megbrain_graph.py 19.2 KB
Newer Older
1 2 3
# -*- coding: utf-8 -*-
# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
#
4
# Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
5 6 7 8 9
#
# 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
10
import json
11
import os
12
import weakref
13 14
from concurrent.futures import ThreadPoolExecutor
from typing import Dict, List, Tuple, Union
15

16 17
import numpy as np

18
from .. import _imperative_rt
19
from .._imperative_rt import GraphOptimizeOptions
20
from .._imperative_rt.core2 import apply, set_cpp_apply_backward_varnode
21
from .._imperative_rt.ops import BackwardGraph
22 23
from .._wrap import device as as_device
from ..ops.builtin import OpDef
24
from .core import TensorBase
25 26


27 28 29 30 31 32 33 34 35 36 37 38 39
def set_priority_to_id(dest_vars):
    """
    For all oprs in the subgraph constructed by dest_vars,
       sets its priority to id if its original priority is zero.
    :param dest_vars: target vars representing the graph.
    """
    dest_vec = []
    for i in dest_vars:
        assert isinstance(i, _imperative_rt.VarNode)
        dest_vec.append(i)
    _imperative_rt.graph._set_priority_to_id(dest_vec)


M
Megvii Engine Team 已提交
40 41 42 43 44 45 46
class Graph(_imperative_rt.ComputingGraph):
    def __init__(self):
        super().__init__()
        self._var_cache = weakref.WeakKeyDictionary()
        self._op_cache = weakref.WeakKeyDictionary()
        self._executor = ThreadPoolExecutor(1)
        self._function = None
47 48
        self._future = None

M
Megvii Engine Team 已提交
49 50 51 52 53
    def _wrap(self, obj):
        if type(obj) is _imperative_rt.VarNode:
            wrapper, cache = VarNode, self._var_cache
        elif type(obj) is _imperative_rt.OperatorNode:
            wrapper, cache = OpNode, self._op_cache
54 55
        else:
            raise TypeError(type(obj))
M
Megvii Engine Team 已提交
56 57 58 59
        if obj not in cache:
            cache[obj] = wrapper(obj)
        return cache[obj]

60 61
    def _set_priority_to_id(self, dest_vars):
        set_priority_to_id(_unwrap(dest_vars))
62

M
Megvii Engine Team 已提交
63 64 65 66
    def compile(self, *args):
        self._function = super().compile(_unwrap(args))
        return self

67 68
    def execute(self, *args):
        assert self._future is None
69 70 71 72 73 74 75 76 77 78

        def wrapped(*args):
            try:
                self._function.execute(*args)
            except Exception as exc:
                for i in self._function._all_rendezvous:
                    i.set_exception(str(exc))
                raise exc

        self._future = self._executor.submit(wrapped, *args)
79 80 81 82 83 84 85 86 87 88 89 90 91 92

    def wait(self):
        assert self._future is not None
        self._future.exception()
        self._function.wait()
        try:
            return self._future.result()
        finally:
            self._future = None

    def __call__(self, *args):
        self.execute(*args)
        return self.wait()

93 94 95 96 97
    def _make_const_for_backward(self, data):
        device = as_device(data.comp_node).to_c()
        data = data.numpy()
        return self._wrap(_imperative_rt.make_const(self, data, device, data.dtype))

98
    def make_const(self, data, dtype=None, device=None, name=None):
M
Megvii Engine Team 已提交
99 100 101 102
        if isinstance(data, _imperative_rt.DeviceTensorND):
            assert dtype is None and device is None
            return self._wrap(_imperative_rt.make_shared(self, data))
        else:
103 104 105 106 107
            data = np.asarray(data, dtype=dtype)
            if data.dtype == np.float64:
                data = data.astype(np.float32)
            elif data.dtype == np.int64:
                data = data.astype(np.int32)
M
Megvii Engine Team 已提交
108
            device = as_device(device).to_c()
109 110 111
            return self._wrap(
                _imperative_rt.make_const(self, data, device, dtype, name)
            )
112

M
Megvii Engine Team 已提交
113 114 115
    def make_input(self, *args: "VarNode", device=None, dtype=None, shape=None):
        opnode = InputNode(*args, device=device, dtype=dtype, shape=shape, graph=self)
        return opnode.outputs[0]
116

117
    def make_h2d(self, *, dtype, device, shape=None, name=None):
118
        device = as_device(device).to_c()
119
        return self._wrap(_imperative_rt.make_h2d(self, device, dtype, shape, name))
120

121 122 123 124 125 126 127 128
    def _to_json(self, filename):
        # debug interface
        if self._function:
            js = json.loads(self._function._to_json())
            json.dump(js, open(filename, "w"))
        else:
            print("this function should be called after compilation.")

