strided_memcpy.h 5.5 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
L
Luo Tao 已提交
2 3 4 5 6 7 8 9 10
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. */
Y
Yu Yang 已提交
11 12

#pragma once
C
chengduo 已提交
13
#include <vector>
W
wanghuancoder 已提交
14

C
chengduo 已提交
15
#include "paddle/fluid/framework/tensor.h"
Y
Yi Wang 已提交
16
#include "paddle/fluid/operators/detail/strided_memcpy.h"
W
wanghuancoder 已提交
17

Y
Yu Yang 已提交
18 19 20
namespace paddle {
namespace operators {

21
// Strided memory copy from src to dst.
Y
Yu Yang 已提交
22
//
23 24 25 26
// The src and dst should be both on dev_ctx.GetPlace(), otherwise, there will
// be a segment fault.
//
// The stride of an array (also referred to as increment, pitch or step size) is
Y
Yu Yang 已提交
27 28 29 30 31 32 33 34 35
// the number of locations in memory between beginnings of successive array
// elements
//
// For example, for tensor like [1, 3, 300, 300]. If there is no padding, the
// stride is [270000, 90000, 300, 1].
//
// NOTE: When use GPU, the memcpy is async. To sync memcpy, please invoke
// `dev_ctx.Wait()`.
template <typename T>
36 37 38 39
inline void StridedMemcpy(const platform::DeviceContext& dev_ctx, const T* src,
                          const framework::DDim& src_stride,
                          const framework::DDim& dst_dim,
                          const framework::DDim& dst_stride, T* dst) {
40 41
  paddle::operators::detail::StridedCopyDimVisitor<T> func(
      dev_ctx, src, src_stride, dst_stride, dst);
S
sneaxiy 已提交
42
  dst_dim.apply_visitor(func);
Y
Yu Yang 已提交
43
}
Y
Yancey1989 已提交
44 45 46 47 48 49 50 51 52

// Strided numel memory copy from src to dst by the specified axis
//
// For example, for a tensor dims [4, 20, 100], the strieded numel is
// [8000, 2000, 100]
//
// NOTE: The src and dst tensor should have the same elements
// except the specified axis.
template <typename T>
Y
fix ci  
Yancey1989 已提交
53
inline void StridedNumelCopyWithAxis(const platform::DeviceContext& ctx,
Y
Yancey1989 已提交
54 55 56
                                     int64_t axis, T* dst,
                                     const framework::DDim& dst_stride_numel,
                                     const T* src,
T
typhoonzero 已提交
57 58
                                     const framework::DDim& src_stride_numel,
                                     int64_t size) {
Y
Yancey1989 已提交
59 60 61 62 63 64
  int64_t before = dst_stride_numel[0] / dst_stride_numel[axis];
  int64_t src_after = src_stride_numel[axis];
  int64_t dst_after = dst_stride_numel[axis];
  auto place = ctx.GetPlace();

  PADDLE_ENFORCE_EQ(src_stride_numel.size(), dst_stride_numel.size(),
65 66 67 68 69
                    platform::errors::InvalidArgument(
                        "Source and destination tensor should have the same "
                        "dimension size, but source tensor dimension size is "
                        "%u, destination tensor size is %u.",
                        src_stride_numel.size(), dst_stride_numel.size()));
Y
Yancey1989 已提交
70 71 72

  for (int64_t i = 0; i < axis; ++i) {
    if (i < axis) {
73 74 75 76 77 78 79 80 81
      PADDLE_ENFORCE_EQ(
          src_stride_numel[i] / src_stride_numel[axis],
          dst_stride_numel[i] / dst_stride_numel[axis],
          platform::errors::InvalidArgument(
              "Source and destination tensor should have the same number of "
              "elements except the specified axis, but the source elements "
              "number is %d, destination elements number is %d.",
              src_stride_numel[i] / src_stride_numel[axis],
              dst_stride_numel[i] / dst_stride_numel[axis]));
Y
Yancey1989 已提交
82 83 84
    } else if (i == axis) {
      continue;
    } else {
85 86 87 88 89 90 91
      PADDLE_ENFORCE_EQ(
          src_stride_numel[i], dst_stride_numel[i],
          platform::errors::InvalidArgument(
              "Source and destination tensor should have the same number of "
              "elements except the specified axis, but the source elements "
              "number is %d, destination elements number is %d.",
              src_stride_numel[i], dst_stride_numel[i]));
Y
Yancey1989 已提交
92 93 94 95 96
    }
  }

  for (int64_t i = 0; i < before; ++i) {
    if (platform::is_cpu_place(place)) {
97
      auto& cpu_place = BOOST_GET_CONST(platform::CPUPlace, place);
Y
Yancey1989 已提交
98
      memory::Copy(cpu_place, dst + i * dst_after, cpu_place,
T
typhoonzero 已提交
99
                   src + i * src_after, sizeof(T) * size);
Y
Yancey1989 已提交
100
    } else {
101
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
102
      auto& gpu_place = BOOST_GET_CONST(platform::CUDAPlace, place);
Y
Yancey1989 已提交
103 104
      auto& cuda_ctx =
          reinterpret_cast<const platform::CUDADeviceContext&>(ctx);
Y
fix ci  
Yancey1989 已提交
105
      memory::Copy(gpu_place, dst + i * dst_after, gpu_place,
T
typhoonzero 已提交
106
                   src + i * src_after, sizeof(T) * size, cuda_ctx.stream());
Y
Yancey1989 已提交
107
#else
108 109
      PADDLE_THROW(platform::errors::PreconditionNotMet(
          "Paddle is not compiled with GPU."));
Y
Yancey1989 已提交
110 111 112 113 114
#endif
    }
  }
}

C
chengduo 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
template <typename T>
inline void StridedMemcpyWithAxis0(
    const platform::DeviceContext& dev_ctx, const framework::Tensor& input,
    const std::vector<const framework::Tensor*>& shape_refer,
    std::vector<framework::Tensor*>* outputs) {
  const framework::DDim in_stride = stride_numel(input.dims());
  const int axis = 0;
  size_t input_offset = 0;

  for (size_t i = 0; i < outputs->size(); ++i) {
    auto out_stride = stride_numel(shape_refer[i]->dims());
    auto out = outputs->at(i);
    if (out != nullptr) {
      StridedNumelCopyWithAxis<T>(dev_ctx, axis, out->data<T>(), out_stride,
                                  input.data<T>() + input_offset, in_stride,
                                  out_stride[axis]);
    }
    input_offset += out_stride[axis];
  }
}

Y
Yu Yang 已提交
136 137
}  // namespace operators
}  // namespace paddle