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

Unify SSE and AVX implementation of pexp

Browse Source
This commit is contained in:
Gael Guennebaud
2018-11-26 16:36:19 +01:00
parent c2f35b1b47
commit cf8b85d5c5
6 changed files with 126 additions and 122 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@@ -110,12 +110,12 @@ template<> struct packet_traits<float> : default_packet_traits
HasSqrt = 1,
HasRsqrt = 1,
HasTanh = EIGEN_FAST_MATH,
HasBlend = 1
HasBlend = 1,
HasFloor = 1
#ifdef EIGEN_VECTORIZE_SSE4_1
,
HasRound = 1,
HasFloor = 1,
HasCeil = 1
#endif
};
@@ -348,6 +348,17 @@ template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { ret
template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); }
template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); }
#else
template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
{
const Packet4f cst_1 = pset1<Packet4f>(1.0f);
Packet4i emm0 = _mm_cvttps_epi32(a);
Packet4f tmp = _mm_cvtepi32_ps(emm0);
/* if greater, substract 1 */
Packet4f mask = _mm_cmpgt_ps(tmp, a);
mask = _mm_and_ps(mask, cst_1);
return psub(tmp, mask);
}
#endif
template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
@@ -536,6 +547,16 @@ template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Pack
return pfrexp_float(a,exponent);
}
template<> EIGEN_STRONG_INLINE Packet4f pcast_and_shiftleft<Packet4f>(Packet4f v, int n)
{
Packet4i vi = _mm_cvttps_epi32(v);
return _mm_castsi128_ps(_mm_slli_epi32(vi, n));
}
template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) {
return pldexp_float(a,exponent);
}
// with AVX, the default implementations based on pload1 are faster
#ifndef __AVX__
template<> EIGEN_STRONG_INLINE void