Source code for jax._src.lax.control_flow

# coding=utf-8
# Copyright 2019 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.
"""
Control flow primitives.
"""


import collections
import functools
from functools import partial
import inspect
import itertools
import operator
import os
from typing import Any, Callable, Optional, Sequence, Tuple, TypeVar

import numpy as np

import jax
from jax._src import api
from jax import core
from jax._src import dtypes
from jax._src import source_info_util
from jax._src import util
from jax._src.lax import lax
from jax._src.lax import slicing
from jax._src.lax import windowed_reductions
from jax import linear_util as lu
from jax.core import ConcreteArray, ShapedArray, raise_to_shaped
from jax._src.api_util import flatten_fun_nokwargs
from jax.interpreters import ad
from jax.interpreters import partial_eval as pe
from jax.interpreters import mlir
from jax.interpreters import xla
from jax.interpreters import batching
from jax.interpreters import masking
from jax._src.lib.mlir import ir
from jax._src.lib.mlir.dialects import mhlo
from jax._src.lib import xla_client
from jax._src.traceback_util import api_boundary
from jax._src.util import (unzip2, unzip3, safe_map, safe_zip,
                           split_list, cache, extend_name_stack)
from jax.tree_util import (tree_flatten, tree_unflatten, treedef_is_leaf,
                           treedef_children, treedef_tuple, tree_multimap,
                           tree_leaves, tree_structure)
from jax._src import ad_util
from jax.config import config

xops = xla_client.ops

_map = safe_map
zip = safe_zip
_reduce = functools.reduce

T = TypeVar('T')
Array = Any

@cache()
def _initial_style_open_jaxpr(fun: Callable, in_tree, in_avals,
                              primitive_name: Optional[str] = None):
  wrapped_fun, out_tree = flatten_fun_nokwargs(lu.wrap_init(fun), in_tree)
  debug = pe.debug_info(fun, in_tree, False, primitive_name or "<unknown>")
  jaxpr, _, consts = pe.trace_to_jaxpr_dynamic(wrapped_fun, in_avals, debug)
  return jaxpr, consts, out_tree()

@cache()
def _initial_style_jaxpr(fun: Callable, in_tree, in_avals,
                         primitive_name: Optional[str] = None):
  jaxpr, consts, out_tree = _initial_style_open_jaxpr(
      fun, in_tree, in_avals, primitive_name)
  closed_jaxpr = core.ClosedJaxpr(pe.convert_constvars_jaxpr(jaxpr), ())
  return closed_jaxpr, consts, out_tree

@cache()
def _initial_style_jaxprs_with_common_consts(
    funs: Sequence[Callable], in_tree, in_avals, primitive_name: str):
  # When staging the branches of a conditional into jaxprs, constants are
  # extracted from each branch and converted to jaxpr arguments. To use the
  # staged jaxprs as the branches to a conditional *primitive*, we need for
  # their (input) signatures to match. This function "joins" the staged jaxprs:
  # for each one, it makes another that accepts *all* constants, but only uses
  # those that it needs (dropping the rest).

  jaxprs, all_consts, all_out_trees = \
      unzip3(_initial_style_open_jaxpr(fun, in_tree, in_avals, primitive_name)
             for fun in funs)

  newvar = core.gensym(jaxprs, suffix='_')
  all_const_avals = [[raise_to_shaped(core.get_aval(c)) for c in consts]
                     for consts in all_consts]
  unused_const_vars = [[newvar(aval) for aval in const_avals]
                       for const_avals in all_const_avals]

  def pad_jaxpr_constvars(i, jaxpr):
    prefix = util.concatenate(unused_const_vars[:i])
    suffix = util.concatenate(unused_const_vars[i + 1:])
    constvars = [*prefix, *jaxpr.constvars, *suffix]
    return core.Jaxpr(constvars=constvars, invars=jaxpr.invars,
                      outvars=jaxpr.outvars, eqns=jaxpr.eqns)

  consts = util.concatenate(all_consts)
  jaxprs = [pad_jaxpr_constvars(i, jaxpr) for i, jaxpr in enumerate(jaxprs)]
  closed_jaxprs = [core.ClosedJaxpr(pe.convert_constvars_jaxpr(jaxpr), ())
                   for jaxpr in jaxprs]
  return closed_jaxprs, consts, all_out_trees

def _abstractify(x):
  return raise_to_shaped(core.get_aval(x))

def _typecheck_param(prim, param, name, msg_required, pred):
  msg = (f'invalid {prim} param {name} of type {type(param).__name__}, '
         f'{msg_required} required:')
  param_str = str(param)
  sep = os.linesep if os.linesep in param_str else ' '
  msg = sep.join([msg, param_str])
  core.typecheck_assert(pred, msg)


### fori_loop and while_loop

def _fori_cond_fun(loop_carry):
  i, upper, _ = loop_carry
  return lax.lt(i, upper)

@cache()
def _fori_body_fun(body_fun):
  def while_body_fun(loop_carry):
    i, upper, x = loop_carry
    return lax.add(i, lax._const(i, 1)), upper, body_fun(i, x)
  return while_body_fun

@cache()
def _fori_scan_body_fun(body_fun):
  def scanned_fun(loop_carry, _):
    i, x = loop_carry
    return (i + 1, body_fun(i, x)), None
  return scanned_fun

