Merge pull request #2040 from gforney/PART

speed up particle loading by sorting tags in parallel

Source/smokeview/IOpart.c

@@ -1177,6 +1177,44 @@ void GeneratePartHistograms(void){
   }
 }
 
+#ifdef pp_SORT_TAGS
+/* ------------------ SortPartTags ------------------------ */
+
+void SortPartTags(partdata *parti){
+  int i;
+  part5data *datacopy_local;
+
+  datacopy_local = parti->data5;
+  for(i = 0; i < parti->ntimes; i++){
+    int class_index;
+
+    for(class_index = 0; class_index < parti->nclasses; class_index++){
+      qsort(datacopy_local->sort_tags, ( size_t )datacopy_local->npoints_file, 2*sizeof(int), CompareTags);
+      datacopy_local++;
+    }
+  }
+}
+
+/* ------------------ SortAllPartTags ------------------------ */
+
+void *SortAllPartTags(void *arg){
+  int i;
+
+  INIT_PRINT_TIMER(timer_sortparttags);
+  for(i = 0; i < npartinfo; i++){
+    partdata *parti;
+
+    parti = partinfo + i;
+    if(parti->loaded == 0)continue;
+    SortPartTags(parti);
+  }
+  PRINT_TIMER(timer_sortparttags, "SortPartTags");
+#ifdef pp_SORT_TAGS_BG
+  THREAD_EXIT(sorttags_threads);
+#endif
+}
+#endif
+
 /* ------------------ GetPartData ------------------------ */
 
 void GetPartData(partdata *parti, int nf_all_arg, FILE_SIZE *file_size_arg){
@@ -1306,7 +1344,9 @@ void GetPartData(partdata *parti, int nf_all_arg, FILE_SIZE *file_size_arg){
             sort_tags_local[2*j]=datacopy_local->tags[j];
             sort_tags_local[2*j+1]=j;
           }
+#ifndef pp_SORT_TAGS
           qsort( sort_tags_local, (size_t)nparts_local, 2*sizeof(int), CompareTags);
+#endif
         }
       }
       else{
@@ -1449,6 +1489,17 @@ partpropdata *GetPartProp(char *label){
   return NULL;
 }
 
+/* ------------------ SetStreakShow ------------------------ */
+
+void SetStreakShow(int show){
+#ifdef pp_SORT_TAGS_BG
+  if(show == 1){
+    THREADcontrol(sorttags_threads, THREAD_JOIN);
+  }
+#endif
+  streak5show = show;
+}
+
 /* ------------------ InitPartProp ------------------------ */
 
 void InitPartProp(void){
@@ -2108,6 +2159,16 @@ void FinalizePartLoad(partdata *parti){
     }
   }
   visParticles = 1;
+#ifdef pp_SORT_TAGS
+#ifdef pp_SORT_TAGS_BG
+  THREADrun(sorttags_threads);
+  if(runscript == 1 || streak5show == 1){
+    THREADcontrol(sorttags_threads, THREAD_JOIN);
+  }
+#else
+  SortAllPartTags(NULL);
+#endif
+#endif
 
   // generate histograms now rather than in the background if a script is running
 
@@ -2229,7 +2290,9 @@ FILE_SIZE ReadPart(char *file_arg, int ifile_arg, int load_flag, int *errorcode_
     return 0.0;
   }
   CheckMemory;
+  INIT_PRINT_TIMER(timer_getpartdata);
   GetPartData(parti, nf_all_local, &file_size_local);
+  PRINT_TIMER(timer_getpartdata, "GetPartData");
   CheckMemory;
   THREADcontrol(partload_threads, THREAD_LOCK);
   parti->loaded = 1;
@@ -2258,7 +2321,7 @@ FILE_SIZE ReadPart(char *file_arg, int ifile_arg, int load_flag, int *errorcode_
     if(parti->finalize == 1){
       INIT_PRINT_TIMER(finalize_part);
       FinalizePartLoad(parti);
-      PRINT_TIMER(finalize_part, "finalize particle time");
+      PRINT_TIMER(finalize_part, "FinalizePartLoad");
     }
     STOP_TIMER(load_time_local);
 #ifndef pp_PARTFRAME

Source/smokeview/glui_bounds.cpp

@@ -6363,10 +6363,10 @@ void PartBoundCB(int var){
   case STREAKLENGTH:
     UpdateStreakValue(float_streak5value-0.001);
     if(float_streak5value==0.0){
-      streak5show=0;
+      SetStreakShow(0);
     }
     else{
-      streak5show=1;
+      SetStreakShow(1);
     }
     updatemenu=1;
     break;

Source/smokeview/menus.c

@@ -3901,14 +3901,14 @@ void ParticleStreakShowMenu(int value){
 
   if(value==-1)return;
   if(value==MENU_STREAK_HIDE){
-    streak5show=0;
+    SetStreakShow(0);
     streak5step=0;
   }
   else if(value==MENU_STREAK_HEAD){
     showstreakhead=1-showstreakhead;
   }
   else{
-    streak5show=1;
+    SetStreakShow(1);
     streak5step=0;
     rvalue=streak_rvalue[value];
     UpdateStreakValue(rvalue-0.001);

