Merge release-4-6 into master

[alexxy/gromacs.git] / src / gromacs / gmxlib / nonbonded / nb_kernel_avx_256_double / nb_kernel_ElecEw_VdwLJ_GeomW3P1_avx_256_double.c

diff --git a/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_VdwLJ_GeomW3P1_avx_256_double.c b/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_VdwLJ_GeomW3P1_avx_256_double.c
index 5dac58b7137c9320288aed35c455564f11af585c..0d88493ded5973e053c072b0da6f47d7c4032b7c 100644 (file)
--- a/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_VdwLJ_GeomW3P1_avx_256_double.c
+++ b/src/gromacs/gmxlib/nonbonded/nb_kernel_avx_256_double/nb_kernel_ElecEw_VdwLJ_GeomW3P1_avx_256_double.c
@@ -401,7 +401,7 @@ nb_kernel_ElecEw_VdwLJ_GeomW3P1_VF_avx_256_double
              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
              */
-            tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
+            tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 
             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));
@@ -988,7 +988,7 @@ nb_kernel_ElecEw_VdwLJ_GeomW3P1_F_avx_256_double
              * This mask will be 0xFFFFFFFF for dummy entries and 0x0 for real ones,
              * so use it as val = _mm_andnot_pd(mask,val) to clear dummy entries.
              */
-            tmpmask0 = gmx_mm_castsi128_pd(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
+            tmpmask0 = gmx_mm_castsi128_ps(_mm_cmplt_epi32(_mm_loadu_si128((const __m128i *)(jjnr+jidx)),_mm_setzero_si128()));
 
             tmpmask1 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(3,3,2,2));
             tmpmask0 = _mm_permute_ps(tmpmask0,_GMX_MM_PERMUTE(1,1,0,0));