/* Copyright (c) 2016 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 "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/core/utils/visit_place.h"

#ifdef PADDLE_WITH_MKLML
#include "paddle/phi/backends/dynload/mklml.h"
#endif

#ifdef PADDLE_USE_OPENBLAS
#include <cblas.h>
#elif PADDLE_USE_ACCELERATE
#include <Accelerate/Accelerate.h>
#endif

#include <memory>
#include <utility>
#include <vector>

#include "paddle/phi/backends/context_pool.h"

#ifdef PADDLE_WITH_XPU
#include "paddle/phi/backends/xpu/xpu_context.h"
#endif

#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/common/data_type.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
#include "paddle/phi/kernels/funcs/math_function_impl.h"
#ifdef PADDLE_WITH_CUSTOM_DEVICE
#include "paddle/phi/api/lib/kernel_dispatch.h"
#include "paddle/phi/core/kernel_factory.h"
#endif

namespace phi::funcs {

template struct SetConstant<CPUContext, phi::float8_e4m3fn>;
template struct SetConstant<CPUContext, phi::float8_e5m2>;
template struct SetConstant<CPUContext, phi::float16>;
template struct SetConstant<CPUContext, phi::bfloat16>;
template struct SetConstant<CPUContext, float>;
template struct SetConstant<CPUContext, double>;
template struct SetConstant<CPUContext, int16_t>;
template struct SetConstant<CPUContext, int>;
template struct SetConstant<CPUContext, int64_t>;
template struct SetConstant<CPUContext, bool>;
template struct SetConstant<CPUContext, uint8_t>;
template struct SetConstant<CPUContext, uint16_t>;
template struct SetConstant<CPUContext, uint32_t>;
template struct SetConstant<CPUContext, uint64_t>;
template struct SetConstant<CPUContext, int8_t>;
template struct SetConstant<CPUContext, phi::complex64>;
template struct SetConstant<CPUContext, phi::complex128>;

#ifdef PADDLE_WITH_XPU
template struct SetConstant<XPUContext, phi::float16>;
template struct SetConstant<XPUContext, phi::bfloat16>;
template struct SetConstant<XPUContext, float>;
template struct SetConstant<XPUContext, double>;
template struct SetConstant<XPUContext, uint8_t>;
template struct SetConstant<XPUContext, int8_t>;
template struct SetConstant<XPUContext, int16_t>;
template struct SetConstant<XPUContext, int>;
template struct SetConstant<XPUContext, int64_t>;
template struct SetConstant<XPUContext, bool>;
template struct SetConstant<XPUContext, phi::complex64>;
template struct SetConstant<XPUContext, phi::complex128>;
#endif  // PADDLE_WITH_XPU

#define DEFINE_CPU_TRANS(RANK)                                                \
  template struct PADDLE_API Transpose<CPUContext, phi::float16, RANK>;       \
  template struct PADDLE_API Transpose<CPUContext, phi::bfloat16, RANK>;      \
  template struct PADDLE_API Transpose<CPUContext, phi::float8_e4m3fn, RANK>; \
  template struct PADDLE_API Transpose<CPUContext, phi::float8_e5m2, RANK>;   \
  template struct PADDLE_API Transpose<CPUContext, float, RANK>;              \
  template struct PADDLE_API Transpose<CPUContext, double, RANK>;             \
  template struct PADDLE_API Transpose<CPUContext, int, RANK>;                \
  template struct PADDLE_API Transpose<CPUContext, int64_t, RANK>;            \
  template struct PADDLE_API Transpose<CPUContext, bool, RANK>;               \
  template struct PADDLE_API Transpose<CPUContext, int16_t, RANK>;            \
  template struct PADDLE_API Transpose<CPUContext, uint8_t, RANK>;            \
  template struct PADDLE_API Transpose<CPUContext, uint16_t, RANK>;           \
  template struct PADDLE_API Transpose<CPUContext, uint32_t, RANK>;           \
  template struct PADDLE_API Transpose<CPUContext, uint64_t, RANK>;           \
  template struct PADDLE_API Transpose<CPUContext, int8_t, RANK>;             \
  template struct PADDLE_API Transpose<CPUContext, phi::complex64, RANK>;     \
  template struct PADDLE_API Transpose<CPUContext, phi::complex128, RANK>;

DEFINE_CPU_TRANS(1);
DEFINE_CPU_TRANS(2);
DEFINE_CPU_TRANS(3);
DEFINE_CPU_TRANS(4);
DEFINE_CPU_TRANS(5);
DEFINE_CPU_TRANS(6);

