FastDeploy/fastdeploy/function/reduce.cc

// Copyright (c) 2022 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.

#include "fastdeploy/function/reduce.h"

#include <limits>
#include <set>

#include "fastdeploy/function/eigen.h"
#include "fastdeploy/function/reduce_functor.h"
#include "fastdeploy/function/transpose.h"
#include "fastdeploy/utils/utils.h"

namespace fastdeploy {
namespace function {
template <typename T, size_t D, size_t R_D, typename Functor>
void ReduceFunctor(const FDTensor& input, FDTensor* output,
                   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;

  auto out_dims = input.shape;
  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];
    out_dims[dims_ref[i]] = 1;
  }
  auto origin_output_dims = out_dims;
  output->Allocate(origin_output_dims, TypeToDataType<T>::dtype);
  // construct the squeezed output tensor
  if (x_rank > 1) {
    const int kDelFlag = -2;
    for (size_t i = 0; i < dims_ref.size(); ++i) {
      out_dims[dims_ref[i]] = kDelFlag;
    }
    out_dims.erase(remove(out_dims.begin(), out_dims.end(), kDelFlag),
                   out_dims.end());
  }

  auto& place = *EigenDeviceWrapper::GetInstance()->GetDevice();
  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);
    if (!keep_dim) {
      output->shape = std::move(out_dims);
    }
  }
}

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

inline void GetShuffledDim(const std::vector<int64_t>& src_dims,
                           std::vector<int64_t>* dst_dims,
                           const std::vector<int64_t>& reduced_dims,
                           std::vector<int64_t>* 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();
  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 OutT>
void GetShuffledInput(const FDTensor& input, FDTensor* shuffled_input,
                      const std::vector<int64_t>& dims) {
  auto shuffled_dims = input.shape;
  std::vector<int64_t> perm_axis(input.shape.size());
  GetShuffledDim(input.shape, &shuffled_dims, dims, &perm_axis);

  shuffled_input->Allocate(shuffled_dims, input.dtype);
  Transpose(input, shuffled_input, perm_axis);
}

//////////////// HandleLargeDim
template <typename OutT, typename Functor>
void HandleLargeDim(const FDTensor& input, FDTensor* output,
                    const std::vector<int64_t>& dims, bool keep_dim) {
  auto out_dims = input.shape;
  std::vector<int64_t> dims_ref = dims;
  auto x_rank = input.shape.size();
  for (size_t i = 0; i < dims_ref.size(); ++i) {
    if (dims_ref[i] < 0) dims_ref[i] = x_rank + dims_ref[i];
    out_dims[dims_ref[i]] = 1;
  }
  if (!keep_dim) {
    const int kDelFlag = -2;
    for (size_t i = 0; i < dims_ref.size(); ++i) {
      out_dims[dims_ref[i]] = kDelFlag;
    }
    out_dims.erase(remove(out_dims.begin(), out_dims.end(), kDelFlag),
                   out_dims.end());
  }
  output->Allocate(out_dims, TypeToDataType<OutT>::dtype);
  //  shuffle the reduced dim to the end
  FDTensor shuffled_input;
  GetShuffledInput<OutT>(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.Allocate({unreduced, reduced}, TypeToDataType<OutT>::dtype);

  output->shape = {unreduced};
  ReduceFunctor<OutT, 2, 1, Functor>(shuffled_input, output, {1}, keep_dim);
  output->shape = out_dims;
}

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

template <typename OutT, typename Functor>
void ReduceKernelImpl(const FDTensor& input, FDTensor* output,
                      const std::vector<int64_t>& dims, bool keep_dim,
                      bool reduce_all) {
  output->Allocate({1}, TypeToDataType<OutT>::dtype);
  const auto& dev = *EigenDeviceWrapper::GetInstance()->GetDevice();
  if (reduce_all) {
    // Flatten and reduce 1-D tensor
    auto x = EigenVector<OutT>::Flatten(input);
    auto out = EigenScalar<OutT>::From(*output);
    auto reduce_dim = Eigen::array<int, 1>({{0}});

