reduce_op.h 23.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
G
guosheng 已提交
2

L
Luo Tao 已提交
3 4 5
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
G
guosheng 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
G
guosheng 已提交
8

L
Luo Tao 已提交
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. */
G
guosheng 已提交
14 15 16

#pragma once

17
#include <algorithm>
18
#include <set>
19
#include <string>
W
whs 已提交
20
#include <vector>
21
#include "paddle/fluid/framework/data_type_transform.h"
22
#include "paddle/fluid/framework/tensor_util.h"
23
#include "paddle/fluid/operators/cast_op.h"
24
#include "paddle/fluid/operators/math/math_function.h"
W
Wu Yi 已提交
25
#include "paddle/fluid/operators/reduce_ops/reduce_op_function.h"
G
guosheng 已提交
26 27 28 29

namespace paddle {
namespace operators {

30 31
#define HANDLE_DIM(NDIM, RDIM)                                            \
  if (ndim == NDIM && rdim == RDIM) {                                     \
32
    ReduceFunctor<DeviceContext, OutT, NDIM, RDIM, Functor>(              \
33 34
        context.template device_context<DeviceContext>(), *input, output, \
        dims, keep_dim);                                                  \
W
whs 已提交
35 36
  }

37
using Tensor = framework::Tensor;
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141
using DDim = framework::DDim;

inline void GetShuffledDim(const DDim& src_dims, DDim* dst_dims,
                           const std::vector<int>& reduced_dims,
                           std::vector<int>* perm_axis) {
  // 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();
  for (size_t i = 0; i < reduce_size; ++i) {
    dst_dims->at(src_size - reduce_size + i) = src_dims[reduced_dims[i]];
    (*perm_axis)[src_size - reduce_size + i] = reduced_dims[i];
    src_dims_check[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 framework::ExecutionContext& context,
                      const Tensor* input, Tensor* shuffled_input,
                      const std::vector<int>& dims) {
  DDim shuffled_dims(input->dims());
  std::vector<int> perm_axis(input->dims().size());
  GetShuffledDim(input->dims(), &shuffled_dims, dims, &perm_axis);

  shuffled_input->Resize(shuffled_dims);
  shuffled_input->mutable_data<OutT>(context.GetPlace());

  math::TransposeNormal<DeviceContext, OutT> trans;
  trans(context.template device_context<DeviceContext>(), *input,
        shuffled_input, perm_axis);
}

inline void GetOriginDimFromShuffled(const DDim& src_dim,
                                     const std::vector<int>& dims,
                                     std::vector<int>* origin_dim) {
  DDim shuffled_dims(src_dim);
  size_t n = src_dim.size();
  std::vector<int> perm_axis(n);
  GetShuffledDim(src_dim, &shuffled_dims, dims, &perm_axis);
  for (size_t i = 0; i < n; ++i) {
    (*origin_dim)[perm_axis[i]] = i;
  }
}

template <typename DeviceContext, typename OutT, typename Functor>
void HandleLargeDim(const framework::ExecutionContext& context,
                    const Tensor* input, Tensor* output,
                    const std::vector<int>& dims, bool keep_dim) {
  //  shuffle the reduced dim to the end
  Tensor shuffled_input;
  GetShuffledInput<DeviceContext, OutT>(context, 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;
  shuffled_input.Resize({unreduced, reduced});
  DDim output_dim = output->dims();
  output->Resize({unreduced});
  ReduceFunctor<DeviceContext, OutT, 2, 1, Functor>(
      context.template device_context<DeviceContext>(), shuffled_input, output,
      {1}, keep_dim);
  output->Resize(output_dim);
}

