rocm_jax/jax/lib/xla_bridge.py

# Copyright 2018 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
#
#     https://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.

"""Interface and utility functions to XLA.

This module wraps the XLA client(s) and builders to standardize their interfaces
and provide some automatic type mapping logic for converting between Numpy and
XLA. There are also a handful of related casting utilities.
"""


from functools import partial, lru_cache
import os
from typing import Callable, Dict, List, Optional, Tuple, Union
import warnings

from absl import logging
# Disable "WARNING: Logging before flag parsing goes to stderr." message
logging._warn_preinit_stderr = 0

import jax.lib
from .._src.config import flags, bool_env
from . import tpu_driver_client
from . import xla_client
from jax._src import util, traceback_util
from jax._src import dtypes
import numpy as np
import threading

traceback_util.register_exclusion(__file__)


xops = xla_client.ops

FLAGS = flags.FLAGS

# TODO(phawkins): Remove jax_xla_backend.
flags.DEFINE_string(
    'jax_xla_backend', '',
    'jax_xla_backend is an alias for jax_platform_name. If both are '
    'provided, --jax_xla_backend takes priority. Prefer --jax_platform_name.')
flags.DEFINE_string(
    'jax_backend_target', 'local',
    'Either "local" or "rpc:address" to connect to a remote service target. '
    'The default is "local".')
flags.DEFINE_string(
    'jax_platform_name',
    os.getenv('JAX_PLATFORM_NAME', '').lower(),
    'Platform name for XLA. The default is to attempt to use a GPU or TPU if '
    'available, but fall back to CPU otherwise. To set the platform manually, '
    'pass "cpu" for CPU, "gpu" for GPU, etc. If intending to use CPU, '
    'setting the platform name to "cpu" can silence warnings that appear with '
    'the default setting.')
flags.DEFINE_bool(
    'jax_disable_most_optimizations',
    bool_env('JAX_DISABLE_MOST_OPTIMIZATIONS', False),
    'Try not to do much optimization work. This can be useful if the cost of '
    'optimization is greater than that of running a less-optimized program.')
flags.DEFINE_string(
    'jax_cpu_backend_variant',
     os.getenv('JAX_CPU_BACKEND_VARIANT', 'tfrt'),
    'Selects CPU backend runtime variant: "stream_executor" or "tfrt". The '
    'default is "tfrt".')

def get_compile_options(
    num_replicas: int,
    num_partitions: int,
    device_assignment=None,
    use_spmd_partitioning: bool = True,
) -> xla_client.CompileOptions:
  """Returns the compile options to use, as derived from flag values.

  Args:
    num_replicas: Number of replicas for which to compile.
    num_partitions: Number of partitions for which to compile.
    device_assignment: Optional tuple of integers indicating the assignment of
      logical replicas to physical devices (default inherited from
      xla_client.CompileOptions). Must be consistent with `num_replicas` and
      `num_partitions`.
    use_spmd_partitioning: boolean indicating whether to enable SPMD or MPMD
      partitioning in XLA.
  """
  compile_options = xla_client.CompileOptions()
  compile_options.num_replicas = num_replicas
  compile_options.num_partitions = num_partitions
  build_options = compile_options.executable_build_options
  build_options.use_spmd_partitioning = use_spmd_partitioning
  if device_assignment is not None:
    logging.vlog(
        2,
        'get_compile_options: num_replicas=%s num_partitions=%s device_assignment=%s',
        num_replicas, num_partitions, device_assignment)
    device_assignment = np.array(device_assignment)

    # Allow 1D device assignment if num_partitions is 1.
    if (device_assignment.ndim == 1) and (num_partitions == 1):
      device_assignment = device_assignment[:, None]

    if num_replicas != device_assignment.shape[0]:
      msg = 'device_assignment does not match num_replicas: {} vs {}.'
      raise ValueError(msg.format(device_assignment, num_replicas))

    if num_partitions != device_assignment.shape[1]:
      msg = 'device_assignment does not match num_partitions: {} vs {}.'
      raise ValueError(msg.format(device_assignment, num_partitions))