[docs]@api_boundary def fori_loop(lower, upper, body_fun, init_val): """Loop from ``lower`` to ``upper`` by reduction to :func:`jax.lax.while_loop`. The type signature in brief is .. code-block:: haskell fori_loop :: Int -> Int -> ((Int, a) -> a) -> a -> a The semantics of ``fori_loop`` are given by this Python implementation:: def fori_loop(lower, upper, body_fun, init_val): val = init_val for i in range(lower, upper): val = body_fun(i, val) return val Unlike that Python version, ``fori_loop`` is implemented in terms of either a call to :func:`jax.lax.while_loop` or a call to :func:`jax.lax.scan`. If the trip count is static (meaning known at tracing time, perhaps because ``lower`` and ``upper`` are Python integer literals) then the ``fori_loop`` is implemented in terms of ``scan`` and reverse-mode autodiff is supported; otherwise, a ``while_loop`` is used and reverse-mode autodiff is not supported. See those functions' docstrings for more information. Also unlike the Python analogue, the loop-carried value ``val`` must hold a fixed shape and dtype across all iterations (and not just be consistent up to NumPy rank/shape broadcasting and dtype promotion rules, for example). In other words, the type ``a`` in the type signature above represents an array with a fixed shape and dtype (or a nested tuple/list/dict container data structure with a fixed structure and arrays with fixed shape and dtype at the leaves). Args: lower: an integer representing the loop index lower bound (inclusive) upper: an integer representing the loop index upper bound (exclusive) body_fun: function of type ``(int, a) -> a``. init_val: initial loop carry value of type ``a``. Returns: Loop value from the final iteration, of type ``a``. """ # TODO(phawkins): perhaps do more type checking here, better error messages. lower_dtype = dtypes.canonicalize_dtype(lax.dtype(lower)) upper_dtype = dtypes.canonicalize_dtype(lax.dtype(upper)) if lower_dtype != upper_dtype: msg = ("lower and upper arguments to fori_loop must have equal types, " "got {} and {}") raise TypeError(msg.format(lower_dtype.name, upper_dtype.name)) # If we can specialize on the trip count, call scan instead of a while_loop # to enable efficient reverse-mode differentiation. if (isinstance(core.get_aval(lower), ConcreteArray) and isinstance(core.get_aval(upper), ConcreteArray)): try: lower_ = int(lower) upper_ = int(upper) except TypeError: use_scan = False else: use_scan = True else: use_scan = False if use_scan: if config.jax_disable_jit and upper_ == lower_: # non-jit implementation of scan does not support length=0 return init_val (_, result), _ = scan(_fori_scan_body_fun(body_fun), (lower_, init_val), None, length=upper_ - lower_) else: _, _, result = while_loop(_fori_cond_fun, _fori_body_fun(body_fun), (lower, upper, init_val)) return result
[docs]@api_boundary def while_loop(cond_fun: Callable[[T], bool], body_fun: Callable[[T], T], init_val: T) -> T: """Call ``body_fun`` repeatedly in a loop while ``cond_fun`` is True. The type signature in brief is .. code-block:: haskell while_loop :: (a -> Bool) -> (a -> a) -> a -> a The semantics of ``while_loop`` are given by this Python implementation:: def while_loop(cond_fun, body_fun, init_val): val = init_val while cond_fun(val): val = body_fun(val) return val Unlike that Python version, ``while_loop`` is a JAX primitive and is lowered to a single XLA While HLO. That makes it useful for reducing compilation times for jit-compiled functions, since native Python loop constructs in an ``@jit`` function are unrolled, leading to large XLA computations. Also unlike the Python analogue, the loop-carried value ``val`` must hold a fixed shape and dtype across all iterations (and not just be consistent up to NumPy rank/shape broadcasting and dtype promotion rules, for example). In other words, the type ``a`` in the type signature above represents an array with a fixed shape and dtype (or a nested tuple/list/dict container data structure with a fixed structure and arrays with fixed shape and dtype at the leaves). Another difference from using Python-native loop constructs is that ``while_loop`` is not reverse-mode differentiable because XLA computations require static bounds on memory requirements. Args: cond_fun: function of type ``a -> Bool``. body_fun: function of type ``a -> a``. init_val: value of type ``a``, a type that can be a scalar, array, or any pytree (nested Python tuple/list/dict) thereof, representing the initial loop carry value. Returns: The output from the final iteration of body_fun, of type ``a``. """ if config.jax_disable_jit: try: val = init_val while cond_fun(val): val = body_fun(val) return val except core.ConcretizationTypeError: # Can't run this while_loop in Python (e.g. because there's a vmap # transformation on it), so we fall back to the primitive version. pass def _create_jaxpr(init_val): init_vals, in_tree = tree_flatten((init_val,)) init_avals = tuple(_map(_abstractify, init_vals)) cond_jaxpr, cond_consts, cond_tree = _initial_style_jaxpr( cond_fun, in_tree, init_avals, "while_cond") body_jaxpr, body_consts, body_tree = _initial_style_jaxpr( body_fun, in_tree, init_avals, "while_loop") if not treedef_is_leaf(cond_tree) or len(cond_jaxpr.out_avals) != 1: msg = "cond_fun must return a boolean scalar, but got pytree {}." raise TypeError(msg.format(cond_tree)) pred_aval = cond_jaxpr.out_avals[0] if (not isinstance(pred_aval, ShapedArray) or pred_aval.strip_weak_type().strip_named_shape() != ShapedArray((), np.bool_)): msg = "cond_fun must return a boolean scalar, but got output type(s) {}." raise TypeError(msg.format(cond_jaxpr.out_avals)) return init_vals, init_avals, body_jaxpr, in_tree, cond_jaxpr, cond_consts, body_consts, body_tree # The body input and output avals must match exactly. However, we want to account for # the case when init contains weakly-typed values (e.g. Python scalars), with avals that # may not match the output despite being compatible by virtue of their weak type. # To do this, we compute the jaxpr in two passes: first with the raw inputs, and if # necessary, a second time with modified init values. init_vals, init_avals, body_jaxpr, in_tree, *rest = _create_jaxpr(init_val) new_init_vals, changed = _promote_weak_typed_inputs(init_vals, init_avals, body_jaxpr.out_avals) if changed: new_init_val, = tree_unflatten(in_tree, new_init_vals) init_vals, init_avals, body_jaxpr, in_tree, *rest = _create_jaxpr(new_init_val) cond_jaxpr, cond_consts, body_consts, body_tree = rest in_tree_children = in_tree.children() assert len(in_tree_children) == 1 _check_tree_and_avals("body_fun output and input", body_tree, body_jaxpr.out_avals, in_tree_children[0], init_avals) outs = while_p.bind(*cond_consts, *body_consts, *init_vals, cond_nconsts=len(cond_consts), cond_jaxpr=cond_jaxpr, body_nconsts=len(body_consts), body_jaxpr=body_jaxpr) return tree_unflatten(body_tree, outs)
def _while_loop_abstract_eval(*args, **kwargs): return _map(raise_to_shaped, kwargs["body_jaxpr"].out_avals) def _while_loop_translation_rule(ctx, avals_in, avals_out, *args, cond_jaxpr, body_jaxpr, cond_nconsts, body_nconsts): c = ctx.builder cond_consts, body_consts, init_vals = split_list(args, [cond_nconsts, body_nconsts]) batched = bool(cond_jaxpr.out_avals[0].shape) # Since jaxprs don't have tuples and have multiple return values, but we need # the HLO While loop to take a single tuple input and output a single boolean # (for the cond computation) or a single tuple output (for the body # computation), we build XLA computations that handle the tuple munging before # generating a Call into the computations formed from the jaxprs. init_carry = xops.Tuple(c, cond_consts + body_consts + init_vals) cond_c = xla_client.XlaBuilder("cond_computation") cond_carry = xla.parameter(cond_c, 0, c.get_shape(init_carry)) cond_carry_elts = [xops.GetTupleElement(cond_carry, i) for i in range(len(args))] x, _, z = split_list(cond_carry_elts, [cond_nconsts, body_nconsts]) cond_ctx = ctx.replace(builder=cond_c, name_stack=extend_name_stack(ctx.name_stack, 'cond')) pred, = xla.jaxpr_subcomp( cond_ctx, cond_jaxpr.jaxpr, _map(partial(xla.pyval_to_ir_constant, cond_c), cond_jaxpr.consts), *(x + z)) if batched: scalar = ShapedArray((), np.bool_) or_ = xla.primitive_subcomputation(ctx.platform, ctx.axis_env, lax.or_p, scalar, scalar) pred = xops.Reduce(cond_c, [pred], [xops.Constant(cond_c, np.array(False))], or_, list(range(cond_jaxpr.out_avals[0].ndim))) body_c = xla_client.XlaBuilder("body_computation") body_carry = xla.parameter(body_c, 0, c.get_shape(init_carry)) body_carry_elts = [xops.GetTupleElement(body_carry, i) for i in range(len(args))] x, y, z = split_list(body_carry_elts, [cond_nconsts, body_nconsts]) body_ctx = ctx.replace(builder=body_c, name_stack=extend_name_stack(ctx.name_stack, 'body')) new_z = xla.jaxpr_subcomp( body_ctx, body_jaxpr.jaxpr, _map(partial(xla.pyval_to_ir_constant, body_c), body_jaxpr.consts), *(y + z)) if batched: body_pred_ctx = body_ctx.replace( name_stack=extend_name_stack(ctx.name_stack, 'body_pred')) body_pred, = xla.jaxpr_subcomp( body_pred_ctx, cond_jaxpr.jaxpr, _map(partial(xla.pyval_to_ir_constant, body_c), cond_jaxpr.consts), *(x + z)) new_z = _map(partial(_pred_bcast_select, body_c, body_pred), new_z, z, body_jaxpr.out_avals) assert _map(body_c.get_shape, new_z) == _map(body_c.get_shape, z) # no broadcast new_carry = xops.Tuple(body_c, [*x, *y, *new_z]) ans = xops.While(cond_c.build(pred), body_c.build(new_carry), init_carry) ans_elts = [xops.GetTupleElement(ans, i) for i in range(len(args))] _, _, z = split_list(ans_elts, [cond_nconsts, body_nconsts]) return z def _pred_bcast_select(c, pred, x, y, x_y_aval: core.AbstractValue): pred_shape = c.get_shape(pred).dimensions() x_shape = c.get_shape(x).dimensions() y_shape = c.get_shape(y).dimensions() assert x_shape == y_shape if x_y_aval is core.abstract_unit: return x elif x_y_aval is core.abstract_token: return xops.AfterAll(c, [x, y]) else: assert pred_shape == x_shape[:len(pred_shape)] == y_shape[:len(pred_shape)], (pred_shape, x_shape, y_shape) bcast_pred = xops.BroadcastInDim(pred, x_shape, list(range(len(pred_shape)))) return xops.Select(bcast_pred, x, y) def _while_loop_batching_rule(axis_size, axis_name, main_type, args, dims, cond_nconsts, cond_jaxpr, body_nconsts, body_jaxpr): orig_batched = [d is not batching.not_mapped for d in dims] cconst_bat, bconst_bat, init_bat = split_list(orig_batched, [cond_nconsts, body_nconsts]) cconsts, bconsts, init = split_list(args, [cond_nconsts, body_nconsts]) cconst_dims, bconst_dims, init_dims = split_list(dims, [cond_nconsts, body_nconsts]) carry_bat = init_bat # Fixpoint computation of which carry are batched: either # batched from init, or the carry out is batched. Each iteration promotes # at least one carry to batched. We need at most len(carry) iterations to # reach a fixpoint. for _ in range(1 + len(carry_bat)): _, carry_bat_out = batching.batch_jaxpr( body_jaxpr, axis_size, bconst_bat + carry_bat, instantiate=carry_bat, axis_name=axis_name, main_type=main_type) if carry_bat == carry_bat_out: break carry_bat = safe_map(operator.or_, carry_bat, carry_bat_out) else: assert False, "Fixpoint not reached" # Knowing how the carry is batched now, we can determine if the predicate is # batched. _, (pred_bat,) = batching.batch_jaxpr( cond_jaxpr, axis_size, cconst_bat + carry_bat, instantiate=False, axis_name=axis_name, main_type=main_type) if pred_bat: # If the predicate is batched, we have to batch *all* of the carry # regardless of if the body needs it. carry_bat = [True] * len(carry_bat) carry_dims = [0] * len(carry_bat) body_jaxpr_batched, _ = batching.batch_jaxpr_axes( body_jaxpr, axis_size, bconst_dims + carry_dims, carry_dims, axis_name=axis_name, main_type=main_type) cond_jaxpr_batched, _ = batching.batch_jaxpr_axes( cond_jaxpr, axis_size, cconst_dims + carry_dims, [0], axis_name=axis_name, main_type=main_type) else: # If the predicate is not batched, we can look at the `cond_jaxpr`'s out # shape to determine the rank of the predicate. From this rank we pick the # dims of the carry to be batched to ensure that the predicate shape is a # prefix of the carry in and out shapes. We can then batch the `body_jaxpr` # according to these new batch dims. cond_rank = len(cond_jaxpr.out_avals[0].shape) carry_dims = [cond_rank if b else None for b in carry_bat] body_jaxpr_batched, _ = batching.batch_jaxpr_axes( body_jaxpr, axis_size, bconst_dims + carry_dims, carry_dims, axis_name=axis_name, main_type=main_type) # Now we need to rebatch the `cond_jaxpr` according to the new dims of the # carry. cond_jaxpr_batched, _ = batching.batch_jaxpr_axes( cond_jaxpr, axis_size, cconst_dims + carry_dims, (None,), axis_name=axis_name, main_type=main_type) # To prepare the `init` to the `while_p`, we broadcast values if they are # unbatched and need to have an out axis. If their current batch axis does not # match the one it needs to be for the translation rule to work, we move it # into place. new_init = [] for x, old_axis, new_axis in zip(init, init_dims, carry_dims): if old_axis is batching.not_mapped and new_axis is not batching.not_mapped: new_init.append(batching.broadcast(x, axis_size, new_axis)) elif old_axis is batching.not_mapped and new_axis is batching.not_mapped: new_init.append(x) else: assert new_axis is not batching.not_mapped new_init.append(batching.moveaxis(x, old_axis, new_axis)) outs = while_p.bind(*(cconsts + bconsts + new_init), cond_nconsts=cond_nconsts, cond_jaxpr=cond_jaxpr_batched, body_nconsts=body_nconsts, body_jaxpr=body_jaxpr_batched) return outs, carry_dims def _while_loop_jvp(primals, tangents, cond_nconsts, cond_jaxpr, body_nconsts, body_jaxpr): nonzeros = [type(t) is not ad_util.Zero for t in tangents] cconst_nz, bconst_nz, init_nz = split_list(nonzeros, [cond_nconsts, body_nconsts]) carry_nz = init_nz for _ in range(1 + len(carry_nz)): body_nonzeros = bconst_nz + carry_nz body_jvp, nonzeros_out = ad.jvp_jaxpr( body_jaxpr, body_nonzeros, instantiate=carry_nz) if nonzeros_out == carry_nz: break carry_nz = _map(operator.or_, carry_nz, nonzeros_out) else: assert False, "Fixpoint not reached" nonzeros = cconst_nz + body_nonzeros tangents = [ad.instantiate_zeros(t) if nz else t for t, nz in zip(tangents, nonzeros)] cconst, bconst, init = split_list(primals, [cond_nconsts, body_nconsts]) _, bconst_dot, init_dot = split_list(tangents, [cond_nconsts, body_nconsts]) bconst_dot = _prune_zeros(bconst_dot) init_dot = _prune_zeros(init_dot) num_carry = len(primals) - cond_nconsts - body_nconsts body_jvp_rearranged = ad.rearrange_binders( body_jvp, [body_nconsts, num_carry], [len(bconst_dot), len(init_dot)], [num_carry], [len(init_dot)]) newvar = core.gensym([cond_jaxpr.jaxpr]) invars_aug = ( cond_jaxpr.jaxpr.invars + [newvar(core.get_aval(x)) for x in init_dot]) cond_jaxpr_augmented = core.Jaxpr(cond_jaxpr.jaxpr.constvars, invars_aug, cond_jaxpr.jaxpr.outvars, cond_jaxpr.jaxpr.eqns) cond_jaxpr_augmented = core.ClosedJaxpr(cond_jaxpr_augmented, cond_jaxpr.consts) out = while_p.bind( *(cconst + bconst + bconst_dot + init + init_dot), cond_nconsts=cond_nconsts, cond_jaxpr=cond_jaxpr_augmented, body_nconsts=len(bconst) + len(bconst_dot), body_jaxpr=body_jvp_rearranged) out_carry, out_carry_dot = split_list(out, [num_carry]) out_tangents_iter = iter(out_carry_dot) out_tangents = [next(out_tangents_iter) if nz else ad_util.Zero.from_value(p) for p, nz in zip(out_carry, nonzeros_out)] return out_carry, out_tangents def _while_partial_eval(trace: pe.JaxprTrace, *tracers: pe.Tracer, cond_nconsts: int, cond_jaxpr: pe.ClosedJaxpr, body_nconsts: int, body_jaxpr: pe.ClosedJaxpr) -> Sequence[pe.Tracer]: """An implementation of partial evaluation for while. As long as some carry (and hence output) are known and the output of `cond_jaxpr` is known, we use a portion of the loop body to compute the known outputs of the `while_loop`. For the unknown outputs we generate Jaxpr to run the whole while, including recomputing the known parts. This means that we don't actually save any computation by partial evaluation if there are unknown outputs. What this achieves is that we can give a proper error for reverse differentiation of `while`, because in that use of partial evaluation the primal inputs are considered "known", and only the tangent computation is unknown (see issue #2129). """ unknowns = [not t.pval.is_known() for t in tracers] params = dict(cond_nconsts=cond_nconsts, cond_jaxpr=cond_jaxpr, body_nconsts=body_nconsts, body_jaxpr=body_jaxpr) cond_consts_uk, body_consts_uk, carry_init_uk = split_list(unknowns, [cond_nconsts, body_nconsts]) # Fixpoint computation of unknown carry. Each iteration promotes # at least one carry to unknown. We need one last iteration to prepare the jaxpr. carry_uk = carry_init_uk for _ in range(1 + len(carry_uk)): body_jaxpr_known, _, carry_out_uk = pe.partial_eval_jaxpr( # type: ignore body_jaxpr, body_consts_uk + carry_uk, instantiate=carry_uk) if carry_out_uk == carry_uk: break else: carry_uk = _map(operator.or_, carry_uk, carry_out_uk) else: assert False, "Fixpoint not reached" cond_jaxpr_known, _, cond_uk = pe.partial_eval_jaxpr( # type: ignore cond_jaxpr, cond_consts_uk + carry_uk, instantiate=False) if cond_uk[0] or all([not uk for uk in unknowns]) or all(unknowns): # If conditional is unknown, or all inputs are known, or all are unknown, # just do the default processing. return trace.default_process_primitive(while_p, tracers, params) # Run the known part of the while. Prepare the inputs, as constants (if known), or # as core.unit. in_consts = [ core.unit if uk else t.pval.get_known() for uk, t in zip(cond_consts_uk + body_consts_uk + carry_uk, tracers)] # There should be no residuals for the cond_jaxpr_known assert 1 == len(cond_jaxpr_known.out_avals) # We ignore the residuals from the body_jaxpr_known, so the type of inputs matches # the type of outputs; residuals are at the end if len(body_jaxpr_known.out_avals) > len(body_jaxpr.out_avals): # TODO(necula): this is not quite enough; we should drop the residual computations also body_jaxpr_known.jaxpr.outvars = body_jaxpr_known.jaxpr.outvars[:len(body_jaxpr.out_avals)] out_known = while_p.bind( *in_consts, cond_nconsts=cond_nconsts, cond_jaxpr=cond_jaxpr_known, body_nconsts=body_nconsts, body_jaxpr=body_jaxpr_known) # Run the whole while_loop to get all the outputs, then merge with known ones out_all: Sequence[pe.Tracer] = trace.default_process_primitive(while_p, tracers, params) out_tracers: Sequence[pe.Tracer] = [ out_unknown if uk else pe.JaxprTracer(trace, pe.PartialVal.known(known), out_unknown.recipe) for uk, out_unknown, known in zip(carry_uk, out_all, out_known)] return out_tracers def _while_transpose_error(*_, **kwargs): raise ValueError("Reverse-mode differentiation does not work for " "lax.while_loop or lax.fori_loop. " "Try using lax.scan instead.") while_p = core.AxisPrimitive('while') while_p.multiple_results = True while_p.def_impl(partial(xla.apply_primitive, while_p)) while_p.def_abstract_eval(_while_loop_abstract_eval) ad.primitive_jvps[while_p] = _while_loop_jvp pe.custom_partial_eval_rules[while_p] = _while_partial_eval xla.register_translation(while_p, _while_loop_translation_rule, initial_style=True) ad.primitive_transposes[while_p] = _while_transpose_error batching.axis_primitive_batchers[while_p] = _while_loop_batching_rule pe.partial_eval_jaxpr_custom_rules[while_p] = pe.partial_eval_jaxpr_custom_rule_not_implemented def _pred_bcast_select_mhlo( pred_aval: core.ShapedArray, pred: ir.Value, xs: Sequence[ir.Value], ys: Sequence[ir.Value], x_y_aval: core.AbstractValue) -> Sequence[ir.Value]: if x_y_aval is core.abstract_unit: return [] elif x_y_aval is core.abstract_token: x, = xs y, = ys return [mhlo.AfterAllOp(mlir.aval_to_ir_type(x_y_aval), [x, y]).result] else: assert isinstance(x_y_aval, core.ShapedArray), x_y_aval x, = xs y, = ys assert x.type == y.type, (x.type, y.type) assert (pred_aval.shape == x_y_aval.shape[:len(pred_aval.shape)]), ( pred_aval.shape, x_y_aval) bcast_pred = mhlo.BroadcastInDimOp( mlir.aval_to_ir_type(x_y_aval.update(dtype=np.dtype(np.bool_))), pred, mlir.dense_int_elements(list(range(len(pred_aval.shape))))).result return mhlo.SelectOp(bcast_pred, x, y).results def _while_lowering(ctx, avals_in, avals_out, *args, cond_jaxpr, body_jaxpr, cond_nconsts, body_nconsts): pred_aval = cond_jaxpr.out_avals[0] batched = bool(pred_aval.shape) # Since jaxprs don't have tuples and have multiple return values, but we need # the HLO While loop to take a single tuple input and output a single boolean # (for the cond computation) or a single tuple output (for the body # computation), we build XLA computations that handle the tuple munging before # generating a Call into the computations formed from the jaxprs. loop_carry_types = _map(mlir.aval_to_ir_types, avals_in) flat_loop_carry_types = util.flatten(loop_carry_types) loop_carry_tuple_type = ir.TupleType.get_tuple(flat_loop_carry_types) flat_args = mlir.flatten_lowering_ir_args(args) init_carry = mhlo.TupleOp(loop_carry_tuple_type, flat_args) while_op = mhlo.WhileOp([loop_carry_tuple_type], [init_carry.result]) # Loop condition cond_block = while_op.regions[0].blocks.append(loop_carry_tuple_type) with ir.InsertionPoint(cond_block): flat_cond_args = [ mhlo.GetTupleElementOp(input_type, cond_block.arguments[0], mlir.i32_attr(i)).result for i, input_type in enumerate(flat_loop_carry_types)] cond_args = util.unflatten(flat_cond_args, _map(len, loop_carry_types)) x, _, z = util.split_list(cond_args, [cond_nconsts, body_nconsts]) cond_ctx = ctx.replace( name_stack=xla.extend_name_stack(ctx.name_stack, 'cond')) (pred,), = mlir.jaxpr_subcomp( cond_ctx, cond_jaxpr.jaxpr, _map(mlir.ir_constants, cond_jaxpr.consts), *(x + z)) if batched: pred, = lax._unary_reduce_lower( mhlo.OrOp, lambda dtype: np.array(False, dtype), ctx, [pred_aval], [pred_aval.update(shape=())], pred, axes=tuple(range(len(pred_aval.shape)))) mhlo.ReturnOp([pred]) # Loop body body_block = while_op.regions[1].blocks.append(loop_carry_tuple_type) with ir.InsertionPoint(body_block): flat_body_args = [ mhlo.GetTupleElementOp(input_type, body_block.arguments[0], mlir.i32_attr(i)).result for i, input_type in enumerate(flat_loop_carry_types)] body_args = util.unflatten(flat_body_args, _map(len, loop_carry_types)) x, y, z = util.split_list(body_args, [cond_nconsts, body_nconsts]) body_ctx = ctx.replace( name_stack=xla.extend_name_stack(ctx.name_stack, 'body')) new_z = mlir.jaxpr_subcomp( body_ctx, body_jaxpr.jaxpr, _map(mlir.ir_constants, body_jaxpr.consts), *(y + z)) if batched: body_pred_ctx = ctx.replace( name_stack=xla.extend_name_stack(ctx.name_stack, 'body_pred')) (body_pred,), = mlir.jaxpr_subcomp( body_pred_ctx, cond_jaxpr.jaxpr, _map(mlir.ir_constants, cond_jaxpr.consts), *(x + z)) new_z = _map(partial(_pred_bcast_select_mhlo, pred_aval, body_pred), new_z, z, body_jaxpr.out_avals) new_carry = mhlo.TupleOp( loop_carry_tuple_type, [*util.flatten(x), *util.flatten(y), *util.flatten(new_z)]) mhlo.ReturnOp([new_carry.result]) outputs = util.unflatten([ mhlo.GetTupleElementOp(output_type, while_op.result, mlir.i32_attr(i)).result for i, output_type in enumerate(flat_loop_carry_types) ], _map(len, loop_carry_types)) _, _, z = util.split_list(outputs, [cond_nconsts, body_nconsts]) return z mlir.register_lowering(while_p, _while_lowering) ### cond and switch # For backward compatibility with a previous switch/cond calling convention, # we allow a single (pytree) `operand` argument to be passed by keyword. We use # a sentinel object as its default value to indicate when it is _not_ passed. _no_operand_sentinel = object()
[docs]@api_boundary def switch(index, branches: Sequence[Callable], *operands, operand=_no_operand_sentinel): """Apply exactly one of ``branches`` given by ``index``. If ``index`` is out of bounds, it is clamped to within bounds. Has the semantics of the following Python:: def switch(index, branches, operand): index = clamp(0, index, len(branches) - 1) return branches[index](operand) Args: index: Integer scalar type, indicating which branch function to apply. branches: Sequence of functions (A -> B) to be applied based on ``index``. operands: Operands (A) input to whichever branch is applied. Returns: Value (B) of ``branch(*operands)`` for the branch that was selected based on ``index``. """ if operand is not _no_operand_sentinel: if operands: raise TypeError("if 'operand' keyword is passed then no positional " f"operands can be passed, got operand={operand} " f"and positional operands {operands}") operands = (operand,) del operand if len(np.shape(index)) != 0: raise TypeError( f"Branch index must be scalar, " f"got {index} of shape {np.shape(index)}.") try: index_dtype = dtypes.result_type(index) except TypeError as err: msg = f"Index type must be an integer, got {index}." raise TypeError(msg) from err if index_dtype.kind not in 'iu': raise TypeError( f"Index type must be an integer, got {index} as {index_dtype}") branches = tuple(branches) if len(branches) == 0: raise ValueError("Empty branch sequence") elif len(branches) == 1: return branches[0](*operands) index = lax.convert_element_type(index, np.int32) lo = np.array(0, np.int32) hi = np.array(len(branches) - 1, np.int32) index = lax.clamp(lo, index, hi) if (config.jax_disable_jit and isinstance(core.get_aval(index), ConcreteArray)): return branches[int(index)](*operands) ops, ops_tree = tree_flatten(operands) ops_avals = tuple(_map(_abstractify, ops)) jaxprs, consts, out_trees = _initial_style_jaxprs_with_common_consts( branches, ops_tree, ops_avals, primitive_name='switch') for i, (out_tree, jaxpr) in enumerate(zip(out_trees[1:], jaxprs[1:])): _check_tree_and_avals(f"branch 0 and {i + 1} outputs", out_trees[0], jaxprs[0].out_avals, out_tree, jaxpr.out_avals) linear = (False,) * (len(consts) + len(ops)) out = cond_p.bind( index, *consts, *ops, branches=tuple(jaxprs), linear=linear) return tree_unflatten(out_trees[0], out)
def _cond(pred, true_fun: Callable, false_fun: Callable, *operands, operand=_no_operand_sentinel): """Conditionally apply ``true_fun`` or ``false_fun``. ``cond()`` has equivalent semantics to this Python implementation:: def cond(pred, true_fun, false_fun, *operands): if pred: return true_fun(*operands) else: return false_fun(*operands) ``pred`` must be a scalar type. Args: pred: Boolean scalar type, indicating which branch function to apply. true_fun: Function (A -> B), to be applied if ``pred`` is True. false_fun: Function (A -> B), to be applied if ``pred`` is False. operands: Operands (A) input to either branch depending on ``pred``. The type can be a scalar, array, or any pytree (nested Python tuple/list/dict) thereof. Returns: Value (B) of either ``true_fun(*operands)`` or ``false_fun(*operands)``, depending on the value of ``pred``. The type can be a scalar, array, or any pytree (nested Python tuple/list/dict) thereof. """ if operand is not _no_operand_sentinel: if operands: raise TypeError("if 'operand' keyword is passed then no positional " f"operands can be passed, got operand={operand} " f"and positional operands {operands}") operands = (operand,) del operand if isinstance(pred, Sequence) or np.ndim(pred) != 0: raise TypeError( f"Pred must be a scalar, got {pred} of " + (f"type {type(pred)}" if isinstance(pred, Sequence) else f"shape {np.shape(pred)}.")) try: pred_dtype = dtypes.result_type(pred) except TypeError as err: msg = ("Pred type must be either boolean or number, got {}.") raise TypeError(msg.format(pred)) from err if pred_dtype.kind != 'b': if pred_dtype.kind in 'iuf': pred = pred != 0 else: msg = ("Pred type must be either boolean or number, got {}.") raise TypeError(msg.format(pred_dtype)) if config.jax_disable_jit and isinstance(core.get_aval(pred), ConcreteArray): if pred: return true_fun(*operands) else: return false_fun(*operands) ops, ops_tree = tree_flatten(operands) ops_avals = tuple(_map(_abstractify, ops)) jaxprs, consts, out_trees = _initial_style_jaxprs_with_common_consts( (true_fun, false_fun), ops_tree, ops_avals, 'cond') true_jaxpr, false_jaxpr = jaxprs out_tree, false_out_tree = out_trees _check_tree_and_avals("true_fun and false_fun output", out_tree, true_jaxpr.out_avals, false_out_tree, false_jaxpr.out_avals) index = lax.convert_element_type(pred, np.int32) linear = (False,) * (len(consts) + len(ops)) out = cond_p.bind( index, *consts, *ops, branches=(false_jaxpr, true_jaxpr), linear=linear) return tree_unflatten(out_tree, out)
[docs]@api_boundary @functools.wraps(_cond) def cond(*args, **kwargs): # detect an attempt to call the former, deprecated cond try: ba = inspect.signature(_cond_with_per_branch_args).bind(*args, **kwargs) except TypeError: pass else: assert not ba.kwargs # no catch-all **kwargs in _cond_with_per_branch _, _, maybe_true_fun, _, maybe_false_fun = ba.args if callable(maybe_true_fun) and callable(maybe_false_fun): return _cond_with_per_branch_args(*ba.args) return _cond(*args, **kwargs)
