tracing.py

# -*- coding: utf-8 -*-
# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
#
# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
#
# Unless required by applicable law or agreed to in writing,
# software distributed under the License is distributed on an
# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
import collections
import contextlib
import functools
import itertools
import json
import os
import typing
import warnings
import weakref

import numpy as np

from ..core._imperative_rt import GraphProfiler
from ..core._imperative_rt.ops import (
    CollectiveComm,
    OprAttr,
    RemoteRecv,
    RemoteSend,
    VirtualDep,
)
from ..core._trace_option import set_symbolic_shape
from ..core._wrap import device as as_device
from ..core.ops.special import Const
from ..core.tensor import megbrain_graph as G
from ..core.tensor.core import OpBase, TensorBase, TensorWrapperBase, apply
from ..core.tensor.raw_tensor import OpDef, RawTensor, as_raw_tensor
from ..core.tensor.tensor import Tensor
from .sublinear_memory_config import SublinearMemoryConfig


def _input_node_use_static_shape():
    return os.environ.get("MEGENGINE_INPUT_NODE_USE_STATIC_SHAPE") is not None


class TraceMismatchError(RuntimeError):
    pass


active_trace = None
skip_tracing = False


def is_tracing():
    if active_trace is None:
        return False
    else:
        return not skip_tracing


@contextlib.contextmanager
def exclude_from_trace():
    global skip_tracing
    if skip_tracing:
        yield
        return
    try:
        skip_tracing = True
        if active_trace is not None:
            active_trace._begin_excluded_region()
        yield
    finally:
        skip_tracing = False


class TensorInfo:
    __slots__ = (
        # collected attributes
        "external",
        "exported",
        "data_read",
        "shape_read",
        "value_read",
        "device",
        "dtype",
        "shape",
        "is_const",
        "bound_data",
        # resources for execution
        "varnode",
        "data_setter",
        "shape_reader",
        "value_reader",
        "data_reader",
    )

    def __init__(self):
        self.exported = None
        self.data_read = None
        self.shape_read = None
        self.value_read = None
        self.bound_data = None

        self.data_setter = None
        self.shape_reader = None
        self.value_reader = None
        self.data_reader = None


_io_op_types = {CollectiveComm, RemoteSend, RemoteRecv}


class trace:
    """
    Wraps a callable and provide:

    * tracing via :meth:`.trace` and :meth:`.dump`
    * accelerated evalutaion via :meth:`.__call__`

    :param function: the function will be traced.
    :param symbolic: whether to apply symbolic execution for tracing. Default: False
    :param capture_as_const: capture global vars or closures as const value. Default: False
    :param sublinear_memory_config: configuration for sublinear memory optimization.
        If not None, it enables sublinear memory optimization with given setting.
    :param profiling: whether to profile compiled trace. Default: False
    :param opt_level: optimization level for compiling trace.
    :param symbolic_shape: whether to use symbolic shape for tracing. Default: True
    """

    def __new__(cls, *args, **kwargs):
        if not args:
            return functools.partial(cls, **kwargs)
        return super().__new__(cls)

    def __init__(
        self,
        function,
        symbolic=False,
        capture_as_const=False,
        sublinear_memory_config: SublinearMemoryConfig = None,
        profiling: bool = False,
        opt_level: int = None,
        symbolic_shape: bool = True,
    ):
        self.__wrapped__ = function
        self._symbolic = symbolic
        self._capture_as_const = capture_as_const
        self._sublinear_memory_config = sublinear_memory_config
        self._profiling = profiling
        self._profiler = None
        self._graph_opt_level = opt_level
        self._symbolic_shape = symbolic_shape

        self._reset()

    def _reset(self):
        self._untraced = True
        self._tinfo = []  # handle -> TensorInfo
        self._seq = []
        self._pc = 0
        self._graph = None
        self._need_reset_nodes = None
        self._lazy_eval_graph = None
        self._lazy_eval_tensors = weakref.WeakSet()
        self._lazy_eval_links = None
        self._active_tensors = weakref.WeakSet()
        self._tensor_remaps = None
        self._inputs_to_restore = None
        self._arg_bindings = None
        self._kwarg_bindings = None
        self._output_bindings = None
        self._output_names = None

    def _new_handle(self):
        handle = len(self._tinfo)
        info = TensorInfo()
        self._tinfo.append(info)
        return handle, info

