pybind.cc 32.3 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6

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

7
http://www.apache.org/licenses/LICENSE-2.0
8 9 10 11 12 13

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. */
L
lgone2000 已提交
14
#include <Python.h>
C
chengduoZH 已提交
15 16
#include <algorithm>
#include <map>
S
sneaxiy 已提交
17
#include <memory>
C
chengduoZH 已提交
18 19 20 21 22
#include <mutex>  // NOLINT // for call_once
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
23

Y
Yi Wang 已提交
24 25 26
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/framework.pb.h"
27
#include "paddle/fluid/framework/ir/pass_builder.h"
Y
Yi Wang 已提交
28 29 30
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
31
#include "paddle/fluid/framework/op_registry.h"
Y
Yu Yang 已提交
32
#include "paddle/fluid/framework/parallel_executor.h"
Y
Yi Wang 已提交
33
#include "paddle/fluid/framework/prune.h"
Y
Refine  
Yu Yang 已提交
34
#include "paddle/fluid/framework/reader.h"
Y
Yi Wang 已提交
35
#include "paddle/fluid/framework/selected_rows.h"
X
Xin Pan 已提交
36
#include "paddle/fluid/framework/version.h"
D
dzhwinter 已提交
37
#include "paddle/fluid/operators/activation_op.h"
S
sneaxiy 已提交
38
#include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
Y
Yi Wang 已提交
39
#include "paddle/fluid/platform/enforce.h"
40
#include "paddle/fluid/platform/init.h"
Y
Yi Wang 已提交
41 42 43 44
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
#include "paddle/fluid/pybind/const_value.h"
#include "paddle/fluid/pybind/exception.h"
45 46
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h"  // NOLINT
Y
Yu Yang 已提交
47
#include "paddle/fluid/pybind/recordio.h"
Y
Yi Wang 已提交
48
#include "paddle/fluid/pybind/tensor_py.h"
Y
Yu Yang 已提交
49

50
#include "paddle/fluid/string/to_string.h"
51

D
Dong Zhihong 已提交
52
#ifdef PADDLE_WITH_CUDA
Y
Yi Wang 已提交
53 54 55
#include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
#include "paddle/fluid/platform/cuda_profiler.h"
#include "paddle/fluid/platform/gpu_info.h"
D
Dong Zhihong 已提交
56 57
#endif

M
minqiyang 已提交
58 59
#include "pybind11/stl.h"

Q
Qiao Longfei 已提交
60 61 62
// disable auto conversion to list in Python
PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);

63
namespace paddle {
64
namespace pybind {
65
bool IsCompiledWithCUDA() {
66
#ifndef PADDLE_WITH_CUDA
Q
qijun 已提交
67 68 69 70 71 72
  return false;
#else
  return true;
#endif
}

Y
update  
Yancey1989 已提交
73
bool IsCompiledWithDIST() {
Y
Yancey1989 已提交
74
#ifdef PADDLE_WITH_DISTRIBUTE
Y
update  
Yancey1989 已提交
75 76 77 78 79 80
  return true;
#else
  return false;
#endif
}

81 82
PYBIND11_PLUGIN(core) {
  py::module m("core", "C++ core of PaddlePaddle");
83

84 85 86 87
  // using framework in this function. Since it is inside a function, it will
  // not cause namespace pollution.
  using namespace paddle::framework;  // NOLINT

