void save_soln_host(const char *userSubroutine,op_set set,op_arg opDat1,op_arg opDat2)
{
size_t blocksPerGrid;
size_t threadsPerBlock;
size_t totalThreadNumber;
size_t dynamicSharedMemorySize;
cl_int errorCode;
cl_event event;
cl_kernel kernelPointer;
int sharedMemoryOffset;
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
blocksPerGrid = 200;
threadsPerBlock = threadsPerBlockSize_save_soln;
totalThreadNumber = threadsPerBlock * blocksPerGrid;
dynamicSharedMemorySize = 0;
dynamicSharedMemorySize = MAX(dynamicSharedMemorySize,sizeof(float ) * 4);
dynamicSharedMemorySize = MAX(dynamicSharedMemorySize,sizeof(float ) * 4);
sharedMemoryOffset = dynamicSharedMemorySize * OP_WARPSIZE;
dynamicSharedMemorySize = dynamicSharedMemorySize * threadsPerBlock;
kernelPointer = getKernel("save_soln_kernel");
errorCode = clSetKernelArg(kernelPointer,0,sizeof(cl_mem ),&opDat1.data_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,1,sizeof(cl_mem ),&opDat2.data_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,2,sizeof(int ),&sharedMemoryOffset);
errorCode = errorCode | clSetKernelArg(kernelPointer,3,sizeof(int ),&set -> size);
//errorCode = errorCode | clSetKernelArg(kernelPointer,4,sizeof(size_t ),&dynamicSharedMemorySize);
errorCode = errorCode | clSetKernelArg(kernelPointer,4,dynamicSharedMemorySize,NULL);
//printf("errorCode after 5: %d\n", errorCode);
assert_m(errorCode == CL_SUCCESS,"Error setting OpenCL kernel arguments save_soln");
errorCode = clEnqueueNDRangeKernel(cqCommandQueue,kernelPointer,1,NULL,&totalThreadNumber,&threadsPerBlock,0,NULL,&event);
assert_m(errorCode == CL_SUCCESS,"Error executing OpenCL kernel save_soln");
errorCode = clFinish(cqCommandQueue);
assert_m(errorCode == CL_SUCCESS,"Error completing device command queue");
#ifdef PROFILE
unsigned long tqueue, tsubmit, tstart, tend, telapsed;
ciErrNum = clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_QUEUED, sizeof(tqueue), &tqueue, NULL );
ciErrNum |= clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_SUBMIT, sizeof(tsubmit), &tsubmit, NULL );
ciErrNum |= clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_START, sizeof(tstart), &tstart, NULL );
ciErrNum |= clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_END, sizeof(tend), &tend, NULL );
assert_m( ciErrNum == CL_SUCCESS, "error getting profiling info" );
OP_kernels[0].queue_time += (tsubmit - tqueue);
OP_kernels[0].wait_time += (tstart - tsubmit);
OP_kernels[0].execution_time += (tend - tstart);
//printf("%20lu\n%20lu\n%20lu\n%20lu\n\n", tqueue, tsubmit, tstart, tend);
//printf("queue: %8.4f\nwait:%8.4f\nexec: %8.4f\n\n", OP_kernels[0].queue_time * 1.0e-9, OP_kernels[0].wait_time * 1.0e-9, OP_kernels[0].execution_time * 1.0e-9 );
#endif
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(0);
OP_kernels[0].name = userSubroutine;
OP_kernels[0].count += 1;
OP_kernels[0].time += wall_t2 - wall_t1;
OP_kernels[0].transfer += (float)set->size * opDat1.size;
OP_kernels[0].transfer += (float)set->size * opDat2.size;
}
void op_par_loop_save_soln(char const *name, op_set set,
op_arg arg0,
op_arg arg1 ) {
if (OP_diags>2) {
printf(" kernel routine w/o indirection: save_soln \n");
}
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// set number of threads
#ifdef _OPENMP
int nthreads = omp_get_max_threads( );
#else
int nthreads = 1;
#endif
// execute plan
#pragma omp parallel for
for (int thr=0; thr<nthreads; thr++) {
int start = (set->size* thr )/nthreads;
int finish = (set->size*(thr+1))/nthreads;
op_x86_save_soln( (float *) arg0.data,
(float *) arg1.data,
start, finish );
}
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(0);
OP_kernels[0].name = name;
OP_kernels[0].count += 1;
OP_kernels[0].time += wall_t2 - wall_t1;
OP_kernels[0].transfer += (float)set->size * arg0.size;
OP_kernels[0].transfer += (float)set->size * arg1.size;
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
int niter;
double rms;
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
char file[] = "new_grid.h5";
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set_hdf5(file, "nodes");
op_set edges = op_decl_set_hdf5(file, "edges");
op_set bedges = op_decl_set_hdf5(file, "bedges");
op_set cells = op_decl_set_hdf5(file, "cells");
op_map pedge = op_decl_map_hdf5(edges, nodes, 2, file, "pedge");
op_map pecell = op_decl_map_hdf5(edges, cells,2, file, "pecell");
op_map pbedge = op_decl_map_hdf5(bedges,nodes,2, file, "pbedge");
op_map pbecell = op_decl_map_hdf5(bedges,cells,1, file, "pbecell");
op_map pcell = op_decl_map_hdf5(cells, nodes,4, file, "pcell");
op_map m_test = op_decl_map_hdf5(cells, nodes,4, file, "m_test");
if (m_test == NULL) printf("m_test not found\n");
op_dat p_bound = op_decl_dat_hdf5(bedges,1,"int" ,file,"p_bound");
op_dat p_x = op_decl_dat_hdf5(nodes ,2,"double",file,"p_x");
op_dat p_q = op_decl_dat_hdf5(cells ,4,"double",file,"p_q");
op_dat p_qold = op_decl_dat_hdf5(cells ,4,"double",file,"p_qold");
op_dat p_adt = op_decl_dat_hdf5(cells ,1,"double",file,"p_adt");
op_dat p_res = op_decl_dat_hdf5(cells ,4,"double",file,"p_res");
op_dat p_test = op_decl_dat_hdf5(cells ,4,"double",file,"p_test");
if (p_test == NULL) printf("p_test not found\n");
op_get_const_hdf5("gam", 1, "double", (char *)&gam, "new_grid.h5");
op_get_const_hdf5("gm1", 1, "double", (char *)&gm1, "new_grid.h5");
op_get_const_hdf5("cfl", 1, "double", (char *)&cfl, "new_grid.h5");
op_get_const_hdf5("eps", 1, "double", (char *)&eps, "new_grid.h5");
op_get_const_hdf5("mach", 1, "double", (char *)&mach, "new_grid.h5");
op_get_const_hdf5("alpha", 1, "double", (char *)&alpha, "new_grid.h5");
op_get_const_hdf5("qinf", 4, "double", (char *)&qinf, "new_grid.h5");
op_decl_const2("gam",1,"double",&gam);
op_decl_const2("gm1",1,"double",&gm1);
op_decl_const2("cfl",1,"double",&cfl);
op_decl_const2("eps",1,"double",&eps);
op_decl_const2("mach",1,"double",&mach);
op_decl_const2("alpha",1,"double",&alpha);
op_decl_const2("qinf",4,"double",qinf);
op_diagnostic_output();
//write back original data just to compare you read the file correctly
//do an h5diff between new_grid_out.h5 and new_grid.h5 to
//compare two hdf5 files
op_dump_to_hdf5("new_grid_out.h5");
op_write_const_hdf5("gam",1,"double",(char *)&gam, "new_grid_out.h5");
op_write_const_hdf5("gm1",1,"double",(char *)&gm1, "new_grid_out.h5");
op_write_const_hdf5("cfl",1,"double",(char *)&cfl, "new_grid_out.h5");
op_write_const_hdf5("eps",1,"double",(char *)&eps, "new_grid_out.h5");
op_write_const_hdf5("mach",1,"double",(char *)&mach, "new_grid_out.h5");
op_write_const_hdf5("alpha",1,"double",(char *)&alpha, "new_grid_out.h5");
op_write_const_hdf5("qinf",4,"double",(char *)qinf, "new_grid_out.h5");
//trigger partitioning and halo creation routines
op_partition("PTSCOTCH", "KWAY", edges, pecell, p_x);
//op_partition("PARMETIS", "KWAY", edges, pecell, p_x);
int g_ncell = op_get_size(cells);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
