Fix excessive GEBP register spilling for 32-bit NEON.

Clang does a poor job of optimizing the GEBP microkernel on 32-bit ARM,
leading to excessive 16-byte register spills, slowing down basic f32
matrix multiplication by approx 50%.

By specializing `gebp_traits`, we can eliminate the register spills.
Volatile inline ASM both acts as a barrier to prevent reordering and
enforces strict register use. In a simple f32 matrix multiply example,
this modification reduces 16-byte spills from 109 instances to zero,
leading to a 1.5x speed increase (search for `16-byte Spill` in the
assembly in https://godbolt.org/z/chsPbE).

This is a replacement of !379.  See there for further discussion.

Also moved `gebp_traits` specializations for NEON to
`Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h` to be alongside
other NEON-specific code.

Fixes #2138.

1 files changed, 183 insertions, 0 deletions

diff --git a/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h b/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
new file mode 100644
index 000000000..3481f337e
--- /dev/null
+++ b/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
@@ -0,0 +1,183 @@
+namespace Eigen {
+namespace internal {
+  
+#if EIGEN_ARCH_ARM && EIGEN_COMP_CLANG
+
+// 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.
+template<>
+struct gebp_traits <float,float,false,false,Architecture::NEON,GEBPPacketFull>
+ : gebp_traits<float,float,false,false,Architecture::Generic,GEBPPacketFull>
+{
+  EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const
+  { 
+    // This volatile inline ASM both acts as a barrier to prevent reordering,
+    // as well as enforces strict register use.
+    asm volatile(
+      "vmla.f32 %q[r], %q[c], %q[alpha]"
+      : [r] "+w" (r)
+      : [c] "w" (c),
+        [alpha] "w" (alpha)
+      : );
+  }
+
+  template <typename LaneIdType>
+  EIGEN_STRONG_INLINE void madd(const Packet4f& a, const Packet4f& b,
+                                Packet4f& c, Packet4f& tmp,
+                                const LaneIdType&) const {
+    acc(a, b, c);
+  }
+  
+  template <typename LaneIdType>
+  EIGEN_STRONG_INLINE void madd(const Packet4f& a, const QuadPacket<Packet4f>& b,
+                                Packet4f& c, Packet4f& tmp,
+                                const LaneIdType& lane) const {
+    madd(a, b.get(lane), c, tmp, lane);
+  }
+};
+
+#endif // EIGEN_ARCH_ARM && EIGEN_COMP_CLANG
+
+#if EIGEN_ARCH_ARM64
+
+template<>
+struct gebp_traits <float,float,false,false,Architecture::NEON,GEBPPacketFull>
+ : gebp_traits<float,float,false,false,Architecture::Generic,GEBPPacketFull>
+{
+  typedef float RhsPacket;
+  typedef float32x4_t RhsPacketx4;
+
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
+  {
+    dest = *b;
+  }
+
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const
+  {
+    dest = vld1q_f32(b);
+  }
+
+  EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacket& dest) const
+  {
+    dest = *b;
+  }
+
+  EIGEN_STRONG_INLINE void updateRhs(const RhsScalar*, RhsPacketx4&) const
+  {}
+
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
+  {
+    loadRhs(b,dest);
+  }
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const
+  {
+    c = vfmaq_n_f32(c, a, b);
+  }
+
+  // NOTE: Template parameter inference failed when compiled with Android NDK:
+  // "candidate template ignored: could not match 'FixedInt<N>' against 'Eigen::internal::FixedInt<0>".
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const
+  { madd_helper<0>(a, b, c); }
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<1>&) const
