int x = index % xres;
int y = index / xres;


As noted by Diogo Teixeira, one way to get rid of the mod/div operations for POT buffers would be:


int precalc_a = ( xresPOT - 1 );
int precalc_b = (int)log2( xresPOT );
int x = index & precalc_a;
int y = index >> precalc_b;


.

I finally took the time to clean up this project and make a public release. In the packages you will find binaries and source code for the application, so if you’re thinking of using it and come up with something nice, let me know. After the first release i have fixed some problems with memory access and crashes. I have also added a simple UI for easy pick of tags from the “data/tags” folder on your hard-drive.

What is it?
- GML Viscosity is an experiment. It is application used to draw GML tags on a viscous liquid rendered purely on the GPU side.

What does it do?
- It liquifies your tags, in a way you can hardly read them but still looks cool (thats graffiti).
- It is possible to load gml files from the disk or directly from the “lots of zeros”book database.

How can I interact?
- Not implemented on this version. This version selects and draw gml files randomly from the web or from a tags folder. For a smart person it should be easy to add mouse support. The code is also prepared to support multitouch so, it should be easy to implement TUIO and create your multitouch version.
- Use ‘r’ key to randomly fetch a new tag from the website and if available it will draw it on the screen.
- Use the GUI window to pick any tag from the folder “tags” inside data folder.

Where can i get it?
Windows Version | MacOSX Version

Have fun.

So imagine we have 2 image buffers Image1, Image2. Usually to change data from Image1, you would simply access it and write directly back to the same positioion. Now, when we’re talking about a shader fragment program we can’t simply do that. You may ask now, what’s the solution? Ping-pong it!

Here is what i’m talking about:

//
// Initialization
//
int W = 100;
int H = 100;
ImageArray = new int[2][W*H];
int CurrActiveBuffer = 0; // Current active buffer index
 
 
//
// Mainloop
//
for( int j=0; j<H; j++ )
{
	for( int i=0; i<W; i++ )
	{
		// ERROR! Write to same buffer. Not possible in gpu shader
		//Image1[i+j*W] = Image1[i+j*W] * 2;  // Mul by 2
 
		// Ping-pong version
		int src = CurrActiveBuffer; // Current active buffer (Input)
		int dest = 1-CurrActiveBuffer;  // Back buffer (Output)
		ImageArray[dest][i+j*W] = ImageArray[src][i+j*W] * 2;  // Mul by 2
	}
}
 
// Swap back and front buffers (read becomes write and vice-versa)
CurrActiveBuffer = 1-CurrActiveBuffer;    // CurrActiveBuffer ? 0 : 1;

As you can see we start with buffer 0. That’s where we get out data from (read) and write it to Buffer 1. Once the operation is done, we swap buffers every frame. This way we’ll be able to use last iteration’s data as input for the next iteration.

Now let’s put this in OpenGL language. I will be using a FrameBufferObject (FBO) and 2 textures here. The framebuffer will be holding to both textures as 2 Color Attachments. So let’s get coding:

//
// Initialization
//
int W = 100;
int H = 100;
int FboID;
int TexID[2];
int CurrActiveBuffer = 0;  // Current active buffer index
 
// Create the frambuffer
glGenFramebuffersEXT( 1, &FboID );
glBindFramebufferEXT( GL_FRAMEBUFFER_EXT, FboID );
 
// Create 2 textures for input/output.
glGenTextures( 2, TexID );
 
for( int i=0; i<2; i++ )
{
	glBindTexture( GL_TEXTURE_2D, TexID[i] );
	glTexParameteri( GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR );
	glTexParameteri( GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR );
	glTexParameteri( GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE );
	glTexParameteri( GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE );
        // RGBA8 buffer
	glTexImage2D( GL_TEXTURE_2D, 0, GL_RGBA, W, H, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL );
	if( _hasMipmapping )  glGenerateMipmapEXT( GL_TEXTURE_2D );
}
 
