mirror of
https://gitlab.com/libeigen/eigen.git
synced 2026-04-10 11:34:33 +08:00
merge Tensor module within Eigen/unsupported and update gemv BLAS wrapper
This commit is contained in:
Gael Guennebaud
@@ -58,7 +58,8 @@ template<> struct packet_traits<float> : default_packet_traits
|
||||
HasCos = 0,
|
||||
HasLog = 0,
|
||||
HasExp = 0,
|
||||
HasSqrt = 0
|
||||
HasSqrt = 0,
|
||||
HasBlend = 1
|
||||
};
|
||||
};
|
||||
template<> struct packet_traits<double> : default_packet_traits
|
||||
@@ -72,7 +73,8 @@ template<> struct packet_traits<double> : default_packet_traits
|
||||
HasHalfPacket = 1,
|
||||
|
||||
HasDiv = 1,
|
||||
HasExp = 0
|
||||
HasExp = 0,
|
||||
HasBlend = 1
|
||||
};
|
||||
};
|
||||
|
||||
@@ -557,6 +559,19 @@ ptranspose(PacketBlock<Packet4d,4>& kernel) {
|
||||
kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49);
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) {
|
||||
const __m256 zero = _mm256_setzero_ps();
|
||||
const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
|
||||
__m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ);
|
||||
return _mm256_blendv_ps(thenPacket, elsePacket, false_mask);
|
||||
}
|
||||
template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) {
|
||||
const __m256d zero = _mm256_setzero_pd();
|
||||
const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
|
||||
__m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ);
|
||||
return _mm256_blendv_pd(thenPacket, elsePacket, false_mask);
|
||||
}
|
||||
|
||||
} // end namespace internal
|
||||
|
||||
} // end namespace Eigen
|
||||
|
||||
75
Eigen/src/Core/arch/CUDA/MathFunctions.h
Normal file
75
Eigen/src/Core/arch/CUDA/MathFunctions.h
Normal file
@@ -0,0 +1,75 @@
|
||||
// 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_MATH_FUNCTIONS_CUDA_H
|
||||
#define EIGEN_MATH_FUNCTIONS_CUDA_H
|
||||
|
||||
namespace Eigen {
|
||||
|
||||
namespace internal {
|
||||
|
||||
// Make sure this is only available when targeting a GPU: we don't want to
|
||||
// introduce conflicts between these packet_traits definitions and the ones
|
||||
// we'll use on the host side (SSE, AVX, ...)
|
||||
#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
float4 plog<float4>(const float4& a)
|
||||
{
|
||||
return make_float4(logf(a.x), logf(a.y), logf(a.z), logf(a.w));
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
double2 plog<double2>(const double2& a)
|
||||
{
|
||||
return make_double2(log(a.x), log(a.y));
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
float4 pexp<float4>(const float4& a)
|
||||
{
|
||||
return make_float4(expf(a.x), expf(a.y), expf(a.z), expf(a.w));
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
double2 pexp<double2>(const double2& a)
|
||||
{
|
||||
return make_double2(exp(a.x), exp(a.y));
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
float4 psqrt<float4>(const float4& a)
|
||||
{
|
||||
return make_float4(sqrtf(a.x), sqrtf(a.y), sqrtf(a.z), sqrtf(a.w));
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
double2 psqrt<double2>(const double2& a)
|
||||
{
|
||||
return make_double2(sqrt(a.x), sqrt(a.y));
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
float4 prsqrt<float4>(const float4& a)
|
||||
{
|
||||
return make_float4(rsqrtf(a.x), rsqrtf(a.y), rsqrtf(a.z), rsqrtf(a.w));
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE
|
||||
double2 prsqrt<double2>(const double2& a)
|
||||
{
|
||||
return make_double2(rsqrt(a.x), rsqrt(a.y));
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
} // end namespace internal
|
||||
|
||||
} // end namespace Eigen
|
||||
|
||||
#endif // EIGEN_MATH_FUNCTIONS_CUDA_H
|
||||
296
Eigen/src/Core/arch/CUDA/PacketMath.h
Normal file
296
Eigen/src/Core/arch/CUDA/PacketMath.h
Normal file
@@ -0,0 +1,296 @@
|
||||
// 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_PACKET_MATH_CUDA_H
|
||||
#define EIGEN_PACKET_MATH_CUDA_H
|
||||
|
||||
namespace Eigen {
|
||||
|
||||
namespace internal {
|
||||
|
||||
// Make sure this is only available when targeting a GPU: we don't want to
|
||||
// introduce conflicts between these packet_traits definitions and the ones
|
||||
// we'll use on the host side (SSE, AVX, ...)
