From b286a1e75c6bd451c27da8c5ebed0b0fb86dfc2a Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Wed, 26 Mar 2014 16:46:36 +0100
Subject: [PATCH] add pbroadcast2/4 generic intrinsics

---
 Eigen/src/Core/GenericPacketMath.h   | 34 ++++++++++++++++++++++++++++
 Eigen/src/Core/arch/SSE/PacketMath.h | 32 ++++++++++++++++++++++++++
 2 files changed, 66 insertions(+)

diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
index 538ab53b2..d07541285 100755
--- a/Eigen/src/Core/GenericPacketMath.h
+++ b/Eigen/src/Core/GenericPacketMath.h
@@ -173,6 +173,40 @@ pset1(const typename unpacket_traits<Packet>::type& a) { return a; }
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
 pload1(const typename unpacket_traits<Packet>::type  *a) { return pset1<Packet>(*a); }
 
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * a2 = pload1(a+2);
+  * a3 = pload1(a+3);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast2
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast4(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1, Packet& a2, Packet& a3)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+  a2 = pload1<Packet>(a+2);
+  a3 = pload1<Packet>(a+3);
+}
+
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast4
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+}
+
 /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */
 template<typename Scalar> inline typename packet_traits<Scalar>::type
 plset(const Scalar& a) { return a; }
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 293fb83e4..9f81a4623 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -391,6 +391,38 @@ template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
   #endif
 }
 
+// with AVX, the default implementations based on pload1 are faster
+#ifndef __AVX__
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec4f_swizzle1(a3, 0,0,0,0);
+  a1 = vec4f_swizzle1(a3, 1,1,1,1);
+  a2 = vec4f_swizzle1(a3, 2,2,2,2);
+  a3 = vec4f_swizzle1(a3, 3,3,3,3);
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+#ifdef EIGEN_VECTORIZE_SSE3
+  a0 = _mm_loaddup_pd(a+0);
+  a1 = _mm_loaddup_pd(a+1);
+  a2 = _mm_loaddup_pd(a+2);
+  a3 = _mm_loaddup_pd(a+3);
+#else
+  a1 = pload<Packet2d>(a);
+  a0 = vec2d_swizzle1(a1, 0,0);
+  a1 = vec2d_swizzle1(a1, 1,1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec2d_swizzle1(a3, 0,0);
+  a3 = vec2d_swizzle1(a3, 1,1);
+#endif
+}
+#endif
+
 EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
 {
   vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));