    device_assignment = xla_client.DeviceAssignment.create(device_assignment)
    assert device_assignment.replica_count() == num_replicas
    assert device_assignment.computation_count() == num_partitions
    compile_options.device_assignment = device_assignment

  debug_options = compile_options.executable_build_options.debug_options
  if jax.lib.cuda_path is not None:
    debug_options.xla_gpu_cuda_data_dir = jax.lib.cuda_path

  if FLAGS.jax_disable_most_optimizations:

    debug_options.xla_backend_optimization_level = 0
    debug_options.xla_llvm_disable_expensive_passes = True
    debug_options.xla_test_all_input_layouts = False

  return compile_options


# Backends

def _make_tpu_driver_client():
  if tpu_driver_client is None:
    logging.info("Remote TPU is not linked into jax; skipping remote TPU.")
    return None
  if FLAGS.jax_backend_target is None:
    logging.info("No --jax_backend_target was provided; skipping remote TPU.")
    return None
  return tpu_driver_client.TpuBackend.create(worker=FLAGS.jax_backend_target)


# Backends, in increasing order of preference.
# We have no particular opinion about how "backends" relate to "devices". For
# example, there could be multiple backends that provide the same kind of
# device.
_backend_factories = {}

def register_backend_factory(name, factory, *, priority=0):
  _backend_factories[name] = (factory, priority)


register_backend_factory('interpreter', xla_client.make_interpreter_client,
                         priority=-100)
if FLAGS.jax_cpu_backend_variant == 'stream_executor':
  register_backend_factory('cpu',
                           partial(xla_client.make_cpu_client, use_tfrt=False),
                           priority=0)
else:
  assert FLAGS.jax_cpu_backend_variant == 'tfrt'
  register_backend_factory('cpu',
                           partial(xla_client.make_cpu_client, use_tfrt=True),
                           priority=0)
register_backend_factory('tpu_driver', _make_tpu_driver_client,
                         priority=100)
register_backend_factory('gpu', xla_client.make_gpu_client,
                         priority=200)
register_backend_factory('tpu', xla_client.make_tpu_client,
                         priority=300)

_default_backend = None
_backends = None
_backend_lock = threading.Lock()


def backends():
  global _backends
  global _default_backend

  with _backend_lock:
    if _backends is not None:
      return _backends

    default_priority = -1000
    _backends = {}
    for name, (factory, priority) in _backend_factories.items():
      logging.vlog(1, "Initializing backend '%s'" % name)
      try:
        backend = factory()
        if backend is not None:
          if backend.device_count() > 0:
            _backends[name] = backend
          util.distributed_debug_log(("Initialized backend", backend.platform),
                                     ("process_index", backend.process_index()),
                                     ("device_count", backend.device_count()),
                                     ("local_devices", backend.local_devices()))
          logging.vlog(1, "Backend '%s' initialized" % name)
          if priority > default_priority:
            _default_backend = backend
            default_priority = priority
      except Exception as err:
        if name in ('cpu', 'interpreter'):
          # We always expect the CPU and interpreter backends to initialize
          # successfully.
          raise
        else:
          # If the backend isn't built into the binary, or if it has no devices,
          # we expect a RuntimeError.
          logging.info("Unable to initialize backend '%s': %s" % (name, err))
          continue
    if _default_backend.platform == "cpu" and FLAGS.jax_platform_name != 'cpu':
      logging.warning('No GPU/TPU found, falling back to CPU. '
                      '(Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)')
    return _backends


@lru_cache(maxsize=None)  # don't use util.memoize because there is no X64 dependence.
def get_backend(platform=None):
  # TODO(mattjj,skyewm): remove this input polymorphism after we clean up how
  # 'backend' values are handled
  if not isinstance(platform, (type(None), str)):
    return platform

  bs = backends()
  platform = (platform or FLAGS.jax_xla_backend or FLAGS.jax_platform_name
              or None)
  if platform is not None:
    backend = bs.get(platform, None)
    if backend is None:
      raise RuntimeError(f"Unknown backend {platform}")
    return backend
  else:
    return _default_backend


def get_device_backend(device=None):
  """Returns the Backend associated with `device`, or the default Backend."""
  if device is not None:
    return device.client
  return get_backend()


def device_count(backend: Optional[str] = None) -> int:
  """Returns the total number of devices.