88
  BindException(&m);
Y
Yu Yang 已提交
89

90 91 92
  py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
      .def_buffer(
          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
Y
yuyang18 已提交
93
      .def("_get_dims",
94
           [](const Tensor &self) { return vectorize(self.dims()); })
Y
yuyang18 已提交
95
      .def("_set_dims",
Q
qijun 已提交
96
           [](Tensor &self, const std::vector<int64_t> &dim) {
Y
Yu Yang 已提交
97
             self.Resize(make_ddim(dim));
Y
Yu Yang 已提交
98
           })
Y
yuyang18 已提交
99
      .def("_set_layout",
D
dzhwinter 已提交
100 101 102
           [](Tensor &self, const std::string &layout) {
             self.set_layout(StringToDataLayout(layout));
           })
Y
yuyang18 已提交
103
      .def("_alloc_float",
D
dzhwinter 已提交
104
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
105
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
106
           })
Y
yuyang18 已提交
107
      .def("_alloc_float",
Y
Yu Yang 已提交
108
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
109
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
110
           })
Y
yuyang18 已提交
111
      .def("_alloc_int",
Y
Yu Yang 已提交
112
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
113
             self.mutable_data<int>(place);
Y
Yu Yang 已提交
114
           })
Y
yuyang18 已提交
115
      .def("_alloc_int",
D
dzhwinter 已提交
116
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
117
             self.mutable_data<int>(place);
Q
qijun 已提交
118
           })
Y
yuyang18 已提交
119
      .def("_alloc_int",
C
chengduoZH 已提交
120 121 122
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<int>(place);
           })
Y
yuyang18 已提交
123
      .def("_alloc_float",
C
chengduoZH 已提交
124 125 126
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<float>(place);
           })
Y
Yu Yang 已提交
127 128
      .def("set", PyCPUTensorSetFromArray<float>)
      .def("set", PyCPUTensorSetFromArray<int>)
129
      .def("set", PyCPUTensorSetFromArray<double>)
130
      .def("set", PyCPUTensorSetFromArray<int64_t>)
Y
Yu Yang 已提交
131
      .def("set", PyCPUTensorSetFromArray<bool>)
132
      .def("set", PyCPUTensorSetFromArray<uint16_t>)
F
fengjiayi 已提交
133
      .def("set", PyCPUTensorSetFromArray<uint8_t>)
Q
qingqing01 已提交
134
      .def("set", PyCPUTensorSetFromArray<int8_t>)
135
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
136 137
      .def("set", PyCUDATensorSetFromArray<float>)
      .def("set", PyCUDATensorSetFromArray<int>)
138
      .def("set", PyCUDATensorSetFromArray<double>)
139
      .def("set", PyCUDATensorSetFromArray<int64_t>)
Y
Yu Yang 已提交
140
      .def("set", PyCUDATensorSetFromArray<bool>)
141
      .def("set", PyCUDATensorSetFromArray<uint16_t>)
F
fengjiayi 已提交
142
      .def("set", PyCUDATensorSetFromArray<uint8_t>)
Q
qingqing01 已提交
143
      .def("set", PyCUDATensorSetFromArray<int8_t>)
C
chengduoZH 已提交
144 145 146 147 148 149
      .def("set", PyCUDAPinnedTensorSetFromArray<float>)
      .def("set", PyCUDAPinnedTensorSetFromArray<int>)
      .def("set", PyCUDAPinnedTensorSetFromArray<double>)
      .def("set", PyCUDAPinnedTensorSetFromArray<int64_t>)
      .def("set", PyCUDAPinnedTensorSetFromArray<bool>)
      .def("set", PyCUDAPinnedTensorSetFromArray<uint16_t>)
F
fengjiayi 已提交
150
      .def("set", PyCUDAPinnedTensorSetFromArray<uint8_t>)
Q
qingqing01 已提交
151
      .def("set", PyCUDAPinnedTensorSetFromArray<int8_t>)
Q
qijun 已提交
152
#endif
153
      .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
Y
yuyang18 已提交
154 155 156 157 158
      .def("_set_float_element", TensorSetElement<float>)
      .def("_get_float_element", TensorGetElement<float>)
      .def("_set_double_element", TensorSetElement<double>)
      .def("_get_double_element", TensorGetElement<double>)
      .def("_dtype", [](Tensor &self) { return ToDataType(self.type()); });
Y
Yu Yang 已提交
159

160
  py::class_<LoDTensor, Tensor>(m, "LoDTensor")
161 162
      .def_buffer(
          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
163 164 165 166 167 168 169 170 171 172 173 174 175 176
      .def("__init__",
           [](LoDTensor &instance, const std::vector<std::vector<size_t>>
                                       &recursive_sequence_lengths) {
             LoD new_lod;
             new_lod.reserve(recursive_sequence_lengths.size());
             std::copy(recursive_sequence_lengths.begin(),
                       recursive_sequence_lengths.end(),
                       std::back_inserter(new_lod));
             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
             PADDLE_ENFORCE(
                 CheckLoD(new_offset_lod, -1),
                 "the provided recursive_sequence_lengths info is invalid");
             new (&instance) LoDTensor(new_offset_lod);
           })
Y
Yu Yang 已提交
177
      .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
G
gongweibao 已提交
178 179 180 181 182
      // We implement offset based LOD in C++ while we use length based with
      // Python API. So we changed set_lod to set_recursive_sequence_lengths to
      // avoid misuse.
      // The discussion is here:
      // https://github.com/PaddlePaddle/Paddle/issues/10855
D
dangqingqing 已提交
183
      .def("set_lod",
184
           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
185
             // the input lod is offset-based level-of-detail info
Y
Yu Yang 已提交
186
             LoD new_lod;
187 188
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
189 190
             PADDLE_ENFORCE(CheckLoD(new_lod, vectorize(self.dims()).front()),
                            "the provided lod info is invalid");
191
             self.set_lod(new_lod);
D
dangqingqing 已提交
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
      .def("set_recursive_sequence_lengths",
           [](LoDTensor &self, const std::vector<std::vector<size_t>>
                                   &recursive_sequence_lengths) {
             // the input recursive_sequence_lengths is length-based
             // level-of-detail info
             LoD new_lod;
             new_lod.reserve(recursive_sequence_lengths.size());
             std::copy(recursive_sequence_lengths.begin(),
                       recursive_sequence_lengths.end(),
                       std::back_inserter(new_lod));
             LoD new_offset_lod = ConvertToOffsetBasedLoD(new_lod);
             PADDLE_ENFORCE(
                 CheckLoD(new_offset_lod, vectorize(self.dims()).front()),
                 "the provided recursive_sequence_lengths info is invalid");
             self.set_lod(new_offset_lod);
           })
      .def("lod",
           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
             // output the offset-based lod info
             LoD lod = self.lod();
             std::vector<std::vector<size_t>> new_lod;
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
             return new_lod;
           })
G
gongweibao 已提交
218
      // Set above comments of set_lod.
219 220 221 222 223 224 225 226 227 228 229 230 231
      .def("recursive_sequence_lengths",
           [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
             // output the length-based lod info
             LoD lod = ConvertToLengthBasedLoD(self.lod());
             std::vector<std::vector<size_t>> new_lod;
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
             return new_lod;
           })
      .def("has_valid_recursive_sequence_lengths", [](LoDTensor &self) -> bool {
        // Check that the lod info is valid and match the outermost
        // dimension of the LoDTensor data
        return CheckLoD(self.lod(), vectorize(self.dims()).front());
D
dangqingqing 已提交
232 233
      });