def _cond_with_per_branch_args(pred, true_operand, true_fun: Callable, false_operand, false_fun: Callable): """Conditionally apply ``true_fun`` or ``false_fun``. Has equivalent semantics to this Python implementation:: def cond(pred, true_operand, true_fun, false_operand, false_fun): if pred: return true_fun(true_operand) else: return false_fun(false_operand) Pred has to be a scalar type, collection types (list, tuple) are not supported """ return _cond(pred, lambda op: true_fun(op[0]), lambda op: false_fun(op[1]), (true_operand, false_operand)) def _cond_abstract_eval(*args, **kwargs): return _map(raise_to_shaped, kwargs["branches"][0].out_avals) def _cond_translation_rule(ctx, avals_in, avals_out, index, *args, branches, linear): del linear # Unused. name_stack = extend_name_stack(ctx.name_stack, "cond") def make_computation(name, jaxpr, op_shape): c = xla_client.XlaBuilder(name + '_comp') op = xla.parameter(c, 0, op_shape) ops = [xops.GetTupleElement(op, i) for i in range(len(jaxpr.in_avals))] subctx = ctx.replace( builder=c, name_stack=extend_name_stack(name_stack, name + '_fun')) outs = xla.jaxpr_subcomp( subctx, jaxpr.jaxpr, _map(partial(xla.pyval_to_ir_constant, c), jaxpr.consts), *ops) return c.build(xops.Tuple(c, outs)) c = ctx.builder op = xops.Tuple(c, args) op_shape = c.get_shape(op) branch_computations = [ make_computation(f'branch_{i}', jaxpr, op_shape) for i, jaxpr in enumerate(branches)] return xla.xla_destructure( c, xops.Conditional(index, branch_computations, [op] * len(branches))) def _select_tree(indices, branch_vals): assert len(branch_vals) > 0 if len(branch_vals) == 1: return branch_vals[0] mid = lax._const(indices, len(branch_vals) // 2) return _bcast_select(lax.lt(indices, mid), _select_tree(indices, branch_vals[:mid]), _select_tree(indices - mid, branch_vals[mid:])) def _bcast_select(pred, on_true, on_false): if np.ndim(pred) != np.ndim(on_true): idx = list(range(np.ndim(pred))) pred = lax.broadcast_in_dim(pred, np.shape(on_true), idx) return lax.select(pred, on_true, on_false) def _cond_batching_rule(axis_size, axis_name, main_type, args, dims, branches, linear): index, *ops = args index_dim, *op_dims = dims if index_dim is not batching.not_mapped: # Convert to a lax.select. While we could get away with not broadcasting # some operands yet, because all outputs must be broadcast together anyway # for the select we broadcast the input operands for simplicity and leave # optimizations to XLA. # TODO(mattjj,frostig): assumes branches are side-effect-free, revise! index, *ops = [batching.bdim_at_front(x, d, axis_size) for x, d in zip(args, dims)] branches_batched = [ batching.batch_jaxpr(jaxpr, axis_size, [True] * len(ops), True, axis_name, main_type)[0] for jaxpr in branches] branch_outs = [] for i, jaxpr in enumerate(branches_batched): # Perform a select on the inputs for safety of reverse-mode autodiff; see # https://github.com/google/jax/issues/1052 ops_ = [_bcast_select(lax.eq(index, lax._const(index, i)), x, lax.stop_gradient(x)) if x is not core.unit else x for x in ops] branch_outs.append(core.jaxpr_as_fun(jaxpr)(*ops_)) out = [_select_tree(index, outs) if outs[0] is not core.unit else outs[0] for outs in zip(*branch_outs)] return out, [0] * len(branch_outs[0]) else: ops_bat = [d is not batching.not_mapped for d in op_dims] ops = [batching.moveaxis(x, d, 0) if b else x for b, x, d in zip(ops_bat, ops, op_dims)] branches_out_bat = [ batching.batch_jaxpr(jaxpr, axis_size, ops_bat, False, axis_name, main_type)[1] for jaxpr in branches] out_bat = [any(bat) for bat in zip(*branches_out_bat)] branches_batched = tuple( batching.batch_jaxpr(jaxpr, axis_size, ops_bat, out_bat, axis_name, main_type)[0] for jaxpr in branches) out_dims = [0 if b else batching.not_mapped for b in out_bat] out = cond_p.bind( index, *ops, branches=branches_batched, linear=linear) return out, out_dims def _cond_jvp(primals, tangents, branches, linear): nonzeros = [type(t) is not ad_util.Zero for t in tangents] index_nz, *ops_nz = nonzeros assert index_nz is False branches_out_nz = [ad.jvp_jaxpr(jaxpr, ops_nz, instantiate=False)[1] for jaxpr in branches] out_nz = [any(nz) for nz in zip(*branches_out_nz)] branches_jvp = tuple(ad.jvp_jaxpr(jaxpr, ops_nz, instantiate=out_nz)[0] for jaxpr in branches) index, *ops = primals _, *ops_dot = tangents ops_dot = _prune_zeros(ops_dot) ops_lin = tuple(linear) linear_jvp = ops_lin + (True,) * len(ops_dot) out = cond_p.bind( index, *ops, *ops_dot, branches=branches_jvp, linear=linear_jvp) out_primals, out_tangents = split_list(out, [len(out_nz)]) out_tangents_iter = iter(out_tangents) out_tangents = [next(out_tangents_iter) if nz else ad_util.Zero.from_value(p) for p, nz in zip(out_primals, out_nz)] return out_primals, out_tangents def _cond_partial_eval(trace, *tracers, branches, linear): unknowns = [t.pval[0] is not None for t in tracers] index_uk, *ops_uk = unknowns if index_uk: # When the branch index is unknown, we stage out the whole cond. params = dict(branches=branches, linear=linear) return trace.default_process_primitive(cond_p, tracers, params) branches_out_uks = [] for branch_jaxpr in branches: _, _, out_uks = pe.partial_eval_jaxpr(branch_jaxpr, ops_uk, instantiate=False) branches_out_uks.append(out_uks) out_uks = [any(uks) for uks in zip(*branches_out_uks)] branches_1, branches_2, branch_res_avals = [], [], [] for branch_jaxpr in branches: branch_jaxpr_1, branch_jaxpr_2, _ = pe.partial_eval_jaxpr( branch_jaxpr, ops_uk, instantiate=out_uks) branch_num_res = len(branch_jaxpr_1.out_avals) - len(out_uks) # move residuals to the front move = [False] * len(ops_uk) + [True] * branch_num_res branch_jaxpr_2 = pe.move_binders_to_front(branch_jaxpr_2, move) # TODO(frostig,mattjj): pe.partial_eval_jaxpr should raise to shaped avals res_avals = _map( raise_to_shaped, branch_jaxpr_2.in_avals[:branch_num_res]) branches_1.append(branch_jaxpr_1) branches_2.append(branch_jaxpr_2) branch_res_avals.append(res_avals) branches_1 = tuple(branches_1) branches_2 = tuple(branches_2) for jaxpr in branches_2[1:]: assert len(jaxpr.out_avals) == len(branches_2[0].out_avals) num_outs = len(branches_2[0].out_avals) all_res_avals, res_avals_per_branch = _merge_branch_residuals( branch_res_avals) branches_1 = _join_cond_outputs( branches_1, all_res_avals, res_avals_per_branch, num_outs) branches_2 = _join_cond_pe_staged_jaxpr_inputs( branches_2, all_res_avals, res_avals_per_branch) # TODO(frostig,mattjj): reinstate this assertion once pe.partial_eval_jaxpr # raises to shaped avals # for j in branches_1[1:]: # assert j.out_avals == branches_1[0].out_avals num_res = len(all_res_avals) _, in_consts = unzip2([t.pval for t in tracers]) out_consts_res = cond_p.bind(*in_consts, branches=branches_1, linear=linear) out_consts, res = split_list(out_consts_res, [len(out_consts_res) - num_res]) # TODO(frostig,mattjj): remove raised_to_shaped of avals once # pe.partial_eval_jaxpr handles it out_avals = _map(raise_to_shaped, branches_2[0].out_avals) out_pvs = [aval if uk else None for aval, uk in zip(out_avals, out_uks)] index_tracer = trace.instantiate_const(tracers[0]) ops_tracers = [trace.instantiate_const(t) if uk else trace.new_instantiated_literal(core.unit) for uk, t in zip(unknowns[1:], tracers[1:])] res_tracers = _map(trace.new_instantiated_const, res) out_tracers = [pe.JaxprTracer(trace, pe.PartialVal((pv, const)), None) for pv, const in zip(out_pvs, out_consts)] linear_2 = (False,) * num_res + linear params = dict(branches=branches_2, linear=linear_2) eqn = pe.new_eqn_recipe( [index_tracer] + res_tracers + ops_tracers, out_tracers, cond_p, params, source_info_util.current()) for t in out_tracers: t.recipe = eqn return out_tracers # When partially evaluating conditionals, each branch produces residuals # depending on the computation carried out by the branch, and a corresponding # staged jaxpr that accepts those residuals as its first few inputs. The # residual-producing branches are staged as jaxprs and bound right away in a # conditional. The residual-consuming jaxprs are assembled together in a jaxpr # conditional. The following helper functions ensure that both collections of # jaxprs (those evaluated and those staged) are valid for joint use under their # respective conditionals. # # In particular, the residuals derived from each original branch may have # distinct types. Because the branches of conditionals must have identical type # signatures, we join residuals together across branches into a common format. # In order to set up a type signature that all branches can conform to, it would # suffice to concatenate all branches' residuals. But concatenation can result # in redundant inputs and outputs, and might lead to memory allocation that # scales unnecessarily with the branch count. This function finds common # residual types across branches for reuse, so as to avoid redundant # allocation. It returns a list L of types (avals) representing the collection # of residuals merged according to type, and, for each branch, a lookup table to # match its residuals to their positions/types in L. Example input/output: # # [x], [y], [x, x] -> [x, y, x], [[0], [1], [0, 2]] # [x], [x], [x, x] -> [x, x], [[0], [0], [0, 1]] # [y, x, x], [x, z, y], [z, x] -> [y, x, x, z], [[0, 1, 2], [1, 3, 0], [3, 1]] def _merge_branch_residuals(branch_res_avals): def enumerate_equal(xs): counts = {v: itertools.count() for v in set(xs)} return [(x, next(counts[x])) for x in xs] branch_res_tagged_avals = _map(enumerate_equal, branch_res_avals) all_tagged_avals = _ordered_unique(util.concatenate(branch_res_tagged_avals)) indices = {v: i for i, v in enumerate(all_tagged_avals)} branch_indices = [ [indices[aval] for aval in avals] for avals in branch_res_tagged_avals] all_avals = [x for x, _ in all_tagged_avals] return all_avals, branch_indices # This function augments branch outputs to agree with the merged residual # format: each branch is made to return zero-filled values in the places of # residual outputs that it does not populate. def _join_cond_outputs(jaxprs, all_res_avals, res_aval_indices_per_jaxpr, num_non_res_outputs): def augment_jaxpr(jaxpr, res_indices): @lu.wrap_init def f_aug(*args): outs_and_residuals = core.jaxpr_as_fun(jaxpr)(*args) outs, residuals = split_list(outs_and_residuals, [num_non_res_outputs]) aug_residuals = _map(ad_util.zeros_like_aval, all_res_avals) aug_residuals = util.subvals(aug_residuals, zip(res_indices, residuals)) return outs + list(aug_residuals) return _make_closed_jaxpr(f_aug, jaxpr.in_avals) return tuple(_map(augment_jaxpr, jaxprs, res_aval_indices_per_jaxpr)) # This function augments branch inputs to agree with the merged residual format: # each branch is made to accept all residuals, even though it will ignore those # that it does not read. def _join_cond_pe_staged_jaxpr_inputs(jaxprs, all_res_avals, res_aval_indices_per_jaxpr): newvar = core.gensym([j.jaxpr for j in jaxprs], suffix='_') all_res_vars = _map(newvar, all_res_avals) def augment_jaxpr(jaxpr, res_indices): num_res = len(res_indices) res_vars = jaxpr.jaxpr.invars[:num_res] non_res_vars = jaxpr.jaxpr.invars[num_res:] aug_res_vars = list(util.subvals(all_res_vars, zip(res_indices, res_vars))) aug_invars = aug_res_vars + non_res_vars jaxpr_aug = core.Jaxpr(jaxpr.jaxpr.constvars, aug_invars, jaxpr.jaxpr.outvars, jaxpr.jaxpr.eqns) jaxpr_aug = core.ClosedJaxpr(jaxpr_aug, jaxpr.consts) return jaxpr_aug return tuple(_map(augment_jaxpr, jaxprs, res_aval_indices_per_jaxpr)) def _ordered_unique(xs): d = collections.OrderedDict((x, None) for x in xs) return list(d.keys()) def _transpose_cond_jaxpr(jaxpr, num_res, reduce_axes): res_avals, primal_avals = split_list(jaxpr.in_avals, [num_res]) primal_avals = _map(raise_to_shaped, primal_avals) @lu.wrap_init def transposed(*args): res, cts_out = split_list(args, [num_res]) primals = res + [ad.UndefinedPrimal(aval) for aval in primal_avals] cts_in = ad.backward_pass( jaxpr.jaxpr, reduce_axes, jaxpr.consts, primals, cts_out) _, cts_in = split_list(cts_in, [num_res]) return _map(ad.instantiate_zeros_aval, primal_avals, cts_in) return _make_closed_jaxpr(transposed, res_avals + jaxpr.out_avals) def _cond_transpose(reduce_axes, cts, *args, branches, linear): index, *ops = args in_avals = _map(raise_to_shaped, branches[0].in_avals) num_res = len(ops) - sum(linear) branches_trans = tuple( _transpose_cond_jaxpr(jaxpr, num_res, reduce_axes) for jaxpr in branches) lin_in_avals = [raise_to_shaped(a, weak_type=False) for a, l in zip(in_avals, linear) if l] assert all(core.typematch(out_aval, lin_in_aval) for jaxpr in branches_trans for out_aval, lin_in_aval in zip(jaxpr.out_avals, lin_in_avals)) res = ops[:num_res] cts = _map(ad.instantiate_zeros_aval, branches[0].out_avals, cts) linear_trans = (False,) * num_res + (True,) * len(cts) out = cond_p.bind( index, *res, *cts, branches=branches_trans, linear=linear_trans) assert all(_map(core.typecheck, lin_in_avals, out)) out_iter = iter(out) out = [next(out_iter) if l else None for l in linear] assert next(out_iter, None) is None return [None] + out def _avals_short(avals): to_str = lambda aval: getattr(aval, 'str_short', partial(str, aval))() return ' '.join(_map(to_str, avals)) def _cond_typecheck(*avals, branches, linear): tc = partial(_typecheck_param, 'cond') tc(branches, 'branches', 'tuple of ClosedJaxpr', type(branches) is tuple and all(type(x) is core.ClosedJaxpr for x in branches)) tc(linear, 'linear', 'tuple of bool', type(linear) is tuple and all(type(x) is bool for x in linear)) core.typecheck_assert( len(branches) > 0, 'cond requires at least one branch function') core.typecheck_assert( len(linear) + 1 == len(avals), f'cond given {len(linear)} linear flags for ' f'{len(avals) - 1} non-predicate operands') jaxpr0 = branches[0] jaxpr0_in_avals_str = _avals_short(jaxpr0.in_avals) jaxpr0_out_avals_str = _avals_short(jaxpr0.out_avals) for i, jaxpr in enumerate(branches[1:]): core.typecheck_assert( len(jaxpr0.in_avals) == len(jaxpr.in_avals), f'cond branch 0 takes {len(jaxpr0.in_avals)} inputs, ' f'branch {i+1} takes {len(jaxpr.in_avals)}') core.typecheck_assert( len(jaxpr0.out_avals) == len(jaxpr.out_avals), f'cond branch 0 outputs {len(jaxpr0.out_avals)} values, ' f'branch {i+1} outputs {len(jaxpr.out_avals)}') core.typecheck_assert( all(_map(core.typematch, jaxpr0.in_avals, jaxpr.in_avals)), f'cond branches 0 and {i+1} have mismatching input types: ' f'{jaxpr0_in_avals_str} vs {_avals_short(jaxpr.in_avals)}') core.typecheck_assert( all(_map(core.typematch, jaxpr0.out_avals, jaxpr.out_avals)), f'cond branches 0 and {i+1} have mismatching output types: ' f'{jaxpr0_out_avals_str} vs {_avals_short(jaxpr.out_avals)}') core.typecheck_assert( len(avals) == 1 + len(jaxpr0.in_avals), f'cond called with {len(avals) - 1} non-predicate operands, ' f'but branches take {len(jaxpr0.in_avals)} inputs') index_aval, *op_avals = avals core.typecheck_assert( index_aval.dtype == np.int32, f'cond called with index of type {index_aval.dtype} instead of int32') core.typecheck_assert( all(_map(core.typecompat, jaxpr0.in_avals, op_avals)), f'cond branches take input types {jaxpr0_in_avals_str}, ' f'called with operands of type {_avals_short(op_avals)}') def cond_bind(*args, branches, linear): if config.jax_enable_checks: avals = _map(core.get_aval, args) _cond_typecheck(*avals, branches=branches, linear=linear) for jaxpr in branches: core.check_jaxpr(jaxpr.jaxpr) return core.AxisPrimitive.bind(cond_p, *args, branches=branches, linear=linear) cond_p = core.AxisPrimitive('cond') cond_p.multiple_results = True cond_p.def_impl(partial(xla.apply_primitive, cond_p)) cond_p.def_abstract_eval(_cond_abstract_eval) cond_p.def_custom_bind(cond_bind) ad.primitive_jvps[cond_p] = _cond_jvp ad.reducing_transposes[cond_p] = _cond_transpose pe.custom_partial_eval_rules[cond_p] = _cond_partial_eval batching.axis_primitive_batchers[cond_p] = _cond_batching_rule xla.register_translation(cond_p, _cond_translation_rule, initial_style=True) core.custom_typechecks[cond_p] = _cond_typecheck pe.partial_eval_jaxpr_custom_rules[cond_p] = pe.partial_eval_jaxpr_custom_rule_not_implemented def _cond_lowering(ctx, avals_in, avals_out, index, *args, branches, linear): del linear # Unused. arg_avals = avals_in[1:] input_types = _map(mlir.aval_to_ir_types, arg_avals) output_types = _map(mlir.aval_to_ir_types, avals_out) flat_input_types = util.flatten(input_types) flat_output_types = util.flatten(output_types) input_tuple_type = ir.TupleType.get_tuple(flat_input_types) output_tuple_type = ir.TupleType.get_tuple(flat_output_types) op = mhlo.TupleOp(input_tuple_type, mlir.flatten_lowering_ir_args(args)).result # TODO(phawkins): avoid build_generic when CaseOp is fixed. case_op = mhlo.CaseOp.build_generic([output_tuple_type], [index] + [op] * len(branches), regions=len(branches)) for i, jaxpr in enumerate(branches): branch = case_op.regions[i].blocks.append(input_tuple_type) with ir.InsertionPoint(branch): args = [mhlo.GetTupleElementOp(input_type, branch.arguments[0], mlir.i32_attr(i)).result for i, input_type in enumerate(flat_input_types)] unflattened_args = util.unflatten(args, _map(len, input_types)) out_vals = mlir.jaxpr_subcomp(ctx, jaxpr.jaxpr, jaxpr.consts, *unflattened_args) out = mhlo.TupleOp(output_tuple_type, util.flatten(out_vals)).results mhlo.ReturnOp(out) results = [mhlo.GetTupleElementOp(output_type, case_op.result, mlir.i32_attr(i)).result for i, output_type in enumerate(flat_output_types)] return util.unflatten(results, _map(len, output_types)) mlir.register_lowering(cond_p, _cond_lowering) ### scan Carry = TypeVar('Carry') X = TypeVar('X') Y = TypeVar('Y')
