Merge release-4-6 into master

[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_sse2_single / nb_kernel_ElecGB_VdwLJ_GeomP1P1_sse2_single.c

diff --git a/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecGB_VdwLJ_GeomP1P1_sse2_single.c b/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecGB_VdwLJ_GeomP1P1_sse2_single.c
index 4b8297188884489b5c5e86fc747833e0d359a150..b2c7e57c5ffd6dc20b31b5a6c8023bbf09a535b8 100644 (file)
--- a/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecGB_VdwLJ_GeomP1P1_sse2_single.c
+++ b/src/gromacs/gmxlib/nonbonded/nb_kernel_sse2_single/nb_kernel_ElecGB_VdwLJ_GeomP1P1_sse2_single.c
@@ -375,6 +375,7 @@ nb_kernel_ElecGB_VdwLJ_GeomP1P1_VF_sse2_single
             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
+            dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;
@@ -780,6 +781,7 @@ nb_kernel_ElecGB_VdwLJ_GeomP1P1_F_sse2_single
             FF               = _mm_add_ps(Fp,_mm_mul_ps(gbeps,_mm_add_ps(G,_mm_add_ps(Heps,Heps))));
             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
             dvdatmp          = _mm_mul_ps(minushalf,_mm_add_ps(vgb,_mm_mul_ps(fgb,r00)));
+            dvdatmp          = _mm_andnot_ps(dummy_mask,dvdatmp);
             dvdasum          = _mm_add_ps(dvdasum,dvdatmp);
             /* The pointers to scratch make sure that this code with compilers that take gmx_restrict seriously (e.g. icc 13) really can't screw things up. */
             fjptrA             = (jnrlistA>=0) ? dvda+jnrA : scratch;