Source/smokeview/options.h

@@ -16,6 +16,8 @@
 #define pp_GPUTHROTTLE      // pp_GPU directive must also be set
 #endif
 //#define pp_FDS            // create a 1 mesh input file
+#define pp_SORT_TAGS        // sort particle tags after after particles are read in
+#define pp_SORT_TAGS_BG     // sort particle tags in background
 
 #ifdef pp_FRAME          // turn on each frame type if pp_FRAME is set
 #define pp_BOUNDFRAME    // turn on frame code for boundary files

Source/smokeview/readsmv.c

@@ -11882,6 +11882,12 @@ int ReadSMV_Configure(){
   THREADrun(ffmpeg_threads);
   PRINT_TIMER(timer_readsmv, "SetupFFMT");
 
+#ifdef pp_SORT_TAGS_BG
+  if(sorttags_threads == NULL){
+    sorttags_threads = THREADinit(&n_sorttags_threads, &use_sorttags_threads, SortAllPartTags);
+  }
+#endif
+
   if(isosurface_threads == NULL){
     isosurface_threads = THREADinit(&n_isosurface_threads, &use_isosurface_threads, SetupAllIsosurfaces);
   }

Source/smokeview/smokeheaders.h

@@ -24,6 +24,9 @@ EXTERNCPP void *ReadVolsmokeAllFramesAllMeshes2(void *arg);
 EXTERNCPP void *SetupAllIsosurfaces(void *arg);
 EXTERNCPP void *MtMergeSmoke3D(void *arg);
 EXTERNCPP void UpdateGluiMergeSmoke(void);
+#ifdef pp_SORT_TAGS
+EXTERNCPP void *SortAllPartTags(void *arg);
+#endif
 
 //*** glui_bounds.cpp headers
 EXTERNCPP int  GLUIGetChopHide(char *label);
@@ -1115,6 +1118,7 @@ EXTERNCPP void SetSliceMax(int set_valmax, float valmax, char *buffer2);
 EXTERNCPP void SetSliceBounds(int set_valmin, float valmin, int set_valmax, float valmax, char *buffer2);
 EXTERNCPP void SetBoundBounds(int set_valmin, float valmin, int set_valmax, float valmax, char *buffer2);
 
+EXTERNCPP void SetStreakShow(int show);
 EXTERNCPP void GetPartColors(partdata *parti, int nlevels, int flag);
 EXTERNCPP void GetBoundaryColors(float *t, int nt, unsigned char *it,
               int settmin, float *tmin, int settmax, float *tmax,

Source/smokeview/smokeviewvars.h

@@ -76,6 +76,12 @@ SVEXTERN int SVDECL(use_partload_threads, 1);
 #endif
 SVEXTERN threaderdata SVDECL(*partload_threads,         NULL);
 
+#ifdef pp_SORT_TAGS_BG
+//*** sorttags
+SVEXTERN int SVDECL(n_sorttags_threads, 1), SVDECL(use_sorttags_threads, 1);
+SVEXTERN threaderdata SVDECL(*sorttags_threads, NULL);
+#endif
+
 //*** patchbounds
 SVEXTERN int SVDECL(n_patchbound_threads, 1), SVDECL(use_patchbound_threads, 1);
 SVEXTERN threaderdata SVDECL(*patchbound_threads, NULL);

vistest/large/part/part.fds → vistest/large/part/part1gb.fds

@@ -1,23 +1,19 @@
-&HEAD CHID='part' /
+&HEAD CHID='part1gb' /
 
 &MESH IJK=10,10,4, XB=0.0,1.0,0.0,1.0,0.0,0.4 /
 
-&TIME T_END=60 /
+&DUMP NFRAMES=550 /
+
+&TIME T_END=55.0 /
 
 &SURF ID='hot', GEOMETRY='SPHERICAL', HEAT_TRANSFER_COEFFICIENT=0., EMISSIVITY=1., TMP_FRONT=800, RADIUS=0.005 /
 
 &PART ID='embers', SURF_ID='hot', SAMPLING_FACTOR=1, STATIC=T, QUANTITIES='PARTICLE TEMPERATURE', PROP_ID='ball shape' /
 
 &PROP ID='ball shape', SMOKEVIEW_ID='SPHERE', SMOKEVIEW_PARAMETERS(1)='D=0.01' /
 
-&INIT PART_ID='embers', N_PARTICLES=800000, XB=0.0,1.0,0.0,1.0,0.0,0.02 /
+&INIT PART_ID='embers', N_PARTICLES=100000, XB=0.0,1.0,0.0,1.0,0.0,0.02 /
 
 &VENT MB='ZMAX', SURF_ID='OPEN' /
 
-&BNDF QUANTITY='GAUGE HEAT FLUX', CELL_CENTERED=T /
-&BNDF QUANTITY='WALL TEMPERATURE', CELL_CENTERED=T /
-
-&SLCF PBY=0.5, QUANTITY='RADIATION LOSS', CELL_CENTERED=T /
-&SLCF PBY=0.5, QUANTITY='INTEGRATED INTENSITY', CELL_CENTERED=T /
-
 &TAIL /