Implement DisceretGen for double in Arm NEON

Author: bab2min

EigenRand/Dists/Discrete.h

@@ -692,8 +692,6 @@ namespace Eigen
 				}
 			}
 
-// Vectorized packetOp for NEON is not supported
-#ifndef EIGEN_VECTORIZE_NEON
 			template<typename Packet, typename Rng>
 			EIGEN_STRONG_INLINE const Packet packetOp(Rng&& rng)
 			{
@@ -778,7 +776,6 @@ namespace Eigen
 	#endif
 				}
 			}
-#endif
 		};
 
 		template<typename> class BinomialGen;
@@ -1946,16 +1943,6 @@ namespace Eigen
 		}
 	}
 
-#ifdef EIGEN_VECTORIZE_NEON
-	namespace internal
-	{
-		template<typename _Scalar, typename Urng, bool _mutable>
-		struct functor_traits<scalar_rng_adaptor<Rand::DiscreteGen<_Scalar, double>, _Scalar, Urng, _mutable> >
-		{
-			enum { Cost = HugeCost, PacketAccess = 0, IsRepeatable = false };
-		};
-	}
-#endif
 }
 
 #endif

EigenRand/arch/NEON/MorePacketMath.h

@@ -307,6 +307,30 @@ namespace Eigen
 			return vld1q_f32(t);
 		}
 
+		template<>
+		EIGEN_STRONG_INLINE Packet2d pgather<Packet4i>(const double* addr, const Packet4i& index, bool upperhalf)
+		{
+			int32_t u[4];
+			vst1q_s32(u, index);
+			double t[2];
+			if (upperhalf)
+			{
+				t[0] = addr[u[2]];
+				t[1] = addr[u[3]];
+			}
+			else
+			{
+				t[0] = addr[u[0]];
+				t[1] = addr[u[1]];
+			}
+			return vld1q_f64(t);
+		}
+
+		EIGEN_STRONG_INLINE Packet4i combine_low32(const Packet4i& a, const Packet4i& b)
+		{
+			return vuzp1q_s32(a, b);
+		}
+
 		template<>
 		EIGEN_STRONG_INLINE int pmovemask<Packet4f>(const Packet4f& a)
 		{