    def _apply_op(self, op, args):
        assert not self._untraced
        # check against trace
        if self._pc >= len(self._seq):
            raise TraceMismatchError("trace should end here, but more op observed")
        record = self._seq[self._pc]
        op_, ihandles, ohandles = record
        if op != op_:
            # FIXME: will be removed once better rng implementation is done
            if isinstance(op, OprAttr) and (
                op.type in ("UniformRNG", "GaussianRNG") and op.type == op_.type
            ):
                if op.param[8:] != op_.param[8:]:
                    raise TraceMismatchError("op different from last time")
            else:
                raise TraceMismatchError("op different from last time")
        if len(ihandles) != len(args):
            raise TraceMismatchError("op input size different from last time")

        for h, x in zip(ihandles, args):
            info = self._tinfo[h]
            if info.external:
                if (
                    x.__class__ is CompiledTensorProxy
                    and not self._tinfo[x._CompiledTensorProxy__handle].exported
                ):
                    raise TraceMismatchError(
                        "failed to capture: input was an external tensor "
                        "last time, got an internal tensor this time"
                    )
                if info.bound_data:
                    if x.__class__ is CompiledTensorProxy:
                        raise TraceMismatchError(
                            "const capture violated: was an external tensor "
                            "last time, got an internal tensor this time"
                        )
                    if x._handle != info.bound_data._handle:
                        if not np.array_equal(x.numpy(), info.bound_data.numpy()):
                            raise TraceMismatchError(
                                "const capture violated: got "
                                "a different tensor this time"
                            )
                else:
                    if info.dtype != x.dtype:
                        raise TraceMismatchError(
                            "failed to capture: different dtype from last time"
                        )
                    if info.device != x.device:
                        raise TraceMismatchError(
                            "failed to capture: different device from last time"
                        )
                    info.data_setter.set_value(x._dev_tensor())
            else:
                if x.__class__ is not CompiledTensorProxy:
                    if x not in self._tensor_remaps:
                        raise TraceMismatchError(
                            "unexpected capture: trying to use an external tensor as "
                            "input, but that input was an internal tensor last time"
                        )
                    else:
                        x = self._tensor_remaps[x]
                if x._CompiledTensorProxy__handle != h:
                    raise TraceMismatchError(
                        "mis-wiring: input edge to an data flow "
                        "graph node is different from last time"
                    )

        self._pc += 1
        outputs = tuple([CompiledTensorProxy(h) for h in ohandles])
        self._active_tensors.update(outputs)
        return outputs

    def _apply_const(self, op, args):
        assert not self._untraced
        # check against trace
        if self._pc >= len(self._seq):
            raise TraceMismatchError("trace should end here, but more op observed")
        record = self._seq[self._pc]
        op_, ihandles, ohandles = record
        assert isinstance(op_, Const)

        eq = op_.value == op.value
        if not isinstance(eq, bool):
            eq = all(eq)
        if not eq:
            raise TraceMismatchError(
                "const tensor violated: got a different tensor this time"
            )

        self._pc += 1
        (h,) = ohandles
        outputs = tuple([self._tinfo[h].bound_data])
        return outputs

    def _record_op(self, op, inputs, outputs):
        if skip_tracing:
            for x in inputs:
                h = getattr(x, "_TraceMixin__handle", None)
                if h is not None:
                    self._tinfo[h].data_read = True
            return

        ihandles = []
        for x in inputs:
            h = getattr(x, "_TraceMixin__handle", None)
            if h is None or (not self._capture_as_const and self._tinfo[h].exported):
                h, info = self._new_handle()
                info.external = True
                info.device = x.device
                info.dtype = x.dtype
                info.shape = x.shape
                if self._capture_as_const:
                    info.bound_data = x

            ihandles.append(h)

        ohandles = []
        for x in outputs:
            h, info = self._new_handle()
            ohandles.append(h)
            info.external = False
            TraceMixin._TraceMixin__inject(x, h)

        self._seq.append((op, tuple(ihandles), tuple(ohandles)))
        self._active_tensors.update(outputs)

    def _record_const(self, op, outputs):
        if skip_tracing:
            (x,) = outputs
            h = getattr(x, "_TraceMixin__handle", None)
            if h is not None:
                self._tinfo[h].data_read = True
            return