// save old flow solution
op_par_loop_save_soln("save_soln",cells,
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_qold,-1,OP_ID,4,"double",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop_adt_calc("adt_calc",cells,
op_arg_dat(p_x,0,pcell,2,"double",OP_READ),
op_arg_dat(p_x,1,pcell,2,"double",OP_READ),
op_arg_dat(p_x,2,pcell,2,"double",OP_READ),
op_arg_dat(p_x,3,pcell,2,"double",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_adt,-1,OP_ID,1,"double",OP_WRITE));
// calculate flux residual
op_par_loop_res_calc("res_calc",edges,
op_arg_dat(p_x,0,pedge,2,"double",OP_READ),
op_arg_dat(p_x,1,pedge,2,"double",OP_READ),
op_arg_dat(p_q,0,pecell,4,"double",OP_READ),
op_arg_dat(p_q,1,pecell,4,"double",OP_READ),
op_arg_dat(p_adt,0,pecell,1,"double",OP_READ),
op_arg_dat(p_adt,1,pecell,1,"double",OP_READ),
op_arg_dat(p_res,0,pecell,4,"double",OP_INC),
op_arg_dat(p_res,1,pecell,4,"double",OP_INC));
op_par_loop_bres_calc("bres_calc",bedges,
op_arg_dat(p_x,0,pbedge,2,"double",OP_READ),
op_arg_dat(p_x,1,pbedge,2,"double",OP_READ),
op_arg_dat(p_q,0,pbecell,4,"double",OP_READ),
op_arg_dat(p_adt,0,pbecell,1,"double",OP_READ),
op_arg_dat(p_res,0,pbecell,4,"double",OP_INC),
op_arg_dat(p_bound,-1,OP_ID,1,"int",OP_READ));
// update flow field
rms = 0.0;
op_par_loop_update("update",cells,
op_arg_dat(p_qold,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_WRITE),
op_arg_dat(p_res,-1,OP_ID,4,"double",OP_RW),
op_arg_dat(p_adt,-1,OP_ID,1,"double",OP_READ),
op_arg_gbl(&rms,1,"double",OP_INC));
}
// print iteration history
rms = sqrt(rms/(double)g_ncell);
if (iter%100 == 0)
op_printf(" %d %10.5e \n",iter,rms);
}
op_timers(&cpu_t2, &wall_t2);
//write given op_dat's indicated segment of data to a memory block in the order it was originally
//arranged (i.e. before partitioning and reordering)
double* q = (double *)op_malloc(sizeof(double)*op_get_size(cells)*4);
op_fetch_data_idx(p_q, q, 0, op_get_size(cells)-1);
free(q);
//write given op_dat's data to hdf5 file in the order it was originally arranged (i.e. before partitioning and reordering)
op_fetch_data_hdf5_file(p_q, "file_name.h5");
//printf("Root process = %d\n",op_is_root());
//output the result dat array to files
//op_dump_to_hdf5("new_grid_out.h5"); //writes data as it is held on each process (under MPI)
//compress using
// ~/hdf5/bin/h5repack -f GZIP=9 new_grid.h5 new_grid_pack.h5
op_timing_output();
op_printf("Max total runtime = %f\n",wall_t2-wall_t1);
op_exit();
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
int *becell, *ecell, *bound, *bedge, *edge, *cell;
double *x, *q, *qold, *adt, *res;
int nnode,ncell,nedge,nbedge,niter;
double rms;
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// read in grid
op_printf("reading in grid \n");
FILE *fp;
if ( (fp = fopen("./new_grid.dat","r")) == NULL) {
op_printf("can't open file new_grid.dat\n"); exit(-1);
}
if (fscanf(fp,"%d %d %d %d \n",&nnode, &ncell, &nedge, &nbedge) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
cell = (int *) malloc(4*ncell*sizeof(int));
edge = (int *) malloc(2*nedge*sizeof(int));
ecell = (int *) malloc(2*nedge*sizeof(int));
bedge = (int *) malloc(2*nbedge*sizeof(int));
becell = (int *) malloc( nbedge*sizeof(int));
bound = (int *) malloc( nbedge*sizeof(int));
x = (double *) malloc(2*nnode*sizeof(double));
q = (double *) malloc(4*ncell*sizeof(double));
qold = (double *) malloc(4*ncell*sizeof(double));
res = (double *) malloc(4*ncell*sizeof(double));
adt = (double *) malloc( ncell*sizeof(double));
for (int n=0; n<nnode; n++) {
if (fscanf(fp,"%lf %lf \n",&x[2*n], &x[2*n+1]) != 2) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<ncell; n++) {
if (fscanf(fp,"%d %d %d %d \n",&cell[4*n ], &cell[4*n+1],
&cell[4*n+2], &cell[4*n+3]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<nedge; n++) {
if (fscanf(fp,"%d %d %d %d \n",&edge[2*n], &edge[2*n+1],
&ecell[2*n],&ecell[2*n+1]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<nbedge; n++) {
if (fscanf(fp,"%d %d %d %d \n",&bedge[2*n],&bedge[2*n+1],
&becell[n], &bound[n]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
fclose(fp);
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
gam = 1.4f;
gm1 = gam - 1.0f;
cfl = 0.9f;
eps = 0.05f;
double mach = 0.4f;
double alpha = 3.0f*atan(1.0f)/45.0f;
double p = 1.0f;
double r = 1.0f;
double u = sqrt(gam*p/r)*mach;
double e = p/(r*gm1) + 0.5f*u*u;
qinf[0] = r;
qinf[1] = r*u;
qinf[2] = 0.0f;
qinf[3] = r*e;
for (int n=0; n<ncell; n++) {
for (int m=0; m<4; m++) {
q[4*n+m] = qinf[m];
res[4*n+m] = 0.0f;
}
}
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set(nnode, "nodes");
op_set edges = op_decl_set(nedge, "edges");
op_set bedges = op_decl_set(nbedge, "bedges");
op_set cells = op_decl_set(ncell, "cells");
op_map pedge = op_decl_map(edges, nodes,2,edge, "pedge");
op_map pecell = op_decl_map(edges, cells,2,ecell, "pecell");
op_map pbedge = op_decl_map(bedges,nodes,2,bedge, "pbedge");
op_map pbecell = op_decl_map(bedges,cells,1,becell,"pbecell");
op_map pcell = op_decl_map(cells, nodes,4,cell, "pcell");
op_dat p_bound = op_decl_dat(bedges,1,"int" ,bound,"p_bound");
op_dat p_x = op_decl_dat(nodes ,2,"double",x ,"p_x");
op_dat p_q = op_decl_dat(cells ,4,"double",q ,"p_q");
//op_dat p_qold = op_decl_dat(cells ,4,"double",qold ,"p_qold");
//op_dat p_adt = op_decl_dat(cells ,1,"double",adt ,"p_adt");
//op_dat p_res = op_decl_dat(cells ,4,"double",res ,"p_res");
// p_res, p_adt and p_qold now declared as a temp op_dats during
// the execution of the time-marching loop
op_decl_const2("gam",1,"double",&gam);
op_decl_const2("gm1",1,"double",&gm1);
op_decl_const2("cfl",1,"double",&cfl);
op_decl_const2("eps",1,"double",&eps);
op_decl_const2("mach",1,"double",&mach);
op_decl_const2("alpha",1,"double",&alpha);
op_decl_const2("qinf",4,"double",qinf);
op_diagnostic_output();
double g_ncell = op_get_size(cells);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
double* tmp_elem = NULL;
op_dat p_res = op_decl_dat_temp(cells ,4,"double",tmp_elem,"p_res");
op_dat p_adt = op_decl_dat_temp(cells ,1,"double",tmp_elem,"p_adt");
op_dat p_qold = op_decl_dat_temp(cells ,4,"double",qold ,"p_qold");
// save old flow solution
op_par_loop_save_soln("save_soln",cells,
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_qold,-1,OP_ID,4,"double",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop_adt_calc("adt_calc",cells,
op_arg_dat(p_x,0,pcell,2,"double",OP_READ),
op_arg_dat(p_x,1,pcell,2,"double",OP_READ),
op_arg_dat(p_x,2,pcell,2,"double",OP_READ),
op_arg_dat(p_x,3,pcell,2,"double",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_adt,-1,OP_ID,1,"double",OP_WRITE));
// calculate flux residual
op_par_loop_res_calc("res_calc",edges,
op_arg_dat(p_x,0,pedge,2,"double",OP_READ),
op_arg_dat(p_x,1,pedge,2,"double",OP_READ),
op_arg_dat(p_q,0,pecell,4,"double",OP_READ),
op_arg_dat(p_q,1,pecell,4,"double",OP_READ),
op_arg_dat(p_adt,0,pecell,1,"double",OP_READ),
