Loading...

Cinder Forums

Recent Topics
All Forums
- Search All
- Tags
- Author
- Advanced Search
Search
New Topic

num3ric

On my way towards the million!

in Gallery • 1 year ago

[UPDATE: You can find the source code for the GPU particle system here: http://github.com/num3ric/gpuPS]

While I'm working on my way towards building a million particle system (and the 2007 update) on the GPU (as it has been mentioned in other discussions already), I have some "byproduct projects" that I might interest some of you.

PS. Artistic value and originality will come later, for now I'm just setting up technical foundations.

For the first project I simply wanted to try out a vertex displacement mapping shader found here. My intention here was to understand how to use textures to modify vertex positions. Some screenshots (nothing new but possibly very useful).

http://github.com/num3ric/vertexSphere

Next up towards the particle system was to use an FBO to store the texture, and a VBO for the geometry (to keep the vertices on the GPU at all time). I believe it's working pretty well, I get the same kind of performance as the ImageHeightField example.

If you're wondering why use an FBO at all, well it's because the plan is to have two of them, which will be used in a ping-pong manner (like the RDiffusion example) to make the vertices dynamic.

1 000 000 static vertices at a very fluid framerate.

http://github.com/num3ric/vertexDisplac

Alright, so... what's next?

1) Figure out how to set vertex positions without loading an image, but simply by writing data into a texture. Maybe using a Surface32f and the setPixel method?

2) Modify shaders to read RBGA and convert into x/y/z and particle weight.

By then, I should be able to position 1 million fixed vertices wherever I want.

3) Add FBO ping-pong AND code a very simple (as in linear) particle motion in the frag shader?

Am I forgetting anything? At this point, it gets blurried in my mind. I'm not immediately reaching for a full-blown PS with dynamic particle generation and complex motion. Simple particle motion and I'll be happy. Also, I have been trying to understand this post which seems to explain most of what I need to do, but it's still hard to clearly understand it.

Question: Should I be using a plain Vbo instead of VboMesh? I'm a bit confused about this.

3

Replies(41)

flight404

Re: On my way towards the million!

1 year ago

Ooooh, very nice. That is quite an intense problem to tackle and I applaud you for taking it on. I played around a bit with an all shader particle system but didn't quite get the results I wanted. I did manage a few hundred thousand point particles, their positions and velocities written into the rgb of two large textures. But If I tried for a million, it would slow enough to make me think I was doing it wrong. I will get back to it eventually (although I will probably look into the binned particles that McDonald showed off).

I will look over my code to see if I have any nuggets of wisdom for you. Anyhow, great progress and I look forward to seeing you progress with this.

eight

Re: On my way towards the million!

1 year ago

Thanks for posting these files. As I am learning Cinder, I try to build and run all examples I like, and I was able to build vertexSphere on Windows 7 using Visual Studio EE.

However, when I run it, all I get is the black uneventful window and 50 % CPU consumption on a dual core. Any idea what maybe happening, where to look for problem?

Thanks.

JIC, VS project http://post.scriptum.ru/svalka/vertexSphere.zip

num3ric

Re: On my way towards the million!

1 year ago

It might be possible that your graphics card does not support "Vertex texture fetching" which is the essence of this code. Google it. You might also verify it yourself by testing the shaders independently in a GLSL development program to see they work there. Also, the files don't have the .glsl extension as in the samples, but that should not be a problem right? I'm assuming the resources are properly included since you're building the app.

One very plausible cause could be the internal format of the texture as indicated on the site from which I took the shaders. Try replacing the texture initialization with these lines:

gl::Texture::Format format;
format.setInternalFormat(GL_RGBA_FLOAT32_ATI); //Maybe GL_RGBA32F_ARB? Something else?
mTexture = gl::Texture( loadImage( loadResource( "nebula.jpg" )), format);

On my mac, both GL_RGBA32F_ARB and GL_RGBA_FLOAT32_ATI are recognized.

Furthermore, I'm testing it with the absolute current version of Cinder, and the removal of the glEnable() calls in the source by Andrew does not seem to break anything.

EDIT: After testing the code with Performance Tools, I was happy to note the transfer to the GPU working well. I get around 13 % CPU with the second example whereas I have more than 85% with vertexSphere.

eight

Re: On my way towards the million!

