How to increase performance using WebGL?

Question

I am rendering a texture using WebGL ,However ,the way I am rendering is I am rendering few lines of data and then moving those lines to right and again drawing another set of lines.

For example : I have a image of 640*480 ,which contains 640*480*4 pixels of RGBA, however I am only filling the alpha values and it is a GrayScale medical Dicom image.

Now ,the issue that I am facing is it is rendering the texture with jerks ,the image rendering is not happening smoothly.

For example, This is what happens : There are 640 lines of data to be rendered.

So, I took a arraybuffer of 640*480*4 and then , Suppose first line came to client via websocket from server to render ,then I will fill the indexes as 3, 640+3, 640*2+3, 640*3+3 and so on until 640*480+3. Then when the second line is received ,I will move first line to second line like 3->7, 640+3->640+7, ......640*480+3->640*480+7. And then the newly received line will be rendered to 3, 640+3, 640*2+3, 640*3+3 and this will continue until the 640th line of image data.

Here's the code that I have done.

Code:

    var renderLineData = function (imageAttr) {
    var data = imageAttr.data;
    var LINES_PER_CHUNK = imageAttr.lines;
    var alpha = 4;
    if(imageAttr.newImage) {
        newBuffer = new ArrayBuffer(imageAttr.width * imageAttr.height * alpha);dataTypedArray = new Uint8Array(newBuffer);
        // provide texture coordinates for the rectangle.
        provideTextureCoordsForRect();
        setParams();
        // Upload the image into the texture.
        // look up uniform locations
        uploadImageToTexture(gl.getUniformLocation(program, 'u_matrix'));
    } else {
        for (var z = imageAttr.index; z > 0; z--) {
            for (i = 0 ; i < LINES_PER_CHUNK; i++) {
                for (j = 0 ; j < imageAttr.height; j++) {
                    dataTypedArray[i * alpha + imageAttr.width*alpha * j + 3 + LINES_PER_CHUNK  * alpha * z] = dataTypedArray[i * alpha + imageAttr.width*alpha * j +  3 + LINES_PER_CHUNK  * alpha * (z-1)];
                }
            }
        }
    }
    for (i = 0, k = imageAttr.height*LINES_PER_CHUNK; i < LINES_PER_CHUNK; i++) {
        for (j = 0 ; j < imageAttr.height; j++) {
            dataTypedArray[i * alpha + imageAttr.width*4 * j + 3] = data[k - imageAttr.height + j];
        }
        k = k - imageAttr.height;
    }
    imageAttrTemp = imageAttr;
    renderImgSlowly(gl, imageAttr, dataTypedArray);
};
function renderImgSlowly (gl, image, dataTypedArray)  {
    gl.clear(gl.COLOR_BUFFER_BIT || gl.DEPTH_BUFFER_BIT);
    gl.texImage2D(gl.TEXTURE_2D, 0, gl.RGBA, image.width, image.height, 0, gl.RGBA, gl.UNSIGNED_BYTE, dataTypedArray);
    //Draw the rectangle.
    gl.drawArrays(gl.TRIANGLES, 0, 6);
}

Answer 1

gl.texImage2D is slow and there is not much that can be done to improve on this. The reason is that texImage2D involves a state change and requires that the GPU halt all rendering and then fetch the data from CPU RAM. Depending on the hardware the interface between the main board and GPU can be very slow (in comparison to RAM access speed)

You also add to the problem with the resolution of the image. All images on the GPU have sizes that are powers of 2 (32,64,128,256,512,1024...) independently for height and width. Sending an image that is 640 by 480 does not fit this rule. To accommodate the bad size the GPU will allocate an image that is W 1024 by H 512 pixels and will thus have to re dimension the image data to fit the internal dimensions (Fast as it they are this is not something they are good at). Depending on hardware this will cause an additional slowdown on top of the already slow data transfer.

You may get a slight improvement if you make your data buffer equal to the powers of two rule (POT) (1024, 512).

Your best option is to avoid the transfer until the entire image has loaded then do it just once.

If you really need it live then I suggest you divide the image into smaller images and send the smaller separate images. For example for the POT image size 1024 by 512 could be divided into 128 by 64 images or 1024 by 8 resulting in 64 smaller images. Only send the small images as they become available and on the GPU reassemble the images as one during render. This will give you an almost 64 times improvement on the time it takes to send the image to the GPU.

