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

First step toward a unification of packet log implementation, currently only SSE and AVX are unified.

Browse Source 
To this end, I added the following functions: pzero, pcmp_*, pfrexp, pset1frombits functions.
This commit is contained in:
Gael Guennebaud
2018-11-26 14:21:24 +01:00
parent 5f6045077c
commit 2c44c40114
6 changed files with 220 additions and 183 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@@ -83,7 +83,7 @@ template<> struct is_arithmetic<__m128d> { enum { value = true }; };
const Packet2d p2d_##NAME = pset1<Packet2d>(X)
#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
const Packet4f p4f_##NAME = _mm_castsi128_ps(pset1<Packet4i>(X))
const Packet4f p4f_##NAME = pset1frombits<Packet4f>(X)
#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
const Packet4i p4i_##NAME = pset1<Packet4i>(X)
@@ -180,6 +180,11 @@ template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { re
template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); }
#endif
template<> EIGEN_STRONG_INLINE Packet4f pset1frombits<Packet4f>(unsigned int from) { return _mm_castsi128_ps(pset1<Packet4i>(from)); }
template<> EIGEN_STRONG_INLINE Packet4f pzero(const Packet4f& /*a*/) { return _mm_setzero_ps(); }
template<> EIGEN_STRONG_INLINE Packet2d pzero(const Packet2d& /*a*/) { return _mm_setzero_pd(); }
// GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction.
// However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203)
// Using inline assembly is also not an option because then gcc fails to reorder properly the instructions.
@@ -328,6 +333,12 @@ template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const
#endif
}
template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(const Packet4f& a, const Packet4f& b) { return _mm_cmple_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4f& b) { return _mm_cmplt_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return _mm_cmpeq_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return _mm_cmpnge_ps(a,b); }
#ifdef EIGEN_VECTORIZE_SSE4_1
template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
@@ -517,6 +528,14 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
#endif
}
template<> EIGEN_STRONG_INLINE Packet4f pshiftright_and_cast(Packet4f a, int n) {
return _mm_cvtepi32_ps(_mm_srli_epi32(_mm_castps_si128(a),n));
}
template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) {
return pfrexp_float(a,exponent);
}
// with AVX, the default implementations based on pload1 are faster
#ifndef __AVX__
template<> EIGEN_STRONG_INLINE void