diff --git a/amrex/docs_html/_downloads/008eb6dbfab802633dff40122ece848c/amrex.pdf b/amrex/docs_html/_downloads/008eb6dbfab802633dff40122ece848c/amrex.pdf
index ad632155b5..f5166de41b 100644
Binary files a/amrex/docs_html/_downloads/008eb6dbfab802633dff40122ece848c/amrex.pdf and b/amrex/docs_html/_downloads/008eb6dbfab802633dff40122ece848c/amrex.pdf differ
diff --git a/amrex/docs_html/doxygen/AMReX__TinyProfiler_8H_source.html b/amrex/docs_html/doxygen/AMReX__TinyProfiler_8H_source.html
index b6ee90c32f..b198adf4bb 100644
--- a/amrex/docs_html/doxygen/AMReX__TinyProfiler_8H_source.html
+++ b/amrex/docs_html/doxygen/AMReX__TinyProfiler_8H_source.html
@@ -270,9 +270,9 @@
Definition: AMReX_TinyProfiler.H:155
~TinyProfileRegion()
Definition: AMReX_TinyProfiler.cpp:943
~TinyProfileRegion()
Definition: AMReX_TinyProfiler.cpp:940
TinyProfileRegion(TinyProfileRegion const &)=delete
TinyProfileRegion(std::string a_regname) noexcept
Definition: AMReX_TinyProfiler.cpp:927
TinyProfileRegion(std::string a_regname) noexcept
Definition: AMReX_TinyProfiler.cpp:924
TinyProfileRegion & operator=(TinyProfileRegion const &)=delete
TinyProfileRegion(TinyProfileRegion &&)=delete
std::string regname
Definition: AMReX_TinyProfiler.H:165
A simple profiler that returns basic performance information (e.g. min, max, and average running time...
Definition: AMReX_TinyProfiler.H:31
static bool memprof_enabled
Definition: AMReX_TinyProfiler.H:143
static std::vector< std::string > regionstack
Definition: AMReX_TinyProfiler.H:134
static bool RegisterArena(const std::string &memory_name, std::map< std::string, MemStat > &memstats) noexcept
Definition: AMReX_TinyProfiler.cpp:495
static bool RegisterArena(const std::string &memory_name, std::map< std::string, MemStat > &memstats) noexcept
Definition: AMReX_TinyProfiler.cpp:492
static void MemoryInitialize() noexcept
Definition: AMReX_TinyProfiler.cpp:335
static bool device_synchronize_around_region
Definition: AMReX_TinyProfiler.H:138
static std::deque< const TinyProfiler * > mem_stack
Definition: AMReX_TinyProfiler.H:121
static std::string const & get_output_file()
Definition: AMReX_TinyProfiler.cpp:961
static void PrintMemStats(std::map< std::string, MemStat > &memstats, std::string const &memname, double dt_max, double t_final, std::ostream *os)
Definition: AMReX_TinyProfiler.cpp:735
static std::string const & get_output_file()
Definition: AMReX_TinyProfiler.cpp:958
static void PrintMemStats(std::map< std::string, MemStat > &memstats, std::string const &memname, double dt_max, double t_final, std::ostream *os)
Definition: AMReX_TinyProfiler.cpp:732
TinyProfiler & operator=(TinyProfiler const &)=delete
static void Finalize(bool bFlushing=false) noexcept
Definition: AMReX_TinyProfiler.cpp:354
static void MemoryFinalize(bool bFlushing=false) noexcept
Definition: AMReX_TinyProfiler.cpp:447
std::vector< Stats * > stats
Definition: AMReX_TinyProfiler.H:119
static std::deque< std::tuple< double, double, std::string * > > ttstack
Definition: AMReX_TinyProfiler.H:135
TinyProfiler(TinyProfiler const &)=delete
static void PrintStats(std::map< std::string, Stats > ®stats, double dt_max, std::ostream *os)
Definition: AMReX_TinyProfiler.cpp:521
static void PrintStats(std::map< std::string, Stats > ®stats, double dt_max, std::ostream *os)
Definition: AMReX_TinyProfiler.cpp:518
static MemStat * memory_alloc(std::size_t nbytes, std::map< std::string, MemStat > &memstats) noexcept
Definition: AMReX_TinyProfiler.cpp:267
void stop() noexcept
Definition: AMReX_TinyProfiler.cpp:155
std::string fname
Definition: AMReX_TinyProfiler.H:116
static void PrintCallStack(std::ostream &os)
Definition: AMReX_TinyProfiler.cpp:950
static void PrintCallStack(std::ostream &os)
Definition: AMReX_TinyProfiler.cpp:947
