pybind.cc 21.3 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 17 18 19 20 21
#include <algorithm>
#include <map>
#include <mutex>  // NOLINT // for call_once
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
22

23
#include "paddle/fluid/framework/channel.h"
Y
Yi Wang 已提交
24 25 26 27 28 29 30
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/init.h"
#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"
D
dzhwinter 已提交
36
#include "paddle/fluid/operators/activation_op.h"
Y
Yi Wang 已提交
37 38 39 40 41
#include "paddle/fluid/platform/enforce.h"
#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"
42 43
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h"  // NOLINT
Y
Yu Yang 已提交
44
#include "paddle/fluid/pybind/recordio.h"
Y
Yi Wang 已提交
45
#include "paddle/fluid/pybind/tensor_py.h"
Y
Yu Yang 已提交
46

47
#include "paddle/fluid/string/to_string.h"
48

D
Dong Zhihong 已提交
49
#ifdef PADDLE_WITH_CUDA
Y
Yi Wang 已提交
50 51 52
#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
53 54
#endif

Q
Qiao Longfei 已提交
55 56 57
// disable auto conversion to list in Python
PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);

58
namespace paddle {
59
namespace pybind {
60
bool IsCompiledWithCUDA() {
61
#ifndef PADDLE_WITH_CUDA
Q
qijun 已提交
62 63 64 65 66 67
  return false;
#else
  return true;
#endif
}

68 69
PYBIND11_PLUGIN(core) {
  py::module m("core", "C++ core of PaddlePaddle");
70

71 72 73 74
  // using framework in this function. Since it is inside a function, it will
  // not cause namespace pollution.
  using namespace paddle::framework;  // NOLINT

75
  BindException(&m);
Y
Yu Yang 已提交
76

77 78 79
  py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
      .def_buffer(
          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
Y
Yu Yang 已提交
80
      .def("get_dims",
81
           [](const Tensor &self) { return vectorize(self.dims()); })
Y
Yu Yang 已提交
82
      .def("set_dims",
Q
qijun 已提交
83
           [](Tensor &self, const std::vector<int64_t> &dim) {
Y
Yu Yang 已提交
84
             self.Resize(make_ddim(dim));
Y
Yu Yang 已提交
85
           })
D
dzhwinter 已提交
86 87 88 89
      .def("set_layout",
           [](Tensor &self, const std::string &layout) {
             self.set_layout(StringToDataLayout(layout));
           })
Y
Yu Yang 已提交
90
      .def("alloc_float",
D
dzhwinter 已提交
91
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
92
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
93
           })
Q
qijun 已提交
94
      .def("alloc_float",
Y
Yu Yang 已提交
95
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
96
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
97 98
           })
      .def("alloc_int",
Y
Yu Yang 已提交
99
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
100
             self.mutable_data<int>(place);
Y
Yu Yang 已提交
101
           })
Q
qijun 已提交
102
      .def("alloc_int",
D
dzhwinter 已提交
103
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
104
             self.mutable_data<int>(place);
Q
qijun 已提交
105
           })
C
chengduoZH 已提交
106 107 108 109 110 111 112 113
      .def("alloc_int",
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<int>(place);
           })
      .def("alloc_float",
           [](Tensor &self, paddle::platform::CUDAPinnedPlace &place) {
             self.mutable_data<float>(place);
           })
Y
Yu Yang 已提交
114 115
      .def("set", PyCPUTensorSetFromArray<float>)
      .def("set", PyCPUTensorSetFromArray<int>)
116
      .def("set", PyCPUTensorSetFromArray<double>)
117
      .def("set", PyCPUTensorSetFromArray<int64_t>)
Y
Yu Yang 已提交
118
      .def("set", PyCPUTensorSetFromArray<bool>)
119
      .def("set", PyCPUTensorSetFromArray<uint16_t>)
120
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
121 122
      .def("set", PyCUDATensorSetFromArray<float>)
      .def("set", PyCUDATensorSetFromArray<int>)
123
      .def("set", PyCUDATensorSetFromArray<double>)
124
      .def("set", PyCUDATensorSetFromArray<int64_t>)
Y
Yu Yang 已提交
125
      .def("set", PyCUDATensorSetFromArray<bool>)
126
      .def("set", PyCUDATensorSetFromArray<uint16_t>)
C
chengduoZH 已提交
127 128 129 130 131 132
      .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>)
Q
qijun 已提交
133
#endif
134
      .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
135 136 137 138 139
      .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
140

