02Credits Blog

Day40 - WebGL Renderer

Description of the WebGL renderer I use for Tetris Attack

2019-03-18

Project Page

Todo

Today I spent a good long while trying to play around with some simple simplex noise functions in Rust. Unfortunately I couldn't get things where I wanted them so instead of describing my failures there I will describe the WebGL renderer I have been using for the Tetris Attack port.

Background

About a year ago a friend of mine and I started working on a silly game project that didn't pan out. In the process however I wrote a 2d game engine in Typescript using WebGL to render sprites very performantly. These days there are only a couple of options for 2d games in JavaScript. You can use the built in 2d html5 canvas functions, you can use a prebuilt library such as PIXI.js, or you can roll your own library. In my experience the 2d canvas APIs are WAY too slow and cumbersome to do anything interesting with. PIXI.js is plenty fast and likely faster than anything I would come up with, but it requires you to specify your graphics in a scene graph which I find limiting and confusing.

So the only remaining option was to roll my own. My engine is an over glorified quad renderer, but supports transparency, arbitrary scale/rotation and somewhere around 10000 sprites at a time. I render the textures to a sprite atlas at startup so that everything can be done in one draw call. The math is all done on the shader side to lighten the load on the browser.

Sprite Atlas

By building the sprite atlas at startup, I don't need to worry about importing a prebuilt atlas format and can play fast and loose with my graphics pipeline. The first step is to load each of the textures into image tags so that they can be drawn to a canvas.

  export async function loadTextures(texturePaths: string[]) {
  let images: { [id: string]: HTMLImageElement } = {};
  for (let path of texturePaths) {
    let image = new Image();
    let loadedPromise = new Promise(resolve => {
      let handler = () => {
        resolve();
        image.removeEventListener("load", handler);
      };
      image.addEventListener("load", handler, false);
      image.src = path;
    });
    await loadedPromise;
    images[path] = image;
  }
  return packTextures(images);
}

Laying out the textures is done in a very naive way by guessing a multiple of 2 size of the texture, laying out the images sorted by height, and trying at a larger size if all of the textures don't fit.

  export function packTextures(images: {
  [id: string]: HTMLImageElement;
}): TextureInfo {
  let imageArray: { image: HTMLImageElement; id: string }[] = [];
  for (let id in images) {
    imageArray.push({ image: images[id], id: id });
  }
  imageArray = imageArray.sort((a, b) => b.image.height - a.image.height);
  let size = 16;
  let correctSize = true;
  let imageLayoutInfo: { [id: string]: number[] };
  do {
    imageLayoutInfo = {};
    correctSize = true;
    let gap = 10;
    size *= 2;
    let x = gap;
    let y = gap;
    let rowHeight = imageArray[0].image.height;
    for (let imageData of imageArray) {
      let image = imageData.image;
      imageLayoutInfo[imageData.id] = [
        x,
        y,
        x + image.width,
        y,
        x,
        y + image.height,
        x + image.width,
        y + image.height
      ];
      x += image.width + gap;
      if (x > size) {
        x = gap;
        y += rowHeight + gap;
        if (y + image.height + gap > size) {
          correctSize = false;
          break;
        }
        rowHeight = image.height + gap;
      }
    }
  } while (!correctSize);

Once a large enough map is attempted, the images are then rendered to the canvas one by one and a WebGL texture is created with the contents.

  let canvas = document.createElement("canvas");
  canvas.width = size;
  canvas.height = size;
  // document.body.appendChild(canvas);
  let ctx = canvas.getContext("2d");
  for (var imageData of imageArray) {
    let info = imageLayoutInfo[imageData.id];
    ctx.drawImage(imageData.image, info[0], info[1]);
  }
  return { size: size, canvas: canvas, texCoords: imageLayoutInfo };
}

Shaders

GLSL shaders most frequently are split into a vertex shader which positions each triangle onto the screen, and a fragment shader which decides which color to assign to each pixel within a triangle. My engine is no different and uses this same structure. In contrast to most graphics engines however, I don't actually use the vertex position to place vertices into the "world" space and instead have various attributes for each vertex specifying where it should be drawn. This way instead of doing a bunch of vertex math on the JavaScript side to position every sprite, the position data is just passed to the shader and handled there. This has the added benefit of requiring marginally less data to be copied about since most of the data doesn't change between frames. Every sprite is passed to the shader as a single unit wide and single unit tall quad of two triangles with data about where to render.

