alias number = f32;

struct MatrixBatch {
  size: vec4<u32>,
  data: array<number>,
}

struct VectorBatch {
  size: vec4<u32>,
  data: array<number>,
}

@group(0) @binding(0)
var<storage, read> m: MatrixBatch;
@group(0) @binding(1)
var<storage, read_write> result: VectorBatch;

@compute @workgroup_size(1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
  result.size.x = m.size.y;
  result.size.z = m.size.z;

  var s = 0f;
  let lo = global_id.z * m.size.x * m.size.y + global_id.x * m.size.x;
  for (var i = 0u; i < m.size.x; i++) {
    s += m.data[lo + i] * m.data[lo + i];
  }

  result.data[global_id.z * result.size.x + global_id.x] = s;
}