reduce.h 9.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
//
// 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
16 17 18

#include <set>

19
#include "paddle/phi/backends/cpu/cpu_context.h"
20
#include "paddle/phi/core/visit_type.h"
21 22 23 24 25
#include "paddle/phi/kernels/cast_kernel.h"

#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
#include "paddle/phi/kernels/funcs/math_function.h"
26 27
// See Note [ Why still include the fluid headers? ]
#include "paddle/fluid/operators/eigen/eigen_function.h"
28
namespace phi {
29

30 31 32 33 34 35
template <typename DeviceContext,
          typename T,
          size_t D,
          size_t R_D,
          typename Functor>
void ReduceFunctor(const DeviceContext& context,
36 37
                   const phi::DenseTensor& input,
                   phi::DenseTensor* output,
38 39 40 41 42 43 44 45 46 47 48 49 50 51
                   const std::vector<int64_t>& dims,
                   bool keep_dim) {
  auto x = EigenTensor<T, D>::From(input);
  auto x_rank = static_cast<int>(x.dimensions().size());
  auto reduce_dim = Eigen::array<int, R_D>();
  std::vector<int64_t> dims_ref = dims;
  for (size_t i = 0; i < dims_ref.size(); ++i) {
    if (dims_ref[i] < 0) dims_ref[i] = x_rank + dims_ref[i];
    reduce_dim[i] = dims_ref[i];
  }
  // construct the squeezed output tensor
  DDim out_dims = output->dims();
  if (keep_dim && x_rank > 1) {
    const int kDelFlag = -2;
52
    auto dims_vector = phi::vectorize(out_dims);
53 54 55 56 57
    for (size_t i = 0; i < dims_ref.size(); ++i) {
      dims_vector[dims_ref[i]] = kDelFlag;
    }
    dims_vector.erase(remove(dims_vector.begin(), dims_vector.end(), kDelFlag),
                      dims_vector.end());
58
    out_dims = phi::make_ddim(dims_vector);
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81
  }
  auto& place = *context.eigen_device();
  Functor functor;

  if (D == 1) {
    auto out = EigenScalar<T>::From(*output);
    functor(place, &x, &out, reduce_dim);
  } else {
    auto out = EigenTensor<T, (D - R_D)>::From(*output, out_dims);
    functor(place, &x, &out, reduce_dim);
  }
}

#define HANDLE_REDUCE_DIM(NDIM, RDIM)                        \
  if (ndim == NDIM && rdim == RDIM) {                        \
    ReduceFunctor<DeviceContext, OutT, NDIM, RDIM, Functor>( \
        dev_ctx, input, output, dims, keep_dim);             \
  }
//////////////// HandleLargeDim

inline void GetShuffledDim(const DDim& src_dims,
                           DDim* dst_dims,
                           const std::vector<int64_t>& reduced_dims,
82
                           std::vector<int>* perm_axis) {
83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
  // check if it's a reduced dim
  std::vector<bool> src_dims_check(src_dims.size(), false);
  size_t src_size = src_dims.size();
  size_t reduce_size = reduced_dims.size();
  std::vector<int64_t> regular_reduced_dims = reduced_dims;
  for (size_t i = 0; i < regular_reduced_dims.size(); i++) {
    if (regular_reduced_dims[i] < 0) {
      regular_reduced_dims[i] = src_size + regular_reduced_dims[i];
    }
  }

  for (size_t i = 0; i < reduce_size; ++i) {
    dst_dims->at(src_size - reduce_size + i) =
        src_dims[regular_reduced_dims[i]];
    (*perm_axis)[src_size - reduce_size + i] = regular_reduced_dims[i];
    src_dims_check[regular_reduced_dims[i]] = true;
  }

  size_t offset = 0;
  for (size_t i = 0; i < src_dims_check.size(); ++i) {
    bool is_reduced = src_dims_check[i];
    if (!is_reduced) {
      (*perm_axis)[offset] = i;
      dst_dims->at(offset++) = src_dims[i];
    }
  }
}

template <typename DeviceContext, typename OutT>
void GetShuffledInput(const DeviceContext& dev_ctx,
113 114
                      const phi::DenseTensor& input,
                      phi::DenseTensor* shuffled_input,
115 116
                      const std::vector<int64_t>& dims) {
  DDim shuffled_dims(input.dims());
117
  std::vector<int> perm_axis(input.dims().size());
118 119
  GetShuffledDim(input.dims(), &shuffled_dims, dims, &perm_axis);

Z
zyfncg 已提交
120
  shuffled_input->Resize(shuffled_dims);
121
  dev_ctx.template Alloc<OutT>(shuffled_input);
122

123
  phi::funcs::TransposeNormal<DeviceContext, OutT> trans;
124 125 126 127 128
  trans(dev_ctx, input, shuffled_input, perm_axis);
}

template <typename DeviceContext, typename OutT, typename Functor>
void HandleLargeDim(const DeviceContext& dev_ctx,
129 130
                    const phi::DenseTensor& input,
                    phi::DenseTensor* output,
131 132 133
                    const std::vector<int64_t>& dims,
                    bool keep_dim) {
  //  shuffle the reduced dim to the end
Z
zyfncg 已提交
134
  phi::DenseTensor shuffled_input;
135 136 137 138 139
  GetShuffledInput<DeviceContext, OutT>(dev_ctx, input, &shuffled_input, dims);

