compiler.py 11.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

import multiprocessing
import os
import six
X
polish  
Xin Pan 已提交
18
import sys
19
from .. import compat as cpt
X
Xin Pan 已提交
20
from . import framework
S
sneaxiy 已提交
21
from .framework import cuda_places, cpu_places
22 23 24

from . import core

X
Xin Pan 已提交
25 26
__all__ = ['CompiledProgram', 'ExecutionStrategy', 'BuildStrategy']

27 28
ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy
BuildStrategy = core.ParallelExecutor.BuildStrategy
F
flame 已提交
29 30
InferNativeConfig = core.NativeConfig
InferAnalysisConfig = core.AnalysisConfig
31 32 33 34 35 36 37 38


def _place_obj(place):
    p = core.Place()
    p.set_place(place)
    return p


39 40
def _is_pserver_mode(main_program):
    main = main_program if main_program \
C
chengduo 已提交
41
        else framework.default_main_program()
42 43 44 45 46 47
    for op in main.global_block().ops:
        if op.type in ["send", "recv"]:
            return True
    return False


X
polish  
Xin Pan 已提交
48
class CompiledProgram(object):
X
polish  
Xin Pan 已提交
49
    """
X
Xin Pan 已提交
50
    Compiles to Graph for execution.
X
polish  
Xin Pan 已提交
51

X
Xin Pan 已提交
52 53 54 55
    1. Users first create the program with layers.
    2. Optionally, users use CompiledProgram to optimize the program before run.
    3. The original program or CompiledProgram is run by executor.

X
polish  
Xin Pan 已提交
56 57 58 59 60 61 62 63
    The CompiledProgram is used to transform a program for various
    optimizations, for example.
      * Pre-compute some logic once so that each run is faster.
      * Transform the program so that it can run in multiple devices.
      * TODO: transform the program for optimized inference or distributed
              training.

    Example:
X
Xin Pan 已提交
64
        .. code-block:: python
X
Xin Pan 已提交
65
            place = fluid.CUDAPlace(0) if use_gpu else fluid.CPUPlace()
X
Xin Pan 已提交
66 67 68 69 70 71 72 73
            exe = fluid.Executor(place)
            exe.run(startup)
            compiled_prog = compiler.CompiledProgram(main).with_data_parallel(
                loss_name=loss.name)
            for i in range(5):
                test_loss, = exe.run(compiled_prog,
                                     feed=feed_dict,
                                     fetch_list=[loss.name])
X
polish  
Xin Pan 已提交
74 75

    Args:
X
Xin Pan 已提交
76 77 78 79 80
        program_or_graph (Graph|Program): If it's Program, it will be first
            lowered to a graph for further optimizations. If it's a graph
            (potentially optimized before), it will be directly used for
            further optimizations. Note: graph is only supported when compiled
            with with_data_parallel option.
X
polish  
Xin Pan 已提交
81 82
    """

X
Xin Pan 已提交
83 84 85 86 87 88 89 90 91 92 93 94
    def __init__(self, program_or_graph):
        if isinstance(program_or_graph, core.Graph):
            self._graph = program_or_graph
            self._program = None
        elif isinstance(program_or_graph, framework.Program):
            self._graph = core.Graph(program_or_graph.desc)
            self._program = program_or_graph
        else:
            raise ValueError("Wrong program_to_graph type: %s" %
                             type(program_or_graph))

        self._program_desc = self._graph.origin_program_desc()
X
polish  
Xin Pan 已提交
95 96 97
        self._scope = None
        self._place = None
        self._executor = None
98 99
        self._compiled = False
        self._is_data_parallel = False
F
flame 已提交
100
        self._is_inference = False
101

X
Xin Pan 已提交
102 103 104 105
    def with_data_parallel(self,
                           loss_name=None,
                           build_strategy=None,
                           exec_strategy=None,
S
sneaxiy 已提交
106 107
                           share_vars_from=None,
                           places=None):
X
Xin Pan 已提交
108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
        """Configs the program to run in data parallel way.