141
  py::class_<LoDTensor, Tensor>(m, "LoDTensor")
142 143
      .def_buffer(
          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
144 145 146
      .def(
          "__init__",
          [](LoDTensor &instance, const std::vector<std::vector<size_t>> &lod) {
D
dzhwinter 已提交
147 148 149 150
            LoD new_lod;
            new_lod.reserve(lod.size());
            std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
            new (&instance) LoDTensor(new_lod);
151
          })
Y
Yu Yang 已提交
152
      .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
D
dangqingqing 已提交
153
      .def("set_lod",
154
           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
Y
Yu Yang 已提交
155
             LoD new_lod;
156 157 158
             new_lod.reserve(lod.size());
             std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
             self.set_lod(new_lod);
D
dangqingqing 已提交
159
           })
160
      .def("lod", [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
D
dzhwinter 已提交
161 162 163 164 165
        auto 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;
D
dangqingqing 已提交
166 167
      });

Q
qijun 已提交
168 169 170 171 172 173 174 175 176 177 178 179 180
  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
181 182 183 184 185 186 187 188 189
      .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
           })
190 191 192 193 194 195 196 197 198 199 200
      .def("rows", [](SelectedRows &self) {
#ifndef PADDLE_WITH_CUDA
        return self.rows();
#else
         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;
#endif
      });
Q
qijun 已提交
201

202
  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
203 204 205

All parameter, weight, gradient are variables in Paddle.
)DOC")
206
      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
207
      .def("set_int",
208 209
           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
210 211 212 213 214 215 216
      .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
217
      .def("get_tensor",
218 219
           [](Variable &self) -> LoDTensor * {
             return self.GetMutable<LoDTensor>();
D
dangqingqing 已提交
220 221
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
222 223 224
      .def("get_lod_rank_table",
           [](Variable &self) { return self.GetMutable<LoDRankTable>(); },
           py::return_value_policy::reference)
Q
qijun 已提交
225 226 227 228 229
      .def("get_selected_rows",
           [](Variable &self) -> SelectedRows * {
             return self.GetMutable<SelectedRows>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
230 231 232
      .def("get_lod_tensor_array",
           [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
           py::return_value_policy::reference)
D
Dong Zhihong 已提交
233 234 235 236 237 238 239
#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
240 241 242 243 244
      .def("get_reader",
           [](Variable &self) -> framework::ReaderHolder * {
             PADDLE_ENFORCE(self.IsType<framework::ReaderHolder>());
             return self.GetMutable<framework::ReaderHolder>();
           },
Y
Yu Yang 已提交
245
           py::return_value_policy::reference);
246

Y
Refine  
Yu Yang 已提交
247 248 249
  py::class_<framework::ReaderHolder>(m, "Reader", "")
      .def("reset", &framework::ReaderHolder::ReInit);

250
  py::class_<Scope>(m, "Scope", "")
D
dongzhihong 已提交
251
      .def("var",
252
           [](Scope &self, const std::string &name) -> Variable * {
D
dongzhihong 已提交
253
             return self.Var(name);
Y
Yu Yang 已提交
254
           },
255
           py::return_value_policy::reference)
256
      .def("find_var", &Scope::FindVar, py::return_value_policy::reference)
Y
Yu Yang 已提交
257
      .def(py::init<>())
258
      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
259
           py::return_value_policy::reference)
Y
Yu Yang 已提交
260
      .def("drop_kids", &Scope::DropKids);
261