Q
qijun 已提交
234 235 236 237 238 239 240 241 242 243 244 245 246
  py::class_<SelectedRows>(m, "SelectedRows")
      .def("__init__",
           [](SelectedRows &instance) { new (&instance) SelectedRows(); })
      .def("__init__",
           [](SelectedRows &instance, const std::vector<int64_t> rows,
              const int64_t &height) {
             new (&instance) SelectedRows(rows, height);
           })
      .def("get_tensor",
           [](SelectedRows &self) { return self.mutable_value(); },
           py::return_value_policy::reference)
      .def("set_height", &SelectedRows::set_height)
      .def("height", &SelectedRows::height)
Q
qijun 已提交
247 248 249 250 251 252 253 254 255
      .def("set_rows",
           [](SelectedRows &self, std::vector<int64_t> rows) {
#ifndef PADDLE_WITH_CUDA
             self.set_rows(rows);
#else
        Vector<int64_t> new_rows(rows);
        self.set_rows(new_rows);
#endif
           })
256
      .def("sync_index", [](SelectedRows &instance) { instance.SyncIndex(); })
257
      .def("rows", [](SelectedRows &self) {
258 259 260 261 262
        auto rows = self.rows();
        std::vector<int64_t> new_rows;
        new_rows.reserve(rows.size());
        std::copy(rows.begin(), rows.end(), std::back_inserter(new_rows));
        return new_rows;
263
      });
Q
qijun 已提交
264

265
  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
266 267 268

All parameter, weight, gradient are variables in Paddle.
)DOC")
269
      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
270
      .def("set_int",
271 272
           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
273 274 275 276 277 278 279
      .def("is_float", [](const Variable &var) { return var.IsType<float>(); })
      .def("set_float",
           [](Variable &var, float val) -> void {
             *var.GetMutable<float>() = val;
           })
      .def("get_float",
           [](const Variable &var) -> float { return var.Get<float>(); })
Y
Yu Yang 已提交
280
      .def("get_tensor",
281 282
           [](Variable &self) -> LoDTensor * {
             return self.GetMutable<LoDTensor>();
D
dangqingqing 已提交
283 284
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
285 286 287
      .def("get_lod_rank_table",
           [](Variable &self) { return self.GetMutable<LoDRankTable>(); },
           py::return_value_policy::reference)
Q
qijun 已提交
288 289 290 291 292
      .def("get_selected_rows",
           [](Variable &self) -> SelectedRows * {
             return self.GetMutable<SelectedRows>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
293 294 295
      .def("get_lod_tensor_array",
           [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
           py::return_value_policy::reference)
D
Dong Zhihong 已提交
296 297 298 299 300 301 302
#ifdef PADDLE_WITH_CUDA
      .def("get_communicator",
           [](Variable &self) -> platform::Communicator * {
             return self.GetMutable<platform::Communicator>();
           },
           py::return_value_policy::reference)
#endif
Y
Refine  
Yu Yang 已提交
303 304 305 306 307
      .def("get_reader",
           [](Variable &self) -> framework::ReaderHolder * {
             PADDLE_ENFORCE(self.IsType<framework::ReaderHolder>());
             return self.GetMutable<framework::ReaderHolder>();
           },
Y
Yu Yang 已提交
308
           py::return_value_policy::reference);
309

Y
Refine  
Yu Yang 已提交
310
  py::class_<framework::ReaderHolder>(m, "Reader", "")
311
      .def("reset", &framework::ReaderHolder::ResetAll);
Y
Refine  
Yu Yang 已提交
312

