From 0fa9e4a15cb93fb6a0ef54ada13979d38d120ab0 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Sat, 2 Jul 2016 17:35:08 +0200
Subject: [PATCH] Fix performance regression in dgemm introduced by changeset
 5d51a7f12c69138ed2a43df240bdf27a5313f7ce

---
 Eigen/src/Core/products/GeneralBlockPanelKernel.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index 2d764727b..c66882012 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -1526,12 +1526,12 @@ void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,Conjuga
           const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
 
           // The following piece of code wont work for 512 bit registers
-          // Moreover it assumes that there is a half packet of the same size
+          // Moreover, if LhsProgress==8 it assumes that there is a half packet of the same size
           // as nr (which is currently 4) for the return type.
           typedef typename unpacket_traits<SResPacket>::half SResPacketHalf;
           if ((SwappedTraits::LhsProgress % 4) == 0 &&
               (SwappedTraits::LhsProgress <= 8) &&
-              unpacket_traits<SResPacketHalf>::size==4)
+              (SwappedTraits::LhsProgress!=8 || unpacket_traits<SResPacketHalf>::size==nr))
           {
             SAccPacket C0, C1, C2, C3;
             straits.initAcc(C0);