alias number = f32;

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

@group(0) @binding(0)
var<storage, read> l: MatrixBatch;
@group(0) @binding(1)
var<storage, read> r: MatrixBatch;
@group(0) @binding(2)
var<storage, read_write> result: MatrixBatch;

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

  var s = 0f;
  var li = global_id.z * l.size.x * l.size.y + global_id.y * l.size.x;
  var ri = global_id.z * r.size.x * r.size.y + global_id.x;
  for (var i = 0u; i < l.size.x; i++) {
    s += l.data[li] * r.data[ri];
    li++;
    ri += r.size.x;
  }

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