def transform_ast(self, node, ctx):
node = qual_names.resolve(node)
node = activity.resolve(node, ctx)
graphs = cfg.build(node)
node = reaching_definitions.resolve(node, ctx, graphs)
node = reaching_fndefs.resolve(node, ctx, graphs)
node = type_inference.resolve(node, ctx, graphs, self.resolver)
return node
def transform(node, ctx):
graphs = cfg.build(node)
node = qual_names.resolve(node)
node = activity.resolve(node, ctx, None)
node = reaching_definitions.resolve(node, ctx, graphs)
node = reaching_fndefs.resolve(node, ctx, graphs)
node = liveness.resolve(node, ctx, graphs)
node = ControlFlowTransformer(ctx).visit(node)
return node
def convert4():
node, ctx = get_node_and_ctx(f4)
node = qual_names.resolve(node)
cfgs = cfg.build(node)
node = activity.resolve(node, ctx)
node = reaching_definitions.resolve(node, ctx, cfgs)
node = reaching_fndefs.resolve(node, ctx, cfgs)
node = liveness.resolve(node, ctx, cfgs)
print('live into `b = a + 1`:', anno.getanno(node.body[0], anno.Static.LIVE_VARS_IN))
print('live into `return b`:', anno.getanno(node.body[1], anno.Static.LIVE_VARS_IN))
def mlir_gen_internal(node, entity_info):
"""Returns mlir module for unprocessed node `node`."""
namer = naming.Namer({})
graphs = cfg.build(node)
ctx = transformer.Context(entity_info, namer, None)
node = qual_names.resolve(node)
node = activity.resolve(node, ctx)
node = reaching_definitions.resolve(node, ctx, graphs)
node = reaching_fndefs.resolve(node, ctx, graphs)
node = liveness.resolve(node, ctx, graphs)
mlir_generator = MLIRGen(ctx)
mlir_generator.visit(node)
return mlir_generator.prog
def _parse_and_analyze(self, test_fn):
# TODO(mdan): Use a custom FunctionTransformer here.
node, source = parser.parse_entity(test_fn, future_features=())
entity_info = transformer.EntityInfo(name=test_fn.__name__,
source_code=source,
source_file=None,
future_features=(),
namespace={})
node = qual_names.resolve(node)
namer = naming.Namer({})
ctx = transformer.Context(entity_info, namer, None)
node = activity.resolve(node, ctx)
graphs = cfg.build(node)
node = reaching_fndefs.resolve(node, ctx, graphs)
node = liveness.resolve(node, ctx, graphs)
return node
def transform_ast(self, node, ctx):
node = _apply_py_to_tf_passes(node, ctx)
# TODO(mdan): Enable this.
# node = anf.transform(node, ctx)
graphs = cfg.build(node)
node = qual_names.resolve(node)
node = activity.resolve(node, ctx)
node = reaching_definitions.resolve(node, ctx, graphs)
node = reaching_fndefs.resolve(node, ctx, graphs)
node = type_inference.resolve(node, ctx, graphs,
TFRTypeResolver(self._op_defs))
mlir_generator = TFRGen(ctx, self._op_defs)
mlir_generator.visit(node)
return mlir_generator.code_buffer
def _live_tensors(f, attr_name="inputs"):
"""Returns the indices of the used inputs.
Note: This currently only handles direct index accesses e.g. op.inputs[1].
If the function has slicing or list comprehension on attr_name then returns
_ALL. This ensure that this is correct even if inefficient.
Args:
f: A grad function, taking the op as first argument.
attr_name: op attr to track. "inputs" or "outputs".
Returns:
Either one of:
* set of integers representing individual indices of inputs used
* the value _ALL, if indices are used but cannot be determined which
* empty set, if no inputs are used
"""
node, _ = parser.parse_entity(f, ())
entity_info = transformer.EntityInfo(
name=f.__name__,
source_code=None,
source_file=None,
future_features=(),
namespace=sys.modules[f.__module__].__dict__)
ctx = transformer.Context(entity_info, None, None)
graphs = cfg.build(node)
node = qual_names.resolve(node)
node = activity.resolve(node, ctx, None)
node = reaching_fndefs.resolve(node, ctx, graphs)
node = liveness.resolve(node, ctx, graphs)
op_arg_name = anno.getanno(node.args.args[0], anno.Basic.QN)
op_inputs_outputs_name = qual_names.QN(op_arg_name, attr=attr_name)
special_tracker = _SubscriptUseTracker(ctx, (op_inputs_outputs_name, ))
node = special_tracker.visit(node)
live_vars_in = anno.getanno(node.body[0], anno.Static.LIVE_VARS_IN)
inputs_outputs_used_qns = set()
for v in special_tracker.complex_reads:
# Complicated patterns like op.inputs[:3]. Could be smarter about them
# if they matter much.
if v == op_inputs_outputs_name:
return _ALL
for v in live_vars_in:
if v in special_tracker.reads:
if (v.has_subscript() and v.parent == op_inputs_outputs_name):
inputs_outputs_used_qns.add(v)
elif v == op_inputs_outputs_name:
# When op.{attr_name} is used directly, assume all tensors are
# used for now. In that case, no point digging further.
# TODO(mdan): We can descend into tuple expansions.
return _ALL
function_calls_tracker = _FunctionCallsTracker(ctx, op_arg_name)
node = function_calls_tracker.visit(node)
input_output_indices = set()
for called_f in function_calls_tracker.calls:
child_indices = _live_tensors(called_f, attr_name=attr_name)
if child_indices is _ALL:
return _ALL
input_output_indices |= child_indices
for v in inputs_outputs_used_qns:
assert v.has_subscript()
_, subscript = v.qn
if not subscript.is_simple():
# Not a number, assuming it can be anything.
return _ALL
subscript_val, = subscript.qn
if (not isinstance(subscript_val, qual_names.Literal)
and not isinstance(subscript_val.value, int)):
# Not a number, assuming it can be anything.
return _ALL
input_output_indices.add(subscript_val.value)
return input_output_indices