129

130
class VarNode(TensorBase):
131
    def __init__(self, node: _imperative_rt.VarNode, isscalar=False):
132
        self._node = node
133
        self._isscalar = isscalar
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221
        if hasattr(self.graph, "_var_cache"):
            self.graph._var_cache[node] = self

    @property
    def graph(self) -> Graph:
        return self._node.graph

    @property
    def op(self):
        if hasattr(self.graph, "_wrap"):
            return self.graph._wrap(self._node.owner)
        else:
            return self._node.owner

    @property
    def name(self):
        return self._node.name

    @property
    def id(self):
        return self._node.id

    @name.setter
    def name(self, name):
        self._node.name = name

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

    @property
    def device(self):
        return as_device(self._node.comp_node)

    @property
    def shape(self):
        return self._node.shape

    @property
    def value(self):
        return self._node.value


class OpNode:
    def __init__(self, node: _imperative_rt.OperatorNode):
        self._node = node
        if hasattr(self.graph, "_op_cache"):
            self.graph._op_cache[node] = self

    @property
    def graph(self) -> Graph:
        return self._node.graph

    @property
    def name(self):
        return self._node.name

    @property
    def id(self):
        return self._node.id

    @name.setter
    def name(self, name):
        self._node.name = name

    @property
    def inputs(self):
        if hasattr(self.graph, "_wrap"):
            return tuple(map(self.graph._wrap, self._node.inputs))
        else:
            return self._node.inputs

    @property
    def outputs(self):
        if hasattr(self.graph, "_wrap"):
            return tuple(map(self.graph._wrap, self._node.outputs))
        else:
            return self._node.outputs

    @property
    def params(self):
        return json.loads(self._node.params)

    @property
    def type(self):
        return self._node.type


222
def optimize_for_inference(dest_vars, **kwargs):
223 224
    r"""
    Applies optimize_for_inference pass for computing graph.
225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267

        :param dest_vars: list of output vars in the computing graph

        :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)
    """
    inference_options = GraphOptimizeOptions()
268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
    inference_optimize_layout_transform_map = {
        "enable_hwcd4": GraphOptimizeOptions.LayoutTransform.NHWCD4,
        "enable_nchw4": GraphOptimizeOptions.LayoutTransform.NCHW4,
        "enable_nchw88": GraphOptimizeOptions.LayoutTransform.NCHW88,
        "enable_nchw32": GraphOptimizeOptions.LayoutTransform.NCHW32,
        "enable_nchw44": GraphOptimizeOptions.LayoutTransform.NCHW44,
        "enable_nchw44_dot": GraphOptimizeOptions.LayoutTransform.NCHW44_DOT,
        "enable_chwn4": GraphOptimizeOptions.LayoutTransform.CHWN4,
    }

    for k, v in inference_optimize_layout_transform_map.items():
        if kwargs.pop(k, False):
            inference_options.layout_transform = v

    if kwargs.pop("enable_io16xc32", False):
        inference_options.f16_io_f32_comp = True
    if kwargs.pop("enable_ioc16", False):
        inference_options.f16_io_comp = True
    if kwargs.pop("enable_fuse_conv_bias_nonlinearity", False):
        inference_options.fuse_conv_bias_nonlinearity = True
    if kwargs.pop("enable_fuse_conv_bias_with_z", False):
        inference_options.fuse_conv_bias_with_z = True

    if kwargs:
        raise ValueError("unknown options: %s" % list(kwargs))
293

294
    dest_vars = _unwrap(dest_vars)
295
    res_vars = _imperative_rt.optimize_for_inference(dest_vars, inference_options)
296
    return _wrap(res_vars)
297 298


299 300 301 302 303 304 305 306 307 308 309 310 311
def modify_opr_algo_strategy_inplace(dest_vars, strategy: str):
    """
    C++ graph version of :func:`~.set_execution_strategy`. Used to inplacely modify
    dumped graph's fast-run strategy.

    :param dest_vars: list of output vars in the computing graph.
    :param strategy: fast-run algorithms strategy.

    """
    dest_vars = _unwrap(dest_vars)
    _imperative_rt.modify_opr_algo_strategy_inplace(dest_vars, strategy)


312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329
CompGraphDumpResult = collections.namedtuple(
    "CompGraphDumpResult",
    [
        "nr_opr",
        "tot_bytes",
        "tensor_value_bytes",
        "content_hash",
        "inputs",
        "outputs",
        "params",
    ],
)


def dump_graph(
    output_vars: Union[Dict[str, VarNode], List[VarNode]],
    *,
    keep_var_name: int = 1,
330
    keep_opr_name: bool = False,
331 332
    keep_param_name: bool = False,
    keep_opr_priority: bool = False,
333 334
    strip_info_file=None,
    append_json=False
335
) -> Tuple[bytes, CompGraphDumpResult]:
336 337
    """
    serialize the computing graph of `output_vars` and get byte result.
338 339 340 341 342 343 344 345 346 347 348 349 350

    :param output_vars: output variables which are the graph's end point.