template <typename DeviceContext, typename T, typename Functor>
void HandleLargeDimGrad(const framework::ExecutionContext& context,
                        const framework::Tensor* x,
                        const framework::Tensor* out,
                        const framework::Tensor* dout, framework::Tensor* dx,
                        const std::vector<int>& dims) {
  const int64_t unreduced = out->numel();
  const int64_t reduced = x->numel() / unreduced;
  DDim out_dim(out->dims());
  DDim x_dim(x->dims());
  // transpose and reshape X
  Tensor shuffled_x;
  GetShuffledInput<DeviceContext, T>(context, x, &shuffled_x, dims);
  DDim shuffled_dim = shuffled_x.dims();
  shuffled_x.Resize({unreduced, reduced});
  // reshape dX {unreduced, reduced}
  dx->Resize({unreduced, reduced});
  ReduceGradFunctor<DeviceContext, T, 2, Functor>(
      context.template device_context<DeviceContext>(), shuffled_x, *out, *dout,
      dx, {1});
  // transpose dX
  std::vector<int> origin_axis(x_dim.size());
  GetOriginDimFromShuffled(x_dim, dims, &origin_axis);
  Tensor dx_tmp;
  framework::TensorCopy(*dx, context.GetPlace(), &dx_tmp);
  dx_tmp.Resize(shuffled_dim);
  dx->Resize(x_dim);
  math::TransposeNormal<DeviceContext, T> trans;
  trans(context.template device_context<DeviceContext>(), dx_tmp, dx,
        origin_axis);
}
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176

template <typename DeviceContext, typename T, typename Functor>
struct ReduceKernelFunctor {
  const Tensor* input;
  Tensor* output;
  std::vector<int> dims;
  bool keep_dim;
  bool reduce_all;
  const framework::ExecutionContext& context;
  ReduceKernelFunctor(const Tensor* input, Tensor* output,
                      const std::vector<int>& dims, bool keep_dim,
                      bool reduce_all,
                      const framework::ExecutionContext& context)
      : input(input),
        output(output),
        dims(dims),
        keep_dim(keep_dim),
        reduce_all(reduce_all),
        context(context) {}

  template <typename OutT>
  void apply() const {
    output->mutable_data<OutT>(context.GetPlace());
    if (reduce_all) {
      // Flatten and reduce 1-D tensor
      auto x = EigenVector<OutT>::Flatten(*input);
      auto out = EigenScalar<OutT>::From(*output);
      auto& place =
          *context.template device_context<DeviceContext>().eigen_device();
      auto reduce_dim = Eigen::array<int, 1>({{0}});
      Functor functor;
      functor(place, &x, &out, reduce_dim);
    } else {
      int ndim = input->dims().size();
      int rdim = dims.size();
177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197
      if (ndim > 6) {
        HandleLargeDim<DeviceContext, OutT, Functor>(context, input, output,
                                                     dims, keep_dim);
      } else {
        HANDLE_DIM(6, 5);
        HANDLE_DIM(6, 4);
        HANDLE_DIM(6, 3);
        HANDLE_DIM(6, 2);
        HANDLE_DIM(6, 1);
        HANDLE_DIM(5, 4);
        HANDLE_DIM(5, 3);
        HANDLE_DIM(5, 2);
        HANDLE_DIM(5, 1);
        HANDLE_DIM(4, 3);
        HANDLE_DIM(4, 2);
        HANDLE_DIM(4, 1);
        HANDLE_DIM(3, 2);
        HANDLE_DIM(3, 1);
        HANDLE_DIM(2, 1);
        HANDLE_DIM(1, 1);
      }
198 199 200
    }
  }
};
Q
QI JUN 已提交
201
template <typename DeviceContext, typename T, typename Functor>
Y
Yu Yang 已提交
202
class ReduceKernel : public framework::OpKernel<T> {
203 204 205 206 207 208 209 210 211
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    bool reduce_all = context.Attr<bool>("reduce_all");
    auto* output = context.Output<Tensor>("Out");
    auto dims = context.Attr<std::vector<int>>("dim");
    bool keep_dim = context.Attr<bool>("keep_dim");
    int out_dtype = context.Attr<int>("out_dtype");
    framework::proto::VarType::Type cast_out_dtype;

212 213 214 215 216 217 218 219 220 221 222 223
    // The dims has full dim, set the reduce_all is True
    const auto& input_dim_size = context.Input<Tensor>("X")->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);

