Merge release-4-6 into master

[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_avx_128_fma_single / nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_128_fma_single.c

diff --git a/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_128_fma_single.c b/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_128_fma_single.c
index d2eb10eb9cdb31339bd18725888fe228682651c8..bddbca0d1855559e23d7848e48f330443fa959d1 100644 (file)
--- a/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_128_fma_single.c
+++ b/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_128_fma_single/nb_kernel_ElecGB_VdwCSTab_GeomP1P1_avx_128_fma_single.c
@@ -421,6 +421,7 @@ nb_kernel_ElecGB_VdwCSTab_GeomP1P1_VF_avx_128_fma_single
             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
+            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;
@@ -887,6 +888,7 @@ nb_kernel_ElecGB_VdwCSTab_GeomP1P1_F_avx_128_fma_single
             FF               = _mm_macc_ps(_mm_macc_ps(twogbeps,H,G),gbeps,Fp);
             fgb              = _mm_mul_ps(gbqqfactor,_mm_mul_ps(FF,gbscale));
             dvdatmp          = _mm_mul_ps(minushalf,_mm_macc_ps(fgb,r00,vgb));
+            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;