generate_default_trace.py

# coding=utf-8
# Copyright 2020 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Generate initial training data from the behavior of the current heuristic."""

import concurrent.futures
import contextlib
import functools
import re
from typing import Dict, List, Optional, Union, Tuple  # pylint:disable=unused-import

from absl import app
from absl import flags
from absl import logging
import gin

import tensorflow as tf

from compiler_opt.distributed import worker
from compiler_opt.distributed import buffered_scheduler
from compiler_opt.distributed.local import local_worker_manager

from compiler_opt.rl import compilation_runner
from compiler_opt.rl import corpus
from compiler_opt.rl import policy_saver
from compiler_opt.rl import registry

from tf_agents.system import system_multiprocessing as multiprocessing

# see https://bugs.python.org/issue33315 - we do need these types, but must
# currently use them as string annotations

_DATA_PATH = flags.DEFINE_string('data_path', None,
                                 'Path to folder containing IR files.')
_POLICY_PATH = flags.DEFINE_string(
    'policy_path', '', 'Path to the policy to generate trace with.')
_OUTPUT_PATH = flags.DEFINE_string(
    'output_path', None, 'Path to the output tfrecord file if not None.')
_OUTPUT_PERFORMANCE_PATH = flags.DEFINE_string(
    'output_performance_path', None,
    'Path to the output performance file if not None.')
_NUM_WORKERS = flags.DEFINE_integer(
    'num_workers', None,
    'Number of parallel workers for compilation. `None` for maximum available.')
_SAMPLING_RATE = flags.DEFINE_float(
    'sampling_rate', 1,
    'Sampling rate of modules, 0.5 means 50% sampling rate that generates data '
    'for half modules.')
_MODULE_FILTER = flags.DEFINE_string(
    'module_filter', None,
    'Regex for module names to include, do not provide one to include all')
_KEY_FILTER = flags.DEFINE_string(
    'key_filter', None,
    'Regex for key names to include, do not provide one to include all')
_GIN_FILES = flags.DEFINE_multi_string(
    'gin_files', [], 'List of paths to gin configuration files.')
_GIN_BINDINGS = flags.DEFINE_multi_string(
    'gin_bindings', [],
    'Gin bindings to override the values set in the config files.')

ResultsQueueEntry = Union[Optional[Tuple[str, List[str],
                                         Dict[str,
                                              compilation_runner.RewardStat]]],
                          BaseException]


class FilteringWorker(worker.Worker):
  """Worker that performs a computation and optionally filters the result.


  Args:
    policy_path: the policy_path to generate trace with.
    key_filter: regex filter for key names to include, or None to include all.
  """

  def __init__(self, policy_path: Optional[str], key_filter: Optional[str],
               runner_type: 'type[compilation_runner.CompilationRunner]',
               runner_kwargs):
    self._policy_path = policy_path
    self._key_filter = re.compile(key_filter) if key_filter else None
    self._runner = runner_type(**runner_kwargs)
    self._policy = policy_saver.Policy.from_filesystem(
        policy_path) if policy_path else None

  def compile_and_filter(
      self, loaded_module_spec: corpus.LoadedModuleSpec
  ) -> Tuple[str, List[str], Dict[str, compilation_runner.RewardStat]]:
    data = self._runner.collect_data(
        loaded_module_spec=loaded_module_spec,
        policy=self._policy,
        reward_stat=None,
        model_id=0)
    if self._key_filter is None:
      return (loaded_module_spec.name, data.serialized_sequence_examples,
              data.reward_stats)
    new_reward_stats = {}
    new_sequence_examples = []
    for k, sequence_example in zip(data.keys,
                                   data.serialized_sequence_examples):
      if not self._key_filter.match(k):
        continue
      new_reward_stats[k] = data.reward_stats[k]
      new_sequence_examples.append(sequence_example)
    return (loaded_module_spec.name, new_sequence_examples, new_reward_stats)


def main(_):
  gin.parse_config_files_and_bindings(
      _GIN_FILES.value, bindings=_GIN_BINDINGS.value, skip_unknown=False)
  logging.info(gin.config_str())
  generate_trace()


def generate_trace(
    worker_manager_class=local_worker_manager.LocalWorkerPoolManager):

  config = registry.get_configuration()

  logging.info('Loading module specs from corpus.')
  module_filter = re.compile(
      _MODULE_FILTER.value) if _MODULE_FILTER.value else None

  cps = corpus.Corpus(
      data_path=_DATA_PATH.value,
      module_filter=lambda name: True
      if not module_filter else module_filter.match(name),
      additional_flags=config.flags_to_add(),
      delete_flags=config.flags_to_delete(),
      replace_flags=config.flags_to_replace())
  logging.info('Done loading module specs from corpus.')

  # Sampling if needed.
  sampled_modules = int(len(cps) * _SAMPLING_RATE.value)
  # sort files by size, to process the large files upfront, hopefully while
  # other smaller files are processed in parallel
  corpus_elements = cps.sample(k=sampled_modules, sort=True)

  tfrecord_context = (
      tf.io.TFRecordWriter(_OUTPUT_PATH.value)
      if _OUTPUT_PATH.value else contextlib.nullcontext())
  performance_context = (
      tf.io.gfile.GFile(_OUTPUT_PERFORMANCE_PATH.value, 'w')
      if _OUTPUT_PERFORMANCE_PATH.value else contextlib.nullcontext())
  work = [
      cps.load_module_spec(corpus_element) for corpus_element in corpus_elements
  ]

  runner_type = config.get_runner_type()
  with tfrecord_context as tfrecord_writer:
    with performance_context as performance_writer:
      with worker_manager_class(
          FilteringWorker,
          _NUM_WORKERS.value,
          policy_path=_POLICY_PATH.value,
          key_filter=_KEY_FILTER.value,
          runner_type=runner_type,
          runner_kwargs=worker.get_full_worker_args(
              runner_type, moving_average_decay_rate=0)) as lwm:

        _, result_futures = buffered_scheduler.schedule_on_worker_pool(
            action=lambda w, j: w.compile_and_filter(j),
            jobs=work,
            worker_pool=lwm)
        total_successful_examples = 0
        total_work = len(corpus_elements)
        total_failed_examples = 0
        total_training_examples = 0
        not_done = result_futures
        while not_done:
          (done, not_done) = concurrent.futures.wait(not_done, 10)
          succeeded = [
              r for r in done if not r.cancelled() and r.exception() is None
          ]
          total_successful_examples += len(succeeded)
          total_failed_examples += (len(done) - len(succeeded))
          for r in succeeded:
            module_name, records, reward_stat = r.result()
            if tfrecord_writer:
              total_training_examples += len(records)
              for r in records:
                tfrecord_writer.write(r)
            if performance_writer:
              for key, value in reward_stat.items():
                performance_writer.write(
                    (f'{module_name},{key},{value.default_reward},'
                     f'{value.moving_average_reward}\n'))
          logging.info('%d success, %d failed out of %d',
                       total_successful_examples, total_failed_examples,
                       total_work)

  print((f'{total_successful_examples} of {len(corpus_elements)} modules '
         f'succeeded, and {total_training_examples} trainining examples '
         'written'))


if __name__ == '__main__':
  flags.mark_flag_as_required('data_path')
  multiprocessing.handle_main(functools.partial(app.run, main))