The vertex shader does the majority of the work taking attribute variables from the WebGL library and doing all of the necessary math to transform the unit quad to the correct shape and location. The position of the sprite is transformed by the center attribute to specify where the center of rotation should be. Then the position is rotate using some simple trig by the rotation attribute. The position is then translated into place to match the passed position. Lastly the texture is transformed by the camera position to place it in its final location. Then the texture data and color are passed to the fragment shader as varying variables which just means they are interpolated between the vertices.

  attribute vec2 a_coord;
attribute vec3 a_position;
attribute vec2 a_texcoord;
attribute float a_rotation;
attribute vec2 a_dimensions;
attribute vec2 a_center;
attribute float a_scale;
attribute vec4 a_color;
varying highp vec2 v_texcoord;
varying highp vec4 v_color;
uniform vec4 u_camera_dimensions;
void main() {
  vec2 relativePosition = (a_coord * a_dimensions - a_dimensions * a_center) * a_scale;
  vec2 rotatedPosition = vec2(cos(a_rotation) * relativePosition.x - sin(a_rotation) * relativePosition.y,
                              sin(a_rotation) * relativePosition.x + cos(a_rotation) * relativePosition.y);
  vec3 worldCoords = vec3(rotatedPosition + a_position.xy, 0);
  gl_Position = vec4((worldCoords.xy - u_camera_dimensions.xy - u_camera_dimensions.zw / 2.0) / (u_camera_dimensions.zw / 2.0), worldCoords.z, 1);
  v_texcoord = a_texcoord;
  v_color = a_color;
}

In contrast, the fragment shader is incredibly simple just sampling the texture in the correct location to match the requested sprite and multiplying by the tint to get the final pixel color.

  precision highp float;
uniform float u_map_dimensions;
uniform sampler2D u_texmap;
varying vec2 v_texcoord;
varying vec4 v_color;
void main() {
  vec4 sampledColor = texture2D(
    u_texmap,
    vec2(v_texcoord.s / u_map_dimensions, v_texcoord.t / u_map_dimensions)
  );
  gl_FragColor = vec4(sampledColor.rgb * v_color.rgb, sampledColor.a * v_color.a);
}

TWGL

After the texture atlas is constructed, the WebGL boilerplate can be started. To simplify things, I use a helper library called TWGL which simplifies compiling the shader code, sending data to the shader, and managing texture settings. Using TWGL, you compile a shader "program" using the glsl shader code, and construct an arrays object containing the data arrays for each of the shader attributes.

  export const canvas = document.createElement("canvas");
canvas.setAttribute("touch-action", "none");
document.body.appendChild(canvas);
const gl = canvas.getContext("webgl", {alpha: false});
let spriteProgram = twgl.createProgramInfo(gl, [vert, frag]);
gl.useProgram(spriteProgram.program);
let maxCount = 800;
let spriteArrays = {
  a_coord: {numComponents: 2, data: new Float32Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW},
  a_position: {numComponents: 3, data: new Float32Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW},
  a_texcoord: {numComponents: 2, data: new Float32Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW},
  a_rotation: {numComponents: 1, data: new Float32Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW},
  a_dimensions: {numComponents: 2, data: new Float32Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW},
  a_center: {numComponents: 2, data: new Float32Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW},
  a_scale: {numComponents: 1, data: new Float32Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW},
  a_color: {numComponents: 4, data: new Float32Array(maxCount), drawType: gl.DYNAMIC_DRAW},
  indices: {numComponents: 3, data: new Uint16Array(maxCount * 2), drawType: gl.DYNAMIC_DRAW}
};
let bufferInfo = twgl.createBufferInfoFromArrays(gl, spriteArrays);

In the above, the vert and frag variables are the text from the vertex and fragment assets imported by Parcel into strings like so:

  import vert from './shaders/vert.glsl';
import frag from './shaders/frag.glsl';

Then every frame I clear the viewport, set the camera position, and resize the attribute arrays if needed.

  export function drawToScreen() {
  gl.clearColor(0, 0, 0, 1);
  gl.clear(gl.COLOR_BUFFER_BIT);
  gl.viewport(0, 0, screenSize.x, screenSize.y);
  twgl.setUniforms(spriteProgram, {
    u_camera_dimensions: [0, 0, screenSize.x, screenSize.y],
    u_texmap: textures.texture,
    u_map_dimensions: textures.size
  });
  for (let id in spriteArrays) {
    let expectedLength = 0;
    if (id == "indices") {
      expectedLength = imagesToDraw.length * spriteArrays[id].numComponents * 2;
    } else {
      expectedLength = imagesToDraw.length * spriteArrays[id].numComponents * 4;
    }
    if (spriteArrays[id].data.length < expectedLength) {
      console.log(expectedLength);
      if (id == "indices") {
        spriteArrays[id].data = new Uint16Array(expectedLength);
      } else {
        spriteArrays[id].data = new Float32Array(expectedLength);
      }
    }
  }