224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251
    if (out_dtype < 0) {
      auto* cast_input = context.Input<Tensor>("X");
      cast_out_dtype =
          static_cast<framework::proto::VarType::Type>(cast_input->type());
      framework::VisitDataType(
          cast_out_dtype,
          ReduceKernelFunctor<DeviceContext, T, Functor>(
              cast_input, output, dims, keep_dim, reduce_all, context));
    } else {
      Tensor tmp_tensor;
      cast_out_dtype = static_cast<framework::proto::VarType::Type>(out_dtype);
      auto* input = context.Input<Tensor>("X");

      tmp_tensor.Resize(input->dims());
      framework::VisitDataType(
          cast_out_dtype,
          CastOpFunctor<DeviceContext, T>(
              input, &tmp_tensor,
              context.template device_context<DeviceContext>()));
      framework::VisitDataType(
          cast_out_dtype,
          ReduceKernelFunctor<DeviceContext, T, Functor>(
              &tmp_tensor, output, dims, keep_dim, reduce_all, context));
    }
  }
};
template <typename DeviceContext, typename OutT, typename Functor>
class BoolReduceKernel : public framework::OpKernel<OutT> {
G
guosheng 已提交
252 253
 public:
  void Compute(const framework::ExecutionContext& context) const override {
254
    bool reduce_all = context.Attr<bool>("reduce_all");
255 256
    auto* input = context.Input<Tensor>("X");
    auto* output = context.Output<Tensor>("Out");
257
    output->mutable_data<OutT>(context.GetPlace());
258 259 260 261

    auto dims = context.Attr<std::vector<int>>("dim");
    bool keep_dim = context.Attr<bool>("keep_dim");

262 263 264 265 266 267 268 269 270 271 272 273
    // The dims has full dim, set the reduce_all is True
    const auto& input_dim_size = context.Input<Tensor>("X")->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);

274 275
    if (reduce_all) {
      // Flatten and reduce 1-D tensor
276 277
      auto x = EigenVector<OutT>::Flatten(*input);
      auto out = EigenScalar<OutT>::From(*output);
278 279 280 281
      auto& place =
          *context.template device_context<DeviceContext>().eigen_device();
      auto reduce_dim = Eigen::array<int, 1>({{0}});
      Functor functor;
282
      functor(place, &x, &out, reduce_dim);
283
    } else {
284 285
      int ndim = input->dims().size();
      int rdim = dims.size();
286
      // comments for accelerating compiling temporarily.
287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307
      if (ndim > 6) {
        HandleLargeDim<DeviceContext, OutT, Functor>(context, input, output,
                                                     dims, keep_dim);
      } else {
        HANDLE_DIM(6, 5);
        HANDLE_DIM(6, 4);
        HANDLE_DIM(6, 3);
        HANDLE_DIM(6, 2);
        HANDLE_DIM(6, 1);
        HANDLE_DIM(5, 4);
        HANDLE_DIM(5, 3);
        HANDLE_DIM(5, 2);
        HANDLE_DIM(5, 1);
        HANDLE_DIM(4, 3);
        HANDLE_DIM(4, 2);
        HANDLE_DIM(4, 1);
        HANDLE_DIM(3, 2);
        HANDLE_DIM(3, 1);
        HANDLE_DIM(2, 1);
        HANDLE_DIM(1, 1);
      }
G
guosheng 已提交
308 309 310
    }
  }
};
311 312
template <typename DeviceContext, typename T, typename Functor,
          bool kNoNeedBufferX = false, bool kNoNeedBufferY = false>
Y
Yu Yang 已提交
313
class ReduceGradKernel : public framework::OpKernel<T> {
G
guosheng 已提交
314
 public:
315 316
  void ComputeFromInput(const Tensor* input2,
                        const framework::ExecutionContext& context) const {
317
    bool reduce_all = context.Attr<bool>("reduce_all");
318 319 320
    auto dims = context.Attr<std::vector<int>>("dim");
    auto* input0 = context.Input<Tensor>("X");
    auto* input1 = context.Input<Tensor>("Out");
321

322 323 324
    auto* output = context.Output<Tensor>(framework::GradVarName("X"));
    output->mutable_data<T>(context.GetPlace());

