From AI and cloud costs to GPU compilers, I make engineering worth every euro.

use system.collections.array

const W = 64
const SIZE = W * W

gpu var src = Array<f32, SIZE>()       // lives in gpu memory
gpu var dst = Array<f32, SIZE>()

// a white square on black: one gpu thread per pixel
forall idx in 0..SIZE
    let y = idx / W
    let x = idx - y * W
    let inside = x >= 24 and x < 40 and y >= 24 and y < 40
    src[idx] = 1.0 if inside else 0.0

// 3×3 box blur: all 4,096 pixels at once
forall idx in 0..SIZE
    let y = idx / W
    let x = idx - y * W
    var sum = 0.0
    var dy = -1
    while dy <= 1
        var dx = -1
        while dx <= 1
            let nx = x + dx
            let ny = y + dy
            if nx >= 0 and nx < W and ny >= 0 and ny < W
                sum = sum + src[ny * W + nx]
            dx = dx + 1
        dy = dy + 1
    dst[idx] = sum / 9.0

let host = dst                         // read back to the cpu
println(f"centre={host[32 * W + 32]} corner={host[24 * W + 24]}")