1 year ago

The Intel card on my laptop does not support it, neither GL_RGBA_FLOAT32_ATI or GL_RGBA32F_ARB options help, however the code runs on a desktop with geforce 8000 card.

Thanks!

Re: On my way towards the million!

1 year ago

Hello:

The application is building correctly but somehow i am having problems when I want to run it:

Running…

2010-07-10 18:03:28.987 vertexSphere[7041:a0f] Error loading /Library/QuickTime/LiveType.component/Contents/MacOS/LiveType: dlopen(/Library/QuickTime/LiveType.component/Contents/MacOS/LiveType, 262): Library not loaded: /Library/Application Support/ProApps/SharedA/Frameworks/LiveType.framework/Versions/A/LiveType

Referenced from: /Library/QuickTime/LiveType.component/Contents/MacOS/LiveType

Reason: no suitable image found. Did find:

/Library/Application Support/ProApps/SharedA/Frameworks/LiveType.framework/Versions/A/LiveType: GC capability mismatch

2010-07-10 18:03:29.052 vertexSphere[7041:a0f] Error loading /Library/QuickTime/Motion.component/Contents/MacOS/Motion: dlopen(/Library/QuickTime/Motion.component/Contents/MacOS/Motion, 262): Library not loaded: /Library/Application Support/ProApps/SharedA/Frameworks/ProCore.framework/Versions/A/ProCore

Referenced from: /Library/QuickTime/Motion.component/Contents/MacOS/Motion

Reason: no suitable image found. Did find:

/Library/Application Support/ProApps/SharedA/Frameworks/ProCore.framework/Versions/A/ProCore: GC capability mismatch

2010-07-10 18:03:29.053 vertexSphere[7041:a0f] Error loading /Library/QuickTime/Motion.component/Contents/MacOS/Motion: dlopen(/Library/QuickTime/Motion.component/Contents/MacOS/Motion, 262): Library not loaded: /Library/Application Support/ProApps/SharedA/Frameworks/ProCore.framework/Versions/A/ProCore

Referenced from: /Library/QuickTime/Motion.component/Contents/MacOS/Motion

Reason: no suitable image found. Did find:

/Library/Application Support/ProApps/SharedA/Frameworks/ProCore.framework/Versions/A/ProCore: GC capability mismatch

Assertion failed: (px != 0), function operator->, file ../boost/boost/smart_ptr/shared_ptr.hpp, line 418.

Program received signal: “SIGABRT”.

sharedlibrary apply-load-rules all

warning: Could not find object file "/Users/andrewfb/Code/cinder_0.8.1_mac/xcode/build/cinder.build/Debug/cinder.build/Objects-normal/i386/App.o" - no debug information available for "/Users/andrewfb/Code/cinder_0.8.1_mac/xcode/../src/cinder/app/App.cpp".

I am getting the "SIGABRT" error all the time, anyone can give me a hint how to fix this problem?

system: 10.6.4 graphic card: NVIDIA GeForce GT 330m

thanks!

Emmanuel

Re: On my way towards the million!

1 year ago

I haven't gotten a chance to look at Éric's project yet, but we might be able to guess at the problem with a little more information. The warnings about garbage collection are due to (harmless) issues with LiveType. The important line is:

Assertion failed: (px != 0), function operator->, file ../boost/boost/smart_ptr/shared_ptr.hpp, line 418.

which is basically just telling us something is NULL that shouldn't be - often due to a resource not loading or something similar. Are you familiar enough with Xcode to give us a stack crawl from the point of failure? If not, just run the app by hitting Cmd-Y, and then when it crashes, type bt into the console and hit enter, and paste the results...

-Andrew

eight

Re: On my way towards the million!

1 year ago

I think I got a (px != 0) assertion exception, when my resources (image and shaders) were not pointed to correctly.

Re: On my way towards the million!

1 year ago

I realized I never answered some questions earlier - sorry about that.

A VBO on its own is just a block of data - its primary advantage is that it can be either permanently stored on the GPU (static data) or can be updated very quickly if it's dynamic data. It has no geometric definition, but they are most commonly used for mesh data, and the gl::VboMesh class can make use of one or more VBOs for this purpose (more than one if there is a combination of both static and dynamic data). You can't gl::draw() a gl::Vbo because it isn't geometry - just data - but you can gl::draw() a VboMesh. In general unless you are doing something reasonably advanced you don't need to use the gl::Vbo class directly.

