Fix typos discovered by codespell (#424)

compiler_opt/benchmark/filter_tests.py

@@ -19,7 +19,7 @@
 in their entirety can sometimes be problematic as some subsets of the tests
 available in the executable might require certain hardware configurations
 such as an X configuration with working graphics drivers, and we would prefer
-to avoid those tests. This exectuable goes through an entire test suite
+to avoid those tests. This executable goes through an entire test suite
 description and returns another test suite description containing only tests
 that pass.
 

compiler_opt/es/blackbox_optimizers.py

@@ -1080,7 +1080,7 @@ def hessv_func(x: FloatArray) -> FloatArray:
         """
         hessv = np.matmul(self.saved_hessian, x)
         # Reminder:
-        # If not using sensing-subspace Hessian, also subract diagonal gs(x)*I
+        # If not using sensing-subspace Hessian, also subtract diagonal gs(x)*I
         hessv /= np.power(self.precision_parameter, 2)
         hessv *= -1
         return hessv
@@ -1107,7 +1107,7 @@ def hessv_func(x: FloatArray) -> FloatArray:
                     np.power(self.precision_parameter, 2))
         hessv /= float(len(self.saved_perturbations))
         # Reminder:
-        # If not using sensing-subspace Hessian, also subract diagonal gs(x)*I
+        # If not using sensing-subspace Hessian, also subtract diagonal gs(x)*I
         hessv /= np.power(self.precision_parameter, 2)
         hessv *= -1
         return hessv

compiler_opt/rl/compilation_runner.py

@@ -334,7 +334,7 @@ def enable(self) -> WorkerFuture:
 class CompilationResultObserver(metaclass=abc.ABCMeta):
   """Abstract base class used to observe compilation results.
 
-  This is indended for users who need to observe compilations while they are in
+  This is intended for users who need to observe compilations while they are in
   the distributed worker pool, rather than after they have been coalesced in
   the collection script.
   """

compiler_opt/rl/distributed/agent.py

@@ -229,7 +229,7 @@ def compute_return_and_advantage(
     advantages = self.compute_advantages(rewards, returns, discounts,
                                          value_preds)
 
-    # TODO(b/171573175): remove the condition once historgrams are
+    # TODO(b/171573175): remove the condition once histograms are
     # supported on TPUs.
     if self._debug_summaries and not self._use_tpu:
       tf.compat.v2.summary.histogram(

compiler_opt/rl/distributed/ppo_train_lib.py

@@ -69,7 +69,7 @@ def train(
     agent.initialize()
 
   # Create the policy saver which saves the initial model now, then it
-  # periodically checkpoints the policy weigths.
+  # periodically checkpoints the policy weights.
   saved_model_dir = os.path.join(root_dir, actor_learner.POLICY_SAVED_MODEL_DIR)
   save_model_trigger = triggers.PolicySavedModelTrigger(
       saved_model_dir, agent, train_step, interval=1000)

compiler_opt/rl/env.py

@@ -311,7 +311,7 @@ def _get_clang_generator(
   Returns:
     A generator of tuples. Each element of the tuple is created with
     clang_session. First argument of the tuple is always an interactive
-    clang session. The second argumnet is a default clang session if
+    clang session. The second argument is a default clang session if
     interactive_only is False and otherwise the exact same interactive
     clang session object as the first argument.
   """
@@ -343,7 +343,7 @@ class MLGOEnvironmentBase:
   """Base implementation for all MLGO environments.
 
   Depending on the RL framework, one may want different implementations of an
-  enviroment (tf_agents: PyEnvironment, jax: dm-env, etc). This class
+  environment (tf_agents: PyEnvironment, jax: dm-env, etc). This class
   implements the core methods that are needed to then implement any of these
   other environments as well.
   """

