Files
eigen/Eigen/src/Core/arch/CUDA/Complex.h
Antonio Sanchez bde6741641 Improved std::complex sqrt and rsqrt.
Replaces `std::sqrt` with `complex_sqrt` for all platforms (previously
`complex_sqrt` was only used for CUDA and MSVC), and implements
custom `complex_rsqrt`.

Also introduces `numext::rsqrt` to simplify implementation, and modified
`numext::hypot` to adhere to IEEE IEC 6059 for special cases.

The `complex_sqrt` and `complex_rsqrt` implementations were found to be
significantly faster than `std::sqrt<std::complex<T>>` and
`1/numext::sqrt<std::complex<T>>`.

Benchmark file attached.
```
GCC 10, Intel Xeon, x86_64:
---------------------------------------------------------------------------
Benchmark                                 Time             CPU   Iterations
---------------------------------------------------------------------------
BM_Sqrt<std::complex<float>>           9.21 ns         9.21 ns     73225448
BM_StdSqrt<std::complex<float>>        17.1 ns         17.1 ns     40966545
BM_Sqrt<std::complex<double>>          8.53 ns         8.53 ns     81111062
BM_StdSqrt<std::complex<double>>       21.5 ns         21.5 ns     32757248
BM_Rsqrt<std::complex<float>>          10.3 ns         10.3 ns     68047474
BM_DivSqrt<std::complex<float>>        16.3 ns         16.3 ns     42770127
BM_Rsqrt<std::complex<double>>         11.3 ns         11.3 ns     61322028
BM_DivSqrt<std::complex<double>>       16.5 ns         16.5 ns     42200711

Clang 11, Intel Xeon, x86_64:
---------------------------------------------------------------------------
Benchmark                                 Time             CPU   Iterations
---------------------------------------------------------------------------
BM_Sqrt<std::complex<float>>           7.46 ns         7.45 ns     90742042
BM_StdSqrt<std::complex<float>>        16.6 ns         16.6 ns     42369878
BM_Sqrt<std::complex<double>>          8.49 ns         8.49 ns     81629030
BM_StdSqrt<std::complex<double>>       21.8 ns         21.7 ns     31809588
BM_Rsqrt<std::complex<float>>          8.39 ns         8.39 ns     82933666
BM_DivSqrt<std::complex<float>>        14.4 ns         14.4 ns     48638676
BM_Rsqrt<std::complex<double>>         9.83 ns         9.82 ns     70068956
BM_DivSqrt<std::complex<double>>       15.7 ns         15.7 ns     44487798

Clang 9, Pixel 2, aarch64:
---------------------------------------------------------------------------
Benchmark                                 Time             CPU   Iterations
---------------------------------------------------------------------------
BM_Sqrt<std::complex<float>>           24.2 ns         24.1 ns     28616031
BM_StdSqrt<std::complex<float>>         104 ns          103 ns      6826926
BM_Sqrt<std::complex<double>>          31.8 ns         31.8 ns     22157591
BM_StdSqrt<std::complex<double>>        128 ns          128 ns      5437375
BM_Rsqrt<std::complex<float>>          31.9 ns         31.8 ns     22384383
BM_DivSqrt<std::complex<float>>        99.2 ns         98.9 ns      7250438
BM_Rsqrt<std::complex<double>>         46.0 ns         45.8 ns     15338689
BM_DivSqrt<std::complex<double>>        119 ns          119 ns      5898944
```
2021-01-17 08:50:57 -08:00

103 lines
4.1 KiB
C++

// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_COMPLEX_CUDA_H
#define EIGEN_COMPLEX_CUDA_H
// clang-format off
#if defined(EIGEN_CUDACC) && defined(EIGEN_GPU_COMPILE_PHASE)
namespace Eigen {
namespace internal {
// Many std::complex methods such as operator+, operator-, operator* and
// operator/ are not constexpr. Due to this, clang does not treat them as device
// functions and thus Eigen functors making use of these operators fail to
// compile. Here, we manually specialize these functors for complex types when
// building for CUDA to avoid non-constexpr methods.
// Sum
template<typename T> struct scalar_sum_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
typedef typename std::complex<T> result_type;
EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
return std::complex<T>(numext::real(a) + numext::real(b),
numext::imag(a) + numext::imag(b));
}
};
template<typename T> struct scalar_sum_op<std::complex<T>, std::complex<T> > : scalar_sum_op<const std::complex<T>, const std::complex<T> > {};
// Difference
template<typename T> struct scalar_difference_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
typedef typename std::complex<T> result_type;
EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
return std::complex<T>(numext::real(a) - numext::real(b),
numext::imag(a) - numext::imag(b));
}
};
template<typename T> struct scalar_difference_op<std::complex<T>, std::complex<T> > : scalar_difference_op<const std::complex<T>, const std::complex<T> > {};
// Product
template<typename T> struct scalar_product_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
enum {
Vectorizable = packet_traits<std::complex<T> >::HasMul
};
typedef typename std::complex<T> result_type;
EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
const T a_real = numext::real(a);
const T a_imag = numext::imag(a);
const T b_real = numext::real(b);
const T b_imag = numext::imag(b);
return std::complex<T>(a_real * b_real - a_imag * b_imag,
a_real * b_imag + a_imag * b_real);
}
};
template<typename T> struct scalar_product_op<std::complex<T>, std::complex<T> > : scalar_product_op<const std::complex<T>, const std::complex<T> > {};
// Quotient
template<typename T> struct scalar_quotient_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
enum {
Vectorizable = packet_traits<std::complex<T> >::HasDiv
};
typedef typename std::complex<T> result_type;
EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
const T a_real = numext::real(a);
const T a_imag = numext::imag(a);
const T b_real = numext::real(b);
const T b_imag = numext::imag(b);
const T norm = T(1) / (b_real * b_real + b_imag * b_imag);
return std::complex<T>((a_real * b_real + a_imag * b_imag) * norm,
(a_imag * b_real - a_real * b_imag) * norm);
}
};
template<typename T> struct scalar_quotient_op<std::complex<T>, std::complex<T> > : scalar_quotient_op<const std::complex<T>, const std::complex<T> > {};
} // namespace internal
} // namespace Eigen
#endif
#endif // EIGEN_COMPLEX_CUDA_H