TinyProfiler(std::string funcname) noexcept
Definition: AMReX_TinyProfiler.cpp:64
bool in_parallel_region
Definition: AMReX_TinyProfiler.H:117
void start() noexcept
Definition: AMReX_TinyProfiler.cpp:94
static double t_init
Definition: AMReX_TinyProfiler.H:137
void memory_start() const noexcept
Definition: AMReX_TinyProfiler.cpp:228
void memory_stop() const noexcept
Definition: AMReX_TinyProfiler.cpp:245
static void StartRegion(std::string regname) noexcept
Definition: AMReX_TinyProfiler.cpp:908
static void StartRegion(std::string regname) noexcept
Definition: AMReX_TinyProfiler.cpp:905
static std::map< std::string, std::map< std::string, Stats > > statsmap
Definition: AMReX_TinyProfiler.H:136
~TinyProfiler()
Definition: AMReX_TinyProfiler.cpp:88
static void StopRegion(const std::string ®name) noexcept
Definition: AMReX_TinyProfiler.cpp:918
static void StopRegion(const std::string ®name) noexcept
Definition: AMReX_TinyProfiler.cpp:915
static void memory_free(std::size_t nbytes, MemStat *stat) noexcept
Definition: AMReX_TinyProfiler.cpp:296
static std::vector< std::map< std::string, MemStat > * > all_memstats
Definition: AMReX_TinyProfiler.H:131
static void Initialize() noexcept
Definition: AMReX_TinyProfiler.cpp:311
TinyProfiler(TinyProfiler &&)=delete
static std::string output_file
Definition: AMReX_TinyProfiler.H:144
static int n_print_tabs
Definition: AMReX_TinyProfiler.H:139
static void DeregisterArena(std::map< std::string, MemStat > &memstats) noexcept
Definition: AMReX_TinyProfiler.cpp:506
static void DeregisterArena(std::map< std::string, MemStat > &memstats) noexcept
Definition: AMReX_TinyProfiler.cpp:503
@ max
Definition: AMReX_ParallelReduce.H:17
Definition: AMReX_Amr.cpp:49
Definition: AMReX_Arena.H:12
[ParallelDescriptor::IOProcessor]()){
-autoconst&ofile=get_output_file();
-if(ofile.empty()){
-os=&([amrex::OutStream]());
-}elseif(ofile!="/dev/null"){
-ofs.open(ofile,std::ios_base::app);
-if(!ofs.is_open()){
-[amrex::Error]("TinyProfilerfailedtoopen"+ofile);
-}
-os=static_cast<std::ostream*>(&ofs);
-}
-}
-
-for(std::size_ti=0;i<all_memstats.size();++i){
-PrintMemStats(*(all_memstats[i]),all_memnames[i],dt_max,t_final,os);
-}
-
-if(!bFlushing){
-all_memstats.clear();
-all_memnames.clear();
-}
+if([ParallelDescriptor::IOProcessor]()){
+autoconst&ofile=get_output_file();
+if(ofile.empty()){
+os=&([amrex::OutStream]());
+}elseif(ofile!="/dev/null"){
+ofs.open(ofile,std::ios_base::app);
+if(!ofs.is_open()){
+[amrex::Error]("TinyProfilerfailedtoopen"+ofile);
+}
+os=static_cast<std::ostream*>(&ofs);
+}
+}
+
+for(std::size_ti=0;i<all_memstats.size();++i){
+PrintMemStats(*(all_memstats[i]),all_memnames[i],dt_max,t_final,os);
+}
+
+if(!bFlushing){
+all_memstats.clear();
+all_memnames.clear();
+}
+}
-if(os){os->precision(oldprec);}
-}
-
-bool
-[TinyProfiler::RegisterArena](conststd::string&memory_name,
-std::map<std::string,MemStat>&memstats)noexcept
-{
-if(!memprof_enabled){returnfalse;}
-
-all_memstats.push_back(&memstats);
-all_memnames.push_back(memory_name);
-returntrue;
-}
-
-void
-[TinyProfiler::DeregisterArena](std::map<std::string,MemStat>&memstats)noexcept
-{
-if(!memprof_enabled){return;}
-
-for(std::size_ti=0;i<all_memstats.size();){
-if(all_memstats[i]==&memstats){
-all_memstats.erase(all_memstats.begin()+i);
-all_memnames.erase(all_memnames.begin()+i);
-}else{
-++i;
-}
-}
-}
-
-void
-[TinyProfiler::PrintStats](std::map<std::string,Stats>®stats,doubledt_max,
-std::ostream*os)
-{
-
-{
-Vector<std::string>localStrings,syncedStrings;
-boolalreadySynced;
-
-for(autoconst&kv:regstats){
-localStrings.push_back(kv.first);
-}
-
-[amrex::SyncStrings](localStrings,syncedStrings,alreadySynced);
-
-if(!alreadySynced){
-for(autoconst&s:syncedStrings){
-if(regstats.find(s)==regstats.end()){