325 326 327 328 329 330 331 332 333 334 335
    // The dims has full dim, set the reduce_all is True
    const auto& input_dim_size = context.Input<Tensor>("X")->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);
336 337 338 339 340 341 342 343 344 345 346
    // NOTE: EigenTensor::From() uses tensor->data()
    // if op has NoNeedBufferVarsInferer, the corresponding kNoNeedBufferX or
    // kNoNeedBufferY should set true
    // and use fake var that has same dims.
    if (kNoNeedBufferX) {
      input0 = output;
    }
    if (kNoNeedBufferY) {
      input1 = input2;
    }

L
lvmengsi 已提交
347 348 349 350
    // NOTE(dengkaipeng): Out is unnecessary in some reduce kernel and
    // not be set as Input in grad Maker, use Out_grad to replace here
    if (!input1) input1 = input2;

351 352
    if (reduce_all) {
      auto x = EigenVector<T>::Flatten(*input0);
353 354
      auto x_reduce = EigenVector<T>::Flatten(*input1);
      auto x_reduce_grad = EigenVector<T>::Flatten(*input2);
355 356 357 358 359 360
      auto x_grad = EigenVector<T>::Flatten(*output);
      auto& place =
          *context.template device_context<DeviceContext>().eigen_device();
      auto broadcast_dim =
          Eigen::array<int, 1>({{static_cast<int>(input0->numel())}});
      Functor functor;
361
      functor(place, &x, &x_reduce, &x_grad, &x_reduce_grad, broadcast_dim,
362 363
              broadcast_dim[0]);
    } else {
364
      int rank = input0->dims().size();
365 366
      switch (rank) {
        case 1:
367 368 369
          ReduceGradFunctor<DeviceContext, T, 1, Functor>(
              context.template device_context<DeviceContext>(), *input0,
              *input1, *input2, output, dims);
370 371
          break;
        case 2:
372 373 374
          ReduceGradFunctor<DeviceContext, T, 2, Functor>(
              context.template device_context<DeviceContext>(), *input0,
              *input1, *input2, output, dims);
375 376
          break;
        case 3:
377 378 379
          ReduceGradFunctor<DeviceContext, T, 3, Functor>(
              context.template device_context<DeviceContext>(), *input0,
              *input1, *input2, output, dims);
380 381
          break;
        case 4:
382 383 384
          ReduceGradFunctor<DeviceContext, T, 4, Functor>(
              context.template device_context<DeviceContext>(), *input0,
              *input1, *input2, output, dims);
385 386
          break;
        case 5:
387 388 389
          ReduceGradFunctor<DeviceContext, T, 5, Functor>(
              context.template device_context<DeviceContext>(), *input0,
              *input1, *input2, output, dims);
390 391
          break;
        case 6:
392 393 394
          ReduceGradFunctor<DeviceContext, T, 6, Functor>(
              context.template device_context<DeviceContext>(), *input0,
              *input1, *input2, output, dims);
395
          break;
396 397 398 399
        default:
          HandleLargeDimGrad<DeviceContext, T, Functor>(context, input0, input1,
                                                        input2, output, dims);
          break;
400
      }
G
guosheng 已提交
401 402
    }
  }
403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422

  void Compute(const framework::ExecutionContext& context) const override {
    int in_dtype = context.Attr<int>("in_dtype");
    if (in_dtype >= 0) {
      Tensor tmp_tensor;
      auto* pre_input = context.Input<Tensor>(framework::GradVarName("Out"));
      auto in_kernel_type =
          framework::OpKernelType(pre_input->type(), context.GetPlace());
      auto out_kernel_type = framework::OpKernelType(
          static_cast<framework::proto::VarType::Type>(in_dtype),
          context.GetPlace());
      framework::TransDataType(in_kernel_type, out_kernel_type, *pre_input,
                               &tmp_tensor);
      ComputeFromInput(&tmp_tensor, context);

    } else {
      auto* input2 = context.Input<Tensor>(framework::GradVarName("Out"));
      ComputeFromInput(input2, context);
    }
  }
423
};
G
guosheng 已提交
424

425 426 427
class ReduceOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
G
guosheng 已提交
428