If you are getting deeper into VBOs, I would recommend http://www.songho.ca/opengl/gl_vbo.html for understanding how Cinder's VBO classes work under the hood.

Also I would use the GL_RGBA32F_ARB constant instead of the ATI variant, though I believe they map to the same value.

Your experiments are looking great - perhaps a group of us can turn them into a VBO guide, similar to the FBO guide.

num3ric

Réf.: On my way towards the million!

1 year ago

Alright, step 1 and 2 are done! Now the goal is in sight and I know I can reach it!

Although, frame rate does appear to slow down with that much density. Hmm... Also, it will be interesting to try it with point sprites but that will probably affect the frame rate even more.

So continuing on the second example, here are the code modifications.

The new texture code for step 1)

//====== Texture ======
mInitPos = Surface32f( SIDE, SIDE, true);
Surface32f::Iter pixelIter = mInitPos.getIter();
while( pixelIter.line() ) {
while( pixelIter.pixel() ) {
// Initial particle positions within the RBG color values. Negative values also work!
mInitPos.setPixel( pixelIter.getPos(), ColorAf( Rand::randFloat(), Rand::randFloat(), Rand::randFloat(), 1.0f ) );
}
}
gl::Texture::Format tFormat;
tFormat.setInternalFormat(GL_RGBA32F_ARB);
mTexture = gl::Texture( mInitPos, tFormat);
mTexture.setWrap( GL_REPEAT, GL_REPEAT );
mTexture.setMinFilter( GL_NEAREST );
mTexture.setMagFilter( GL_NEAREST );

I used random values for simplicity here, but it might already be interesting to model more complex geometry only out of dots.

Step 2 was too easy. Here's the new (and even simpler) vertex shader:

uniform sampler2D displacementMap;
void main()
{
float scale = 600.0;
vec4 newVertexPos;
vec4 dv;
dv = texture2D( displacementMap, gl_MultiTexCoord0.xy );
newVertexPos = vec4(scale*dv.x, scale*dv.y, scale*dv.z, 1);
gl_Position = gl_ModelViewProjectionMatrix * newVertexPos;
}

The hardest part remains : ping-pong with motion code in the frag shader. Hence, the next update will take a bit more time.

By the way, thanks a lot guys for your comments/support.

Éric

flight404

Re: On my way towards the million!

1 year ago

Exciting!

BTW, for ping-ponging FBOs, check the RDiffusion sample if you want to see how I set it up. Might save you some time.

flight404

Re: On my way towards the million!

1 year ago

I just looked at my earlier attempt to hit the million mark. Using 1000x1000 FBOs, creating 1,000,000 particles in the shader that move to a 1000x1000 res Perlin noise texture, I got 11 fps (MacBook Pro). What kind of speeds are you seeing?

BTW, you can use getAverageFps() to find out (though I am finding that to be a little buggy and sometimes it gives me super super low values... if that happens to you, just restart the project a couple times).

num3ric

Réf.: On my way towards the million!

1 year ago

As I said, I did notice a drop in the fps on the last iteration when everything is condensed in the middle of the screen. By fixing the view (removing all arcball/camera rotations), I obtain ~~19 fps~~ 29 fps on my 2008 (old model) MBP with a GeForce 8600M GT. Otherwise, it can drop to 12fps and it's really noticeable. This probably means you had equivalent code all along darn it! ;)

And my ping-pong is not even coded yet. :(

One thing to note : Trying out my vertexDisplac example on a brand new 17" MBP, I noticed that the frame rate dropped to around 10 fps while the application was absolutely fluid. So indeed, this function might not be perfectly accurate all the time.

Also, I just went back to this article by Memo, only to read that he achieves 100-200 fps with OpenCL!?! What the hell, I don't get it. 2D = 10x faster? I seriously doubt it. So how's that possible?

Oh well, it's still impressive to me. And if someone wants to try using OpenCL (when it's ready), thou shalt try and share the results.

num3ric

Réf.: On my way towards the million!

1 year ago

I can't make further progress right now because something, which I hadn't noticed at first, is blocking me. It seems that RGB values which I use as x/y/z positions are not properly interpreted. Is the RGB scale linear or not? If not that could be the cause, but I don't see how to fix it. And in every paper about this type of particle system, nothing concerning the linearity of the color space is mentioned.