        (x,) = outputs
        h, info = self._new_handle()
        ohandles = [h]
        info.external = True
        info.device = x.device
        info.dtype = x.dtype
        info.shape = x.shape
        info.bound_data = x
        info.is_const = True
        TraceMixin._TraceMixin__inject(x, h)
        self._seq.append((op, tuple(), tuple(ohandles)))

    def _set_active(self, active: bool):
        global active_trace
        if active:
            if active_trace:
                raise NotImplementedError("sorry, not implemented: nested trace")
            active_trace = self
        else:
            assert active_trace is self
            active_trace = None

    def _init_trace(self, symbolic: bool):
        apply.enable(apply_with_tracing)
        apply.enable(apply_const_with_tracing)
        if symbolic:
            apply.enable(apply_symbolic_mode)
            apply.enable(apply_const_symbolic_mode)
            self._lazy_eval_graph = G.Graph()
            self._apply_graph_options(self._lazy_eval_graph)
            self._lazy_eval_links = ()

    def _take_escaped_tensors(self):
        escaped_tensors = tuple(self._active_tensors)
        self._active_tensors.clear()
        return escaped_tensors

    def _lazy_eval(self, lazy_eval_graph, lazy_eval_tensors, lazy_eval_links):
        readers = [
            G.OutputNode(x._LazyEvalTensor__varnode).outputs[0]
            for x in lazy_eval_tensors
        ]
        self._apply_graph_options(lazy_eval_graph)
        # FIXME
        if self._graph_opt_level is not None:
            lazy_eval_graph.options.graph_opt_level = self._graph_opt_level
        else:
            lazy_eval_graph.options.graph_opt_level = 2
        lazy_eval_graph.compile(*lazy_eval_links, *readers)
        lazy_eval_graph()
        for r, x in zip(readers, lazy_eval_tensors):
            assign_raw_tensor(x, as_raw_tensor(r.op.get_value()))

    @contextlib.contextmanager
    def _setup(self):
        interrupted = False

        def do_enter():
            self._save_symbolic_shape = set_symbolic_shape(self._symbolic_shape)
            self._set_active(True)
            if self._untraced:
                self._init_trace(self._symbolic)
            else:
                apply.enable(apply_compiled_mode)
                apply.enable(apply_const_compiled_mode)
                if self._graph is None:
                    self._compile()
                self._graph.execute()

        def do_finalize():
            escaped_tensors = self._take_escaped_tensors()
            if self._untraced:
                for x in escaped_tensors:
                    info = self._tinfo[x._TraceMixin__handle]
                    info.data_read = True
                    x._TraceMixin__restore()
                if self._inputs_to_restore:
                    for x in self._inputs_to_restore:
                        x._TraceMixin__restore()
                if self._symbolic and (
                    self._lazy_eval_tensors or self._lazy_eval_links
                ):
                    # eval lazy eval tensors
                    self._lazy_eval(
                        self._lazy_eval_graph,
                        tuple(self._lazy_eval_tensors),
                        self._lazy_eval_links,
                    )
                    self._lazy_eval_graph = None
                    self._lazy_eval_tensors = None
                    self._lazy_eval_links = None
                self._untraced = False
            else:
                # compiled_tensor leaks
                if self._pc == len(self._seq):
                    for x in escaped_tensors:
                        try:
                            assign_raw_tensor(x, as_raw_tensor(x._dev_tensor()))
                        except TraceMismatchError:
                            # TraceMismatchError thrown in do_exit
                            pass
                    self._graph.wait()
                    self._reset_exec_env()

            # reset status
            self._pc = 0
            self._tensor_remaps = None
            apply.disable(apply_with_tracing)
            apply.disable(apply_const_with_tracing)
            apply.disable(apply_symbolic_mode)
            apply.disable(apply_const_symbolic_mode)
            apply.disable(apply_compiled_mode)
            apply.disable(apply_const_compiled_mode)
            self._set_active(False)
            # Restore global variable
            set_symbolic_shape(self._save_symbolic_shape)

        def do_exit():
            if not self._untraced and self._pc != len(self._seq):
                raise TraceMismatchError("premature end")
            if not self._symbolic or not self._untraced:
                for x in self._active_tensors:
                    x._dev_tensor()

        try:
            do_enter()
            yield
            do_exit()
        except:
            interrupted = True
            raise
        finally:
            do_finalize()
            if interrupted:
                self._reset()

    def _begin_excluded_region(self):
        if self._capture_as_const:
            raise RuntimeError(
                "exclude_from_trace cannot be used with capture_as_const"
            )
        if self._untraced:
            # conditionally reading a compiled tensor in excluded region
            # is permitted, so we have to assume every tensor might be read
            for x in self._active_tensors:
                info = self._tinfo[x._TraceMixin__handle]
                info.exported = True
                info.data_read = True

    def _apply_graph_options(self, graph):