-regstats.insert(std::make_pair(s,Stats()));
-}
-}
-}
-}
-
-if(regstats.empty()){return;}
+bool
+[TinyProfiler::RegisterArena](conststd::string&memory_name,
+std::map<std::string,MemStat>&memstats)noexcept
+{
+if(!memprof_enabled){returnfalse;}
+
+all_memstats.push_back(&memstats);
+all_memnames.push_back(memory_name);
+returntrue;
+}
+
+void
+[TinyProfiler::DeregisterArena](std::map<std::string,MemStat>&memstats)noexcept
+{
+if(!memprof_enabled){return;}
+
+for(std::size_ti=0;i<all_memstats.size();){
+if(all_memstats[i]==&memstats){
+all_memstats.erase(all_memstats.begin()+i);
+all_memnames.erase(all_memnames.begin()+i);
+}else{
+++i;
+}
+}
+}
+
+void
+[TinyProfiler::PrintStats](std::map<std::string,Stats>®stats,doubledt_max,
+std::ostream*os)
+{
+
+{
+Vector<std::string>localStrings,syncedStrings;
+boolalreadySynced;
+
+for(autoconst&kv:regstats){
+localStrings.push_back(kv.first);
+}
+
+[amrex::SyncStrings](localStrings,syncedStrings,alreadySynced);
+
+if(!alreadySynced){
+for(autoconst&s:syncedStrings){
+if(regstats.find(s)==regstats.end()){
+regstats.insert(std::make_pair(s,Stats()));
+}
+}
+}
+}
+
+if(regstats.empty()){return;}
+
+intnprocs=[ParallelDescriptor::NProcs]();
+intioproc=[ParallelDescriptor::IOProcessorNumber]();
-intnprocs=[ParallelDescriptor::NProcs]();
-intioproc=[ParallelDescriptor::IOProcessorNumber]();
-
-std::vector<ProcStats>allprocstats;
-intmaxfnamelen=0;
-Longmaxncalls=0;
-
-
-for(constauto®stat:regstats)
-{
-Longn=regstat.second.n;
-doubledts[2]={regstat.second.dtin,regstat.second.dtex};
+std::vector<ProcStats>allprocstats;
+intmaxfnamelen=0;
+Longmaxncalls=0;
+
+
+for(constauto®stat:regstats)
+{
+Longn=regstat.second.n;
+doubledts[2]={regstat.second.dtin,regstat.second.dtex};
+
+std::vector<Long>ncalls(nprocs);
+std::vector<double>dtdt(2*nprocs);
-std::vector<Long>ncalls(nprocs);
-std::vector<double>dtdt(2*nprocs);
-
-if([ParallelDescriptor::NProcs]()==1)
-{
-ncalls[0]=n;
-dtdt[0]=dts[0];
-dtdt[1]=dts[1];
-}else
-{
-[ParallelDescriptor::Gather](&n,1,ncalls.data(),1,ioproc);
-[ParallelDescriptor::Gather](dts,2,dtdt.data(),2,ioproc);
-}
-
-if([ParallelDescriptor::IOProcessor]()){
-ProcStatspst;
-for(inti=0;i<nprocs;++i){
-pst.nmin=[std::min](pst.nmin,ncalls[i]);
-pst.navg+=ncalls[i];
-pst.nmax=[std::max](pst.nmax,ncalls[i]);
-pst.dtinmin=[std::min](pst.dtinmin,dtdt[2*i]);
-pst.dtinavg+=dtdt[2*i];
-pst.dtinmax=[std::max](pst.dtinmax,dtdt[2*i]);
-pst.dtexmin=[std::min](pst.dtexmin,dtdt[2*i+1]);
-pst.dtexavg+=dtdt[2*i+1];
-pst.dtexmax=[std::max](pst.dtexmax,dtdt[2*i+1]);
-}
-pst.navg/=nprocs;
-pst.dtinavg/=nprocs;
-pst.dtexavg/=nprocs;
-pst.fname=regstat.first;
-allprocstats.push_back(pst);
-maxfnamelen=[std::max](maxfnamelen,int(pst.fname.size()));
-maxncalls=[std::max](maxncalls,pst.nmax);
-}
-}
-
-if([ParallelDescriptor::IOProcessor]()&&os)
-{
-IOFormatSaveriofmtsaver(*os);
+if([ParallelDescriptor::NProcs]()==1)
+{
+ncalls[0]=n;
+dtdt[0]=dts[0];
+dtdt[1]=dts[1];
+}else
+{
+[ParallelDescriptor::Gather](&n,1,ncalls.data(),1,ioproc);
+[ParallelDescriptor::Gather](dts,2,dtdt.data(),2,ioproc);
+}
+
+if([ParallelDescriptor::IOProcessor]()){
+ProcStatspst;
+for(inti=0;i<nprocs;++i){
+pst.nmin=[std::min](pst.nmin,ncalls[i]);
+pst.navg+=ncalls[i];
+pst.nmax=[std::max](pst.nmax,ncalls[i]);
+pst.dtinmin=[std::min](pst.dtinmin,dtdt[2*i]);
+pst.dtinavg+=dtdt[2*i];
+pst.dtinmax=[std::max](pst.dtinmax,dtdt[2*i]);
+pst.dtexmin=[std::min](pst.dtexmin,dtdt[2*i+1]);
+pst.dtexavg+=dtdt[2*i+1];
+pst.dtexmax=[std::max](pst.dtexmax,dtdt[2*i+1]);
+}
+pst.navg/=nprocs;
+pst.dtinavg/=nprocs;
+pst.dtexavg/=nprocs;
+pst.fname=regstat.first;
+allprocstats.push_back(pst);
+maxfnamelen=[std::max](maxfnamelen,int(pst.fname.size()));
+maxncalls=[std::max](maxncalls,pst.nmax);
+}
+}
+
+if([ParallelDescriptor::IOProcessor]()&&os)
