patch them over with a call to a subroutine.
2) Emulation via the illegal instruction exception.

_________________
comrade (comrade64@live.com; http://comrade.ownz.com/)

24 Mar 2013, 08:44

revolution
When all else fails, read the source

Joined: 24 Aug 2004
Posts: 20299
Location: In your JS exploiting you and your system

revolution 24 Mar 2013, 09:54

comrade wrote:

1) Dynamic code rewriting. Some of these extended instructions are so long that you can easily patch them over with a call to a subroutine.
2) Emulation via the illegal instruction exception.

Well you can combine both by rewriting upon an exception. However, with AVs being particularly fussy and inaccurate, writing to your code section can cause problems with false triggering. !@#@$% AVs, should be banned anyway, they are useless, provide no protection, gives users a false sense of security with bad programs, gives users a false sense of danger for perfectly good programs, they harm performance and create havoc when they detect your own OS as malware. And many people even pay to put up with such nonsense!

24 Mar 2013, 09:54

randall

Joined: 03 Dec 2011
Posts: 155
Location: Poland

randall 24 Mar 2013, 11:46

comrade wrote:

An interesting project (somebody was talking about compos on this board) is emulating extended instruction sets on older CPUs. There are several approaches to try out performance wise (similar to VMs):

1) Dynamic code rewriting. Some of these extended instructions are so long that you can easily patch them over with a call to a subroutine.
2) Emulation via the illegal instruction exception.

Intel has such a tool:
http://software.intel.com/en-us/articles/intel-software-development-emulator

24 Mar 2013, 11:46

randall

Joined: 03 Dec 2011
Posts: 155
Location: Poland

randall 03 Apr 2013, 11:01

I have updated my program. Now it is fully multithreaded. See starting post for more details.

03 Apr 2013, 11:01

HaHaAnonymous

Joined: 02 Dec 2012
Posts: 1178
Location: Unknown

HaHaAnonymous 03 Apr 2013, 16:01

[ Post removed by author. ]

Last edited by HaHaAnonymous on 28 Feb 2015, 21:07; edited 1 time in total

03 Apr 2013, 16:01

keantoken

Joined: 19 Mar 2008
Posts: 69

keantoken 13 Apr 2013, 14:52

Well then how does one protect their computer? You have to do it somehow.

I will certainly study the multithreading code! Thanks a lot Randall!

Here's the results with my FX8350 at 4GHz with 1880MHz RAM:

EDIT: Actually here's the results:

$ time ./qjulia -v

real 0m1.200s
user 0m8.864s
sys 0m0.016s

And here's the results with a 2560x1440 resolution like the original version:

$ time ./qjulia -v

real 0m4.671s
user 0m35.442s
sys 0m0.043s

13 Apr 2013, 14:52

Turbo Lover

Joined: 22 Feb 2013
Posts: 32

Turbo Lover 14 Apr 2013, 17:06

this is awesome! how long did it take you to develop this??? you must be very experienced in asm!

14 Apr 2013, 17:06

randall

Joined: 03 Dec 2011
Posts: 155
Location: Poland

randall 15 Apr 2013, 16:44

Thanks! It took my three or four days for single threaded version and another three days for multithreading code. keantoken motivated me to implement multithreaded version (thanks Smile

).

The code is rather simple. You just have to know the technique. This is 'standard' distance field raymarching:
http://www.iquilezles.org/www/articles/raymarchingdf/raymarchingdf.htm

I knew the algo because some time ago I have written similar program using C++ and OpenGL. Using GPUs it can run in real-time.

15 Apr 2013, 16:44

tthsqe

Joined: 20 May 2009
Posts: 767

tthsqe 07 Jun 2013, 00:32

Ok, I just want to get some verification on how this works:
So, at point (X,Y,Z) in three space,
You form the quaternions

Code:

z'=1+0i+0j+0k
z=0+Xi+Yj+Zk

Then you iterate

Code:

z'=2*z'*z
z=z*z+c

After a few iterations, you estimate the distance from the point (X,Y,Z) to the surface by

Code:

DE = 0.5* log(|z|)*|z|/|z'|

I see that c is your g_Quat parameter, but is the choice

Code:

z=0+Xi+Yj+Zk

arbitrary?