Apart from that there is not much else that can be done. GPUs are good at rendering, GPUs suck when it comes to mainboard IO, avoid this at all costs (during rendering) to get the most out of the graphics hardware.

Answer 2

First off, nothing you are doing is likely to be the speed issue. A 640x320 image is not that large and the amount of processing you're doing in JavaScript is unlikely to be the bottleneck.

On top of that WebGL will have no trouble drawing a single quad which is all you're drawing. Nor will it have a problem uploading a 640x480 texture.

The bottleneck is the network. Sending chunks over the network is slow.

On the other hand, if you want to optimize, why are you shifting the data around in JavaScript? Just put it in the correct place in the texture to start with with gl.texSubImage2D. If you only want to draw the part that has had data put in it then adjust the texture coordinates to select that part of the texture

Also, why are you using RGBA if you only need one channel? Use LUMINANCE.

  if (imageAttr.newImage) {
      destColumn = imageAttr.width;
      gl.texImage2D(gl.TEXTURE_2D, 0, gl.LUMINANCE, imageAttr.width, imageAttr.height, 0,
                    gl.LUMINANCE, gl.UNSIGNED_BYTE, null);
  }
  destColumn -= imageAttr.lines;
  // should check it destColumn does not go negative!
  gl.texSubImage2D(gl.TEXTURE_2D, 0, destColumn, 0, imageAttr.lines, imageAttr.height,
                   gl.LUMINANCE, gl.UNSIGNED_BYTE, imageAttr.data);

  var srcX = destColumn;
  var srcY = 0;
  var srcWidth  = imageAttr.width - destColumn;
  var srcHeight = imageAttr.height;

  var dstX = destColumn * gl.canvas.width / imageAttr.width;
  var dstY = 0;
  var dstWidth  = srcWidth * gl.canvas.width / imageAttr.width;
  var dstHeight = srcHeight;

  var texWidth     = imageAttr.width;
  var texHeight    = imageAttr.height;
  var targetWidth  = gl.canvas.width;
  var targetHeight = gl.canvas.height;

  drawImageInWebGL(
    tex, texWidth, texHeight,  
    srcX, srcY, srcWidth, srcHeight,
    dstX, dstY, dstWidth, dstHeight,
    targetWidth, targetHeight);
}

Here's an example

var m4 = twgl.m4;
var gl = document.getElementById("c").getContext("webgl");
// compiles shader, links and looks up locations
var programInfo = twgl.createProgramInfo(gl, ["vs", "fs"]);

// a unit quad
var arrays = {
  position: { 
    numComponents: 2, 
    data: [
      0, 0,  
      1, 0, 
      0, 1, 
      0, 1, 
      1, 0,  
      1, 1,
    ],
  },
};
// calls gl.createBuffer, gl.bindBuffer, gl.bufferData for each array
var bufferInfo = twgl.createBufferInfoFromArrays(gl, arrays);
      
// we're only using 1 texture so just make and bind it now
var tex = gl.createTexture();
gl.bindTexture(gl.TEXTURE_2D, tex);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_S, gl.CLAMP_TO_EDGE);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_T, gl.CLAMP_TO_EDGE);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MIN_FILTER, gl.LINEAR);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MAG_FILTER, gl.NEAREST);
      
      
var destColumn = 0;
      
// We're using 1 byte wide texture pieces so we need to 
// set UNPACK_ALIGNMENT to 1 as it defaults to 4
gl.pixelStorei(gl.UNPACK_ALIGNMENT, 1);
            
simulateSendingAnImageNColumnsAtATime(1, 1, addLinesToImageAndDraw);
  
function addLinesToImageAndDraw(imageAttr) {
  if (imageAttr.newImage) {
      destColumn = imageAttr.width;
      gl.texImage2D(gl.TEXTURE_2D, 0, gl.LUMINANCE, imageAttr.width, imageAttr.height, 0,
                    gl.LUMINANCE, gl.UNSIGNED_BYTE, null);
  }
  destColumn -= imageAttr.lines;
  // should check it destColumn does not go negative!
  gl.texSubImage2D(gl.TEXTURE_2D, 0, destColumn, 0, imageAttr.lines, imageAttr.height,
                   gl.LUMINANCE, gl.UNSIGNED_BYTE, imageAttr.data);
  