op_arg_dat(p_adt,1,pecell,1,"double",OP_READ),
op_arg_dat(p_res,0,pecell,4,"double",OP_INC),
op_arg_dat(p_res,1,pecell,4,"double",OP_INC));
op_par_loop_bres_calc("bres_calc",bedges,
op_arg_dat(p_x,0,pbedge,2,"double",OP_READ),
op_arg_dat(p_x,1,pbedge,2,"double",OP_READ),
op_arg_dat(p_q,0,pbecell,4,"double",OP_READ),
op_arg_dat(p_adt,0,pbecell,1,"double",OP_READ),
op_arg_dat(p_res,0,pbecell,4,"double",OP_INC),
op_arg_dat(p_bound,-1,OP_ID,1,"int",OP_READ));
// update flow field
rms = 0.0;
op_par_loop_update("update",cells,
op_arg_dat(p_qold,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_WRITE),
op_arg_dat(p_res,-1,OP_ID,4,"double",OP_RW),
op_arg_dat(p_adt,-1,OP_ID,1,"double",OP_READ),
op_arg_gbl(&rms,1,"double",OP_INC));
}
// print iteration history
rms = sqrt(rms/(double)g_ncell );
if (iter%100 == 0)
op_printf(" %d %10.5e \n",iter,rms);
if (iter%1000 == 0 && g_ncell == 720000){ //defailt mesh -- for validation testing
//op_printf(" %d %3.16f \n",iter,rms);
double diff=fabs((100.0*(rms/0.0001060114637578))-100.0);
op_printf("\n\nTest problem with %d cells is within %3.15E %% of the expected solution\n",720000, diff);
if(diff < 0.00001) {
op_printf("This test is considered PASSED\n");
}
else {
op_printf("This test is considered FAILED\n");
}
}
if (op_free_dat_temp(p_res) < 0)
op_printf("Error: temporary op_dat %s cannot be removed\n",p_res->name);
if (op_free_dat_temp(p_adt) < 0)
op_printf("Error: temporary op_dat %s cannot be removed\n",p_adt->name);
if (op_free_dat_temp(p_qold) < 0)
op_printf("Error: temporary op_dat %s cannot be removed\n",p_qold->name);
}
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_printf("Max total runtime = %f\n",wall_t2-wall_t1);
op_exit();
free(cell);
free(edge);
free(ecell);
free(bedge);
free(becell);
free(bound);
free(x);
free(q);
free(qold);
free(res);
free(adt);
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
int *becell, *ecell, *bound, *bedge, *edge, *cell;
float *x, *q, *qold, *adt, *res;
int nnode,ncell,nedge,nbedge,niter;
float rms;
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// read in grid
op_printf("reading in grid \n");
FILE *fp;
if ( (fp = fopen("./new_grid.dat","r")) == NULL) {
op_printf("can't open file new_grid.dat\n"); exit(-1);
}
if (fscanf(fp,"%d %d %d %d \n",&nnode, &ncell, &nedge, &nbedge) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
cell = (int *) malloc(4*ncell*sizeof(int));
edge = (int *) malloc(2*nedge*sizeof(int));
ecell = (int *) malloc(2*nedge*sizeof(int));
bedge = (int *) malloc(2*nbedge*sizeof(int));
becell = (int *) malloc( nbedge*sizeof(int));
bound = (int *) malloc( nbedge*sizeof(int));
x = (float *) malloc(2*nnode*sizeof(float));
q = (float *) malloc(4*ncell*sizeof(float));
qold = (float *) malloc(4*ncell*sizeof(float));
res = (float *) malloc(4*ncell*sizeof(float));
adt = (float *) malloc( ncell*sizeof(float));
for (int n=0; n<nnode; n++) {
if (fscanf(fp,"%f %f \n",&x[2*n], &x[2*n+1]) != 2) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<ncell; n++) {
if (fscanf(fp,"%d %d %d %d \n",&cell[4*n ], &cell[4*n+1],
&cell[4*n+2], &cell[4*n+3]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<nedge; n++) {
if (fscanf(fp,"%d %d %d %d \n",&edge[2*n], &edge[2*n+1],
&ecell[2*n],&ecell[2*n+1]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<nbedge; n++) {
if (fscanf(fp,"%d %d %d %d \n",&bedge[2*n],&bedge[2*n+1],
&becell[n], &bound[n]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
fclose(fp);
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
gam = 1.4f;
gm1 = gam - 1.0f;
cfl = 0.9f;
eps = 0.05f;
float mach = 0.4f;
float alpha = 3.0f*atan(1.0f)/45.0f;
float p = 1.0f;
float r = 1.0f;
float u = sqrt(gam*p/r)*mach;
float e = p/(r*gm1) + 0.5f*u*u;
qinf[0] = r;
qinf[1] = r*u;
qinf[2] = 0.0f;
qinf[3] = r*e;
for (int n=0; n<ncell; n++) {
for (int m=0; m<4; m++) {
q[4*n+m] = qinf[m];
res[4*n+m] = 0.0f;
}
}
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set(nnode, "nodes");
op_set edges = op_decl_set(nedge, "edges");
op_set bedges = op_decl_set(nbedge, "bedges");
op_set cells = op_decl_set(ncell, "cells");
op_map pedge = op_decl_map(edges, nodes,2,edge, "pedge");
op_map pecell = op_decl_map(edges, cells,2,ecell, "pecell");
op_map pbedge = op_decl_map(bedges,nodes,2,bedge, "pbedge");
op_map pbecell = op_decl_map(bedges,cells,1,becell,"pbecell");
op_map pcell = op_decl_map(cells, nodes,4,cell, "pcell");
op_dat p_bound = op_decl_dat(bedges,1,"int" ,bound,"p_bound");
op_dat p_x = op_decl_dat(nodes ,2,"float",x ,"p_x");
op_dat p_q = op_decl_dat(cells ,4,"float",q ,"p_q");
op_dat p_qold = op_decl_dat(cells ,4,"float",qold ,"p_qold");
op_dat p_adt = op_decl_dat(cells ,1,"float",adt ,"p_adt");
op_dat p_res = op_decl_dat(cells ,4,"float",res ,"p_res");
op_decl_const2("gam",1,"float",&gam);
op_decl_const2("gm1",1,"float",&gm1);
op_decl_const2("cfl",1,"float",&cfl);
op_decl_const2("eps",1,"float",&eps);
op_decl_const2("mach",1,"float",&mach);
op_decl_const2("alpha",1,"float",&alpha);
op_decl_const2("qinf",4,"float",qinf);
op_diagnostic_output();
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
// save old flow solution
op_par_loop_save_soln("save_soln",cells,
op_arg_dat(p_q,-1,OP_ID,4,"float",OP_READ),
op_arg_dat(p_qold,-1,OP_ID,4,"float",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop_adt_calc("adt_calc",cells,
op_arg_dat(p_x,0,pcell,2,"float",OP_READ),
op_arg_dat(p_x,1,pcell,2,"float",OP_READ),
op_arg_dat(p_x,2,pcell,2,"float",OP_READ),
op_arg_dat(p_x,3,pcell,2,"float",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"float",OP_READ),
op_arg_dat(p_adt,-1,OP_ID,1,"float",OP_WRITE));
// calculate flux residual
op_par_loop_res_calc("res_calc",edges,
op_arg_dat(p_x,0,pedge,2,"float",OP_READ),
op_arg_dat(p_x,1,pedge,2,"float",OP_READ),
op_arg_dat(p_q,0,pecell,4,"float",OP_READ),
op_arg_dat(p_q,1,pecell,4,"float",OP_READ),
op_arg_dat(p_adt,0,pecell,1,"float",OP_READ),
op_arg_dat(p_adt,1,pecell,1,"float",OP_READ),
op_arg_dat(p_res,0,pecell,4,"float",OP_INC),
op_arg_dat(p_res,1,pecell,4,"float",OP_INC));
op_par_loop_bres_calc("bres_calc",bedges,
op_arg_dat(p_x,0,pbedge,2,"float",OP_READ),
op_arg_dat(p_x,1,pbedge,2,"float",OP_READ),
op_arg_dat(p_q,0,pbecell,4,"float",OP_READ),
op_arg_dat(p_adt,0,pbecell,1,"float",OP_READ),
op_arg_dat(p_res,0,pbecell,4,"float",OP_INC),
op_arg_dat(p_bound,-1,OP_ID,1,"int",OP_READ));
// update flow field
rms = 0.0;
op_par_loop_update("update",cells,
op_arg_dat(p_qold,-1,OP_ID,4,"float",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"float",OP_WRITE),
op_arg_dat(p_res,-1,OP_ID,4,"float",OP_RW),
op_arg_dat(p_adt,-1,OP_ID,1,"float",OP_READ),
op_arg_gbl(&rms,1,"float",OP_INC));
}
// print iteration history
rms = sqrt(rms/(float) op_get_size(cells));
if (iter%100 == 0)
op_printf(" %d %10.5e \n",iter,rms);
}
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_printf("Max total runtime = \n%f\n",wall_t2-wall_t1);
op_exit();
free(cell);
free(edge);
free(ecell);