|
||||
#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
|
||||
template<> struct is_arithmetic<float4> { enum { value = true }; };
|
||||
template<> struct is_arithmetic<double2> { enum { value = true }; };
|
||||
|
||||
|
||||
template<> struct packet_traits<float> : default_packet_traits
|
||||
{
|
||||
typedef float4 type;
|
||||
typedef float4 half;
|
||||
enum {
|
||||
Vectorizable = 1,
|
||||
AlignedOnScalar = 1,
|
||||
size=4,
|
||||
HasHalfPacket = 0,
|
||||
|
||||
HasDiv = 1,
|
||||
HasSin = 0,
|
||||
HasCos = 0,
|
||||
HasLog = 1,
|
||||
HasExp = 1,
|
||||
HasSqrt = 1,
|
||||
HasRsqrt = 1,
|
||||
|
||||
HasBlend = 0,
|
||||
};
|
||||
};
|
||||
|
||||
template<> struct packet_traits<double> : default_packet_traits
|
||||
{
|
||||
typedef double2 type;
|
||||
typedef double2 half;
|
||||
enum {
|
||||
Vectorizable = 1,
|
||||
AlignedOnScalar = 1,
|
||||
size=2,
|
||||
HasHalfPacket = 0,
|
||||
|
||||
HasDiv = 1,
|
||||
HasLog = 1,
|
||||
HasExp = 1,
|
||||
HasSqrt = 1,
|
||||
HasRsqrt = 1,
|
||||
|
||||
HasBlend = 0,
|
||||
};
|
||||
};
|
||||
|
||||
|
||||
template<> struct unpacket_traits<float4> { typedef float type; enum {size=4}; typedef float4 half; };
|
||||
template<> struct unpacket_traits<double2> { typedef double type; enum {size=2}; typedef double2 half; };
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pset1<float4>(const float& from) {
|
||||
return make_float4(from, from, from, from);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pset1<double2>(const double& from) {
|
||||
return make_double2(from, from);
|
||||
}
|
||||
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 plset<float>(const float& a) {
|
||||
return make_float4(a, a+1, a+2, a+3);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 plset<double>(const double& a) {
|
||||
return make_double2(a, a+1);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 padd<float4>(const float4& a, const float4& b) {
|
||||
return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 padd<double2>(const double2& a, const double2& b) {
|
||||
return make_double2(a.x+b.x, a.y+b.y);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 psub<float4>(const float4& a, const float4& b) {
|
||||
return make_float4(a.x-b.x, a.y-b.y, a.z-b.z, a.w-b.w);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 psub<double2>(const double2& a, const double2& b) {
|
||||
return make_double2(a.x-b.x, a.y-b.y);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pnegate(const float4& a) {
|
||||
return make_float4(-a.x, -a.y, -a.z, -a.w);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pnegate(const double2& a) {
|
||||
return make_double2(-a.x, -a.y);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pconj(const float4& a) { return a; }
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pconj(const double2& a) { return a; }
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmul<float4>(const float4& a, const float4& b) {
|
||||
return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmul<double2>(const double2& a, const double2& b) {
|
||||
return make_double2(a.x*b.x, a.y*b.y);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pdiv<float4>(const float4& a, const float4& b) {
|
||||
return make_float4(a.x/b.x, a.y/b.y, a.z/b.z, a.w/b.w);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pdiv<double2>(const double2& a, const double2& b) {
|
||||
return make_double2(a.x/b.x, a.y/b.y);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmin<float4>(const float4& a, const float4& b) {
|
||||
return make_float4(fminf(a.x, b.x), fminf(a.y, b.y), fminf(a.z, b.z), fminf(a.w, b.w));
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmin<double2>(const double2& a, const double2& b) {
|
||||
return make_double2(fmin(a.x, b.x), fmin(a.y, b.y));
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmax<float4>(const float4& a, const float4& b) {
|
||||
return make_float4(fmaxf(a.x, b.x), fmaxf(a.y, b.y), fmaxf(a.z, b.z), fmaxf(a.w, b.w));
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmax<double2>(const double2& a, const double2& b) {
|
||||
return make_double2(fmax(a.x, b.x), fmax(a.y, b.y));
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pload<float4>(const float* from) {