  On most platforms, this is the same as :py:func:`jax.local_device_count`.
  However, on multi-process platforms where different devices are associated
  with different processes, this will return the total number of devices across
  all processes.

  Args:
    backend: This is an experimental feature and the API is likely to change.
      Optional, a string representing the xla backend: ``'cpu'``, ``'gpu'``, or
      ``'tpu'``.

  Returns:
    Number of devices.

  """
  return int(get_backend(backend).device_count())


def local_device_count(backend: Optional[str] = None) -> int:
  """Returns the number of devices addressable by this process."""
  return int(get_backend(backend).local_device_count())


def devices(backend: Optional[str] = None) -> List[xla_client.Device]:
  """Returns a list of all devices for a given backend.

  Each device is represented by a subclass of :class:`Device` (e.g.
  :class:`CpuDevice`, :class:`GpuDevice`). The length of the returned list is
  equal to ``device_count(backend)``. Local devices can be identified by
  comparing :meth:`Device.process_index` to the value returned by
  :py:func:`jax.process_index`.

  If ``backend`` is ``None``, returns all the devices from the default backend.
  The default backend is generally ``'gpu'`` or ``'tpu'`` if available,
  otherwise ``'cpu'``.

  Args:
    backend: This is an experimental feature and the API is likely to change.
      Optional, a string representing the xla backend: ``'cpu'``, ``'gpu'``, or
      ``'tpu'``.

  Returns:
    List of Device subclasses.
  """
  return get_backend(backend).devices()


def default_backend() -> str:
  """Returns the platform name of the default XLA backend."""
  return get_backend(None).platform


def local_devices(process_index: Optional[int] = None,
                  backend: Optional[str] = None,
                  host_id: Optional[int] = None) -> List[xla_client.Device]:
  """Like :py:func:`jax.devices`, but only returns devices local to a given process.

  If ``process_index`` is ``None``, returns devices local to this process.

  Args:
    process_index: the integer index of the process. Process indices can be
      retrieved via ``len(jax.process_count())``.
    backend: This is an experimental feature and the API is likely to change.
      Optional, a string representing the xla backend: ``'cpu'``, ``'gpu'``, or
      ``'tpu'``.

  Returns:
    List of Device subclasses.
  """
  if host_id is not None:
    warnings.warn(
        "The argument to jax.local_devices has been renamed from `host_id` to "
        "`process_index`. This alias will eventually be removed; please update "
        "your code.")
    process_index = host_id
  if process_index is None:
    process_index = get_backend(backend).process_index()
  if not (0 <= process_index < process_count()):
    raise ValueError(f"Unknown process_index {process_index}")
  return [d for d in devices(backend) if d.process_index == process_index]


def process_index(backend: Optional[str] = None) -> int:
  """Returns the integer process index of this process.

  On most platforms, this will always be 0. This will vary on multi-process
  platforms though.

  Args:
    backend: This is an experimental feature and the API is likely to change.
      Optional, a string representing the xla backend: ``'cpu'``, ``'gpu'``, or
      ``'tpu'``.

  Returns:
    Integer process index.
  """
  return get_backend(backend).process_index()


# TODO: remove this sometime after jax 0.2.13 is released
def host_id(backend=None):
  warnings.warn(
      "jax.host_id has been renamed to jax.process_index. This alias "
      "will eventually be removed; please update your code.")
  return process_index(backend)


def process_count(backend: Optional[str] = None) -> int:
  """Returns the number of JAX processes associated with the backend."""
  return max(d.process_index for d in devices(backend)) + 1


# TODO: remove this sometime after jax 0.2.13 is released
def host_count(backend=None):
  warnings.warn(
      "jax.host_count has been renamed to jax.process_count. This alias "
      "will eventually be removed; please update your code.")
  return process_count(backend)


# TODO: remove this sometime after jax 0.2.13 is released
def host_ids(backend=None):
  warnings.warn(
      "jax.host_ids has been deprecated; please use range(jax.process_count()) "
      "instead. jax.host_ids will eventually be removed; please update your "
      "code.")
  return list(range(process_count(backend)))