free(bedge);
free(becell);
free(bound);
free(x);
free(q);
free(qold);
free(res);
free(adt);
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
int *becell, *ecell, *bound, *bedge, *edge, *cell;
double *x, *q, *qold, *adt, *res;
int nnode,ncell,nedge,nbedge;
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// read in airfoil grid
op_printf("reading in data \n");
FILE *fp;
if ( (fp = fopen("./new_grid.dat","r")) == NULL) {
op_printf("can't open file new_grid.dat\n"); exit(-1);
}
if (fscanf(fp,"%d %d %d %d \n",&nnode, &ncell, &nedge, &nbedge) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
cell = (int *) malloc(4*ncell*sizeof(int));
edge = (int *) malloc(2*nedge*sizeof(int));
ecell = (int *) malloc(2*nedge*sizeof(int));
bedge = (int *) malloc(2*nbedge*sizeof(int));
becell = (int *) malloc( nbedge*sizeof(int));
bound = (int *) malloc( nbedge*sizeof(int));
x = (double *) malloc(2*nnode*sizeof(double));
q = (double *) malloc(4*ncell*sizeof(double));
qold = (double *) malloc(4*ncell*sizeof(double));
res = (double *) malloc(4*ncell*sizeof(double));
adt = (double *) malloc( ncell*sizeof(double));
for (int n=0; n<nnode; n++) {
if (fscanf(fp,"%lf %lf \n",&x[2*n], &x[2*n+1]) != 2) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<ncell; n++) {
if (fscanf(fp,"%d %d %d %d \n",&cell[4*n ], &cell[4*n+1],
&cell[4*n+2], &cell[4*n+3]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<nedge; n++) {
if (fscanf(fp,"%d %d %d %d \n",&edge[2*n], &edge[2*n+1],
&ecell[2*n],&ecell[2*n+1]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<nbedge; n++) {
if (fscanf(fp,"%d %d %d %d \n",&bedge[2*n],&bedge[2*n+1],
&becell[n], &bound[n]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
fclose(fp);
// declare sets, pointers, datasets
op_set edges = op_decl_set(nedge, "edges");
op_set cells = op_decl_set(ncell, "cells");
op_map pecell = op_decl_map(edges, cells,2,ecell, "pecell");
op_dat p_res = op_decl_dat(cells ,4,"double",res ,"p_res");
int count;
op_diagnostic_output();
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
//indirect reduction
count = 0;
op_par_loop_res_calc("res_calc",edges,
op_arg_dat(p_res,0,pecell,4,"double",OP_INC),
op_arg_gbl(&count,1,"int",OP_INC));
op_printf("number of edges:: %d should be: %d \n",count,nedge);
if (count != nedge) op_printf("indirect reduction FAILED\n");
else op_printf("indirect reduction PASSED\n");
//direct reduction
count = 0;
op_par_loop_update("update",cells,
op_arg_dat(p_res,-1,OP_ID,4,"double",OP_RW),
op_arg_gbl(&count,1,"int",OP_INC));
op_printf("number of cells: %d should be: %d \n",count,ncell);
if (count != ncell) op_printf("direct reduction FAILED\n");
else op_printf("direct reduction PASSED\n");
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_exit();
free(cell);
free(edge);
free(ecell);
free(bedge);
free(becell);
free(bound);
free(x);
free(q);
free(qold);
free(res);
free(adt);
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
//MPI for user I/O
int my_rank;
int comm_size;
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
int *pp;
double *A, *r, *u, *du;
int nnode, nedge;
/**------------------------BEGIN I/O and PARTITIONING ---------------------**/
int g_nnode, g_nedge, g_n, g_e;
g_nnode = (NN-1)*(NN-1);
g_nedge = (NN-1)*(NN-1) + 4*(NN-1)*(NN-2);
int *g_pp = 0;
double *g_A = 0, *g_r = 0, *g_u = 0, *g_du = 0;
op_printf("Global number of nodes, edges = %d, %d\n",g_nnode,g_nedge);
if(my_rank == MPI_ROOT) {
g_pp = (int *)malloc(sizeof(int)*2*g_nedge);
g_A = (double *)malloc(sizeof(double)*g_nedge);
g_r = (double *)malloc(sizeof(double)*g_nnode);
g_u = (double *)malloc(sizeof(double)*g_nnode);
g_du = (double *)malloc(sizeof(double)*g_nnode);
// create matrix and r.h.s., and set coordinates needed for renumbering / partitioning
g_e = 0;
for (int i=1; i<NN; i++) {
for (int j=1; j<NN; j++) {
g_n = i-1 + (j-1)*(NN-1);
g_r[g_n] = 0.0f;
g_u[g_n] = 0.0f;
g_du[g_n] = 0.0f;
g_pp[2*g_e] = g_n;
g_pp[2*g_e+1] = g_n;
g_A[g_e] = -1.0f;
g_e++;
for (int pass=0; pass<4; pass++) {
int i2 = i;
int j2 = j;
if (pass==0) i2 += -1;
if (pass==1) i2 += 1;
if (pass==2) j2 += -1;
if (pass==3) j2 += 1;
if ( (i2==0) || (i2==NN) || (j2==0) || (j2==NN) ) {
g_r[g_n] += 0.25f;
}
else {
g_pp[2*g_e] = g_n;
g_pp[2*g_e+1] = i2-1 + (j2-1)*(NN-1);
g_A[g_e] = 0.25f;
g_e++;
}
}
}
}
}
/* Compute local sizes */
nnode = compute_local_size (g_nnode, comm_size, my_rank);
nedge = compute_local_size (g_nedge, comm_size, my_rank);
op_printf("Number of nodes, edges on process %d = %d, %d\n"
,my_rank,nnode,nedge);
/*Allocate memory to hold local sets, mapping tables and data*/
pp = (int *)malloc(2*sizeof(int)*nedge);
A = (double *) malloc(nedge*sizeof(double));
r = (double *) malloc(nnode*sizeof(double));
u = (double *) malloc(nnode*sizeof(double));
du = (double *) malloc(nnode*sizeof(double));
/* scatter sets, mappings and data on sets*/
scatter_int_array(g_pp, pp, comm_size, g_nedge,nedge, 2);
scatter_double_array(g_A, A, comm_size, g_nedge,nedge, 1);
scatter_double_array(g_r, r, comm_size, g_nnode,nnode, 1);
scatter_double_array(g_u, u, comm_size, g_nnode,nnode, 1);
scatter_double_array(g_du, du, comm_size, g_nnode,nnode, 1);
/*Freeing memory allocated to gloabal arrays on rank 0
after scattering to all processes*/
if(my_rank == MPI_ROOT) {
free(g_pp);
free(g_A);
free(g_r);
free(g_u);
free(g_du);
}
/**------------------------END I/O and PARTITIONING ---------------------**/
// declare sets, pointers, and datasets
op_set nodes = op_decl_set(nnode,"nodes");
op_set edges = op_decl_set(nedge,"edges");
op_map ppedge = op_decl_map(edges,nodes,2,pp, "ppedge");
op_dat p_A = op_decl_dat(edges,1,"double", A, "p_A" );
op_dat p_r = op_decl_dat(nodes,1,"double", r, "p_r" );
op_dat p_u = op_decl_dat(nodes,1,"double", u, "p_u" );
op_dat p_du = op_decl_dat(nodes,1,"double", du,"p_du");
alpha = 1.0f;
op_decl_const(1,"double",&alpha);
op_diagnostic_output();
//trigger partitioning and halo creation routines
op_partition("PTSCOTCH", "KWAY", NULL, NULL, NULL);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main iteration loop
double u_sum, u_max, beta = 1.0f;
for (int iter=0; iter<NITER; iter++) {
op_par_loop(res,"res", edges,
op_arg_dat(p_A, -1,OP_ID, 1,"double", OP_READ),
op_arg_dat(p_u, 1,ppedge, 1,"double", OP_READ),
op_arg_dat(p_du, 0,ppedge, 1,"double", OP_INC),
op_arg_gbl(&beta, 1,"double", OP_READ));
u_sum = 0.0f;
u_max = 0.0f;
op_par_loop(update,"update", nodes,
op_arg_dat(p_r, -1,OP_ID, 1,"double",OP_READ),
op_arg_dat(p_du, -1,OP_ID, 1,"double",OP_RW),
op_arg_dat(p_u, -1,OP_ID, 1,"double",OP_INC),
op_arg_gbl(&u_sum,1,"double",OP_INC),
op_arg_gbl(&u_max,1,"double",OP_MAX));
op_printf("\n u max/rms = %f %f \n\n",u_max, sqrt(u_sum/g_nnode));
}
op_timers(&cpu_t2, &wall_t2);
//get results data array
op_dat temp = op_mpi_get_data(p_u);
//output the result dat array to files
print_dat_tofile(temp, "out_grid.dat"); //ASCI
//print_dat_tobinfile(temp, "out_grid.bin"); //Binary