  var srcX = destColumn;
  var srcY = 0;
  var srcWidth  = imageAttr.width - destColumn;
  var srcHeight = imageAttr.height;

  var dstX = destColumn * gl.canvas.width / imageAttr.width;
  var dstY = 0;
  var dstWidth  = srcWidth * gl.canvas.width / imageAttr.width;
  var dstHeight = gl.canvas.height;

  var texWidth     = imageAttr.width;
  var texHeight    = imageAttr.height;
  var targetWidth  = gl.canvas.width;
  var targetHeight = gl.canvas.height;
      
  drawImage(
    tex, texWidth, texHeight,  
    srcX, srcY, srcWidth, srcHeight,
    dstX, dstY, dstWidth, dstHeight,
    targetWidth, targetHeight);
}
      

// we pass in texWidth and texHeight because unlike images
// we can't look up the width and height of a texture

// we pass in targetWidth and targetHeight to tell it
// the size of the thing we're drawing too. We could look 
// up the size of the canvas with gl.canvas.width and
// gl.canvas.height but maybe we want to draw to a framebuffer
// etc.. so might as well pass those in.

// srcX, srcY, srcWidth, srcHeight are in pixels 
// computed from texWidth and texHeight

// dstX, dstY, dstWidth, dstHeight are in pixels
// computed from targetWidth and targetHeight
function drawImage(
    tex, texWidth, texHeight,
    srcX, srcY, srcWidth, srcHeight,
    dstX, dstY, dstWidth, dstHeight,
    targetWidth, targetHeight) {
  var mat  = m4.identity();
  var tmat = m4.identity();
  
  var uniforms = {
    matrix: mat,
    textureMatrix: tmat,
    texture: tex,
  };

  // these adjust the unit quad to generate texture coordinates
  // to select part of the src texture

  // NOTE: no check is done that srcX + srcWidth go outside of the
  // texture or are in range in any way. Same for srcY + srcHeight

  m4.translate(tmat, [srcX / texWidth, srcY / texHeight, 0], tmat);
  m4.scale(tmat, [srcWidth / texWidth, srcHeight / texHeight, 1], tmat);

  // these convert from pixels to clip space
  m4.ortho(0, targetWidth, targetHeight, 0, -1, 1, mat)

  // these move and scale the unit quad into the size we want
  // in the target as pixels
  m4.translate(mat, [dstX, dstY, 0], mat);
  m4.scale(mat, [dstWidth, dstHeight, 1], mat);

  gl.useProgram(programInfo.program);
  // calls gl.bindBuffer, gl.enableVertexAttribArray, gl.vertexAttribPointer
  twgl.setBuffersAndAttributes(gl, programInfo, bufferInfo);
  // calls gl.uniformXXX, gl.activeTexture, gl.bindTexture
  twgl.setUniforms(programInfo, uniforms);
  // calls gl.drawArray or gl.drawElements
  twgl.drawBufferInfo(gl, gl.TRIANGLES, bufferInfo);
  
}


// =====================================================================
// Everything below this line represents stuff from the server.
// so it's irrelevant to the answer
//

function simulateSendingAnImageNColumnsAtATime(minColumnsPerChunk, maxColumnsPerChunk, callback) {
  var imageData = createImageToSend(640, 480);
  
  // cut data into columns at start because this work would be done on
  // the server
  var columns = [];
  var x = 0;
  while (x < imageData.width) {
    // how many columns are left?
    var maxWidth = imageData.width - x;
    
