commit e2ea866515fc35c835a191ad0d1916dd2df4ed2f
author Rasmus Munk Larsen <rmlarsen@google.com> Thu Sep 22 23:56:34 2022 +0000
committer Rasmus Munk Larsen <rmlarsen@google.com> Thu Sep 22 23:56:34 2022 +0000
tree eff7b1f538a045e36940876da6656f380dd2290d
parent 23299632c246b77937fb78e8607863a2f02e191b

Add a macro to set the nr trait in the BEBP kernel for NEON.

Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h

diff --git a/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h b/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
index 5022205..1d5beb9 100644
--- a/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
@@ -5,6 +5,10 @@
   
 #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.
@@ -49,11 +53,8 @@
 {
   typedef float RhsPacket;
   typedef float32x4_t RhsPacketx4;
-  enum {
-    nr = 8
-  };
-  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
-  {
+  enum { nr = EIGEN_NEON_GEBP_NR };
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const {
     dest = *b;
   }
 
@@ -115,9 +116,7 @@
  : gebp_traits<double,double,false,false,Architecture::Generic>
 {
   typedef double RhsPacket;
-  enum {
-    nr = 8
-  };
+  enum { nr = EIGEN_NEON_GEBP_NR };
   struct RhsPacketx4 {
     float64x2_t B_0, B_1;
   };
@@ -193,9 +192,8 @@
   typedef half RhsPacket;
   typedef float16x4_t RhsPacketx4;
   typedef float16x4_t PacketHalf;
-  enum {
-    nr = 8
-  };
+  enum { nr = EIGEN_NEON_GEBP_NR };
+
   EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
   {
     dest = *b;