pybind.cc 20.5 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 36 37 38 39 40
#include "paddle/fluid/framework/selected_rows.h"
#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"
41 42
#include "paddle/fluid/pybind/protobuf.h"
#include "paddle/fluid/pybind/pybind.h"  // NOLINT
Y
Yu Yang 已提交
43
#include "paddle/fluid/pybind/recordio.h"
Y
Yi Wang 已提交
44
#include "paddle/fluid/pybind/tensor_py.h"
Y
Yu Yang 已提交
45

46
#include "paddle/fluid/string/to_string.h"
47

D
Dong Zhihong 已提交
48
#ifdef PADDLE_WITH_CUDA
Y
Yi Wang 已提交
49 50 51
#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
52 53
#endif

Q
Qiao Longfei 已提交
54 55 56
// disable auto conversion to list in Python
PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);

57
namespace paddle {
58
namespace pybind {
59
bool IsCompiledWithCUDA() {
60
#ifndef PADDLE_WITH_CUDA
Q
qijun 已提交
61 62 63 64 65 66
  return false;
#else
  return true;
#endif
}

67 68
PYBIND11_PLUGIN(core) {
  py::module m("core", "C++ core of PaddlePaddle");
69

70 71 72 73
  // using framework in this function. Since it is inside a function, it will
  // not cause namespace pollution.
  using namespace paddle::framework;  // NOLINT

74
  BindException(&m);
Y
Yu Yang 已提交
75

76 77 78
  py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
      .def_buffer(
          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
Y
Yu Yang 已提交
79
      .def("get_dims",
80
           [](const Tensor &self) { return vectorize(self.dims()); })
Y
Yu Yang 已提交
81
      .def("set_dims",
Q
qijun 已提交
82
           [](Tensor &self, const std::vector<int64_t> &dim) {
Y
Yu Yang 已提交
83
             self.Resize(make_ddim(dim));
Y
Yu Yang 已提交
84
           })
D
dzhwinter 已提交
85 86 87 88
      .def("set_layout",
           [](Tensor &self, const std::string &layout) {
             self.set_layout(StringToDataLayout(layout));
           })
Y
Yu Yang 已提交
89
      .def("alloc_float",
D
dzhwinter 已提交
90
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
91
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
92
           })
Q
qijun 已提交
93
      .def("alloc_float",
Y
Yu Yang 已提交
94
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
95
             self.mutable_data<float>(place);
Y
Yu Yang 已提交
96 97
           })
      .def("alloc_int",
Y
Yu Yang 已提交
98
           [](Tensor &self, paddle::platform::CPUPlace &place) {
Q
qijun 已提交
99
             self.mutable_data<int>(place);
Y
Yu Yang 已提交
100
           })
Q
qijun 已提交
101
      .def("alloc_int",
D
dzhwinter 已提交
102
           [](Tensor &self, paddle::platform::CUDAPlace &place) {
Q
qijun 已提交
103
             self.mutable_data<int>(place);
Q
qijun 已提交
104
           })
C
chengduoZH 已提交
105 106 107 108 109 110 111 112
      .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
113 114
      .def("set", PyCPUTensorSetFromArray<float>)
      .def("set", PyCPUTensorSetFromArray<int>)
115
      .def("set", PyCPUTensorSetFromArray<double>)
116
      .def("set", PyCPUTensorSetFromArray<int64_t>)
Y
Yu Yang 已提交
117
      .def("set", PyCPUTensorSetFromArray<bool>)
118
      .def("set", PyCPUTensorSetFromArray<uint16_t>)
119
#ifdef PADDLE_WITH_CUDA
Y
Yu Yang 已提交
120 121
      .def("set", PyCUDATensorSetFromArray<float>)
      .def("set", PyCUDATensorSetFromArray<int>)
122
      .def("set", PyCUDATensorSetFromArray<double>)
123
      .def("set", PyCUDATensorSetFromArray<int64_t>)
Y
Yu Yang 已提交
124
      .def("set", PyCUDATensorSetFromArray<bool>)
125
      .def("set", PyCUDATensorSetFromArray<uint16_t>)
C
chengduoZH 已提交
126 127 128 129 130 131
      .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
132
#endif
133
      .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
134 135 136 137 138
      .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
139