    // how many columns should we send
    var columnWidth = Math.min(maxWidth, rand(minColumnsPerChunk, maxColumnsPerChunk + 1));
    
    var data = createImageChunk(imageData, imageData.width - x - columnWidth, 0, columnWidth, imageData.height);    
    
    columns.push({
      newImage: x === 0,
      lines: columnWidth,
      width: imageData.width,
      height: imageData.height,
      data: data,
    });
    
    x += columnWidth;
  }
  
  var columnNdx = 0;
  sendNextColumn();
  
  function sendNextColumn() {
    if (columnNdx < columns.length) {
      callback(columns[columnNdx++]);
      if (columnNdx < columns.length) {
        // should we make this random to siumlate network speed
        var timeToNextChunkMS = 17;
        setTimeout(sendNextColumn, timeToNextChunkMS);
      }
    }
  }
}

function createImageChunk(imageData, x, y, width, height) {
  var data = new Uint8Array(width * height);
  for (var yy = 0; yy < height; ++yy) {
    for (var xx = 0; xx < width; ++xx) {
      var srcOffset = ((yy + y) * imageData.width + xx + x) * 4;
      var dstOffset = yy * width + xx;
      // compute gray scale
      var gray = Math.max(imageData.data[srcOffset], imageData.data[srcOffset + 1], imageData.data[srcOffset + 2]);
      data[dstOffset] = gray;
    }
  }
  return data;
}

function rand(min, max) {
  return Math.floor(Math.random() * max - min) + min;
}

function createImageToSend(width, height) {
  // create a texture using a canvas so we don't have to download one
  var ctx = document.createElement("canvas").getContext("2d");
  ctx.width = width;
  ctx.height = height;
  ctx.fillStyle = "#222";
  ctx.fillRect(0, 0, ctx.canvas.width, ctx.canvas.height);
  ctx.lineWidth = 20;
  ["#AAA", "#888", "#666"].forEach(function(color, ndx, array) {
    ctx.strokeStyle = color;
    ctx.beginPath();
    ctx.arc((ndx + 1) / (array.length + 1) * ctx.canvas.width, ctx.canvas.height / 2,
            ctx.canvas.height * 0.4, 0, Math.PI * 2, false);
    ctx.stroke();
  });
  ctx.fillStyle = "white";
  ctx.font = "40px sans-serif";
  ctx.textAlign = "center";
  ctx.textBaseline = "middle";
  ctx.fillText("Some Image", ctx.canvas.width / 2, ctx.canvas.height / 2);
  return ctx.getImageData(0, 0, ctx.canvas.width, ctx.canvas.height);
}

canvas { border: 1px solid black; }

<script src="https://twgljs.org/dist/twgl-full.min.js"></script>
<script id="vs" type="not-js">
// we will always pass a 0 to 1 unit quad
// and then use matrices to manipulate it
attribute vec4 position;   

uniform mat4 matrix;
uniform mat4 textureMatrix;

varying vec2 texcoord;

void main () {
  gl_Position = matrix * position;
  
  texcoord = (textureMatrix * position).xy;
}
</script>
<script id="fs" type="not-js">
precision mediump float;

varying vec2 texcoord;
uniform sampler2D texture;

void main() {
  gl_FragColor = texture2D(texture, texcoord);
}
</script>
<canvas id="c" width="640" height="480"></canvas>

NOTE: This will not be smooth because it is using setTimeout to simulate receiving network data but that's exactly what you're likely seeing.

Here's a sample that rotates the image independently of updating the texture. You can see it runs perfectly smooth. The slowness is not WebGL, the slowness is networking (as simulated by setTimeout)

var m4 = twgl.m4;
var gl = document.getElementById("c").getContext("webgl");
// compiles shader, links and looks up locations
var programInfo = twgl.createProgramInfo(gl, ["vs", "fs"]);

// a unit quad
var arrays = {
  position: { 
    numComponents: 2, 
    data: [
      0, 0,  
      1, 0, 
      0, 1, 
      0, 1, 
      1, 0,  
      1, 1,
    ],
  },
};
// calls gl.createBuffer, gl.bindBuffer, gl.bufferData for each array
var bufferInfo = twgl.createBufferInfoFromArrays(gl, arrays);
      
// we're only using 1 texture so just make and bind it now
var tex = gl.createTexture();
gl.bindTexture(gl.TEXTURE_2D, tex);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_S, gl.CLAMP_TO_EDGE);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_WRAP_T, gl.CLAMP_TO_EDGE);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MIN_FILTER, gl.LINEAR);
gl.texParameteri(gl.TEXTURE_2D, gl.TEXTURE_MAG_FILTER, gl.NEAREST);
      
      
var destColumn = 0;
var imageWidth;
var imageHeight;
      