S
sneaxiy 已提交
313 314 315 316
  using LoDTensorBlockingQueue =
      ::paddle::operators::reader::LoDTensorBlockingQueue;
  using LoDTensorBlockingQueueHolder =
      ::paddle::operators::reader::LoDTensorBlockingQueueHolder;
S
sneaxiy 已提交
317 318
  py::class_<LoDTensorBlockingQueue, std::shared_ptr<LoDTensorBlockingQueue>>(
      m, "LoDTensorBlockingQueue", "")
S
sneaxiy 已提交
319
      .def("push",
S
sneaxiy 已提交
320
           [](LoDTensorBlockingQueue &self,
S
sneaxiy 已提交
321
              const std::vector<framework::LoDTensor> &lod_tensor_vec) {
S
sneaxiy 已提交
322
             pybind11::gil_scoped_release release;
S
sneaxiy 已提交
323
             return self.Push(lod_tensor_vec);
S
sneaxiy 已提交
324
           })
S
sneaxiy 已提交
325 326 327 328
      .def("size", &LoDTensorBlockingQueue::Size)
      .def("capacity", &LoDTensorBlockingQueue::Cap)
      .def("close", &LoDTensorBlockingQueue::Close)
      .def("is_closed", &LoDTensorBlockingQueue::IsClosed);
S
sneaxiy 已提交
329

S
sneaxiy 已提交
330
  m.def("init_lod_tensor_blocking_queue",
S
sneaxiy 已提交
331
        [](Variable &var, size_t capacity,
S
sneaxiy 已提交
332
           const std::vector<std::vector<int64_t>> &shapes)
S
sneaxiy 已提交
333
            -> std::shared_ptr<LoDTensorBlockingQueue> {
S
sneaxiy 已提交
334 335 336 337 338 339 340
              std::vector<DDim> dims(shapes.size());
              std::transform(shapes.begin(), shapes.end(), dims.begin(),
                             [](const std::vector<int64_t> &shape) {
                               return make_ddim(shape);
                             });
              auto *holder = var.GetMutable<LoDTensorBlockingQueueHolder>();
              holder->InitOnce(capacity, dims);
S
sneaxiy 已提交
341
              return holder->GetQueue();
S
sneaxiy 已提交
342
            },
S
sneaxiy 已提交
343
        py::return_value_policy::copy);
S
sneaxiy 已提交
344

345
  py::class_<Scope>(m, "Scope", "")
D
dongzhihong 已提交
346
      .def("var",
347
           [](Scope &self, const std::string &name) -> Variable * {
D
dongzhihong 已提交
348
             return self.Var(name);
Y
Yu Yang 已提交
349
           },
350
           py::return_value_policy::reference)
351
      .def("find_var", &Scope::FindVar, py::return_value_policy::reference)
Y
Yu Yang 已提交
352
      .def(py::init<>())
353
      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
354
           py::return_value_policy::reference)
Y
Yu Yang 已提交
355
      .def("drop_kids", &Scope::DropKids);
356

