Weird behavior if DataTextures are not square (1:1)
Asked Answered
E

1

6

I have a pair of shader programs where everything works great if my DataTextures are square (1:1), but if one or both are 2:1 (width:height) ratio the behavior gets messed up. I can extend each of the buffers with unused filler to make sure they are always square, but this seems unnecessarily costly (memory-wise) in the long run, as one of the two buffer sizes is quite large to start. Is there a way to handle a 2:1 buffer in this scenario?

I have a pair of shader programs:

  1. The first is a single frag shader used to calculate the physics for my program (it writes out a texture tPositions to be read by the second set of shaders). It is driven by Three.js's GPUComputeRenderer script (resolution set at the size of my largest buffer.)
  2. The second pair of shaders (vert and frag) use the data texture tPositions produced by the first shader program to then render out the visualization (resolution set at the window size).

The visualization is a grid of variously shaped particle clouds. In the shader programs, there are textures of two different sizes: The smaller sized textures contain information for each of the particle clouds (one texel per cloud), larger sized textures contain information for each particle in all of the clouds (one texel per particle). Both have a certain amount of unused filler tacked on the end to fill them out to a power of 2.

Texel-per-particle sized textures (large): tPositions, tOffsets

Texel-per-cloud sized textures (small): tGridPositionsAndSeeds, tSelectionFactors

As I said before, the problem is that when these two buffer sizes (the large and the small) are at a 1:1 (width: height) ratio, the programs work just fine; however, when one or both are at a 2:1 (width:height) ratio the behavior is a mess. What accounts for this, and how can I address it? Thanks in advance!

UPDATE: Could the problem be related to my housing the texel coords to read the tPosition texture in the shader's position attribute in the second shader program? If so, perhaps this Github issue regarding texel coords in the position attribute may be related, though I can't find a corresponding question/answer here on SO.

UPDATE 2: I'm also looking into whether this could be an unpack alignment issue. Thoughts?

Here's the set up in Three.js for the first shader program:

function initComputeRenderer() {

    textureData = MotifGrid.getBufferData();

    gpuCompute = new GPUComputationRenderer( textureData.uPerParticleBufferWidth, textureData.uPerParticleBufferHeight, renderer );

    dtPositions = gpuCompute.createTexture();
    dtPositions.image.data = textureData.tPositions;

    offsetsTexture = new THREE.DataTexture( textureData.tOffsets, textureData.uPerParticleBufferWidth, textureData.uPerParticleBufferHeight, THREE.RGBAFormat, THREE.FloatType );
    offsetsTexture.needsUpdate = true;

    gridPositionsAndSeedsTexture = new THREE.DataTexture( textureData.tGridPositionsAndSeeds, textureData.uPerMotifBufferWidth, textureData.uPerMotifBufferHeight, THREE.RGBAFormat, THREE.FloatType );
    gridPositionsAndSeedsTexture.needsUpdate = true;

    selectionFactorsTexture = new THREE.DataTexture( textureData.tSelectionFactors, textureData.uPerMotifBufferWidth, textureData.uPerMotifBufferHeight, THREE.RGBAFormat, THREE.FloatType );
    selectionFactorsTexture.needsUpdate = true;

    positionVariable = gpuCompute.addVariable( "tPositions", document.getElementById( 'position_fragment_shader' ).textContent, dtPositions );
    positionVariable.wrapS = THREE.RepeatWrapping; // repeat wrapping for use only with bit powers: 8x8, 16x16, etc.
    positionVariable.wrapT = THREE.RepeatWrapping;

    gpuCompute.setVariableDependencies( positionVariable, [ positionVariable ] );

    positionUniforms = positionVariable.material.uniforms;
    positionUniforms.tOffsets = { type: "t", value: offsetsTexture };
    positionUniforms.tGridPositionsAndSeeds = { type: "t", value: gridPositionsAndSeedsTexture };
    positionUniforms.tSelectionFactors = { type: "t", value: selectionFactorsTexture };
    positionUniforms.uPerMotifBufferWidth = { type : "f", value : textureData.uPerMotifBufferWidth };
    positionUniforms.uPerMotifBufferHeight = { type : "f", value : textureData.uPerMotifBufferHeight };
    positionUniforms.uTime = { type: "f", value: 0.0 };
    positionUniforms.uXOffW = { type: "f", value: 0.5 };

