Added support for vectorized type casting of tensors

This commit is contained in:
Benoit Steiner
2015-02-27 08:46:04 -08:00
parent f41b1f1666
commit 573b377110
9 changed files with 413 additions and 3 deletions

View File

@@ -98,6 +98,28 @@ template<typename T> struct packet_traits : default_packet_traits
template<typename T> struct packet_traits<const T> : packet_traits<T> { };
template <typename Src, typename Tgt> struct type_casting_traits {
enum {
VectorizedCast = 0,
SrcCoeffRatio = 1,
TgtCoeffRatio = 1
};
};
/** \internal \returns static_cast<TgtType>(a) (coeff-wise) */
template <typename SrcPacket, typename TgtPacket>
EIGEN_DEVICE_FUNC inline TgtPacket
pcast(const SrcPacket& a) {
return static_cast<TgtPacket>(a);
}
template <typename SrcPacket, typename TgtPacket>
EIGEN_DEVICE_FUNC inline TgtPacket
pcast(const SrcPacket& a, const SrcPacket& /*b*/) {
return static_cast<TgtPacket>(a);
}
/** \internal \returns a + b (coeff-wise) */
template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
padd(const Packet& a,

View File

@@ -0,0 +1,51 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2015 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_TYPE_CASTING_AVX_H
#define EIGEN_TYPE_CASTING_AVX_H
namespace Eigen {
namespace internal {
// For now we use SSE to handle integers, so we can't use AVX instructions to cast
// from int to float
template <>
struct type_casting_traits<float, int> {
enum {
VectorizedCast = 0,
SrcCoeffRatio = 1,
TgtCoeffRatio = 1
};
};
template <>
struct type_casting_traits<int, float> {
enum {
VectorizedCast = 0,
SrcCoeffRatio = 1,
TgtCoeffRatio = 1
};
};
template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
return _mm256_cvtps_epi32(a);
}
template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
return _mm256_cvtepi32_ps(a);
}
} // end namespace internal
} // end namespace Eigen
#endif // EIGEN_TYPE_CASTING_AVX_H

View File

@@ -0,0 +1,77 @@
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2015 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_TYPE_CASTING_SSE_H
#define EIGEN_TYPE_CASTING_SSE_H
namespace Eigen {
namespace internal {
template <>
struct type_casting_traits<float, int> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 1,
TgtCoeffRatio = 1
};
};
template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) {
return _mm_cvtps_epi32(a);
}
template <>
struct type_casting_traits<int, float> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 1,
TgtCoeffRatio = 1
};
};
template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) {
return _mm_cvtepi32_ps(a);
}
template <>
struct type_casting_traits<double, float> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 2,
TgtCoeffRatio = 1
};
};
template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet2d, Packet4f>(const Packet2d& a, const Packet2d& b) {
return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6));
}
template <>
struct type_casting_traits<float, double> {
enum {
VectorizedCast = 1,
SrcCoeffRatio = 1,
TgtCoeffRatio = 2
};
};
template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet4f, Packet2d>(const Packet4f& a) {
// Simply discard the second half of the input
return _mm_cvtps_pd(a);
}
} // end namespace internal
} // end namespace Eigen
#endif // EIGEN_TYPE_CASTING_SSE_H