gpu-ez

Easy GLSL parallel computing on GPU.JS

• for Web-browsers and Node.js

Power!

GPU.JS. Nuff said. The tons of wrapping, workarounds and several useful helpers we can use in GLSL code.

Format

GLSL

GLSL content must contain string of the following pattern in comment:

GLSL: functionName[vectorSize](functionArgs, ...) {dimensions} {maxIterations}

where:

• functionName -- identifier: a name of entry point function, which must be declared in GLSL code as float functionName() {...};

• vectorSize -- optional number from 2 to 4, indicating that result is a vector of specified size rather than a float number;

• functionArgs, ... -- comma-separated identifiers; can be numbers, JS arrays, or GPU arrays ("textures", returned by default from JS end of GLSL function); arrays and textures must be followed by a number in curly braces, indicating the number of dimensions; inside GLSL code, arguments get prefixed by user_ prefix; arrays are accessible via parentheses, f.x. 3D array declared as array{3} can be referred as array(z,y,x) from GLSL code; if arg name is appended by [vectorSize], it's type is a vector of specified size rather than 'float' (f.x. array[4]{1} -- one-dimensional array of 4-sized vectors);

• dimensions is representing dimensions of resulting array: comma-separated key: value pairs of format: x: ... for 1D array, y: ..., x: ... for 2D array, and z: ..., y: ..., x: ... for 3D array. ... can be JS expressions of numbers and functionArgs, representing size of particular dimension (points will appear as integer numbers in [0 ... size - 1] range). Entry function will be called parallelly to obtain corresponding value of resulting array at each point, represented by threadId structure in GLSL code; threadId.x represents the most inner dimension of resulting array;

• maxIterations must be set to JS expression of numbers and arguments, representing number of total iterations of the most inner loops inside GLSL code.

For example:

// GLSL: getCoords[3](map{2}, mapsize) {y: mapsize, x: mapsize} {1}
vec3 getCoords(void) { return vec3(threadId.x, map(threadId.x, 0), map(0, threadId.x)); }

/* GLSL: add(a, b) {x: 1} {1} */
float add(void) { return user_a + user_b; }

JS

In JS, to create JS endpoint of GPU kernel, use:

• .glsl(content) to load GLSL directly from text content;

• .glslFile(filename) to load it from file;

• .glslDir(dirname)(basename) to simplify loading of several files under the same directory.

Note that glslFile and glslDir available only from Node.js.

Call the JS endpoint passing the needed parameters, described by functionArgs above:

• endpoint(arguments, ...) -- returns GPU array ("texture");

• endpoint.arrayOut(arguments, ...) -- returns JS array.

See examples below.

Example

Classic example: Matrix Multiplication

matrix-multiply.c

// GLSL: matrixMultiply(m1{2}, m2{2}, m1h, m2w, size) {y: m1h, x: m2w} {size}

float matrixMultiply() {
  #define m1 user_m1
  #define m2 user_m2
  int size = int(user_size);

  int x = threadId.x;
  int y = threadId.y;

  float sum = 0.0;

  for (int i = 0; i < size; i++) {
    sum += m1(y, i) * m2(i, x);
  }

  return sum;
}

matrix-multiply.js

const {glslDir} = require('gpu-ez');
const glsl = glslDir(__dirname);

(async () => {
  const matrixMultiply = await glsl('matrix-multiply.c');
  const m1 = [[1, 2], [3, 4], [5, 6]];
  const m2 = [[7, 8], [9, 10]];
  const result = matrixMultiply.arrayOut(m1, m2, m1.length, m2[0].length, m2.length);
  console.log(result);
})();

More complex example: Inverse matrix

See files:

• for Node.js: example.js;
• for Web-browser: example.html.

Enjoy!

Happy scaling!