Y
Yu Yang 已提交
357 358
  //! @note: Be careful! PyBind will return std::string as an unicode, not
  //! Python str. If you want a str object, you should cast them in Python.
Y
Yu Yang 已提交
359 360
  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
    std::vector<py::bytes> ret_values;
361 362 363 364 365 366 367 368 369 370
    for (auto &iter : OpInfoMap::Instance().map()) {
      auto &info = iter.second;
      if (info.HasOpProtoAndChecker()) {
        std::string str;
        PADDLE_ENFORCE(
            info.Proto().SerializeToString(&str),
            "Serialize OpProto Error. This could be a bug of Paddle.");
        ret_values.emplace_back(str);
      }
    }
Y
Yu Yang 已提交
371 372
    return ret_values;
  });
373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388
  m.def(
      "get_grad_op_desc", [](const OpDesc &op_desc,
                             const std::unordered_set<std::string> &no_grad_set,
                             const std::vector<BlockDesc *> &grad_sub_block) {
        std::unordered_map<std::string, std::string> grad_to_var;
        std::vector<std::unique_ptr<OpDesc>> grad_op_descs =
            framework::OpInfoMap::Instance()
                .Get(op_desc.Type())
                .GradOpMaker()(op_desc, no_grad_set, &grad_to_var,
                               grad_sub_block);
        std::vector<OpDesc *> grad_op_desc_ptrs(grad_op_descs.size());
        std::transform(grad_op_descs.begin(), grad_op_descs.end(),
                       grad_op_desc_ptrs.begin(),
                       [](std::unique_ptr<OpDesc> &p) { return p.release(); });
        return std::make_pair(grad_op_desc_ptrs, grad_to_var);
      });
Y
Yu Yang 已提交
389
  m.def("prune", [](const ProgramDesc &origin,
390
                    const std::vector<std::array<size_t, 2>> &targets) {
Y
Yu Yang 已提交
391
    ProgramDesc prog_with_targets(origin);
392
    for (const auto &t : targets) {
393
      prog_with_targets.MutableBlock(t[0])->Op(t[1])->SetIsTarget(true);
394
    }
395
    proto::ProgramDesc pruned_desc;
396
    Prune(*prog_with_targets.Proto(), &pruned_desc);
Y
Yu Yang 已提交
397
    return new ProgramDesc(pruned_desc);
398
  });
399 400 401 402
  m.def("empty_var_name",
        []() { return std::string(framework::kEmptyVarName); });
  m.def("grad_var_suffix",
        []() { return std::string(framework::kGradVarSuffix); });
403 404 405
  m.def_submodule(
       "var_names",
       "The module will return special predefined variable name in Paddle")
Y
Yi Wang 已提交
406 407
      .def("empty", []() { return kEmptyVarName; })
      .def("temp", []() { return kTempVarName; });
Q
qijun 已提交
408
  // clang-format off
Y
Yu Yang 已提交
409
  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
Q
qijun 已提交
410 411
      .def_static("create",
                  [](paddle::platform::CPUPlace& place)
Q
qijun 已提交
412
                      -> paddle::platform::DeviceContext* {
Q
qijun 已提交
413 414 415
                    return new paddle::platform::CPUDeviceContext();
                  })
      .def_static("create",
D
dzhwinter 已提交
416
                  [](paddle::platform::CUDAPlace& place)
Q
qijun 已提交
417
                      -> paddle::platform::DeviceContext* {
418
#ifndef PADDLE_WITH_CUDA
D
dzhwinter 已提交
419
                    PADDLE_THROW("CUDAPlace is not supported in CPU device.");
Q
qijun 已提交
420
#else
Q
qijun 已提交
421
                    return new paddle::platform::CUDADeviceContext(place);
Q
qijun 已提交
422
#endif
C
chengduoZH 已提交
423 424 425 426 427 428 429 430 431 432 433
                  })
          .def_static("create",
                [](paddle::platform::CUDAPinnedPlace& place)
                        -> paddle::platform::DeviceContext* {
#ifndef PADDLE_WITH_CUDA
                  PADDLE_THROW(
                        "CUDAPinnedPlace is not supported in CPU device.");
#else
                  return new paddle::platform::CUDAPinnedDeviceContext(place);
#endif
                });;
D
Dong Zhihong 已提交
434 435 436 437
// clang-format on
#ifdef PADDLE_WITH_CUDA
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
#endif
D
dzhwinter 已提交
438
  py::class_<platform::CUDAPlace>(m, "CUDAPlace")
439
      .def(py::init<int>())
D
dzhwinter 已提交
440
      .def("__str__", string::to_string<const platform::CUDAPlace &>);
Q
qijun 已提交
441

442 443 444
  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
      .def(py::init<>())
      .def("__str__", string::to_string<const platform::CPUPlace &>);
Y
Yu Yang 已提交
445

C
chengduoZH 已提交
446 447 448 449
  py::class_<paddle::platform::CUDAPinnedPlace>(m, "CUDAPinnedPlace")
      .def(py::init<>())
      .def("__str__", string::to_string<const platform::CUDAPinnedPlace &>);

Y
Yu Yang 已提交
450 451 452 453 454 455 456
  py::class_<platform::Place>(m, "Place")
      .def(py::init<>())
      .def("set_place",
           [](platform::Place &self, const platform::CPUPlace &cpu_place) {
             self = cpu_place;
           })
      .def("set_place",
D
dzhwinter 已提交
457
           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
Y
Yu Yang 已提交
458
             self = gpu_place;
C
chengduoZH 已提交
459 460
           })
      .def("set_place", [](platform::Place &self,
C
chengduoZH 已提交
461 462
                           const platform::CUDAPinnedPlace &cuda_pinned_place) {
        self = cuda_pinned_place;
C
chengduoZH 已提交
463
      });
Y
Yu Yang 已提交
464

Y
Yu Yang 已提交
465 466 467
  py::class_<OperatorBase>(m, "Operator")
      .def_static("create",
                  [](py::bytes protobin) {
468
                    proto::OpDesc desc;
Y
Yu Yang 已提交
469 470 471 472 473
                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
                                   "Cannot parse user input to OpDesc");
                    PADDLE_ENFORCE(desc.IsInitialized(),
                                   "User OpDesc is not initialized, reason %s",
                                   desc.InitializationErrorString());
474
                    return OpRegistry::CreateOp(desc);
Y
Yu Yang 已提交
475
                  })
476
      .def("run",
477
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
478 479 480
              const platform::CPUPlace &place) { self.Run(scope, place); })
      .def("run",
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
481
              const platform::CUDAPlace &place) { self.Run(scope, place); })
C
chengduoZH 已提交
482 483 484 485 486
      .def("run",
           [](OperatorBase &self, const Scope &scope,
              const platform::CUDAPinnedPlace &place) {
             self.Run(scope, place);
           })
Y
Yu Yang 已提交
487 488 489 490 491 492 493
      .def("type",
           [](const OperatorBase &op) -> std::string { return op.Type(); })
      .def("outputs",
           [](const OperatorBase &op)
               -> std::map<std::string, std::vector<std::string>> {
                 return op.Outputs();
               })
Q
qijun 已提交
494 495
      .def("output_vars",
           [](const OperatorBase &op) { return op.OutputVars(true); })
Y
Yu Yang 已提交
496
      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
Q
qijun 已提交
497
      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
Y
Yu Yang 已提交
498 499 500 501
      .def("__str__", &OperatorBase::DebugString)
      .def("no_intermediate_outputs",
           [](const OperatorBase &op) { return op.OutputVars(false); })
      .def("support_gpu", &OperatorBase::SupportGPU);
Y
Yu Yang 已提交
502