Here's how it shows. Let's say I want to draw a circle using particles by passing the positions into the RGB values as I did earlier. Easy right?

ang += *** any increment here *** ;
mInitPos.setPixel( pixelIter.getPos(), ColorAf( cosf(ang), 0.0f, sinf(ang), 1.0f ) );

So we should have a perfect circle with radius 1.0.

Scale that in the shader:

float scale = 600.0;
[.....]
newVertexPos = vec4(scale*dv.x, scale*dv.y, scale*dv.z, 1);

So the radius is now 600.0.

Here's what shows up. Where is my fixed radius?

( With increment ang += Rand::randFloat(0.525f,0.530f); and format

tFormat.setInternalFormat(GL_RGBA16F_ARB) or GL_RGBA32F_ARB )

( Same increment, format GL_RGBA16 ).

Lovely patterns, but very far from what I want. In fact, just by using ColorAf( Rand::randFloat(), Rand::randFloat(), Rand::randFloat(), 1.0f) , you can see that the distribution is not random but centered (as I should have noted earlier).

If you have any idea for a possible solution, I'd be very keen to know. Here is the source:

[ Edited, link to older version : http://github.com/num3ric/gpuPS/tree/3196f34ed06da0abb345d6d1d2f4a91878b71b5a ]

flight404

Re: On my way towards the million!

1 year ago

It certainly is some odd behavior but I took a peek and one thing I noticed is you have 4 times the number of indices you need for your VBO. Of course I should preface this by saying VBOs are still pretty confusing to me so I might be waaaay off here.

Looks like you planned on drawing quads but changed to GL_POINTS but didnt change how you numbered the indices. I am able to get it to mostly look right.

mVboMesh = gl::VboMesh( totalVertices, totalVertices, layout, GL_POINTS);
for( int x = 0; x < SIDE; ++x ) {
for( int y = 0; y < SIDE; ++y ) {
indices.push_back( x * SIDE + y );
texCoords.push_back( Vec2f( x/(float)SIDE, y/(float)SIDE ) );
vCoords.push_back( Vec3f( x/(float)SIDE, 0, y/(float)SIDE ) );
}
}

I still get some stragglers though. If I draw them in a ring and set the y position = to a fraction of arg, i get the following spiral... note the arc of incorrect radius. Still not sure about that.

num3ric

Réf.: On my way towards the million!

1 year ago

This thread is getting long and messy but... I found the solution so I might as well update it. This was missing (the format of the FBO):

gl::Fbo::Format format;
[...]
format.setMinFilter( GL_NEAREST );
format.setMagFilter( GL_NEAREST );

After removing the extra indices (thanks), I also removed the position buffer in the dummy VBO since it's not used and its removal doesn't break anything.

Now everything ***should*** be behaving normally:

flight404

Re: On my way towards the million!

1 year ago

Sweet! So whats next?

num3ric

Re: On my way towards the million!

1 year ago

I don't want to overload your mailbox but if you're asking... (Writing this post also helps to structure my thoughts.)

After having re-read the 2007 slides, there are a few available simulation options. Euler is where I'd like to go as it should be more polyvalent and might **eventually** allow for more unique movements. It's similar to what we always do : update the positions by adding the velocities (which themselves can be updated with an acceleration value).

1) new_v = v + a

2) new_p = p + new_v

There are two ways of doing so (see p.19):

1) Have 4 FBOs and 2 separate shaders. I would ping-pong the velocities and then ping-pong the positions reading in the new velocity values in the pixel shader.

2) Multiple render targets with 2 FBOs (each containing both pos and vel textures) and 1 shader would be more efficient (both in terms of coding and performance). It's not currently supported ( http://forum.libcinder.org/#Topic/23286000000153124).

As for Vertlet, I'm not too sure how to implement the formula but maybe it could be done in a single ping-pong operation and 2+1 FBOs, which means no multiple render targets issue.

In the end (!!!),

Since I'd rather not make any compromise, I'll try to modify the Fbo source. I have no idea if I can get anywhere but I'm already learning a lot how just by reading the (awesome) source code so it will be interesting.