140
  py::class_<LoDTensor, Tensor>(m, "LoDTensor")
141 142
      .def_buffer(
          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
143 144 145
      .def(
          "__init__",
          [](LoDTensor &instance, const std::vector<std::vector<size_t>> &lod) {
D
dzhwinter 已提交
146 147 148 149
            LoD new_lod;
            new_lod.reserve(lod.size());
            std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod));
            new (&instance) LoDTensor(new_lod);
150
          })
Y
Yu Yang 已提交
151
      .def("__init__", [](LoDTensor &instance) { new (&instance) LoDTensor(); })
D
dangqingqing 已提交
152
      .def("set_lod",
153
           [](LoDTensor &self, const std::vector<std::vector<size_t>> &lod) {
Y
Yu Yang 已提交
154
             LoD new_lod;
155 156 157
             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
158
           })
159
      .def("lod", [](LoDTensor &self) -> std::vector<std::vector<size_t>> {
D
dzhwinter 已提交
160 161 162 163 164
        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
165 166
      });

Q
qijun 已提交
167 168 169 170 171 172 173 174 175 176 177 178 179
  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
180 181 182 183 184 185 186 187 188
      .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
           })
189 190 191 192 193 194 195 196 197 198 199
      .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
200

201
  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
202 203 204

All parameter, weight, gradient are variables in Paddle.
)DOC")
205
      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
206
      .def("set_int",
207 208
           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
209 210 211 212 213 214 215
      .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
216
      .def("get_tensor",
217 218
           [](Variable &self) -> LoDTensor * {
             return self.GetMutable<LoDTensor>();
D
dangqingqing 已提交
219 220
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
221 222 223
      .def("get_lod_rank_table",
           [](Variable &self) { return self.GetMutable<LoDRankTable>(); },
           py::return_value_policy::reference)
Q
qijun 已提交
224 225 226 227 228
      .def("get_selected_rows",
           [](Variable &self) -> SelectedRows * {
             return self.GetMutable<SelectedRows>();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
229 230 231
      .def("get_lod_tensor_array",
           [](Variable &self) { return self.GetMutable<LoDTensorArray>(); },
           py::return_value_policy::reference)
D
Dong Zhihong 已提交
232 233 234 235 236 237 238
#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
239 240 241 242 243
      .def("get_reader",
           [](Variable &self) -> framework::ReaderHolder * {
             PADDLE_ENFORCE(self.IsType<framework::ReaderHolder>());
             return self.GetMutable<framework::ReaderHolder>();
           },
Y
Yu Yang 已提交
244
           py::return_value_policy::reference);
245

Y
Refine  
Yu Yang 已提交
246 247 248
  py::class_<framework::ReaderHolder>(m, "Reader", "")
      .def("reset", &framework::ReaderHolder::ReInit);

249
  py::class_<Scope>(m, "Scope", "")
D
dongzhihong 已提交
250
      .def("var",
251
           [](Scope &self, const std::string &name) -> Variable * {
D
dongzhihong 已提交
252
             return self.Var(name);
Y
Yu Yang 已提交
253
           },
254
           py::return_value_policy::reference)
255
      .def("find_var", &Scope::FindVar, py::return_value_policy::reference)
Y
Yu Yang 已提交
256
      .def(py::init<>())
257
      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
258
           py::return_value_policy::reference)
Y
Yu Yang 已提交
259
      .def("drop_kids", &Scope::DropKids);
260