429
  void InferShape(framework::InferShapeContext* ctx) const override {
430 431
    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "ReduceOp");
    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "ReduceOp");
432 433 434
    auto x_dims = ctx->GetInputDim("X");
    auto x_rank = x_dims.size();
    auto dims = ctx->Attrs().Get<std::vector<int>>("dim");
435 436 437 438 439 440
    PADDLE_ENFORCE_GT(dims.size(), 0,
                      platform::errors::InvalidArgument(
                          "The input dim dimensions of ReduceOp "
                          "should be greater than 0. But received the dim "
                          "dimesions of Reduce = %d.",
                          dims.size()));
441

442
    for (size_t i = 0; i < dims.size(); ++i) {
443
      PADDLE_ENFORCE_LT(dims[i], x_rank,
444 445 446 447 448
                        platform::errors::InvalidArgument(
                            "The reduce dim index %d should be in the "
                            "range [-dimension(X), dimension(X)] "
                            "which dimesion = %d. But received dim index = %d.",
                            i, x_rank, dims[i]));
449 450 451 452 453 454
      PADDLE_ENFORCE_GE(dims[i], -x_rank,
                        platform::errors::InvalidArgument(
                            "The reduce dim index %d should be in the "
                            "range [-dimension(X), dimension(X)] "
                            "which dimesion = %d. But received dim index = %d.",
                            i, x_rank, dims[i]));
455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480
      if (dims[i] < 0) dims[i] = x_rank + dims[i];
    }
    sort(dims.begin(), dims.end());
    bool reduce_all = ctx->Attrs().Get<bool>("reduce_all");
    bool keep_dim = ctx->Attrs().Get<bool>("keep_dim");
    if (reduce_all) {
      if (keep_dim)
        ctx->SetOutputDim(
            "Out", framework::make_ddim(std::vector<int64_t>(x_rank, 1)));
      else
        ctx->SetOutputDim("Out", {1});
    } else {
      auto dims_vector = vectorize(x_dims);
      if (keep_dim) {
        for (size_t i = 0; i < dims.size(); ++i) {
          dims_vector[dims[i]] = 1;
        }
      } else {
        const int kDelFlag = -2;
        for (size_t i = 0; i < dims.size(); ++i) {
          dims_vector[dims[i]] = kDelFlag;
        }
        dims_vector.erase(
            remove(dims_vector.begin(), dims_vector.end(), kDelFlag),
            dims_vector.end());
      }
481 482 483
      if (!keep_dim && dims_vector.size() == 0) {
        dims_vector.push_back(1);
      }
484 485
      auto out_dims = framework::make_ddim(dims_vector);
      ctx->SetOutputDim("Out", out_dims);
486
      if (dims.size() > 0 && dims[0] != 0) {
487 488 489 490 491 492 493
        // Only pass LoD when not reducing on the first dim.
        ctx->ShareLoD("X", /*->*/ "Out");
      }
    }
  }
};

G
Guo Sheng 已提交
494 495 496 497 498 499 500 501 502 503 504 505 506
class ReduceOpUseInputPlace : public ReduceOp {
 public:
  using ReduceOp::ReduceOp;

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    framework::OpKernelType kt = OperatorWithKernel::GetExpectedKernelType(ctx);
    kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
    return kt;
  }
};

507 508 509
class ReduceGradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
W
whs 已提交
510

511
  void InferShape(framework::InferShapeContext* ctx) const override {
512 513 514
    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "ReduceOp");
    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
                   "Out@GRAD", "ReduceOp");
515 516 517
    auto x_dims = ctx->GetInputDim("X");
    auto x_rank = x_dims.size();
    auto dims = ctx->Attrs().Get<std::vector<int>>("dim");
W
whs 已提交
518
    for (size_t i = 0; i < dims.size(); ++i) {
519
      PADDLE_ENFORCE_LT(dims[i], x_rank,
520 521 522 523 524
                        platform::errors::InvalidArgument(
                            "The reduce dim index %d should be in the "
                            "range [-dimension(X), dimension(X)], "
                            "which dimesion = %d. But received dim index = %d.",
                            i, x_rank, dims[i]));
W
whs 已提交
525
      if (dims[i] < 0) dims[i] = x_rank + dims[i];
526 527 528 529 530 531
    }
    sort(dims.begin(), dims.end());
    auto x_grad_name = framework::GradVarName("X");
    if (ctx->HasOutput(x_grad_name)) {
      ctx->SetOutputDim(x_grad_name, x_dims);
      ctx->ShareLoD("X", /*->*/ x_grad_name);
W
whs 已提交
532
    }
533
  }
534 535 536 537