        .. note::

            The underlying C++ API only accepts a var list. If a dict is given,
            the vars would be renamed to the given names.

    :param keep_var_name: level for keeping variable names:

        * 0: none of the names are kept
        * 1: (default)keep names of output vars
        * 2: keep names of all (output and internal) vars
351
    :param keep_opr_name: whether to keep operator names.
352 353 354 355 356
    :param keep_param_name: whether to keep param names, so param values can be
        easily manipulated after loading model
    :param keep_opr_priority: whether to keep priority setting for operators
    :param strip_info_file: a string for path or a file handler. if is not None,
        then the dump information for code strip would be written to ``strip_info_file``
357 358 359
    :param append_json: will be check when `strip_info_file` is not None. if set
        true, the information for code strip will be append to strip_info_file.
        if set false, will rewrite strip_info_file
360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378
    :return: dump result as byte string, and an instance of namedtuple
        :class:`CompGraphDumpResult`, whose fields are:

            * ``nr_opr`` number of operators dumped
            * ``tot_bytes`` total bytes for the whole graph
            * ``tensor_value_bytes`` bytes consumed for dumping tensor values
            * ``inputs`` names of input tensors
            * ``params`` list of names of dumped params
            * ``outputs`` names of output vars
    """
    if isinstance(output_vars, dict):
        used_vars = set()
        for name, var in output_vars.items():
            assert var.id not in used_vars, (
                "var name is associated with a var object, so we can not have "
                "two names given to the same var: {}".format(var)
            )
            used_vars.add(var.id)
            var.name = name
379
        output_vars = list(output_vars.values())
380
    else:
381 382 383
        output_vars = list(output_vars)

    ov = _unwrap(output_vars)
384 385 386 387 388 389 390 391 392

    stat = []
    inputs = []
    outputs = []
    params = []

    dump_content = _imperative_rt.dump_graph(
        ov,
        keep_var_name,
393
        keep_opr_name,
394 395 396 397 398 399 400 401 402 403 404 405
        keep_param_name,
        keep_opr_priority,
        stat,
        inputs,
        outputs,
        params,
    )

    dump_info = CompGraphDumpResult(*stat, inputs, outputs, params)

    if strip_info_file is not None:
        if isinstance(strip_info_file, str):
406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
            if not os.path.exists(strip_info_file):
                os.mknod(strip_info_file)
            strip_info_file = open(strip_info_file, "r+")
        new_strip_dict = json.loads(_imperative_rt.get_info_for_strip(ov))
        ori_strip_dict = new_strip_dict
        json_content = strip_info_file.read()
        if append_json and len(json_content) != 0:
            # if there are contents in json file. Read them first and then append new information
            ori_strip_dict = json.loads(json_content)
            for k in ori_strip_dict:
                new_strip_dict_v = new_strip_dict.get(k)
                if new_strip_dict_v is not None:
                    for value in new_strip_dict_v:
                        if not value in ori_strip_dict[k]:
                            ori_strip_dict[k].append(value)
        ori_strip_dict["hash"] = dump_info.content_hash
        strip_info_file.seek(0)
        strip_info_file.truncate()
        json.dump(ori_strip_dict, strip_info_file)
425 426

    return dump_content, dump_info
427

428

429 430 431 432 433
CompGraphLoadResult = collections.namedtuple(
    "CompGraphLoadResult", ["graph", "output_vars_dict", "output_vars_list"]
)


434
def load_graph(fpath) -> CompGraphLoadResult:
435 436
    """
    Load a serialized computing graph from file.
437

438
    :param fpath: Path or Handle of the input file
439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456
    :return: An instance of namedtuple :class:`CompGraphLoadResult`,
        whose fields are:

            * ``graph`` loaded CompGraph
            * ``output_vars_dict`` A Python dict, mapping name to output SymbolVar
            * ``output_vars_list`` A Python list, containing output vars in the
                                   order passed to serialize_comp_graph_to_file
    """
    output_vars_map = []
    output_vars_list = []
    if isinstance(fpath, str):
        buf = open(fpath, "rb").read()
    else:
        buf = fpath.read()
    cg = _imperative_rt.load_graph(buf, output_vars_map, output_vars_list)
    return CompGraphLoadResult(cg, dict(output_vars_map), output_vars_list)


457
def _wrap(x):
M
Megvii Engine Team 已提交
458
    if isinstance(x, collections.abc.Sequence):
459
        return type(x)(map(_wrap, x))
460 461 462 463
    if hasattr(x.graph, "_wrap"):
        return x.graph._wrap(x)
    else:
        return x
464 465 466


def _unwrap(x):
M
Megvii Engine Team 已提交
467
    if isinstance(x, collections.abc.Sequence):
468
        return type(x)(map(_unwrap, x))
469 470 471 472
    if isinstance(x, VarNode):
        return x._node
    else:
        return x
473 474


