GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-06-04 18:54:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Disable 3pX4 kernel on Altivec: despite this platform has 32 registers, this version seems significantly slower.

Browse Source
This commit is contained in:
Gael Guennebaud 2014-04-25 11:48:22 +02:00
parent ae4d9434e2
commit c9788d55b9
1 changed files with 13 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
Eigen/src/Core/products/GeneralBlockPanelKernel.h
View File

@ -188,7 +188,7 @@ public:
nr = 4, nr = 4,
// register block size along the M direction (currently, this one cannot be modified) // register block size along the M direction (currently, this one cannot be modified)
#ifdef EIGEN_HAS_FUSED_MADD #if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC)
// we assume 16 registers // we assume 16 registers
mr = 3*LhsPacketSize, mr = 3*LhsPacketSize,
#else #else
@ -296,7 +296,7 @@ public:
NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS, NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
nr = 4, nr = 4,
#ifdef EIGEN_HAS_FUSED_MADD #if defined(EIGEN_HAS_FUSED_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC)
// we assume 16 registers // we assume 16 registers
mr = 3*LhsPacketSize, mr = 3*LhsPacketSize,
#else #else
@ -759,29 +759,29 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,mr,nr,ConjugateLhs,ConjugateRhs>
for(Index k=0; k<peeled_kc; k+=pk) for(Index k=0; k<peeled_kc; k+=pk)
{ {
EIGEN_ASM_COMMENT("begin gegp micro kernel 3p x 4"); EIGEN_ASM_COMMENT("begin gegp micro kernel 3p x 4");
RhsPacket B_0; RhsPacket B_0, T0;
LhsPacket A2; LhsPacket A2;
#define EIGEN_GEBGP_ONESTEP(K) \ #define EIGEN_GEBGP_ONESTEP(K) \
internal::prefetch(blA+(3*K+16)*LhsProgress); \ internal::prefetch(blA+(3*K+16)*LhsProgress); \
traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0); \ traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0); \
traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1); \ traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1); \
traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2); \ traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2); \
traits.loadRhs(&blB[(0+4*K)*RhsProgress], B_0); \ traits.loadRhs(&blB[(0+4*K)*RhsProgress], B_0); \
traits.madd(A0, B_0, C0, B_0); \ traits.madd(A0, B_0, C0, T0); \
traits.madd(A1, B_0, C4, B_0); \ traits.madd(A1, B_0, C4, T0); \
traits.madd(A2, B_0, C8, B_0); \ traits.madd(A2, B_0, C8, B_0); \
traits.loadRhs(&blB[1+4*K*RhsProgress], B_0); \ traits.loadRhs(&blB[1+4*K*RhsProgress], B_0); \
traits.madd(A0, B_0, C1, B_0); \ traits.madd(A0, B_0, C1, T0); \
traits.madd(A1, B_0, C5, B_0); \ traits.madd(A1, B_0, C5, T0); \
traits.madd(A2, B_0, C9, B_0); \ traits.madd(A2, B_0, C9, B_0); \
traits.loadRhs(&blB[2+4*K*RhsProgress], B_0); \ traits.loadRhs(&blB[2+4*K*RhsProgress], B_0); \
traits.madd(A0, B_0, C2, B_0); \ traits.madd(A0, B_0, C2, T0); \
traits.madd(A1, B_0, C6, B_0); \ traits.madd(A1, B_0, C6, T0); \
traits.madd(A2, B_0, C10, B_0); \ traits.madd(A2, B_0, C10, B_0); \
traits.loadRhs(&blB[3+4*K*RhsProgress], B_0); \ traits.loadRhs(&blB[3+4*K*RhsProgress], B_0); \
traits.madd(A0, B_0, C3 , B_0); \ traits.madd(A0, B_0, C3 , T0); \
traits.madd(A1, B_0, C7, B_0); \ traits.madd(A1, B_0, C7, T0); \
traits.madd(A2, B_0, C11, B_0) traits.madd(A2, B_0, C11, B_0)
internal::prefetch(blB+(48+0)); internal::prefetch(blB+(48+0));
@ -802,7 +802,7 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,mr,nr,ConjugateLhs,ConjugateRhs>
// process remaining peeled loop // process remaining peeled loop
for(Index k=peeled_kc; k<depth; k++) for(Index k=peeled_kc; k<depth; k++)
{ {
RhsPacket B_0; RhsPacket B_0, T0;
LhsPacket A2; LhsPacket A2;
EIGEN_GEBGP_ONESTEP(0); EIGEN_GEBGP_ONESTEP(0);
blB += 4*RhsProgress; blB += 4*RhsProgress;