|
||||
return *reinterpret_cast<const float4*>(from);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pload<double2>(const double* from) {
|
||||
return *reinterpret_cast<const double2*>(from);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 ploadu<float4>(const float* from) {
|
||||
return make_float4(from[0], from[1], from[2], from[3]);
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploadu<double2>(const double* from) {
|
||||
return make_double2(from[0], from[1]);
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE float4 ploaddup<float4>(const float* from) {
|
||||
return make_float4(from[0], from[0], from[1], from[1]);
|
||||
}
|
||||
template<> EIGEN_STRONG_INLINE double2 ploaddup<double2>(const double* from) {
|
||||
return make_double2(from[0], from[0]);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<float>(float* to, const float4& from) {
|
||||
*reinterpret_cast<float4*>(to) = from;
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<double>(double* to, const double2& from) {
|
||||
*reinterpret_cast<double2*>(to) = from;
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const float4& from) {
|
||||
to[0] = from.x;
|
||||
to[1] = from.y;
|
||||
to[2] = from.z;
|
||||
to[3] = from.w;
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const double2& from) {
|
||||
to[0] = from.x;
|
||||
to[1] = from.y;
|
||||
}
|
||||
|
||||
#ifdef __CUDA_ARCH__
|
||||
template<>
|
||||
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const float* from) {
|
||||
return __ldg((const float4*)from);
|
||||
}
|
||||
template<>
|
||||
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const double* from) {
|
||||
return __ldg((const double2*)from);
|
||||
}
|
||||
|
||||
template<>
|
||||
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const float* from) {
|
||||
return make_float4(__ldg(from+0), __ldg(from+1), __ldg(from+2), __ldg(from+3));
|
||||
}
|
||||
template<>
|
||||
EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Unaligned>(const double* from) {
|
||||
return make_double2(__ldg(from+0), __ldg(from+1));
|
||||
}
|
||||
#endif
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, int stride) {
|
||||
return make_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, int stride) {
|
||||
return make_double2(from[0*stride], from[1*stride]);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, int stride) {
|
||||
to[stride*0] = from.x;
|
||||
to[stride*1] = from.y;
|
||||
to[stride*2] = from.z;
|
||||
to[stride*3] = from.w;
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, int stride) {
|
||||
to[stride*0] = from.x;
|
||||
to[stride*1] = from.y;
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline float pfirst<float4>(const float4& a) {
|
||||
return a.x;
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC inline double pfirst<double2>(const double2& a) {
|
||||
return a.x;
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline float predux<float4>(const float4& a) {
|
||||
return a.x + a.y + a.z + a.w;
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC inline double predux<double2>(const double2& a) {
|
||||
return a.x + a.y;
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline float predux_max<float4>(const float4& a) {
|
||||
return fmaxf(fmaxf(a.x, a.y), fmaxf(a.z, a.w));
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC inline double predux_max<double2>(const double2& a) {
|
||||
return fmax(a.x, a.y);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline float predux_min<float4>(const float4& a) {
|
||||
return fminf(fminf(a.x, a.y), fminf(a.z, a.w));
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC inline double predux_min<double2>(const double2& a) {
|
||||
return fmin(a.x, a.y);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline float4 pabs<float4>(const float4& a) {
|
||||
return make_float4(fabs(a.x), fabs(a.y), fabs(a.z), fabs(a.w));
|
||||
}
|
||||
template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) {
|
||||
return make_double2(abs(a.x), abs(a.y));
|
||||
}
|
||||
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline void
|
||||
ptranspose(PacketBlock<float4,4>& kernel) {
|
||||
