Merge release-4-6 into master

[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse2_double / nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse2_double.c

diff --git a/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse2_double.c b/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse2_double.c
index a18d1c18495a1a7b68f3c5fc70515194a566fa7c..d0aa93072ab6a152315ae762952e87eb3a7d6015 100644 (file)
--- a/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse2_double.c
+++ b/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_double/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_sse2_double.c
@@ -354,6 +354,7 @@ nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_sse2_double
             FF               = _mm_add_pd(Fp,_mm_mul_pd(gbeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
             dvdatmp          = _mm_mul_pd(minushalf,_mm_add_pd(vgb,_mm_mul_pd(fgb,r00)));
+            dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
             velec            = _mm_mul_pd(qq00,rinv00);
@@ -750,6 +751,7 @@ nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_sse2_double
             FF               = _mm_add_pd(Fp,_mm_mul_pd(gbeps,_mm_add_pd(G,_mm_add_pd(Heps,Heps))));
             fgb              = _mm_mul_pd(gbqqfactor,_mm_mul_pd(FF,gbscale));
             dvdatmp          = _mm_mul_pd(minushalf,_mm_add_pd(vgb,_mm_mul_pd(fgb,r00)));
+            dvdatmp          = _mm_unpacklo_pd(dvdatmp,_mm_setzero_pd());
             dvdasum          = _mm_add_pd(dvdasum,dvdatmp);
             gmx_mm_increment_1real_pd(dvda+jnrA,_mm_mul_pd(dvdatmp,_mm_mul_pd(isaj0,isaj0)));
             velec            = _mm_mul_pd(qq00,rinv00);