        Args:
            loss_name (str): The loss name must set in training. Default None.
            build_strategy(BuildStrategy): build_strategy is used to
                build the graph so it can run on multiple devices/cores with
                optimized topology.
                For more information, please refer to fluid.BuildStrategy.
                Default None.
            exec_strategy(ExecutionStrategy): exec_strategy is used to
                to select the a way to execute the graph, for example how many
                threads are used, how many iterations to clean up the temp
                variables. For more information, please refer
                to fluid.ExecutionStrategy. Default None.
            share_vars_from(CompiledProgram): If provide, this CompiledProgram
                will share variables from `share_vars_from`. `share_vars_from`
                must be run by the executor before this CompiledProgram so that
                vars are ready.
S
sneaxiy 已提交
126 127 128 129 130 131
            places(list(CUDAPlace)|list(CPUPlace)|None): If provide, only compile
                program in the given places. Otherwise, the places used when compiled 
                is determined by the Executor, and the places used are controlled 
                by environment variables: FLAGS_selected_gpus or CUDA_VISIBLE_DEVICES
                if using GPU; or CPU_NUM if using CPU.  

X
Xin Pan 已提交
132 133 134
        Returns:
            self
        """
135
        assert not self._is_data_parallel, "Already compiled with parallel."
X
Xin Pan 已提交
136
        assert not self._is_inference, "Cannot compile both data parallel and inference"
137 138 139 140
        self._is_data_parallel = True
        self._build_strategy = build_strategy
        self._exec_strategy = exec_strategy
        self._loss_name = loss_name
X
polish  
Xin Pan 已提交
141
        self._share_vars_from = share_vars_from
X
fix  
Xin Pan 已提交
142 143 144 145
        if self._exec_strategy is None:
            self._exec_strategy = ExecutionStrategy()
        if self._build_strategy is None:
            self._build_strategy = BuildStrategy()
S
sneaxiy 已提交
146 147 148 149 150 151
        if places is not None:
            if not isinstance(places, (list, tuple)):
                places = [places]
            self._places = [_place_obj(p) for p in places]
        else:
            self._places = None
Q
Qiao Longfei 已提交
152 153
        self._build_strategy.is_distribution = framework.is_pserver_mode(
            self._program)
154 155
        return self

F
flame 已提交
156 157 158 159 160 161 162 163
    def with_inference_optimize(self, config):
        """ Add inference optimize

        Args:
            config: instance of `NativeConfig` or `AnalysisConfig` to create predictor
        Returns:
            self
        """
X
Xin Pan 已提交
164
        assert not self._is_data_parallel, "Cannot compile both data parallel and inference"
X
Xin Pan 已提交
165 166
        assert not self._is_inference, "Already compiled with inference"

F
flame 已提交
167 168 169 170 171 172 173
        assert any([
            isinstance(config, InferNativeConfig),
            isinstance(config, InferAnalysisConfig)
        ])
        self._is_inference = True
        self._infer_config = config
        return self
X
polish  
Xin Pan 已提交
174

F
flame 已提交
175
    def _with_distributed(self):
X
polish  
Xin Pan 已提交
176 177
        raise NotImplementedError()

178
    def _compile_data_parallel(self, use_cuda=False, scope=None):
X
polish  
Xin Pan 已提交
179
        if self._share_vars_from:
180
            if scope:
X
polish  
Xin Pan 已提交
181 182 183 184 185 186 187 188 189 190
                sys.stderr.write("share_vars_from is set, scope is ignored.\n")
            if not self._share_vars_from._is_data_parallel:
                raise ValueError("share_vars_from is not data parallel. Cannot "
                                 "share vars from it.")
            if self._share_vars_from._executor is None:
                raise ValueError(
                    "share_vars_from is not compiled and run, so there is no "
                    "var to share.")
            self._local_scopes = self._share_vars_from._executor.local_scopes()
        else:
191
            assert scope is not None, ""
X
polish  
Xin Pan 已提交
192
            self._local_scopes = []
193 194

