elementwise_functor.h

/* 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

#include "paddle/phi/common/complex.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/enforce.h"
#include "paddle/phi/core/hostdevice.h"

namespace phi {
namespace funcs {

// Define the binary functors used in elementwise ops.
// Note: InverseXxxFunctor is needed when calling ElementwiseComputeEx on CPU.

// Add
template <typename T>
struct AddFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return a + b; }
};
template <typename T>
struct InverseAddFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return b + a; }
};

// Subtract
template <typename T>
struct SubtractFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return a - b; }
};
template <typename T>
struct InverseSubtractFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return b - a; }
};

// Multiply
template <typename T>
struct MultiplyFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return a * b; }
};
template <>
struct MultiplyFunctor<bool> {
  inline HOSTDEVICE bool operator()(const bool a, const bool b) const {
    return a && b;
  }
};
template <typename T>
struct InverseMultiplyFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return b * a; }
};
template <>
struct InverseMultiplyFunctor<bool> {
  inline HOSTDEVICE bool operator()(const bool a, const bool b) const {
    return b && a;
  }
};

// Divide
#define DIV_ERROR_INFO                                             \
  "InvalidArgumentError: Integer division by zero encountered in " \
  "(floor) divide. Please check the input value."

template <typename T, typename Enable = void>
struct DivideFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return a / b; }
};

template <typename T>
struct DivideFunctor<
    T,
    typename std::enable_if<std::is_integral<T>::value>::type> {
  inline HOSTDEVICE T operator()(const T a, const T b) const {
    // For int32/int64, need to check whether the divison is zero.
    PADDLE_ENFORCE(b != 0, DIV_ERROR_INFO);
    return a / b;
  }
};

template <typename T, typename Enable = void>
struct InverseDivideFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return b / a; }
};

template <typename T>
using ComplexType = phi::dtype::complex<T>;

template <typename InT, typename OutT>
struct DivGradXYFunctor {
  inline HOSTDEVICE phi::Array<OutT, 2> operator()(const InT a,
                                                   const InT b,
                                                   const InT c) {
    // dx = dout / y
    // dy = - dout * out / y
    phi::Array<OutT, 2> outs;
    outs[0] = a / c;
    outs[1] = -a * b / c;
    return outs;
  }
};

template <typename InT, typename OutT>
struct DivGradXYFunctor<ComplexType<InT>, ComplexType<OutT>> {
  inline HOSTDEVICE phi::Array<ComplexType<OutT>, 2> operator()(
      const ComplexType<InT> a,
      const ComplexType<InT> b,
      const ComplexType<InT> c) {
    phi::Array<ComplexType<OutT>, 2> outs;
    ComplexType<InT> c_conj(c.real, -c.imag);
    ComplexType<InT> out_div_c_conj((b / c).real, -(b / c).imag);
    outs[0] = a / c_conj;
    outs[1] = -a * out_div_c_conj;
    return outs;
  }
};

// Float div grad
template <typename T>
struct DivGradXFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return a / b; }
};

// ComplexType div grad
template <typename T>
struct DivGradXFunctor<ComplexType<T>> {
  inline HOSTDEVICE ComplexType<T> operator()(const ComplexType<T> a,
                                              const ComplexType<T> b) const {
    ComplexType<T> b_conj(b.real, -b.imag);
    return a / b_conj;
  }
};

// Float mul and div
template <typename T>
struct DivGradYFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b, const T c) const {
    return -a * b / c;
  }
};

// ComplexType mul and div
template <typename T>
struct DivGradYFunctor<ComplexType<T>> {
  inline HOSTDEVICE ComplexType<T> operator()(const ComplexType<T> a,
                                              const ComplexType<T> b,
                                              const ComplexType<T> c) const {
    ComplexType<T> out_div_c_conj((b / c).real, -(b / c).imag);
    return -a * out_div_c_conj;
  }
};
// Fmin
template <typename T>
struct FMinFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const {
    return std::fmin(a, b);
  }
};

template <>
struct FMinFunctor<dtype::float16> {
  inline HOSTDEVICE dtype::float16 operator()(const dtype::float16 a,
                                              const dtype::float16 b) const {
    float float_a = static_cast<float>(a);
    float float_b = static_cast<float>(b);
    auto result = std::fmin(float_a, float_b);
    return static_cast<dtype::float16>(result);
  }
};

template <>
struct FMinFunctor<int> {
  inline HOSTDEVICE int operator()(const int a, const int b) const {
    float float_a = static_cast<float>(a);
    float float_b = static_cast<float>(b);
    auto result = std::fmin(float_a, float_b);
    return std::lrint(result);
  }
};

template <>
struct FMinFunctor<int64_t> {
  inline HOSTDEVICE int64_t operator()(const int64_t a, const int64_t b) const {
    double double_a = static_cast<double>(a);
    double double_b = static_cast<double>(b);
    auto result = std::fmin(double_a, double_b);
    return std::llrint(result);
  }
};