        graph.options.no_force_inplace = True
        graph.options.seq_opt.enable_seq_comp_node_opt = False
        # graph opt level
        # if self._graph_opt_level is not None:
        #     graph.options.graph_opt_level = self._graph_opt_level
        # FIXME
        graph.options.graph_opt_level = 0
        # sublinear
        if self._sublinear_memory_config is not None:
            graph.options.enable_sublinear_memory_opt = True
            sublinear_config = graph.options.sublinear_mem_config
            sublinear_config.lb_memory = self._sublinear_memory_config.lb_memory
            sublinear_config.genetic_nr_iter = (
                self._sublinear_memory_config.genetic_nr_iter
            )
            sublinear_config.genetic_pool_size = (
                self._sublinear_memory_config.genetic_pool_size
            )
            sublinear_config.thresh_nr_try = self._sublinear_memory_config.thresh_nr_try
            sublinear_config.num_worker = self._sublinear_memory_config.num_worker
        # profile
        if self._profiling:
            self._profiler = GraphProfiler(graph)

    def _compile(self):
        graph = self._graph = G.Graph()
        graph.options.async_exec_level = 0b100
        self._apply_graph_options(graph)
        # graph.options.graph_opt_level = 0
        need_reset_nodes = self._need_reset_nodes = []
        # links enforce ordering of I/O nodes
        links = ()
        readers = []

        if self._capture_as_const:
            for h in itertools.chain(self._arg_bindings, self._kwarg_bindings.values()):
                info = self._tinfo[h]
                opnode = info.data_setter = G.InputNode(
                    device=info.device,
                    dtype=info.dtype,
                    shape=info.shape or (1,),
                    graph=graph,
                    use_static_shape=_input_node_use_static_shape(),
                )
                need_reset_nodes.append(opnode)
                info.varnode = opnode.outputs[0]
                links += opnode.outputs[1:]

        for op, ihandles, ohandles in self._seq:
            if isinstance(op, Const):
                assert len(ihandles) == 0
                (h,) = ohandles
                info = self._tinfo[h]
                if not hasattr(info, "varnode"):
                    assert info.external
                    assert info.bound_data
                    info.varnode = graph.make_const(
                        info.bound_data.numpy(),
                        info.bound_data.dtype,
                        info.bound_data.device,
                    )
                continue

            require_links = type(op) in _io_op_types
            ivars = []
            for i, h in enumerate(ihandles):
                info = self._tinfo[h]
                if not hasattr(info, "varnode"):
                    assert info.external
                    if info.bound_data:
                        if hasattr(info, "is_const") and info.is_const:
                            info.varnode = graph.make_const(
                                info.bound_data.numpy(),
                                info.bound_data.dtype,
                                info.bound_data.device,
                            )
                        else:
                            info.varnode = graph.make_const(
                                info.bound_data._dev_tensor()
                                # info.bound_data.numpy()
                            )
                    else:
                        opnode = info.data_setter = G.InputNode(
                            *links,
                            device=info.device,
                            dtype=info.dtype,
                            shape=info.shape or (1,),
                            graph=graph,
                            use_static_shape=_input_node_use_static_shape(),
                        )
                        need_reset_nodes.append(opnode)
                        info.varnode, *links = opnode.outputs
                if require_links and i == 0 and len(links) > 0:
                    info.varnode = apply(VirtualDep(), info.varnode, *links)[0]
                    links = (info.varnode,)

                ivars.append(info.varnode)
            ovars = apply(op, *ivars)
            if require_links and len(ovars) > 0:
                links = (ovars[0],)
            assert len(ovars) == len(ohandles)
            for h, v in zip(ohandles, ovars):
                info = self._tinfo[h]
                info.varnode = v

                def add_reader(opnode):
                    nonlocal links
                    need_reset_nodes.append(opnode)
                    readers.append(opnode.outputs[0])
                    links = opnode.outputs

                if info.data_read:
                    # Shape can be obtained from data so doesn't need its own
                    # output node. On the other hand, value is read separately
                    # to leverage eager h2d copy
                    info.shape_read = False
                    opnode = info.data_reader = G.OutputNode(v, *links)
                    add_reader(opnode)
                if info.value_read:
                    opnode = info.value_reader = G.ValueOutputNode(v, *links)
                    add_reader(opnode)
                if info.shape_read:
                    opnode = info.shape_reader = G.AttrOutputNode(v, *links)
                    add_reader(opnode)
        # FIXME
        if self._graph_opt_level is not None:
            graph.options.graph_opt_level = self._graph_opt_level
        else:
            graph.options.graph_opt_level = 2
        graph.compile(*readers, *links)