        self._exec_strategy.use_cuda = isinstance(self._place, core.CUDAPlace)
S
sneaxiy 已提交
195 196 197 198 199 200
        has_set_place = (self._places is not None)
        if has_set_place:
            desire_place = _place_obj(self._place)
            for p in self._places:
                assert p._type() == desire_place._type(), \
                    "Place type not match. You may set the wrong type of places"
201
        else:
S
sneaxiy 已提交
202 203 204
            places = cuda_places(
            ) if self._exec_strategy.use_cuda else cpu_places()
            self._places = [_place_obj(p) for p in places]
205 206 207 208 209 210 211 212
        assert self._places, "no place for execution"

        if self._exec_strategy.num_threads == 0:
            if self._exec_strategy.use_cuda:
                # Experiments on se-resnext shows that too many threads hurt
                # performance. Worth tunning for other models in the future.
                self._exec_strategy.num_threads = len(self._places) * 4
            else:
S
sneaxiy 已提交
213
                self._exec_strategy.num_threads = len(self._places) * 2
214

D
dzhwinter 已提交
215 216
        # FIXME(dzhwinter): enable_inplace should be after memory_optimize
        # if turn on python memory optimize, turn off the inplace_pass.
217 218 219 220 221 222
        # memory_optimize and enable_inplace default are True, but we can disable them on purpose
        if self._program and self._program._is_mem_optimized:
            self._build_strategy.memory_optimize = False

        if self._program and self._program._is_mem_optimized:
            self._build_strategy.enable_inplace = False
X
Xin Pan 已提交
223 224 225 226 227 228

        # TODO(wuyi): trainer endpoings should be passed in through
        # build_strategy, not program.xxx.
        if self._program and self._build_strategy.num_trainers > 1 and \
                self._program._trainers_endpoints:
            tps = self._program._trainers_endpoints
D
dzhwinter 已提交
229

230
            assert self._build_strategy.num_trainers == len(
X
Xin Pan 已提交
231 232 233 234
                tps), "num_trainers == len(end_points)"
            self._build_strategy.trainers_endpoints = tps

        self._persistable_vars = []
Z
Zhen Wang 已提交
235 236 237 238
        for node in self._graph.nodes():
            if node.is_var() and node.var() is not None and node.var().persistable() and \
                    node.var().type() != core.VarDesc.VarType.RAW:
                self._persistable_vars.append(cpt.to_text(node.name()))
239 240

        places = list(map(_place_obj, self._places))
Y
Yan Xu 已提交
241 242 243 244 245 246 247 248 249 250 251
        # ParallelExecutor would broadcast all the parameters during initializing.
        # The parameters of each process should be in the same ordered for the data-parallelism
        # distributed training to keep the broadcast correct.
        self._persistable_vars = list(set(self._persistable_vars))
        self._persistable_vars.sort()

        return core.ParallelExecutor(
            places, self._persistable_vars,
            cpt.to_text(self._loss_name)
            if self._loss_name else six.u(''), self._scope, self._local_scopes,
            self._exec_strategy, self._build_strategy, self._graph)
252

F
flame 已提交
253 254 255
    def _compile_inference(self):
        return core.create_paddle_predictor(self._infer_config)

256
    def _compile(self, scope, place):
X
Xin Pan 已提交
257 258 259 260 261 262 263 264 265 266
        """Compile the program based on the configs.

        Args:
            scope: The variables (resources) that are associated with
               this compiled program.
            place: The location that the compiled program will be run on.

        Returns:
            self
        """
267
        if self._compiled:
X
polish  
Xin Pan 已提交
268 269
            if scope and self._scope != scope:
                raise ValueError("Cannot compile with different scope")
S
sneaxiy 已提交
270
            if place and not self._place._equals(place):
X
polish  
Xin Pan 已提交
271
                raise ValueError("Cannot compile with different place")
272
            return self
X
fix  
Xin Pan 已提交
273
        self._compiled = True
274 275 276 277

        self._scope = scope
        self._place = place
        if self._is_data_parallel:
278 279 280
            self._executor = self._compile_data_parallel(
                use_cuda=isinstance(self._place, core.CUDAPlace),
                scope=self._scope)
F
flame 已提交
281 282
        elif self._is_inference:
            self._executor = self._compile_inference()
283 284 285 286
        else:
            p = _place_obj(self._place)
            self._executor = core.Executor(p)
        return self