I'm therefore putting this project on hold for a moment.

eight

Re: On my way towards the million!

1 year ago

FWIW: I am getting consistently around 5 fps for these examples on NVIDIA GeFORCE 8400, Intel Core 2 Quad CPU Q9400 @ 2.66 GHz, 8 GB RAM, Windows 7 64 bit, compiled with MSVS 2008. It seems it is half the speed of num3ric's. Any comments on how to speed this up? Thanks.

num3ric

Re: On my way towards the million!

1 year ago

( Haha, that's what we call a detailed description. :P )

If I fix everything (no Arcball/camera) as it might be when I have particle movement, I actually get around 60 fps now with no enableDepthRead (always with 1 million points)! Otherwise I get 29 fps so 5 fps is definitely quite slow.

Are you using the last version of the gpuPS which has its VBO cleaned out? That helped quite a bit.

The main hardware that should matter here is your graphics card. You have a GF 8400, mine is a 8600M GT. I'm far from an expert, but at the rate at which these cards are evolving, it might be part of the answer.

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

Edit: Are you reading your fps upside-down? I didn't take the time to figure out a proper way to display it...

eight

Re: On my way towards the million!

1 year ago

Yes, fps are upside-down, which is funny, but OK.

gpuPS as is runs at 7 fps, with enableDepthRead 3-4 fps.

I guess I'll get one of the newer NVIDIAs to try.

num3ric

Réf.: On my way towards the million!

1 year ago

DONE!

I updated the source code with a simple rotation motion - not like in previous images/video because it's not a good starting point for personal explorations and is also quite messy. ~~You can have a look but~~ ~~IT WON'T WORK ON YOUR MACHINE~~ ~~until Cinder has the multiple render targets Fbo source code. ( Although you could use my Fbo.h/Fbo.cpp to recompile Cinder.) It is currently in development :~~ http://forum.libcinder.org/#Topic/23286000000153124

http://github.com/num3ric/gpuPS

Further things to try:

- Point sprites

- Passing in different uniforms as points of attraction or repulsion

- Perlin on the shaders?

- Particle generation

... and the list goes on! If you do anything with this, I'd be pleased to hear about it.

I've had my dose of particles for now... I'll explore other things for a while. :)

Again, thanks to the Barbarians for everything!

Éric

eight

Réf.: On my way towards the million!

1 year ago

Looks great. What framerate did you get?

flight404

Re: On my way towards the million!

1 year ago

Very very exciting. I look forward to having some down time so I can learn what you have done here. It looks and behaves great and the density of the particles is inspiring some ideas for sure.

num3ric

Réf.: On my way towards the million!

1 year ago

@bobabe

Just as before (with camera rotation - no alpha blending - no depthRead), ~ 30 fps on my end-of-2007 MBP. It does slow down sometimes when there's very high density. With the version I showcased here in the pictures, it doesn't slow down much and it's very nice to play with in real-time.

In general, I noticed that the Arcball rotation is always perfectly fluid while the particles might slow down. As I said earlier, fixing the view would make things twice as fast but it kills the mood. ( Unless there was a way to rotate the scene in the shader, which could be possible... ) There could also be things that I'm doing wrong or further optimizations that I'm not aware of.

If all fails, you can always drop to 512 x 512 = 262 144 particles but then you're not reaching the goal of this thread. :)

eight

Réf.: On my way towards the million!

1 year ago

30 fps -- nice! I am getting my GTX 460 next week, your code will be first to run.

Re: On my way towards the million!

1 year ago

This is awesome - really nice work...

num3ric

On my way towards the million!

1 year ago

~~ Thanks.

Only now do I find this:

http://code.google.com/p/creativecomputing/source/browse/trunk/cc.creativecomputing.gpu.particles/

I didn't think there was source code of available for this kind of shader p-systems... Turns out there was, from two years ago haha! I didn't take the time to really look at it, but there might be some very useful lessons in there.

Here's his video: http://www.vimeo.com/2826441 (EDIT: Hmm, it's cg shaders though.)

Oh well, what can I say, struggling my way through this project was more rewarding... :)

eight

Re: On my way towards the million!

1 year ago

NVIDIA GTX 460*.......... gpuPS 58 fps

!!!

^{*EVGA 768-P3-1362-AR GeForce GTX 460 Graphics Card - PCI Express 2.0 x16 - 768 MB GDDR5 SDRAM}

num3ric

On my way towards the million!