// Fmax
template <typename T>
struct FMaxFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const {
    return std::fmax(a, b);
  }
};

template <>
struct FMaxFunctor<dtype::float16> {
  inline HOSTDEVICE dtype::float16 operator()(const dtype::float16 a,
                                              const dtype::float16 b) const {
    float float_a = static_cast<float>(a);
    float float_b = static_cast<float>(b);
    auto result = std::fmax(float_a, float_b);
    return static_cast<dtype::float16>(result);
  }
};

template <>
struct FMaxFunctor<int> {
  inline HOSTDEVICE int operator()(const int a, const int b) const {
    float float_a = static_cast<float>(a);
    float float_b = static_cast<float>(b);
    auto result = std::fmax(float_a, float_b);
    return std::lrint(result);
  }
};

template <>
struct FMaxFunctor<int64_t> {
  inline HOSTDEVICE int64_t operator()(const int64_t a, const int64_t b) const {
    double double_a = static_cast<double>(a);
    double double_b = static_cast<double>(b);
    auto result = std::fmax(double_a, double_b);
    return std::llrint(result);
  }
};

template <typename T>
struct FMaxGradDx {
  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
    return dout * static_cast<T>((x >= y) || isnan(y));
  }
};

template <>
struct FMaxGradDx<dtype::float16> {
  HOSTDEVICE dtype::float16 operator()(dtype::float16 x,
                                       dtype::float16 y,
                                       dtype::float16 out,
                                       dtype::float16 dout) const {
    return dout * static_cast<dtype::float16>((x >= y) || dtype::isnan(y));
  }
};

template <>
struct FMaxGradDx<int> {
  HOSTDEVICE int operator()(int x, int y, int out, int dout) const {
    return dout * static_cast<int>((x >= y));
  }
};

template <>
struct FMaxGradDx<int64_t> {
  HOSTDEVICE int64_t operator()(int64_t x,
                                int64_t y,
                                int64_t out,
                                int64_t dout) const {
    return dout * static_cast<int64_t>((x >= y));
  }
};

template <typename T>
struct FMaxGradDy {
  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
    return dout * static_cast<T>(!((x >= y) || isnan(y)));
  }
};

template <>
struct FMaxGradDy<dtype::float16> {
  HOSTDEVICE dtype::float16 operator()(dtype::float16 x,
                                       dtype::float16 y,
                                       dtype::float16 out,
                                       dtype::float16 dout) const {
    return dout * static_cast<dtype::float16>(!((x >= y) || dtype::isnan(y)));
  }
};

template <>
struct FMaxGradDy<int64_t> {
  HOSTDEVICE int64_t operator()(int64_t x,
                                int64_t y,
                                int64_t out,
                                int64_t dout) const {
    return dout * static_cast<int64_t>(!((x >= y)));
  }
};

template <>
struct FMaxGradDy<int> {
  HOSTDEVICE int operator()(int x, int y, int out, int dout) const {
    return dout * static_cast<int>(!((x >= y)));
  }
};

template <typename T>
struct FMinGradDx {
  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
    return dout * static_cast<T>((x <= y) || isnan(y));
  }
};

template <>
struct FMinGradDx<dtype::float16> {
  HOSTDEVICE dtype::float16 operator()(dtype::float16 x,
                                       dtype::float16 y,
                                       dtype::float16 out,
                                       dtype::float16 dout) const {
    return dout * static_cast<dtype::float16>((x <= y) || dtype::isnan(y));
  }
};

template <>
struct FMinGradDx<int> {
  HOSTDEVICE int operator()(int x, int y, int out, int dout) const {
    return dout * static_cast<int>((x <= y));
  }
};

template <>
struct FMinGradDx<int64_t> {
  HOSTDEVICE int64_t operator()(int64_t x,
                                int64_t y,
                                int64_t out,
                                int64_t dout) const {
    return dout * static_cast<int64_t>((x <= y));
  }
};

template <typename T>
struct FMinGradDy {
  HOSTDEVICE T operator()(T x, T y, T out, T dout) const {
    return dout * static_cast<T>(!((x <= y) || isnan(y)));
  }
};

template <>
struct FMinGradDy<dtype::float16> {
  HOSTDEVICE dtype::float16 operator()(dtype::float16 x,
                                       dtype::float16 y,
                                       dtype::float16 out,
                                       dtype::float16 dout) const {
    return dout * static_cast<dtype::float16>(!((x <= y) || dtype::isnan(y)));
  }
};

template <>
struct FMinGradDy<int> {
  HOSTDEVICE int operator()(int x, int y, int out, int dout) const {
    return dout * static_cast<int>(!((x <= y)));
  }
};

template <>
struct FMinGradDy<int64_t> {
  HOSTDEVICE int64_t operator()(int64_t x,
                                int64_t y,
                                int64_t out,
                                int64_t dout) const {
    return dout * static_cast<int64_t>(!((x <= y)));
  }
};

template <typename T>
struct MultiplyGradFunctor {
  inline HOSTDEVICE T operator()(const T a, const T b) const { return a * b; }
};
template <typename T>
struct MultiplyGradFunctor<ComplexType<T>> {
  inline HOSTDEVICE ComplexType<T> operator()(const ComplexType<T> a,
                                              const ComplexType<T> b) const {
    ComplexType<T> b_conj(b.real, -b.imag);
    return a * b_conj;
  }
};

template <typename InT, typename OutT>
struct MultiplyGradXYFunctor {
  inline HOSTDEVICE phi::Array<OutT, 2> operator()(const InT a,
                                                   const InT b,
                                                   const InT c) {
    phi::Array<OutT, 2> outs;
    // dx = dout * y
    outs[0] = a * b;
    // dy = dout * x
    outs[1] = a * c;
    return outs;
  }
};

template <typename InT, typename OutT>
struct MultiplyGradXYFunctor<ComplexType<InT>, ComplexType<OutT>> {
  inline HOSTDEVICE phi::Array<ComplexType<OutT>, 2> operator()(
      const ComplexType<InT> a,
      const ComplexType<InT> b,
      const ComplexType<InT> c) {
    phi::Array<ComplexType<OutT>, 2> outs;
    // dx = dout * y
    ComplexType<InT> b_conj(b.real, -b.imag);
    outs[0] = a * b_conj;
    // dy = dout * x
    ComplexType<InT> c_conj(c.real, -c.imag);
    outs[1] = a * c_conj;
    return outs;
  }
};

}  // namespace funcs
}  // namespace phi