+{
+IOFormatSaveriofmtsaver(*os);
+
+*os<<std::setfill('')<<std::setprecision(4);
+intwt=9;
-*os<<std::setfill('')<<std::setprecision(4);
-intwt=9;
-
-intwnc=([int])std::log10((double)maxncalls)+1;
-wnc=[std::max](wnc,int(std::string("NCalls").[size]()));
-wt=[std::max](wt,int(std::string("Excl.Min").[size]()));
-intwp=6;
-wp=[std::max](wp,int(std::string("Max%").[size]()));
-
-conststd::stringhline(maxfnamelen+wnc+2+(wt+2)*3+wp+2,'-');
+intwnc=([int])std::log10((double)maxncalls)+1;
+wnc=[std::max](wnc,int(std::string("NCalls").[size]()));
+wt=[std::max](wt,int(std::string("Excl.Min").[size]()));
+intwp=6;
+wp=[std::max](wp,int(std::string("Max%").[size]()));
+
+conststd::stringhline(maxfnamelen+wnc+2+(wt+2)*3+wp+2,'-');
+
+ProcStatsother_procstat;
+boolprint_other_procstat=false;
-ProcStatsother_procstat;
-boolprint_other_procstat=false;
-
-
-if([print_threshold]>0.){
-
-other_procstat.nmin=0;
-other_procstat.dtinmin=0.;
-other_procstat.dtexmin=0.;
-other_procstat.fname="Other";
-intnum_procstats_in_other=0;
-
-
-std::sort(allprocstats.begin(),allprocstats.end(),[ProcStats::compin]);
-for(Longi=static_cast<Long>(allprocstats.size())-1;i>=0;--i){
-
-if((other_procstat.dtinmax+allprocstats[i].dtinmax)*(100.0/dt_max)
-<[print_threshold]){
-allprocstats[i].do_print=false;
-++num_procstats_in_other;
-
-
-
-other_procstat.nmin+=allprocstats[i].nmin;
-other_procstat.navg+=allprocstats[i].navg;
-other_procstat.nmax+=allprocstats[i].nmax;
-
-other_procstat.dtinmin+=allprocstats[i].dtinmin;
-other_procstat.dtinavg+=allprocstats[i].dtinavg;
-other_procstat.dtinmax+=allprocstats[i].dtinmax;
-
-other_procstat.dtexmin+=allprocstats[i].dtexmin;
-other_procstat.dtexavg+=allprocstats[i].dtexavg;
-other_procstat.dtexmax+=allprocstats[i].dtexmax;
-}else{
-break;
-}
-}
-
-if(num_procstats_in_other==1){
-
-
-allprocstats.back().do_print=true;
-}elseif(num_procstats_in_other>=2){
-print_other_procstat=true;
-}
-}
-
-
-std::sort(allprocstats.begin(),allprocstats.end(),[ProcStats::compex]);
-if(print_other_procstat){
-
-allprocstats.push_back(other_procstat);
-}
-*os<<"\n"<<hline<<"\n";
-*os<<std::left
-<<std::setw(maxfnamelen)<<"Name"
-<<std::right
-<<std::setw(wnc+2)<<"NCalls"
-<<std::setw(wt+2)<<"Excl.Min"
-<<std::setw(wt+2)<<"Excl.Avg"
-<<std::setw(wt+2)<<"Excl.Max"
-<<std::setw(wp+2)<<"Max%"
-<<"\n"<<hline<<"\n";
-for(constauto&allprocstat:allprocstats)
-{
-if(!allprocstat.do_print){
-continue;
-}
-*os<<std::setprecision(4)<<std::left
-<<std::setw(maxfnamelen)<<allprocstat.fname
-<<std::right
-<<std::setw(wnc+2)<<allprocstat.navg
-<<std::setw(wt+2)<<allprocstat.dtexmin
-<<std::setw(wt+2)<<allprocstat.dtexavg
-<<std::setw(wt+2)<<allprocstat.dtexmax
-<<std::setprecision(2)<<std::setw(wp+1)<<std::fixed
-<<allprocstat.dtexmax*(100.0/dt_max)<<"%";
-os->unsetf(std::ios_base::fixed);
-*os<<"\n";
-}
-*os<<hline<<"\n";
-if(print_other_procstat){
-allprocstats.pop_back();
-}
-
-
-std::sort(allprocstats.begin(),allprocstats.end(),[ProcStats::compin]);
-if(print_other_procstat){
-
-allprocstats.push_back(other_procstat);
-}
-*os<<"\n"<<hline<<"\n";
-*os<<std::left
-<<std::setw(maxfnamelen)<<"Name"
-<<std::right
-<<std::setw(wnc+2)<<"NCalls"
-<<std::setw(wt+2)<<"Incl.Min"
-<<std::setw(wt+2)<<"Incl.Avg"
-<<std::setw(wt+2)<<"Incl.Max"
-<<std::setw(wp+2)<<"Max%"
-<<"\n"<<hline<<"\n";
-for(constauto&allprocstat:allprocstats)
-{
-if(!allprocstat.do_print){
-continue;
-}
-*os<<std::setprecision(4)<<std::left
-<<std::setw(maxfnamelen)<<allprocstat.fname
-<<std::right
-<<std::setw(wnc+2)<<allprocstat.navg
-<<std::setw(wt+2)<<allprocstat.dtinmin
-<<std::setw(wt+2)<<allprocstat.dtinavg
-<<std::setw(wt+2)<<allprocstat.dtinmax
-<<std::setprecision(2)<<std::setw(wp+1)<<std::fixed
-<<allprocstat.dtinmax*(100.0/dt_max)<<"%";
