Warp (subgroup) operations

Threads execute in hardware groups — warps (NVIDIA), waves (AMD), SIMD-groups (Apple) — that can exchange registers without touching shared memory. kernel.warp exposes the portable subset: kernel.warp.size, kernel.warp.lane_id, and kernel.warp.shuffle_down(v, n) (the shuffle offset is a compile-time literal). The canonical use is a register-only tree reduction:

use system.gpu
use system.collections.array

gpu fn warp_reduce_sum(input Array<int, 32>, dst out Array<int, 1>)
    let lane = kernel.warp.lane_id
    var v = input[lane]

    v = v + kernel.warp.shuffle_down(v, 16)
    v = v + kernel.warp.shuffle_down(v, 8)
    v = v + kernel.warp.shuffle_down(v, 4)
    v = v + kernel.warp.shuffle_down(v, 2)
    v = v + kernel.warp.shuffle_down(v, 1)

    if lane == 0
        dst[0] = v

fn main()
    gpu let input = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32]
    gpu var dst = Array<int,1>()
    warp_reduce_sum(input, dst).launch(Dim3(1, 1, 1), Dim3(32, 1, 1))
    let result = dst
    println(f'{result[0]}')

Output: 528 — the sum of 1..32, reduced in five shuffle steps with no shared memory and no barrier. Warp ops require adapter subgroup support. A device without it refuses the kernel before dispatch rather than miscomputing.