XLA: Optimizing Compiler for Machine Learning

XLA (Accelerated Linear Algebra) is a domain-specific compiler for linear algebra that can accelerate TensorFlow models with potentially no source code changes.

The results are improvements in speed and memory usage: most internal benchmarks run ~1.15x faster after XLA is enabled. The dataset below is evaluated on a single NVidia V100 GPU:

Introduction

When a TensorFlow program is run, all of the operations are executed individually by the TensorFlow executor. Each TensorFlow operation has a precompiled GPU kernel implementation that the executor dispatches to.

XLA provides an alternative mode of running models: it compiles the TensorFlow graph into a sequence of computation kernels generated specifically for the given model. Because these kernels are unique to the model, they can exploit model-specific information for optimization. For example, let's look at an optimization XLA does in the context of a simple TensorFlow computation:

def model_fn(x, y, z):
  return tf.reduce_sum(x + y * z)

Run without XLA, the graph launches three kernels: one for the multiplication, one for the addition and one for the reduction. However, XLA can optimize the graph so that it computes the result in a single kernel launch. It does this by "fusing" the addition, multiplication and reduction into a single GPU kernel. Moreover, this fused operation does not write out the intermediate values produced by y*z and x+y*z to memory; instead it "streams" the results of these intermediate computations directly to their users while keeping them entirely in GPU registers. Fusion is XLA's single most important optimization. Memory bandwidth is typically the scarcest resource on hardware accelerators, so removing memory operations is one of the best ways to improve performance.

Enable XLA for TensorFlow models

Auto-clustering

A simplest way to start using XLA in TensorFlow models is to enable auto-clustering, which automatically finds clusters (connected subgraphs) within the TensorFlow graph which can be compiled and executed using XLA. Auto-clustering on GPU can be enabled by setting the TF_XLA_FLAGS environment variable:

$ TF_XLA_FLAGS=--tf_xla_auto_jit=2 path/to/your/tf/program

Auto-clustering is currently optimized for GPU workloads, but it can also be enabled on CPU by additionally using the flag --tf_xla_cpu_global_jit:

$ TF_XLA_FLAGS="--tf_xla_auto_jit=2 --tf_xla_cpu_global_jit" path/to/your/program

For a detailed usage example see the auto-clustering tutorial colab.

Explicit compilation with tf.function

Auto-clustering is a great tool for making the model faster without any changes to the code, but it may be hard to understand what changes have been performed.

Explicit compilation API offers a more fine-grained control for choosing which functions should be compiled. For example, the following TensorFlow function which performs the MNIST training is compiled with XLA:

@tf.function(experimental_compile=True)
def train_mnist(images, labels):
    images, labels = cast(images, labels)

    with tf.GradientTape() as tape:
      predicted_labels = layer(images)
      loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
          logits=predicted_labels, labels=labels
      ))
    layer_variables = layer.trainable_variables
    grads = tape.gradient(loss, layer_variables)
    optimizer.apply_gradients(zip(grads, layer_variables))

The experimental_compile API has must-compile semantics: either the entire function is compiled with XLA, or an errors.InvalidArgumentError exception is thrown. XLA can not currently compile functions where dimensions are not inferrable: that is, if it's not possible to infer the dimensions of all tensors without running the entire computation. For example, the following function will not compile:

@tf.function
def not_compilable(x):
  return tf.unique(x)

Shapes can vary across the runs though:

@tf.function(experimental_compile=True)
def recompiled_on_launch(a, b):
  return a + b

recompiled_on_launch(tf.ones([1, 10]), tf.ones([1, 10]))
recompiled_on_launch(tf.ones([1, 100]), tf.ones([1, 100]))

See the tutorial colab for a more detailed usage example.

AOT (Ahead-of-time) compilation for CPU with `tfcompile`

You can also use a standalone tfcompile tool, which converts TensorFlow graph into executable code (for x86-64 CPU only).

Inspect compiled programs

XLA provides introspection facilities which let you inspect the generated programs. To dump the generated programs, use the environment variable XLA_FLAGS:

$ XLA_FLAGS="--xla_dump_to=/tmp/generated" TF_XLA_FLAGS="--tf_xla_auto_jit=2" my/tensorflow/program

After the dumping is performed, you can find the following files in /tmp/generated:

module_XXXX.*_optimizations.txt Generated XLA programs, one per each compiled cluster. Attaching those when submitting XLA bug reports is extremely helpful!
module_XXXX.ir-*.ll Generated files in LLVM intermediate representation, with NVPTX intrinsics.
module_XXXX.ptx Generated PTX files.

You can also dump the graph visualizing the embedding of XLA clusters inside of the TensorFlow graph with:

$ TF_DUMP_GRAPH_PREFIX=/tmp/generated TF_XLA_FLAGS="--tf_xla_clustering_debug"

Reproducible bug reports

A bug report is much easier to reproduce if it includes dumps for the generated XLA programs and the used auto-clustering embedding. To generate them for a TensorFlow program running with auto-clustering, launch:

$ TF_DUMP_GRAPH_PREFIX=/tmp/generated \
  TF_XLA_FLAGS="--tf_xla_clustering_debug --tf_xla_auto_jit=2" \
  XLA_FLAGS="--xla_dump_hlo_as_text --xla_dump_to=/tmp/generated" \
    my/tensorflow/program"

When filing bugs, attach the contents of the /tmp/generated directory (referenced above).

If possible, try to isolate a bug to a single XLA program by using the replay_computation and iteratively running it on generated programs.

Known Issues

Compilation with XLA can greatly improve the performance of your programs, but the TensorFlow interop has a number of known sharp corners.

TensorArray TF/XLA Interconversion

The problem manifests itself as an error message Support for TensorList crossing the XLA/TF boundary is not implemented.

XLA supports tf.TensorArray. However, the interconversion between TF and XLA representations is not implemented yet. This error often arises when the TensorArray is used inside the compiled block, but the derivative is taken outside.

Workaround: compile the outermost scope which is taking the derivative.

Random Number Generation

XLA currently ignores TF seeds to random operations. This affects stateful TF random operations, such as tf.random.normal, or tf.nn.dropout. XLA will behave as if the compilation was seeded with a new unique seed at each run. This limitation does not apply to stateless random ops.

XLA Frontends

Apart from TensorFlow, XLA programs can be generated by:

JAX: Composable transformations of Python+NumPy programs
Julia: The Julia language for scientific computing
PyTorch: PyTorch framework