[docs]@api_boundary def scan(f: Callable[[Carry, X], Tuple[Carry, Y]], init: Carry, xs: X, length: Optional[int] = None, reverse: bool = False, unroll: int = 1) -> Tuple[Carry, Y]: """Scan a function over leading array axes while carrying along state. The type signature in brief is .. code-block:: haskell scan :: (c -> a -> (c, b)) -> c -> [a] -> (c, [b]) where we use [t] here to denote the type t with an additional leading axis. That is, if t is an array type then [t] represents the type with an additional leading axis, and if t is a pytree (container) type with array leaves then [t] represents the type with the same pytree structure and corresponding leaves each with an additional leading axis. When ``a`` is an array type or None, and ``b`` is an array type, the semantics of ``scan`` are given roughly by this Python implementation:: def scan(f, init, xs, length=None): if xs is None: xs = [None] * length carry = init ys = [] for x in xs: carry, y = f(carry, x) ys.append(y) return carry, np.stack(ys) Unlike that Python version, both ``a`` and ``b`` may be arbitrary pytree types, and so multiple arrays can be scanned over at once and produce multiple output arrays. (None is actually an empty pytree.) Also unlike that Python version, ``scan`` is a JAX primitive and is lowered to a single XLA While HLO. That makes it useful for reducing compilation times for jit-compiled functions, since native Python loop constructs in an ``@jit`` function are unrolled, leading to large XLA computations. Finally, the loop-carried value ``carry`` must hold a fixed shape and dtype across all iterations (and not just be consistent up to NumPy rank/shape broadcasting and dtype promotion rules, for example). In other words, the type ``c`` in the type signature above represents an array with a fixed shape and dtype (or a nested tuple/list/dict container data structure with a fixed structure and arrays with fixed shape and dtype at the leaves). Args: f: a Python function to be scanned of type ``c -> a -> (c, b)``, meaning that ``f`` accepts two arguments where the first is a value of the loop carry and the second is a slice of ``xs`` along its leading axis, and that ``f`` returns a pair where the first element represents a new value for the loop carry and the second represents a slice of the output. init: an initial loop carry value of type ``c``, which can be a scalar, array, or any pytree (nested Python tuple/list/dict) thereof, representing the initial loop carry value. This value must have the same structure as the first element of the pair returned by ``f``. xs: the value of type ``[a]`` over which to scan along the leading axis, where ``[a]`` can be an array or any pytree (nested Python tuple/list/dict) thereof with consistent leading axis sizes. length: optional integer specifying the number of loop iterations, which must agree with the sizes of leading axes of the arrays in ``xs`` (but can be used to perform scans where no input ``xs`` are needed). reverse: optional boolean specifying whether to run the scan iteration forward (the default) or in reverse, equivalent to reversing the leading axes of the arrays in both ``xs`` and in ``ys``. unroll: optional positive int specifying, in the underlying operation of the scan primitive, how many scan iterations to unroll within a single iteration of a loop. Returns: A pair of type ``(c, [b])`` where the first element represents the final loop carry value and the second element represents the stacked outputs of the second output of ``f`` when scanned over the leading axis of the inputs. """ xs_flat, xs_tree = tree_flatten(xs) try: lengths = [x.shape[0] for x in xs_flat] except AttributeError as err: msg = "scan got value with no leading axis to scan over: {}." raise ValueError( msg.format(', '.join(str(x) for x in xs_flat if not hasattr(x, 'shape')))) from err if length is not None: length = int(length) if not all(length == l for l in lengths): msg = ("scan got `length` argument of {} which disagrees with " "leading axis sizes {}.") raise ValueError(msg.format(length, [x.shape[0] for x in xs_flat])) else: unique_lengths = set(lengths) if len(unique_lengths) > 1: msg = "scan got values with different leading axis sizes: {}." raise ValueError(msg.format(', '.join(str(x.shape[0]) for x in xs_flat))) elif len(unique_lengths) == 0: msg = "scan got no values to scan over and `length` not provided." raise ValueError(msg) else: length, = unique_lengths if config.jax_disable_jit: if length == 0: raise ValueError("zero-length scan is not supported in disable_jit() mode because the output type is unknown.") carry = init ys = [] maybe_reversed = reversed if reverse else lambda x: x for i in maybe_reversed(range(length)): xs_slice = [_index_array(i, core.get_aval(x), x) for x in xs_flat] carry, y = f(carry, tree_unflatten(xs_tree, xs_slice)) ys.append(y) stack = lambda y, *ys: (y if core.get_aval(y) is core.abstract_unit else jax.numpy.stack((y, *ys))) stacked_y = tree_multimap(stack, *maybe_reversed(ys)) return carry, stacked_y x_shapes = [masking.padded_shape_as_value(x.shape[1:]) for x in xs_flat] x_dtypes = [x.dtype for x in xs_flat] x_avals = tuple(_map(ShapedArray, x_shapes, x_dtypes)) def _create_jaxpr(init): init_flat, init_tree = tree_flatten(init) in_flat, in_tree = tree_flatten((init, xs)) carry_avals = tuple(_map(_abstractify, init_flat)) jaxpr, consts, out_tree = _initial_style_jaxpr( f, in_tree, carry_avals + x_avals, "scan") out_tree_children = out_tree.children() if len(out_tree_children) != 2: msg = "scan body output must be a pair, got {}." raise TypeError(msg.format(tree_unflatten(out_tree, jaxpr.out_avals))) carry_avals_out = jaxpr.out_avals[:out_tree_children[0].num_leaves] return init_flat, carry_avals, carry_avals_out, init_tree, in_flat, jaxpr, consts, out_tree, out_tree_children # The carry input and output avals must match exactly. However, we want to account for # the case when init contains weakly-typed values (e.g. Python scalars), with avals that # may not match the output despite being compatible by virtue of their weak type. # To do this, we compute the jaxpr in two passes: first with the raw inputs, and if # necessary, a second time with modified init values. init_flat, carry_avals, carry_avals_out, init_tree, *rest = _create_jaxpr(init) new_init_flat, changed = _promote_weak_typed_inputs(init_flat, carry_avals, carry_avals_out) if changed: new_init = tree_unflatten(init_tree, new_init_flat) init_flat, carry_avals, carry_avals_out, init_tree, *rest = _create_jaxpr(new_init) in_flat, jaxpr, consts, out_tree, out_tree_children = rest _check_tree_and_avals("scan carry output and input", # Extract the subtree and avals for the first element of the return tuple out_tree_children[0], carry_avals_out, init_tree, carry_avals) out = scan_p.bind(*consts, *in_flat, reverse=reverse, length=length, jaxpr=jaxpr, num_consts=len(consts), num_carry=len(init_flat), linear=(False,) * (len(consts) + len(in_flat)), unroll=unroll) return tree_unflatten(out_tree, out)
def _scan_impl_unrolled(*args, reverse, length, num_consts, num_carry, linear, f_impl, x_avals, y_avals): consts, init, xs = split_list(args, [num_consts, num_carry]) carry = init ys = [] for i in range(length): i_ = length - i - 1 if reverse else i x = _map(partial(_index_array, i_), x_avals, xs) out = f_impl(*consts, *carry, *x) carry, y = split_list(out, [num_carry]) ys.append(y) ys = list(reversed(ys)) if reverse else ys ys = list(zip(*ys)) ys = _map(_stack, y_avals, ys) return (*carry, *ys) def _scan_impl_loop(*args, reverse, length, num_consts, num_carry, linear, f_impl, x_avals, y_avals): consts, init, xs = split_list(args, [num_consts, num_carry]) def cond_fun(vals): i, *_ = vals return i < length def body_fun(vals): [i], carry, ys = split_list(vals, [1, num_carry]) i_ = length - i - 1 if reverse else i x = _map(partial(_dynamic_index_array, i_), x_avals, xs) out_flat = f_impl(*consts, *carry, *x) carry_out, y_updates = split_list(out_flat, [num_carry]) ys_out = _map(partial(_update_array, i_), y_avals, ys, y_updates) return [i + 1] + carry_out + ys_out ys_init = _map(partial(_empty_array, length), y_avals) if length == 0: return init + ys_init else: init_val = [lax._const(length, 0)] + init + ys_init _, *outs = while_loop(cond_fun, body_fun, init_val) return outs def _scan_impl_block_unrolled(*args, reverse, length, num_consts, num_carry, linear, block_length, f_impl, x_avals, y_avals): consts, init, xs = split_list(args, [num_consts, num_carry]) num_blocks, rem = divmod(length, block_length) assert rem == 0 partition = partial(_partition_leading, num_blocks, block_length) xs_block = _map(partition, x_avals, xs) prepend_aval = partial(_prepend_dim_to_aval, block_length) x_block_avals = _map(prepend_aval, x_avals) y_block_avals = _map(prepend_aval, y_avals) f_impl_block = partial( _scan_impl_unrolled, reverse=reverse, length=block_length, num_consts=num_consts, num_carry=num_carry, linear=linear, f_impl=f_impl, x_avals=x_avals, y_avals=y_avals) outs = _scan_impl_loop( *consts, *init, *xs_block, reverse=reverse, length=num_blocks, num_consts=num_consts, num_carry=num_carry, linear=linear, f_impl=f_impl_block, x_avals=x_block_avals, y_avals=y_block_avals) carry, ys_blocks = split_list(outs, [num_carry]) combine = partial(_combine_leading, num_blocks, block_length) ys = _map(combine, y_avals, ys_blocks) return (*carry, *ys) def _scan_impl(*args, reverse, length, num_consts, num_carry, jaxpr, linear, unroll): _, _, x_avals = split_list(jaxpr.in_avals, [num_consts, num_carry]) _, y_avals = split_list(jaxpr.out_avals, [num_carry]) f_impl = core.jaxpr_as_fun(jaxpr) if unroll == 1: return _scan_impl_loop( *args, reverse=reverse, length=length, num_consts=num_consts, num_carry=num_carry, linear=linear, f_impl=f_impl, x_avals=x_avals, y_avals=y_avals) consts, init, xs = split_list(args, [num_consts, num_carry]) num_blocks, rem = divmod(length, unroll) length_div = num_blocks * unroll if rem > 0: if reverse: split = partial(_split_leading_dim, rem) xs_rem, xs = unzip2(_map(split, x_avals, xs)) else: split = partial(_split_leading_dim, length_div) xs, xs_rem = unzip2(_map(split, x_avals, xs)) outs = _scan_impl_block_unrolled( *consts, *init, *xs, reverse=reverse, length=length_div, num_consts=num_consts, num_carry=num_carry, linear=linear, block_length=unroll, f_impl=f_impl, x_avals=x_avals, y_avals=y_avals) carry, ys = split_list(outs, [num_carry]) if rem > 0: outs = _scan_impl_unrolled( *consts, *carry, *xs_rem, reverse=reverse, length=rem, num_consts=num_consts, num_carry=num_carry, linear=linear, f_impl=f_impl, x_avals=x_avals, y_avals=y_avals) carry, ys_rem = split_list(outs, [num_carry]) if reverse: ys = _map(_concatenate, y_avals, ys_rem, ys) else: ys = _map(_concatenate, y_avals, ys, ys_rem) return (*carry, *ys) def _stack(aval, vals): if aval is core.abstract_unit: return core.unit else: vals = [lax.expand_dims(x, (0,)) for x in vals] return lax.concatenate(vals, 0) def _concatenate(aval, x1, x2): if aval is core.abstract_unit: return core.unit else: return lax.concatenate([x1, x2], 0) def _split_leading_dim(i, aval, x): if aval is core.abstract_unit: return (core.unit, core.unit) else: assert x.ndim >= 1 return (slicing.slice_in_dim(x, 0, i), slicing.slice_in_dim(x, i, x.shape[0])) def _dynamic_index_array(i, aval, x): if aval is core.abstract_unit: return core.unit else: return slicing.dynamic_index_in_dim(x, i, keepdims=False) def _index_array(i, aval, x): if aval is core.abstract_unit: return core.unit else: return slicing.index_in_dim(x, i, keepdims=False) def _empty_array(sz, aval): if aval is core.abstract_unit: return core.unit else: return lax.full((sz,) + aval.shape, 0, aval.dtype) def _update_array(i, aval, xs, x): if aval is core.abstract_unit: return core.unit else: return slicing.dynamic_update_index_in_dim(xs, x, i, 0) def _partition_leading(sz0, sz1, aval, x): if aval is core.abstract_unit: return core.unit else: assert x.ndim >= 1 assert x.shape[0] == sz0 * sz1 return lax.reshape(x, (sz0, sz1, *x.shape[1:])) def _combine_leading(sz0, sz1, aval, x): if aval is core.abstract_unit: return core.unit else: assert x.ndim >= 2 assert x.shape[0] == sz0 assert x.shape[1] == sz1 return lax.collapse(x, 0, 2) def _prepend_dim_to_aval(sz, aval): if aval is core.abstract_unit: return aval elif isinstance(aval, ShapedArray): return aval.update(shape=(sz, *aval.shape), weak_type=False) else: raise TypeError(f'Prepending dim {sz} to aval {aval}') def _scan_abstract_eval(*args, reverse, length, num_consts, num_carry, jaxpr, linear, unroll): carry_avals, y_avals = split_list(jaxpr.out_avals, [num_carry]) ys_avals = _map(partial(_prepend_dim_to_aval, length), y_avals) return carry_avals + ys_avals def _scan_jvp(primals, tangents, reverse, length, jaxpr, num_consts, num_carry, linear, unroll): num_xs = len(jaxpr.in_avals) - num_carry - num_consts num_ys = len(jaxpr.out_avals) - num_carry nonzeros = [type(t) is not ad_util.Zero for t in tangents] const_nz, init_nz, xs_nz = split_list(nonzeros, [num_consts, num_carry]) # Fixpoint computation of which carry are not ad.zero: either # non-zero from init, or the carry out is non-zero. Each iteration promotes # at least one carry to non-zero. We need at most len(carry) iterations, # but we need one last iteration to prepare the jaxpr based on the final # carry_nz. carry_nz = init_nz for _ in range(1 + len(carry_nz)): nonzeros = const_nz + carry_nz + xs_nz jaxpr_jvp, nonzeros_out = ad.jvp_jaxpr( jaxpr, nonzeros, instantiate=carry_nz + [False] * num_ys) carry_nz_out, _ = nonzeros_out[:num_carry], nonzeros_out[num_carry:] if carry_nz_out == carry_nz: break else: carry_nz = _map(operator.or_, carry_nz, carry_nz_out) else: assert False, "Fixpoint not reached" tangents = [ad.instantiate_zeros(t) if nz else t for t, nz in zip(tangents, nonzeros)] consts, init, xs = split_list(primals, [num_consts, num_carry]) all_tangents = split_list(tangents, [num_consts, num_carry]) consts_dot, init_dot, xs_dot = _map(_prune_zeros, all_tangents) jaxpr_jvp_rearranged = ad.rearrange_binders( jaxpr_jvp, [num_consts, num_carry, num_xs], [len(consts_dot), len(init_dot), len(xs_dot)], [num_carry, num_ys], [len(init_dot), sum(nonzeros_out) - len(init_dot)]) consts_linear, init_linear, xs_linear = split_list(linear, [num_consts, num_carry]) jaxpr_jvp_linear = tuple(consts_linear + [True] * len(consts_dot) + init_linear + [True] * len(init_dot) + xs_linear + [True] * len(xs_dot)) out_flat = scan_p.bind( *(consts + consts_dot + init + init_dot + xs + xs_dot), reverse=reverse, length=length, jaxpr=jaxpr_jvp_rearranged, num_consts=num_consts + len(consts_dot), num_carry=num_carry + len(init_dot), linear=jaxpr_jvp_linear, unroll=unroll) carry, carry_dot, ys, ys_dot = split_list(out_flat, [num_carry, len(init_dot), num_ys]) primals_out = carry + ys tangents_out_iter = iter(carry_dot + ys_dot) tangents_out = [next(tangents_out_iter) if nz else ad_util.Zero.from_value(p) for p, nz in zip(primals_out, nonzeros_out)] return primals_out, tangents_out def _prune_zeros(ts): return [t for t in ts if type(t) is not ad_util.Zero] def _scan_partial_eval(trace, *tracers, reverse, length, num_consts, num_carry, jaxpr, linear, unroll): num_ys = len(jaxpr.out_avals) - num_carry unknowns = [t.pval[0] is not None for t in tracers] const_uk, init_uk, xs_uk = split_list(unknowns, [num_consts, num_carry]) # Fixpoint computation of which carry are unknown (not a constant): either # unknown from init, or the carry out is unknown. Each iteration promotes # at least one carry to unknown. We need at most len(carry) iterations, # but we need one last iteration to prepare the jaxpr based on the final # carry_uk. carry_uk = init_uk for _ in range(1 + len(carry_uk)): unknowns = const_uk + carry_uk + xs_uk jaxpr_1, jaxpr_2, out_uk = pe.partial_eval_jaxpr( jaxpr, unknowns, instantiate=carry_uk + [False] * num_ys) carry_uk_out = out_uk[:num_carry] if carry_uk_out == carry_uk: break else: carry_uk = _map(operator.or_, carry_uk, carry_uk_out) else: assert False, "Fixpoint not reached" num_res = len(jaxpr_1.out_avals) - len(jaxpr_2.out_avals) # The residuals are treated as extensive outputs of jaxpr_1 (and extensive # inputs to jaxpr_2), but residuals that are loop-invariant can be hoisted. # TODO(mattjj): hoist other loop-invariant values here too (instantiate=False) invariant_pvals = [pe.PartialVal.known(core.unit if uk else t.pval[1]) for uk, t in zip(unknowns[:num_consts], tracers[:num_consts])] other_pvals = [pe.PartialVal.unknown(a) for a in jaxpr_1.in_avals[num_consts:]] in_pvals_1 = invariant_pvals + other_pvals jaxpr_1_opt, out_pvals_1, consts_1 = pe.trace_to_jaxpr( lu.wrap_init(core.jaxpr_as_fun(jaxpr_1)), in_pvals_1, instantiate=[True] * (num_carry + num_ys) + [False] * num_res) jaxpr_1_opt = pe.ClosedJaxpr(pe.convert_constvars_jaxpr(jaxpr_1_opt), ()) num_consts_1 = num_consts + len(consts_1) # any now-known residuals are intensive, so we want to revise jaxpr_2 to take # those inputs as constants rather than as extensive inputs _, _, res_pvals = split_list(out_pvals_1, [num_carry, num_ys]) intensive_residuals = [const for pv, const in res_pvals if pv is None] move = [False] * len(jaxpr_1.in_avals) + [pv is None for pv, _ in res_pvals] jaxpr_2_opt = pe.move_binders_to_front(jaxpr_2, move) num_consts_2 = num_consts + len(intensive_residuals) # As another optimization, for any extensive inputs that are just forwarded to # extensive outputs, to avoid a copy (looping over dynamic-update-slice) we'd # rather just forward the input tracer. That means pruning some extensive # outputs from the jaxpr here, and updating out_flat below. extensive_invars = jaxpr_1_opt.jaxpr.invars[num_consts_1 + num_carry:] extensive_outvars = jaxpr_1_opt.jaxpr.outvars[num_carry:] extensive_avals = [core.unmapped_aval(length, core.no_axis_name, 0, core.raise_to_shaped(v.aval)) for v in extensive_outvars] fwd_extensive = [num_consts + num_carry + extensive_invars.index(v) if v in extensive_invars else None for v in extensive_outvars] jaxpr_1_opt.jaxpr.outvars = ( jaxpr_1_opt.jaxpr.outvars[:num_carry] + [v for i, v in zip(fwd_extensive, extensive_outvars) if i is None]) in_consts = (list(consts_1) + [core.unit] * num_consts + [core.unit if uk else t.pval[1] for uk, t in zip(unknowns[num_consts:], tracers[num_consts:])]) linear_1 = ([False] * len(consts_1) + [True] * num_consts + [lin or uk for uk, lin in zip(unknowns[num_consts:], linear[num_consts:])]) out_flat = scan_p.bind( *in_consts, reverse=reverse, length=length, jaxpr=jaxpr_1_opt, num_consts=num_consts_1, num_carry=num_carry, linear=tuple(linear_1), unroll=unroll) # Propagate the forwarded extensive outputs using fwd_extensive. Any # numpy.ndarray inputs should be converted to JAX DeviceArrays. out_carry, out_extensive = split_list(out_flat, [num_carry]) out_extensive_iter = iter(out_extensive) out_extensive = [next(out_extensive_iter) if i is None else _maybe_device_put(tracers[i].pval[1]) if tracers[i].is_known() else tracers[i] for i in fwd_extensive] assert all(a.strip_named_shape() == core.raise_to_shaped( core.get_aval(out)).strip_named_shape() for a, out in zip(extensive_avals, out_extensive)) out_flat = out_carry + out_extensive out_carry, ys, res_and_units = split_list(out_flat, [num_carry, num_ys]) extensive_residuals = [r for r, (pv, _) in zip(res_and_units, res_pvals) if pv is not None] new_tracers = [trace.instantiate_const(t) if uk else trace.new_instantiated_literal(core.unit) for uk, t in zip(unknowns, tracers)] carry_avals, y_avals = split_list(jaxpr.out_avals, [num_carry]) ys_avals = _map(partial(_prepend_dim_to_aval, length), y_avals) out_avals = carry_avals + ys_avals out_pvs = [aval if uk else None for aval, uk in zip(out_avals, out_uk)] out_consts = out_carry + ys int_res_tracers = _map(trace.new_instantiated_const, intensive_residuals) ext_res_tracers = _map(trace.new_instantiated_const, extensive_residuals) out_tracers = [pe.JaxprTracer(trace, pe.PartialVal((pv, const)), None) for pv, const in zip(out_pvs, out_consts)] linear_2 = ([False] * len(int_res_tracers) + [lin or not uk for uk, lin in zip(unknowns, linear)] + [False] * len(ext_res_tracers)) eqn = pe.new_eqn_recipe(int_res_tracers + new_tracers + ext_res_tracers, out_tracers, scan_p, dict(reverse=reverse, length=length, jaxpr=jaxpr_2_opt, num_consts=num_consts_2, num_carry=num_carry, linear=tuple(linear_2), unroll=unroll), source_info_util.current()) for t in out_tracers: t.recipe = eqn return out_tracers def _maybe_device_put(x): if isinstance(x, np.ndarray): return lax._device_put_raw(x) else: return x def _scan_transpose(reduce_axes, cts, *args, reverse, length, num_consts, num_carry, jaxpr, linear, unroll): # we've only implemented transposing scans with specific lin/nonlin patterns consts_lin, init_lin, xs_lin = split_list(linear, [num_consts, num_carry]) num_ires = len(consts_lin) - sum(consts_lin) num_eres = len(xs_lin) - sum(xs_lin) if consts_lin != [False] * num_ires + [True] * (len(consts_lin) - num_ires): raise NotImplementedError if xs_lin != [True] * (len(xs_lin) - num_eres) + [False] * num_eres: raise NotImplementedError if not all(init_lin): pass # TODO(mattjj): error check https://github.com/google/jax/issues/1963 consts, _, xs = split_list(args, [num_consts, num_carry]) ires, _ = split_list(consts, [num_ires]) _, eres = split_list(xs, [sum(xs_lin)]) assert not any(ad.is_undefined_primal(r) for r in ires) assert not any(ad.is_undefined_primal(r) for r in eres) carry_avals, y_avals = split_list(jaxpr.out_avals, [num_carry]) ys_avals = _map(partial(_prepend_dim_to_aval, length), y_avals) ct_carry, ct_ys = split_list(cts, [num_carry]) ct_carry = _map(ad.instantiate_zeros_aval, carry_avals, ct_carry) ct_ys = _map(ad.instantiate_zeros_aval, ys_avals, ct_ys) ct_consts = _map(ad_util.zeros_like_aval, jaxpr.in_avals[num_ires:num_consts]) # jaxpr :: [ires, T d] -> [T c] -> [T a, eres] -> ([T c], [T b]) # jaxpr_trans :: [ires] -> [CT d, CT c] -> [CT b, eres] -> ([CT d, CT c], [CT a]) jaxpr_trans = _transpose_scan_jaxpr( num_ires, num_consts - num_ires, num_eres, jaxpr, reduce_axes) linear_trans = ([False] * num_ires + [True] * (len(ct_consts) + len(ct_carry) + len(ct_ys)) + [False] * num_eres) outs = scan_p.bind( *(ires + ct_consts + ct_carry + ct_ys + eres), reverse=not reverse, length=length, jaxpr=jaxpr_trans, num_consts=num_ires, num_carry=num_consts-num_ires+num_carry, linear=tuple(linear_trans), unroll=unroll) ct_consts, ct_init, ct_xs = split_list(outs, [num_consts - num_ires, num_carry]) return [None] * num_ires + ct_consts + ct_init + ct_xs + [None] * num_eres # transpose_scan_jaxpr :: ([res1, c, a, res2] -> b) # -> ([res1, CT c, CT b, res2] -> [CT c, CT a]) def _transpose_scan_jaxpr(num_res1, num_c, num_res2, jaxpr, reduce_axes): num_a = len(jaxpr.in_avals) - num_res1 - num_c - num_res2 # TODO: allow input cotangent avals to be batched relative to jaxpr.in_avals # if an axis isn't reduced res1_avals, c_avals, a_avals, res2_avals = split_list( jaxpr.in_avals, [num_res1, num_c, num_a]) num_b = len(jaxpr.out_avals) b_avals = list(jaxpr.out_avals) @lu.wrap_init def transposed(*res1_cbar_bbar_res2): res1, c_bar, b_bar, res2 = split_list( res1_cbar_bbar_res2, [num_res1, num_c, num_b]) primals = (res1 + [ad.UndefinedPrimal(aval) for aval in c_avals] + [ad.UndefinedPrimal(aval) for aval in a_avals] + res2) cbar_abar = ad.backward_pass(jaxpr.jaxpr, reduce_axes, jaxpr.consts, primals, b_bar) _, new_c_bar, a_bar, _ = split_list(cbar_abar, [num_res1, num_c, num_a]) a_bar = _map(ad.instantiate_zeros_aval, a_avals, a_bar) c_bar = _map(ad.instantiate_zeros_aval, c_avals, _map(ad.add_tangents, c_bar, new_c_bar)) return c_bar + a_bar return _make_closed_jaxpr(transposed, res1_avals + c_avals + b_avals + res2_avals) def _make_closed_jaxpr(traceable: lu.WrappedFun, in_avals: Sequence[core.AbstractValue]): jaxpr, _, consts = pe.trace_to_jaxpr_dynamic(traceable, in_avals) return core.ClosedJaxpr(jaxpr, consts) def _scan_batching_rule(axis_size, axis_name, main_type, args, dims, reverse, length, jaxpr, num_consts, num_carry, linear, unroll): num_ys = len(jaxpr.out_avals) - num_carry orig_batched = [d is not batching.not_mapped for d in dims] const_batched, init_batched, xs_batched = split_list(orig_batched, [num_consts, num_carry]) # Fixpoint computation of which carry are batched: either # batched from init, or the carry out is batched. Each iteration promotes # at least one carry to batched. We need at most len(carry) iterations, # but we need one last iteration to prepare the jaxpr based on the final # carry_batched. carry_batched = init_batched for _ in range(1 + len(carry_batched)): batched = const_batched + carry_batched + xs_batched jaxpr_batched, batched_out = batching.batch_jaxpr( jaxpr, axis_size, batched, instantiate=carry_batched + [False] * num_ys, axis_name=axis_name, main_type=main_type) carry_batched_out, ys_batched = batched_out[:num_carry], batched_out[num_carry:] if carry_batched_out == carry_batched: break else: carry_batched = _map(operator.or_, carry_batched, carry_batched_out) else: assert False, "Fixpoint not reached" consts, init, xs = split_list(args, [num_consts, num_carry]) consts_bdims, init_bdims, xs_bdims = split_list(dims, [num_consts, num_carry]) new_consts = [batching.moveaxis(x, d, 0) if d is not batching.not_mapped and d != 0 else x for x, d in zip(consts, consts_bdims)] new_init = [batching.broadcast(x, axis_size, 0) if now_batched and not was_batched else batching.moveaxis(x, d, 0) if now_batched else x for x, d, was_batched, now_batched in zip(init, init_bdims, init_batched, carry_batched)] new_xs = [batching.moveaxis(x, d, 1) if d is not batching.not_mapped and d != 1 else x for x, d in zip(xs, xs_bdims)] new_args = new_consts + new_init + new_xs outs = scan_p.bind( *new_args, reverse=reverse, length=length, jaxpr=jaxpr_batched, num_consts=num_consts, num_carry=num_carry, linear=linear, unroll=unroll) carry_bdims = [0 if b else batching.not_mapped for b in carry_batched] ys_bdims = [1 if b else batching.not_mapped for b in ys_batched] return outs, carry_bdims + ys_bdims def _scan_masking_rule(padded_vals, logical_shapes, reverse, length, jaxpr, num_consts, num_carry, linear, unroll): dynamic_length, = masking.shape_as_value((length,)) masked_jaxpr = _masked_scan_jaxpr(jaxpr, num_consts, num_carry) consts, init, xs = split_list(padded_vals, [num_consts, num_carry]) max_length, = {x.shape[0] for x in xs} const_linear, init_linear, xs_linear = split_list(linear, [num_consts, num_carry]) dynamic_length = lax.convert_element_type(dynamic_length, dtypes.int_) out_vals = scan_p.bind(dynamic_length, *consts, dtypes.int_(0), *init, *xs, reverse=reverse, length=max_length, jaxpr=masked_jaxpr, num_consts=1 + num_consts, num_carry=1 + num_carry, linear=tuple([False] + const_linear + [False] + init_linear + xs_linear), unroll=unroll) return out_vals[1:] def _masked_scan_jaxpr(jaxpr, num_consts, num_carry): fun = core.jaxpr_as_fun(jaxpr) @lu.wrap_init def masked(*args): [dynamic_length], consts, [i], carry, xs = split_list( args, [1, num_consts, 1, num_carry]) out = fun(*(consts + carry + xs)) new_carry, ys = split_list(out, [num_carry]) new_carry = [lax.select(i < dynamic_length, new_c, c) for new_c, c in zip(new_carry, carry)] return [i + 1] + new_carry + ys aval = ShapedArray((), dtypes.int_) const_avals, carry_avals, x_avals = split_list(jaxpr.in_avals, [num_consts, num_carry]) return _make_closed_jaxpr(masked, [aval] + const_avals + [aval] + carry_avals + x_avals) def _scan_typecheck(bind_time, *avals, reverse, length, num_consts, num_carry, jaxpr, linear, unroll): tc = partial(_typecheck_param, 'scan') tc(reverse, 'reverse', 'bool', type(reverse) is bool) tc(num_consts, 'num_consts', 'non-negative int', type(num_consts) is int and num_consts >= 0) tc(num_carry, 'num_carry', 'non-negative int', type(num_carry) is int and num_carry >= 0) tc(jaxpr, 'jaxpr', 'ClosedJaxpr', type(jaxpr) is core.ClosedJaxpr) tc(linear, 'linear', 'tuple of bool', type(linear) is tuple and all(type(x) is bool for x in linear)) tc(unroll, 'unroll', 'positive int', type(unroll) is int and unroll > 0) length_types = (int, masking.Poly) if bind_time else (int,) tc(length, 'length', 'non-negative int', type(length) in length_types and length >= 0) core.typecheck_assert( len(linear) == len(avals), f'scan param linear has length {len(linear)} for {len(avals)} operands') const_avals, init_avals, x_avals = split_list(avals, [num_consts, num_carry]) const_avals_jaxpr, init_avals_jaxpr, x_avals_jaxpr = split_list( jaxpr.in_avals, [num_consts, num_carry]) carry_avals_jaxpr, _ = split_list(jaxpr.out_avals, [num_carry]) x_avals_mapped = _map(partial(core.mapped_aval, length, 0), x_avals) core.typecheck_assert( all(_map(core.typematch, init_avals_jaxpr, carry_avals_jaxpr)), f'scan input carry input and output types mismatch: ' f'\n{_avals_short(init_avals_jaxpr)}\nvs\n{_avals_short(carry_avals_jaxpr)}') core.typecheck_assert( all(_map(core.typecompat, const_avals_jaxpr, const_avals)), f'scan jaxpr takes input const types\n{_avals_short(const_avals_jaxpr)},\n' f'called with consts of type\n{_avals_short(const_avals)}') core.typecheck_assert( all(_map(core.typecompat, init_avals_jaxpr, init_avals)), f'scan jaxpr takes input carry types\n{_avals_short(init_avals_jaxpr)},\n' f'called with initial carry of type\n{_avals_short(init_avals)}') core.typecheck_assert( all(_map(core.typecompat, x_avals_jaxpr, x_avals_mapped)), f'scan jaxpr takes input sequence types\n{_avals_short(x_avals_jaxpr)},\n' f'called with sequence of type\n{_avals_short(x_avals)}') def scan_bind(*args, **params): if config.jax_enable_checks: avals = _map(core.get_aval, args) _scan_typecheck(True, *avals, **params) core.check_jaxpr(params['jaxpr'].jaxpr) return core.AxisPrimitive.bind(scan_p, *args, **params) scan_p = core.AxisPrimitive("scan") scan_p.multiple_results = True scan_p.def_custom_bind(scan_bind) scan_p.def_impl(partial(xla.apply_primitive, scan_p)) scan_p.def_abstract_eval(_scan_abstract_eval) ad.primitive_jvps[scan_p] = _scan_jvp ad.reducing_transposes[scan_p] = _scan_transpose pe.custom_partial_eval_rules[scan_p] = _scan_partial_eval xla.register_translation(scan_p, xla.lower_fun(_scan_impl, new_style=True, multiple_results=True), initial_style=True) batching.axis_primitive_batchers[scan_p] = _scan_batching_rule masking.masking_rules[scan_p] = _scan_masking_rule core.custom_typechecks[scan_p] = partial(_scan_typecheck, False) pe.partial_eval_jaxpr_custom_rules[scan_p] = pe.partial_eval_jaxpr_custom_rule_not_implemented mlir.register_lowering(scan_p, mlir.lower_fun(_scan_impl, multiple_results=True))
