mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-02-05 17:50:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

bug #1636: fix gemm performance issue with gcc>=6 and no FMA

Browse Source
This commit is contained in:
Gael Guennebaud 2018-12-07 09:15:46 +01:00
parent cbf2f4b7a0
commit 4e7746fe22
1 changed files with 10 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -1387,10 +1387,18 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
EIGEN_ASM_COMMENT("begin gebp micro kernel 2pX4");
RhsPacket B_0, B1, B2, B3, T0;
#define EIGEN_GEBGP_ONESTEP(K) \
// NOTE: the begin/end asm comments below work around bug 935!
// but they are not enough for gcc>=6 without FMA (bug 1637)
#if EIGEN_GNUC_AT_LEAST(6,0)
#define EIGEN_GEBP_2PX4_SPILLING_WORKAROUND asm("" : [a0] "+x" (A0), [a1] "+x" (A1) );
#else
#define EIGEN_GEBP_2PX4_SPILLING_WORKAROUND
#endif
#define EIGEN_GEBGP_ONESTEP(K) \
do { \
EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX4"); \
EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
EIGEN_GEBP_2PX4_SPILLING_WORKAROUND \
traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0); \
traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1); \
traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], B_0, B1, B2, B3); \