    Functor functor;
    functor(dev, &x, &out, reduce_dim);
  } else {
    int ndim = input.shape.size();
    int rdim = dims.size();
    if (ndim > 4) {
      HandleLargeDim<OutT, Functor>(input, output, dims, keep_dim);
    } else {
      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);
    }
  }
}

template <typename OutT, typename Functor>
void BoolReduceKernel(const FDTensor& input, FDTensor* output,
                      const std::vector<int64_t>& dims, bool keep_dim,
                      bool reduce_all) {
  // The dims has full dim, set the reduce_all is True
  const auto& input_dim_size = input.shape.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<bool, Functor>(input, output, dims, keep_dim, reduce_all);
}

template <typename Functor>
void Reduce(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
            bool keep_dim, bool reduce_all) {
  // If the dims has full dim, set the reduce_all is True
  const int& input_dim_size = x.shape.size();
  std::set<int> dims_set(dims.begin(), dims.end());
  bool full_dim = true;
  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()) {
      full_dim = false;
      break;
    }
  }
  reduce_all = (reduce_all || full_dim);

  FD_VISIT_INT_FLOAT_TYPES(x.dtype, "ReduceKernelImpl", ([&] {
                             ReduceKernelImpl<data_t, Functor>(
                                 x, out, dims, keep_dim, reduce_all);
                           }));
}

enum ArgMinMaxType { kArgMin, kArgMax };

template <typename T, typename Tout, int64_t Rank, ArgMinMaxType argMinMaxValue>
struct ArgMinMaxFunctor {};

#define DECLARE_ARG_MIN_MAX_FUNCTOR(eigen_op_type, enum_argminmax_value) \
  template <typename T, typename Tout, int64_t Rank>                     \
  struct ArgMinMaxFunctor<T, Tout, Rank, enum_argminmax_value> {         \
    void operator()(const FDTensor& in, FDTensor* out,                   \
                    const std::vector<int64_t>& x_dims, int64_t axis,    \
                    bool keepdims, bool flatten) {                       \
      const auto& dev = *EigenDeviceWrapper::GetInstance()->GetDevice(); \
      auto in_eigen = EigenTensor<T, Rank>::From(in, x_dims);            \
      if (keepdims) {                                                    \
        if (!flatten) {                                                  \
          auto out_eigen = EigenTensor<Tout, Rank>::From(*out);          \
          out_eigen.device(dev) =                                        \
              in_eigen.eigen_op_type(axis).template cast<Tout>();        \
        } else {                                                         \
          auto out_eigen = EigenScalar<Tout>::From(*out);                \
          out_eigen.device(dev) =                                        \
              in_eigen.eigen_op_type(axis).template cast<Tout>();        \
        }                                                                \
      } else {                                                           \
        auto out_eigen = EigenTensor<Tout, Rank - 1>::From(*out);        \
        out_eigen.device(dev) =                                          \
            in_eigen.eigen_op_type(axis).template cast<Tout>();          \
      }                                                                  \
    }                                                                    \
  }

DECLARE_ARG_MIN_MAX_FUNCTOR(argmin, ArgMinMaxType::kArgMin);
DECLARE_ARG_MIN_MAX_FUNCTOR(argmax, ArgMinMaxType::kArgMax);

template <typename T, typename Tout, ArgMinMaxType EnumArgMinMaxValue>
void ArgMinMaxKernel(const FDTensor& x, FDTensor* out, int64_t axis,
                     bool keepdims, bool flatten) {
  bool new_keepdims = keepdims | flatten;
  // if flatten, will construct the new dims for the cacluate
  std::vector<int64_t> x_dims;
  int new_axis = axis;
  if (flatten) {
    x_dims = {x.Numel()};
    // if flatten, the axis just as 0
    new_axis = 0;
  } else {
    x_dims = x.shape;
    if (axis < 0) new_axis = axis + x_dims.size();
  }
#define CALL_ARG_MINMAX_FUNCTOR(rank)                                \
  ArgMinMaxFunctor<T, Tout, rank, EnumArgMinMaxValue> functor##rank; \
  functor##rank(x, out, x_dims, new_axis, new_keepdims, flatten)

  switch (x_dims.size()) {
    case 1:
      CALL_ARG_MINMAX_FUNCTOR(1);
      break;
    case 2:
      CALL_ARG_MINMAX_FUNCTOR(2);
      break;
    case 3:
      CALL_ARG_MINMAX_FUNCTOR(3);
      break;
    case 4:
      CALL_ARG_MINMAX_FUNCTOR(4);
      break;
    case 5:
      CALL_ARG_MINMAX_FUNCTOR(5);
      break;
    case 6:
      CALL_ARG_MINMAX_FUNCTOR(6);
      break;
    default:
      FDASSERT(x_dims.size() <= 6,
               "%s operator doesn't supports tensors whose ranks are greater "
               "than 6.",
               (EnumArgMinMaxValue == kArgMin ? "argmin" : "argmax"));
      break;
#undef CALL_ARG_MINMAX_FUNCTOR
  }
}

