Add support for the "expanded cost features" model.

This commit includes two separate changes:

  1) the addition of extra features which contribute to the "inlining
  cost" estimate, and
  2) extending the size and training duration of the model to account
  for the additional features.

Author: jacob-hegna

compiler_opt/rl/feature_ops.py

@@ -42,6 +42,7 @@ def _build_quantile_map(quantile_file_dir):
 
 @gin.configurable
 def get_observation_processing_layer_creator(quantile_file_dir,
+                                             with_sqrt=False,
                                              with_z_score_normalization=False,
                                              eps=1e-8):
   """Wrapper for observation_processing_layer."""
@@ -67,7 +68,9 @@ def normalization(obs):
         x = tf.cast(
             tf.raw_ops.Bucketize(input=expanded_obs, boundaries=quantile),
             tf.float32) / len(quantile)
-        features = [x, tf.sqrt(x), x * x]
+        features = [x, x * x]
+        if with_sqrt:
+          features.append(np.sqrt(x))
         if with_z_score_normalization:
           y = tf.cast(expanded_obs, tf.float32)
           y = (y - mean) / (std + eps)

compiler_opt/rl/feature_ops_test.py

@@ -18,11 +18,18 @@
 import os
 
 from absl.testing import parameterized
+import numpy as np
 import tensorflow as tf
 
 from compiler_opt.rl import constant
 from compiler_opt.rl import feature_ops
 
+_WITH_Z_SCORE_SQRT_PRODUCT_VALUES = [('with_sqrt_with_z_score', True, True),
+                                     ('with_sqrt_without_z_score', True, False),
+                                     ('without_sqrt_with_z_score', False, True),
+                                     ('without_sqrt_without_z_score', False,
+                                      False)]
+
 
 class FeatureUtilsTest(tf.test.TestCase, parameterized.TestCase):
 
@@ -50,12 +57,11 @@ def test_build_quantile_map_from_config(self):
     # std
     self.assertAllClose(14.988885, std)
 
-  @parameterized.named_parameters(('with_z_score', True),
-                                  ('without_z_score', False))
-  def test_create_observation_processing_layer(self, with_z_score):
+  @parameterized.named_parameters(*_WITH_Z_SCORE_SQRT_PRODUCT_VALUES)
+  def test_create_observation_processing_layer(self, with_z_score, with_sqrt):
     observation_processing_layer = (
         feature_ops.get_observation_processing_layer_creator(
-            self._quantile_file_dir, with_z_score))
+            self._quantile_file_dir, with_sqrt, with_z_score))
 
     obs_spec = tf.TensorSpec(dtype=tf.float32, shape=(), name='edge_count')
     processing_layer = observation_processing_layer(obs_spec)
@@ -65,24 +71,28 @@ def test_create_observation_processing_layer(self, with_z_score):
 
     outputs = self.evaluate(outputs)
 
+    expected_shape = [2, 1, 2]
+    expected = np.array([[[0.333333, 0.111111]], [[0.777778, 0.604938]]])
+
+    if with_sqrt:
+      expected_shape[2] += 1
+      expected = np.concatenate([expected, [[[0.57735]], [[0.881917]]]],
+                                axis=-1)
+
     if with_z_score:
-      self.assertAllEqual([2, 1, 4], outputs.shape)
-      self.assertAllClose([[[0.333333, 0.57735, 0.111111, -0.555968]],
-                           [[0.777778, 0.881917, 0.604938, -0.155671]]],
-                          outputs)
-    else:
-      self.assertAllEqual([2, 1, 3], outputs.shape)
-      self.assertAllClose(
-          [[[0.333333, 0.57735, 0.111111]], [[0.777778, 0.881917, 0.604938]]],
-          outputs)
-
-  @parameterized.named_parameters(('with_z_score', True),
-                                  ('without_z_score', False))
+      expected_shape[2] += 1
+      expected = np.concatenate([expected, [[[-0.555968]], [[-0.155671]]]],
+                                axis=-1)
+
+    self.assertAllEqual(expected_shape, outputs.shape)
+    self.assertAllClose(expected.tolist(), outputs)
+
+  @parameterized.named_parameters(*_WITH_Z_SCORE_SQRT_PRODUCT_VALUES)
   def test_create_observation_processing_layer_for_dummy_features(
-      self, with_z_score):
+      self, with_z_score, with_sqrt):
     observation_processing_layer = (
         feature_ops.get_observation_processing_layer_creator(
-            self._quantile_file_dir, with_z_score))
+            self._quantile_file_dir, with_sqrt, with_z_score))
 
