devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Vectorize cast

Browse Source
This commit is contained in:
Charles Schlosser
2023-04-26 02:50:13 +00:00
committed by Rasmus Munk Larsen
parent 3918768be1
commit eb5ff1861a
8 changed files with 254 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
Eigen/src/Core/arch/AVX/TypeCasting.h
View File

@@ -62,6 +62,15 @@ struct type_casting_traits<float, bool> {
TgtCoeffRatio = 1
};
};
template <>
struct type_casting_traits<float, double> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 1,
TgtCoeffRatio = 2
};
};
#endif // EIGEN_VECTORIZE_AVX512
template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
@@ -80,6 +89,10 @@ template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet4d, Packet8i>(const Packet4d
return _mm256_set_m128i(_mm256_cvttpd_epi32(b), _mm256_cvttpd_epi32(a));
}
template<> EIGEN_STRONG_INLINE Packet4d pcast<Packet8f, Packet4d>(const Packet8f& a) {
return _mm256_cvtps_pd(_mm256_castps256_ps128(a));
}
template <>
EIGEN_STRONG_INLINE Packet16b pcast<Packet8f, Packet16b>(const Packet8f& a,
const Packet8f& b) {

45
Eigen/src/Core/arch/Default/Half.h
View File

@@ -1014,4 +1014,49 @@ struct hash<Eigen::half> {
} // end namespace std
#endif
namespace Eigen {
namespace internal {
template <>
struct cast_impl<float, half> {
EIGEN_DEVICE_FUNC
static inline half run(const float& a) {
#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
return __float2half(a);
#else
return Eigen::half(a);
#endif
}
};
template <>
struct cast_impl<int, half> {
EIGEN_DEVICE_FUNC
static inline half run(const int& a) {
#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
return __float2half(static_cast<float>(a));
#else
return half(static_cast<float>(a));
#endif
}
};
template <>
struct cast_impl<half, float> {
EIGEN_DEVICE_FUNC
static inline float run(const half& a) {
#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
return __half2float(a);
#else
return static_cast<float>(a);
#endif
}
};
} // namespace internal
} // namespace Eigen
#endif // EIGEN_HALF_H

116
Eigen/src/Core/arch/Default/TypeCasting.h
View File

@@ -1,116 +0,0 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
// Copyright (C) 2019 Rasmus Munk Larsen <rmlarsen@google.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_GENERIC_TYPE_CASTING_H
#define EIGEN_GENERIC_TYPE_CASTING_H
#include "../../InternalHeaderCheck.h"
namespace Eigen {
namespace internal {
template<>
struct scalar_cast_op<float, Eigen::half> {
typedef Eigen::half result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
return __float2half(a);
#else
return Eigen::half(a);
#endif
}
};
template<>
struct functor_traits<scalar_cast_op<float, Eigen::half> >
{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
template<>
struct scalar_cast_op<int, Eigen::half> {
typedef Eigen::half result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
return __float2half(static_cast<float>(a));
#else
return Eigen::half(static_cast<float>(a));
#endif
}
};
template<>
struct functor_traits<scalar_cast_op<int, Eigen::half> >
{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
template<>
struct scalar_cast_op<Eigen::half, float> {
typedef float result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
return __half2float(a);
#else
return static_cast<float>(a);
#endif
}
};
template<>
struct functor_traits<scalar_cast_op<Eigen::half, float> >
{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
template<>
struct scalar_cast_op<float, Eigen::bfloat16> {
typedef Eigen::bfloat16 result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::bfloat16 operator() (const float& a) const {
return Eigen::bfloat16(a);
}
};
template<>
struct functor_traits<scalar_cast_op<float, Eigen::bfloat16> >
{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
template<>
struct scalar_cast_op<int, Eigen::bfloat16> {
typedef Eigen::bfloat16 result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::bfloat16 operator() (const int& a) const {
return Eigen::bfloat16(static_cast<float>(a));
}
};
template<>
struct functor_traits<scalar_cast_op<int, Eigen::bfloat16> >
{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
template<>
struct scalar_cast_op<Eigen::bfloat16, float> {
typedef float result_type;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::bfloat16& a) const {
return static_cast<float>(a);
}
};
template<>
struct functor_traits<scalar_cast_op<Eigen::bfloat16, float> >
{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
}
}
#endif // EIGEN_GENERIC_TYPE_CASTING_H