}

Here is the first shader program (only a frag for physics calculations):

   // tPositions is handled by the GPUCompute script
    uniform sampler2D tOffsets; 
    uniform sampler2D tGridPositionsAndSeeds;
    uniform sampler2D tSelectionFactors;
    uniform float uPerMotifBufferWidth;
    uniform float uPerMotifBufferHeight;
    uniform float uTime;
    uniform float uXOffW;

    [...skipping a noise function for brevity...]

    void main() {

        vec2 uv = gl_FragCoord.xy / resolution.xy;

        vec4 offsets = texture2D( tOffsets, uv ).xyzw;
        float alphaMass = offsets.z;
        float cellIndex = offsets.w;

        if (cellIndex >= 0.0) {

            float damping = 0.98;

            float texelSizeX = 1.0 / uPerMotifBufferWidth;
            float texelSizeY = 1.0 / uPerMotifBufferHeight;
            vec2 perMotifUV = vec2( mod(cellIndex, uPerMotifBufferWidth)*texelSizeX, floor(cellIndex / uPerMotifBufferHeight)*texelSizeY );
            perMotifUV += vec2(0.5*texelSizeX, 0.5*texelSizeY);

            vec4 selectionFactors = texture2D( tSelectionFactors, perMotifUV ).xyzw;
            float swapState = selectionFactors.x;
            vec4 gridPosition = texture2D( tGridPositionsAndSeeds, perMotifUV ).xyzw;
            vec2 noiseSeed = gridPosition.zw;
            vec4 nowPos;
            vec2 velocity;

            nowPos = texture2D( tPositions, uv ).xyzw;
            velocity = vec2(nowPos.z, nowPos.w);

            if ( swapState == 0.0 ) {
                nowPos = texture2D( tPositions, uv ).xyzw;
                velocity = vec2(nowPos.z, nowPos.w);
            } else { // if swapState == 1
                //nowPos = vec4( -(uTime) + gridPosition.x + offsets.x, gridPosition.y + offsets.y, 0.0, 0.0 );
                nowPos = vec4( -(uTime) + offsets.x, offsets.y, 0.0, 0.0 );
                velocity = vec2(0.0, 0.0);
            }

            [...skipping the physics for brevity...]

            vec2 newPosition = vec2(nowPos.x - velocity.x, nowPos.y - velocity.y);
            // Write new position out
            gl_FragColor = vec4(newPosition.x, newPosition.y, velocity.x, velocity.y);
   }

Here is the setup for the second shader program: Note: The renderer for this section is a WebGLRenderer at window size

function makePerParticleReferencePositions() {

    var positions = new Float32Array( perParticleBufferSize * 3 );

    var texelSizeX = 1 / perParticleBufferDimensions.width;
    var texelSizeY = 1 / perParticleBufferDimensions.height;

    for ( var j = 0, j3 = 0; j < perParticleBufferSize; j ++, j3 += 3 ) {

        positions[ j3 + 0 ] = ( ( j % perParticleBufferDimensions.width ) / perParticleBufferDimensions.width ) + ( 0.5 * texelSizeX );
        positions[ j3 + 1 ] = ( Math.floor( j / perParticleBufferDimensions.height ) / perParticleBufferDimensions.height ) + ( 0.5 * texelSizeY );
        positions[ j3 + 2 ] = j * 0.0001; // this is the real z value for the particle display

    }

    return positions;
}

var positions = makePerParticleReferencePositions();

...