     obs_spec = tf.TensorSpec(dtype=tf.float32, shape=(), name='dummy_feature')
     processing_layer = observation_processing_layer(obs_spec)

compiler_opt/rl/inlining/config.py

@@ -27,6 +27,7 @@ def get_inlining_signature_spec():
   """Returns (time_step_spec, action_spec) for LLVM inlining."""
   observation_spec = dict(
       (key, tf.TensorSpec(dtype=tf.int64, shape=(), name=key)) for key in (
+          # Base features
           'caller_basic_block_count',
           'caller_conditionally_executed_blocks',
           'caller_users',
@@ -38,6 +39,33 @@ def get_inlining_signature_spec():
           'edge_count',
           'callsite_height',
           'cost_estimate',
+
+          # Expanded cost features
+          'sroa_savings',
+          'sroa_losses',
+          'load_elimination',
+          'call_penalty',
+          'call_argument_setup',
+          'load_relative_intrinsic',
+          'lowered_call_arg_setup',
+          'indirect_call_penalty',
+          'jump_table_penalty',
+          'case_cluster_penalty',
+          'switch_penalty',
+          'unsimplified_common_instructions',
+          'num_loops',
+          'dead_blocks',
+          'simplified_instructions',
+          'constant_args',
+          'constant_offset_ptr_args',
+          'callsite_cost',
+          'cold_cc_penalty',
+          'last_call_to_static_bonus',
+          'is_multiple_blocks',
+          'nested_inlines',
+          'nested_inline_cost_estimate',
+          'threshold',
+
           # inlining_default is not used as feature in training.
           'inlining_default'))
   reward_spec = tf.TensorSpec(dtype=tf.float32, shape=(), name='reward')

compiler_opt/rl/inlining/gin_configs/behavioral_cloning_nn_agent.gin

@@ -11,13 +11,14 @@ train_eval.batch_size=64
 train_eval.train_sequence_length=1
 
 get_observation_processing_layer_creator.quantile_file_dir='compiler_opt/rl/inlining/vocab'
+get_observation_processing_layer_creator.with_sqrt = False
 get_observation_processing_layer_creator.with_z_score_normalization = False
 
 create_agent.policy_network = @q_network.QNetwork
 
 [email protected]()
-QNetwork.fc_layer_params=(40, 20)
-QNetwork.dropout_layer_params=(0.2, 0.2)
+QNetwork.fc_layer_params=(40, 40, 20)
+QNetwork.dropout_layer_params=(0.2, 0.2, 0.2)
 [email protected]
 
 tf.train.AdamOptimizer.learning_rate = 0.001

compiler_opt/rl/inlining/gin_configs/ppo_nn_agent.gin

@@ -9,19 +9,20 @@ [email protected]_inlining_signature_spec
 train_eval.agent_name='ppo'
 train_eval.warmstart_policy_dir=''
 train_eval.num_policy_iterations=3000
-train_eval.num_iterations=200
+train_eval.num_iterations=300
 train_eval.batch_size=128
 train_eval.train_sequence_length=16
 train_eval.deploy_policy_name='saved_collect_policy'
 train_eval.use_stale_results=False
 
 get_observation_processing_layer_creator.quantile_file_dir='compiler_opt/rl/inlining/vocab'
+get_observation_processing_layer_creator.with_sqrt = False
 get_observation_processing_layer_creator.with_z_score_normalization = False
 
 create_agent.policy_network = @actor_distribution_network.ActorDistributionNetwork
 
 ActorDistributionNetwork.preprocessing_combiner=@tf.keras.layers.Concatenate()
-ActorDistributionNetwork.fc_layer_params=(40, 20)
+ActorDistributionNetwork.fc_layer_params=(40, 40, 20)
 ActorDistributionNetwork.dropout_layer_params=None
 ActorDistributionNetwork.activation_fn=@tf.keras.activations.relu