// -*- C++ -*-
//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
//
//===----------------------------------------------------------------------===//

// Modifications Copyright (c) 2025 Advanced Micro Devices, Inc.
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.

#ifndef _LIBCUDACXX___CUDA_COMPLEX_NVFP16_H
#define _LIBCUDACXX___CUDA_COMPLEX_NVFP16_H

#include <cuda/std/detail/__config>

#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
#  pragma GCC system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
#  pragma clang system_header
#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
#  pragma system_header
#endif // no system header

#if _LIBCUDACXX_HAS_NVFP16()

#  include <cuda/std/__cmath/nvfp16.h>
#  include <cuda/std/__complex/vector_support.h>
#  include <cuda/std/__floating_point/nvfp_types.h>
#  include <cuda/std/__fwd/get.h>
#  include <cuda/std/__type_traits/enable_if.h>
#  include <cuda/std/__type_traits/integral_constant.h>
#  include <cuda/std/__type_traits/is_constructible.h>
#  include <cuda/std/__type_traits/is_extended_floating_point.h>
#  include <cuda/std/cmath>
#  include <cuda/std/complex>

#  if !_CCCL_COMPILER(NVRTC) && !defined(_CCCL_COMPILER_HIPRTC)
#    include <sstream> // for std::basic_ostringstream
#  endif // !_CCCL_COMPILER(NVRTC)

// This is a workaround against the user defining macros __CUDA_NO_HALF_CONVERSIONS__ __CUDA_NO_HALF_OPERATORS__
namespace __cccl_internal
{
template <>
struct __is_non_narrowing_convertible<__half, float>
{
  static constexpr bool value = true;
};

template <>
struct __is_non_narrowing_convertible<__half, double>
{
  static constexpr bool value = true;
};

template <>
struct __is_non_narrowing_convertible<float, __half>
{
  static constexpr bool value = true;
};

template <>
struct __is_non_narrowing_convertible<double, __half>
{
  static constexpr bool value = true;
};
} // namespace __cccl_internal

_LIBCUDACXX_BEGIN_NAMESPACE_STD

template <>
struct __complex_alignment<__half> : integral_constant<size_t, alignof(__half2)>
{};

template <>
struct __type_to_vector<__half>
{
  using __type = __half2;
};

template <>
struct __cccl_complex_overload_traits<__half, false, false>
{
  using _ValueType   = __half;
  using _ComplexType = complex<__half>;
};

template <class _Tp>
_LIBCUDACXX_HIDE_FROM_ABI __half __convert_to_half(const _Tp& __value) noexcept
{
  return __value;
}

_LIBCUDACXX_HIDE_FROM_ABI __half __convert_to_half(const float& __value) noexcept
{
  return __float2half(__value);
}

_LIBCUDACXX_HIDE_FROM_ABI __half __convert_to_half(const double& __value) noexcept
{
  // NOTE(HIP/AMD): fp16 currently does not provide a __double2half conversion
  // (see https://github.com/ROCm/clr/blob/amd-staging/hipamd/include/hip/amd_detail/amd_hip_fp16.h)
  // corresponding ticket SWDEV-529927
  return __float2half(static_cast<float>(__value));
}

template <>
class _CCCL_TYPE_VISIBILITY_DEFAULT _CCCL_ALIGNAS(alignof(__half2)) complex<__half>
{
  __half2 __repr_;

  template <class _Up>
  friend class complex;

  template <class _Up>
  friend struct __get_complex_impl;

public:
  using value_type = __half;

  _LIBCUDACXX_HIDE_FROM_ABI complex(const value_type& __re = value_type(), const value_type& __im = value_type())
      : __repr_(__re, __im)
  {}

  template <class _Up, enable_if_t<__cccl_internal::__is_non_narrowing_convertible<value_type, _Up>::value, int> = 0>
  _LIBCUDACXX_HIDE_FROM_ABI complex(const complex<_Up>& __c)
      : __repr_(__convert_to_half(__c.real()), __convert_to_half(__c.imag()))
  {}

  template <class _Up,
            enable_if_t<!__cccl_internal::__is_non_narrowing_convertible<value_type, _Up>::value, int> = 0,
            enable_if_t<_CCCL_TRAIT(is_constructible, value_type, _Up), int>                           = 0>
  _LIBCUDACXX_HIDE_FROM_ABI explicit complex(const complex<_Up>& __c)
      : __repr_(__convert_to_half(__c.real()), __convert_to_half(__c.imag()))
  {}