// Add attributes to the BufferGeometry: 
gridOfMotifs.geometry.addAttribute( 'position', new THREE.BufferAttribute( positions, 3 ) );
gridOfMotifs.geometry.addAttribute( 'aTextureIndex', new THREE.BufferAttribute( motifGridAttributes.aTextureIndex, 1 ) );
gridOfMotifs.geometry.addAttribute( 'aAlpha', new THREE.BufferAttribute( motifGridAttributes.aAlpha, 1 ) );
gridOfMotifs.geometry.addAttribute( 'aCellIndex', new THREE.BufferAttribute(
        motifGridAttributes.aCellIndex, 1 ) );

uniformValues = {};
uniformValues.tSelectionFactors = motifGridAttributes.tSelectionFactors;
uniformValues.uPerMotifBufferWidth = motifGridAttributes.uPerMotifBufferWidth;
uniformValues.uPerMotifBufferHeight = motifGridAttributes.uPerMotifBufferHeight;

gridOfMotifs.geometry.computeBoundingSphere();

...

function makeCustomUniforms( uniformValues ) {

    selectionFactorsTexture = new THREE.DataTexture( uniformValues.tSelectionFactors, uniformValues.uPerMotifBufferWidth, uniformValues.uPerMotifBufferHeight, THREE.RGBAFormat, THREE.FloatType );
    selectionFactorsTexture.needsUpdate = true;

    var customUniforms = {
        tPositions : { type : "t", value : null },
        tSelectionFactors : { type : "t", value : selectionFactorsTexture },
        uPerMotifBufferWidth : { type : "f", value : uniformValues.uPerMotifBufferWidth },
        uPerMotifBufferHeight : { type : "f", value : uniformValues.uPerMotifBufferHeight },
        uTextureSheet : { type : "t", value : texture }, // this is a sprite sheet of all 10 strokes
        uPointSize : { type : "f", value : 18.0 }, // the radius of a point in WebGL units, e.g. 30.0
        // Coords for the hatch textures:
        uTextureCoordSizeX : { type : "f", value : 1.0 / numTexturesInSheet },
        uTextureCoordSizeY : { type : "f", value : 1.0 }, // the size of a texture in the texture map ( they're square, thus only one value )
    };
    return customUniforms;
}

And here is the corresponding shader program (vert & frag):

Vertex shader:

    uniform sampler2D tPositions;
    uniform sampler2D tSelectionFactors;
    uniform float uPerMotifBufferWidth;
    uniform float uPerMotifBufferHeight;
    uniform sampler2D uTextureSheet;
    uniform float uPointSize; // the radius size of the point in WebGL units, e.g. "30.0"
    uniform float uTextureCoordSizeX; // vertical dimension of each texture given the full side = 1
    uniform float uTextureCoordSizeY; // horizontal dimension of each texture given the full side = 1

    attribute float aTextureIndex;
    attribute float aAlpha;
    attribute float aCellIndex;

    varying float vCellIndex;
    varying vec2 vTextureCoords;
    varying vec2 vTextureSize;
    varying float vAlpha;
    varying vec3 vColor;
    varying float vDensity;

   [...skipping noise function for brevity...]

    void main() {

        vec4 tmpPos = texture2D( tPositions, position.xy );
        vec2 pos = tmpPos.xy;
        vec2 vel = tmpPos.zw;

        vCellIndex = aCellIndex;

        if (aCellIndex >= 0.0) { // buffer filler cell indexes are -1

            float texelSizeX = 1.0 / uPerMotifBufferWidth;
            float texelSizeY = 1.0 / uPerMotifBufferHeight;
            vec2 perMotifUV = vec2( mod(aCellIndex, uPerMotifBufferWidth)*texelSizeX, floor(aCellIndex / uPerMotifBufferHeight)*texelSizeY );
            perMotifUV += vec2(0.5*texelSizeX, 0.5*texelSizeY);

            vec4 selectionFactors = texture2D( tSelectionFactors, perMotifUV ).xyzw;
            float aSelectedMotif = selectionFactors.x;
            float aColor = selectionFactors.y;
            float fadeFactor = selectionFactors.z;

            vTextureCoords = vec2( aTextureIndex * uTextureCoordSizeX, 0 );
            vTextureSize = vec2( uTextureCoordSizeX, uTextureCoordSizeY );

            vAlpha = aAlpha * fadeFactor;
            vDensity = vel.x + vel.y;
            vAlpha *= abs( vDensity * 3.0 );

            vColor = vec3( 1.0, aColor, 1.0 ); // set RGB color associated to vertex; use later in fragment shader.