[docs]@api_boundary def map(f, xs): """Map a function over leading array axes. Like Python's builtin map, except inputs and outputs are in the form of stacked arrays. Consider using the ``jax.vmap`` transform instead, unless you need to apply a function element by element for reduced memory usage or heterogeneous computation with other control flow primitives. When ``xs`` is an array type, the semantics of ``map`` are given by this Python implementation:: def map(f, xs): return np.stack([f(x) for x in xs]) Like ``scan``, ``map`` is implemented in terms of JAX primitives so many of the same advantages over a Python loop apply: ``xs`` may be an arbitrary nested pytree type, and the mapped computation is compiled only once. Args: f: a Python function to apply element-wise over the first axis or axes of ``xs``. xs: values over which to map along the leading axis. Returns: Mapped values. """ g = lambda _, x: ((), f(x)) _, ys = scan(g, (), xs) return ys
def _concat_masking_rule(padded_vals, logical_shapes, dimension): result = lax.concatenate(padded_vals, dimension) # fragmented offset = 0 for padded_val, logical_shape in zip(padded_vals, logical_shapes): result = _memcpy(dimension, logical_shape[dimension], padded_val, result, offset) offset = offset + logical_shape[dimension] return result def _memcpy(axis, num, src, dst, offset): def body(i, dst): update = slicing.dynamic_index_in_dim(src, i, axis) return slicing.dynamic_update_index_in_dim(dst, update, i + offset, axis) return fori_loop(0, num, body, dst) masking.masking_rules[lax.concatenate_p] = _concat_masking_rule # type: ignore def _rng_bit_generator_batching_rule(batched_args, batch_dims, *, shape, dtype, algorithm): """Calls RBG in a loop and stacks the results.""" key, = batched_args bd, = batch_dims if bd is batching.not_mapped: return lax.rng_bit_generator_p.bind(key, shape=shape, dtype=dtype, algorithm=algorithm), (None, None) key = batching.moveaxis(key, bd, 0) map_body = lambda k: lax.rng_bit_generator_p.bind(k, shape=shape, dtype=dtype, algorithm=algorithm) stacked_keys, stacked_bits = map(map_body, key) return (stacked_keys, stacked_bits), (0, 0) batching.primitive_batchers[lax.rng_bit_generator_p] = _rng_bit_generator_batching_rule def _show_diff(array1, array2): if core.typematch(array1, array2): return f"{array1}" return f"DIFFERENT {array1} vs. {array2}" def _check_tree_and_avals(what, tree1, avals1, tree2, avals2): """Raises TypeError if (tree1, avals1) does not match (tree2, avals2). Corresponding `tree` and `avals` must match in the sense that the number of leaves in `tree` must be equal to the length of `avals`. `what` will be prepended to details of the mismatch in TypeError. """ if tree1 != tree2: raise TypeError( f"{what} must have same type structure, got {tree1} and {tree2}.") if not all(_map(core.typematch, avals1, avals2)): diff = tree_multimap(_show_diff, tree_unflatten(tree1, avals1), tree_unflatten(tree2, avals2)) raise TypeError(f"{what} must have identical types, got\n{diff}.") def _check_tree(func_name, expected_name, actual_tree, expected_tree, has_aux=False): if has_aux: actual_tree_children = actual_tree.children() if len(actual_tree_children) == 2: # select first child as result tree actual_tree = tree_structure(actual_tree_children[0]) else: raise ValueError( f"{func_name}() produced a pytree with structure " f"{actual_tree}, but a pytree tuple with auxiliary " f"output was expected because has_aux was set to True.") if actual_tree != expected_tree: raise TypeError( f"{func_name}() output pytree structure must match {expected_name}, " f"got {actual_tree} and {expected_tree}.") def _promote_weak_typed_inputs(in_vals, in_avals, out_avals): """Promote weakly-typed in_vals to be compatible with out_avals. Args: in_vals : flattened list of input values. in_avals : corresponding list of avals. out_avals : list of target output avals. Returns: in_vals_new : flattened list of modified in_vals with no weak types. changed : bool; true if in_vals required modification. """ if len(in_vals) != len(in_avals) or len(in_avals) != len(out_avals): # Calling function is responsible for catching this. return in_vals, False weak_mismatches = [i for i, (a1, a2) in enumerate(zip(in_avals, out_avals)) if getattr(a1, 'weak_type', False) and not core.typematch(a1, a2)] if not weak_mismatches: return in_vals, False for i in weak_mismatches: new_dtype = dtypes.result_type(in_vals[i], out_avals[i]) in_vals[i] = lax.convert_element_type(in_vals[i], new_dtype) return in_vals, True def _stop_gradient_fun(f): """Create a version of f() that stops all gradients.""" def wrapper(*args, **kwargs): args_flat, in_args_tree = tree_flatten((args, kwargs)) args_avals = tuple(_map(_abstractify, args_flat)) g = lambda a, b: f(*a, **b) jaxpr, consts, out_tree = _initial_style_jaxpr(g, in_args_tree, args_avals) all_args = _map(lax.stop_gradient, (*consts, *args_flat)) out = core.jaxpr_as_fun(jaxpr)(*all_args) return tree_unflatten(out_tree, out) return wrapper _RootTuple = collections.namedtuple('_RootTuple', 'f, solve, l_and_s') def _split_root_args(args, const_lengths): params_list = split_list(args, list(const_lengths)) return _RootTuple(*params_list[:-1]), params_list[-1]
[docs]@api_boundary def custom_root(f, initial_guess, solve, tangent_solve, has_aux=False): """Differentiably solve for a roots of a function. This is a low-level routine, mostly intended for internal use in JAX. Gradients of custom_root() are defined with respect to closed-over variables from the provided function ``f`` via the implicit function theorem: https://en.wikipedia.org/wiki/Implicit_function_theorem Args: f: function for which to find a root. Should accept a single argument, return a tree of arrays with the same structure as its input. initial_guess: initial guess for a zero of f. solve: function to solve for the roots of f. Should take two positional arguments, f and initial_guess, and return a solution with the same structure as initial_guess such that func(solution) = 0. In other words, the following is assumed to be true (but not checked):: solution = solve(f, initial_guess) error = f(solution) assert all(error == 0) tangent_solve: function to solve the tangent system. Should take two positional arguments, a linear function ``g`` (the function ``f`` linearized at its root) and a tree of array(s) ``y`` with the same structure as initial_guess, and return a solution ``x`` such that ``g(x)=y``: - For scalar ``y``, use ``lambda g, y: y / g(1.0)``. - For vector ``y``, you could use a linear solve with the Jacobian, if dimensionality of ``y`` is not too large: ``lambda g, y: np.linalg.solve(jacobian(g)(y), y)``. has_aux: bool indicating whether the ``solve`` function returns auxiliary data like solver diagnostics as a second argument. Returns: The result of calling solve(f, initial_guess) with gradients defined via implicit differentiation assuming ``f(solve(f, initial_guess)) == 0``. """ guess_flat, in_args_tree = tree_flatten((initial_guess,)) guess_avals = tuple(_map(_abstractify, guess_flat)) f_jaxpr, f_consts, out_tree = _initial_style_jaxpr( f, in_args_tree, guess_avals) in_tree, = treedef_children(in_args_tree) _check_tree("f", "initial_guess", out_tree, in_tree, False) solve_jaxpr, solve_consts, solution_tree = _initial_style_jaxpr( partial(solve, _stop_gradient_fun(f)), in_args_tree, guess_avals) _check_tree("solve", "initial_guess", solution_tree, in_tree, has_aux) def linearize_and_solve(x, b): unchecked_zeros, f_jvp = jax.linearize(f, x) return tangent_solve(f_jvp, b) l_and_s_jaxpr, l_and_s_consts, out_tree = _initial_style_jaxpr( linearize_and_solve, treedef_tuple((in_tree,) * 2), guess_avals * 2) _check_tree("tangent_solve", "x", out_tree, in_tree, False) all_consts = [f_consts, solve_consts, l_and_s_consts] const_lengths = _RootTuple(*_map(len, all_consts)) jaxprs = _RootTuple(f_jaxpr, solve_jaxpr, l_and_s_jaxpr) solution_flat = _custom_root( const_lengths, jaxprs, *(_flatten(all_consts) + guess_flat)) return tree_unflatten(solution_tree, solution_flat)
@partial(jax.custom_jvp, nondiff_argnums=(0, 1)) def _custom_root(const_lengths, jaxprs, *args): params, initial_guess = _split_root_args(args, const_lengths) solution = core.jaxpr_as_fun(jaxprs.solve)(*(params.solve + initial_guess)) return solution @_custom_root.defjvp def _root_jvp(const_lengths, jaxprs, primals, tangents): params, _ = _split_root_args(primals, const_lengths) sol = _custom_root(const_lengths, jaxprs, *primals) f_out_vals = len(jaxprs.f.out_avals) solution, aux = split_list(sol, [f_out_vals]) params_dot, _ = _split_root_args(tangents, const_lengths) # F(m, u) = 0 # system of equations in u, parameterized by m # # solution is u*(m) defined in a neighborhood # F(m, u*(m)) = 0 # satisfied in a neighborhood # # ∂_0 F(m, u*(m)) + ∂_1 F(m, u*(m)) ∂ u*(m) = 0 # implied by line above # ∂ u*(m) = - (∂_1 F(m, u*(m)))^{-1} ∂_0 F(m, u*(m)) # rearrange # # ∂ u*(m)[v] = - (∂_1 F(m, u*(m)))^{-1} [∂_0 F(m, u*(m))[v]] # jvp f = core.jaxpr_as_fun(jaxprs.f) linearize_and_solve = partial( core.jaxpr_as_fun(jaxprs.l_and_s), *params.l_and_s) f_at_solution = lambda *params: f(*params, *solution) _, rhs = ad.jvp(lu.wrap_init(f_at_solution)).call_wrapped( params.f, params_dot.f) solution_dot = _map( operator.neg, linearize_and_solve(*solution, *rhs)) # append aux, create symbolic zero tangents for the aux values solution += aux solution_dot += _map(lax.zeros_like_array, aux) return solution, solution_dot class _LinearSolveTuple(collections.namedtuple( '_LinearSolveTuple', 'matvec, vecmat, solve, transpose_solve')): def transpose(self): return type(self)(self.vecmat, self.matvec, self.transpose_solve, self.solve) def _split_linear_solve_args(args, const_lengths): params_list = split_list(args, list(const_lengths)) return _LinearSolveTuple(*params_list[:-1]), params_list[-1] def _transpose_one_output(linear_fun, primals): transpose_fun = jax.linear_transpose(linear_fun, primals) def transposed_fun(x): (y,) = transpose_fun(x) return y return transposed_fun def _flatten(args): return [x for arg in args for x in arg] def _check_shapes(func_name, expected_name, actual, expected): actual_shapes = _map(np.shape, tree_leaves(actual)) expected_shapes = _map(np.shape, tree_leaves(expected)) if actual_shapes != expected_shapes: raise ValueError( f"{func_name}() output shapes must match {expected_name}, " f"got {actual_shapes} and {expected_shapes}")
[docs]@api_boundary def custom_linear_solve( matvec, b, solve, transpose_solve=None, symmetric=False, has_aux=False): """Perform a matrix-free linear solve with implicitly defined gradients. This function allows for overriding or defining gradients for a linear solve directly via implicit differentiation at the solution, rather than by differentiating *through* the solve operation. This can sometimes be much faster or more numerically stable, or differentiating through the solve operation may not even be implemented (e.g., if ``solve`` uses ``lax.while_loop``). Required invariant:: x = solve(matvec, b) # solve the linear equation assert matvec(x) == b # not checked Args: matvec: linear function to invert. Must be differentiable. b: constant right handle side of the equation. May be any nested structure of arrays. solve: higher level function that solves for solution to the linear equation, i.e., ``solve(matvec, x) == x`` for all ``x`` of the same form as ``b``. This function need not be differentiable. transpose_solve: higher level function for solving the transpose linear equation, i.e., ``transpose_solve(vecmat, x) == x``, where ``vecmat`` is the transpose of the linear map ``matvec`` (computed automatically with autodiff). Required for backwards mode automatic differentiation, unless ``symmetric=True``, in which case ``solve`` provides the default value. symmetric: bool indicating if it is safe to assume the linear map corresponds to a symmetric matrix, i.e., ``matvec == vecmat``. has_aux: bool indicating whether the ``solve`` and ``transpose_solve`` functions return auxiliary data like solver diagnostics as a second argument. Returns: Result of ``solve(matvec, b)``, with gradients defined assuming that the solution ``x`` satisfies the linear equation ``matvec(x) == b``. """ if transpose_solve is None and symmetric: transpose_solve = solve b_flat, in_args_tree = tree_flatten((b,)) b_avals = tuple(_map(_abstractify, b_flat)) tree, = treedef_children(in_args_tree) def _shape_checked(fun, name, has_aux): def f(x): y = fun(x) _check_shapes(name, "b", y, b_flat) return y def f_aux(x): y, aux = fun(x) _check_shapes(name, "b", y, b_flat) return y, aux return f_aux if has_aux else f # no auxiliary data assumed for matvec matvec_jaxpr, matvec_consts, out_tree = _initial_style_jaxpr( _shape_checked(matvec, "matvec", False), in_args_tree, b_avals, 'custom_linear_solve') _check_tree("matvec", "b", out_tree, tree, False) solve_jaxpr, solve_consts, out_tree = _initial_style_jaxpr( _shape_checked(partial(solve, matvec), "solve", has_aux), in_args_tree, b_avals, 'custom_linear_solve') _check_tree("solve", "b", out_tree, tree, has_aux) if transpose_solve is None: vecmat_jaxpr = tr_solve_jaxpr = None vecmat_consts = tr_solve_consts = [] else: if symmetric: vecmat = matvec vecmat_jaxpr = matvec_jaxpr vecmat_consts = matvec_consts else: vecmat = _transpose_one_output(matvec, b) vecmat_jaxpr, vecmat_consts, out_tree = _initial_style_jaxpr( vecmat, in_args_tree, b_avals, 'custom_linear_solve') assert out_tree == tree tr_solve_jaxpr, tr_solve_consts, out_tree = _initial_style_jaxpr( _shape_checked(partial(transpose_solve, vecmat), "transpose_solve", has_aux), in_args_tree, b_avals, 'custom_linear_solve') _check_tree("transpose_solve", "b", out_tree, tree, has_aux) all_consts = [matvec_consts, vecmat_consts, solve_consts, tr_solve_consts] const_lengths = _LinearSolveTuple(*_map(len, all_consts)) jaxprs = _LinearSolveTuple( matvec_jaxpr, vecmat_jaxpr, solve_jaxpr, tr_solve_jaxpr) out_flat = linear_solve_p.bind( *(_flatten(all_consts) + b_flat), const_lengths=const_lengths, jaxprs=jaxprs) return tree_unflatten(out_tree, out_flat)