  _LIBCUDACXX_HIDE_FROM_ABI complex& operator=(const value_type& __re)
  {
    __repr_.x = __re;
    __repr_.y = value_type();
    return *this;
  }

  template <class _Up>
  _LIBCUDACXX_HIDE_FROM_ABI complex& operator=(const complex<_Up>& __c)
  {
    __repr_.x = __convert_to_half(__c.real());
    __repr_.y = __convert_to_half(__c.imag());
    return *this;
  }

#  if !_CCCL_COMPILER(NVRTC) && !defined(_CCCL_COMPILER_HIPRTC)
  template <class _Up>
  _LIBCUDACXX_HIDE_FROM_ABI complex(const ::std::complex<_Up>& __other)
      : __repr_(_LIBCUDACXX_ACCESS_STD_COMPLEX_REAL(__other), _LIBCUDACXX_ACCESS_STD_COMPLEX_IMAG(__other))
  {}

  template <class _Up>
  _LIBCUDACXX_HIDE_FROM_ABI complex& operator=(const ::std::complex<_Up>& __other)
  {
    __repr_.x = _LIBCUDACXX_ACCESS_STD_COMPLEX_REAL(__other);
    __repr_.y = _LIBCUDACXX_ACCESS_STD_COMPLEX_IMAG(__other);
    return *this;
  }

  _CCCL_HOST operator ::std::complex<value_type>() const
  {
    return {__repr_.x, __repr_.y};
  }
#  endif // !_CCCL_COMPILER(NVRTC)

  _LIBCUDACXX_HIDE_FROM_ABI value_type real() const
  {
    return __repr_.x;
  }
  _LIBCUDACXX_HIDE_FROM_ABI value_type imag() const
  {
    return __repr_.y;
  }

  _LIBCUDACXX_HIDE_FROM_ABI void real(value_type __re)
  {
    __repr_.x = __re;
  }
  _LIBCUDACXX_HIDE_FROM_ABI void imag(value_type __im)
  {
    __repr_.y = __im;
  }

  // Those additional volatile overloads are meant to help with reductions in thrust
  _LIBCUDACXX_HIDE_FROM_ABI value_type real() const volatile
  {
    return __repr_.x;
  }
  _LIBCUDACXX_HIDE_FROM_ABI value_type imag() const volatile
  {
    return __repr_.y;
  }

  _LIBCUDACXX_HIDE_FROM_ABI complex& operator+=(const value_type& __re)
  {
    __repr_.x = __hadd(__repr_.x, __re);
    return *this;
  }
  _LIBCUDACXX_HIDE_FROM_ABI complex& operator-=(const value_type& __re)
  {
    __repr_.x = __hsub(__repr_.x, __re);
    return *this;
  }
  _LIBCUDACXX_HIDE_FROM_ABI complex& operator*=(const value_type& __re)
  {
    __repr_.x = __hmul(__repr_.x, __re);
    __repr_.y = __hmul(__repr_.y, __re);
    return *this;
  }
  _LIBCUDACXX_HIDE_FROM_ABI complex& operator/=(const value_type& __re)
  {
    __repr_.x = __hdiv(__repr_.x, __re);
    __repr_.y = __hdiv(__repr_.y, __re);
    return *this;
  }

  // We can utilize vectorized operations for those operators
  _LIBCUDACXX_HIDE_FROM_ABI friend complex& operator+=(complex& __lhs, const complex& __rhs) noexcept
  {
    __lhs.__repr_ = __hadd2(__lhs.__repr_, __rhs.__repr_);
    return __lhs;
  }

  _LIBCUDACXX_HIDE_FROM_ABI friend complex& operator-=(complex& __lhs, const complex& __rhs) noexcept
  {
    __lhs.__repr_ = __hsub2(__lhs.__repr_, __rhs.__repr_);
    return __lhs;
  }

