// SPDX-License-Identifier: AGPL-3.0-only
//
// Subgroup-accelerated tree reduction: uses subgroupAdd for warp/wavefront-level
// reduction, then shared memory for cross-subgroup accumulation.
//
// Absorbed from hotSpring V0632 (March 2026).
//
// RTX 3090: subgroup_size=32 -> 8 subgroups per WG, 3 barrier steps
// RX 6950 XT: subgroup_size=64 -> 4 subgroups per WG, 2 barrier steps
// vs 8 barrier steps in the shared-memory-only version.
//
// NOTE: Do NOT add `enable subgroups;` — naga 28 generates broken SPIR-V
// when the enable directive is present, causing all subgroup ops to return 0.
// wgpu's SUBGROUP device feature is sufficient; the directive is redundant.
struct ReduceParams {
size: u32,
_pad1: u32,
_pad2: u32,
_pad3: u32,
}
@group(0) @binding(0) var<storage, read> input: array<f64>;
@group(0) @binding(1) var<storage, read_write> output: array<f64>;
@group(0) @binding(2) var<uniform> params: ReduceParams;
var<workgroup> wg_partial: array<f64, 8>;
@compute @workgroup_size(256)
fn main(
@builtin(global_invocation_id) global_id: vec3<u32>,
@builtin(local_invocation_id) local_id: vec3<u32>,
@builtin(workgroup_id) workgroup_id: vec3<u32>,
@builtin(num_workgroups) nwg: vec3<u32>,
@builtin(subgroup_invocation_id) sg_id: u32,
@builtin(subgroup_size) sg_size: u32,
) {
let tid = local_id.x;
let gid = global_id.x + global_id.y * nwg.x * 256u;
var val: f64 = f64(0.0);
if gid < params.size {
val = input[gid];
}
let sg_sum = subgroupAdd(val);
let sg_idx = tid / sg_size;
if sg_id == 0u {
wg_partial[sg_idx] = sg_sum;
}
workgroupBarrier();
let n_subgroups = 256u / sg_size;
if tid < n_subgroups {
val = wg_partial[tid];
} else {
val = f64(0.0);
}
if tid < sg_size {
let final_sum = subgroupAdd(val);
if tid == 0u {
let wg_linear = workgroup_id.x + workgroup_id.y * nwg.x;
output[wg_linear] = final_sum;
}
}
}