F
fengjiayi 已提交
503
  py::class_<framework::Executor>(m, "Executor")
D
dzhwinter 已提交
504
      .def(py::init<const platform::Place &>())
Y
Yancey1989 已提交
505
      .def("close", &Executor::Close)
S
sneaxiy 已提交
506 507 508 509 510
      .def("run", [](Executor &self, const ProgramDesc &prog, Scope *scope,
                     int block_id, bool create_local_scope, bool create_vars) {
        pybind11::gil_scoped_release release;
        self.Run(prog, scope, block_id, create_local_scope, create_vars);
      });
S
sneaxiy 已提交
511

D
dzhwinter 已提交
512
  m.def("init_gflags", framework::InitGflags);
Y
Yang Yu 已提交
513
  m.def("init_glog", framework::InitGLOG);
X
Xin Pan 已提交
514 515
  m.def("init_devices",
        [](bool init_p2p) { framework::InitDevices(init_p2p); });
516

517
  m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
Y
update  
Yancey1989 已提交
518
  m.def("is_compiled_with_dist", IsCompiledWithDIST);
519 520 521 522 523 524
#ifdef PADDLE_WITH_CUDA
  m.def("is_float16_supported", [](const platform::CUDAPlace &place) -> bool {
    // Only GPUs with Compute Capability >= 53 support float16
    return platform::GetCUDAComputeCapability(place.device) >= 53;
  });
#endif
525

526
  m.def("set_feed_variable", framework::SetFeedVariable);
Q
qijun 已提交
527
  m.def("get_fetch_variable", framework::GetFetchVariable);
Q
qijun 已提交
528

X
Xin Pan 已提交
529 530
  m.def("_is_program_version_supported", IsProgramVersionSupported);

531 532 533 534 535
  BindProgramDesc(&m);
  BindBlockDesc(&m);
  BindVarDsec(&m);
  BindOpDesc(&m);
  BindConstValue(&m);
Y
Yu Yang 已提交
536

Y
Yu Yang 已提交
537 538 539 540 541 542 543 544 545
  py::class_<framework::LoDRankTable>(m, "LodRankTable")
      .def("items", [](framework::LoDRankTable &table) {
        std::vector<std::pair<size_t, size_t>> res;
        for (auto &item : table.items()) {
          res.push_back({item.index, item.length});
        }
        return res;
      });

Y
Yu Yang 已提交
546
  py::class_<LoDTensorArray>(m, "LoDTensorArray")
S
sneaxiy 已提交
547 548
      .def("__init__",
           [](LoDTensorArray &instance) { new (&instance) LoDTensorArray(); })
Y
Yu Yang 已提交
549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564
      .def("__getitem__",
           [](LoDTensorArray &self, size_t i) { return &self.at(i); },
           py::return_value_policy::reference)
      .def("__len__", [](LoDTensorArray &self) { return self.size(); })
      .def("__setitem__",
           [](LoDTensorArray &self, size_t i, const LoDTensor &t) {
             PADDLE_ENFORCE_LT(i, self.size());
             self[i].ShareDataWith(t);
             self[i].set_lod(t.lod());
           })
      .def("append", [](LoDTensorArray &self, const LoDTensor &t) {
        self.emplace_back();
        self.back().ShareDataWith(t);
        self.back().set_lod(t.lod());
      });

D
dzhwinter 已提交
565 566 567
  m.def("IsInplace",
        [](std::string op) -> bool { return operators::IsInplace(op); });

Y
Yu Yang 已提交
568
  m.def("op_support_gpu", OpSupportGPU);
D
Dong Zhihong 已提交
569
#ifdef PADDLE_WITH_CUDA
D
Dong Zhihong 已提交
570
  m.def("get_cuda_device_count", platform::GetCUDADeviceCount);
D
dangqingqing 已提交
571 572 573 574

  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
D
Dong Zhihong 已提交
575
#endif
Y
Yu Yang 已提交
576