// We're using 1 byte wide texture pieces so we need to 
// set UNPACK_ALIGNMENT to 1 as it defaults to 4
gl.pixelStorei(gl.UNPACK_ALIGNMENT, 1);
            
simulateSendingAnImageNColumnsAtATime(1, 1, addLinesToImageAndDraw);
  
function addLinesToImageAndDraw(imageAttr) {
  if (imageAttr.newImage) {
      destColumn  = imageAttr.width;
      imageWidth  = imageAttr.width;
      imageHeight = imageAttr.height;
      gl.texImage2D(gl.TEXTURE_2D, 0, gl.LUMINANCE, imageAttr.width, imageAttr.height, 0,
                    gl.LUMINANCE, gl.UNSIGNED_BYTE, null);
  }
  destColumn -= imageAttr.lines;
  // should check it destColumn does not go negative!
  gl.texSubImage2D(gl.TEXTURE_2D, 0, destColumn, 0, imageAttr.lines, imageAttr.height,
                   gl.LUMINANCE, gl.UNSIGNED_BYTE, imageAttr.data);
}
      
function render(time) {
  if (imageWidth) {
    var srcX = destColumn;
    var srcY = 0;
    var srcWidth  = imageWidth - destColumn;
    var srcHeight = imageHeight;

    var dstX = destColumn * gl.canvas.width / imageWidth;
    var dstY = 0;
    var dstWidth  = srcWidth * gl.canvas.width / imageWidth;
    var dstHeight = gl.canvas.height;

    var texWidth     = imageWidth;
    var texHeight    = imageHeight;
    var targetWidth  = gl.canvas.width;
    var targetHeight = gl.canvas.height;

    drawImageWithRotation(
      time * 0.001,
      tex, texWidth, texHeight,  
      srcX, srcY, srcWidth, srcHeight,
      dstX, dstY, dstWidth, dstHeight,
      targetWidth, targetHeight);
  }
  requestAnimationFrame(render);
}
requestAnimationFrame(render);
      

// we pass in texWidth and texHeight because unlike images
// we can't look up the width and height of a texture

// we pass in targetWidth and targetHeight to tell it
// the size of the thing we're drawing too. We could look 
// up the size of the canvas with gl.canvas.width and
// gl.canvas.height but maybe we want to draw to a framebuffer
// etc.. so might as well pass those in.

// srcX, srcY, srcWidth, srcHeight are in pixels 
// computed from texWidth and texHeight

// dstX, dstY, dstWidth, dstHeight are in pixels
// computed from targetWidth and targetHeight
function drawImageWithRotation(
    rotation,
    tex, texWidth, texHeight,
    srcX, srcY, srcWidth, srcHeight,
    dstX, dstY, dstWidth, dstHeight,
    targetWidth, targetHeight) {
  var mat  = m4.identity();
  var tmat = m4.identity();
  
  var uniforms = {
    matrix: mat,
    textureMatrix: tmat,
    texture: tex,
  };

  // these adjust the unit quad to generate texture coordinates
  // to select part of the src texture

  // NOTE: no check is done that srcX + srcWidth go outside of the
  // texture or are in range in any way. Same for srcY + srcHeight
  m4.translate(tmat, [srcX / texWidth, srcY / texHeight, 0], tmat);
  m4.scale(tmat, [srcWidth / texWidth, srcHeight / texHeight, 1], tmat);
    
  // convert from pixels to clipspace
  m4.ortho(0, targetWidth, targetHeight, 0, -1, 1, mat);
      
  // rotate around center of canvas
  m4.translate(mat, [targetWidth / 2, targetHeight / 2, 0], mat);
  m4.rotateZ(mat, rotation, mat);
  m4.translate(mat, [-targetWidth / 2, -targetHeight / 2, 0], mat);
      