07 Jun 2013, 00:32

tthsqe

Joined: 20 May 2009
Posts: 767

tthsqe 07 Jun 2013, 02:51

Also, how do you know that the iteration for z' is not

Code:

z'=z'*z+z*z'

07 Jun 2013, 02:51

randall

Joined: 03 Dec 2011
Posts: 155
Location: Poland

randall 07 Jun 2013, 09:53

'c' parameter is arbitrary, different values generate different shapes just like in standard 2D Julia Sets.

It can be shown that derivative of Zn+1 = Zn*Zn + c is simply Z'n+1 = 2*Z'n*Zn.

Theory:
http://www.iquilezles.org/www/articles/juliasets3d/juliasets3d.htm
http://paulbourke.net/fractals/quatjulia/
http://devmaster.net/forums/topic/3432-ray-tracing-quaternion-julia-sets-on-the-gpu/

07 Jun 2013, 09:53

tthsqe

Joined: 20 May 2009
Posts: 767

tthsqe 07 Jun 2013, 18:51

Ah, sorry for my long post. Even though (z*z)' = 2*z*z' is not correct for quaternions, it seems that is has been shown that the iteration of z' = 2*z*z' does give a correct lower bound on the distance to the set. My bad.

In that case, you can probably save a lot of time by not keeping track of z', but only its modulus |z'|, or even better the square of its modulus.

07 Jun 2013, 18:51

randall

Joined: 03 Dec 2011
Posts: 155
Location: Poland

randall 09 Jun 2013, 12:05

I have tested my program on Haswell CPU (Core i7-4770K CPU @ 3.50GHz).

It takes 870 ms to generate 1280x720 image. It is almost 10x faster than my previous CPU (Core2 Duo @ 1.86GHz).

09 Jun 2013, 12:05

Melissa

Joined: 12 Apr 2012
Posts: 125

Melissa 09 Jun 2013, 14:29

bmaxa@maxa:~/fasm/examples/qjulia$ time ./qjulia

real 0m1.190s
user 0m4.628s
sys 0m0.000s

Wow 50% faster then i5 ivy bridge Wink

09 Jun 2013, 14:29

keantoken

Joined: 19 Mar 2008
Posts: 69

keantoken 02 Aug 2017, 17:07

AMD Ryzen 5 1500X 3.5GHz:

$ time ./qjulia

real 0m0.768s
user 0m5.970s
sys 0m0.004s

02 Aug 2017, 17:07

randall

Joined: 03 Dec 2011
Posts: 155
Location: Poland

randall 02 Aug 2017, 18:11

keantoken wrote:

AMD Ryzen 5 1500X 3.5GHz:

$ time ./qjulia

real 0m0.768s
user 0m5.970s
sys 0m0.004s

Nice, Ryzen seems quite fast.

_________________
https://github.com/michal-z

02 Aug 2017, 18:11

Furs

Joined: 04 Mar 2016
Posts: 2493

Furs 02 Aug 2017, 21:37

This is pretty amazing stuff. I've always been fascinated by quaternions since I can't really understand them (I'm not much of a math guy). Smile

02 Aug 2017, 21:37

keantoken

Joined: 19 Mar 2008
Posts: 69

keantoken 03 Aug 2017, 05:29

Using a tile size of 146 (default 80) gives about 30% performance gain for my CPU:

$ time ./qjulia

real 0m0.668s
user 0m4.683s
sys 0m0.004s

03 Aug 2017, 05:29

randall

Joined: 03 Dec 2011
Posts: 155
Location: Poland

randall 03 Aug 2017, 09:29

keantoken wrote:

Using a tile size of 146 (default 80) gives about 30% performance gain for my CPU:

$ time ./qjulia

real 0m0.668s
user 0m4.683s
sys 0m0.004s

Ryzen 5 has 4 cores or more?

_________________
https://github.com/michal-z

03 Aug 2017, 09:29

keantoken

Joined: 19 Mar 2008
Posts: 69

keantoken 04 Aug 2017, 18:03

It has 4 cores, 2 threads per core. I checked the core counter in qjulia and it reported 8. Oddly enough however, the CPU only reports 25% usage while qjulia is running.

EDIT: Nevermind, it uses the whole CPU, my CPU usage program showed incorrectly.

04 Aug 2017, 18:03

Goto page Previous 1, 2, 3 Next

< Last Thread | Next Thread >

Forum Rules:

You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
You cannot attach files in this forum
You can download files in this forum