//print each mpi process's timing info for each kernel
op_timing_output();
//print total time for niter interations
op_printf("Max total runtime = %f\n",wall_t2-wall_t1);
op_exit();
}
void op_par_loop_adt_calc(char const *name, op_set set,
op_arg arg0,
op_arg arg1,
op_arg arg2,
op_arg arg3,
op_arg arg4,
op_arg arg5 ){
int nargs = 6;
op_arg args[6] = {arg0,arg1,arg2,arg3,arg4,arg5};
int ninds = 1;
int inds[6] = {0,0,0,0,-1,-1};
int sent[6] = {0,0,0,0,0,0};
if(ninds > 0) //indirect loop
{
for(int i = 0; i<nargs; i++)
{
if(args[i].argtype == OP_ARG_DAT)
{
if (OP_diags==1) reset_halo(args[i]);
sent[0] = exchange_halo(args[i]);
if(sent[0] == 1)wait_all(args[i]);
}
}
}
if (OP_diags>2) {
printf(" kernel routine with indirection: adt_calc \n");
}
// get plan
#ifdef OP_PART_SIZE_1
int part_size = OP_PART_SIZE_1;
#else
int part_size = OP_part_size;
#endif
op_plan *Plan = op_plan_get(name,set,part_size,nargs,args,ninds,inds);
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// set number of threads
#ifdef _OPENMP
int nthreads = omp_get_max_threads( );
#else
int nthreads = 1;
#endif
// execute plan
int block_offset = 0;
for (int col=0; col < Plan->ncolors; col++) {
int nblocks = Plan->ncolblk[col];
#pragma omp parallel for
for (int blockIdx=0; blockIdx<nblocks; blockIdx++)
op_x86_adt_calc( blockIdx,
(double *)arg0.data, Plan->ind_maps[0],
Plan->loc_maps[0],
Plan->loc_maps[1],
Plan->loc_maps[2],
Plan->loc_maps[3],
(double *)arg4.data,
(double *)arg5.data,
Plan->ind_sizes,
Plan->ind_offs,
block_offset,
Plan->blkmap,
Plan->offset,
Plan->nelems,
Plan->nthrcol,
Plan->thrcol);
block_offset += nblocks;
}
//set dirty bit on direct/indirect datasets with access OP_INC,OP_WRITE, OP_RW
for(int i = 0; i<nargs; i++)
if(args[i].argtype == OP_ARG_DAT)
set_dirtybit(args[i]);
//performe any global operations
// - NONE
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(1);
OP_kernels[1].name = name;
OP_kernels[1].count += 1;
OP_kernels[1].time += wall_t2 - wall_t1;
OP_kernels[1].transfer += Plan->transfer;
OP_kernels[1].transfer2 += Plan->transfer2;
}
int main(int argc, char **argv) {
// OP initialisation
op_init(argc, argv, 5);
// timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
int nnode, nedge, n, e;
nnode = (NN - 1) * (NN - 1);
nedge = (NN - 1) * (NN - 1) + 4 * (NN - 1) * (NN - 2);
int *pp = (int *)malloc(sizeof(int) * 2 * nedge);
double *A = (double *)malloc(sizeof(double) * nedge);
double *r = (double *)malloc(sizeof(double) * nnode);
double *u = (double *)malloc(sizeof(double) * nnode);
double *du = (double *)malloc(sizeof(double) * nnode);
// create matrix and r.h.s., and set coordinates needed for renumbering /
// partitioning
e = 0;
for (int i = 1; i < NN; i++) {
for (int j = 1; j < NN; j++) {
n = i - 1 + (j - 1) * (NN - 1);
r[n] = 0.0f;
u[n] = 0.0f;
du[n] = 0.0f;
pp[2 * e] = n;
pp[2 * e + 1] = n;
A[e] = -1.0f;
e++;
for (int pass = 0; pass < 4; pass++) {
int i2 = i;
int j2 = j;
if (pass == 0)
i2 += -1;
if (pass == 1)
i2 += 1;
if (pass == 2)
j2 += -1;
if (pass == 3)
j2 += 1;
if ((i2 == 0) || (i2 == NN) || (j2 == 0) || (j2 == NN)) {
r[n] += 0.25f;
} else {
pp[2 * e] = n;
pp[2 * e + 1] = i2 - 1 + (j2 - 1) * (NN - 1);
A[e] = 0.25f;
e++;
}
}
}
}
// declare sets, pointers, and datasets
op_set nodes = op_decl_set(nnode, "nodes");
op_set edges = op_decl_set(nedge, "edges");
op_map ppedge = op_decl_map(edges, nodes, 2, pp, "ppedge");
op_dat p_A = op_decl_dat(edges, 1, "double", A, "p_A");
op_dat p_r = op_decl_dat(nodes, 1, "double", r, "p_r");
op_dat p_u = op_decl_dat(nodes, 1, "double", u, "p_u");
op_dat p_du = op_decl_dat(nodes, 1, "double", du, "p_du");
alpha = 1.0f;
op_decl_const2("alpha", 1, "double", &alpha);
op_diagnostic_output();
// initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main iteration loop
double u_sum, u_max, beta = 1.0f;
for (int iter = 0; iter < NITER; iter++) {
op_par_loop_res("res", edges,
op_arg_dat(p_A, -1, OP_ID, 1, "double", OP_READ),
op_arg_dat(p_u, 1, ppedge, 1, "double", OP_READ),
op_arg_dat(p_du, 0, ppedge, 1, "double", OP_INC),
op_arg_gbl(&beta, 1, "double", OP_READ));
u_sum = 0.0f;
u_max = 0.0f;
op_par_loop_update("update", nodes,
op_arg_dat(p_r, -1, OP_ID, 1, "double", OP_READ),
op_arg_dat(p_du, -1, OP_ID, 1, "double", OP_RW),
op_arg_dat(p_u, -1, OP_ID, 1, "double", OP_INC),
op_arg_gbl(&u_sum, 1, "double", OP_INC),
op_arg_gbl(&u_max, 1, "double", OP_MAX));
op_printf("\n u max/rms = %f %f \n\n", u_max, sqrt(u_sum / nnode));
}
op_timers(&cpu_t2, &wall_t2);
// print out results
op_printf("\n Results after %d iterations:\n\n", NITER);
op_fetch_data(p_u, u);
for (int pass = 0; pass < 1; pass++) {
for (int j = NN - 1; j > 0; j--) {
for (int i = 1; i < NN; i++) {
if (pass == 0)
op_printf(" %7.4f", u[i - 1 + (j - 1) * (NN - 1)]);
else if (pass == 1)
op_printf(" %7.4f", du[i - 1 + (j - 1) * (NN - 1)]);
else if (pass == 2)
op_printf(" %7.4f", r[i - 1 + (j - 1) * (NN - 1)]);
}
op_printf("\n");
}
op_printf("\n");
}
op_timing_output();
// print total time for niter interations
op_printf("Max total runtime = %f\n", wall_t2 - wall_t1);
int result = check_result<double>(u, NN, TOLERANCE);
op_exit();
free(pp);
free(A);
free(u);
free(du);
free(r);
return result;
}
void op_par_loop_update(char const *name, op_set set,
op_arg arg0,
op_arg arg1,
op_arg arg2,
op_arg arg3,
op_arg arg4 ){
int ninds = 0;
int nargs = 5;
op_arg args[5] = {arg0,arg1,arg2,arg3,arg4};
double *arg4h = (double *)arg4.data;
if (OP_diags>2) {
printf(" kernel routine w/o indirection: update \n");
}
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// set number of threads
#ifdef _OPENMP
int nthreads = omp_get_max_threads( );
#else
int nthreads = 1;
#endif
// allocate and initialise arrays for global reduction
double arg4_l[1+64*64];
for (int thr=0; thr<nthreads; thr++)
for (int d=0; d<1; d++) arg4_l[d+thr*64]=ZERO_double;
// execute plan
#pragma omp parallel for
for (int thr=0; thr<nthreads; thr++) {
int start = (set->size* thr )/nthreads;
int finish = (set->size*(thr+1))/nthreads;
op_x86_update( (double *) arg0.data,
(double *) arg1.data,
(double *) arg2.data,
(double *) arg3.data,
arg4_l + thr*64,
start, finish );
}
// combine reduction data
for (int thr=0; thr<nthreads; thr++)
for(int d=0; d<1; d++) arg4h[d] += arg4_l[d+thr*64];
//set dirty bit on direct/indirect datasets with access OP_INC,OP_WRITE, OP_RW
for(int i = 0; i<nargs; i++)
if(args[i].argtype == OP_ARG_DAT)
set_dirtybit(args[i]);
//performe any global operations
for(int i = 0; i<nargs; i++)
if(args[i].argtype == OP_ARG_GBL)
global_reduce(&args[i]);
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(4);
OP_kernels[4].name = name;
OP_kernels[4].count += 1;
OP_kernels[4].time += wall_t2 - wall_t1;
OP_kernels[4].transfer += (double)set->size * arg0.size;
OP_kernels[4].transfer += (double)set->size * arg1.size;
OP_kernels[4].transfer += (double)set->size * arg2.size * 2.0f;
OP_kernels[4].transfer += (double)set->size * arg3.size;
}
