memory_optimization_transpiler.py 21.6 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# 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.

15 16
from __future__ import print_function

17
import six
18
from collections import defaultdict, MutableSet
19
from .. import core
M
minqiyang 已提交
20
from ... import compat as cpt
21
from ..framework import Program, default_main_program, Parameter, Variable, core
22
from ..backward import _rename_arg_
23
from functools import reduce
24
from six.moves import range
25 26

dtype_to_size = {
27 28 29 30 31 32
    core.VarDesc.VarType.FP16: 2,
    core.VarDesc.VarType.FP32: 4,
    core.VarDesc.VarType.FP64: 8,
    core.VarDesc.VarType.INT16: 2,
    core.VarDesc.VarType.INT32: 4,
    core.VarDesc.VarType.INT64: 8,
33 34
    core.VarDesc.VarType.BOOL: 1,
    core.VarDesc.VarType.UINT8: 1,
35
}
36

37
SUB_BLOCK_OPS = [
X
Xin Pan 已提交
38
    "while", "while_grad", "conditional_block", "conditional_block_grad"
39
]
40

X
Xin Pan 已提交
41
SUB_BLOCK_PAIR = [("while", "while_grad"),
42 43
                  ("conditional_block", "conditional_block_grad")]

Q
qiaolongfei 已提交
44
PRINT_LOG = False
D
dzhwinter 已提交
45
FLAGS_memory_optimize = ""
Q
qiaolongfei 已提交
46

47

48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
class OrderedSet(MutableSet):
    def __init__(self, iterable=None):
        self.end = end = []
        end += [None, end, end]  # sentinel node for doubly linked list
        self.map = {}  # key --> [key, prev, next]
        if iterable is not None:
            self |= iterable

    def __len__(self):
        return len(self.map)

    def __contains__(self, key):
        return key in self.map

    def add(self, key):
        if key not in self.map:
            end = self.end
            curr = end[1]
            curr[2] = end[1] = self.map[key] = [key, curr, end]

    def update(self, other):
        for e in other:
            self.add(e)

    def discard(self, key):
        if key in self.map:
            key, prev, next = self.map.pop(key)
            prev[2] = next
            next[1] = prev

    def remove(self, key):
        self.discard(key)

    def __iter__(self):
        end = self.end
        curr = end[2]
        while curr is not end:
            yield curr[0]
            curr = curr[2]

    def __reversed__(self):
        end = self.end
        curr = end[1]
        while curr is not end:
            yield curr[0]
            curr = curr[1]

    def pop(self, last=True):
        if not self:
            raise KeyError('set is empty')
        key = self.end[1][0] if last else self.end[2][0]
        self.discard(key)
        return key

    def __repr__(self):
        if not self:
            return '%s()' % (self.__class__.__name__, )
        return '%s(%r)' % (self.__class__.__name__, list(self))

    def __eq__(self, other):
        if isinstance(other, OrderedSet):
            return len(self) == len(other) and list(self) == list(other)
        return set(self) == set(other)


113
class ControlFlowGraph(object):
114 115
    def __init__(self, program, ops, forward_num, skip_opt):
        self._program = program
116 117
        self._ops = ops
        self._forward_num = forward_num
118 119 120 121 122 123
        self._successors = defaultdict(OrderedSet)
        self._presuccessors = defaultdict(OrderedSet)
        self._uses = defaultdict(OrderedSet)
        self._defs = defaultdict(OrderedSet)
        self._live_in = defaultdict(OrderedSet)
        self._live_out = defaultdict(OrderedSet)
D
dzhwinter 已提交
124

125
        self._skip_opt = skip_opt
D
dzhwinter 已提交
126
        self.pool = []
127 128

    def _add_connections(self, connections):
129
        """Populates _successors and _presuccessors for two neighbor nodes."""
130 131 132 133
        for node1, node2 in connections:
            self._add(node1, node2)

    def _add(self, node1, node2):
134 135
        self._successors[node1].add(node2)
        self._presuccessors[node2].add(node1)
136