577 578 579 580
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
      .value("kCUDA", platform::ProfilerState::kCUDA)
581
      .value("kAll", platform::ProfilerState::kAll)
582 583 584 585 586 587 588 589 590 591 592 593 594
      .export_values();

  py::enum_<platform::EventSortingKey>(m, "EventSortingKey", py::arithmetic())
      .value("kDefault", platform::EventSortingKey::kDefault)
      .value("kCalls", platform::EventSortingKey::kCalls)
      .value("kTotal", platform::EventSortingKey::kTotal)
      .value("kMin", platform::EventSortingKey::kMin)
      .value("kMax", platform::EventSortingKey::kMax)
      .value("kAve", platform::EventSortingKey::kAve)
      .export_values();

  m.def("enable_profiler", platform::EnableProfiler);
  m.def("disable_profiler", platform::DisableProfiler);
X
Xin Pan 已提交
595
  m.def("is_profiler_enabled", platform::IsProfileEnabled);
596
  m.def("reset_profiler", platform::ResetProfiler);
Y
Yu Yang 已提交
597

598 599 600 601 602 603 604
  py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
  pass.def(py::init())
      .def("set_str", [](ir::Pass &self, const std::string &name,
                         const std::string &attr) {
        self.Set<std::string>(name, new std::string(attr));
      });

X
fix  
Xin Pan 已提交
605 606
  py::class_<ir::PassBuilder, std::shared_ptr<ir::PassBuilder>> pb(
      m, "PassBuilder");
607 608 609 610 611 612 613 614 615 616 617 618 619 620
  pb.def(py::init())
      .def("append_pass",
           [](ir::PassBuilder &self,
              const std::string &pass_type) -> std::shared_ptr<ir::Pass> {
             return self.AppendPass(pass_type);
           })
      .def("all_passes", [](ir::PassBuilder &self) { return self.AllPasses(); })
      .def("insert_pass",
           [](ir::PassBuilder &self, size_t idx, const std::string &pass_type) {
             return self.InsertPass(idx, pass_type);
           })
      .def("remove_pass",
           [](ir::PassBuilder &self, size_t idx) { self.RemovePass(idx); });

Y
yuyang18 已提交
621
  // -- python binds for parallel executor.
Y
yuyang18 已提交
622
  py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
C
chengduo 已提交
623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639
  py::class_<ExecutionStrategy> exec_strategy(pe, "ExecutionStrategy", R"DOC(
    ExecutionStrategy allows the user to more preciously control how to run
    the program in ParallelExecutor by setting the property.

    The available properties include:
        use_cuda (bool): Whether to use CUDA or not. Default True.
        num_threads (int): The number of threads that used to run the
            operators in ParallelExecutor. If it is not set, it will be
            set in ParallelExecutor according to the device count.
            Default 0.
        allow_op_delay (bool): Whether to delay the communication operators
            to run. Default False.
        num_iteration_per_drop_scope (int): how many iterations between
            the two dropping local scopes. Default 100.

        )DOC");

Y
yuyang18 已提交
640
  exec_strategy.def(py::init())
Y
yuyang18 已提交
641 642 643 644 645 646 647
      .def_property(
          "num_threads",
          [](const ExecutionStrategy &self) { return self.num_threads_; },
          [](ExecutionStrategy &self, size_t num_threads) {
            self.num_threads_ = num_threads;
          })
      .def_property(
648 649 650 651
          "use_cuda",
          [](const ExecutionStrategy &self) { return self.use_cuda_; },
          [](ExecutionStrategy &self, bool use_cuda) {
            self.use_cuda_ = use_cuda;
Y
yuyang18 已提交
652 653 654 655 656 657
          })
      .def_property(
          "allow_op_delay",
          [](const ExecutionStrategy &self) { return self.allow_op_delay_; },
          [](ExecutionStrategy &self, bool allow_op_delay) {
            self.allow_op_delay_ = allow_op_delay;
Y
yuyang18 已提交
658 659 660 661 662 663 664 665
          })
      .def_property(
          "num_iteration_per_drop_scope",
          [](const ExecutionStrategy &self) {
            return self.num_iteration_per_drop_scope_;
          },
          [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) {
            self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope;
Y
yuyang18 已提交
666
          });
Y
yuyang18 已提交
667
  exec_strategy.def_property(
Y
yuyang18 已提交
668 669 670 671 672 673 674
      "use_experimental_executor",
      [](const ExecutionStrategy &self) {
        return self.type_ == ExecutionStrategy::kExperimental;
      },
      [](ExecutionStrategy &self, bool experimental) {
        self.type_ = experimental ? ExecutionStrategy::kExperimental
                                  : ExecutionStrategy::kDefault;
Y
yuyang18 已提交
675 676
      });

C
chengduo 已提交
677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695
  py::class_<BuildStrategy> build_strategy(pe, "BuildStrategy", R"DOC(
    BuildStrategy allows the user to more preciously control how to
    build the SSA Graph in ParallelExecutor by setting the property.