compiler_opt/rl/imitation_learning/generate_bc_trajectories_lib.py

@@ -154,7 +154,7 @@ def add_feature_list(seq_example: tf.train.SequenceExample,
       np.dtype(np.float32),
       str,
   ]):
-    raise AssertionError((f'Unsupported type for feautre {feature_name}'
+    raise AssertionError((f'Unsupported type for feature {feature_name}'
                           f' of type {type(feature_list[0])}. '
                           'Supported types are np.int64, np.float32, str'))
   if isinstance(feature_list[0], np.float32):
@@ -382,7 +382,7 @@ def compile_module(
     Returns:
       sequence_example: a tf.train.SequenceExample containing the trajectory
         from compilation. In addition to the features returned from the env
-        tbe sequence_example adds the following extra features: action,
+        the sequence_example adds the following extra features: action,
         reward and module_name. action is the action taken at any given step,
         reward is the reward specified by reward_key, not necessarily the
         reward returned by the environment and module_name is the name of
@@ -933,7 +933,7 @@ def gen_trajectories(
       ModuleWorker.select_best-exploration
     worker_wait_sec: max number of seconds to wait for a worker to terminate
     worker_class_type: the class that will process each module
-    worker_class_type: allows for overrriding ModuleWorker
+    worker_class_type: allows for overriding ModuleWorker
     worker_manager_class: A pool of workers hosted on the local machines, each
       in its own process.
   """

compiler_opt/rl/imitation_learning/weighted_bc_trainer_lib.py

@@ -55,7 +55,7 @@
 class TrainingWeights:
   """Class for computing weights for training.
   
-  To use, create an instance by specifying the paritions used in
+  To use, create an instance by specifying the partitions used in
   collecting the data with generate_bc_trajectories. Next, run multiple steps
   of update_weights with collected profiles from generate_bc_trajectories,
   where each step corresponds to one pair of a pollicy profile and a
@@ -120,7 +120,7 @@ def create_new_profile(self,
     
     The regret is measured as the difference between the loss of the data_eval
     profiles and of the data_comparator profiles. The reward is the negative
-    regret normalized by the loss of hte data_comparator profiles.
+    regret normalized by the loss of the data_comparator profiles.
 
     Args:
       data_comparator: baseline profiles to measure improvement against

compiler_opt/rl/inlining/inlining_runner.py

@@ -53,7 +53,7 @@ def compile_fn(
 
     Args:
       command_line: the fully qualified command line.
-      tf_policy_path: path to TF policy direcoty on local disk.
+      tf_policy_path: path to TF policy directory on local disk.
       reward_only: whether only return native size.
 
     Returns:

compiler_opt/rl/policy_saver.py

@@ -48,7 +48,7 @@ def _split_tensor_name(name: str) -> Tuple[str, int]:
 def _get_non_identity_op(tensor):
   """Get the true output op aliased by Identity `tensor`.
 
-  Output signature tensors are in a Function that refrences the true call
+  Output signature tensors are in a Function that references the true call
   in the base SavedModel metagraph.  Traverse the function upwards until
   we find this true output op and tensor and return that.
 

compiler_opt/rl/random_net_distillation.py

@@ -31,7 +31,7 @@ def __init__(self,
                fc_layer_params=(32,),
                initial_intrinsic_reward_scale=1.0,
                half_decay_steps=10000):
-    """Initilization for RandomNetworkDistillation class.
+    """Initialization for RandomNetworkDistillation class.
 
     Args:
       time_step_spec: the time step spec for raw observation
@@ -81,7 +81,7 @@ def __init__(self,
     self._external_reward_mean = tf.keras.metrics.Mean()
 
   def _get_intrinsic_reward(self, observation):
-    """Compute the intrisic reward.
+    """Compute the intrinsic reward.
 
     Args:
       observation: raw observation in observation_spec format
@@ -91,7 +91,7 @@ def _get_intrinsic_reward(self, observation):
     """
     with tf.GradientTape() as tape:
       # make the predict network parameters trainable
-      # Compute the feature embedding loss (for next obseravtion trajectory)
+      # Compute the feature embedding loss (for next observation trajectory)
       feature_target, _ = self._target_net(observation)
       feature_predict, _ = self._predict_net(observation)
       feature_target = tf.stop_gradient(feature_target)

compiler_opt/rl/regalloc/regalloc_runner.py

@@ -48,7 +48,7 @@ def compile_fn(
 
     Args:
       command_line: the fully qualified command line.
-      tf_policy_path: path to TF policy direcoty on local disk.
+      tf_policy_path: path to TF policy directory on local disk.
       reward_only: whether only return reward.
 