1 year ago

Not bad, not bad. :)

Also, I just updated the gpuPS with the latest version of Cinder (thanks Andrew!). Street-coding that is; I'm actually using internet right off the street to push things on Github.

http://github.com/num3ric/gpuPS

~~I also noticed that the frame rate takes a bit more time to reach full speed, but it's a minor difference.~~ Nevermind. :)

PlumCantaloupe

Re: On my way towards the million!

1 year ago

Wow, amazing great job! I can't wait to play with it - thanks for posting the source.

Re: On my way towards the million!

1 year ago

Just found this thread and saw you mentioning my source code, I decided to use cg over glsl because it has two nice features that make it really handy in context of particle systems, interfaces with structs + dynamic arrays, as you might have seen in my code I have an interface force and different structs implementing it, using dynamic arrays you can now plug together forces dynamically, you can also put forces together for blending. I find this approach also more handy than opencl haven't checked it for performance though. Once you have gotten the concept of handling data in textures you can really do great stuff, so I also managed to add springs to create dynamic trails.

The real master of particles is Matt Swoboda http://directtovideo.wordpress.com/ you have to see the crazy fairlight demos

http://www.youtube.com/watch?v=ON4N0yGz4n8

http://www.youtube.com/watch?v=Grqb1aIa_4s

You can get some concepts of his particle system here:

http://directtovideo.wordpress.com/2009/10/06/a-thoroughly-modern-particle-system/

You can also get some background info here:

http://vimeo.com/14458079

letsgooutside

Re: On my way towards the million!

9 months ago

Revisiting this code with some new hardware and a few tweaks.

How about 6,000,000 at over 30 fps?

Here is most I could load given the memory available... ~8.2m at 20fps.

http://www.BanTheRewind.com

num3ric

Re: On my way towards the million!

9 months ago

I just yell "WAAAAAAAAA!"

This is fantastic, you're pushing it!

I'd love to get a pull request with your tweaks! I need to revisit the code and add more detailed comments too.

Oh Cinder, I miss you...

Edit: We now need to find awesome ideas to put this sh*tload of particles to good use!

letsgooutside

Re: On my way towards the million!

9 months ago

Will clean up and share when I get further along. I'm positive I can break into eight digits using two textures / VBOs. Just working on getting them to play nice together. For the next month or so, I have a machine with 6GB vram and 3.6GHz GPU using ATi Crossfire technology. It isn't as pretty as NVidia and has horrible vertical sync issues, but it really opens the door for some experimentation.

http://www.BanTheRewind.com

eight

Re: On my way towards the million!

9 months ago

Any plans for parallelization using CUDA or opencl?

--8

letsgooutside

Re: On my way towards the million!

9 months ago

8,479,744 (2912 ^ 2) particles. The alpha is set to 0.1 on each point so might look a little more sparse than it actually is.

8,479,744 Particles from fashionbuddha on Vimeo.

http://www.BanTheRewind.com

eight

Re: On my way towards the million!

9 months ago

Nice! In your crossfire setup do you have two cards, and then create two textures/VBO's for each card? If so, this is different from, say CUDA, which parallelizes with regards to the GPUs on each card, and so conceivably, could be more efficient. Right?

--8

letsgooutside

Re: On my way towards the million!

9 months ago

I have three interlinked HD 6950s on this machine. Crossfire does a frame on each GPU, whereas with OpenCL you can explicitly divide tasks across the hardware. Just installed the OpenCL driver today and will be taking a crack at it.

http://www.BanTheRewind.com

Sable Raf

Réf.: On my way towards the million!

3 months ago

This is really great. Is the code available somewhere?

num3ric

Réf.: On my way towards the million!

3 months ago

If you're talking about my source code, I have updated the first post with the link to my github repo (so that no one has to browse through this mess to find the link).

If your comment concerns letsgoooutside's project, then I can't comment.

Éric
http://num3ric.com/

Sable Raf

Réf.: On my way towards the million!

2 months ago

Thank you for your answer. I'll check out the repo. If letsgooutside ever releases his own improvements, I'm also interested.

© 2012 ZOHO Corp.
All Rights Reserved.
Privacy
Terms & Conditions
RSS
Powered by