### utility functions

@util.memoize
def dtype_to_etype(dtype):
  """Convert from dtype to canonical etype (reading config.x64_enabled)."""
  return xla_client.dtype_to_etype(dtypes.canonicalize_dtype(dtype))


@util.memoize
def supported_numpy_dtypes():
  return {dtypes.canonicalize_dtype(dtype)
          for dtype in xla_client.XLA_ELEMENT_TYPE_TO_DTYPE.values()}


# TODO(mattjj,frostig): try to remove this function
def normalize_to_xla_dtypes(val):
  """Normalize dtypes in a value."""
  if hasattr(val, '__array__') or np.isscalar(val):
    return np.asarray(val, dtype=dtypes.canonicalize_dtype(dtypes.result_type(val)))
  elif isinstance(val, (tuple, list)):
    return tuple(normalize_to_xla_dtypes(x) for x in val)
  raise TypeError('Can\'t convert to XLA: {}'.format(val))

def _numpy_array_constant(builder, value, canonicalize_types=True):
  if canonicalize_types:
    value = normalize_to_xla_dtypes(value)
  return [xops.ConstantLiteral(builder, value)]

def parameter(builder, num, shape, name=None, replicated=None):
  if name is None:
    name = ''
  if replicated is None:
    replicated = []
  elif isinstance(replicated, bool):
    replicated = [replicated] * shape.leaf_count()

  return xops.Parameter(builder, num,
                        shape.with_major_to_minor_layout_if_absent(), name,
                        replicated)


def constant_general(builder, py_val, canonicalize_types=True):
  """Translate a general constant `py_val` to a constant, canonicalizing its dtype.

  Args:
    py_val: a Python value to be translated to a constant.

  Returns:
    A representation of the constant as a list of xla ops.
  """
  for t in type(py_val).mro():
    handler = _constant_handlers.get(t)
    if handler: return handler(builder, py_val, canonicalize_types)
  if hasattr(py_val, '__jax_array__'):
    return constant(builder, py_val.__jax_array__(), canonicalize_types)
  raise TypeError("No constant handler for type: {}".format(type(py_val)))

def constant(builder, py_val, canonicalize_types=True):
  """Translate constant `py_val` to a constant, canonicalizing its dtype.

  Args:
    py_val: a Python value to be translated to a constant.

  Returns:
    A representation of the constant, either a ComputationDataHandle or None
  """
  const = constant_general(builder, py_val, canonicalize_types=canonicalize_types)
  assert len(const) == 1, f"Internal error: cannot create constant from object of type {type(py_val)}"
  return const[0]

# HLO instructions optionally can be annotated to say how the output should be
# spatially partitioned (represented in XLA as OpSharding protos, see
# _sharding_to_proto). For array outputs, the annotation is either an int per
# dimension specifying the number of ways that dimension divided (i.e. the total
# number of shards is the product), or None to indicate the array should be
# replicated. Tuple outputs are represented as tuples thereof. XLA supports
# arbitrary tuple nesting, but JAX only uses one level of tupling (and our type
# checkers don't support recursive types), so we only represent one level of
# nesting in this type definition.
SpatialSharding = Union[Tuple[int, ...],
                        None,
                        Tuple[Union[Tuple[int, ...], None], ...]]

def _sharding_to_proto(sharding: SpatialSharding):
  """Converts a SpatialSharding to an OpSharding.

  See
  https://github.com/tensorflow/tensorflow/blob/main/tensorflow/compiler/xla/xla_data.proto#L601
  for details on the OpSharding proto.
  """
  proto = xla_client.OpSharding()
  if isinstance(sharding, tuple) and not isinstance(sharding[0], int):
    assert all(s is None or isinstance(s, tuple) for s in sharding)
    return tuple_sharding_proto(list(map(_sharding_to_proto, sharding)))  # type: ignore

  if sharding is None:
    proto.type = xla_client.OpSharding.Type.REPLICATED
  else:
    proto.type = xla_client.OpSharding.Type.OTHER
    proto.tile_assignment_dimensions = list(sharding)
    proto.tile_assignment_devices = list(range(np.product(sharding)))
  return proto