475
def apply_normal_varnode(op: OpDef, *args: VarNode):
476 477 478
    # for PyOp like RemoteSend/Recv
    if getattr(op, "op", None):
        op = op.op
479 480 481 482
    outputs = _imperative_rt.invoke_op(op, _unwrap(args))
    return _wrap(outputs)


483
def apply_backward_varnode(op: BackwardGraph, *args: VarNode):
484 485
    assert args
    graph = args[0].graph
486 487
    outputs = op.interpret(
        op,
488
        lambda op, args: apply_normal_varnode(op, *args),
489 490
        graph._make_const_for_backward,
        args,
491
    )
492
    return outputs
493 494 495


set_cpp_apply_backward_varnode(apply_backward_varnode)
496 497


M
Megvii Engine Team 已提交
498
def input_callback(callback, *args, device=None, dtype=None, shape=None, graph=None):
499
    outputs = _imperative_rt.input_callback(
M
Megvii Engine Team 已提交
500
        callback, as_device(device).to_c(), dtype, shape, _unwrap(args), graph=graph
501 502 503 504 505 506
    )
    value, dummy = _wrap(outputs)
    return value, dummy


class InputNode(OpNode):
507 508 509 510 511 512 513 514 515
    def __init__(
        self,
        *args: VarNode,
        device=None,
        dtype=None,
        shape=None,
        graph=None,
        use_static_shape=False
    ):
516 517 518 519
        r = _imperative_rt.DeviceTensorNDRendezvous()
        if device is not None:
            device = as_device(device).to_c()
        outputs = _imperative_rt.input_callback(
520 521 522 523 524 525 526
            r,
            device,
            dtype,
            shape,
            _unwrap(args),
            graph=graph,
            use_static_shape=use_static_shape,
527 528 529 530 531 532 533 534 535 536 537 538 539
        )
        super().__init__(outputs[0].owner)
        self._rendezvous = r

    def set_value(self, value):
        assert isinstance(value, _imperative_rt.DeviceTensorND)
        self._rendezvous.set(value)

    def reset(self):
        self._rendezvous.reset()

    @property
    def device(self):
540 541 542 543 544
        var = self.outputs[0]
        if isinstance(var, VarNode):
            return var.device
        else:
            return var.comp_node
545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567

    @property
    def dtype(self):
        return self.outputs[0].dtype


def output_callback(callback, var, *args):
    args = (var,) + args
    dummy = _imperative_rt.output_callback(callback, _unwrap(args))
    return _wrap(dummy)


class OutputNode(OpNode):
    def __init__(self, var, *args):
        args = (var,) + args
        r = _imperative_rt.DeviceTensorNDRendezvous()
        dummy = _imperative_rt.output_callback(r, _unwrap(args))
        super().__init__(dummy.owner)
        self._rendezvous = r

    def get_value(self):
        return self._rendezvous.get()

M
Megvii Engine Team 已提交
568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590
    def drop_value(self):
        self._rendezvous.drop()

    def reset(self):
        self._rendezvous.reset()


class ValueOutputNode(OpNode):
    def __init__(self, var, *args):
        args = (var,) + args
        r = _imperative_rt.HostTensorNDRendezvous()
        dummy = _imperative_rt.value_output_callback(r, _unwrap(args))
        super().__init__(dummy.owner)
        self._rendezvous = r

    def get_value(self):
        hostnd, event = self._rendezvous.get()
        event.wait()
        return hostnd.numpy()

    def drop_value(self):
        self._rendezvous.drop()

591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613
    def reset(self):
        self._rendezvous.reset()


class TensorAttr:
    def __init__(self, shape, dtype, device):
        self.shape = shape
        self.dtype = dtype
        self.device = device


class AttrOutputNode(OpNode):
    def __init__(self, var, *args):
        args = (var,) + args
        r = _imperative_rt.TensorAttrRendezvous()
        dummy = _imperative_rt.attr_output_callback(r, _unwrap(args))
        super().__init__(dummy.owner)
        self._rendezvous = r

    def get_value(self):
        attr = self._rendezvous.get()
        return TensorAttr(attr.shape, attr.dtype, as_device(attr.comp_node))

M
Megvii Engine Team 已提交
614 615 616
    def drop_value(self):
        self._rendezvous.drop()

617 618
    def reset(self):
        self._rendezvous.reset()
619 620 621 622 623 624


class VirtualDepNode(OpNode):
    def __init__(self, vars, device=""):
        out = _imperative_rt.virtual_dep(_unwrap(vars), device)
        super().__init__(out)