Y
Yu Yang 已提交
261 262
  //! @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
263 264
  m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
    std::vector<py::bytes> ret_values;
265 266 267 268 269 270 271 272 273 274
    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
275 276
    return ret_values;
  });
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
  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
293
  m.def("prune", [](const ProgramDesc &origin,
294
                    const std::vector<std::array<size_t, 2>> &targets) {
Y
Yu Yang 已提交
295
    ProgramDesc prog_with_targets(origin);
296
    for (const auto &t : targets) {
297
      prog_with_targets.MutableBlock(t[0])->Op(t[1])->MarkAsTarget();
298
    }
299
    proto::ProgramDesc pruned_desc;
300
    Prune(*prog_with_targets.Proto(), &pruned_desc);
Y
Yu Yang 已提交
301
    return new ProgramDesc(pruned_desc);
302
  });
Y
Yu Yang 已提交
303
  m.def("inference_optimize", [](ProgramDesc &origin) {
304
    proto::ProgramDesc pruned_desc;
305
    InferenceOptimize(*(origin.Proto()), &pruned_desc);
Y
Yu Yang 已提交
306
    return new ProgramDesc(pruned_desc);
307
  });
F
fengjiayi 已提交
308 309
  m.def("empty_var_name", []() { return framework::kEmptyVarName; });
  m.def("grad_var_suffix", []() { return framework::kGradVarSuffix; });
310 311 312
  m.def_submodule(
       "var_names",
       "The module will return special predefined variable name in Paddle")
Y
Yi Wang 已提交
313 314
      .def("empty", []() { return kEmptyVarName; })
      .def("temp", []() { return kTempVarName; });
Q
qijun 已提交
315
  // clang-format off
Y
Yu Yang 已提交
316
  py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
Q
qijun 已提交
317 318
      .def_static("create",
                  [](paddle::platform::CPUPlace& place)
Q
qijun 已提交
319
                      -> paddle::platform::DeviceContext* {
Q
qijun 已提交
320 321 322
                    return new paddle::platform::CPUDeviceContext();
                  })
      .def_static("create",
D
dzhwinter 已提交
323
                  [](paddle::platform::CUDAPlace& place)
Q
qijun 已提交
324
                      -> paddle::platform::DeviceContext* {
325
#ifndef PADDLE_WITH_CUDA
D
dzhwinter 已提交
326
                    PADDLE_THROW("CUDAPlace is not supported in CPU device.");
Q
qijun 已提交
327
#else
Q
qijun 已提交
328
                    return new paddle::platform::CUDADeviceContext(place);
Q
qijun 已提交
329
#endif
C
chengduoZH 已提交
330 331 332 333 334 335 336 337 338 339 340
                  })
          .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
341 342 343 344
// clang-format on
#ifdef PADDLE_WITH_CUDA
  py::class_<platform::Communicator>(m, "Communicator").def(py::init<>());
#endif
D
dzhwinter 已提交
345
  py::class_<platform::CUDAPlace>(m, "CUDAPlace")
346
      .def(py::init<int>())
D
dzhwinter 已提交
347
      .def("__str__", string::to_string<const platform::CUDAPlace &>);
Q
qijun 已提交
348

349 350 351
  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
      .def(py::init<>())
      .def("__str__", string::to_string<const platform::CPUPlace &>);
Y
Yu Yang 已提交
352

C
chengduoZH 已提交
353 354 355 356
  py::class_<paddle::platform::CUDAPinnedPlace>(m, "CUDAPinnedPlace")
      .def(py::init<>())
      .def("__str__", string::to_string<const platform::CUDAPinnedPlace &>);

Y
Yu Yang 已提交
357 358 359 360 361 362 363
  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
364
           [](platform::Place &self, const platform::CUDAPlace &gpu_place) {
Y
Yu Yang 已提交
365
             self = gpu_place;
C
chengduoZH 已提交
366 367
           })
      .def("set_place", [](platform::Place &self,
C
chengduoZH 已提交
368 369
                           const platform::CUDAPinnedPlace &cuda_pinned_place) {
        self = cuda_pinned_place;
C
chengduoZH 已提交
370
      });
Y
Yu Yang 已提交
371

Y
Yu Yang 已提交
372 373 374
  py::class_<OperatorBase>(m, "Operator")
      .def_static("create",
                  [](py::bytes protobin) {
375
                    proto::OpDesc desc;
Y
Yu Yang 已提交
376 377 378 379 380
                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
                                   "Cannot parse user input to OpDesc");
                    PADDLE_ENFORCE(desc.IsInitialized(),
                                   "User OpDesc is not initialized, reason %s",
                                   desc.InitializationErrorString());
381
                    return OpRegistry::CreateOp(desc);
Y
Yu Yang 已提交
382
                  })
383
      .def("run",
384
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
385 386 387
              const platform::CPUPlace &place) { self.Run(scope, place); })
      .def("run",
           [](OperatorBase &self, const Scope &scope,
D
dzhwinter 已提交
388
              const platform::CUDAPlace &place) { self.Run(scope, place); })
C
chengduoZH 已提交
389 390 391 392 393
      .def("run",
           [](OperatorBase &self, const Scope &scope,
              const platform::CUDAPinnedPlace &place) {
             self.Run(scope, place);
           })
Y
Yu Yang 已提交
394 395 396 397 398 399 400
      .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
401 402
      .def("output_vars",
           [](const OperatorBase &op) { return op.OutputVars(true); })
Y
Yu Yang 已提交
403
      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
Q
qijun 已提交
404
      .def("input_vars", [](const OperatorBase &op) { return op.InputVars(); })
Y
Yu Yang 已提交
405 406 407 408
      .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
409