137 138
    # TODO(panyx0718): We need to have a unified way of building intermediate
    # representation.
139
    def _build_graph(self):
140 141
        """Build a graph based on op sequence.
        """
142
        self.op_size = len(self._ops)
143 144 145
        op_node_connections = [(i, i + 1) for i in range(self.op_size - 1)]
        self._add_connections(op_node_connections)
        for i in range(self.op_size):
146 147
            self._uses[i].update(self._ops[i].input_arg_names())
            self._defs[i].update(self._ops[i].output_arg_names())
148

149 150 151 152 153 154 155 156 157 158
    def _update_graph(self, old_name, new_name, begin_idx=0):
        for i in range(begin_idx, self.op_size):
            if old_name in self._uses[i]:
                self._uses[i].remove(old_name)
                self._uses[i].add(new_name)
            if old_name in self._defs[i]:
                self._defs[i].remove(old_name)
                self._defs[i].add(new_name)
            if old_name in self._live_in[i]:
                self._live_in[i].remove(old_name)
D
dzhwinter 已提交
159
                self._live_in[i].add(new_name)
160 161 162 163
            if old_name in self._live_out[i]:
                self._live_out[i].remove(old_name)
                self._live_out[i].add(new_name)

164 165 166
    def _dataflow_analyze(self):
        self._build_graph()
        live_in = defaultdict(set)
D
dzhwinter 已提交
167 168 169 170 171 172 173 174
        worklist = list(range(len(self._ops) - 1, -1, -1))
        while worklist:
            i = worklist.pop(0)
            live_in[i] = set(self._live_in[i])
            for s in self._successors[i]:
                self._live_out[i] |= self._live_in[s]
            self._live_in[i] = self._uses[i] | (
                self._live_out[i] - self._defs[i])
D
dongzhihong 已提交
175
            if live_in[i] != set(self._live_in[i]):
D
dzhwinter 已提交
176 177
                for d in self._presuccessors[i]:
                    worklist.append(d)
178

D
dzhwinter 已提交
179
    def _fill_pool(self, i, is_forward):
D
dzhwinter 已提交
180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
        def comparator(x, cache):
            x_shape = x[1]
            cache_shape = cache[1]
            x_size = abs(reduce(lambda x, y: x * y, x_shape))
            cache_size = abs(reduce(lambda x, y: x * y, cache_shape))
            if (x_shape[0] == -1 and cache_shape[0] == -1) or \
               (x_shape[0] != -1 and cache_shape[0] != -1) :
                return x_size <= cache_size
            else:
                return False

        def find_var_in_block(x):
            known_vars = set()
            for op in self._ops:
                known_vars.update(op.output_arg_names())
            return x in known_vars

D
dzhwinter 已提交
197 198
        block_desc = self._ops[i].block()
        in_diff, _ = self._get_diff(self._live_in[i], self._live_out[i])
199 200
        # NOTE: must sort the in_diff set for cases that get different cache var.
        # FIXME(typhoonzero): maybe use a "sorted set" is better than this.
D
dzhwinter 已提交
201
        can_optimize = [
D
dzhwinter 已提交
202
            x for x in sorted(in_diff)
D
dzhwinter 已提交
203 204 205 206
            if self._check_var_validity(block_desc, x, is_forward)
        ]
        if can_optimize:
            for var_name in can_optimize:
D
dzhwinter 已提交
207 208
                cache = (var_name, self._find_var(block_desc, var_name,
                                                  is_forward).shape())
D
dzhwinter 已提交
209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
                if cache not in self.pool and find_var_in_block(var_name):
                    i = 0
                    while i < len(self.pool):
                        mycache = self.pool[i]
                        mysize = mycache[1][0]
                        cache_size = cache[1][0]
                        if (mysize == -1 and cache_size == -1) or \
                           (mysize != -1 and cache_size != -1):
                            if comparator(mycache, cache):
                                i += 1
                            else:
                                break
                        elif mysize == -1 and cache_size != -1:
                            i += 1
                        elif mysize != -1 and cache_size == -1:
                            break
                    self.pool.insert(i, cache)
226 227 228 229 230

    def _get_diff(self, a, b):
        u = a & b
        return a - u, b - u

231 232
    def _has_var(self, block_desc, var_name, is_forward):
        if is_forward:
M
minqiyang 已提交
233
            return block_desc.has_var(cpt.to_bytes(var_name))
234
        else:
M
minqiyang 已提交
235
            return block_desc.has_var_recursive(cpt.to_bytes(var_name))
236 237 238

    def _find_var(self, block_desc, var_name, is_forward):
        if is_forward:
M
minqiyang 已提交
239
            return block_desc.find_var(cpt.to_bytes(var_name))
240
        else:
M
minqiyang 已提交
241
            return block_desc.find_var_recursive(cpt.to_bytes(var_name))
242

243 244 245 246 247 248 249 250 251 252 253 254 255 256 257
    def _check_var_validity(self, block_desc, x, is_forward):
        if str(x) == "@EMPTY@":
            return False
        if not self._has_var(block_desc, x, is_forward):
            return False
        if self._find_var(block_desc, x, is_forward).persistable():
            return False
        if self._find_var(block_desc, x,
                          is_forward).type() != core.VarDesc.VarType.LOD_TENSOR:
            return False
        if x in self._skip_opt:
            return False
        if not self._find_var(block_desc, x, is_forward).shape():
            return False
        return True
258

259 260
    # TODO(panyx0718): This needs to be less hacky. It seems memory optimization
    # doesn't consider vars copied between cpu and gpu.
261 262 263
    def _update_skip_opt_set(self):
        for i in range(self.op_size):
            op = self._ops[i]
D
dzhwinter 已提交
264
            if op.has_attr("force_cpu") and op.attr("force_cpu") == True:
265 266
                self._skip_opt.update(op.output_arg_names())

267
    def release_memory(self, skip_opt_set=None):
268
        self._dataflow_analyze()
269
        self._update_skip_opt_set()
270 271
        if skip_opt_set:
            self._skip_opt.update(skip_opt_set)
272 273 274 275
        fwd_id = 0
        bwd_id = 0
        for i in range(self.op_size):
            op = self._ops[i]
276
            if op.type() in SUB_BLOCK_OPS:
277 278 279 280 281
                continue
            block_desc = op.block()
            is_forward = i < self._forward_num
            in_diff, out_diff = self._get_diff(self._live_in[i],
                                               self._live_out[i])
282 283 284 285
            can_optimize = [
                x for x in in_diff
                if self._check_var_validity(block_desc, x, is_forward)
            ]
286 287
            if can_optimize:
                index = i + fwd_id + 1 if is_forward else i - self._forward_num + bwd_id + 1
W
Wu Yi 已提交
288
                delete_op = block_desc._insert_op(index)
289 290 291 292 293 294 295
                delete_op.set_type("delete_var")
                delete_op.set_input("X", can_optimize)
                if is_forward:
                    fwd_id += 1
                else:
                    bwd_id += 1

296
    def memory_optimize(self, skip_opt_set=None, level=0):
297 298 299
        def compare_shape(x_shape, cache_shape, opt_level):
            if opt_level == 0:
                return x_shape == cache_shape
300
            elif opt_level == 1:
301 302 303 304 305 306
                if (x_shape[0] == -1) ^ (cache_shape[0] == -1):
                    return False
                x_size = abs(reduce(lambda x, y: x * y, x_shape))
                cache_size = abs(reduce(lambda x, y: x * y, cache_shape))
                if x_size <= cache_size:
                    return True
307 308
            else:
                raise ValueError("only support opt_level 0 or 1.")
309 310 311 312
            return False

        self._dataflow_analyze()
        self._update_skip_opt_set()
313 314 315
        # update skip set to meet users' demand
        if skip_opt_set:
            self._skip_opt.update(skip_opt_set)
D
dzhwinter 已提交
316
        counter = 0
317
        for i in range(self.op_size):
318
            op = self._ops[i]
319
            if op.type() in SUB_BLOCK_OPS:
320 321 322
                continue
            block_desc = op.block()
            is_forward = i < self._forward_num
323
            if self.pool:
324
                # NOTE: must sort the in_diff set for cases that get different cache var.
325
                defs_can_optimize = [
326
                    x for x in self._defs[i]
327 328
                    if self._check_var_validity(block_desc, x, is_forward)
                ]
329 330 331 332
                out_pair = [
                    (x, self._find_var(block_desc, x, is_forward).shape())
                    for x in defs_can_optimize
                ]
333
                for x, x_shape in out_pair:
334 335 336
                    # If x is both in uses and defs, it can not be optimized!
                    if x in self._uses[i]:
                        continue
D
dzhwinter 已提交
337 338 339
                    if x == FLAGS_memory_optimize:
                        print("start match var ", x, " of op ", op.type())
                        print(self.pool)
340 341 342
                    for index, cache_pair in enumerate(self.pool):
                        cache_var = cache_pair[0]
                        cache_shape = cache_pair[1]
343
                        if not self._has_var(block_desc, cache_var, is_forward):
D
"rerun"  
dzhwinter 已提交
344 345 346
                            if PRINT_LOG:
                                print("cache %s not exists!" %
                                      (cpt.to_text(cache_var)))
347
                            continue
D
dzhwinter 已提交
348
                        if x == cache_var:
D
"rerun"  
dzhwinter 已提交
349 350 351 352
                            if PRINT_LOG:
                                print("x : ", cpt.to_text(x), " cache : ",
                                      cpt.to_text(cache_var), " is same var!")
                            break
353 354 355 356 357

                        x_dtype = self._find_var(block_desc, x,
                                                 is_forward).dtype()
                        cache_dtype = self._find_var(block_desc, cache_var,
                                                     is_forward).dtype()
D
dzhwinter 已提交
358 359 360 361
                        if x_dtype != cache_dtype:
                            if PRINT_LOG:
                                print("x_dtype and cache_dtyp are different")
                            continue
D
dzhwinter 已提交
362 363 364

                        if not compare_shape(x_shape, cache_shape, level):
                            continue
D
dongzhihong 已提交
365
                        # TODO(qijun): dtype_to_size[x_dtype] and dtype_to_size[cache_dtype]
366
                        if PRINT_LOG:
D
dzhwinter 已提交
367 368 369 370 371 372
                            print(
                                ("!!! %d,  %s => %s, cache idx %d, pool size %d"
                                 % (counter, x + str(x_shape),
                                    cache_var + str(cache_shape), index,
                                    len(self.pool))))
                            counter += 1
373 374 375 376
                        self.pool.pop(index)
                        # Rename the var to the cache var already with
                        # memory allocated in order to reuse the memory.
                        _rename_arg_(self._ops, x, cache_var, begin_idx=i)
M
minqiyang 已提交
377 378 379
                        self._program.block(block_desc.id).var(cpt.to_text(
                            x)).desc = self._find_var(block_desc, cache_var,
                                                      is_forward)
380 381
                        self._program.block(block_desc.id).vars[cpt.to_text(x)] = \
                            Variable(self._program.block(block_desc.id), name=cpt.to_text(x))
382 383
                        self._update_graph(x, cache_var, begin_idx=i)
                        break
D
dzhwinter 已提交
384
            self._fill_pool(i, is_forward)
385 386


387
def _process_sub_block_pair(pdesc, sub_block_pair):
388 389 390 391 392 393 394 395 396 397 398 399 400
    """Creates a list of tuple each of which tracks info of a subblock.

      Note: this function doesn't handle nested subblocks yet.
      TODO(panyx0718): assert if case nested subblocks happen.

    :param pdesc: ProgramDesc.
    :param sub_block_pair: A list op pairs. Each op pair is the forward
        op and backward op. The ops in the list are special that they contain
        a subblock of ops.
    :return: A list of tuples, each tuple is (all ops in a subblock pair
        including forward and backward, number of forward ops,
        all output args names of the ops in the subblock pairs).
    """
401 402 403
    ops_list = []
    block_desc = pdesc.block(0)
    op_size = block_desc.op_size()
404 405 406 407 408 409 410 411 412 413 414 415 416
    for fwd_op, bwd_op in sub_block_pair:
        sub_block_ids = []
        grad_sub_block_ids = []
        sub_block_id_pair = []
        sub_op_dict = {}
        for i in range(op_size):
            op = block_desc.op(i)
            if op.type() == fwd_op:
                sub_block_ids.append(op.attr("sub_block").id)
                sub_op_dict[op.attr("sub_block").id] = op
            elif op.type() == bwd_op:
                grad_sub_block_ids.append(op.attr("sub_block").id)
                sub_op_dict[op.attr("sub_block").id] = op
417

418 419
        # Find fwd_op/bwd_op block pair
        for grad_id in grad_sub_block_ids:
Q
qijun 已提交
420 421 422 423
            fwd_id = pdesc.block(grad_id).get_forward_block_idx()
            if fwd_id in sub_block_ids:
                sub_block_id_pair.append((fwd_id, grad_id))
                sub_block_ids.remove(fwd_id)
424

425
        # Get fwd_op/bwd_op block ops
Q
qijun 已提交
426
        for fwd_id, grad_id in sub_block_id_pair:
427
            sub_block_ops = []
Q
qijun 已提交
428
            sub_block = pdesc.block(fwd_id)
429 430 431
            block_op_size = sub_block.op_size()
            for i in range(block_op_size):
                sub_block_ops.append(sub_block.op(i))
432

433 434 435 436
            grad_sub_block = pdesc.block(grad_id)
            grad_sub_block_op_size = grad_sub_block.op_size()
            for i in range(grad_sub_block_op_size):
                sub_block_ops.append(grad_sub_block.op(i))
437

438
            sub_op_output = set()
Q
qijun 已提交
439
            sub_op_output.update(sub_op_dict[fwd_id].output_arg_names())
440
            sub_op_output.update(sub_op_dict[grad_id].output_arg_names())
441 442
            sub_op_output.update(sub_op_dict[fwd_id].input_arg_names())
            sub_op_output.update(sub_op_dict[grad_id].input_arg_names())
443
            ops_list.append((sub_block_ops, block_op_size, sub_op_output))
444

445
        # Process rest fwd_op block ops
Q
qijun 已提交
446
        for fwd_id in sub_block_ids:
447
            sub_block_ops = []
Q
qijun 已提交
448
            sub_block = pdesc.block(fwd_id)
449 450 451 452
            sub_block_op_size = sub_block.op_size()
            for i in range(sub_block_op_size):
                sub_block_ops.append(sub_block.op(i))
            sub_op_output = set()
Q
qijun 已提交
453
            sub_op_output.update(sub_op_dict[fwd_id].output_arg_names())
454
            sub_op_output.update(sub_op_dict[fwd_id].input_arg_names())
455 456
            ops_list.append((sub_block_ops, sub_block_op_size, sub_op_output))
    return ops_list
457

458

459
def _get_cfgs(input_program):
460 461 462 463 464
    """Process each block and create ControlFlowGraph for each of them.

    :param input_program: Program object.
    :return: A list of ControlFlowGraph, each corresponds to a block.
    """
465
    ops_list = []
W
Wu Yi 已提交
466
    pdesc = input_program._get_desc()
467 468
    block_desc = pdesc.block(0)
    op_size = block_desc.op_size()
469

470 471
    # Only process one level of nested subblock.
    ops_list.extend(_process_sub_block_pair(pdesc, SUB_BLOCK_PAIR))
472

473 474 475 476 477 478 479
    skip_opt_set = set()
    for _, _, skip_opt in ops_list:
        skip_opt_set.update(skip_opt)

    # Get global block ops
    ops_list.insert(
        0, ([block_desc.op(i) for i in range(op_size)], op_size, skip_opt_set))
480 481 482 483
    cfgs = [
        ControlFlowGraph(input_program, ops, forward_num, skip_opt)
        for ops, forward_num, skip_opt in ops_list
    ]
484
    return cfgs
485 486


487 488 489 490 491 492 493 494 495 496 497 498 499
def _is_opt_role_op(op):
    op_maker = core.op_proto_and_checker_maker
    optimize_role = core.op_proto_and_checker_maker.OpRole.Optimize
    if op_maker.kOpRoleAttrName() in op.attr_names and \
            int(op.all_attrs()[op_maker.kOpRoleAttrName()]) == int(optimize_role):
        return True


def memory_optimize(input_program,
                    skip_opt_set=None,
                    print_log=False,
                    level=0,
                    skip_grads=False):
500 501 502 503
    """Optimize memory by reusing var memory.

      Note: it doesn't not support subblock nested in subblock.

D
"rerun"  
dzhwinter 已提交
504 505 506 507 508 509 510
    Args:
        input_program(str): Input Program
        skip_opt_set(set): vars wil be skipped in memory optimze
        print_log(bool): whether to print debug log.
        level(int): If level=0, reuse if the shape is completely equal, o
    Returns:
        None
511
    """
512 513 514 515 516 517 518 519 520 521 522

    def to_name_str(var):
        if isinstance(var, Variable):
            return var.desc.name()
        elif isinstance(var, str):
            return var
        elif isinstance(var, six.string_types):
            return str(var)
        else:
            raise TypeError(str(var) + " should be Variable or str")

523 524
    if level != 0 and level != 1:
        raise ValueError("only support opt_level 0 or 1.")
D
dzhwinter 已提交
525 526 527 528 529
    if skip_opt_set is not None:
        if isinstance(skip_opt_set, set) or isinstance(skip_opt_set, list):
            skip_opt_set = set(skip_opt_set)
        else:
            raise ValueError("only support skip_opt_set as set.")
Q
qiaolongfei 已提交
530 531
    global PRINT_LOG
    PRINT_LOG = print_log
532 533 534 535 536 537 538 539 540 541 542 543
    if skip_grads:
        grad_set = set()
        OP_ROLE_VAR = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
        for op in input_program.global_block().ops:
            if _is_opt_role_op(op):
                if op.attr(OP_ROLE_VAR):
                    grad_name = op.attr(OP_ROLE_VAR)[1]
                    grad_set.add(grad_name)
        if not skip_opt_set:
            skip_opt_set = grad_set
        else:
            skip_opt_set.update(grad_set)
544 545
    if skip_opt_set is not None:
        skip_opt_set = set(map(to_name_str, skip_opt_set))
546
    cfgs = _get_cfgs(input_program)
D
dzhwinter 已提交
547
    input_program._is_mem_optimized = True
548
    for cfg in cfgs:
549
        cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level)
550 551


552
def release_memory(input_program, skip_opt_set=None):
Y
yuyang18 已提交
553 554 555 556 557 558 559 560 561
    """
    Modify the input program and insert :code:`delete_op` to early drop not used
    variables. The modification will be performed inplace.

    Notes: This is an experimental API and could be removed in next few
    releases. Users should not use this API.

    Args:
        input_program(Program): The program will be inserted :code:`delete_op`.
D
"rerun"  
dzhwinter 已提交
562 563 564
        skip_opt_set(set): vars wil be skipped in memory optimze
    Returns:
        None
Y
yuyang18 已提交
565
    """
566
    cfgs = _get_cfgs(input_program)
D
dzhwinter 已提交
567
    input_program._is_mem_optimized = True
568
    for cfg in cfgs:
569
        cfg.release_memory(skip_opt_set=skip_opt_set)