Y
Yu Yang 已提交
262 263
  //! @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
264 265
  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
    std::vector<py::bytes> ret_values;
266 267 268 269 270 271 272 273 274 275
    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
276 277
    return ret_values;
  });
278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293
  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
294
  m.def("prune", [](const ProgramDesc &origin,
295
                    const std::vector<std::array<size_t, 2>> &targets) {
Y
Yu Yang 已提交
296
    ProgramDesc prog_with_targets(origin);
297
    for (const auto &t : targets) {
298
      prog_with_targets.MutableBlock(t[0])->Op(t[1])->SetIsTarget(true);
299
    }
300
    proto::ProgramDesc pruned_desc;
301
    Prune(*prog_with_targets.Proto(), &pruned_desc);
Y
Yu Yang 已提交
302
    return new ProgramDesc(pruned_desc);
303
  });
Y
Yu Yang 已提交
304
  m.def("inference_optimize", [](ProgramDesc &origin) {
305
    proto::ProgramDesc pruned_desc;
306
    InferenceOptimize(*(origin.Proto()), &pruned_desc);
Y
Yu Yang 已提交
307
    return new ProgramDesc(pruned_desc);
308
  });
F
fengjiayi 已提交
309 310
  m.def("empty_var_name", []() { return framework::kEmptyVarName; });
  m.def("grad_var_suffix", []() { return framework::kGradVarSuffix; });
311 312 313
  m.def_submodule(
       "var_names",
       "The module will return special predefined variable name in Paddle")
Y
Yi Wang 已提交
314 315
      .def("empty", []() { return kEmptyVarName; })
      .def("temp", []() { return kTempVarName; });
Q
qijun 已提交
316
  // clang-format off
Y
Yu Yang 已提交
317
  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
Q
qijun 已提交
318 319
      .def_static("create",
                  [](paddle::platform::CPUPlace& place)
Q
qijun 已提交
320
                      -> paddle::platform::DeviceContext* {
Q
qijun 已提交
321 322 323
                    return new paddle::platform::CPUDeviceContext();
                  })
      .def_static("create",
D
dzhwinter 已提交
324
                  [](paddle::platform::CUDAPlace& place)
Q
qijun 已提交
325
                      -> paddle::platform::DeviceContext* {
326
#ifndef PADDLE_WITH_CUDA
D
dzhwinter 已提交
327
                    PADDLE_THROW("CUDAPlace is not supported in CPU device.");
Q
qijun 已提交
328
#else
Q
qijun 已提交
329
                    return new paddle::platform::CUDADeviceContext(place);
Q
qijun 已提交
330
#endif
C
chengduoZH 已提交
331 332 333 334 335 336 337 338 339 340 341
                  })
          .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
342 343 344 345
// clang-format on
#ifdef PADDLE_WITH_CUDA
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
#endif
D
dzhwinter 已提交
346
  py::class_<platform::CUDAPlace>(m, "CUDAPlace")
347
      .def(py::init<int>())
D
dzhwinter 已提交
348
      .def("__str__", string::to_string<const platform::CUDAPlace &>);
Q
qijun 已提交
349

350 351 352
  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
      .def(py::init<>())
      .def("__str__", string::to_string<const platform::CPUPlace &>);
Y
Yu Yang 已提交
353

C
chengduoZH 已提交
354 355 356 357
  py::class_<paddle::platform::CUDAPinnedPlace>(m, "CUDAPinnedPlace")
      .def(py::init<>())
      .def("__str__", string::to_string<const platform::CUDAPinnedPlace &>);

Y
Yu Yang 已提交
358 359 360 361 362 363 364
  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
365
           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
Y
Yu Yang 已提交
366
             self = gpu_place;
C
chengduoZH 已提交
367 368
           })
      .def("set_place", [](platform::Place &self,
C
chengduoZH 已提交
369 370
                           const platform::CUDAPinnedPlace &cuda_pinned_place) {
        self = cuda_pinned_place;
C
chengduoZH 已提交
371
      });
Y
Yu Yang 已提交
372

Y
Yu Yang 已提交
373 374 375
  py::class_<OperatorBase>(m, "Operator")
      .def_static("create",
                  [](py::bytes protobin) {
376
                    proto::OpDesc desc;
Y
Yu Yang 已提交
377 378 379 380 381
                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
                                   "Cannot parse user input to OpDesc");
                    PADDLE_ENFORCE(desc.IsInitialized(),
                                   "User OpDesc is not initialized, reason %s",
                                   desc.InitializationErrorString());
382
                    return OpRegistry::CreateOp(desc);
Y
Yu Yang 已提交
383
                  })
384
      .def("run",
385
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
386 387 388
              const platform::CPUPlace &place) { self.Run(scope, place); })
      .def("run",
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
389
              const platform::CUDAPlace &place) { self.Run(scope, place); })
C
chengduoZH 已提交
390 391 392 393 394
      .def("run",
           [](OperatorBase &self, const Scope &scope,
              const platform::CUDAPinnedPlace &place) {
             self.Run(scope, place);
           })
Y
Yu Yang 已提交
395 396 397 398 399 400 401
      .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
402 403
      .def("output_vars",
           [](const OperatorBase &op) { return op.OutputVars(true); })
Y
Yu Yang 已提交
404
      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
Q
qijun 已提交
405
      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
Y
Yu Yang 已提交
406 407 408 409
      .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
410