During the draw portion of every frame, modules in the game can call an image function passing an image name, and location/tint information to draw in this frame. Since a draw call draws the vertices in order, the image function acts somewhat like an immediate mode API which simplifies thinking about how things will be drawn to the screen. However the images are not drawn immediately and instead are stored in a list to be drawn at the end of the frame.

  let imagesToDraw = [];
export function image(imageUrl, position, dimensions, rotation = 0, color = Color.white, center = Vector.half) {
  imagesToDraw.push({ imageUrl, position, dimensions, rotation, color, center });
}

I then splice in data for each of the images at the expected location. Since some browsers still do not support instanced data effectively, I use a helper method to copy the vertex data for each of the 4 vertexes of a quad made from two triangles. spliceData repeats the passed in array of data 4 times while splice array just copies the single array slice into the destination array.

  export function spliceArray(dest: Uint16Array | Float32Array, offset: number, data: number[]) {
  for (let i = 0; i < data.length; i++) {
    dest[offset + i] = data[i];
  }
}
export function spliceData(array: {numComponents: number, data: Float32Array | Uint16Array}, entityIndex: number, data: number[]) {
  let expectedCount = array.numComponents * 4;
  for (let i = 0; i < expectedCount; i += data.length) {
    spliceArray(array.data, entityIndex * expectedCount + i, data);
  }
}

I suspect that there are standard optimized versions of these functions which would speed things up somewhat, but I have yet to encounter performance problems with the engine as is, so I probably will not worry too much about it.

These helpers are used to copy the data into the attribute arrays like so:

  imagesToDraw.sort((a, b) => a.position.z - b.position.z);
let index = 0;
for (let imageToDraw of imagesToDraw) {
  spliceData(spriteArrays.a_coord, index, [ 0, 1, 1, 1, 0, 0, 1, 0 ]);
  spliceData(spriteArrays.a_position, index, [
    imageToDraw.position.x,
    imageToDraw.position.y,
    imageToDraw.position.z
  ]);
  spliceData(spriteArrays.a_texcoord, index, textures.texCoords[imageToDraw.imageUrl]);
  spliceData(spriteArrays.a_rotation, index, [imageToDraw.rotation || 0]);
  spliceData(spriteArrays.a_dimensions, index, [
    imageToDraw.dimensions.x,
    imageToDraw.dimensions.y
  ]);
  spliceData(spriteArrays.a_center, index, [
    imageToDraw.center.x,
    imageToDraw.center.y
  ]);
  spliceData(spriteArrays.a_scale, index, [1]);
  spliceData(spriteArrays.a_color, index, [imageToDraw.color.r, imageToDraw.color.g, imageToDraw.color.b, imageToDraw.color.a]);
  let offset = index * 4;
  spliceArray(spriteArrays.indices.data, index * 6,
              [offset + 0, offset + 1, offset + 2, offset + 2, offset + 1, offset + 3]);
  index++;
}

With the attribute arrays filled, all that is left is to bind the arrays to the gpu memory, make the draw call, and clear the imagesToDraw list for the next frame.

  for (let id in spriteArrays) {
  if (id != "indices") {
    twgl.setAttribInfoBufferFromArray(gl, bufferInfo.attribs[id], spriteArrays[id]);
  } else {
    gl.bindBuffer(gl.ELEMENT_ARRAY_BUFFER, bufferInfo.indices);
    gl.bufferData(gl.ELEMENT_ARRAY_BUFFER, spriteArrays[id].data, spriteArrays[id].drawType);
  }
}
twgl.setBuffersAndAttributes(gl, spriteProgram, bufferInfo);
twgl.drawBufferInfo(gl, bufferInfo, gl.TRIANGLES, imagesToDraw.length * 6);
imagesToDraw = [];

And thats it! I've found this method of drawing textures to the screen to be incredibly flexible and fast. I've used this same basic structure in a number of experiments and basic game engines to good effect. At some point I will reassess the browser support for instanced geometry which would speed things up even further by not requiring copies of the quad and repeat data in the attributes buffers for each vertex, but until then this works fine and has great support on all of the browsers I've tried it on. As an added benefit, I have stronger more consistent control over how the pixels are sampled which is a huge plus as the normal canvas APIs do not have a clean way to do nearest neighbor sampling cross platform.

Thats about it for today, not a very visually interesting post, but hopefully somebody finds this useful!

Till tomorrow,
Keith