    def _reset_exec_env(self):
        for opnode in self._need_reset_nodes:
            opnode.reset()

    def _require_shape(self, handle):
        info = self._tinfo[handle]
        info.shape_read = True

    def _require_value(self, handle):
        info = self._tinfo[handle]
        info.value_read = True

    def _require_data(self, handle):
        info = self._tinfo[handle]
        info.data_read = True

    def __call__(self, *args, **kwargs):
        if is_tracing():
            return self.__wrapped__(*args, **kwargs)
        with self._setup():
            if self._capture_as_const:
                self._process_inputs(*args, **kwargs)
            outputs = self.__wrapped__(*args, **kwargs)
            if self._capture_as_const:
                self._process_outputs(outputs)
            return outputs

    def dump(
        self,
        file,
        *,
        arg_names=None,
        output_names=None,
        append=False,
        optimize_for_inference=True,
        **kwargs
    ):
        r"""
        Serializes trace to file system.

        :param file: output file, could be file object or filename.
        :param arg_names: names of the input tensors in the traced function.
        :param output_names: names of the output tensors in the traced function,
            use the default name if not specified.
        :param append: whether output is appended to ``file``.
            Only works when ``file`` is str.
        :param optimize_for_inference: enbale optmizations,
            will skip all optimize options if this is False. Default: True

        :Keyword Arguments:

            * enable_io16xc32 --
                whether to use float16 for I/O between oprs and use
                float32 as internal computation precision. Note the output var would be
                changed to float16.
            * enable_ioc16 --
                whether to use float16 for both I/O and computation
                precision.

            * enable_hwcd4 --
                whether to use NHWCD4 data layout. This is faster on some
                OpenCL backend.
            * enable_nchw88 --
                whether to use NCHW88 data layout, currently
                used in X86 AVX backend.
            * enable_nchw44 --
                whether to use NCHW44 data layout, currently
                used in arm backend.
            * enable_nchw44_dot --
                whether to use NCHW44_dot data layout, currently
                used in armv8.2+dotprod backend.
            * enable_nchw4 --
                whether to use NCHW4 data layout, currently
                used in nvidia backend(based on cudnn).
            * enable_nchw32 --
                whether to use NCHW32 data layout, currently
                used in nvidia backend with tensorcore(based on cudnn).
            * enable_chwn4 --
                whether to use CHWN4 data layout, currently
                used in nvidia backend with tensorcore.

            * enable_fuse_conv_bias_nonlinearity: whether to fuse conv+bias+nonlinearty
                into one opr.
            * enable_fuse_conv_bias_with_z: whether to fuse conv_bias with z
                input for inference on nvidia backend(this optimization pass will
                result in mismatch of the precision of output of training and
                inference)
        """
        if not self._capture_as_const:
            raise ValueError(
                "you must specify capture_as_const=True at __init__ to use dump"
            )
        if self._untraced:
            raise RuntimeError("should run at least once before calling dump")
        if self._output_names and output_names:
            raise TypeError(
                "cannot specify output_names when output is already in dict format"
            )
        if output_names and not isinstance(output_names, collections.abc.Sequence):
            output_names = (output_names,)
        if output_names and len(output_names) != len(self._output_bindings):
            raise ValueError(
                "wrong number of output_names, should be {} values".format(
                    len(self._output_bindings)
                )
            )
        if arg_names is None:
            arg_names = ["arg_%d" % i for i in range(len(self._arg_bindings))]
        if arg_names and not isinstance(arg_names, collections.abc.Sequence):
            arg_names = (arg_names,)
        if arg_names and len(arg_names) != len(self._arg_bindings):
            raise ValueError(
                "wrong number of arg_names, should be {} values".format(
                    len(self._arg_bindings)
                )
            )
        output_names = output_names or self._output_names

        dumped_device = as_device("xpux")

        h2v = {}
        graph = G.Graph()
        # only graph_opt_level takes effect in dump
        self._apply_graph_options(graph)

        for i, h in enumerate(self._arg_bindings):
            info = self._tinfo[h]
            h2v[h] = graph.make_h2d(
                dtype=info.dtype,
                device=dumped_device,
                shape=info.shape or (1,),
                name=arg_names[i] if arg_names else None,
            )
        for k, h in self._kwarg_bindings.items():
            info = self._tinfo[h]
            h2v[h] = graph.make_h2d(
                dtype=info.dtype, device=dumped_device, shape=info.shape or (1,), name=k
            )