     Returns:

compiler_opt/tools/feature_importance_graphs.py

@@ -46,7 +46,7 @@ def load_shap_values(file_name: str) -> DataType:
 
 
 def init_shap_for_notebook():
-  """Initalizes some JS code for interactive feature importance plots."""
+  """Initializes some JS code for interactive feature importance plots."""
   shap.initjs()
 
 

compiler_opt/tools/feature_importance_utils.py

@@ -89,7 +89,7 @@ def flatten_input(to_flatten: types.NestedTensorSpec,
   """Flattens problem instance data into a flat array for shap
 
   Args:
-    to_flatten: A nested tensor spec of data that needs to be flattend into
+    to_flatten: A nested tensor spec of data that needs to be flattened into
       an array
     array_size: An integer representing the size of the output array. Used for
       allocating the flat array to place all the data in.
@@ -167,7 +167,7 @@ def get_max_part_size(input_signature: SignatureType) -> int:
 
 def create_run_model_function(action_fn: Callable,
                               input_sig: SignatureType) -> Callable:
-  """Returns a function that takes in a flattend input array and returns the
+  """Returns a function that takes in a flattened input array and returns the
   model output as a scalar.
 
   Args:

docs/benchmarking.md

@@ -78,7 +78,7 @@ subdirectory within that repository.
 * `--llvm_test_suite_build_path` - The path to place the build for the
 llvm-test-suite. Similar behavior to the LLVM build path.
 * `llvm_use_incremental` - Whether or not to do an incremental build of LLVM.
-If you alread have all the correct compilation flags setup for running MLGO
+If you already have all the correct compilation flags setup for running MLGO
 with LLVM, you can set this flag and you should get an extremely fast LLVM
 build as the only thing changing is the release mode regalloc model.
 * `model_path` - The path to the regalloc model. If this is set to "download",

docs/inlining-demo/demo.md

@@ -186,7 +186,7 @@ yours). The `--args=clang_embed_bitcode=true` option above adds the compilation
 flag `-Xclang=-fembed-bitcode=all`. This can be seen in the compilation database.
 The effect of this is that the object files have the llvm bytecode produced by
 clang, before the optimization passes, and the clang command line, captured in
-the .llvmbc and .llvmcmd sections, respectivelly. This is the mechanism by which
+the .llvmbc and .llvmcmd sections, respectively. This is the mechanism by which
 we extract our corpus.
 
 Naturally, the effect of this is that the object files, and the linked binaries,
@@ -346,7 +346,7 @@ compiler_opt/tools/generate_default_trace.py \
 ## Deploying and using the new policy
 
 We need to build the 'release' mode of the compiler. Currently, that means
-overwritting the model in `llvm/lib/Analysis/models/inliner`.
+overwriting the model in `llvm/lib/Analysis/models/inliner`.
 
 ```shell
 cd $LLVM_SRCDIR

docs/regalloc-demo/demo.md

@@ -89,7 +89,7 @@ gclient runhooks
 ## Install Dependencies
 
 If you're working in a Debian based docker container, it will most likely
-not come by default with `sudo`. It isn't stricly necessary to install,
+not come by default with `sudo`. It isn't strictly necessary to install,
 but it makes it easier to copypaste the installation commands below and it
 also enables the use of the Chromium dependency auto-installation script:
 
@@ -110,7 +110,7 @@ $WORKING_DIR/chromium/src/build/install-build-deps.sh
 ```
 
 **Note:** These installation commands are all designed to be run on Debian
-based distros. However, adapating to other distros with alternative package
+based distros. However, adapting to other distros with alternative package
 management systems should not be too difficult. The packages for the first
 command should be very similarly named and the
 [official Chromium documentation](https://chromium.googlesource.com/chromium/src/+/main/docs/linux/build_instructions.md)
@@ -176,7 +176,7 @@ symbol_level=0
 enable_nacl=false
 ```
 
-Immedaitely after closing the editor, `gn` will generate all of the files
+Immediately after closing the editor, `gn` will generate all of the files
 necessary so that `ninja` can execute all the necessary compilation steps.
 However, to extract a corpus for ML training, we also need a database of
 compilation commands. This can be obtained by running the following command: