memory_optimization_transpiler.py 16.0 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
from collections import defaultdict
16
from .. import core
W
Wu Yi 已提交
17
from ..framework import Program, default_main_program, Parameter
18
from ..backward import _rename_arg_
19
from functools import reduce
20
from six.moves import range
21 22

dtype_to_size = {
23 24 25 26 27 28
    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,
29 30
    core.VarDesc.VarType.BOOL: 1,
    core.VarDesc.VarType.UINT8: 1,
31
}
32

33
SUB_BLOCK_OPS = [
34 35 36
    "while", "while_grad", "parallel_do", "parallel_do_grad",
    "conditional_block", "conditional_block_grad"
]
37

38 39 40
SUB_BLOCK_PAIR = [("while", "while_grad"), ("parallel_do", "parallel_do_grad"),
                  ("conditional_block", "conditional_block_grad")]

Q
qiaolongfei 已提交
41 42
PRINT_LOG = False

43 44

class ControlFlowGraph(object):
45 46
    def __init__(self, program, ops, forward_num, skip_opt):
        self._program = program
47 48 49 50
        self._ops = ops
        self._forward_num = forward_num
        self._successors = defaultdict(set)
        self._presuccessors = defaultdict(set)
51 52 53 54
        self._uses = defaultdict(set)
        self._defs = defaultdict(set)
        self._live_in = defaultdict(set)
        self._live_out = defaultdict(set)
55
        self._skip_opt = skip_opt
56 57

    def _add_connections(self, connections):
58
        """Populates _successors and _presuccessors for two neighbor nodes."""
59 60 61 62
        for node1, node2 in connections:
            self._add(node1, node2)

    def _add(self, node1, node2):
63 64
        self._successors[node1].add(node2)
        self._presuccessors[node2].add(node1)
65

66 67
    # TODO(panyx0718): We need to have a unified way of building intermediate
    # representation.
68
    def _build_graph(self):
69 70
        """Build a graph based on op sequence.
        """
71
        self.op_size = len(self._ops)
72 73 74
        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):
75 76
            self._uses[i].update(self._ops[i].input_arg_names())
            self._defs[i].update(self._ops[i].output_arg_names())
77

78 79 80 81 82 83 84 85 86 87 88 89 90 91 92
    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)
                self._live_out[i].add(new_name)
            if old_name in self._live_out[i]:
                self._live_out[i].remove(old_name)
                self._live_out[i].add(new_name)

93
    def _reach_fixed_point(self, live_in, live_out):
94
        """Check if the liveness set has stablized."""
95 96 97 98 99
        if len(live_in) != len(self._live_in):
            return False
        if len(live_out) != len(self._live_out):
            return False
        for i in range(self.op_size):
100 101
            if (live_in[i] != self._live_in[i] or
                    live_out[i] != self._live_out[i]):
102 103 104 105 106 107 108
                return False
        return True

    def _dataflow_analyze(self):
        self._build_graph()
        live_in = defaultdict(set)
        live_out = defaultdict(set)
109 110
        # Repeatedly apply liveness updates until the algorithm stablize
        # on a complete set live input vars and live output vars.
111
        while True:
112
            for i in reversed(list(range(self.op_size))):
113 114
                live_in[i] = set(self._live_in[i])
                live_out[i] = set(self._live_out[i])
115
                for s in self._successors[i]:
116
                    self._live_out[i] |= self._live_in[s]
Q
QI JUN 已提交
117 118
                self._live_in[i] = self._uses[i] | (
                    self._live_out[i] - self._defs[i])
119 120 121 122 123 124 125
            if self._reach_fixed_point(live_in, live_out):
                break

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

126 127 128 129 130 131 132 133 134 135 136 137
    def _has_var(self, block_desc, var_name, is_forward):
        if is_forward:
            return block_desc.has_var(str(var_name))
        else:
            return block_desc.has_var_recursive(str(var_name))

    def _find_var(self, block_desc, var_name, is_forward):
        if is_forward:
            return block_desc.find_var(str(var_name))
        else:
            return block_desc.find_var_recursive(str(var_name))

138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
    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
153

154 155
    # TODO(panyx0718): This needs to be less hacky. It seems memory optimization
    # doesn't consider vars copied between cpu and gpu.
156 157 158 159 160 161
    def _update_skip_opt_set(self):
        for i in range(self.op_size):
            op = self._ops[i]
            if op.type() == "fill_constant" and op.attr("force_cpu") == True:
                self._skip_opt.update(op.output_arg_names())

162
    def release_memory(self, skip_opt_set=None):
163
        self._dataflow_analyze()
164
        self._update_skip_opt_set()
165 166
        if skip_opt_set:
            self._skip_opt.update(skip_opt_set)
167 168 169 170
        fwd_id = 0
        bwd_id = 0
        for i in range(self.op_size):
            op = self._ops[i]
171
            if op.type() in SUB_BLOCK_OPS:
172 173 174 175 176
                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])
177 178 179 180
            can_optimize = [
                x for x in in_diff
                if self._check_var_validity(block_desc, x, is_forward)
            ]
181 182
            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
183
                delete_op = block_desc._insert_op(index)
184 185 186 187 188 189 190
                delete_op.set_type("delete_var")
                delete_op.set_input("X", can_optimize)
                if is_forward:
                    fwd_id += 1
                else:
                    bwd_id += 1

191
    def memory_optimize(self, skip_opt_set=None, level=0):
192 193 194
        def compare_shape(x_shape, cache_shape, opt_level):
            if opt_level == 0:
                return x_shape == cache_shape
195
            elif opt_level == 1:
196 197 198 199 200 201
                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
202 203
            else:
                raise ValueError("only support opt_level 0 or 1.")
204 205 206 207
            return False

        self._dataflow_analyze()
        self._update_skip_opt_set()
208 209 210
        # update skip set to meet users' demand
        if skip_opt_set:
            self._skip_opt.update(skip_opt_set)
211 212
        self.pool = []
        for i in range(self.op_size):
213
            op = self._ops[i]
214
            if op.type() in SUB_BLOCK_OPS:
215 216 217
                continue
            block_desc = op.block()
            is_forward = i < self._forward_num
218
            if self.pool:
219 220 221 222
                defs_can_optimize = [
                    x for x in self._defs[i]
                    if self._check_var_validity(block_desc, x, is_forward)
                ]
223 224 225 226
                out_pair = [
                    (x, self._find_var(block_desc, x, is_forward).shape())
                    for x in defs_can_optimize
                ]
227
                for x, x_shape in out_pair:
228 229 230
                    # If x is both in uses and defs, it can not be optimized!
                    if x in self._uses[i]:
                        continue
231 232 233
                    for index, cache_pair in enumerate(self.pool):
                        cache_var = cache_pair[0]
                        cache_shape = cache_pair[1]
234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249
                        if not compare_shape(x_shape, cache_shape, level):
                            continue

                        if not self._has_var(block_desc, cache_var, is_forward):
                            continue

                        x_dtype = self._find_var(block_desc, x,
                                                 is_forward).dtype()
                        cache_dtype = self._find_var(block_desc, cache_var,
                                                     is_forward).dtype()
                        # TODO(qijun): actually, we should compare
                        # dtype_to_size[x_dtype] and dtype_to_size[cache_dtype]
                        if x_dtype != cache_dtype:
                            continue

                        if PRINT_LOG:
250 251 252 253 254
                            print(("Hit Cache !!!! cache pool index "
                                   "is %d, var name is %s, "
                                   "cached var name is %s, "
                                   "var shape is %s ") % (index, x, cache_var,
                                                          str(cache_shape)))
255 256 257 258 259 260 261 262 263 264 265 266 267
                        self.pool.pop(index)
                        if x == cache_var:
                            break
                        # 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)
                        self._program.block(block_desc.id).var(str(
                            x)).desc = self._find_var(block_desc, cache_var,
                                                      is_forward)
                        self._update_graph(x, cache_var, begin_idx=i)
                        break

            in_diff, _ = self._get_diff(self._live_in[i], self._live_out[i])
268 269 270 271
            can_optimize = [
                x for x in in_diff
                if self._check_var_validity(block_desc, x, is_forward)
            ]
272 273
            if can_optimize:
                for var_name in can_optimize:
274 275 276 277
                    self.pool.append((var_name, self._find_var(
                        block_desc, var_name, is_forward).shape()))


