# HG changeset patch
# User Benoit Steiner <benoit.steiner.goog@gmail.com>
# Date 1387306183 28800
# Node ID 58831f55ae353766d70b656254cf0dbdab885926
# Parent  2e597973fede8cd01abefd5857d1dc2b6a01371e
Use SSE instructions to pack the rhs matrix whenever possible.

diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -1252,16 +1252,17 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Sca
     if(PanelMode) count += (stride-offset-depth);
   }
 }
 
 // this version is optimized for row major matrices
 template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
 struct gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
 {
+  typedef typename packet_traits<Scalar>::type Packet;
   enum { PacketSize = packet_traits<Scalar>::size };
   EIGEN_DONT_INLINE void operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride=0, Index offset=0);
 };
 
 template<typename Scalar, typename Index, int nr, bool Conjugate, bool PanelMode>
 EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, nr, RowMajor, Conjugate, PanelMode>
   ::operator()(Scalar* blockB, const Scalar* rhs, Index rhsStride, Index depth, Index cols, Index stride, Index offset)
 {
@@ -1271,22 +1272,28 @@ EIGEN_DONT_INLINE void gemm_pack_rhs<Sca
   Index packet_cols = (cols/nr) * nr;
   Index count = 0;
   for(Index j2=0; j2<packet_cols; j2+=nr)
   {
     // skip what we have before
     if(PanelMode) count += nr * offset;
     for(Index k=0; k<depth; k++)
     {
-      const Scalar* b0 = &rhs[k*rhsStride + j2];
-                blockB[count+0] = cj(b0[0]);
-                blockB[count+1] = cj(b0[1]);
-      if(nr==4) blockB[count+2] = cj(b0[2]);
-      if(nr==4) blockB[count+3] = cj(b0[3]);
-      count += nr;
+      if (nr == PacketSize) {
+        Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]);
+        pstoreu(blockB+count, cj.pconj(A));
+        count += PacketSize;
+      } else {
+        const Scalar* b0 = &rhs[k*rhsStride + j2];
+                  blockB[count+0] = cj(b0[0]);
+                  blockB[count+1] = cj(b0[1]);
+        if(nr==4) blockB[count+2] = cj(b0[2]);
+        if(nr==4) blockB[count+3] = cj(b0[3]);
+        count += nr;
+      }
     }
     // skip what we have after
     if(PanelMode) count += nr * (stride-offset-depth);
   }
   // copy the remaining columns one at a time (nr==1)
   for(Index j2=packet_cols; j2<cols; ++j2)
   {
     if(PanelMode) count += offset;