tensor_bind.h 4.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

   http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License. */

#pragma once
Q
qijun 已提交
16
#include <string>
Q
qijun 已提交
17 18 19 20
#include "paddle/framework/tensor.h"
#include "paddle/memory/memcpy.h"
#include "pybind11/numpy.h"
#include "pybind11/pybind11.h"
21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44

namespace py = pybind11;

namespace paddle {

namespace pybind {

namespace details {

template <bool less, size_t I, typename... ARGS>
struct CastToPyBufferImpl;

template <size_t I, typename... ARGS>
struct CastToPyBufferImpl<false, I, ARGS...> {
  py::buffer_info operator()(framework::Tensor &tensor) {
    PADDLE_THROW("This type of tensor cannot be expose to Python");
    return py::buffer_info();
  }
};

template <size_t I, typename... ARGS>
struct CastToPyBufferImpl<true, I, ARGS...> {
  using CUR_TYPE = typename std::tuple_element<I, std::tuple<ARGS...>>::type;
  py::buffer_info operator()(framework::Tensor &tensor) {
45 46 47 48
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(tensor.holder_->place()),
                   "Only CPU tensor can cast to numpy array");

    if (std::type_index(typeid(CUR_TYPE)) == tensor.holder_->type()) {
49 50 51 52 53 54 55 56 57 58 59 60
      auto dim_vec = framework::vectorize(tensor.dims());
      std::vector<size_t> dims_outside;
      std::vector<size_t> strides;
      dims_outside.resize(dim_vec.size());
      strides.resize(dim_vec.size());

      size_t prod = 1;
      for (size_t i = dim_vec.size(); i != 0; --i) {
        dims_outside[i - 1] = (size_t)dim_vec[i - 1];
        strides[i - 1] = sizeof(CUR_TYPE) * prod;
        prod *= dims_outside[i - 1];
      }
Q
qijun 已提交
61
      framework::Tensor dst_tensor;
Q
qijun 已提交
62
      if (paddle::platform::is_gpu_place(tensor.holder_->place())) {
Q
qijun 已提交
63
        dst_tensor.CopyFrom<CUR_TYPE>(tensor, platform::CPUPlace());
Q
qijun 已提交
64
      } else if (paddle::platform::is_cpu_place(tensor.holder_->place())) {
Q
qijun 已提交
65 66
        dst_tensor = tensor;
      }
67
      return py::buffer_info(
Q
qijun 已提交
68
          dst_tensor.mutable_data<CUR_TYPE>(dst_tensor.holder_->place()),
69 70
          sizeof(CUR_TYPE),
          py::format_descriptor<CUR_TYPE>::format(),
Q
qijun 已提交
71
          (size_t)framework::arity(dst_tensor.dims()),
72 73
          dims_outside,
          strides);
74 75 76 77 78 79 80 81 82 83 84 85 86
    } else {
      constexpr bool less = I + 1 < std::tuple_size<std::tuple<ARGS...>>::value;
      return CastToPyBufferImpl<less, I + 1, ARGS...>()(tensor);
    }
  }
};
}  // namespace details
inline py::buffer_info CastToPyBuffer(framework::Tensor &tensor) {
  auto buffer_info = details::CastToPyBufferImpl<true, 0, float, int>()(tensor);
  return buffer_info;
}

template <typename T>
Q
qijun 已提交
87
void PyCPUTensorSetFromArray(
88
    framework::Tensor &self,
Q
qijun 已提交
89 90
    py::array_t<T, py::array::c_style | py::array::forcecast> array,
    paddle::platform::CPUPlace &place) {
91 92 93 94 95 96
  std::vector<int> dims;
  dims.reserve(array.ndim());
  for (size_t i = 0; i < array.ndim(); ++i) {
    dims.push_back((int)array.shape()[i]);
  }

F
fengjiayi 已提交
97
  self.Resize(framework::make_ddim(dims));
Q
qijun 已提交
98 99 100 101 102 103 104 105 106 107 108 109 110
  auto *dst = self.mutable_data<T>(place);
  std::memcpy(dst, array.data(), sizeof(T) * array.size());
}

template <typename T>
void PyCUDATensorSetFromArray(
    framework::Tensor &self,
    py::array_t<T, py::array::c_style | py::array::forcecast> array,
    paddle::platform::GPUPlace &place) {
  std::vector<int> dims;
  dims.reserve(array.ndim());
  for (size_t i = 0; i < array.ndim(); ++i) {
    dims.push_back((int)array.shape()[i]);
Q
qijun 已提交
111
  }
Q
qijun 已提交
112 113 114 115 116 117

  self.Resize(framework::make_ddim(dims));
  auto *dst = self.mutable_data<T>(place);
  std::memcpy(dst, array.data(), sizeof(T) * array.size());
  paddle::platform::GpuMemcpySync(
      dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice);
118 119 120 121
}

}  // namespace pybind
}  // namespace paddle