def tuple_sharding_proto(elems):
  proto = xla_client.OpSharding()
  assert all(isinstance(e, type(proto)) for e in elems)
  proto.type = xla_client.OpSharding.Type.TUPLE
  proto.tuple_shardings = elems
  return proto

def set_sharding_proto(builder, op, sharding_proto):
  """Uses CustomCall to annotate a value as sharded."""
  # "Sharding" is a built-in custom call target that acts like an identity
  # function, and is used to attach an OpSharding to.
  return with_sharding_proto(builder, sharding_proto, xops.CustomCall,
                             builder, b"Sharding", [op], builder.get_shape(op))

def with_sharding_proto(builder, sharding_proto, op_fn, *args, **kwargs):
  """Builds op_fn(*args, **kwargs) with sharding annotation."""
  builder.set_sharding(sharding_proto)
  try:
    return op_fn(*args, **kwargs)
  finally:
    builder.clear_sharding()

def set_sharding(builder, op, sharding: SpatialSharding):
  """Uses CustomCall to annotate a value as sharded."""
  return set_sharding_proto(builder, op, _sharding_to_proto(sharding))

def with_sharding(builder, sharding: SpatialSharding, op_fn, *args, **kwargs):
  """Builds op_fn(*args, **kwargs) with sharding annotation."""
  return with_sharding_proto(builder, _sharding_to_proto(sharding), op_fn, *args, **kwargs)

def make_computation_builder(name):
  return xla_client.XlaBuilder(name)


def register_constant_handler(type_, handler_fun):
  _constant_handlers[type_] = handler_fun
_constant_handlers: Dict[type, Callable] = {}


def _ndarray_constant_handler(c, val, canonicalize_types=True):
  """Constant handler for ndarray literals, handling zero-size strides.

  This function essentially calls _numpy_array_constant(val) except it has
  special handling of arrays with any strides of size zero: for those, it
  generates appropriate calls to NumpyArrayConstant, Broadcast, and Transpose
  to avoid staging in large literals that might arise from np.zeros or np.ones
  or the output of lax.broadcast (which uses np.broadcast_to which in turn
  uses size-zero strides).

  Args:
    c: an XlaBuilder
    val: an ndarray.

  Returns:
    An XLA ComputationDataHandle / XlaOp representing the constant ndarray
    staged into the XLA Computation.
  """
  # TODO(mattjj): revise this to use xops.BroadcastInDim rather than Transpose
  if dtypes.result_type(val) == dtypes.float0:
    return _numpy_array_constant(c, np.zeros(val.shape, dtype=np.bool_))
  elif np.any(np.equal(0, val.strides)) and val.size > 0:
    zero_stride_axes, = np.where(np.equal(0, val.strides))
    other_axes, = np.where(np.not_equal(0, val.strides))
    collapsed_val = val[tuple(0 if ax in zero_stride_axes else slice(None)
                              for ax in range(val.ndim))]
    xla_val = xops.Broadcast(
        _numpy_array_constant(c, collapsed_val, canonicalize_types)[0],
        np.take(val.shape, zero_stride_axes))
    permutation = np.argsort(tuple(zero_stride_axes) + tuple(other_axes))
    return [xops.Transpose(xla_val, permutation)]
  else:
    return _numpy_array_constant(c, val, canonicalize_types)
register_constant_handler(np.ndarray, _ndarray_constant_handler)


def _scalar_constant_handler(c, val, canonicalize_types=True):
  return _numpy_array_constant(c, val, canonicalize_types)

for scalar_type in [np.int8, np.int16, np.int32, np.int64,
                    np.uint8, np.uint16, np.uint32, np.uint64,
                    np.float16, np.float32, np.float64,
                    np.bool_, np.longlong,
                    xla_client.bfloat16]:
  register_constant_handler(scalar_type, _scalar_constant_handler)

# https://github.com/winpython/winpython/issues/613#issuecomment-380121523
if hasattr(np, "float128"):
  register_constant_handler(np.float128, _scalar_constant_handler)

def _python_scalar_handler(dtype, c, val, canonicalize_dtypes=True):
  return _numpy_array_constant(c, dtype.type(val))

for ptype, dtype in dtypes.python_scalar_dtypes.items():
  register_constant_handler(ptype, partial(_python_scalar_handler, dtype))