-os->unsetf(std::ios_base::fixed);
-*os<<"\n";
-}
-*os<<hline<<"\n\n";
-}
-}
-
-void
-[TinyProfiler::PrintMemStats](std::map<std::string,MemStat>&memstats,
-std::stringconst&memname,doubledt_max,
-doublet_final,std::ostream*os)
-{
-
-{
-Vector<std::string>localStrings,syncedStrings;
-boolalreadySynced;
-
-for(autoconst&kv:memstats){
-localStrings.push_back(kv.first);
-}
-
-[amrex::SyncStrings](localStrings,syncedStrings,alreadySynced);
-
-if(!alreadySynced){
-for(autoconst&s:syncedStrings){
-if(memstats.find(s)==memstats.end()){
-memstats[s];
-}
-}
-}
-}
-
-if(memstats.empty()){return;}
+
+if([print_threshold]>0.){
+
+other_procstat.nmin=0;
+other_procstat.dtinmin=0.;
+other_procstat.dtexmin=0.;
+other_procstat.fname="Other";
+intnum_procstats_in_other=0;
+
+
+std::sort(allprocstats.begin(),allprocstats.end(),[ProcStats::compin]);
+for(Longi=static_cast<Long>(allprocstats.size())-1;i>=0;--i){
+
+if((other_procstat.dtinmax+allprocstats[i].dtinmax)*(100.0/dt_max)
+<[print_threshold]){
+allprocstats[i].do_print=false;
+++num_procstats_in_other;
+
+
+
+other_procstat.nmin+=allprocstats[i].nmin;
+other_procstat.navg+=allprocstats[i].navg;
+other_procstat.nmax+=allprocstats[i].nmax;
+
+other_procstat.dtinmin+=allprocstats[i].dtinmin;
+other_procstat.dtinavg+=allprocstats[i].dtinavg;
+other_procstat.dtinmax+=allprocstats[i].dtinmax;
+
+other_procstat.dtexmin+=allprocstats[i].dtexmin;
+other_procstat.dtexavg+=allprocstats[i].dtexavg;
+other_procstat.dtexmax+=allprocstats[i].dtexmax;
+}else{
+break;
+}
+}
+
+if(num_procstats_in_other==1){
+
+
+allprocstats.back().do_print=true;
+}elseif(num_procstats_in_other>=2){
+print_other_procstat=true;
+}
+}
+
+
+std::sort(allprocstats.begin(),allprocstats.end(),[ProcStats::compex]);
+if(print_other_procstat){
+
+allprocstats.push_back(other_procstat);
+}
+*os<<"\n"<<hline<<"\n";
+*os<<std::left
+<<std::setw(maxfnamelen)<<"Name"
+<<std::right
+<<std::setw(wnc+2)<<"NCalls"
+<<std::setw(wt+2)<<"Excl.Min"
+<<std::setw(wt+2)<<"Excl.Avg"
+<<std::setw(wt+2)<<"Excl.Max"
+<<std::setw(wp+2)<<"Max%"
+<<"\n"<<hline<<"\n";
+for(constauto&allprocstat:allprocstats)
+{
+if(!allprocstat.do_print){
+continue;
+}
+*os<<std::setprecision(4)<<std::left
+<<std::setw(maxfnamelen)<<allprocstat.fname
+<<std::right
+<<std::setw(wnc+2)<<allprocstat.navg
+<<std::setw(wt+2)<<allprocstat.dtexmin
+<<std::setw(wt+2)<<allprocstat.dtexavg
+<<std::setw(wt+2)<<allprocstat.dtexmax
+<<std::setprecision(2)<<std::setw(wp+1)<<std::fixed
+<<allprocstat.dtexmax*(100.0/dt_max)<<"%";
+os->unsetf(std::ios_base::fixed);
+*os<<"\n";
+}
+*os<<hline<<"\n";
+if(print_other_procstat){
+allprocstats.pop_back();
+}
+
+
+std::sort(allprocstats.begin(),allprocstats.end(),[ProcStats::compin]);
+if(print_other_procstat){
+
+allprocstats.push_back(other_procstat);
+}
+*os<<"\n"<<hline<<"\n";
+*os<<std::left
+<<std::setw(maxfnamelen)<<"Name"
+<<std::right
+<<std::setw(wnc+2)<<"NCalls"
+<<std::setw(wt+2)<<"Incl.Min"
+<<std::setw(wt+2)<<"Incl.Avg"
+<<std::setw(wt+2)<<"Incl.Max"
+<<std::setw(wp+2)<<"Max%"
+<<"\n"<<hline<<"\n";
+for(constauto&allprocstat:allprocstats)
+{
+if(!allprocstat.do_print){
+continue;
+}
+*os<<std::setprecision(4)<<std::left
+<<std::setw(maxfnamelen)<<allprocstat.fname
+<<std::right
+<<std::setw(wnc+2)<<allprocstat.navg
+<<std::setw(wt+2)<<allprocstat.dtinmin
+<<std::setw(wt+2)<<allprocstat.dtinavg
+<<std::setw(wt+2)<<allprocstat.dtinmax
+<<std::setprecision(2)<<std::setw(wp+1)<<std::fixed
+<<allprocstat.dtinmax*(100.0/dt_max)<<"%";
+os->unsetf(std::ios_base::fixed);
+*os<<"\n";
+}
+*os<<hline<<"\n\n";
+}
+}
+
+void
+[TinyProfiler::PrintMemStats](std::map<std::string,MemStat>&memstats,
+std::stringconst&memname,doubledt_max,
+doublet_final,std::ostream*os)
+{
+
+{
+Vector<std::string>localStrings,syncedStrings;