        for op, ihandles, ohandles in self._seq:
            if isinstance(op, Const):
                assert len(ihandles) == 0
                (h,) = ohandles
                info = self._tinfo[h]
                if h not in h2v:
                    assert info.external
                    assert info.bound_data
                    h2v[h] = graph.make_const(
                        info.bound_data.numpy(), dtype=info.dtype, device=info.device,
                    )
                continue
            ivars = []
            for h in ihandles:
                info = self._tinfo[h]
                if h not in h2v:
                    assert info.external
                    assert info.bound_data
                    h2v[h] = graph.make_const(
                        info.bound_data.numpy(), dtype=info.dtype, device=dumped_device
                    )
                ivars.append(h2v[h])
            ovars = apply(op, *ivars)
            assert len(ovars) == len(ohandles)
            h2v.update(zip(ohandles, ovars))

        dest_vars = []
        for i, h in enumerate(self._output_bindings):
            v = h2v[h]
            if output_names:
                v.name = output_names[i]
            dest_vars.append(v)

        if optimize_for_inference:
            dest_vars = G.optimize_for_inference(dest_vars, **kwargs)

        if isinstance(file, str):
            permission = "wb" if append == False else "ab"
            file = open(file, permission)
        dump_content, dump_info = G.dump_graph(dest_vars)
        file.write(dump_content)
        return dump_info

    def _process_inputs(self, *args, **kwargs):
        if self._untraced:
            self._inputs_to_restore = []

            def record_input(x):
                if x is None:
                    return
                h, info = self._new_handle()
                info.external = False
                info.device = x.device
                info.dtype = x.dtype
                info.shape = x.shape
                TraceMixin._TraceMixin__inject(x, h)
                self._inputs_to_restore.append(x)
                return h

            self._arg_bindings = []
            for i, x in enumerate(args):
                x = find_raw_tensor(x)
                if x is None:
                    raise TypeError(
                        "positional arguments should all be tensor "
                        "but args[%d] cannot be recognized as one" % i
                    )
                self._arg_bindings.append(record_input(x))

            self._kwarg_bindings = {}
            for k, x in kwargs.items():
                x = find_raw_tensor(x)
                if x is not None:
                    self._kwarg_bindings[k] = record_input(x)
        else:
            if len(args) != len(self._arg_bindings):
                raise TraceMismatchError("positional argument length mismatch")

            self._tensor_remaps = {}

            for i, (h, x) in enumerate(zip(self._arg_bindings, args)):
                x = find_raw_tensor(x)
                if x is None:
                    raise TypeError(
                        "positional arguments should all be tensor "
                        "but args[%d] cannot be recognized as one" % i
                    )
                info = self._tinfo[h]
                if x.dtype != info.dtype:
                    raise TypeError("args[%d].dtype different from last time" % i)
                if x.device != info.device:
                    raise TypeError("args[%d].device different from last time" % i)
                info.data_setter.set_value(x._dev_tensor())
                self._tensor_remaps[x] = CompiledTensorProxy(h)

            kwargs_tensors = {}
            for k, x in kwargs.items():
                x = find_raw_tensor(x)
                if x is not None:
                    kwargs_tensors[k] = x
            if set(kwargs_tensors) != set(self._kwarg_bindings):
                too_many = set(kwargs_tensors) - set(self._kwarg_bindings)
                too_few = set(self._kwarg_bindings) - set(kwargs_tensors)
                if too_many:
                    raise TraceMismatchError(
                        "keyword arguments found to be tensor this time "
                        "but were non-tensor previously: %s" % " ".join(too_many)
                    )
                if too_few:
                    raise TraceMismatchError(
                        "keyword arguments found to be non-tensor this time "
                        "but were tensor previously: %s" % " ".join(too_few)
                    )
            for k, h in self._kwarg_bindings.items():
                x = kwargs_tensors[k]
                info = self._tinfo[h]
                if x.dtype != info.dtype:
                    raise TypeError("kwargs[%s].dtype different from last time" % k)
                if x.device != info.device:
                    raise TypeError("kwargs[%s].device different from last time" % k)
                info.data_setter.set_value(x._dev_tensor())
                self._tensor_remaps[x] = CompiledTensorProxy(h)

    def _process_outputs(self, outputs):
        output_names = None
        if isinstance(outputs, collections.abc.Mapping):
            output_names, outputs = zip(*sorted(outputs.items()))
        elif not isinstance(outputs, collections.abc.Sequence):
            outputs = (outputs,)

