GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-04 03:00:39 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove custom unaligned loads for SSE. They were only useful for core2 CPU.

Browse Source
This commit is contained in:
Gael Guennebaud 2016-02-08 14:29:12 +01:00
parent d039c88096
commit f32ad79b41
1 changed files with 4 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

44
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@ -235,63 +235,27 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { E
return _mm_loadu_ps(from); return _mm_loadu_ps(from);
#endif #endif
} }
template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); }
template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); }
#else #else
// Fast unaligned loads. Note that here we cannot directly use intrinsics: this would
// require pointer casting to incompatible pointer types and leads to invalid code
// because of the strict aliasing rule. The "dummy" stuff are required to enforce
// a correct instruction dependency.
// TODO: do the same for MSVC (ICC is compatible)
// NOTE: with the code below, MSVC's compiler crashes! // NOTE: with the code below, MSVC's compiler crashes!
#if defined(__GNUC__) && defined(__i386__)
// bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd
#define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
#elif defined(__clang__)
// bug 201: Segfaults in __mm_loadh_pd with clang 2.8
#define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
#else
#define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0
#endif
template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
{ {
EIGEN_DEBUG_UNALIGNED_LOAD EIGEN_DEBUG_UNALIGNED_LOAD
#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
return _mm_loadu_ps(from); return _mm_loadu_ps(from);
#else
__m128d res;
res = _mm_load_sd((const double*)(from)) ;
res = _mm_loadh_pd(res, (const double*)(from+2)) ;
return _mm_castpd_ps(res);
#endif
} }
#endif
template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
{ {
EIGEN_DEBUG_UNALIGNED_LOAD EIGEN_DEBUG_UNALIGNED_LOAD
#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
return _mm_loadu_pd(from); return _mm_loadu_pd(from);
#else
__m128d res;
res = _mm_load_sd(from) ;
res = _mm_loadh_pd(res,from+1);
return res;
#endif
} }
template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
{ {
EIGEN_DEBUG_UNALIGNED_LOAD EIGEN_DEBUG_UNALIGNED_LOAD
#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from));
#else
__m128d res;
res = _mm_load_sd((const double*)(from)) ;
res = _mm_loadh_pd(res, (const double*)(from+2)) ;
return _mm_castpd_si128(res);
#endif
} }
#endif
template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
{ {