void bres_calc_host(const char *userSubroutine,op_set set,op_arg opDat1,op_arg opDat2,op_arg opDat3,op_arg opDat4,op_arg opDat5,op_arg opDat6)
{
size_t blocksPerGrid;
size_t threadsPerBlock;
size_t totalThreadNumber;
size_t dynamicSharedMemorySize;
cl_int errorCode;
cl_event event;
cl_kernel kernelPointer;
int i3;
op_arg opDatArray[6];
int indirectionDescriptorArray[6];
op_plan *planRet;
int blockOffset;
opDatArray[0] = opDat1;
opDatArray[1] = opDat2;
opDatArray[2] = opDat3;
opDatArray[3] = opDat4;
opDatArray[4] = opDat5;
opDatArray[5] = opDat6;
indirectionDescriptorArray[0] = 0;
indirectionDescriptorArray[1] = 0;
indirectionDescriptorArray[2] = 1;
indirectionDescriptorArray[3] = 2;
indirectionDescriptorArray[4] = 3;
indirectionDescriptorArray[5] = -1;
planRet = op_plan_get(userSubroutine,set,setPartitionSize_bres_calc,6,opDatArray,4,indirectionDescriptorArray);
cl_mem gm1_d;
gm1_d = op_allocate_constant(&gm1,sizeof(float ));
cl_mem qinf_d;
qinf_d = op_allocate_constant(&qinf,4 * sizeof(float));
cl_mem eps_d;
eps_d = op_allocate_constant(&eps,sizeof(float ));
blockOffset = 0;
double cpu_t1;
double cpu_t2;
double wall_t1;
op_timers(&cpu_t1, &wall_t1);
double wall_t2;
for (i3 = 0; i3 < planRet -> ncolors; ++i3) {
blocksPerGrid = planRet -> ncolblk[i3];
dynamicSharedMemorySize = planRet -> nshared;
threadsPerBlock = threadsPerBlockSize_bres_calc;
totalThreadNumber = threadsPerBlock * blocksPerGrid;
kernelPointer = getKernel("bres_calc_kernel");
errorCode = clSetKernelArg(kernelPointer,0,sizeof(cl_mem ),&opDat1.data_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,1,sizeof(cl_mem ),&opDat3.data_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,2,sizeof(cl_mem ),&opDat4.data_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,3,sizeof(cl_mem ),&opDat5.data_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,4,sizeof(cl_mem ),&opDat6.data_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,5,sizeof(cl_mem ),&planRet -> ind_maps[0]);
errorCode = errorCode | clSetKernelArg(kernelPointer,6,sizeof(cl_mem ),&planRet -> ind_maps[1]);
errorCode = errorCode | clSetKernelArg(kernelPointer,7,sizeof(cl_mem ),&planRet -> ind_maps[2]);
errorCode = errorCode | clSetKernelArg(kernelPointer,8,sizeof(cl_mem ),&planRet -> ind_maps[3]);
errorCode = errorCode | clSetKernelArg(kernelPointer,9,sizeof(cl_mem ),&planRet -> loc_maps[0]);
errorCode = errorCode | clSetKernelArg(kernelPointer,10,sizeof(cl_mem ),&planRet -> loc_maps[1]);
errorCode = errorCode | clSetKernelArg(kernelPointer,11,sizeof(cl_mem ),&planRet -> loc_maps[2]);
errorCode = errorCode | clSetKernelArg(kernelPointer,12,sizeof(cl_mem ),&planRet -> loc_maps[3]);
errorCode = errorCode | clSetKernelArg(kernelPointer,13,sizeof(cl_mem ),&planRet -> loc_maps[4]);
errorCode = errorCode | clSetKernelArg(kernelPointer,14,sizeof(cl_mem ),&planRet -> ind_sizes);
errorCode = errorCode | clSetKernelArg(kernelPointer,15,sizeof(cl_mem ),&planRet -> ind_offs);
errorCode = errorCode | clSetKernelArg(kernelPointer,16,sizeof(cl_mem ),&planRet -> blkmap);
errorCode = errorCode | clSetKernelArg(kernelPointer,17,sizeof(cl_mem ),&planRet -> offset);
errorCode = errorCode | clSetKernelArg(kernelPointer,18,sizeof(cl_mem ),&planRet -> nelems);
errorCode = errorCode | clSetKernelArg(kernelPointer,19,sizeof(cl_mem ),&planRet -> nthrcol);
errorCode = errorCode | clSetKernelArg(kernelPointer,20,sizeof(cl_mem ),&planRet -> thrcol);
errorCode = errorCode | clSetKernelArg(kernelPointer,21,sizeof(int ),&blockOffset);
errorCode = errorCode | clSetKernelArg(kernelPointer,22,dynamicSharedMemorySize,NULL);
errorCode = errorCode | clSetKernelArg(kernelPointer,23,sizeof(cl_mem ),&gm1_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,24,sizeof(cl_mem ),&qinf_d);
errorCode = errorCode | clSetKernelArg(kernelPointer,25,sizeof(cl_mem ),&eps_d);
assert_m(errorCode == CL_SUCCESS,"Error setting OpenCL kernel arguments");
errorCode = clEnqueueNDRangeKernel(cqCommandQueue,kernelPointer,1,NULL,&totalThreadNumber,&threadsPerBlock,0,NULL,&event);
assert_m(errorCode == CL_SUCCESS,"Error executing OpenCL kernel");
errorCode = clFinish(cqCommandQueue);
assert_m(errorCode == CL_SUCCESS,"Error completing device command queue");
blockOffset += blocksPerGrid;
}
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(0);
OP_kernels[1].name = userSubroutine;
OP_kernels[1].count = OP_kernels[1].count + 1;
}
//#define AUTO_BLOCK_SIZE
void op_par_loop_save_soln(char const *name, op_set set,
op_arg arg0,
op_arg arg1 ){
cl_int ciErrNum;
cl_event ceEvent;
if (OP_diags>2) {
printf(" kernel routine w/o indirection: save_soln \n");
}
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// set CUDA execution parameters
#ifdef AUTO_BLOCK_SIZE
const size_t nthread = 1024;
#else
#ifdef OP_BLOCK_SIZE_0
const size_t nthread = OP_BLOCK_SIZE_0;
#else
// int nthread = OP_block_size;
const size_t nthread = 128;
#endif
#endif
const size_t nblocks = 200;
const size_t n_tot_thread = nblocks * nthread;
// work out shared memory requirements per element
int nshared = 0;
nshared = MAX(nshared,sizeof(float)*4);
nshared = MAX(nshared,sizeof(float)*4);
// execute plan
int offset_s = nshared*OP_WARPSIZE;
nshared = nshared*nthread;
cl_kernel hKernel = getKernel( "op_cuda_save_soln" );
//nshared *= 4;
//offset_s *= 4;
int i = 0;
ciErrNum = clSetKernelArg( hKernel, i++, sizeof(cl_mem), &(arg0.data_d) );
ciErrNum |= clSetKernelArg( hKernel, i++, sizeof(cl_mem), &(arg1.data_d) );
ciErrNum |= clSetKernelArg( hKernel, i++, sizeof(int), &offset_s );
ciErrNum |= clSetKernelArg( hKernel, i++, sizeof(int), &set->size );
ciErrNum |= clSetKernelArg( hKernel, i++, nshared, NULL );
assert_m( ciErrNum == CL_SUCCESS, "error setting kernel arguments" );
#ifdef AUTO_BLOCK_SIZE
ciErrNum = clEnqueueNDRangeKernel( cqCommandQueue, hKernel, 1, NULL, &n_tot_thread, NULL, 0, NULL, &ceEvent );
#else
ciErrNum = clEnqueueNDRangeKernel( cqCommandQueue, hKernel, 1, NULL, &n_tot_thread, &nthread, 0, NULL, &ceEvent );
#endif
assert_m( ciErrNum == CL_SUCCESS, "error executing kernel" );
#ifndef ASYNC
ciErrNum = clFinish( cqCommandQueue );
assert_m( ciErrNum == CL_SUCCESS, "error completing device commands" );
#ifdef PROFILE
unsigned long tqueue, tsubmit, tstart, tend, telapsed;
ciErrNum = clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_QUEUED, sizeof(tqueue), &tqueue, NULL );
ciErrNum |= clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_SUBMIT, sizeof(tsubmit), &tsubmit, NULL );
ciErrNum |= clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_START, sizeof(tstart), &tstart, NULL );
ciErrNum |= clGetEventProfilingInfo( ceEvent, CL_PROFILING_COMMAND_END, sizeof(tend), &tend, NULL );
assert_m( ciErrNum == CL_SUCCESS, "error getting profiling info" );
OP_kernels[0].queue_time += (tsubmit - tqueue);
OP_kernels[0].wait_time += (tstart - tsubmit);
OP_kernels[0].execution_time += (tend - tstart);
//printf("%20lu\n%20lu\n%20lu\n%20lu\n\n", tqueue, tsubmit, tstart, tend);