        if not self._untraced:
            if output_names != self._output_names:
                too_many = set(output_names) - set(self._output_names)
                too_few = set(self._output_names) - set(output_names)
                if too_many:
                    raise TraceMismatchError(
                        "output has more keys than last time: %s" % " ".join(too_many)
                    )
                if too_few:
                    raise TraceMismatchError(
                        "output has less keys than last time: %s" % " ".join(too_few)
                    )
            if len(outputs) != len(self._output_bindings):
                raise TraceMismatchError("output size differs from last time")
        else:
            self._output_names = output_names
            self._output_bindings = []

        for i, x in enumerate(outputs):
            x = find_raw_tensor(x)
            if x is None:
                raise TypeError("every item of return value should be tensor")
            if self._untraced:
                if not isinstance(x, TraceMixin):
                    raise RuntimeError("output is not computed from inputs")
                h = x._TraceMixin__handle
                self._output_bindings.append(h)
            else:
                if not isinstance(x, CompiledTensorProxy):
                    raise RuntimeError("output is not computed from inputs")
                h = x._CompiledTensorProxy__handle
                if h != self._output_bindings[i]:
                    raise TraceMismatchError(
                        "retval[%s] is a different tensor than last time"
                        % (output_names and output_names[i] or i)
                    )

    def get_profile(self):
        """
        Get profiling result for compiled trace.

        :return: a json compatible object.
        """
        if not self._profiler:
            raise RuntimeError("trace is not set with profiling=True")
        return json.loads(self._profiler.get())

    def trace(self, *args, **kwargs):
        raise NotImplementedError(
            "trace is deemed unbeneficial with the new "
            "tracing mechanism. You should alwasy use __call__."
        )


class CompiledTensorProxy(RawTensor):
    """
    Duck-typed RawTensor
    """

    def __init__(self, handle):
        self.__handle = handle
        self._isscalar = False
        self.__info = active_trace._tinfo[handle]
        self.__shape = None
        self.__data = None
        self.__value = None

    @property
    def dtype(self):
        return self.__info.varnode.dtype

    @property
    def device(self):
        return self.__info.varnode.device

    @property
    def shape(self):
        if self._isscalar:
            return ()
        if self.__shape is None:
            if self.__info.shape_read:
                self.__shape = self.__info.shape_reader.get_value().shape
            elif self.__info.data_read:
                self.__shape = self._dev_tensor().shape
            else:
                raise TraceMismatchError("shape of this tensor is not read in trace")
        return self.__shape

    def numpy(self):
        if self.__value is None:
            if self.__info.value_read:
                self.__value = self.__info.value_reader.get_value()
            elif self.__info.data_read:
                self.__value = self._dev_tensor().numpy()
            else:
                raise TraceMismatchError("value of this tensor is not read in trace")
            if self._isscalar:
                self.__value = self.__value.squeeze()
        return self.__value

    def _dev_tensor(self):
        if self.__data is None:
            if not self.__info.data_read:
                raise TraceMismatchError("raw data of this tensor is not read in trace")
            self.__data = self.__info.data_reader.get_value()
        return self.__data

    def __del__(self):
        if self.__info.shape_read and self.__shape is not None:
            self.__info.shape_reader.drop_value()
        if self.__info.value_read and self.__value is not None:
            self.__info.value_reader.drop_value()
        if self.__info.data_read and self.__data is not None:
            self.__info.data_reader.drop_value()


class LazyEvalTensor(RawTensor):
    def __init__(self, varnode, isscalar=False):
        super().__init__()
        self.__varnode = varnode
        self._isscalar = isscalar

    @property
    def dtype(self):
        return self.__varnode.dtype

    @property
    def device(self):
        return self.__varnode.device

    @property
    def shape(self):
        if self._isscalar:
            return ()
        return self.__varnode.shape

    def numpy(self):
        ret = self.__varnode.value
        if self._isscalar:
            ret = ret.squeeze()
        return ret

    def _dev_tensor(self):
        raise RuntimeError("cannot access data during symbolic tracing")


class TraceMixin:
    __subclass_cache = {}

    def __inject(self, handle):
        cache = __class__.__subclass_cache
        cls = self.__class__
        subcls = cache.get(cls)
        if subcls is None:
            subcls = cache[cls] = type("Traced" + cls.__name__, (__class__, cls), {})
        self.__class__ = subcls
        self.__handle = handle
        self.__cls = cls
        return self

    def __restore(self):
        cls = self.__cls
        del self.__handle
        del self.__cls
        self.__class__ = cls
        return self

