alias number = f32;

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

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

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

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

  var vc = result.size.x / 4;
  var resi = global_id.y * vc + global_id.x;
  result.data[resi] = vec4(0f);

  var mi = resi;
  for (var z = 0u; z < m.size.z; z++) {
    result.data[resi] += m.data[mi] / vec4(f32(m.size.z));
    mi += m.size.y * vc;
  }
}