+boolalreadySynced;
+
+for(autoconst&kv:memstats){
+localStrings.push_back(kv.first);
+}
+
+[amrex::SyncStrings](localStrings,syncedStrings,alreadySynced);
+
+if(!alreadySynced){
+for(autoconst&s:syncedStrings){
+if(memstats.find(s)==memstats.end()){
+memstats[s];
+}
+}
+}
+}
+
+if(memstats.empty()){return;}
+
+constintnprocs=[ParallelDescriptor::NProcs]();
+constintioproc=[ParallelDescriptor::IOProcessorNumber]();
-constintnprocs=[ParallelDescriptor::NProcs]();
-constintioproc=[ParallelDescriptor::IOProcessorNumber]();
-
-std::vector<MemProcStats>allprocstats;
-
-
-for(constauto&it:memstats)
-{
-Longnalloc=it.second.nalloc;
-Longnfree=it.second.nfree;
-
-Longavgmem=static_cast<Long>(
-(it.second.avgmem+static_cast<double>(it.second.currentmem)*t_final)/dt_max);
-Longmaxmem=it.second.maxmem;
-
-std::vector<Long>nalloc_vec(nprocs);
-std::vector<Long>nfree_vec(nprocs);
-std::vector<Long>avgmem_vec(nprocs);
-std::vector<Long>maxmem_vec(nprocs);
-
-if(nprocs==1)
-{
-nalloc_vec[0]=nalloc;
-nfree_vec[0]=nfree;
-avgmem_vec[0]=avgmem;
-maxmem_vec[0]=maxmem;
-}else
-{
-[ParallelDescriptor::Gather](&nalloc,1,nalloc_vec.data(),1,ioproc);
-[ParallelDescriptor::Gather](&nfree,1,nfree_vec.data(),1,ioproc);
-[ParallelDescriptor::Gather](&maxmem,1,maxmem_vec.data(),1,ioproc);
-[ParallelDescriptor::Gather](&avgmem,1,avgmem_vec.data(),1,ioproc);
-}
-
-if([ParallelDescriptor::IOProcessor]()){
-MemProcStatspst;
-for(inti=0;i<nprocs;++i){
-
-pst.nalloc+=nalloc_vec[i];
-pst.nfree+=nfree_vec[i];
-pst.avgmem_min=[std::min](pst.avgmem_min,avgmem_vec[i]);
-pst.avgmem_avg+=avgmem_vec[i];
-pst.avgmem_max=[std::max](pst.avgmem_max,avgmem_vec[i]);
-pst.maxmem_min=[std::min](pst.maxmem_min,maxmem_vec[i]);
-pst.maxmem_avg+=maxmem_vec[i];
-pst.maxmem_max=[std::max](pst.maxmem_max,maxmem_vec[i]);
-}
-pst.avgmem_avg/=nprocs;
-pst.maxmem_avg/=nprocs;
-pst.fname=it.first;
-allprocstats.push_back(pst);
-}
-}
-
-std::sort(allprocstats.begin(),allprocstats.end(),[MemProcStats::compmem]);
-
-std::vector<std::vector<std::string>>allstatsstr;
-
-if(nprocs==1){
-allstatsstr.push_back({"Name","Nalloc","Nfree","AvgMem","MaxMem"});
-}else{
-allstatsstr.push_back({"Name","Nalloc","Nfree",
-"AvgMemmin","AvgMemavg","AvgMemmax",
-"MaxMemmin","MaxMemavg","MaxMemmax"});
-}
-
-automem_to_string=[](Longnbytes){
-std::stringunit="B";
-if(nbytes>=10000){
-nbytes/=1024;
-unit="KiB";
-}
-if(nbytes>=10000){
-nbytes/=1024;
-unit="MiB";
-}
-if(nbytes>=10000){
-nbytes/=1024;
-unit="GiB";
-}
-if(nbytes>=10000){
-nbytes/=1024;
-unit="TiB";
-}
-returnstd::to_string(nbytes)+unit;
-};
-
-for(auto&stat:allprocstats){
-if(stat.nalloc!=0||stat.nfree!=0||stat.maxmem_max!=0){
-if(nprocs==1){
-allstatsstr.push_back({stat.fname,
-std::to_string(stat.nalloc),
-std::to_string(stat.nfree),
-mem_to_string(stat.avgmem_max),
-mem_to_string(stat.maxmem_max)});
-}else{
-allstatsstr.push_back({stat.fname,
-std::to_string(stat.nalloc),
-std::to_string(stat.nfree),
-mem_to_string(stat.avgmem_min),
-mem_to_string(stat.avgmem_avg),
-mem_to_string(stat.avgmem_max),
-mem_to_string(stat.maxmem_min),
-mem_to_string(stat.maxmem_avg),
-mem_to_string(stat.maxmem_max)});
-}
-}
-}
-
-std::vector<int>maxlen(allstatsstr[0].[size](),0);
-for(auto&strvec:allstatsstr){
-for(std::size_ti=0;i<maxlen.size();++i){
-maxlen[i]=[std::max](maxlen[i],static_cast<int>(strvec[i].[size]()));
-}
-}
-
-for(std::size_ti=1;i<maxlen.size();++i){
-maxlen[i]+=2;
-}
-
-if(allstatsstr.size()==1||!os){return;}
+std::vector<MemProcStats>allprocstats;
+
+
+for(constauto&it:memstats)
+{
+Longnalloc=it.second.nalloc;
+Longnfree=it.second.nfree;
+
+Longavgmem=static_cast<Long>(
+(it.second.avgmem+static_cast<double>(it.second.currentmem)*t_final)/dt_max);
+Longmaxmem=it.second.maxmem;
+
+std::vector<Long>nalloc_vec(nprocs);
