Merge release-5-0 into master

Change-Id: Ifc6c3180fe5803bc9a23a0be84be7c93ecb7f472

diff --git a/src/gromacs/mdlib/nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn_inner.h b/src/gromacs/mdlib/nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn_inner.h
index 5663260589ca2f8af2c509f58cbdc615503d8b1e..99385ff3cac17b1fe9e1fc6cb3210b588c870f0f 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn_inner.h
+++ b/src/gromacs/mdlib/nbnxn_kernels/simd_2xnn/nbnxn_kernel_simd_2xnn_inner.h
@@ -65,7 +65,7 @@
 /* Without exclusions and energies we only need to mask the cut-off,
  * this can be faster with blendv.
  */
-#if !(defined CHECK_EXCLS || defined CALC_ENERGIES) && defined GMX_SIMD_HAVE_BLENDV && !defined COUNT_PAIRS
+#if !(defined CHECK_EXCLS || defined CALC_ENERGIES || defined LJ_EWALD_GEOM) && defined GMX_SIMD_HAVE_BLENDV
 /* With RF and tabulated Coulomb we replace cmp+and with sub+blendv.
  * With gcc this is slower, except for RF on Sandy Bridge.
  * Tested with gcc 4.6.2, 4.6.3 and 4.7.1.
@@ -328,10 +328,10 @@
         tmp = gmx_simd_align_r(tmpa);
         for (i = 0; i < UNROLLI; i += 2)
         {
-            gmx_simd_store_r(tmp, i == 0 ? wco_S0 : wco_S2);
+            gmx_simd_store_r(tmp, gmx_simd_sub_r(rc2_S, i == 0 ? rsq_S0 : rsq_S2));
             for (j = 0; j < 2*UNROLLJ; j++)
             {
-                if (!(tmp[j] == 0))
+                if (tmp[j] >= 0)
                 {
                     npair++;
                 }
@@ -780,9 +780,10 @@
 #endif
 #endif
 
-        cr2_S0        = gmx_simd_mul_r(lje_c2_S, rsq_S0);
+        /* Mask for the cut-off to avoid overflow in gmx_simd_exp_r */
+        cr2_S0        = gmx_simd_mul_r(lje_c2_S, gmx_simd_blendzero_r(rsq_S0, wco_vdw_S0));
 #ifndef HALF_LJ
-        cr2_S2        = gmx_simd_mul_r(lje_c2_S, rsq_S2);
+        cr2_S2        = gmx_simd_mul_r(lje_c2_S, gmx_simd_blendzero_r(rsq_S2, wco_vdw_S2));
 #endif
         expmcr2_S0    = gmx_simd_exp_r(gmx_simd_mul_r(mone_S, cr2_S0));
 #ifndef HALF_LJ

diff --git a/src/gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h b/src/gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h
index 4b7de1678fae416dfdf61dc275fa3f199342375f..c567589ed5319c66148843e9195a12b913ef6fea 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h
+++ b/src/gromacs/mdlib/nbnxn_kernels/simd_4xn/nbnxn_kernel_simd_4xn_inner.h
@@ -53,7 +53,7 @@
  * this can be faster when we have defined gmx_simd_blendv_r, i.e. an instruction
  * that selects from two SIMD registers based on the contents of a third.
  */
-#if !(defined CHECK_EXCLS || defined CALC_ENERGIES) && defined GMX_SIMD_HAVE_BLENDV
+#if !(defined CHECK_EXCLS || defined CALC_ENERGIES || defined LJ_EWALD_GEOM) && defined GMX_SIMD_HAVE_BLENDV
 /* With RF and tabulated Coulomb we replace cmp+and with sub+blendv.
  * With gcc this is slower, except for RF on Sandy Bridge.
  * Tested with gcc 4.6.2, 4.6.3 and 4.7.1.
@@ -1014,11 +1014,12 @@
 #endif
 #endif
 