// Now attach textures to FBO
int src = CurrActiveBuffer;
int dest = 1-CurrActiveBuffer;
glFramebufferTexture2DEXT( GL_FRAMEBUFFER_EXT, 
                           GL_COLOR_ATTACHMENT0_EXT, 
                           GL_TEXTURE_2D, TexID[src], 0 );
glFramebufferTexture2DEXT( GL_FRAMEBUFFER_EXT, 
                           GL_COLOR_ATTACHMENT1_EXT, 
                           GL_TEXTURE_2D, TexID[dest], 0 );
);
 
 
 
//
// Mainloop
//
int src = CurrActiveBuffer;
int dest = 1-CurrActiveBuffer;
 
FBO.Bind();
 
glDrawBuffer( dest );
 
glBindTexture( GL_TEXTURE_2D, TexID[src] );
ShaderProgram.SetTextureUniform( 0 );
RenderScene();
 
FBO.Unbind();
 
 
// Swap back and front buffers (read becomes write and vice-versa)
CurrActiveBuffer = 1-CurrActiveBuffer;    // CurrActiveBuffer ? 0 : 1;

Why would you want to ping-pong? Well for instance imagine you are doing a water effect in the gpu. You’ll need to access data from your previous buffer right? Using the CPU that would be trivial as you have the last frame’s memory buffer allocated somewhere which you can read/write access directly. That doesn’t really work on the gpu side that way (but we’re getting closer with CL, CUDA, DC)

Have fun.

On this case we will need 4 points: 2 anchor points (the line end points) and 2 control points. The control points won’t really touch the curve, they work more as directional information on the curve itself.

As we need to send this data to the shader i have decided to use  LINES as input primitive, 2 points define a line so it’s perfect, we’ll use that for the 2 anchor points. As for the control points 2 different texture units (glMultiTexCoord3f) attached to the line’s vertex data will do. Using geometry shaders besides setting the input primitive type we also need to set the output type. LINE_STRIP is fine, as it works perfectly for what we’re doing. That’s all on the application side.

On the vertex shader side it’s pretty simple:

[VERTEX SHADER]
void main( void )
{
    gl_FrontColor = gl_Color;
    ControlPoint1 = gl_MultiTexCoord0.xyz;
    ControlPoint2 = gl_MultiTexCoord1.xyz;
 
    gl_Position = gl_Vertex;
}

What the code is doing is sending the data further down the pipeline to the geometry shader, where all the magic happens. At this stage, having both Anchor and Control points we can now define the curve by a given detail. Think of detail as a number of step-points along the curve which makes it look smoother or flatten (tesselation, subdivision, smoothing, etc).

Both control points are sent by the application for the geometry shader, still, as in the the vertex shader comes before the geometry shader, we will need to send them down on the vertex-shader, otherwise the GS won’t be able to “see” them (i know, hurray for Cg). We’re now almost done. By using the function from the above link(s) and as shown below we compute the curve with a given detail on the geometry shader, by generating new vertices along the curve purely on the gpu side. I think it to be pretty straightforward and the code should be self-explanatory.

[GEOMETRY SHADER]
uniform int g_Detail;
varying in vec3 ControlPoint1[];
varying in vec3 ControlPoint2[];
 
// Found in nvidia sdk
vec3 evaluateBezierPosition( vec3 v[4], float t )
{
    vec3 p;
    float OneMinusT = 1.0 - t;
    float b0 = OneMinusT*OneMinusT*OneMinusT;
    float b1 = 3.0*t*OneMinusT*OneMinusT;
    float b2 = 3.0*t*t*OneMinusT;
    float b3 = t*t*t;
    return b0*v[0] + b1*v[1] + b2*v[2] + b3*v[3];
}
 
void main()
{
    vec3 pos[4];
    pos[0] = gl_PositionIn[0].xyz;
    pos[1] = ControlPoint1[0];
    pos[2] = ControlPoint2[0];
    pos[3] = gl_PositionIn[1].xyz;
    float OneOverDetail = 1.0 / float(g_Detail-1.0);
    for( int i=0; i<g_Detail; i++ )
    {
        float t = i * OneOverDetail;
        vec3 p = evaluateBezierPosition( pos, t );
        gl_FrontColor = gl_FrontColorIn[0]; 
        gl_Position = gl_ModelViewProjectionMatrix * vec4( p.xyz, 1.0 );
        EmitVertex();
    }
 
    EndPrimitive();
}

What’s next? That is up to you. You’re not going to leave me with all the work, are you ?

