Re-enable Arm Neon Eigen::half packets of size 8

- Add predux_half_dowto4
- Remove explicit casts in Half.h to match the behaviour of BFloat16.h
- Enable more packetmath tests for Eigen::half

2 files changed, 26 insertions, 49 deletions

diff --git a/Eigen/src/Core/arch/Default/Half.h b/Eigen/src/Core/arch/Default/Half.h
index fda38bcb0..91d3bc51b 100644
--- a/Eigen/src/Core/arch/Default/Half.h
+++ b/Eigen/src/Core/arch/Default/Half.h
@@ -159,6 +159,10 @@ struct half : public half_impl::half_base {
   explicit EIGEN_DEVICE_FUNC half(std::complex<RealScalar> c)
       : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(c.real()))) {}
 
+   EIGEN_DEVICE_FUNC operator float() const {  // NOLINT: Allow implicit conversion to float, because it is lossless.
+    return half_impl::half_to_float(*this);
+  }
+
   EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const {
     // +0.0 and -0.0 become false, everything else becomes true.
   #if defined(EIGEN_HAS_ARM64_FP16_SCALAR_ARITHMETIC)
@@ -167,47 +171,6 @@ struct half : public half_impl::half_base {
     return (x & 0x7fff) != 0;
   #endif
   }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const {
-    return static_cast<signed char>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const {
-    return static_cast<unsigned char>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const {
-    return static_cast<short>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(numext::uint16_t) const {
-    return static_cast<numext::uint16_t>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const {
-    return static_cast<int>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const {
-    return static_cast<unsigned int>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const {
-    return static_cast<long>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const {
-    return static_cast<unsigned long>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const {
-    return static_cast<long long>(half_impl::half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const {
-    return static_cast<unsigned long long>(half_to_float(*this));
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const {
-    return half_impl::half_to_float(*this);
-  }
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const {
-    return static_cast<double>(half_impl::half_to_float(*this));
-  }
-
-  template<typename RealScalar>
-  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(std::complex<RealScalar>) const {
-    return std::complex<RealScalar>(static_cast<RealScalar>(*this), RealScalar(0));
-  }
 };
 
 } // end namespace Eigen
@@ -686,6 +649,12 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) {
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
   return half(::tanhf(float(a)));
 }
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half asin(const half& a) {
+  return half(::asinf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half acos(const half& a) {
+  return half(::acosf(float(a)));
+}
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
 #if (EIGEN_CUDA_SDK_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300) || \
   defined(EIGEN_HIP_DEVICE_COMPILE)
@@ -694,6 +663,9 @@ EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
   return half(::floorf(float(a)));
 #endif
 }
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half rint(const half& a) {
+  return half(::rintf(float(a)));
+}
 EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
 #if (EIGEN_CUDA_SDK_VER >= 80000 && defined EIGEN_CUDA_ARCH && EIGEN_CUDA_ARCH >= 300) || \
   defined(EIGEN_HIP_DEVICE_COMPILE)
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 30edd7097..b13cbe5ec 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -3849,16 +3849,15 @@ template<> EIGEN_STRONG_INLINE Packet2d psqrt(const Packet2d& _x){ return vsqrtq
 typedef float16x4_t Packet4hf;
 typedef float16x8_t Packet8hf;
 
-// TODO(tellenbach): Enable packets of size 8 as soon as the GEBP can handle them
 template <>
 struct packet_traits<Eigen::half> : default_packet_traits {
-  typedef Packet4hf type;
+  typedef Packet8hf type;
   typedef Packet4hf half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
-    size = 4,
-    HasHalfPacket = 0,
+    size = 8,
+    HasHalfPacket = 1,
 
     HasCmp = 1,
     HasCast = 1,
@@ -3904,7 +3903,7 @@ struct unpacket_traits<Packet4hf> {
 template <>
 struct unpacket_traits<Packet8hf> {
   typedef Eigen::half type;
-  typedef Packet8hf half;
+  typedef Packet4hf half;
   enum {
     size = 8,
     alignment = Aligned16,
@@ -3914,6 +3913,11 @@ struct unpacket_traits<Packet8hf> {
   };
 };
 
+template<>
+EIGEN_DEVICE_FUNC Packet4hf predux_half_dowto4<Packet8hf>(const Packet8hf& a) {
+  return vadd_f16(vget_low_f16(a), vget_high_f16(a));
+}
+
 template <>
 EIGEN_STRONG_INLINE Packet8hf pset1<Packet8hf>(const Eigen::half& from) {
   return vdupq_n_f16(from.x);
@@ -4418,7 +4422,8 @@ EIGEN_STRONG_INLINE Eigen::half predux_max<Packet4hf>(const Packet4hf& a) {
   return h;
 }
 
-EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel) {
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel)
+{
   EIGEN_ALIGN16 Eigen::half in[4][8];
 
   pstore<Eigen::half>(in[0], kernel.packet[0]);
@@ -4432,11 +4437,11 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8hf, 4>& kernel) {
   for (int i = 0; i < 4; ++i) {
     EIGEN_UNROLL_LOOP
     for (int j = 0; j < 4; ++j) {
-      out[i][j] = in[j][2*i];
+      out[i][j] = in[j][2 * i];
     }
     EIGEN_UNROLL_LOOP
     for (int j = 0; j < 4; ++j) {
-      out[i][j+4] = in[j][2*i+1];
+      out[i][j + 4] = in[j][2 * i + 1];
     }
   }