-        cr2_S0        = gmx_simd_mul_r(lje_c2_S, rsq_S0);
-        cr2_S1        = gmx_simd_mul_r(lje_c2_S, rsq_S1);
+        /* Mask for the cut-off to avoid overflow in gmx_simd_exp_r */
+        cr2_S0        = gmx_simd_mul_r(lje_c2_S, gmx_simd_blendzero_r(rsq_S0, wco_vdw_S0));
+        cr2_S1        = gmx_simd_mul_r(lje_c2_S, gmx_simd_blendzero_r(rsq_S1, wco_vdw_S1));
 #ifndef HALF_LJ
-        cr2_S2        = gmx_simd_mul_r(lje_c2_S, rsq_S2);
-        cr2_S3        = gmx_simd_mul_r(lje_c2_S, rsq_S3);
+        cr2_S2        = gmx_simd_mul_r(lje_c2_S, gmx_simd_blendzero_r(rsq_S2, wco_vdw_S2));
+        cr2_S3        = gmx_simd_mul_r(lje_c2_S, gmx_simd_blendzero_r(rsq_S3, wco_vdw_S3));
 #endif
         expmcr2_S0    = gmx_simd_exp_r(gmx_simd_mul_r(mone_S, cr2_S0));
         expmcr2_S1    = gmx_simd_exp_r(gmx_simd_mul_r(mone_S, cr2_S1));

diff --git a/src/gromacs/mdlib/nbnxn_search.c b/src/gromacs/mdlib/nbnxn_search.c
index 0ee8b4d4d09303d14c4b06c927caf3ddc9227749..0826017bab62cad0285c13546270b01a9f8b85a8 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_search.c
+++ b/src/gromacs/mdlib/nbnxn_search.c
@@ -416,6 +416,12 @@ static real grid_atom_density(int n, rvec corner0, rvec corner1)
 {
     rvec size;
 
+    if (n == 0)
+    {
+        /* To avoid zero density we use a minimum of 1 atom */
+        n = 1;
+    }
+
     rvec_sub(corner1, corner0, size);
 
     return n/(size[XX]*size[YY]*size[ZZ]);
@@ -436,6 +442,8 @@ static int set_grid_size_xy(const nbnxn_search_t nbs,
 
     if (n > grid->na_sc)
     {
+        assert(atom_density > 0);
+
         /* target cell length */
         if (grid->bSimple)
         {
@@ -1810,7 +1818,8 @@ void nbnxn_put_on_grid(nbnxn_search_t nbs,
         nbs->ePBC = ePBC;
         copy_mat(box, nbs->box);
 
-        if (atom_density >= 0)
+        /* Avoid zero density */
+        if (atom_density > 0)
         {
             grid->atom_density = atom_density;
         }
@@ -1826,12 +1835,21 @@ void nbnxn_put_on_grid(nbnxn_search_t nbs,
          * for the local atoms (dd_zone=0).
          */
         nbs->natoms_nonlocal = a1 - nmoved;
+
+        if (debug)
+        {
+            fprintf(debug, "natoms_local = %5d atom_density = %5.1f\n",
+                    nbs->natoms_local, grid->atom_density);
+        }
     }
     else
     {
         nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1);
     }
 
+    /* We always use the home zone (grid[0]) for setting the cell size,
+     * since determining densities for non-local zones is difficult.
+     */
     nc_max_grid = set_grid_size_xy(nbs, grid,
                                    dd_zone, n-nmoved, corner0, corner1,
                                    nbs->grid[0].atom_density);

diff --git a/src/gromacs/mdlib/nbnxn_search.h b/src/gromacs/mdlib/nbnxn_search.h
index 12576a096c94e4f235fc01abc1359478753613fa..6b3ab7c8d2c8744dad189d8c331cfbc02b14c79a 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_search.h
+++ b/src/gromacs/mdlib/nbnxn_search.h
@@ -65,7 +65,7 @@ void nbnxn_init_search(nbnxn_search_t    * nbs_ptr,
 /* Put the atoms on the pair search grid.
  * Only atoms a0 to a1 in x are put on the grid.
  * The atom_density is used to determine the grid size.
- * When atom_density=-1, the density is determined from a1-a0 and the corners.
+ * When atom_density<=0, the density is determined from a1-a0 and the corners.
  * With domain decomposition part of the n particles might have migrated,
  * but have not been removed yet. This count is given by nmoved.
  * When move[i] < 0 particle i has migrated and will not be put on the grid.