+std::vector<Long>nfree_vec(nprocs);
+std::vector<Long>avgmem_vec(nprocs);
+std::vector<Long>maxmem_vec(nprocs);
+
+if(nprocs==1)
+{
+nalloc_vec[0]=nalloc;
+nfree_vec[0]=nfree;
+avgmem_vec[0]=avgmem;
+maxmem_vec[0]=maxmem;
+}else
+{
+[ParallelDescriptor::Gather](&nalloc,1,nalloc_vec.data(),1,ioproc);
+[ParallelDescriptor::Gather](&nfree,1,nfree_vec.data(),1,ioproc);
+[ParallelDescriptor::Gather](&maxmem,1,maxmem_vec.data(),1,ioproc);
+[ParallelDescriptor::Gather](&avgmem,1,avgmem_vec.data(),1,ioproc);
+}
+
+if([ParallelDescriptor::IOProcessor]()){
+MemProcStatspst;
+for(inti=0;i<nprocs;++i){
+
+pst.nalloc+=nalloc_vec[i];
+pst.nfree+=nfree_vec[i];
+pst.avgmem_min=[std::min](pst.avgmem_min,avgmem_vec[i]);
+pst.avgmem_avg+=avgmem_vec[i];
+pst.avgmem_max=[std::max](pst.avgmem_max,avgmem_vec[i]);
+pst.maxmem_min=[std::min](pst.maxmem_min,maxmem_vec[i]);
+pst.maxmem_avg+=maxmem_vec[i];
+pst.maxmem_max=[std::max](pst.maxmem_max,maxmem_vec[i]);
+}
+pst.avgmem_avg/=nprocs;
+pst.maxmem_avg/=nprocs;
+pst.fname=it.first;
+allprocstats.push_back(pst);
+}
+}
+
+std::sort(allprocstats.begin(),allprocstats.end(),[MemProcStats::compmem]);
+
+std::vector<std::vector<std::string>>allstatsstr;
+
+if(nprocs==1){
+allstatsstr.push_back({"Name","Nalloc","Nfree","AvgMem","MaxMem"});
+}else{
+allstatsstr.push_back({"Name","Nalloc","Nfree",
+"AvgMemmin","AvgMemavg","AvgMemmax",
+"MaxMemmin","MaxMemavg","MaxMemmax"});
+}
+
+automem_to_string=[](Longnbytes){
+std::stringunit="B";
+if(nbytes>=10000){
+nbytes/=1024;
+unit="KiB";
+}
+if(nbytes>=10000){
+nbytes/=1024;
+unit="MiB";
+}
+if(nbytes>=10000){
+nbytes/=1024;
+unit="GiB";
+}
+if(nbytes>=10000){
+nbytes/=1024;
+unit="TiB";
+}
+returnstd::to_string(nbytes)+unit;
+};
+
+for(auto&stat:allprocstats){
+if(stat.nalloc!=0||stat.nfree!=0||stat.maxmem_max!=0){
+if(nprocs==1){
+allstatsstr.push_back({stat.fname,
+std::to_string(stat.nalloc),
+std::to_string(stat.nfree),
+mem_to_string(stat.avgmem_max),
+mem_to_string(stat.maxmem_max)});
+}else{
+allstatsstr.push_back({stat.fname,
+std::to_string(stat.nalloc),
+std::to_string(stat.nfree),
+mem_to_string(stat.avgmem_min),
+mem_to_string(stat.avgmem_avg),
+mem_to_string(stat.avgmem_max),
+mem_to_string(stat.maxmem_min),
+mem_to_string(stat.maxmem_avg),
+mem_to_string(stat.maxmem_max)});
+}
+}
+}
+
+std::vector<int>maxlen(allstatsstr[0].[size](),0);
+for(auto&strvec:allstatsstr){
+for(std::size_ti=0;i<maxlen.size();++i){
+maxlen[i]=[std::max](maxlen[i],static_cast<int>(strvec[i].[size]()));
+}
+}
+
+for(std::size_ti=1;i<maxlen.size();++i){
+maxlen[i]+=2;
+}
+
+if(allstatsstr.size()==1||!os){return;}
+
+IOFormatSaveriofmtsaver(*os);
+*os<<std::setfill('');
-IOFormatSaveriofmtsaver(*os);
-*os<<std::setfill('');
-
-intlenhline=0;
-for(autoi:maxlen){
-lenhline+=i;
-}
-conststd::stringhline(lenhline,'-');
-
-*os<<memname<<"Usage:\n";
-*os<<hline<<"\n";
-for(std::size_ti=0;i<allstatsstr.size();++i){
-*os<<std::left<<std::setw(maxlen[0])<<allstatsstr[i][0];
-for(std::size_tj=1;j<maxlen.size();++j){
-*os<<std::right<<std::setw(maxlen[j])<<allstatsstr[i][j];
-}
-*os<<'\n';
-if(i==0){
-*os<<hline<<"\n";
-}
-}
-*os<<hline<<"\n\n";
-}
-
-void
-[TinyProfiler::StartRegion](std::stringregname)noexcept
-{
-if(!enabled){return;}
-
-if(std::find(regionstack.begin(),regionstack.end(),regname)==regionstack.end()){
-regionstack.emplace_back(std::move(regname));
-}
-}
-
-void
-[TinyProfiler::StopRegion](conststd::string®name)noexcept
-{
-if(!enabled){return;}
-
-if(regname==regionstack.back()){
-regionstack.pop_back();
-}
-}
-
-[TinyProfileRegion::TinyProfileRegion](std::stringa_regname)noexcept
-:regname(std::move(a_regname)),
-tprof(std::string("REG::")+regname,false)
-{
-[TinyProfiler::StartRegion](regname);
-tprof.start();
-}
-