+  { madd_helper<1>(a, b, c); }
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<2>&) const
+  { madd_helper<2>(a, b, c); }
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<3>&) const
+  { madd_helper<3>(a, b, c); }
+
+ private:
+  template<int LaneID>
+  EIGEN_STRONG_INLINE void madd_helper(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c) const
+  {
+    #if EIGEN_COMP_GNUC_STRICT && !(EIGEN_GNUC_AT_LEAST(9,0))
+    // workaround gcc issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89101
+    // vfmaq_laneq_f32 is implemented through a costly dup
+         if(LaneID==0)  asm("fmla %0.4s, %1.4s, %2.s[0]\n" : "+w" (c) : "w" (a), "w" (b) :  );
+    else if(LaneID==1)  asm("fmla %0.4s, %1.4s, %2.s[1]\n" : "+w" (c) : "w" (a), "w" (b) :  );
+    else if(LaneID==2)  asm("fmla %0.4s, %1.4s, %2.s[2]\n" : "+w" (c) : "w" (a), "w" (b) :  );
+    else if(LaneID==3)  asm("fmla %0.4s, %1.4s, %2.s[3]\n" : "+w" (c) : "w" (a), "w" (b) :  );
+    #else
+    c = vfmaq_laneq_f32(c, a, b, LaneID);
+    #endif
+  }
+};
+
+
+template<>
+struct gebp_traits <double,double,false,false,Architecture::NEON>
+ : gebp_traits<double,double,false,false,Architecture::Generic>
+{
+  typedef double RhsPacket;
+
+  struct RhsPacketx4 {
+    float64x2_t B_0, B_1;
+  };
+
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
+  {
+    dest = *b;
+  }
+
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketx4& dest) const
+  {
+    dest.B_0 = vld1q_f64(b);
+    dest.B_1 = vld1q_f64(b+2);
+  }
+
+  EIGEN_STRONG_INLINE void updateRhs(const RhsScalar* b, RhsPacket& dest) const
+  {
+    loadRhs(b,dest);
+  }
+
+  EIGEN_STRONG_INLINE void updateRhs(const RhsScalar*, RhsPacketx4&) const
+  {}
+
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
+  {
+    loadRhs(b,dest);
+  }
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const
+  {
+    c = vfmaq_n_f64(c, a, b);
+  }
+
+  // NOTE: Template parameter inference failed when compiled with Android NDK:
+  // "candidate template ignored: could not match 'FixedInt<N>' against 'Eigen::internal::FixedInt<0>".
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const
+  { madd_helper<0>(a, b, c); }
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<1>&) const
+  { madd_helper<1>(a, b, c); }
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<2>&) const
+  { madd_helper<2>(a, b, c); }
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<3>&) const
+  { madd_helper<3>(a, b, c); }
+
+ private:
+  template <int LaneID>
+  EIGEN_STRONG_INLINE void madd_helper(const LhsPacket& a, const RhsPacketx4& b, AccPacket& c) const
+  {
+    #if EIGEN_COMP_GNUC_STRICT && !(EIGEN_GNUC_AT_LEAST(9,0))
+    // workaround gcc issue https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89101
+    // vfmaq_laneq_f64 is implemented through a costly dup
+         if(LaneID==0)  asm("fmla %0.2d, %1.2d, %2.d[0]\n" : "+w" (c) : "w" (a), "w" (b.B_0) :  );
+    else if(LaneID==1)  asm("fmla %0.2d, %1.2d, %2.d[1]\n" : "+w" (c) : "w" (a), "w" (b.B_0) :  );
+    else if(LaneID==2)  asm("fmla %0.2d, %1.2d, %2.d[0]\n" : "+w" (c) : "w" (a), "w" (b.B_1) :  );
+    else if(LaneID==3)  asm("fmla %0.2d, %1.2d, %2.d[1]\n" : "+w" (c) : "w" (a), "w" (b.B_1) :  );
+    #else
+         if(LaneID==0) c = vfmaq_laneq_f64(c, a, b.B_0, 0);
+    else if(LaneID==1) c = vfmaq_laneq_f64(c, a, b.B_0, 1);
+    else if(LaneID==2) c = vfmaq_laneq_f64(c, a, b.B_1, 0);
+    else if(LaneID==3) c = vfmaq_laneq_f64(c, a, b.B_1, 1);
+    #endif
+  }
+};
+
+#endif // EIGEN_ARCH_ARM64
+
+}  // namespace internal
+}  // namespace Eigen