#ifdef PADDLE_WITH_XPU
#define DEFINE_XPU_TRANS(RANK)                                     \
  template struct PADDLE_API Transpose<XPUContext, bool, RANK>;    \
  template struct PADDLE_API Transpose<XPUContext, float, RANK>;   \
  template struct PADDLE_API Transpose<XPUContext, int, RANK>;     \
  template struct PADDLE_API Transpose<XPUContext, int64_t, RANK>; \
  template struct PADDLE_API Transpose<XPUContext, phi::complex64, RANK>;

DEFINE_XPU_TRANS(1);
DEFINE_XPU_TRANS(2);
DEFINE_XPU_TRANS(3);
DEFINE_XPU_TRANS(4);
DEFINE_XPU_TRANS(5);
DEFINE_XPU_TRANS(6);
#endif  // PADDLE_WITH_XPU

template <typename DeviceContext, typename T>
void TransposeNormal<DeviceContext, T>::operator()(
    const DeviceContext& dev_ctx UNUSED,
    const DenseTensor& in,
    DenseTensor* out,
    const std::vector<int>& axis) {
  const int rank = static_cast<const int>(axis.size());
  auto in_stride = common::stride(in.dims());
  auto out_stride = common::stride(out->dims());
  const T* in_ptr = in.data<T>();
  T* out_ptr = out->data<T>();

  auto transpose_helper = [&](int64_t beg, int64_t end) {
    for (int64_t out_idx = beg; out_idx < end; ++out_idx) {
      int64_t in_idx = 0;
      int64_t tmp_idx = out_idx;
      // calculate the input index
      for (int i = 0; i < rank; ++i) {
        const int64_t coordinate = tmp_idx / out_stride[i];
        tmp_idx -= coordinate * out_stride[i];
        in_idx += coordinate * in_stride[axis[i]];
      }
      out_ptr[out_idx] = in_ptr[in_idx];
    }
  };
  transpose_helper(0, out->numel());
}

// define transpose normal
#define DEFINE_CPU_TRANS_NORMAL(TYPE) \
  template struct TransposeNormal<CPUContext, TYPE>
DEFINE_CPU_TRANS_NORMAL(phi::float8_e4m3fn);
DEFINE_CPU_TRANS_NORMAL(phi::float8_e5m2);
DEFINE_CPU_TRANS_NORMAL(phi::float16);
DEFINE_CPU_TRANS_NORMAL(phi::bfloat16);
DEFINE_CPU_TRANS_NORMAL(float);
DEFINE_CPU_TRANS_NORMAL(double);
DEFINE_CPU_TRANS_NORMAL(int);
DEFINE_CPU_TRANS_NORMAL(int64_t);
DEFINE_CPU_TRANS_NORMAL(bool);
DEFINE_CPU_TRANS_NORMAL(int16_t);
DEFINE_CPU_TRANS_NORMAL(uint8_t);
DEFINE_CPU_TRANS_NORMAL(uint16_t);
DEFINE_CPU_TRANS_NORMAL(uint32_t);
DEFINE_CPU_TRANS_NORMAL(uint64_t);
DEFINE_CPU_TRANS_NORMAL(int8_t);
DEFINE_CPU_TRANS_NORMAL(phi::complex64);
DEFINE_CPU_TRANS_NORMAL(phi::complex128);

#ifdef PADDLE_WITH_XPU
#define DEFINE_XPU_TRANS_NORMAL(TYPE) \
  template struct TransposeNormal<XPUContext, TYPE>
DEFINE_XPU_TRANS_NORMAL(bool);
DEFINE_XPU_TRANS_NORMAL(float);
DEFINE_XPU_TRANS_NORMAL(int);
DEFINE_XPU_TRANS_NORMAL(int64_t);
DEFINE_XPU_TRANS_NORMAL(phi::complex64);
#endif  // PADDLE_WITH_XPU

struct TensorSetConstantCPU {
  TensorSetConstantCPU(DenseTensor* tensor, float value)
      : tensor_(tensor), value_(value) {}
  template <typename T>
  void apply() const {
    auto cpu = CPUPlace();
    auto* begin = tensor_->mutable_data<T>(cpu);
    std::fill(begin, begin + tensor_->numel(), static_cast<T>(value_));
  }
  DenseTensor* tensor_;
  float value_;
};

template <>
void set_constant_with_place<XPUPlace>(const DeviceContext& dev_ctx,
                                       DenseTensor* tensor,
                                       float value) {
#ifdef PADDLE_WITH_XPU
  phi::VisitDataType(
      tensor->dtype(),
      TensorSetConstantXPU<float>(tensor, value, tensor->place()));
#else
  PADDLE_THROW(common::errors::PreconditionNotMet("Not compiled with XPU!"));
#endif
}

template <>
void set_constant_with_place<phi::IPUPlace>(const DeviceContext& dev_ctx,
                                            DenseTensor* tensor,
                                            float value) {
  PADDLE_THROW(common::errors::Unimplemented("IPUPlace is not supported"));
}