template <typename T, ArgMinMaxType EnumArgMinMaxValue>
void ArgMinMax(const FDTensor& x, FDTensor* out, int64_t axis,
               FDDataType output_dtype, bool keepdims, bool flatten) {
  const auto& x_dims = x.shape;
  int64_t x_rank = x_dims.size();
  FDASSERT(axis >= -x_rank,
           "'axis'(%lld) must be greater than or equal to -Rank(X)(%lld).",
           axis, -x_rank);
  FDASSERT(axis < x_rank,
           "'axis'(%lld) must be less than or equal to Rank(X)(%lld).", axis,
           x_rank);
  FDASSERT(output_dtype == FDDataType::INT32 || FDDataType::INT64,
           "The attribute of dtype in argmin/argmax must be [%s] or [%s], but "
           "received [%s].",
           Str(FDDataType::INT32).c_str(), Str(FDDataType::INT64).c_str(),
           Str(output_dtype).c_str());
  if (axis < 0) axis += x_rank;
  if (output_dtype == FDDataType::INT32) {
    int64_t all_element_num = 0;
    if (flatten) {
      all_element_num = x.Numel();

    } else {
      all_element_num = x_dims[axis];
    }
    FDASSERT(all_element_num <= (std::numeric_limits<int>::max)(),
             "The element num of the argmin/argmax input at axis is "
             "%lld, is larger than int32 maximum value:%d, you must "
             "set the dtype of argmin/argmax to 'int64'.",
             all_element_num, (std::numeric_limits<int>::max)());
  }
  std::vector<int64_t> vec;
  if (flatten) {
    vec.emplace_back(static_cast<int64_t>(1));
  } else {
    for (int64_t i = 0; i < axis; i++) vec.emplace_back(x_dims[i]);
    if (keepdims) {
      vec.emplace_back(static_cast<int64_t>(1));
    }
    for (int64_t i = axis + 1; i < x_rank; i++) vec.emplace_back(x_dims[i]);
  }
  out->Allocate(vec, output_dtype);

  FD_VISIT_INT_TYPES(output_dtype, "ArgMinMaxKernel", ([&] {
                       ArgMinMaxKernel<T, data_t, EnumArgMinMaxValue>(
                           x, out, axis, keepdims, flatten);
                     }));
}

void Max(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
         bool keep_dim, bool reduce_all) {
  Reduce<MaxFunctor>(x, out, dims, keep_dim, reduce_all);
}

void Min(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
         bool keep_dim, bool reduce_all) {
  Reduce<MinFunctor>(x, out, dims, keep_dim, reduce_all);
}

void Sum(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
         bool keep_dim, bool reduce_all) {
  Reduce<SumFunctor>(x, out, dims, keep_dim, reduce_all);
}

void All(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
         bool keep_dim, bool reduce_all) {
  BoolReduceKernel<bool, AllFunctor>(x, out, dims, keep_dim, reduce_all);
}

void Any(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
         bool keep_dim, bool reduce_all) {
  BoolReduceKernel<bool, AnyFunctor>(x, out, dims, keep_dim, reduce_all);
}

void Mean(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
          bool keep_dim, bool reduce_all) {
  Reduce<MeanFunctor>(x, out, dims, keep_dim, reduce_all);
}

void Prod(const FDTensor& x, FDTensor* out, const std::vector<int64_t>& dims,
          bool keep_dim, bool reduce_all) {
  Reduce<ProdFunctor>(x, out, dims, keep_dim, reduce_all);
}

void ArgMax(const FDTensor& x, FDTensor* out, int64_t axis,
            FDDataType output_dtype, bool keep_dim, bool flatten) {
  FD_VISIT_INT_FLOAT_TYPES(x.dtype, "ArgMaxKernel", ([&] {
                             ArgMinMax<data_t, kArgMax>(
                                 x, out, axis, output_dtype, keep_dim, flatten);
                           }));
}

void ArgMin(const FDTensor& x, FDTensor* out, int64_t axis,
            FDDataType output_dtype, bool keep_dim, bool flatten) {
  FD_VISIT_INT_FLOAT_TYPES(x.dtype, "ArgMaxKernel", ([&] {
                             ArgMinMax<data_t, kArgMin>(
                                 x, out, axis, output_dtype, keep_dim, flatten);
                           }));
}

}  // namespace function
}  // namespace fastdeploy