simd

TREX-CoE · Feb 14, 2024 · be2a719 · be2a719
1 parent 2228ab2
commit be2a719
Show file tree

Hide file tree

Showing 2 changed files with 51 additions and 16 deletions.
diff --git a/org/qmckl_ao.org b/org/qmckl_ao.org
@@ -5882,7 +5882,9 @@ qmckl_compute_ao_value_hpc_gaussian (const qmckl_context context,
             }
           }
 
-IVDEP
+#ifdef HAVE_OPENMP
+#pragma omp simd
+#endif
           for (int64_t iprim = 0 ; iprim < nidx ; ++iprim) {
             exp_mat[iprim] = exp(-ar2[iprim]);
           }
@@ -5899,12 +5901,16 @@ IVDEP
             }
           }
 
-IVDEP
+#ifdef HAVE_OPENMP
+#pragma omp simd
+#endif
           for (int64_t iprim = 0 ; iprim < nidx ; ++iprim) {
             exp_mat_sp[iprim] = expf(-ar2_sp[iprim]);
           }
 
-IVDEP
+#ifdef HAVE_OPENMP
+#pragma omp simd
+#endif
           for (int64_t iprim = 0 ; iprim < nidx ; ++iprim) {
             exp_mat[iprim] = exp_mat_sp[iprim];
           }
@@ -5930,16 +5936,16 @@ IVDEP
 
           const double s1 = ce_mat[ishell-ishell_start];
 
+          if (s1 == 0.0) {
+            continue;
+          }
+
           const int64_t k = ao_index[ishell];
           double* restrict const ao_value_1 = ao_value + ipoint_ao_num + k;
 
           const int32_t l = shell_ang_mom[ishell];
           const int32_t n = lstart[l+1]-lstart[l];
 
-          if (s1 == 0.0) {
-            continue;
-          }
-
           double* restrict poly_vgl_1 = NULL;
           if (nidx > 0) {
             const double* restrict f = ao_factor + k;
@@ -5976,15 +5982,6 @@ IVDEP
               }
               break;
             }
-          } else {
-/*
-#ifdef HAVE_OPENMP
-#pragma omp simd
-#endif
-            for (int64_t il=0 ; il<n ; ++il) {
-              ao_value_1[il] = 0.0;
-            }
-*/
           }
         }
       }
@@ -6685,6 +6682,9 @@ qmckl_compute_ao_vgl_hpc_gaussian (
             }
           }
 
+#ifdef HAVE_OPENMP
+#pragma omp simd
+#endif
           for (int64_t iprim = 0 ; iprim < nidx ; ++iprim) {
             exp_mat[iprim][0] = exp(-ar2[iprim]);
           }
@@ -6702,16 +6702,25 @@ qmckl_compute_ao_vgl_hpc_gaussian (
             }
           }
 
+#ifdef HAVE_OPENMP
+#pragma omp simd
+#endif
           for (int64_t iprim = 0 ; iprim < nidx ; ++iprim) {
             exp_mat_sp[iprim] = expf(-ar2_sp[iprim]);
           }
 
+#ifdef HAVE_OPENMP
+#pragma omp simd
+#endif
           for (int64_t iprim = 0 ; iprim < nidx ; ++iprim) {
             exp_mat[iprim][0] = exp_mat_sp[iprim];
           }
 
         }
 
+#ifdef HAVE_OPENMP
+#pragma omp simd
+#endif
         for (int64_t iprim = 0 ; iprim < nidx ; ++iprim) {
           double f = qmckl_mat(expo_per_nucleus, iprim, inucl) * exp_mat[iprim][0];
           f = -f-f;

diff --git a/org/qmckl_jastrow_champ.org b/org/qmckl_jastrow_champ.org
@@ -6647,6 +6647,19 @@ end function qmckl_compute_jastrow_champ_factor_een_rescaled_e_gl_f
          double* const een_rescaled_e_gl );
     #+end_src
 
+  #+begin_src c :comments org :tangle (eval h_private_func) :noweb yes :exports none
+   qmckl_exit_code qmckl_compute_jastrow_champ_factor_een_rescaled_e_gl_doc (
+         const qmckl_context context,
+         const int64_t walk_num,
+         const int64_t elec_num,
+         const int64_t cord_num,
+         const double rescale_factor_ee,
+         const double* coord_ee,
+         const double* ee_distance,
+         const double* een_rescaled_e,
+         double* const een_rescaled_e_gl );
+ #+end_src
+
 
      #+CALL: generate_c_interface(table=qmckl_factor_een_rescaled_e_gl_args,rettyp=get_value("CRetType"),fname=get_value("Name"))
 
@@ -6692,6 +6705,19 @@ end function qmckl_compute_jastrow_champ_factor_een_rescaled_e_gl_f
     end function qmckl_compute_jastrow_champ_factor_een_rescaled_e_gl_doc
      #+end_src
 
+  #+begin_src c :comments org :tangle (eval h_private_func) :noweb yes :exports none
+   qmckl_exit_code qmckl_compute_jastrow_champ_factor_een_rescaled_e_gl_hpc (
+         const qmckl_context context,
+         const int64_t walk_num,
+         const int64_t elec_num,
+         const int64_t cord_num,
+         const double rescale_factor_ee,
+         const double* coord_ee,
+         const double* ee_distance,
+         const double* een_rescaled_e,
+         double* const een_rescaled_e_gl );
+ #+end_src
+
  #+begin_src c :comments org :tangle (eval c) :noweb yes
 qmckl_exit_code qmckl_compute_jastrow_champ_factor_een_rescaled_e_gl_hpc (
          const qmckl_context context,