Scope-based usage tracking in the render pass

[kvark]

Currently, we are creating a new resource tracker for each recorded render pass. It's short-lived and it only accumulates the "current" state of every used resource as a union of all the usages needed by the pass. Once a pass has finished recording, we'll stitch the current state of each resource before the pass with the combined usage of a pass.

It's unfortunate to re-populate a resource tracker from the ground up. I wonder if it would make sense to just extend the number of states stored per resource in the tracker of a command buffer. Currently it's just (initial, current). We could make it (initial, current, temporary). When a pass is started, we'd do something like:

for resource in command_buffer.trackers.resources {
  resource.enter_scope(); 
}
impl Resource {
  fn enter_scope(&mut self) {
    self.temporary = self.current;
    self.current = empty;
  }
}

Then we'd merge all the usage we encounter in a pass (into current) and at the end of the pass do:

for resource in command_buffer.trackers.resources {
  resource.leave_scope(); 
}
impl Resource {
  fn leave_scope(&mut self) {
    if self.current != empty {
      transition(self.temp, self.current);
    }
    self.current = self.temp;
  }
}

The benefit is - having less stuff done for a render pass, so lower overhead.

[cwfitzgerald]

Obsoleted by #2662