cpu - skip duplicate output rstr

CEED · Aug 20, 2024 · 49a690b · 49a690b
1 parent d56950a
commit 49a690b
Showing 1 changed file with 43 additions and 0 deletions.
diff --git a/backends/opt/ceed-opt-operator.c b/backends/opt/ceed-opt-operator.c
@@ -182,6 +182,49 @@ static int CeedOperatorSetupFields_Opt(CeedQFunction qf, CeedOperator op, bool i
       }
     }
   }
+  // Drop duplicate restrictions
+  if (is_input) {
+    for (CeedInt i = 0; i < num_fields; i++) {
+      CeedVector          vec_i;
+      CeedElemRestriction rstr_i;
+
+      CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec_i));
+      CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr_i));
+      for (CeedInt j = i + 1; j < num_fields; j++) {
+        CeedVector          vec_j;
+        CeedElemRestriction rstr_j;
+
+        CeedCallBackend(CeedOperatorFieldGetVector(op_fields[j], &vec_j));
+        CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[j], &rstr_j));
+        if (vec_i == vec_j && rstr_i == rstr_j) {
+          CeedCallBackend(CeedVectorReferenceCopy(e_vecs[i], &e_vecs[j]));
+          CeedCallBackend(CeedVectorReferenceCopy(e_vecs_full[i + start_e], &e_vecs_full[j + start_e]));
+          skip_rstr[j] = true;
+        }
+      }
+    }
+  } else {
+    for (CeedInt i = num_fields - 1; i >= 0; i--) {
+      CeedVector          vec_i;
+      CeedElemRestriction rstr_i;
+
+      CeedCallBackend(CeedOperatorFieldGetVector(op_fields[i], &vec_i));
+      CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[i], &rstr_i));
+      for (CeedInt j = i - 1; j >= 0; j--) {
+        CeedVector          vec_j;
+        CeedElemRestriction rstr_j;
+
+        CeedCallBackend(CeedOperatorFieldGetVector(op_fields[j], &vec_j));
+        CeedCallBackend(CeedOperatorFieldGetElemRestriction(op_fields[j], &rstr_j));
+        if (vec_i == vec_j && rstr_i == rstr_j) {
+          CeedCallBackend(CeedVectorReferenceCopy(e_vecs[i], &e_vecs[j]));
+          CeedCallBackend(CeedVectorReferenceCopy(e_vecs_full[i + start_e], &e_vecs_full[j + start_e]));
+          skip_rstr[j]       = true;
+          apply_add_basis[i] = true;
+        }
+      }
+    }
+  }
   return CEED_ERROR_SUCCESS;
 }