  // transpose to 2D tensor whose shape is {unreduced, reduced}.
  const int64_t unreduced = output->numel();
  const int64_t reduced = shuffled_input.numel() / unreduced;
140
  shuffled_input.ResizeAndAllocate({unreduced, reduced});
141
  DDim output_dim = output->dims();
142
  output->ResizeAndAllocate({unreduced});
143 144
  ReduceFunctor<DeviceContext, OutT, 2, 1, Functor>(
      dev_ctx, shuffled_input, output, {1}, keep_dim);
145
  output->ResizeAndAllocate(output_dim);
146 147 148 149 150 151
}

////////////// ReduceKernel

template <typename DeviceContext, typename T, typename OutT, typename Functor>
void ReduceKernelImpl(const DeviceContext& dev_ctx,
152 153
                      const phi::DenseTensor& input,
                      phi::DenseTensor* output,
154 155 156
                      const std::vector<int64_t>& dims,
                      bool keep_dim,
                      bool reduce_all) {
157
  dev_ctx.template Alloc<OutT>(output);
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195

  if (reduce_all) {
    // Flatten and reduce 1-D tensor
    auto x = EigenVector<OutT>::Flatten(input);
    auto out = EigenScalar<OutT>::From(*output);
    auto& dev = *dev_ctx.eigen_device();
    auto reduce_dim = Eigen::array<int, 1>({{0}});

    Functor functor;
    functor(dev, &x, &out, reduce_dim);
  } else {
    int ndim = input.dims().size();
    int rdim = dims.size();
    if (ndim > 6) {
      HandleLargeDim<DeviceContext, OutT, Functor>(
          dev_ctx, input, output, dims, keep_dim);

    } else {
      HANDLE_REDUCE_DIM(6, 5);
      HANDLE_REDUCE_DIM(6, 4);
      HANDLE_REDUCE_DIM(6, 3);
      HANDLE_REDUCE_DIM(6, 2);
      HANDLE_REDUCE_DIM(6, 1);
      HANDLE_REDUCE_DIM(5, 4);
      HANDLE_REDUCE_DIM(5, 3);
      HANDLE_REDUCE_DIM(5, 2);
      HANDLE_REDUCE_DIM(5, 1);
      HANDLE_REDUCE_DIM(4, 3);
      HANDLE_REDUCE_DIM(4, 2);
      HANDLE_REDUCE_DIM(4, 1);
      HANDLE_REDUCE_DIM(3, 2);
      HANDLE_REDUCE_DIM(3, 1);
      HANDLE_REDUCE_DIM(2, 1);
      HANDLE_REDUCE_DIM(1, 1);
    }
  }
}

196 197 198 199 200 201 202 203 204
template <typename DeviceContext, typename T, typename Functor>
void Reduce(const DeviceContext& dev_ctx,
            const DenseTensor& x,
            bool reduce_all,
            const std::vector<int64_t>& dims,
            bool keep_dim,
            DataType out_dtype,
            DenseTensor* out) {
  // If the dims has full dim, set the reduce_all is True
205
  const int& input_dim_size = x.dims().size();
206 207
  std::set<int> dims_set(dims.begin(), dims.end());
  bool full_dim = true;
208 209 210
  for (int i = 0; i < input_dim_size; ++i) {
    if (dims_set.find(i) == dims_set.end() &&
        dims_set.find(i - input_dim_size) == dims_set.end()) {
211 212 213 214 215 216 217
      full_dim = false;
      break;
    }
  }
  reduce_all = (reduce_all || full_dim);

  // no need to cast dtype
218
  if (out_dtype == phi::DataType::UNDEFINED || out_dtype == x.dtype()) {
219 220
    // do reduce sum
    PD_VISIT_ALL_TYPES(
221
        x.dtype(), "ReduceKernelImpl", ([&] {
222
          phi::ReduceKernelImpl<DeviceContext, T, data_t, Functor>(
223 224 225
              dev_ctx, x, out, dims, keep_dim, reduce_all);
        }));
  } else {
226
    // cast x tensor to out_dtype
227
    auto tmp_tensor = phi::Cast<T, DeviceContext>(dev_ctx, x, out_dtype);
228 229 230 231

    // do reduce sum
    PD_VISIT_ALL_TYPES(
        out_dtype, "ReduceKernelImpl", ([&] {
232
          phi::ReduceKernelImpl<DeviceContext, T, data_t, Functor>(
233 234 235 236 237
              dev_ctx, tmp_tensor, out, dims, keep_dim, reduce_all);
        }));
  }
}

238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
template <typename DeviceContext, typename OutT, typename Functor>
void BoolReduceKernel(const DeviceContext& dev_ctx,
                      const phi::DenseTensor& input,
                      const std::vector<int64_t>& dims,
                      bool keep_dim,
                      bool reduce_all,
                      phi::DenseTensor* output) {
  dev_ctx.template Alloc<OutT>(output);

  // The dims has full dim, set the reduce_all is True
  const auto& input_dim_size = input.dims().size();
  std::set<int> dims_set(dims.begin(), dims.end());
  bool full_dim = true;
  for (auto i = 0; i < input_dim_size; i++) {
    if (dims_set.find(i) == dims_set.end()) {
      full_dim = false;
      break;
    }
  }
  reduce_all = (reduce_all || full_dim);

  ReduceKernelImpl<DeviceContext, bool, OutT, Functor>(
      dev_ctx, input, output, dims, keep_dim, reduce_all);
}

263
}  // namespace phi