double tmp = kernel.packet[0].y;
|
||||
kernel.packet[0].y = kernel.packet[1].x;
|
||||
kernel.packet[1].x = tmp;
|
||||
|
||||
tmp = kernel.packet[0].z;
|
||||
kernel.packet[0].z = kernel.packet[2].x;
|
||||
kernel.packet[2].x = tmp;
|
||||
|
||||
tmp = kernel.packet[0].w;
|
||||
kernel.packet[0].w = kernel.packet[3].x;
|
||||
kernel.packet[3].x = tmp;
|
||||
|
||||
tmp = kernel.packet[1].z;
|
||||
kernel.packet[1].z = kernel.packet[2].y;
|
||||
kernel.packet[2].y = tmp;
|
||||
|
||||
tmp = kernel.packet[1].w;
|
||||
kernel.packet[1].w = kernel.packet[3].y;
|
||||
kernel.packet[3].y = tmp;
|
||||
|
||||
tmp = kernel.packet[2].w;
|
||||
kernel.packet[2].w = kernel.packet[3].z;
|
||||
kernel.packet[3].z = tmp;
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline void
|
||||
ptranspose(PacketBlock<double2,2>& kernel) {
|
||||
double tmp = kernel.packet[0].y;
|
||||
kernel.packet[0].y = kernel.packet[1].x;
|
||||
kernel.packet[1].x = tmp;
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
} // end namespace internal
|
||||
|
||||
} // end namespace Eigen
|
||||
|
||||
|
||||
#endif // EIGEN_PACKET_MATH_CUDA_H
|
||||
@@ -44,7 +44,8 @@ template<> struct packet_traits<std::complex<float> > : default_packet_traits
|
||||
HasAbs2 = 0,
|
||||
HasMin = 0,
|
||||
HasMax = 0,
|
||||
HasSetLinear = 0
|
||||
HasSetLinear = 0,
|
||||
HasBlend = 1
|
||||
};
|
||||
};
|
||||
#endif
|
||||
@@ -472,6 +473,11 @@ ptranspose(PacketBlock<Packet2cf,2>& kernel) {
|
||||
kernel.packet[1].v = tmp;
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
|
||||
__m128d result = pblend(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v));
|
||||
return Packet2cf(_mm_castpd_ps(result));
|
||||
}
|
||||
|
||||
} // end namespace internal
|
||||
|
||||
} // end namespace Eigen
|
||||
|
||||
@@ -108,7 +108,8 @@ template<> struct packet_traits<float> : default_packet_traits
|
||||
HasCos = EIGEN_FAST_MATH,
|
||||
HasLog = 1,
|
||||
HasExp = 1,
|
||||
HasSqrt = 1
|
||||
HasSqrt = 1,
|
||||
HasBlend = 1
|
||||
};
|
||||
};
|
||||
template<> struct packet_traits<double> : default_packet_traits
|
||||
@@ -123,7 +124,8 @@ template<> struct packet_traits<double> : default_packet_traits
|
||||
|
||||
HasDiv = 1,
|
||||
HasExp = 1,
|
||||
HasSqrt = 1
|
||||
HasSqrt = 1,
|
||||
HasBlend = 1
|
||||
};
|
||||
};
|
||||
#endif
|
||||
@@ -135,7 +137,9 @@ template<> struct packet_traits<int> : default_packet_traits
|
||||
// FIXME check the Has*
|
||||
Vectorizable = 1,
|
||||
AlignedOnScalar = 1,
|
||||
size=4
|
||||
size=4,
|
||||
|
||||
HasBlend = 1
|
||||
};
|
||||
};
|
||||
|
||||
@@ -809,6 +813,37 @@ ptranspose(PacketBlock<Packet4i,4>& kernel) {
|
||||
kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
|
||||
}
|
||||
|
||||
template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
|
||||
const __m128i zero = _mm_setzero_si128();
|
||||
const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
|
||||
__m128i false_mask = _mm_cmpeq_epi32(select, zero);
|
||||
#ifdef EIGEN_VECTORIZE_SSE4_1
|
||||
return _mm_blendv_epi8(thenPacket, elsePacket, false_mask);
|
||||
#else
|
||||
return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket));
|
||||
#endif
|
||||
}
|
||||
template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
|
||||
const __m128 zero = _mm_setzero_ps();
|
||||
const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
|
||||
__m128 false_mask = _mm_cmpeq_ps(select, zero);
|
||||
#ifdef EIGEN_VECTORIZE_SSE4_1
|
||||
return _mm_blendv_ps(thenPacket, elsePacket, false_mask);
|
||||
#else
|
||||
return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket));
|
||||
#endif
|
||||
}
|
||||
template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
|
||||
const __m128d zero = _mm_setzero_pd();
|
||||
const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]);
|
||||
__m128d false_mask = _mm_cmpeq_pd(select, zero);
|
||||
#ifdef EIGEN_VECTORIZE_SSE4_1
|
||||
return _mm_blendv_pd(thenPacket, elsePacket, false_mask);
|
||||
#else
|
||||
return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket));
|
||||
#endif
|
||||
}
|
||||
|
||||
} // end namespace internal
|
||||
|
||||
} // end namespace Eigen
|
||||
|
||||
Reference in New Issue
Block a user