F
fengjiayi 已提交
410
  py::class_<framework::Executor>(m, "Executor")
D
dzhwinter 已提交
411
      .def(py::init<const platform::Place &>())
412 413 414
      .def("run",
           (void (Executor::*)(const ProgramDesc &, Scope *, int, bool, bool)) &
               Executor::Run);
F
fengjiayi 已提交
415

D
dzhwinter 已提交
416
  m.def("init_gflags", framework::InitGflags);
Y
Yang Yu 已提交
417
  m.def("init_glog", framework::InitGLOG);
X
Xin Pan 已提交
418 419
  m.def("init_devices",
        [](bool init_p2p) { framework::InitDevices(init_p2p); });
420

421
  m.def("is_compiled_with_cuda", IsCompiledWithCUDA);
422 423 424 425 426 427
#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
428

429
  m.def("set_feed_variable", framework::SetFeedVariable);
Q
qijun 已提交
430
  m.def("get_fetch_variable", framework::GetFetchVariable);
Q
qijun 已提交
431

432 433 434 435 436
  BindProgramDesc(&m);
  BindBlockDesc(&m);
  BindVarDsec(&m);
  BindOpDesc(&m);
  BindConstValue(&m);
Y
Yu Yang 已提交
437

Y
Yu Yang 已提交
438 439 440 441 442 443 444 445 446
  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
447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
  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());
      });

Y
Yu Yang 已提交
464
  m.def("op_support_gpu", OpSupportGPU);
D
Dong Zhihong 已提交
465
#ifdef PADDLE_WITH_CUDA
D
Dong Zhihong 已提交
466
  m.def("get_cuda_device_count", platform::GetCUDADeviceCount);
D
dangqingqing 已提交
467 468 469 470

  m.def("nvprof_init", platform::CudaProfilerInit);
  m.def("nvprof_start", platform::CudaProfilerStart);
  m.def("nvprof_stop", platform::CudaProfilerStop);
D
Dong Zhihong 已提交
471
#endif
Y
Yu Yang 已提交
472

473 474 475 476
  py::enum_<platform::ProfilerState>(m, "ProfilerState", py::arithmetic())
      .value("kDisabled", platform::ProfilerState::kDisabled)
      .value("kCPU", platform::ProfilerState::kCPU)
      .value("kCUDA", platform::ProfilerState::kCUDA)
477
      .value("kAll", platform::ProfilerState::kAll)
478 479 480 481 482 483 484 485 486 487 488 489 490 491
      .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
492

Y
Yu Yang 已提交
493
  py::class_<ParallelExecutor>(m, "ParallelExecutor")
Y
Yu Yang 已提交
494
      .def("__init__",
495
           [](ParallelExecutor &self, size_t num_threads, bool use_event,
Y
Yu Yang 已提交
496 497
              const std::vector<platform::Place> &places,
              const std::unordered_set<std::string> &params,
498
              const std::unordered_set<std::string> &bcast_vars,
Y
Yu Yang 已提交
499
              const ProgramDesc &main_program, const std::string &loss_var_name,
500 501 502 503 504 505
              Scope *scope, std::vector<Scope *> &local_scopes,
              bool allow_op_delay) {
             new (&self)
                 ParallelExecutor(num_threads, use_event, places, params,
                                  bcast_vars, main_program, loss_var_name,
                                  scope, local_scopes, allow_op_delay);
Y
Yu Yang 已提交
506
           })
T
typhoonzero 已提交
507
      .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
508 509 510 511 512
      .def("local_scopes",
           [](ParallelExecutor &self) -> std::vector<Scope *> * {
             return &self.GetLocalScopes();
           },
           py::return_value_policy::reference)
Y
Yu Yang 已提交
513
      .def("run", &ParallelExecutor::Run);
Y
Yu Yang 已提交
514

515
  BindRecordIOWriter(&m);
516
  return m.ptr();
L
Luo Tao 已提交
517
}
518
}  // namespace pybind
519
}  // namespace paddle