Download the example + source.
You will also need to install Vitamin 0.5.6 as the project is built with it.

Next step was to optimize the cube generation by lowering from a 24 vertex cube to 14 vertex triangle strip. Things got better, but nothing close to my expectations. I was not satisfied, i mean, i had alot of cubes on screen (100K which was not that much) and that was it, nothing else. We’re talking about 20fps or so, for 100.000 cubes (around 1.2 million triangles per frame). Later on i added vertex normals to the geometry shader and started to work on some lighting/shadowing, but i ended up going back on the rendering side of the job. Meanwhile, i was speaking with a friend of mine about this idea and we were discussing ways to compute lighting but i couldn’t stop thinking about my real problem. So it came to me.

Previously, i have done some experiments with opengl hardware instancing, but never got  to do much about it .  What better time than now, so i grabbed the project and took it for a spin. After a few hours i had the same amount of cubes on screen with a much much better framerate. Quickly implemented some eye-candy (coloring, texturing, vertex lighting), some tweaking here and there and as i was listening to Mr. Peter Broderick (hi, i love you man) added some audio analysis to the feature list.
Last but not least, a kind of “Brownian Motion” was used to generate points in space, increased the cube count to 512*512 and watched it flow ( at 20fps ).

In conclusion, Hardware instancing was much easier to implement  and performance seems much better at first sight. Above is a video of 262.144 audio-reactive cubes with GPU animation and basic lighting at around 20fps. For my video card i think that is very good. On a sidenote, i have not given up on the geometry shaders. I am not sure what will be my next step regarding the subject (back to geometry shaders?) but for now this is it. Hardware Instancing kicked Geometry Shaders in the ass.

Download: http://victamin.googlecode.com/files/JavaCL_1_4b.zip

Have fun!

To achieve this i’ve picked Vertex Texture Fetch. It allows you to access texture data on the vertex shader. Pass a texture to the vertex shader and offset the surface vertices by the pixel’s intensity. To improve the looks, bump-mapping with multiple lights have been used. Everything presented here is generated on the GPU, even the objects. Runs nicely on a Nvidia 8200M. Lots of  space for improvement.

]]> http://www.pixelnerve.com/v/2009/11/04/hmm-greebles/feed/ 0 http://www.pixelnerve.com/v/2009/10/30/opencl-4-java/ http://www.pixelnerve.com/v/2009/10/30/opencl-4-java/#comments Fri, 30 Oct 2009 18:09:42 +0000 victormartins http://www.pixelnerve.com/v/?p=834 I finally took  the time to play with OpenCL. I was motivated by the particle example from Rui Madeira. After speaking with him, he gave me a few other links on more examples like Memo Akten‘s 1.000.000 particles running with mouse interaction on the GPU, the very NVidia’s first OpenCL application example, etc. I was intriged!  So i took the day to play with OpenCL4Java and ported Rui’s example to Java running on Processing’s IDE.  I’ve tested it with a Intel Quad core and i found the Rui’s sample to crash with GTX280 videocard. I didn’t gave it much thought but it might be for that for-loop in the program on the particle-particle iteration. In the other hand, Memo’s example ran easily on the gpu side. I should have made an example of mine, but it was easier to just port their examples and get things running. That was the main goal.

I have only tested this under Windows XP 64bit with ATI Stream SDK and Nvidia drivers. If you find any problem, please report.
One other thing: Memo’s example isn’t the real thing. It’s simply the CL program. So you won’t be able to get all the fuzzy million particles around. Not yet.

Library/Examples available here: http://victamin.googlecode.com/files/OpenCL4Java_1b.zip

Installation steps:
1. Copy the library to processing’s libraries folder
2. Install ATI Stream SDK (i have packed the OpenCL.dll file, still you might have to install the whole pack). If you’re using Nvidia, install the OpenCL drivers and toolkit
3. Open the example and run it.

Enjoy and have fun!

Fur experiment on the gpu. I have found a paper on the internet speaking about fur-rendering, i fount it to be easy so i wanted to give it a try. I believe the paper can be found at ATI’s website. Works pretty well. Next step is adding forces to improve the fur animation (it already animates if you look closely).