Move EIGEN_NEON_GEBP_NR macro to the right place in GeneralBlockPanelKernel.h

This commit is contained in:
Rasmus Munk Larsen 2022-09-23 02:24:27 +00:00
parent e2ea866515
commit ed8cda3ce4

View File

@ -5,10 +5,6 @@ namespace internal {
#if EIGEN_ARCH_ARM && EIGEN_COMP_CLANG
#ifndef EIGEN_NEON_GEBP_NR
#define EIGEN_NEON_GEBP_NR 8
#endif
// Clang seems to excessively spill registers in the GEBP kernel on 32-bit arm.
// Here we specialize gebp_traits to eliminate these register spills.
// See #2138.
@ -47,6 +43,10 @@ struct gebp_traits <float,float,false,false,Architecture::NEON,GEBPPacketFull>
#if EIGEN_ARCH_ARM64
#ifndef EIGEN_NEON_GEBP_NR
#define EIGEN_NEON_GEBP_NR 8
#endif
template<>
struct gebp_traits <float,float,false,false,Architecture::NEON,GEBPPacketFull>
: gebp_traits<float,float,false,false,Architecture::Generic,GEBPPacketFull>