def _linear_solve_abstract_eval(*args, const_lengths, jaxprs): args_to_raise = args[sum(const_lengths):] # raise aux_args to shaped arrays as well if present # number of aux args is the difference in out_avals # of solve and matvec (since they map to the same vector space) num_aux = len(jaxprs.solve.out_avals) - len(jaxprs.matvec.out_avals) if num_aux > 0: args_to_raise += tuple(jaxprs.solve.out_avals[-num_aux:]) return _map(raise_to_shaped, args_to_raise) def _custom_linear_solve_impl(*args, const_lengths, jaxprs): params, b = _split_linear_solve_args(args, const_lengths) x = core.jaxpr_as_fun(jaxprs.solve)(*(params.solve + b)) return x def _tangent_linear_map(func, params, params_dot, *x): """Compute the tangent of a linear map. Assuming ``func(*params, *x)`` is linear in ``x`` and computes ``A @ x``, this function computes ``∂A @ x``. """ assert any(type(p) is not ad_util.Zero for p in params_dot) zeros = _map(ad_util.Zero.from_value, x) _, out_tangent = ad.jvp(lu.wrap_init(func)).call_wrapped( params + list(x), params_dot + zeros) return out_tangent def _custom_linear_solve_jvp(primals, tangents, const_lengths, jaxprs): # A x - b = 0 # ∂A x + A ∂x - ∂b = 0 # ∂x = A^{-1} (∂b - ∂A x) kwargs = dict(const_lengths=const_lengths, jaxprs=jaxprs) x = linear_solve_p.bind(*primals, **kwargs) params, _ = _split_linear_solve_args(primals, const_lengths) params_dot, b_dot = _split_linear_solve_args(tangents, const_lengths) num_x_leaves = len(b_dot) # x is a flat tree with possible aux values appended # since x_tree == b_tree == b_dot_tree, we can cut off # aux values with len info provided by b_dot tree here x_leaves, _ = split_list(x, [num_x_leaves]) if all(type(p) is ad_util.Zero for p in params_dot.matvec): # no need to evaluate matvec_tangents rhs = b_dot else: matvec_tangents = _tangent_linear_map( core.jaxpr_as_fun(jaxprs.matvec), params.matvec, params_dot.matvec, *x_leaves) rhs = _map(ad.add_tangents, b_dot, _map(operator.neg, matvec_tangents)) x_dot = linear_solve_p.bind(*(_flatten(params) + rhs), **kwargs) # split into x tangents and aux tangents (these become zero) dx_leaves, daux_leaves = split_list(x_dot, [num_x_leaves]) daux_leaves = _map(ad_util.Zero.from_value, daux_leaves) x_dot = dx_leaves + daux_leaves return x, x_dot def _linear_solve_transpose_rule(cotangent, *primals, const_lengths, jaxprs): if jaxprs.transpose_solve is None: raise TypeError('transpose_solve required for backwards mode automatic ' 'differentiation of custom_linear_solve') params, b = _split_linear_solve_args(primals, const_lengths) # split off symbolic zeros in the cotangent if present x_cotangent, _ = split_list(cotangent, [len(b)]) assert all(ad.is_undefined_primal(x) for x in b) cotangent_b_full = linear_solve_p.bind( *(_flatten(params.transpose()) + x_cotangent), const_lengths=const_lengths.transpose(), jaxprs=jaxprs.transpose()) # drop aux values in cotangent computation cotangent_b, _ = split_list(cotangent_b_full, [len(b)]) return [None] * sum(const_lengths) + cotangent_b def _linear_solve_batching_rule(axis_size, axis_name, main_type, args, dims, const_lengths, jaxprs): orig_bat = [d is not batching.not_mapped for d in dims] params, b = _split_linear_solve_args(args, const_lengths) params_dims, b_dims = _split_linear_solve_args(dims, const_lengths) params_bat, orig_b_bat = _split_linear_solve_args(orig_bat, const_lengths) (matvec, vecmat, solve, solve_t) = jaxprs (matvec_bat, vecmat_bat, solve_bat, solve_t_bat) = params_bat num_aux = len(solve.out_avals) - len(matvec.out_avals) # Fixpoint computation of which parts of x and b are batched; we need to # ensure this is consistent between all four jaxprs b_bat = orig_b_bat x_bat = [False] * len(solve.out_avals) for i in range(1 + len(orig_b_bat) + len(solve.out_avals)): # Apply vecmat and solve -> new batched parts of x solve_jaxpr_batched, solve_x_bat = batching.batch_jaxpr( solve, axis_size, solve_bat + b_bat, instantiate=x_bat, axis_name=axis_name, main_type=main_type) if vecmat is None: vecmat_jaxpr_batched = None x_bat_out = solve_x_bat else: vecmat_jaxpr_batched, vecmat_x_bat = batching.batch_jaxpr( vecmat, axis_size, vecmat_bat + b_bat, instantiate=x_bat, axis_name=axis_name, main_type=main_type) # batch all aux data by default x_bat_out = _map(operator.or_, vecmat_x_bat + [True] * num_aux, solve_x_bat) # Apply matvec and solve_t -> new batched parts of b matvec_jaxpr_batched, matvec_b_bat = batching.batch_jaxpr( matvec, axis_size, matvec_bat + x_bat_out, instantiate=b_bat, axis_name=axis_name, main_type=main_type) if solve_t is None: solve_t_jaxpr_batched = None b_bat_out = _map(operator.or_, matvec_b_bat, orig_b_bat) else: solve_t_jaxpr_batched, solve_t_b_aux_bat = batching.batch_jaxpr( solve_t, axis_size, solve_t_bat + x_bat_out, instantiate=b_bat, axis_name=axis_name, main_type=main_type) assert len(solve_t_b_aux_bat) == len(orig_b_bat) + num_aux solve_t_b_bat, _ = split_list(solve_t_b_aux_bat, [len(orig_b_bat)]) b_bat_out = _map(lambda m, s, o: m or s or o, matvec_b_bat, solve_t_b_bat, orig_b_bat) if x_bat_out == x_bat and b_bat_out == b_bat: break else: x_bat = x_bat_out b_bat = b_bat_out else: assert False, "Fixedpoint not reached" batched_jaxprs = _LinearSolveTuple(matvec_jaxpr_batched, vecmat_jaxpr_batched, solve_jaxpr_batched, solve_t_jaxpr_batched) # Move batched axes to the front new_params = [ batching.moveaxis(x, d, 0) if d is not batching.not_mapped and d != 0 else x for x, d in zip(_flatten(params), _flatten(params_dims)) ] # Broadcast out b if necessary new_b = [ batching.broadcast(x, axis_size, 0) if now_bat and not was_bat else batching.moveaxis(x, d, 0) if now_bat and d != 0 else x for x, d, was_bat, now_bat in zip(b, b_dims, orig_b_bat, b_bat) ] outs = linear_solve_p.bind( *(new_params + new_b), const_lengths=const_lengths, jaxprs=batched_jaxprs) out_dims = [0 if batched else batching.not_mapped for batched in solve_x_bat] return outs, out_dims linear_solve_p = core.AxisPrimitive('custom_linear_solve') linear_solve_p.multiple_results = True linear_solve_p.def_impl(_custom_linear_solve_impl) linear_solve_p.def_abstract_eval(_linear_solve_abstract_eval) ad.primitive_jvps[linear_solve_p] = _custom_linear_solve_jvp xla.register_translation( linear_solve_p, xla.lower_fun(_custom_linear_solve_impl, new_style=True, multiple_results=True), initial_style=True) ad.primitive_transposes[linear_solve_p] = _linear_solve_transpose_rule batching.axis_primitive_batchers[linear_solve_p] = _linear_solve_batching_rule pe.partial_eval_jaxpr_custom_rules[linear_solve_p] = pe.partial_eval_jaxpr_custom_rule_not_implemented def _interleave(a, b, axis): """Given two Tensors of static shape, interleave them along the first axis.""" assert a.shape[axis] == b.shape[axis] or a.shape[axis] == b.shape[axis] + 1 a_pad = [(0, 0, 0)] * a.ndim b_pad = [(0, 0, 0)] * b.ndim a_pad[axis] = (0, 1 if a.shape[axis] == b.shape[axis] else 0, 1) b_pad[axis] = (1, 0 if a.shape[axis] == b.shape[axis] else 1, 1) op = lax.bitwise_or if a.dtype == np.bool_ else lax.add return op(lax.pad(a, lax._const(a, 0), a_pad), lax.pad(b, lax._const(b, 0), b_pad))
[docs]@api_boundary def associative_scan(fn: Callable, elems, reverse: bool = False, axis: int = 0): """Performs a scan with an associative binary operation, in parallel. For an introduction to associative scans, see [BLE1990]_. Args: fn: A Python callable implementing an associative binary operation with signature ``r = fn(a, b)``. Function `fn` must be associative, i.e., it must satisfy the equation ``fn(a, fn(b, c)) == fn(fn(a, b), c)``. The inputs and result are (possibly nested Python tree structures of) array(s) matching ``elems``. Each array has a dimension in place of the ``axis`` dimension. `fn` should be applied elementwise over the ``axis`` dimension (for example, by using :func:`jax.vmap` over the elementwise function.) The result ``r`` has the same shape (and structure) as the two inputs ``a`` and ``b``. elems: A (possibly nested Python tree structure of) array(s), each with an ``axis`` dimension of size ``num_elems``. reverse: A boolean stating if the scan should be reversed with respect to the ``axis`` dimension. axis: an integer identifying the axis over which the scan should occur. Returns: A (possibly nested Python tree structure of) array(s) of the same shape and structure as ``elems``, in which the ``k``'th element of ``axis`` is the result of recursively applying ``fn`` to combine the first ``k`` elements of ``elems`` along ``axis``. For example, given ``elems = [a, b, c, ...]``, the result would be ``[a, fn(a, b), fn(fn(a, b), c), ...]``. Example 1: partial sums of an array of numbers: >>> lax.associative_scan(jnp.add, jnp.arange(0, 4)) DeviceArray([0, 1, 3, 6], dtype=int32) Example 2: partial products of an array of matrices >>> mats = jax.random.uniform(jax.random.PRNGKey(0), (4, 2, 2)) >>> partial_prods = lax.associative_scan(jnp.matmul, mats) >>> partial_prods.shape (4, 2, 2) Example 3: reversed partial sums of an array of numbers >>> lax.associative_scan(jnp.add, jnp.arange(0, 4), reverse=True) DeviceArray([6, 6, 5, 3], dtype=int32) .. [BLE1990] Blelloch, Guy E. 1990. "Prefix Sums and Their Applications.", Technical Report CMU-CS-90-190, School of Computer Science, Carnegie Mellon University. """ elems_flat, tree = tree_flatten(elems) if reverse: elems_flat = [lax.rev(elem, [axis]) for elem in elems_flat] def combine(a_flat, b_flat): # Lower `fn` to operate on flattened sequences of elems. a = tree_unflatten(tree, a_flat) b = tree_unflatten(tree, b_flat) c = fn(a, b) c_flat, _ = tree_flatten(c) return c_flat # Check that all inputs have a consistent leading dimension `num_elems`. axis = util.canonicalize_axis(axis, elems_flat[0].ndim) num_elems = int(elems_flat[0].shape[axis]) if not all(int(elem.shape[axis]) == num_elems for elem in elems_flat[1:]): raise ValueError('Array inputs to associative_scan must have the same ' 'first dimension. (saw: {})' .format([elem.shape for elem in elems_flat])) # Summary of algorithm: # # Consider elements of `_scan(elems)` at odd indices. That's the same as first # summing successive pairs of elements of `elems` and performing a scan on # that half sized tensor. We perform the latter scan by recursion. # # Now consider the even elements of `_scan(elems)`. These can be computed # from the odd elements of `_scan(elems)` by adding each odd element of # `_scan(elems)` to the matching even element in the original `elems`. # # We return the odd and even elements interleaved. # # For the base case of the recursion we return the first element # of `elems` followed by the sum of the first two elements computed as # a (small two-down-to-one) reduction step. def _scan(elems): """Perform scan on `elems`.""" num_elems = elems[0].shape[axis] if num_elems < 2: return elems # Combine adjacent pairs of elements. reduced_elems = combine( [slicing.slice_in_dim(elem, 0, -1, stride=2, axis=axis) for elem in elems], [slicing.slice_in_dim(elem, 1, None, stride=2, axis=axis) for elem in elems]) # Recursively compute scan for partially reduced tensors. odd_elems = _scan(reduced_elems) if num_elems % 2 == 0: even_elems = combine( [slicing.slice_in_dim(e, 0, -1, axis=axis) for e in odd_elems], [slicing.slice_in_dim(e, 2, None, stride=2, axis=axis) for e in elems]) else: even_elems = combine( odd_elems, [slicing.slice_in_dim(e, 2, None, stride=2, axis=axis) for e in elems]) # The first element of a scan is the same as the first element # of the original `elems`. even_elems = [ lax.concatenate([slicing.slice_in_dim(elem, 0, 1, axis=axis), result], dimension=axis) for (elem, result) in zip(elems, even_elems)] return list(_map(partial(_interleave, axis=axis), even_elems, odd_elems)) scans = _scan(elems_flat) if reverse: scans = [lax.rev(scanned, [axis]) for scanned in scans] return tree_unflatten(tree, scans)
# Cumulative reductions.
[docs]def cumsum(operand: Array, axis: int = 0, reverse: bool = False) -> Array: """Computes a cumulative sum along `axis`.""" return cumsum_p.bind(operand, axis=int(axis), reverse=bool(reverse))
[docs]def cumprod(operand: Array, axis: int = 0, reverse: bool = False) -> Array: """Computes a cumulative product along `axis`.""" return cumprod_p.bind(operand, axis=int(axis), reverse=bool(reverse))
[docs]def cummax(operand: Array, axis: int = 0, reverse: bool = False) -> Array: """Computes a cumulative maximum along `axis`.""" return cummax_p.bind(operand, axis=int(axis), reverse=bool(reverse))
[docs]def cummin(operand: Array, axis: int = 0, reverse: bool = False) -> Array: """Computes a cumulative minimum along `axis`.""" return cummin_p.bind(operand, axis=int(axis), reverse=bool(reverse))
def _cumred_shape_rule(x, *, axis: int, reverse: bool): if axis < 0 or axis >= x.ndim: raise ValueError( "axis {} is out of bounds for array of shape {}".format(axis, x.shape)) return x.shape def _cumsum_transpose_rule(t, operand, *, axis: int, reverse: bool): return [cumsum(t, axis=axis, reverse=not reverse)] def _cumred_tpu_translation_rule(window_reduce: Callable, x, *, axis: int, reverse: bool): # On TPU, an implementation using reduce_window is handled specially by the # compiler and is efficient. On other backends, it is O(n^2). n = x.shape[axis] if n == 0: return x padding = [(0, 0)] * x.ndim padding[axis] = (0, n - 1) if reverse else (n - 1, 0) strides = [1] * x.ndim window_dims = [1] * x.ndim window_dims[axis] = n return window_reduce(x, window_dims, strides, padding) def _cumred_batch_rule(prim, batched_args, batch_dims, *, axis: int, reverse: bool): operand, = batched_args bdim, = batch_dims axis = axis if axis < bdim else axis + 1 return prim.bind(operand, axis=axis, reverse=reverse), bdim def _cumred_dtype_rule(name, operand, *args, **kw): if not dtypes.issubdtype(operand.dtype, np.number): raise TypeError("{} does not accept dtype {}. Accepted dtypes are subtypes " "of number.".format(name, np.dtype(operand.dtype).name)) return dtypes.canonicalize_dtype(operand.dtype) def _cumulative_reduction_primitive(name, reduce_fn, tpu_reduce_window_fn): reducer_p = lax.standard_primitive( _cumred_shape_rule, partial(_cumred_dtype_rule, name), name, translation_rule=xla.lower_fun( partial(associative_scan, reduce_fn), multiple_results=False, new_style=True)) xla.register_translation(reducer_p, xla.lower_fun( partial(_cumred_tpu_translation_rule, tpu_reduce_window_fn), multiple_results=False, new_style=True), platform='tpu') batching.primitive_batchers[reducer_p] = partial(_cumred_batch_rule, reducer_p) mlir.register_lowering( reducer_p, mlir.lower_fun(partial(associative_scan, reduce_fn), multiple_results=False)) mlir.register_lowering( reducer_p, mlir.lower_fun(partial(_cumred_tpu_translation_rule, tpu_reduce_window_fn), multiple_results=False), platform='tpu') return reducer_p cumsum_p = _cumulative_reduction_primitive("cumsum", lax.add, windowed_reductions._reduce_window_sum) ad.deflinear2(cumsum_p, _cumsum_transpose_rule) cumprod_p = _cumulative_reduction_primitive("cumprod", lax.mul, windowed_reductions._reduce_window_prod) cummax_p = _cumulative_reduction_primitive("cummax", lax.max, windowed_reductions._reduce_window_max) cummin_p = _cumulative_reduction_primitive("cummin", lax.min, windowed_reductions._reduce_window_min) def _cumulative_jvp_rule(primals, tangents, *, axis: int, reverse: bool, combine_fn: Callable): # Irrespective of backend, we always use the parallel prefix scan # implementation when differentiating because reduce_window is not # arbitrarily differentiable. return api.jvp(partial(associative_scan, combine_fn, axis=axis, reverse=reverse), primals, tangents) ad.primitive_jvps[cumprod_p] = partial(_cumulative_jvp_rule, combine_fn=lax.mul) ad.primitive_jvps[cummin_p] = partial(_cumulative_jvp_rule, combine_fn=lax.min) ad.primitive_jvps[cummax_p] = partial(_cumulative_jvp_rule, combine_fn=lax.max)