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
538 539 540 541 542 543
    int in_dtype = ctx.Attr<int>("in_dtype");
    if (in_dtype >= 0) {
      return framework::OpKernelType(
          static_cast<framework::proto::VarType::Type>(in_dtype),
          ctx.GetPlace());
    }
544 545 546
    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
                                       ctx, framework::GradVarName("Out")),
                                   ctx.GetPlace());
547
  }
548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571
};

class ReduceOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() final {
    AddInput("X",
             "(Tensor) The input tensor. Tensors with rank at most 6 are "
             "supported.");
    AddOutput("Out", "(Tensor) The result tensor.");
    AddAttr<std::vector<int>>(
        "dim",
        "(list<int>, default {0}) The dimensions to reduce. "
        "Must be in the range [-rank(input), rank(input)). "
        "If `dim[i] < 0`, the dims[i] to reduce is `rank + dims[i]`. "
        "Note that reducing on the first dim will make the LoD info lost.")
        .SetDefault({0});
    AddAttr<bool>("keep_dim",
                  "(bool, default false) "
                  "If true, retain the reduced dimension with length 1.")
        .SetDefault(false);
    AddAttr<bool>("reduce_all",
                  "(bool, default false) "
                  "If true, output a scalar reduced along all dimensions.")
        .SetDefault(false);
572 573 574 575 576 577 578 579 580 581
    AddAttr<int>("in_dtype",
                 "(int, default -1)"
                 "The dtype of input, default value is -1, the user could not "
                 "set this value.")
        .SetDefault(-1);
    AddAttr<int>(
        "out_dtype",
        "(int, default -1)"
        "The dtype of output, default value is -1, the dtype is same as intput")
        .SetDefault(-1);
582 583
    AddComment(string::Sprintf(R"DOC(
%s Operator.
W
whs 已提交
584

585 586 587
This operator computes the %s of input tensor along the given dimension.
The result tensor has 1 fewer dimension than the input unless keep_dim is true.
If reduce_all is true, just reduce along all dimensions and output a scalar.
W
whs 已提交
588

589 590
)DOC",
                               GetOpType(), GetName()));
G
guosheng 已提交
591
  }
592 593 594 595

 protected:
  virtual std::string GetName() const = 0;
  virtual std::string GetOpType() const = 0;
G
guosheng 已提交
596 597 598 599
};

}  // namespace operators
}  // namespace paddle
600

601 602
namespace ops = paddle::operators;

H
hong 已提交
603 604 605 606 607 608 609 610 611 612 613 614 615 616
#define REGISTER_REDUCE_OP(op_name)                                           \
  class __##op_name##Maker__ : public ops::ReduceOpMaker {                    \
   protected:                                                                 \
    virtual std::string GetName() const { return #op_name; }                  \
    virtual std::string GetOpType() const { return "Reduce " #op_name; }      \
  };                                                                          \
  REGISTER_OPERATOR(                                                          \
      op_name, ops::ReduceOp, __##op_name##Maker__,                           \
      paddle::framework::DefaultGradOpMaker<paddle::framework::OpDesc, true>, \
      paddle::framework::DefaultGradOpMaker<paddle::imperative::OpBase,       \
                                            true>);                           \
  REGISTER_OPERATOR(op_name##_grad, ops::ReduceGradOp)

#define REGISTER_REDUCE_OP_WITHOUT_GRAD(op_name, ...)                    \
617 618 619 620 621
  class __##op_name##Maker__ : public ops::ReduceOpMaker {               \
   protected:                                                            \
    virtual std::string GetName() const { return #op_name; }             \
    virtual std::string GetOpType() const { return "Reduce " #op_name; } \
  };                                                                     \
H
hong 已提交
622 623 624 625
  REGISTER_OPERATOR(                                                     \
      op_name, ops::ReduceOp##__VA_ARGS__, __##op_name##Maker__,         \
      paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,    \
      paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>);