  // these move and scale the unit quad into the size we want
  // in the target as pixels
  m4.translate(mat, [dstX, dstY, 0], mat);
  m4.scale(mat, [dstWidth, dstHeight, 1], mat);

  gl.useProgram(programInfo.program);
  // calls gl.bindBuffer, gl.enableVertexAttribArray, gl.vertexAttribPointer
  twgl.setBuffersAndAttributes(gl, programInfo, bufferInfo);
  // calls gl.uniformXXX, gl.activeTexture, gl.bindTexture
  twgl.setUniforms(programInfo, uniforms);
  // calls gl.drawArray or gl.drawElements
  twgl.drawBufferInfo(gl, bufferInfo);
  
}


// =====================================================================
// Everything below this line represents stuff from the server.
// so it's irrelevant to the answer
//

function simulateSendingAnImageNColumnsAtATime(minColumnsPerChunk, maxColumnsPerChunk, callback) {
  var imageData = createImageToSend(640, 480);
  
  // cut data into columns at start because this work would be done on
  // the server
  var columns = [];
  var x = 0;
  while (x < imageData.width) {
    // how many columns are left?
    var maxWidth = imageData.width - x;
    
    // how many columns should we send
    var columnWidth = Math.min(maxWidth, rand(minColumnsPerChunk, maxColumnsPerChunk + 1));
    
    var data = createImageChunk(imageData, imageData.width - x - columnWidth, 0, columnWidth, imageData.height);    
    
    columns.push({
      newImage: x === 0,
      lines: columnWidth,
      width: imageData.width,
      height: imageData.height,
      data: data,
    });
    
    x += columnWidth;
  }
  
  var columnNdx = 0;
  sendNextColumn();
  
  function sendNextColumn() {
    if (columnNdx < columns.length) {
      callback(columns[columnNdx++]);
      if (columnNdx < columns.length) {
        // should we make this random to siumlate network speed
        var timeToNextChunkMS = 17;
        setTimeout(sendNextColumn, timeToNextChunkMS);
      }
    }
  }
}

function createImageChunk(imageData, x, y, width, height) {
  var data = new Uint8Array(width * height);
  for (var yy = 0; yy < height; ++yy) {
    for (var xx = 0; xx < width; ++xx) {
      var srcOffset = ((yy + y) * imageData.width + xx + x) * 4;
      var dstOffset = yy * width + xx;
      // compute gray scale
      var gray = Math.max(imageData.data[srcOffset], imageData.data[srcOffset + 1], imageData.data[srcOffset + 2]);
      data[dstOffset] = gray;
    }
  }
  return data;
}

function rand(min, max) {
  return Math.floor(Math.random() * max - min) + min;
}

function createImageToSend(width, height) {
  // create a texture using a canvas so we don't have to download one
  var ctx = document.createElement("canvas").getContext("2d");
  ctx.width = width;
  ctx.height = height;
  ctx.fillStyle = "#222";
  ctx.fillRect(0, 0, ctx.canvas.width, ctx.canvas.height);
  ctx.lineWidth = 20;
  ["#AAA", "#888", "#666"].forEach(function(color, ndx, array) {
    ctx.strokeStyle = color;
    ctx.beginPath();
    ctx.arc((ndx + 1) / (array.length + 1) * ctx.canvas.width, ctx.canvas.height / 2,
            ctx.canvas.height * 0.4, 0, Math.PI * 2, false);
    ctx.stroke();
  });
  ctx.fillStyle = "white";
  ctx.font = "40px sans-serif";
  ctx.textAlign = "center";
  ctx.textBaseline = "middle";
  ctx.fillText("Some Image", ctx.canvas.width / 2, ctx.canvas.height / 2);
  return ctx.getImageData(0, 0, ctx.canvas.width, ctx.canvas.height);
}

canvas { border: 1px solid black; }

<script src="https://twgljs.org/dist/3.x/twgl-full.min.js"></script>
<script id="vs" type="not-js">
// we will always pass a 0 to 1 unit quad
// and then use matrices to manipulate it
attribute vec4 position;   

uniform mat4 matrix;
uniform mat4 textureMatrix;

varying vec2 texcoord;

void main () {
  gl_Position = matrix * position;
  
  texcoord = (textureMatrix * position).xy;
}
</script>
<script id="fs" type="not-js">
precision mediump float;

varying vec2 texcoord;
uniform sampler2D texture;

void main() {
  gl_FragColor = texture2D(texture, texcoord);
}
</script>
<canvas id="c" width="640" height="480"></canvas>