    @property
    def shape(self):
        if not skip_tracing:
            active_trace._require_shape(self.__handle)
        return super().shape

    def numpy(self):
        if not skip_tracing:
            active_trace._require_value(self.__handle)
        return super().numpy()

    def _dev_tensor(self):
        if not skip_tracing:
            active_trace._require_data(self.__handle)
        return super()._dev_tensor()


class TracedRawTensor(TraceMixin, RawTensor):
    pass


class TracedLazyTensor(TraceMixin, LazyEvalTensor):
    pass


def assign_raw_tensor(lhs, rhs):
    handle = rhs._handle
    # Keep isscalar of lhs
    isscalar = lhs._isscalar
    rhs.__dict__.clear()
    lhs.__dict__.clear()
    lhs.__class__ = RawTensor
    lhs.__init__(handle, isscalar=isscalar)


# this hook turns RawTensor into LazyEvalTensor
@apply.register()
def apply_symbolic_mode(op: OpDef, *args: RawTensor):
    graph = active_trace._lazy_eval_graph
    ivars = []
    for x in args:
        var = getattr(x, "_LazyEvalTensor__varnode", None)
        if var:
            ivars.append(var)
        else:
            data_setter = G.InputNode(
                device=x.device,
                dtype=x.dtype,
                shape=x.shape or (1,),
                graph=graph,
                use_static_shape=True,
            )
            var = data_setter.outputs[0]
            ivars.append(var)
            data_setter.set_value(x._dev_tensor())

    require_links = type(op) in _io_op_types

    if require_links and active_trace._lazy_eval_links:
        assert len(ivars) > 0, "op should has at least one input"
        ivars[0] = apply(VirtualDep(), ivars[0], *active_trace._lazy_eval_links)[0]
        active_trace._lazy_eval_links = (ivars[0],)

    ovars = apply(op, *ivars)

    if require_links:
        active_trace._lazy_eval_links = (ovars[0],)

    outputs = [LazyEvalTensor(v) for v in ovars]
    active_trace._lazy_eval_tensors.update(outputs)
    return outputs


apply.disable(apply_symbolic_mode)


@apply.register()
def apply_const_symbolic_mode(op: Const, *args: RawTensor):
    graph = active_trace._lazy_eval_graph
    ret = LazyEvalTensor(
        graph.make_const(op.value, dtype=op.dtype, device=op.device), isscalar=True
    )
    active_trace._lazy_eval_tensors.add(ret)
    return (ret,)


apply.disable(apply_const_symbolic_mode)


@apply.register()
def apply_compiled_mode(op: OpDef, *args: RawTensor):
    if skip_tracing:
        args = [
            as_raw_tensor(x._dev_tensor()) if x.__class__ is CompiledTensorProxy else x
            for x in args
        ]
        return apply.super(op, *args)
    return active_trace._apply_op(op, args)


apply.disable(apply_compiled_mode)


@apply.register()
def apply_const_compiled_mode(op: Const, *args: RawTensor):
    if skip_tracing:
        args = [
            as_raw_tensor(x._dev_tensor()) if x.__class__ is CompiledTensorProxy else x
            for x in args
        ]
        return apply.super(op, *args)
    return active_trace._apply_const(op, args)


apply.disable(apply_const_compiled_mode)


# this hook injects TraceMixin
@apply.register()
def apply_with_tracing(op: OpDef, *args: RawTensor):
    outputs = apply.super(op, *args)
    active_trace._record_op(op, args, outputs)
    return outputs


apply.disable(apply_with_tracing)


@apply.register()
def apply_const_with_tracing(op: Const, *args: RawTensor):
    outputs = apply.super(op, *args)
    active_trace._record_const(op, outputs)
    return outputs


apply.disable(apply_const_with_tracing)


class BrokenRawTensor(RawTensor):
    def __getattribute__(self, _):
        raise RuntimeError("broken due to misuse of tracing")

    def __setattr__(self, *_):
        raise RuntimeError("broken due to misuse of tracing")


@functools.singledispatch
def find_raw_tensor(x):
    return None


@find_raw_tensor.register(RawTensor)
def _(x):
    return x


@find_raw_tensor.register(TensorWrapperBase)
def _(x):
    x = getattr(x, "__wrapped__", None)
    if x is not None:
        return find_raw_tensor(x)


@find_raw_tensor.register(Tensor)
def _(x):
    x = getattr(x, "_data", None)
    if x is not None:
        return find_raw_tensor(x)