amoffat / gpgpu.js

gpgpu.js

This is a toy utility for this blog post. It is designed to abstract away WebGL as a graphics technology and make it behave more like a general-purpose computing technology, like CUDA or OpenCL.

I will probably not be expanding on this utility unless there is substantial interest.

Benchmarks

Running the following kernel on 100M floats:

• CPU Kernel: (num) -> num + Math.tan(Math.cos(Math.sin(num * num)))
• GPU Kernel: dst = src + tan(cos(sin(src * src)));

CPU: 6851.25ms
GPU Total: 1449.29ms
GPU Execution: 30.64ms
GPU IO: 1418.65ms
Theoretical Speedup: 223.59x
Actual Speedup: 4.73x

Basic Usage

root = $("#gpgpu")
engine = getEngine root

kernel = "dst = src + tan(cos(sin(src * src)));"
inputs = new Float32Array [1, 2, 3, 4, 5]

outputs = engine.execute kernel, inputs

User-defined functions

root = $("#gpgpu")
engine = getEngine root

inputs = new Float32Array [1, 2, 3, 4, 5]
preamble = """
float cube(float v) {
    return v * v * v;
}
"""
kernel = "dst = cube(src);"
outputs = engine.execute inputs, kernel, preamble

Uniforms

inputData = new Float32Array [1, 2, 3, 4, 5]
inputs = Storage.fromData engine, inputData

kernel = "dst = src * float(num);"
preamble = "uniform int num;"

comp = engine.createComputation inputData.length, kernel, preamble
outputs = comp.execute inputs, {num:["1i", 3]}