struct MatrixBatch {
  size: vec4<u32>,
  data: array<f32>,
}

@group(0) @binding(0)
var<storage, read> e: MatrixBatch;
@group(0) @binding(1)
var<storage, read> dlds: MatrixBatch;
@group(0) @binding(2)
var<storage, read_write> gtattw: MatrixBatch;

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

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