profiler.py 10.3 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 . import core
D
dangqingqing 已提交
16
from contextlib import contextmanager
17
import os
D
dangqingqing 已提交
18

X
Xin Pan 已提交
19 20 21 22
__all__ = [
    'cuda_profiler', 'reset_profiler', 'profiler', 'start_profiler',
    'stop_profiler'
]
D
dangqingqing 已提交
23

D
dangqingqing 已提交
24
NVPROF_CONFIG = [
25 26 27 28 29 30
    "gpustarttimestamp",
    "gpuendtimestamp",
    "gridsize3d",
    "threadblocksize",
    "streamid",
    "enableonstart 0",
D
dangqingqing 已提交
31
    "conckerneltrace",
32 33 34
]


D
dangqingqing 已提交
35 36 37 38 39 40 41 42 43 44
@contextmanager
def cuda_profiler(output_file, output_mode=None, config=None):
    """The CUDA profiler.
    This fuctions is used to profile CUDA program by CUDA runtime application
    programming interface. The profiling result will be written into
    `output_file` with Key-Value pair format or Comma separated values format.
    The user can set the output mode by `output_mode` argument and set the
    counters/options for profiling by `config` argument. The default config
    is ['gpustarttimestamp', 'gpustarttimestamp', 'gridsize3d',
    'threadblocksize', 'streamid', 'enableonstart 0', 'conckerneltrace'].
D
Dang Qingqing 已提交
45 46 47
    Then users can use NVIDIA Visual Profiler
    (https://developer.nvidia.com/nvidia-visual-profiler) tools to load this
    this output file to visualize results.
D
dangqingqing 已提交
48 49 50 51 52 53

    Args:
        output_file (string) : The output file name, the result will be
            written into this file.
        output_mode (string) : The output mode has Key-Value pair format and
            Comma separated values format. It should be 'kvp' or 'csv'.
54 55
        config (list of string) : The profiler options and counters can refer
            to "Compute Command Line Profiler User Guide".
D
Dang Qingqing 已提交
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

    Raises:
        ValueError: If `output_mode` is not in ['kvp', 'csv'].

    Examples:

        .. code-block:: python

            import paddle.fluid as fluid
            import paddle.fluid.profiler as profiler

            epoc = 8
            dshape = [4, 3, 28, 28]
            data = fluid.layers.data(name='data', shape=[3, 28, 28], dtype='float32')
            conv = fluid.layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1])

            place = fluid.CUDAPlace(0)
            exe = fluid.Executor(place)
            exe.run(fluid.default_startup_program())

            output_file = 'cuda_profiler.txt'
            with profiler.cuda_profiler(output_file, 'csv') as nvprof:
                for i in range(epoc):
                    input = np.random.random(dshape).astype('float32')
                    exe.run(fluid.default_main_program(), feed={'data': input})
            # then use  NVIDIA Visual Profiler (nvvp) to load this output file
            # to visualize results.
D
dangqingqing 已提交
83 84 85
    """
    if output_mode is None:
        output_mode = 'csv'
D
dangqingqing 已提交
86 87 88
    if output_mode not in ['kvp', 'csv']:
        raise ValueError("The output mode must be 'kvp' or 'csv'.")
    config = NVPROF_CONFIG if config is None else config
89 90 91 92
    config_file = 'nvprof_config_file'
    with open(config_file, 'wb') as fp:
        fp.writelines(["%s\n" % item for item in config])
    core.nvprof_init(output_file, output_mode, config_file)
D
dangqingqing 已提交
93
    # Enables profiler collection by the active CUDA profiling tool.
D
dangqingqing 已提交
94
    core.nvprof_start()
D
dangqingqing 已提交
95 96
    yield
    # Disables profiler collection.
D
dangqingqing 已提交
97
    core.nvprof_stop()
98
    os.remove(config_file)
99 100 101


def reset_profiler():
D
Dang Qingqing 已提交
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117
    """
    Clear the previous time record. This interface does not work for
    `fluid.profiler.cuda_profiler`, it only works for
    `fluid.profiler.start_profiler`, `fluid.profiler.stop_profiler`,
    and `fluid.profiler.profiler`.

    Examples:

        .. code-block:: python

            import paddle.fluid.profiler as profiler
            with profiler.profiler(state, 'total', '/tmp/profile'):
                for iter in range(10):
                    if iter == 2:
                        profiler.reset_profiler()
                    # ...
118
    """
119 120 121
    core.reset_profiler()


X
Xin Pan 已提交
122
def start_profiler(state):
D
Dang Qingqing 已提交
123 124 125 126
    """
    Enable the profiler. Uers can use `fluid.profiler.start_profiler` and
    `fluid.profiler.stop_profiler` to insert the code, except the usage of
    `fluid.profiler.profiler` interface.
X
Xin Pan 已提交
127 128 129 130 131

    Args:
        state (string) : The profiling state, which should be 'CPU', 'GPU'
            or 'All'. 'CPU' means only profile CPU. 'GPU' means profiling
            GPU as well. 'All' also generates timeline.
D
Dang Qingqing 已提交
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147

    Raises:
        ValueError: If `state` is not in ['CPU', 'GPU', 'All'].