//printf("queue: %8.4f\nwait:%8.4f\nexec: %8.4f\n\n", OP_kernels[0].queue_time * 1.0e-9, OP_kernels[0].wait_time * 1.0e-9, OP_kernels[0].execution_time * 1.0e-9 );
#endif
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(0);
OP_kernels[0].name = name;
OP_kernels[0].count += 1;
OP_kernels[0].time += wall_t2 - wall_t1;
OP_kernels[0].transfer += (float)set->size * arg0.size;
OP_kernels[0].transfer += (float)set->size * arg1.size;
#endif
}
void op_par_loop_update(char const *name, op_set set,
op_arg arg0,
op_arg arg1,
op_arg arg2,
op_arg arg3,
op_arg arg4 ){
float *arg3h = (float *)arg3.data;
float *arg4h = (float *)arg4.data;
if (OP_diags>2) {
printf(" kernel routine w/o indirection: update \n");
}
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// set number of threads
#ifdef _OPENMP
int nthreads = omp_get_max_threads( );
#else
int nthreads = 1;
#endif
// allocate and initialise arrays for global reduction
float arg3_l[1+64*64];
for (int thr=0; thr<nthreads; thr++)
for (int d=0; d<1; d++) arg3_l[d+thr*64]=ZERO_float;
float arg4_l[1+64*64];
for (int thr=0; thr<nthreads; thr++)
for (int d=0; d<1; d++) arg4_l[d+thr*64]=arg4h[d];
// execute plan
#pragma omp parallel for
for (int thr=0; thr<nthreads; thr++) {
int start = (set->size* thr )/nthreads;
int finish = (set->size*(thr+1))/nthreads;
op_x86_update( (float *) arg0.data,
(float *) arg1.data,
(float *) arg2.data,
arg3_l + thr*64,
arg4_l + thr*64,
start, finish );
}
// combine reduction data
for (int thr=0; thr<nthreads; thr++)
for(int d=0; d<1; d++) arg3h[d] += arg3_l[d+thr*64];
for (int thr=0; thr<nthreads; thr++)
for(int d=0; d<1; d++) arg4h[d] = MAX(arg4h[d],arg4_l[d+thr*64]);
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(1);
OP_kernels[1].name = name;
OP_kernels[1].count += 1;
OP_kernels[1].time += wall_t2 - wall_t1;
OP_kernels[1].transfer += (float)set->size * arg0.size;
OP_kernels[1].transfer += (float)set->size * arg1.size * 2.0f;
OP_kernels[1].transfer += (float)set->size * arg2.size * 2.0f;
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
int niter;
double rms;
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
gam = 1.4f;
gm1 = gam - 1.0f;
cfl = 0.9f;
eps = 0.05f;
double mach = 0.4f;
double alpha = 3.0f*atan(1.0f)/45.0f;
double p = 1.0f;
double r = 1.0f;
double u = sqrt(gam*p/r)*mach;
double e = p/(r*gm1) + 0.5f*u*u;
qinf[0] = r;
qinf[1] = r*u;
qinf[2] = 0.0f;
qinf[3] = r*e;
char file[] = "new_grid.h5";//"new_grid-26mil.h5";
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set_hdf5(file, "nodes");
op_set edges = op_decl_set_hdf5(file, "edges");
op_set bedges = op_decl_set_hdf5(file, "bedges");
op_set cells = op_decl_set_hdf5(file, "cells");
op_map pedge = op_decl_map_hdf5(edges, nodes, 2, file, "pedge");
op_map pecell = op_decl_map_hdf5(edges, cells,2, file, "pecell");
op_map pbedge = op_decl_map_hdf5(bedges,nodes,2, file, "pbedge");
op_map pbecell = op_decl_map_hdf5(bedges,cells,1, file, "pbecell");
op_map pcell = op_decl_map_hdf5(cells, nodes,4, file, "pcell");
op_dat p_bound = op_decl_dat_hdf5(bedges,1,"int" ,file,"p_bound");
op_dat p_x = op_decl_dat_hdf5(nodes ,2,"double",file,"p_x");
op_dat p_q = op_decl_dat_hdf5(cells ,4,"double",file,"p_q");
op_dat p_qold = op_decl_dat_hdf5(cells ,4,"double",file,"p_qold");
op_dat p_adt = op_decl_dat_hdf5(cells ,1,"double",file,"p_adt");
op_dat p_res = op_decl_dat_hdf5(cells ,4,"double",file,"p_res");
op_decl_const(1,"double",&gam );
op_decl_const(1,"double",&gm1 );
op_decl_const(1,"double",&cfl );
op_decl_const(1,"double",&eps );
op_decl_const(1,"double",&mach );
op_decl_const(1,"double",&alpha);
op_decl_const(4,"double",qinf );
op_diagnostic_output();
int g_ncell = op_get_size(cells);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
// save old flow solution
op_par_loop(save_soln,"save_soln", cells,
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop(adt_calc,"adt_calc",cells,
op_arg_dat(p_x, 0,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 1,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 2,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 3,pcell, 2,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_adt,-1,OP_ID, 1,"double",OP_WRITE));
// calculate flux residual
op_par_loop(res_calc,"res_calc",edges,
op_arg_dat(p_x, 0,pedge, 2,"double",OP_READ),
op_arg_dat(p_x, 1,pedge, 2,"double",OP_READ),
op_arg_dat(p_q, 0,pecell,4,"double",OP_READ),
op_arg_dat(p_q, 1,pecell,4,"double",OP_READ),
op_arg_dat(p_adt, 0,pecell,1,"double",OP_READ),
op_arg_dat(p_adt, 1,pecell,1,"double",OP_READ),
op_arg_dat(p_res, 0,pecell,4,"double",OP_INC ),
op_arg_dat(p_res, 1,pecell,4,"double",OP_INC ));
op_par_loop(bres_calc,"bres_calc",bedges,
op_arg_dat(p_x, 0,pbedge, 2,"double",OP_READ),
op_arg_dat(p_x, 1,pbedge, 2,"double",OP_READ),
op_arg_dat(p_q, 0,pbecell,4,"double",OP_READ),
op_arg_dat(p_adt, 0,pbecell,1,"double",OP_READ),
op_arg_dat(p_res, 0,pbecell,4,"double",OP_INC ),
op_arg_dat(p_bound,-1,OP_ID ,1,"int", OP_READ));
// update flow field
rms = 0.0;
op_par_loop(update,"update",cells,
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_WRITE),
op_arg_dat(p_res, -1,OP_ID, 4,"double",OP_RW ),
op_arg_dat(p_adt, -1,OP_ID, 1,"double",OP_READ ),
op_arg_gbl(&rms,1,"double",OP_INC));
}
// print iteration history
rms = sqrt(rms/(double)g_ncell);
if (iter%100 == 0)
op_printf(" %d %10.5e \n",iter,rms);
}
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_printf("Max total runtime = \n%f\n",wall_t2-wall_t1);
op_exit();
}
int main(int argc, char **argv) {
// OP initialisation
op_init(argc, argv, 2);
// MPI for user I/O
int my_rank;
int comm_size;
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
// timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
int *becell, *ecell, *bound, *bedge, *edge, *cell;
double *x, *q, *qold, *adt, *res;
int nnode, ncell, nedge, nbedge, niter;
/**------------------------BEGIN I/O and PARTITIONING -------------------**/
op_timers(&cpu_t1, &wall_t1);
/* read in grid from disk on root processor */
FILE *fp;
if ((fp = fopen("new_grid.dat", "r")) == NULL) {
op_printf("can't open file new_grid.dat\n");
exit(-1);
}
int g_nnode, g_ncell, g_nedge, g_nbedge;
check_scan(
fscanf(fp, "%d %d %d %d \n", &g_nnode, &g_ncell, &g_nedge, &g_nbedge), 4);
int *g_becell = 0, *g_ecell = 0, *g_bound = 0, *g_bedge = 0, *g_edge = 0,
*g_cell = 0;
double *g_x = 0, *g_q = 0, *g_qold = 0, *g_adt = 0, *g_res = 0;
op_printf("reading in grid \n");
op_printf("Global number of nodes, cells, edges, bedges = %d, %d, %d, %d\n",
g_nnode, g_ncell, g_nedge, g_nbedge);
if (my_rank == MPI_ROOT) {
g_cell = (int *)malloc(4 * g_ncell * sizeof(int));
g_edge = (int *)malloc(2 * g_nedge * sizeof(int));
g_ecell = (int *)malloc(2 * g_nedge * sizeof(int));
g_bedge = (int *)malloc(2 * g_nbedge * sizeof(int));
g_becell = (int *)malloc(g_nbedge * sizeof(int));
g_bound = (int *)malloc(g_nbedge * sizeof(int));
g_x = (double *)malloc(2 * g_nnode * sizeof(double));
