[ETFeeder] Resolve deps based on not only data_deps, but also ctrl_deps

[changhai0109]

Summary

In ETFeederNode, add fields of all_deps, which is a complete set of data_deps and ctrl_deps. Add unreleased_deps to track the parents of a node which is not issued yet. Update deps resolving reference to all_deps instead of data_deps.

Test Plan

Add following function to et_generator

def two_comp_nodes_ctrl_dependent(num_npus: int, runtime: int) -> None:
    for npu_id in range(num_npus):
        output_filename = f"two_comp_nodes_ctrl_dependent.{npu_id}.et"
        with open(output_filename, "wb") as et:
            encode_message(et, GlobalMetadata(version="0.0.4"))

            parent_node = get_node("COMP_NODE", COMP_NODE)
            parent_node.duration_micros = runtime
            parent_node.attr.append(ChakraAttr(name="is_cpu_op", bool_val=False))
            encode_message(et, parent_node)

            child_node = get_node("COMP_NODE", COMP_NODE)
            child_node.duration_micros = runtime
            child_node.ctrl_deps.append(parent_node.id)
            child_node.attr.append(ChakraAttr(name="is_cpu_op", bool_val=False))
            encode_message(et, child_node)

register this function in main as follows

def main():
   ...
  two_comp_nodes_ctrl_dependent(args.num _npus, args.default_runtime)
  ...

Run et_generator and get two_comp_nodes_ctrl_dependent.{npu_id}.et, then run with astrasim.

[github-actions[bot]]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