278
def _process_sub_block_pair(pdesc, sub_block_pair):
279 280 281 282 283 284 285 286 287 288 289 290 291
    """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).
    """
292 293 294
    ops_list = []
    block_desc = pdesc.block(0)
    op_size = block_desc.op_size()
295 296 297 298 299 300 301 302 303 304 305 306 307
    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
308

309 310
        # Find fwd_op/bwd_op block pair
        for grad_id in grad_sub_block_ids:
Q
qijun 已提交
311 312 313 314
            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)
315

316
        # Get fwd_op/bwd_op block ops
Q
qijun 已提交
317
        for fwd_id, grad_id in sub_block_id_pair:
318
            sub_block_ops = []
Q
qijun 已提交
319
            sub_block = pdesc.block(fwd_id)
320 321 322
            block_op_size = sub_block.op_size()
            for i in range(block_op_size):
                sub_block_ops.append(sub_block.op(i))
323

324 325 326 327
            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))
328

329
            sub_op_output = set()
Q
qijun 已提交
330
            sub_op_output.update(sub_op_dict[fwd_id].output_arg_names())
331
            sub_op_output.update(sub_op_dict[grad_id].output_arg_names())
332 333
            sub_op_output.update(sub_op_dict[fwd_id].input_arg_names())
            sub_op_output.update(sub_op_dict[grad_id].input_arg_names())
334
            ops_list.append((sub_block_ops, block_op_size, sub_op_output))
335

336
        # Process rest fwd_op block ops
Q
qijun 已提交
337
        for fwd_id in sub_block_ids:
338
            sub_block_ops = []
Q
qijun 已提交
339
            sub_block = pdesc.block(fwd_id)
340 341 342 343
            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
344
            sub_op_output.update(sub_op_dict[fwd_id].output_arg_names())
345
            sub_op_output.update(sub_op_dict[fwd_id].input_arg_names())
346 347
            ops_list.append((sub_block_ops, sub_block_op_size, sub_op_output))
    return ops_list
348

349

350
def _get_cfgs(input_program):
351 352 353 354 355
    """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.
    """
356 357 358 359
    ops_list = []
    pdesc = input_program.get_desc()
    block_desc = pdesc.block(0)
    op_size = block_desc.op_size()
360

361 362
    # Only process one level of nested subblock.
    ops_list.extend(_process_sub_block_pair(pdesc, SUB_BLOCK_PAIR))
363

364 365 366 367 368 369 370
    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))
371 372 373 374
    cfgs = [
        ControlFlowGraph(input_program, ops, forward_num, skip_opt)
        for ops, forward_num, skip_opt in ops_list
    ]
375
    return cfgs
376 377


378
def memory_optimize(input_program, skip_opt_set=None, print_log=False, level=0):
379 380 381 382 383 384 385 386 387 388 389
    """Optimize memory by reusing var memory.

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

    :param input_program: Input Program
    :param print_log: whether to print debug log.
    :param level: If level=0, reuse if the shape is completely equal, o
    :return:
    """
    if level != 0 and level != 1:
        raise ValueError("only support opt_level 0 or 1.")
Q
qiaolongfei 已提交
390 391
    global PRINT_LOG
    PRINT_LOG = print_log
392
    cfgs = _get_cfgs(input_program)
393
    for cfg in cfgs:
394
        cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level)
395 396


397
def release_memory(input_program, skip_opt_set=None):
Y
yuyang18 已提交
398 399 400 401 402 403 404 405 406 407
    """
    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`.
    """
408 409
    cfgs = _get_cfgs(input_program)
    for cfg in cfgs:
410
        cfg.release_memory(skip_opt_set=skip_opt_set)