-[TinyProfileRegion::TinyProfileRegion](constchar*a_regname)noexcept
-:regname(a_regname),
-tprof(std::string("REG::")+std::string(a_regname),false)
-{
-[TinyProfiler::StartRegion](a_regname);
-tprof.start();
-}
-
-[TinyProfileRegion::~TinyProfileRegion]()
-{
-[tprof].[stop]();
-[TinyProfiler::StopRegion]([regname]);
-}
-
-void
-[TinyProfiler::PrintCallStack](std::ostream&os)
-{
-if(![enabled]){return;}
-
-os<<"=====TinyProfilers======\n";
-for(autoconst&x:[ttstack]){
-os<<*(std::get<2>(x))<<"\n";
-}
-}
-
-std::stringconst&
-[TinyProfiler::get_output_file]()
-{
-
-
-staticboolfirst=true;
-if(first){
-first=false;
+intlenhline=0;
+for(autoi:maxlen){
+lenhline+=i;
+}
+conststd::stringhline(lenhline,'-');
+
+*os<<memname<<"Usage:\n";
+*os<<hline<<"\n";
+for(std::size_ti=0;i<allstatsstr.size();++i){
+*os<<std::left<<std::setw(maxlen[0])<<allstatsstr[i][0];
+for(std::size_tj=1;j<maxlen.size();++j){
+*os<<std::right<<std::setw(maxlen[j])<<allstatsstr[i][j];
+}
+*os<<'\n';
+if(i==0){
+*os<<hline<<"\n";
+}
+}
+*os<<hline<<"\n\n";
+}
+
+void
+[TinyProfiler::StartRegion](std::stringregname)noexcept
+{
+if(!enabled){return;}
+
+if(std::find(regionstack.begin(),regionstack.end(),regname)==regionstack.end()){
+regionstack.emplace_back(std::move(regname));
+}
+}
+
+void
+[TinyProfiler::StopRegion](conststd::string®name)noexcept
+{
+if(!enabled){return;}
+
+if(regname==regionstack.back()){
+regionstack.pop_back();
+}
+}
+
+[TinyProfileRegion::TinyProfileRegion](std::stringa_regname)noexcept
+:regname(std::move(a_regname)),
+tprof(std::string("REG::")+regname,false)
+{
+[TinyProfiler::StartRegion](regname);
+tprof.start();
+}
+
+[TinyProfileRegion::TinyProfileRegion](constchar*a_regname)noexcept
+:regname(a_regname),
+tprof(std::string("REG::")+std::string(a_regname),false)
+{
+[TinyProfiler::StartRegion](a_regname);
+tprof.start();
+}
+
+[TinyProfileRegion::~TinyProfileRegion]()
+{
+[tprof].[stop]();
+[TinyProfiler::StopRegion]([regname]);
+}
+
+void
+[TinyProfiler::PrintCallStack](std::ostream&os)
+{
+if(![enabled]){return;}
+
+os<<"=====TinyProfilers======\n";
+for(autoconst&x:[ttstack]){
+os<<*(std::get<2>(x))<<"\n";
+}
+}
+
+std::stringconst&
+[TinyProfiler::get_output_file]()
+{
+
+
+staticboolfirst=true;
+if(first){
+first=false;
+
+[amrex::ParmParse][pp]("tiny_profiler");
+[pp].[query]("output_file",[output_file]);
-[amrex::ParmParse][pp]("tiny_profiler");
-[pp].[query]("output_file",[output_file]);
-
-if([ParallelDescriptor::IOProcessor]()){
-if(![output_file].empty()&&[output_file]!="/dev/null"){
-if([FileSystem::Exists]([output_file])){
-[FileSystem::Remove]([output_file]);
-}
-}
-}
-}
+if([ParallelDescriptor::IOProcessor]()){
+if(![output_file].empty()&&[output_file]!="/dev/null"){
+if([FileSystem::Exists]([output_file])){
+[FileSystem::Remove]([output_file]);
+}
+}
+}
+}
+
+return[output_file];
+}
-return[output_file];
-}
-
-}
+}
diff --git a/amrex/docs_xml/doxygen/classamrex_1_1TinyProfileRegion.xml b/amrex/docs_xml/doxygen/classamrex_1_1TinyProfileRegion.xml
index 4aba217583..b18bbc360e 100644
--- a/amrex/docs_xml/doxygen/classamrex_1_1TinyProfileRegion.xml
+++ b/amrex/docs_xml/doxygen/classamrex_1_1TinyProfileRegion.xml
@@ -47,7 +47,7 @@
-
+
@@ -64,7 +64,7 @@
-
+
@@ -141,7 +141,7 @@
-
+
diff --git a/amrex/docs_xml/doxygen/classamrex_1_1TinyProfiler.xml b/amrex/docs_xml/doxygen/classamrex_1_1TinyProfiler.xml
index 44de1921dd..c2224e56ae 100644
--- a/amrex/docs_xml/doxygen/classamrex_1_1TinyProfiler.xml
+++ b/amrex/docs_xml/doxygen/classamrex_1_1TinyProfiler.xml
@@ -563,7 +563,7 @@
-
+
bool
@@ -584,7 +584,7 @@
-
+
void
@@ -601,7 +601,7 @@
-
+
void
@@ -618,7 +618,7 @@
-
+
void
@@ -635,7 +635,7 @@
-
+
void
@@ -652,7 +652,7 @@
-
+
@@ -667,7 +667,7 @@
-
+
void
@@ -692,7 +692,7 @@
-
+
void
@@ -725,7 +725,7 @@
-
+