Changeset 16430

Timestamp:

10/16/13 11:14:22 (11 years ago)

Author:

Eric.Larour

Message:

CHG: improved speed on assemble routine.

File:

• issm/trunk-jpl/src/c/toolkits/issm/IssmMpiSparseMat.h (modified) (3 diffs)

issm/trunk-jpl/src/c/toolkits/issm/IssmMpiSparseMat.h

@@ -194 +194 @@
                         int*          numvalues_perrow    = NULL;
                         int           row;
+                        
+                        doubletype**  values_perrow       = NULL;
+                        int**         cols_perrow         = NULL;
+                        int**         mods_perrow         = NULL;
+                        int*          counters_perrow     = NULL;
+                        int           counter;
 
                         /*Early exit: */
@@ -284 +290 @@
                         }
 
-                        
-                        /*Deal with sparse row one at a time: */
-                        for(row=0;row<this->m;row++){
-
-                                /*Allocate: */
-                                this_row_numvalues=numvalues_perrow[row];
-                                this_row_cols=xNew<int>(this_row_numvalues);
-                                this_row_values=xNew<doubletype>(this_row_numvalues);
-                                this_row_mods=xNew<int>(this_row_numvalues);
-
-                                /*Collect: */
-                                count=0;
-                                for(i=0;i<num_procs;i++){
-                                        int  numvalues=numvalues_fromcpu[i];
-                                        int* rows=row_indices_fromcpu[i];
-                                        int* cols=col_indices_fromcpu[i];
-                                        doubletype* values=values_fromcpu[i];
-                                        int* mods=modes_fromcpu[i];
-
-                                        for(j=0;j<numvalues;j++){
-                                                if((rows[j]-lower_row)==row){
-                                                        this_row_cols[count]=cols[j];
-                                                        this_row_values[count]=values[j];
-                                                        this_row_mods[count]=mods[j];
-                                                        count++;
-                                                }
-                                        }
-                                }
-                                if(count!=this_row_numvalues)_error_("counter error during assembly of values!");
-
-                                /*Now fill in the sparse row: */
-                                this->matrix[row]->SetValues(this_row_numvalues,this_row_cols,this_row_values,this_row_mods);
-
-                                /*Free ressources: */
-                                xDelete<int>(this_row_cols);
-                                xDelete<doubletype>(this_row_values);
-                                xDelete<int>(this_row_mods);
+                        /*Allocate all the values, cols and mods from each cpu: */
+                        values_perrow=xNew<doubletype*>(this->m);
+                        cols_perrow=xNew<int*>(this->m);
+                        mods_perrow=xNew<int*>(this->m);
+                        counters_perrow=xNewZeroInit<int>(this->m);
+
+                        for(i=0;i<this->m;i++){
+                                values_perrow[i]=xNewZeroInit<doubletype>(numvalues_perrow[i]);
+                                cols_perrow[i]=xNewZeroInit<int>(numvalues_perrow[i]);
+                                mods_perrow[i]=xNewZeroInit<int>(numvalues_perrow[i]);
+                        }
+
+                        /*collect:*/
+                        for(i=0;i<num_procs;i++){
+                                int  numvalues=numvalues_fromcpu[i];
+                                int* rows=row_indices_fromcpu[i];
+                                int* cols=col_indices_fromcpu[i];
+                                doubletype* values=values_fromcpu[i];
+                                int* mods=modes_fromcpu[i];
+
+                                for(j=0;j<numvalues;j++){
+                                        row=rows[j]-lower_row;
+                                        counter=counters_perrow[row];
+                                        values_perrow[row][counter]=values[j];
+                                        cols_perrow[row][counter]=cols[j];
+                                        mods_perrow[row][counter]=mods[j];
+                                        counter=counters_perrow[row]++;
+                                }
+                        }
+                                        
+                        /*Plug into matrix: */
+                        for(i=0;i<this->m;i++){
+                                this->matrix[i]->SetValues(numvalues_perrow[i],cols_perrow[i],values_perrow[i],mods_perrow[i]);
                         }
                         /*}}}*/
@@ -358 +360 @@
                         xDelete<int>(sendcnts);
                         xDelete<int>(displs);
+                        
+                        for(i=0;i<this->m;i++){
+                                doubletype* values=values_perrow[i]; xDelete<doubletype>(values);
+                                int* cols=cols_perrow[i]; xDelete<int>(cols);
+                                int* mods=mods_perrow[i]; xDelete<int>(mods);
+                        }
+                        xDelete<int>(counters_perrow);
+                        xDelete<doubletype*>(values_perrow);
+                        xDelete<int*>(cols_perrow);
+                        xDelete<int*>(mods_perrow);
                         /*}}}*/