GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-20 00:29:38 +08:00
Code Issues Packages Projects Releases Wiki Activity

optimize non fused MADD, and add a flatten attribute macro to enforce

Browse Source 
inlining within a function
This commit is contained in:
Gael Guennebaud 2010-07-13 15:16:34 +02:00
parent b72b7ab76f
commit 36d9b51a44
3 changed files with 33 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Eigen/src/Core/arch/SSE/Complex.h
View File

@ -284,7 +284,7 @@ template<> EIGEN_STRONG_INLINE Packet1cd ei_por <Packet1cd>(const Packet1cd&
template<> EIGEN_STRONG_INLINE Packet1cd ei_pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); }
template<> EIGEN_STRONG_INLINE Packet1cd ei_pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); }
// FIXME force unaligned load, this is a temporary fix
// FIXME force unaligned load, this is a temporary fix
template<> EIGEN_STRONG_INLINE Packet1cd ei_pload <Packet1cd>(const std::complex<double>* from)
{ EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(ei_ploadu<Packet2d>((const double*)from)); }
template<> EIGEN_STRONG_INLINE Packet1cd ei_ploadu<Packet1cd>(const std::complex<double>* from)

31
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -133,14 +133,34 @@ inline void computeProductBlockingSizes(std::ptrdiff_t& k, std::ptrdiff_t& m, st
computeProductBlockingSizes<LhsScalar,RhsScalar,1>(k, m, n);
}
// FIXME
#ifndef EIGEN_HAS_FUSE_CJMADD
#define EIGEN_HAS_FUSE_CJMADD
#endif
#ifdef EIGEN_HAS_FUSE_CJMADD
#define MADD(CJ,A,B,C,T) C = CJ.pmadd(A,B,C);
#else
#define MADD(CJ,A,B,C,T) T = B; T = CJ.pmul(A,T); C = ei_padd(C,T);
// FIXME (a bit overkill maybe ?)
template<typename CJ, typename A, typename B, typename C, typename T> struct ei_gebp_madd_selector {
EIGEN_STRONG_INLINE EIGEN_ALWAYS_INLINE_ATTRIB static void run(const CJ& cj, A& a, B& b, C& c, T& /*t*/)
{
c = cj.pmadd(a,b,c);
}
};
template<typename CJ, typename T> struct ei_gebp_madd_selector<CJ,T,T,T,T> {
EIGEN_STRONG_INLINE EIGEN_ALWAYS_INLINE_ATTRIB static void run(const CJ& cj, T& a, T& b, T& c, T& t)
{
t = b; t = cj.pmul(a,t); c = ei_padd(c,t);
}
};
template<typename CJ, typename A, typename B, typename C, typename T>
EIGEN_STRONG_INLINE void ei_gebp_madd(const CJ& cj, A& a, B& b, C& c, T& t)
{
ei_gebp_madd_selector<CJ,A,B,C,T>::run(cj,a,b,c,t);
}
#define MADD(CJ,A,B,C,T) ei_gebp_madd(CJ,A,B,C,T);
// #define MADD(CJ,A,B,C,T) T = B; T = CJ.pmul(A,T); C = ei_padd(C,T);
#endif
/* optimized GEneral packed Block * packed Panel product kernel
@ -170,6 +190,7 @@ struct ei_gebp_kernel
typedef typename ei_meta_if<Vectorizable,_RhsPacket,RhsScalar>::ret RhsPacket;
typedef typename ei_meta_if<Vectorizable,_ResPacket,ResScalar>::ret ResPacket;
EIGEN_FLATTEN_ATTRIB
void operator()(ResScalar* res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index rows, Index depth, Index cols, ResScalar alpha,
Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0, RhsScalar* unpackedB = 0)
{

6
Eigen/src/Core/util/Macros.h
View File

@ -147,6 +147,12 @@
#define EIGEN_ALWAYS_INLINE_ATTRIB
#endif
#if EIGEN_GNUC_AT_LEAST(4,0)
#define EIGEN_FLATTEN_ATTRIB __attribute__((flatten))
#else
#define EIGEN_FLATTEN_ATTRIB
#endif
// EIGEN_FORCE_INLINE means "inline as much as possible"
#if (defined _MSC_VER) || (defined __intel_compiler)
#define EIGEN_STRONG_INLINE __forceinline