    The available properties include:
        reduce_strategy (str): There are two reduce strategies, 'AllReduce'
            and 'Reduce'. If you want that all parameters will be optimized
            on all devices, you can choose 'AllReduce'; if you choose
            'Reduce', all parameters will be evenly allocated to different
            devices for optimization, and then broadcast the optimized
            parameter to other devices. Default 'AllReduce'.
        gradient_scale_strategy (str): There are two ways of defining loss@grad,
            'CoeffNumDevice' and 'Customized'. By default, ParallelExecutor
            sets the loss@grad according to the number of devices. If you want
            to customize loss@grad, you can choose 'Customized'.
            Default 'CoeffNumDevice'.
        debug_graphviz_path (str): Whether to write the SSA Graph to file in the
            form of graphviz. It is useful for debugging. Default "".
)DOC");
Y
yuyang18 已提交
696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719

  py::enum_<BuildStrategy::ReduceStrategy>(build_strategy, "ReduceStrategy")
      .value("Reduce", BuildStrategy::ReduceStrategy::kReduce)
      .value("AllReduce", BuildStrategy::ReduceStrategy::kAllReduce);
  py::enum_<BuildStrategy::GradientScaleStrategy>(build_strategy,
                                                  "GradientScaleStrategy")
      .value("CoeffNumDevice",
             BuildStrategy::GradientScaleStrategy::kCoeffNumDevice)
      .value("One", BuildStrategy::GradientScaleStrategy::kOne)
      .value("Customized", BuildStrategy::GradientScaleStrategy::kCustomized);

  build_strategy.def(py::init())
      .def_property(
          "reduce_strategy",
          [](const BuildStrategy &self) { return self.reduce_; },
          [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) {
            self.reduce_ = strategy;
          })
      .def_property(
          "gradient_scale_strategy",
          [](const BuildStrategy &self) { return self.gradient_scale_; },
          [](BuildStrategy &self,
             BuildStrategy::GradientScaleStrategy strategy) {
            self.gradient_scale_ = strategy;
Y
yuyang18 已提交
720 721 722 723 724 725
          })
      .def_property(
          "debug_graphviz_path",
          [](const BuildStrategy &self) { return self.debug_graphviz_path_; },
          [](BuildStrategy &self, const std::string &path) {
            self.debug_graphviz_path_ = path;
F
fengjiayi 已提交
726 727 728 729
          })
      .def_property(
          "enable_data_balance",
          [](const BuildStrategy &self) { return self.enable_data_balance_; },
C
chengduo 已提交
730 731 732 733 734 735 736
          [](BuildStrategy &self, bool b) { self.enable_data_balance_ = b; })
      .def_property("fuse_elewise_add_act_ops",
                    [](const BuildStrategy &self) {
                      return self.fuse_elewise_add_act_ops_;
                    },
                    [](BuildStrategy &self, bool b) {
                      self.fuse_elewise_add_act_ops_ = b;
737
                    })
738
      .def("_create_passes_from_strategy",
X
fix  
Xin Pan 已提交
739 740 741
           [](BuildStrategy &self) -> std::shared_ptr<ir::PassBuilder> {
             return self.CreatePassesFromStrategy();
           });
Y
yuyang18 已提交
742 743 744 745

  pe.def(py::init<const std::vector<platform::Place> &,
                  const std::unordered_set<std::string> &,
                  const std::unordered_set<std::string> &, const ProgramDesc &,
Y
yuyang18 已提交
746
                  const std::string &, Scope *, std::vector<Scope *> &,
747 748
                  const ExecutionStrategy &, const BuildStrategy &, size_t,
                  size_t>())
Y
Yu Yang 已提交
749 750 751 752
      // NOTE: even we return a vec<Scope*>* to Python use reference policy.
      // We still cannot get local_scope from this vector, since the element
      // of vec<Scope*> will be freed by Python GC. We can only return Scope*
      // one by one and mark them as reference.
753 754 755 756 757
      .def("local_scopes",
           [](ParallelExecutor &self) -> std::vector<Scope *> * {
             return &self.GetLocalScopes();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
758 759 760 761
      .def("feed_tensors_into_local_scopes",
           &ParallelExecutor::FeedTensorsIntoLocalScopes)
      .def("feed_and_split_tensor_into_local_scopes",
           &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes)
S
sneaxiy 已提交
762 763 764 765 766 767
      .def("run", [](ParallelExecutor &self,
                     const std::vector<std::string> &fetch_tensors,
                     const std::string &fetched_var_name) {
        pybind11::gil_scoped_release release;
        self.Run(fetch_tensors, fetched_var_name);
      });
Y
Yu Yang 已提交
768

769
  BindRecordIOWriter(&m);
770
  return m.ptr();
L
Luo Tao 已提交
771
}
772
}  // namespace pybind
773
}  // namespace paddle