    Examples:

        .. code-block:: python

            import paddle.fluid.profiler as profiler

            profiler.start_profiler('GPU')
            for iter in range(10):
                if iter == 2:
                    profiler.reset_profiler()
                # except each iteration
            profiler.stop_profiler('total', '/tmp/profile')
X
Xin Pan 已提交
148 149 150
    """
    if core.is_profiler_enabled():
        return
X
Xin Pan 已提交
151 152 153 154 155 156 157 158 159 160 161 162
    if state not in ['CPU', 'GPU', "All"]:
        raise ValueError("The state must be 'CPU' or 'GPU' or 'All'.")
    if state == "GPU":
        prof_state = core.ProfilerState.kCUDA
    elif state == "CPU":
        prof_state = core.ProfilerState.kCPU
    else:
        prof_state = core.ProfilerState.kAll
    core.enable_profiler(prof_state)


def stop_profiler(sorted_key=None, profile_path='/tmp/profile'):
D
Dang Qingqing 已提交
163 164 165 166
    """
    Stop the profiler. Uers can use `fluid.profiler.start_profiler` and
    `fluid.profiler.stop_profiler` to insert the code, except the usage of
    `fluid.profiler.profiler` interface.
X
Xin Pan 已提交
167 168 169 170 171 172 173 174 175 176 177 178 179

    Args:
        sorted_key (string) : If None, the profiling results will be printed
            in the order of first end time of events. Otherwise, the profiling
            results will be sorted by the this flag. This flag should be one
            of 'calls', 'total', 'max', 'min' or 'ave'.
            The `calls` means sorting by the number of calls.
            The `total` means sorting by the total execution time.
            The `max` means sorting by the maximum execution time.
            The `min` means sorting by the minimum execution time.
            The `ave` means sorting by the average execution time.
        profile_path (string) : If state == 'All', it will write a profile
            proto output file.
D
Dang Qingqing 已提交
180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196

    Raises:
        ValueError: If `sorted_key` is not in
            ['calls', 'total', 'max', 'min', 'ave'].

    Examples:

        .. code-block:: python

            import paddle.fluid.profiler as profiler

            profiler.start_profiler('GPU')
            for iter in range(10):
                if iter == 2:
                    profiler.reset_profiler()
                # except each iteration
            profiler.stop_profiler('total', '/tmp/profile')
X
Xin Pan 已提交
197 198 199
    """
    if not core.is_profiler_enabled():
        return
X
Xin Pan 已提交
200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216
    sorted_key = 'default' if sorted_key is None else sorted_key
    if sorted_key not in ['default', 'calls', 'total', 'max', 'min', 'ave']:
        raise ValueError("The sorted_key must be None or in 'calls', 'total', "
                         "'max', 'min' and 'ave'")
    key_map = {
        'default': core.EventSortingKey.kDefault,
        'calls': core.EventSortingKey.kCalls,
        'total': core.EventSortingKey.kTotal,
        'max': core.EventSortingKey.kMax,
        'min': core.EventSortingKey.kMin,
        'ave': core.EventSortingKey.kAve,
    }
    # TODO(qingqing) : redirect C++ ostream to Python stream.
    # with core.ostream_redirect(stdout=True, stderr=True):
    core.disable_profiler(key_map[sorted_key], profile_path)


217
@contextmanager
X
Xin Pan 已提交
218
def profiler(state, sorted_key=None, profile_path='/tmp/profile'):
219
    """The profiler interface.
220 221 222
    Different from cuda_profiler, this profiler can be used to profile both CPU
    and GPU program. By defalut, it records the CPU and GPU operator kernels,
    if you want to profile other program, you can refer the profiling tutorial
D
Dang Qingqing 已提交
223 224 225 226
    to add more records in C++ code.

    If the state == 'All', a profile proto file will be written to
    `profile_path`. This file records timeline information during the execution.
227
    Then users can visualize this file to see the timeline, please refer
D
Dang Qingqing 已提交
228
    https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/timeline.md
229 230

    Args:
D
dangqingqing 已提交
231 232 233 234 235
        state (string) : The profiling state, which should be 'CPU' or 'GPU',
            telling the profiler to use CPU timer or GPU timer for profiling.
            Although users may have already specified the execution place
            (CPUPlace/CUDAPlace) in the begining, for flexibility the profiler
            would not inherit this place.
236 237 238 239 240
        sorted_key (string) : If None, the profiling results will be printed
            in the order of first end time of events. Otherwise, the profiling
            results will be sorted by the this flag. This flag should be one
            of 'calls', 'total', 'max', 'min' or 'ave'.
            The `calls` means sorting by the number of calls.
241 242 243 244
            The `total` means sorting by the total execution time.
            The `max` means sorting by the maximum execution time.
            The `min` means sorting by the minimum execution time.
            The `ave` means sorting by the average execution time.
X
Xin Pan 已提交
245 246
        profile_path (string) : If state == 'All', it will write a profile
            proto output file.
D
Dang Qingqing 已提交
247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265

    Raises:
        ValueError: If `state` is not in ['CPU', 'GPU', 'All']. If `sorted_key` is
            not in ['calls', 'total', 'max', 'min', 'ave'].

    Examples:

        .. code-block:: python

            import paddle.fluid.profiler as profiler

            with profiler.profiler('All', 'total', '/tmp/profile') as prof:
                for pass_id in range(pass_num):
                    for batch_id, data in enumerate(train_reader()):
                        exe.run(fluid.default_main_program(),
                                feed=feeder.feed(data),
                                fetch_list=[],
                                use_program_cache=True)
                        # ...
266
    """
X
Xin Pan 已提交
267
    start_profiler(state)
268
    yield
X
Xin Pan 已提交
269
    stop_profiler(sorted_key, profile_path)