  _LIBCUDACXX_HIDE_FROM_ABI friend bool operator==(const complex& __lhs, const complex& __rhs) noexcept
  {
    return __hbeq2(__lhs.__repr_, __rhs.__repr_);
  }
};

template <> // complex<float>
template <> // complex<__half>
_LIBCUDACXX_HIDE_FROM_ABI complex<float>::complex(const complex<__half>& __c)
    : __re_(__half2float(__c.real()))
    , __im_(__half2float(__c.imag()))
{}

template <> // complex<double>
template <> // complex<__half>
_LIBCUDACXX_HIDE_FROM_ABI complex<double>::complex(const complex<__half>& __c)
    : __re_(__half2float(__c.real()))
    , __im_(__half2float(__c.imag()))
{}

template <> // complex<float>
template <> // complex<__half>
_LIBCUDACXX_HIDE_FROM_ABI complex<float>& complex<float>::operator=(const complex<__half>& __c)
{
  __re_ = __half2float(__c.real());
  __im_ = __half2float(__c.imag());
  return *this;
}

template <> // complex<double>
template <> // complex<__half>
_LIBCUDACXX_HIDE_FROM_ABI complex<double>& complex<double>::operator=(const complex<__half>& __c)
{
  __re_ = __half2float(__c.real());
  __im_ = __half2float(__c.imag());
  return *this;
}

_LIBCUDACXX_HIDE_FROM_ABI __half arg(__half __re)
{
  // NOTE(HIP/AMD): fp16 currently does not provide __int2half_rn for host code; this is implementated for floating point alternatives
  // (device only see https://github.com/ROCm/clr/blob/amd-staging/hipamd/include/hip/amd_detail/amd_hip_fp16.h)
  // corresponding ticket SWDEV-529927
  return _CUDA_VSTD::atan2(__float2half_rn(0.0f), __re);
}

// We have performance issues with some trigonometric functions with __half
template <>
_LIBCUDACXX_HIDE_FROM_ABI complex<__half> asinh(const complex<__half>& __x)
{
  return complex<__half>{_CUDA_VSTD::asinh(complex<float>{__x})};
}
template <>
_LIBCUDACXX_HIDE_FROM_ABI complex<__half> acosh(const complex<__half>& __x)
{
  return complex<__half>{_CUDA_VSTD::acosh(complex<float>{__x})};
}
template <>
_LIBCUDACXX_HIDE_FROM_ABI complex<__half> atanh(const complex<__half>& __x)
{
  return complex<__half>{_CUDA_VSTD::atanh(complex<float>{__x})};
}
template <>
_LIBCUDACXX_HIDE_FROM_ABI complex<__half> acos(const complex<__half>& __x)
{
  return complex<__half>{_CUDA_VSTD::acos(complex<float>{__x})};
}

template <>
struct __get_complex_impl<__half>
{
  template <size_t _Index>
  static _LIBCUDACXX_HIDE_FROM_ABI constexpr __half& get(complex<__half>& __z) noexcept
  {
    return (_Index == 0) ? __z.__repr_.x : __z.__repr_.y;
  }

  template <size_t _Index>
  static _LIBCUDACXX_HIDE_FROM_ABI constexpr __half&& get(complex<__half>&& __z) noexcept
  {
    return _CUDA_VSTD::move((_Index == 0) ? __z.__repr_.x : __z.__repr_.y);
  }

  template <size_t _Index>
  static _LIBCUDACXX_HIDE_FROM_ABI constexpr const __half& get(const complex<__half>& __z) noexcept
  {
    return (_Index == 0) ? __z.__repr_.x : __z.__repr_.y;
  }

  template <size_t _Index>
  static _LIBCUDACXX_HIDE_FROM_ABI constexpr const __half&& get(const complex<__half>&& __z) noexcept
  {
    return _CUDA_VSTD::move((_Index == 0) ? __z.__repr_.x : __z.__repr_.y);
  }
};

#  if !defined(_LIBCUDACXX_HAS_NO_LOCALIZATION) && !_CCCL_COMPILER(NVRTC) && !defined(_CCCL_COMPILER_HIPRTC)
template <class _CharT, class _Traits>
::std::basic_istream<_CharT, _Traits>& operator>>(::std::basic_istream<_CharT, _Traits>& __is, complex<__half>& __x)
{
  ::std::complex<float> __temp;
  __is >> __temp;
  __x = __temp;
  return __is;
}

template <class _CharT, class _Traits>
::std::basic_ostream<_CharT, _Traits>&
operator<<(::std::basic_ostream<_CharT, _Traits>& __os, const complex<__half>& __x)
{
  return __os << complex<float>{__x};
}
#  endif // !_LIBCUDACXX_HAS_NO_LOCALIZATION && !_CCCL_COMPILER(NVRTC)

_LIBCUDACXX_END_NAMESPACE_STD

#endif // _LIBCUDACXX_HAS_NVFP16()

#endif // _LIBCUDACXX___CUDA_COMPLEX_NVFP16_H