Nils Thuerey
818bfcca63
compilation. Currently, it is not yet active by default, but Genscher wanted to do some tests. It can be used to distribute the computation load onto multiple shared- memory CPUs by splitting the domain along the y-axis (assuming a gravity force along z). However, there is no load balancing: so if there's fluid only in one of the y-axis halves you will not get a speedup for 2 CPUs. - Added a fix for the memory allocation bugs #7120 and #6775. In solver_init.cpp there are now several variables max___MemChunk (line 692+), that set upper limits for various systems. The same problem existed for mac & linux, but the limit is higher, so it probably went by undetected. The windows limit is currently 1GB, if the strange 700MB limit problems mentioned in the bug regports the bugs persist, this could be further reduced. For 64bit compilations this problem shouldn't exist anyway. What's still missing is a display of how much the resolution was reduced to fit into memory... - And some minor solver code cleanup.
181 lines
4.9 KiB
C++
181 lines
4.9 KiB
C++
|
|
// advance pointer in main loop
|
|
#define ADVANCE_POINTERS(p) \
|
|
ccel += (QCELLSTEP*(p)); \
|
|
tcel += (QCELLSTEP*(p)); \
|
|
pFlagSrc+= (p); \
|
|
pFlagDst+= (p); \
|
|
i+= (p);
|
|
|
|
#define MAX_CALC_ARR 4
|
|
|
|
// >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
|
|
// init region vars
|
|
#define GRID_REGION_INIT() \
|
|
const int istart = -1+gridLoopBound; \
|
|
const int iend = mLevel[mMaxRefine].lSizex-1-gridLoopBound; \
|
|
LbmFloat calcCurrentMass=0; \
|
|
LbmFloat calcCurrentVolume=0; \
|
|
int calcCellsFilled=0; \
|
|
int calcCellsEmptied=0; \
|
|
int calcNumUsedCells=0; \
|
|
|
|
|
|
|
|
|
|
// -----------------------------------------------------------------------------------
|
|
// serial stuff
|
|
#if PARALLEL!=1
|
|
|
|
#define PERFORM_USQRMAXCHECK USQRMAXCHECK(usqr,ux,uy,uz, mMaxVlen, mMxvx,mMxvy,mMxvz);
|
|
#define LIST_EMPTY(x) mListEmpty.push_back( x );
|
|
#define LIST_FULL(x) mListFull.push_back( x );
|
|
#define FSGR_ADDPART(x) mpParticles->addFullParticle( x );
|
|
|
|
// >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
|
|
#define GRID_REGION_START() \
|
|
{ /* main_region */ \
|
|
int kstart=getForZMinBnd(), kend=getForZMaxBnd(mMaxRefine); \
|
|
if(gridLoopBound>0){ kstart=getForZMin1(), kend=getForZMax1(mMaxRefine); } \
|
|
int kdir = 1; \
|
|
int jstart = gridLoopBound; \
|
|
int jend = mLevel[mMaxRefine].lSizey-gridLoopBound; \
|
|
const int id=0; \
|
|
LbmFloat *ccel = NULL, *tcel = NULL; \
|
|
CellFlagType *pFlagSrc=NULL, *pFlagDst=NULL; \
|
|
if(mLevel[mMaxRefine].setCurr==1) { \
|
|
kdir = -1; \
|
|
int temp = kend; \
|
|
kend = kstart-1; \
|
|
kstart = temp-1; \
|
|
temp = id; /* dummy remove warning */ \
|
|
} \
|
|
|
|
|
|
|
|
|
|
|
|
// -----------------------------------------------------------------------------------
|
|
#else // PARALLEL==1
|
|
|
|
//#include "paraloop.h"
|
|
#define PERFORM_USQRMAXCHECK USQRMAXCHECK(usqr,ux,uy,uz, calcMaxVlen, calcMxvx,calcMxvy,calcMxvz);
|
|
#define LIST_EMPTY(x) calcListEmpty.push_back( x );
|
|
#define LIST_FULL(x) calcListFull.push_back( x );
|
|
#define FSGR_ADDPART(x) calcListParts.push_back( x );
|
|
|
|
|
|
// parallel region
|
|
//was: # pragma omp parallel default(shared)
|
|
#if COMPRESSGRIDS!=1
|
|
// requires compressed grids...!
|
|
ERROR!
