On-device reduction

.reduce(init, fold) on a gpu-resident array runs as a single-launch tree reduction over workgroup shared memory — no host round-trip per element:

use system.gpu

gpu let data = [1, 2, 3, 4, 5, 6, 7, 8]
let sum = data.reduce(0, fn(a int, b int) int: a + b)
println(f'sum = {sum}')

Output: sum = 36. The result of a device reduce is a gpu-resident scalar. Binding it to a host let (as above) is the fence — a single-element readback brings the value over. Bind it with gpu let instead and it stays on the device:

use system.gpu
use system.io
use system.collections.array

fn main()
    gpu var data = [1, 2, 3, 4]
    gpu let sum = data.reduce(0, fn(a i32, b i32) i32: a + b)
    let host_sum = sum
    println(f'{host_sum}')

Output: 10. The fold must be an associative binary operator — + or * over the two parameters. init is folded in exactly once. Anything else is rejected at compile time. On a host array the same .reduce call falls through to the ordinary CPU fold, so one spelling covers both residencies.