Custom seed for Math.random() in GPU.js

[jerrygreen]

Under the hood, GPU.js uses seed for the Math.random, but can I give it my own seed? I want to use my own seed.

[hiSandog]

A custom seed would be useful, but deterministic behavior needs to be defined across CPU and GPU backends. Should the seed be per kernel creation, per kernel invocation, or per thread/index inside the kernel? A small example showing two runs producing identical output would make the expected API much clearer.

[jerrygreen]

Any way to pass a seed would be nice, whatever is easier.

I see the project isn’t really updated so my hopes for the change are abysmal.

There are million ways to make your own prng function but this adds a lot of friction – and it complicates the user codebase a lot for things that I feel should be part of library code (or more precisely, package code, since it’s the common term in js to call a library).

Having some kind of setup for creating kernel which will change how Math.random() works (putting user-defined seed instead of hardcoding a random value from CPU) – such option would be great.

If you want a concrete example how I imagine this should look like, here is an example updated from readme:

<script src="dist/gpu-browser.min.js"></script>
<script>
    // GPU is a constructor and namespace for browser
    const gpu = new GPU();
+   gpu.setSeed(99999)
    const multiplyMatrix = gpu.createKernel(function(a, b) {
        let sum = 0;
        for (let i = 0; i < 512; i++) {
-            sum += a[this.thread.y][i] * b[i][this.thread.x];
+            sum += a[this.thread.y][i] * Math.random();
        }
        return sum;
    }).setOutput([512, 512]);

    const c = multiplyMatrix(a, b);
</script>

Or maybe something like this:

<script src="dist/gpu-browser.min.js"></script>
<script>
    // GPU is a constructor and namespace for browser
    const gpu = new GPU();
    const multiplyMatrix = gpu.createKernel(function(a, b) {
        let sum = 0;
        for (let i = 0; i < 512; i++) {
-            sum += a[this.thread.y][i] * b[i][this.thread.x];
+            sum += a[this.thread.y][i] * Math.random();
        }
        return sum;
-    }).setOutput([512, 512]);
+    }).setSeed(99999).setOutput([512, 512]);

    const c = multiplyMatrix(a, b);
</script>

The latter one probably makes more sense since kernel already has options and methods such as setOutput. So it’s only a matter of adding yet another option/method that is applied as seed value instead of the one library hardcoded to receive from CPU.