template <>
void set_constant_with_place<CustomPlace>(const DeviceContext& dev_ctx,
                                          DenseTensor* tensor,
                                          float value) {
#ifdef PADDLE_WITH_CUSTOM_DEVICE
  auto kernel_result = phi::KernelFactory::Instance().SelectKernelOrThrowError(
      "full",
      {paddle::experimental::ParseBackend(tensor->place()),
       DataLayout::ALL_LAYOUT,
       paddle::experimental::ParseDataType(tensor->dtype())});
  const auto& kernel = kernel_result.kernel;
  using kernel_signature = void (*)(const DeviceContext&,
                                    const phi::IntArray&,
                                    const phi::Scalar&,
                                    DataType,
                                    DenseTensor*);
  auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
  (*kernel_fn)(dev_ctx,
               phi::IntArray(vectorize(tensor->dims())),
               phi::Scalar(value),
               tensor->dtype(),
               tensor);
#else
  PADDLE_THROW(common::errors::Unimplemented("CustomPlace is not supported"));
#endif
}

template <>
void set_constant_with_place<CPUPlace>(const DeviceContext& dev_ctx,
                                       DenseTensor* tensor,
                                       float value) {
  phi::VisitDataType(tensor->dtype(), TensorSetConstantCPU(tensor, value));
}

template <>
void set_constant_with_place<phi::GPUPinnedPlace>(const DeviceContext& dev_ctx,
                                                  DenseTensor* tensor,
                                                  float value) {
  phi::VisitDataType(tensor->dtype(), TensorSetConstantCPU(tensor, value));
}

struct TensorSetConstantWithPlace {
  using argument_type = phi::Place;
  using result_type = void;
  TensorSetConstantWithPlace(const DeviceContext& dev_ctx,
                             DenseTensor* tensor,
                             float value)
      : dev_ctx_(dev_ctx), tensor_(tensor), value_(value) {}

  template <typename Place>
  void operator()(Place place UNUSED) const {
    set_constant_with_place<Place>(dev_ctx_, tensor_, value_);
  }

  const DeviceContext& dev_ctx_;
  DenseTensor* tensor_;
  float value_;
};

void set_constant(const DeviceContext& dev_ctx,
                  DenseTensor* tensor,
                  float value) {
  TensorSetConstantWithPlace func(dev_ctx, tensor, value);
#ifdef PADDLE_WITH_CUSTOM_DEVICE
  if (dev_ctx.GetPlace().GetType() == AllocationType::CUSTOM) {
    func(CustomPlace());
    return;
  }
#endif
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  // tensor->place().apply_visitor(func);
  phi::VisitPlace(tensor->place(), func);
#elif defined(PADDLE_WITH_XPU)
  if (dev_ctx.GetPlace().GetType() == AllocationType::XPU) {
    func(XPUPlace());
    return;
  } else {
    func(CPUPlace());
  }
#else
  func(CPUPlace());
#endif
}

template struct ColwiseSum<CPUContext, float>;
template struct ColwiseSum<CPUContext, double>;
template struct ColwiseSum<CPUContext, int>;
template struct ColwiseSum<CPUContext, int64_t>;

template struct RowwiseMean<CPUContext, float>;
template struct RowwiseMean<CPUContext, double>;

template <typename T>
struct RowwiseAdd<CPUContext, T> {
  void operator()(const CPUContext& dev_ctx UNUSED,
                  const DenseTensor& input,
                  const DenseTensor& vector,
                  DenseTensor* output) {
    auto in_dims = input.dims();
    const auto& out_dims = output->dims();
    auto size = input.numel() / in_dims[0];
    PADDLE_ENFORCE_EQ(
        vector.numel(),
        size,
        common::errors::InvalidArgument(
            "The input vector size"
            " should be equal to the size of each row of input tensor."
            " Expected vector size=%d, but received %d",
            size,
            vector.numel()));
    PADDLE_ENFORCE_EQ(out_dims,
                      in_dims,
                      common::errors::InvalidArgument(
                          "The output tensor shape should be same as the input"
                          " tensor shape. Expected output tensor shape: %s,"
                          " but received %s",
                          in_dims.to_str().c_str(),
                          out_dims.to_str().c_str()));

    auto in = EigenMatrix<T>::From(input);
    auto vec = EigenVector<T>::Flatten(vector);
    auto out = EigenMatrix<T>::From(*output);

    for (int64_t i = 0; i < in_dims[0]; ++i) {
      out.chip(i, 0) = in.chip(i, 0) + vec;
    }
  }
};

template struct RowwiseAdd<CPUContext, float>;
template struct RowwiseAdd<CPUContext, double>;

}  // namespace phi::funcs