|
|
#endif
|
|
|
|
// loop start
|
|
#define GRID_REGION_START() \
|
|
{ \
|
|
\
|
|
\
|
|
if(mSizez<2) { \
|
|
mPanic = 1; \
|
|
errFatal("ParaLoop::2D","Not valid...!", SIMWORLD_GENERICERROR); \
|
|
} \
|
|
\
|
|
\
|
|
vector<LbmPoint> calcListFull; \
|
|
vector<LbmPoint> calcListEmpty; \
|
|
vector<ParticleObject> calcListParts; \
|
|
LbmFloat calcMxvx, calcMxvy, calcMxvz, calcMaxVlen; \
|
|
calcMxvx = calcMxvy = calcMxvz = calcMaxVlen = 0.0; \
|
|
calcListEmpty.reserve(mListEmpty.capacity() / omp_get_num_threads() ); \
|
|
calcListFull.reserve( mListFull.capacity() / omp_get_num_threads() ); \
|
|
calcListParts.reserve(mSizex); \
|
|
\
|
|
\
|
|
const int id = omp_get_thread_num(); \
|
|
const int Nthrds = omp_get_num_threads(); \
|
|
\
|
|
\
|
|
\
|
|
\
|
|
\
|
|
int kdir = 1; \
|
|
\
|
|
int kstart=getForZMinBnd(), kend=getForZMaxBnd(mMaxRefine); \
|
|
if(gridLoopBound>0){ kstart=getForZMin1(); kend=getForZMax1(mMaxRefine); } \
|
|
LbmFloat *ccel = NULL, *tcel = NULL; \
|
|
CellFlagType *pFlagSrc=NULL, *pFlagDst=NULL; \
|
|
\
|
|
\
|
|
if(mLevel[mMaxRefine].setCurr==1) { \
|
|
kdir = -1; \
|
|
int temp = kend; \
|
|
kend = kstart-1; \
|
|
kstart = temp-1; \
|
|
} \
|
|
\
|
|
const int Nj = mLevel[mMaxRefine].lSizey; \
|
|
int jstart = 0+( id * (Nj / Nthrds) ); \
|
|
int jend = 0+( (id+1) * (Nj / Nthrds) ); \
|
|
if( ((Nj/Nthrds) *Nthrds) != Nj) { \
|
|
errMsg("LbmFsgrSolver","Invalid domain size Nj="<<Nj<<" Nthrds="<<Nthrds); \
|
|
} \
|
|
\
|
|
if(jstart<gridLoopBound) jstart = gridLoopBound; \
|
|
if(jend>mLevel[mMaxRefine].lSizey-gridLoopBound) jend = mLevel[mMaxRefine].lSizey-gridLoopBound; \
|
|
\
|
|
debMsgStd("ParaLoop::OMP",DM_MSG,"Thread:"<<id<<" i:"<<istart<<"-"<<iend<<" j:"<<jstart<<"-"<<jend<<", k:"<<kstart<<"-"<<kend<<" ", 1); \
|
|
\
|
|
|
|
|
|
|
|
|
|
// para GRID LOOP END is parainc3
|
|
|
|
#endif // PARALLEL==1
|
|
|
|
|
|
// -----------------------------------------------------------------------------------
|
|
|
|
// >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
|
|
#define GRID_LOOP_START() \
|
|
for(int k=kstart;k!=kend;k+=kdir) { \
|
|
pFlagSrc = &RFLAG(lev, istart, jstart, k, SRCS(lev)); \
|
|
pFlagDst = &RFLAG(lev, istart, jstart, k, TSET(lev)); \
|
|
ccel = RACPNT(lev, istart, jstart, k, SRCS(lev)); \
|
|
tcel = RACPNT(lev, istart, jstart, k, TSET(lev)); \
|
|
for(int j=jstart;j!=jend;++j) { \
|
|
/* for(int i=0;i<mLevel[lev].lSizex-2; ) { */ \
|
|
for(int i=istart;i!=iend; ) { \
|
|
ADVANCE_POINTERS(1); \
|
|
|
|
|
|
|
|
|
|
// >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
|
|
#define GRID_LOOPREG_END() \
|
|
\
|
|
} /* i */ \
|
|
int i=0; \
|
|
ADVANCE_POINTERS(2*gridLoopBound); \
|
|
} /* j */ \
|
|
/* COMPRESSGRIDS!=1 */ \
|
|
/* int i=0; */ \
|
|
/* ADVANCE_POINTERS(mLevel[lev].lSizex*2); */ \
|
|
} /* all cell loop k,j,i */ \
|
|
if(doReduce) { } /* dummy remove warning */ \
|
|
} /* main_region */ \
|
|
\
|
|
|
|
|
|
|
|
|
|
// old loop for COMPRESSGRIDS==0
|
|
#define old__GRID_LOOP_START() \
|
|
for(int k=kstart;k<kend;++k) { \
|
|
for(int j=1;j<mLevel[lev].lSizey-1;++j) { \
|
|
for(int i=0;i<mLevel[lev].lSizex-2; ) {
|
|
|
|
|