F
fengjiayi 已提交
411
  py::class_<framework::Executor>(m, "Executor")
D
dzhwinter 已提交
412
      .def(py::init<const platform::Place &>())
413 414 415
      .def("run",
           (void (Executor::*)(const ProgramDesc &, Scope *, int, bool, bool)) &
               Executor::Run);
F
fengjiayi 已提交
416

D
dzhwinter 已提交
417
  m.def("init_gflags", framework::InitGflags);
Y
Yang Yu 已提交
418
  m.def("init_glog", framework::InitGLOG);
X
Xin Pan 已提交
419 420
  m.def("init_devices",
        [](bool init_p2p) { framework::InitDevices(init_p2p); });
421

422
  m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
423 424 425 426 427 428
#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
429

430
  m.def("set_feed_variable", framework::SetFeedVariable);
Q
qijun 已提交
431
  m.def("get_fetch_variable", framework::GetFetchVariable);
Q
qijun 已提交
432

433 434 435 436 437
  BindProgramDesc(&m);
  BindBlockDesc(&m);
  BindVarDsec(&m);
  BindOpDesc(&m);
  BindConstValue(&m);
Y
Yu Yang 已提交
438

Y
Yu Yang 已提交
439 440 441 442 443 444 445 446 447
  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
448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464
  py::class_<LoDTensorArray>(m, "LoDTensorArray")
      .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
465 466 467
  m.def("IsInplace",
        [](std::string op) -> bool { return operators::IsInplace(op); });

Y
Yu Yang 已提交
468
  m.def("op_support_gpu", OpSupportGPU);
D
Dong Zhihong 已提交
469
#ifdef PADDLE_WITH_CUDA
D
Dong Zhihong 已提交
470
  m.def("get_cuda_device_count", platform::GetCUDADeviceCount);
D
dangqingqing 已提交
471 472 473 474

  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
D
Dong Zhihong 已提交
475
#endif
Y
Yu Yang 已提交
476

477 478 479 480
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
      .value("kCUDA", platform::ProfilerState::kCUDA)
481
      .value("kAll", platform::ProfilerState::kAll)
482 483 484 485 486 487 488 489 490 491 492 493 494 495
      .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);
  m.def("reset_profiler", platform::ResetProfiler);
Y
Yu Yang 已提交
496

Y
Yu Yang 已提交
497
  py::class_<ParallelExecutor>(m, "ParallelExecutor")
Y
Yu Yang 已提交
498
      .def("__init__",
499
           [](ParallelExecutor &self, size_t num_threads, bool use_event,
Y
Yu Yang 已提交
500 501
              const std::vector<platform::Place> &places,
              const std::unordered_set<std::string> &params,
502
              const std::unordered_set<std::string> &bcast_vars,
Y
Yu Yang 已提交
503
              const ProgramDesc &main_program, const std::string &loss_var_name,
504
              Scope *scope, std::vector<Scope *> &local_scopes,
C
chengduoZH 已提交
505 506
              bool allow_op_delay, bool use_default_grad_scale,
              bool balance_parameter_opt_between_cards) {
507 508 509
             new (&self) ParallelExecutor(
                 num_threads, use_event, places, params, bcast_vars,
                 main_program, loss_var_name, scope, local_scopes,
C
chengduoZH 已提交
510 511
                 allow_op_delay, use_default_grad_scale,
                 balance_parameter_opt_between_cards);
Y
Yu Yang 已提交
512
           })
T
typhoonzero 已提交
513
      .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
Y
Yu Yang 已提交
514 515 516 517
      // 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.
518 519 520 521 522
      .def("local_scopes",
           [](ParallelExecutor &self) -> std::vector<Scope *> * {
             return &self.GetLocalScopes();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
523 524 525 526
      .def("feed_tensors_into_local_scopes",
           &ParallelExecutor::FeedTensorsIntoLocalScopes)
      .def("feed_and_split_tensor_into_local_scopes",
           &ParallelExecutor::FeedAndSplitTensorIntoLocalScopes)
Y
Yu Yang 已提交
527
      .def("run", &ParallelExecutor::Run);
Y
Yu Yang 已提交
528

529
  BindRecordIOWriter(&m);
530
  return m.ptr();
L
Luo Tao 已提交
531
}
532
}  // namespace pybind
533
}  // namespace paddle