            gl_PointSize = uPointSize;

        } else { // if this is a filler cell index (-1)
            vAlpha = 0.0;
            vDensity = 0.0;
            vColor = vec3(0.0, 0.0, 0.0);
            gl_PointSize = 0.0;
        }
        gl_Position = projectionMatrix * modelViewMatrix * vec4( pos.x, pos.y, position.z, 1.0 ); // position holds the real z value. The z value of "color" is a component of velocity
    }

Fragment shader:

    uniform sampler2D tPositions;
    uniform sampler2D uTextureSheet;

    varying float vCellIndex;
    varying vec2 vTextureCoords;
    varying vec2 vTextureSize;
    varying float vAlpha;
    varying vec3 vColor;
    varying float vDensity;  

    void main() {
        gl_FragColor = vec4( vColor, vAlpha );

        if (vCellIndex >= 0.0) { // only render out the texture if this point is not a buffer filler
            vec2 realTexCoord = vTextureCoords + ( gl_PointCoord * vTextureSize );
            gl_FragColor = gl_FragColor * texture2D( uTextureSheet, realTexCoord );
        }
    }

Expected Behavior: I can achieve this by forcing all the DataTextures to be 1:1 expected behavior - shaped particle clouds correctly assembled and displayed in an offset grid

Weird Behavior: When the smaller DataTextures are 2:1 those perfectly horizontal clouds in the top right of the picture below form and have messed up physics. When the larger DataTextures are 2:1, the grid is skewed, and the clouds appear to be missing parts (as seen below). When both the small and large textures are 2:1, both odd behaviors happen (this is the case in the image below). weird behavior - particle clouds incorrectly assembled and displayed in a skewed grid

Enugu answered 14/3, 2017 at 15:0 Comment(0)
E
0

Thanks to an answer to my related question here, I now know what was going wrong. The problem was in the way I was using the arrays of indexes (1,2,3,4,5...) to access the DataTextures' texels in the shader.

In this function (and the one for the larger DataTextures)...

float texelSizeX = 1.0 / uPerMotifBufferWidth;
float texelSizeY = 1.0 / uPerMotifBufferHeight;
vec2 perMotifUV = vec2( 
    mod(aCellIndex, uPerMotifBufferWidth)*texelSizeX, 
    floor(aCellIndex / uPerMotifBufferHeight)*texelSizeY );
perMotifUV += vec2(0.5*texelSizeX, 0.5*texelSizeY);

...I assumed that in order to create the y value for my custom uv, perMotifUV, I would need to divide the aCellIndex by the height of the buffer, uPerMotifBufferHeight (it's "vertical" dimension). However, as explained in the SO Q&A here the indices should, of course, be divided by the buffer's width, which would then tell you how many rows down you are!

Thus, the function should be revised to...

float texelSizeX = 1.0 / uPerMotifBufferWidth;
float texelSizeY = 1.0 / uPerMotifBufferHeight;
vec2 perMotifUV = vec2( 
    mod(aCellIndex, uPerMotifBufferWidth)*texelSizeX, 
    floor(aCellIndex / uPerMotifBufferWidth)*texelSizeY ); **Note the change to uPerMotifBufferWidth here
perMotifUV += vec2(0.5*texelSizeX, 0.5*texelSizeY);

The reason my program worked on square DataTextures (1:1) is that in such cases the height and width were equal, so my function was effectively dividing by width in the incorrect line because height=width!

Enugu answered 5/5, 2017 at 14:57 Comment(0)

© 2022 - 2024 — McMap. All rights reserved.