// SPDX-License-Identifier: AGPL-3.0-or-later
//
// ESN readout kernel (f32)
//
// output[i] = W_out[i,:] · state (matrix-vector product)
//
// Separated from reservoir update so readout can run on CPU
// while reservoir runs on GPU/NPU. In practice, readout is
// cheap (output_size << reservoir_size) and often stays on host.
//
// Provenance: hotSpring v0.6.0 (Stanton-Murillo transport, ESN D*/η*/λ* prediction)
//
// Bindings:
// 0: w_out [O*R] f32, read — readout weights (row-major, O × R)
// 1: state [R] f32, read — reservoir state
// 2: output [O] f32, rw — prediction (written)
// 3: params [4] f32, read — [reservoir_size, output_size, _, _]
@group(0) @binding(0) var<storage, read> w_out: array<f32>;
@group(0) @binding(1) var<storage, read> state: array<f32>;
@group(0) @binding(2) var<storage, read_write> output: array<f32>;
@group(0) @binding(3) var<storage, read> params: array<f32>;
@compute @workgroup_size(64)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
let i = gid.x;
let R = u32(params[0]);
let O = u32(params[1]);
if (i >= O) { return; }
var sum: f32 = 0.0;
let row = i * R;
for (var j: u32 = 0u; j < R; j = j + 1u) {
sum += w_out[row + j] * state[j];
}
output[i] = sum;
}