g_q = (double *)malloc(4 * g_ncell * sizeof(double));
g_qold = (double *)malloc(4 * g_ncell * sizeof(double));
g_res = (double *)malloc(4 * g_ncell * sizeof(double));
g_adt = (double *)malloc(g_ncell * sizeof(double));
for (int n = 0; n < g_nnode; n++) {
check_scan(fscanf(fp, "%lf %lf \n", &g_x[2 * n], &g_x[2 * n + 1]), 2);
}
for (int n = 0; n < g_ncell; n++) {
check_scan(fscanf(fp, "%d %d %d %d \n", &g_cell[4 * n],
&g_cell[4 * n + 1], &g_cell[4 * n + 2],
&g_cell[4 * n + 3]),
4);
}
for (int n = 0; n < g_nedge; n++) {
check_scan(fscanf(fp, "%d %d %d %d \n", &g_edge[2 * n],
&g_edge[2 * n + 1], &g_ecell[2 * n],
&g_ecell[2 * n + 1]),
4);
}
for (int n = 0; n < g_nbedge; n++) {
check_scan(fscanf(fp, "%d %d %d %d \n", &g_bedge[2 * n],
&g_bedge[2 * n + 1], &g_becell[n], &g_bound[n]),
4);
}
// initialise flow field and residual
}
fclose(fp);
nnode = compute_local_size(g_nnode, comm_size, my_rank);
ncell = compute_local_size(g_ncell, comm_size, my_rank);
nedge = compute_local_size(g_nedge, comm_size, my_rank);
nbedge = compute_local_size(g_nbedge, comm_size, my_rank);
op_printf(
"Number of nodes, cells, edges, bedges on process %d = %d, %d, %d, %d\n",
my_rank, nnode, ncell, nedge, nbedge);
/*Allocate memory to hold local sets, mapping tables and data*/
cell = (int *)malloc(4 * ncell * sizeof(int));
edge = (int *)malloc(2 * nedge * sizeof(int));
ecell = (int *)malloc(2 * nedge * sizeof(int));
bedge = (int *)malloc(2 * nbedge * sizeof(int));
becell = (int *)malloc(nbedge * sizeof(int));
bound = (int *)malloc(nbedge * sizeof(int));
x = (double *)malloc(2 * nnode * sizeof(double));
q = (double *)malloc(4 * ncell * sizeof(double));
qold = (double *)malloc(4 * ncell * sizeof(double));
res = (double *)malloc(4 * ncell * sizeof(double));
adt = (double *)malloc(ncell * sizeof(double));
/* scatter sets, mappings and data on sets*/
scatter_int_array(g_cell, cell, comm_size, g_ncell, ncell, 4);
scatter_int_array(g_edge, edge, comm_size, g_nedge, nedge, 2);
scatter_int_array(g_ecell, ecell, comm_size, g_nedge, nedge, 2);
scatter_int_array(g_bedge, bedge, comm_size, g_nbedge, nbedge, 2);
scatter_int_array(g_becell, becell, comm_size, g_nbedge, nbedge, 1);
scatter_int_array(g_bound, bound, comm_size, g_nbedge, nbedge, 1);
scatter_double_array(g_x, x, comm_size, g_nnode, nnode, 2);
scatter_double_array(g_q, q, comm_size, g_ncell, ncell, 4);
scatter_double_array(g_qold, qold, comm_size, g_ncell, ncell, 4);
scatter_double_array(g_res, res, comm_size, g_ncell, ncell, 4);
scatter_double_array(g_adt, adt, comm_size, g_ncell, ncell, 1);
/*Freeing memory allocated to gloabal arrays on rank 0
after scattering to all processes*/
if (my_rank == MPI_ROOT) {
free(g_cell);
free(g_edge);
free(g_ecell);
free(g_bedge);
free(g_becell);
free(g_bound);
free(g_x);
free(g_q);
free(g_qold);
free(g_adt);
free(g_res);
}
op_timers(&cpu_t2, &wall_t2);
op_printf("Max total file read time = %f\n", wall_t2 - wall_t1);
/**------------------------END I/O and PARTITIONING -----------------------**/
op_set edges = op_decl_set(nedge, "edges");
op_set cells = op_decl_set(ncell, "cells");
op_map pecell = op_decl_map(edges, cells, 2, ecell, "pecell");
op_dat p_res = op_decl_dat(cells, 4, "double", res, "p_res");
int count;
// trigger partitioning and halo creation routines
op_partition("PTSCOTCH", "KWAY", cells, pecell, NULL);
op_diagnostic_output();
// initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// indirect reduction
count = 0;
op_par_loop_res_calc("res_calc", edges,
op_arg_dat(p_res, 0, pecell, 4, "double", OP_INC),
op_arg_gbl(&count, 1, "int", OP_INC));
op_printf("number of edges:: %d should be: %d \n", count, g_nedge);
if (count != g_nedge)
op_printf("indirect reduction FAILED\n");
else
op_printf("indirect reduction PASSED\n");
// direct reduction
count = 0;
op_par_loop_update("update", cells,
op_arg_dat(p_res, -1, OP_ID, 4, "double", OP_RW),
op_arg_gbl(&count, 1, "int", OP_INC));
op_printf("number of cells: %d should be: %d \n", count, g_ncell);
if (count != g_ncell)
op_printf("direct reduction FAILED\n");
else
op_printf("direct reduction PASSED\n");
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_exit();
free(cell);
free(edge);
free(ecell);
free(bedge);
free(becell);
free(bound);
free(x);
free(q);
free(qold);
free(res);
free(adt);
}
int main(int argc, char **argv) {
// OP initialisation
op_init(argc, argv, 2);
int niter;
float rms;
// timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
char file[] = "new_grid.h5";
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set_hdf5(file, "nodes");
op_set edges = op_decl_set_hdf5(file, "edges");
op_set bedges = op_decl_set_hdf5(file, "bedges");
op_set cells = op_decl_set_hdf5(file, "cells");
op_map pedge = op_decl_map_hdf5(edges, nodes, 2, file, "pedge");
op_map pecell = op_decl_map_hdf5(edges, cells, 2, file, "pecell");
op_map pbedge = op_decl_map_hdf5(bedges, nodes, 2, file, "pbedge");
op_map pbecell = op_decl_map_hdf5(bedges, cells, 1, file, "pbecell");
op_map pcell = op_decl_map_hdf5(cells, nodes, 4, file, "pcell");
op_dat p_bound = op_decl_dat_hdf5(bedges, 1, "int", file, "p_bound");
op_dat p_x = op_decl_dat_hdf5(nodes, 2, "float", file, "p_x");
op_dat p_q = op_decl_dat_hdf5(cells, 4, "float", file, "p_q");
op_dat p_qold = op_decl_dat_hdf5(cells, 4, "float", file, "p_qold");
op_dat p_adt = op_decl_dat_hdf5(cells, 1, "float", file, "p_adt");
op_dat p_res = op_decl_dat_hdf5(cells, 4, "float", file, "p_res");
op_get_const_hdf5("gam", 1, "float", (char *)&gam, "new_grid.h5");
op_get_const_hdf5("gm1", 1, "float", (char *)&gm1, "new_grid.h5");
op_get_const_hdf5("cfl", 1, "float", (char *)&cfl, "new_grid.h5");
op_get_const_hdf5("eps", 1, "float", (char *)&eps, "new_grid.h5");
op_get_const_hdf5("mach", 1, "float", (char *)&mach, "new_grid.h5");
op_get_const_hdf5("alpha", 1, "float", (char *)&alpha, "new_grid.h5");
op_get_const_hdf5("qinf", 4, "float", (char *)&qinf, "new_grid.h5");
op_decl_const2("gam", 1, "float", &gam);
op_decl_const2("gm1", 1, "float", &gm1);
op_decl_const2("cfl", 1, "float", &cfl);
op_decl_const2("eps", 1, "float", &eps);
op_decl_const2("mach", 1, "float", &mach);
op_decl_const2("alpha", 1, "float", &alpha);
op_decl_const2("qinf", 4, "float", qinf);
if (op_is_root())
op_diagnostic_output();
// trigger partitioning and halo creation routines
op_partition("PTSCOTCH", "KWAY", edges, pecell, p_x);
// op_partition("PARMETIS", "KWAY", edges, pecell, p_x);
int g_ncell = op_get_size(cells);
// initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 1000;
for (int iter = 1; iter <= niter; iter++) {
// save old flow solution
op_par_loop_save_soln("save_soln", cells,
op_arg_dat(p_q, -1, OP_ID, 4, "float", OP_READ),
op_arg_dat(p_qold, -1, OP_ID, 4, "float", OP_WRITE));
// predictor/corrector update loop
for (int k = 0; k < 2; k++) {
// calculate area/timstep
op_par_loop_adt_calc("adt_calc", cells,
op_arg_dat(p_x, 0, pcell, 2, "float", OP_READ),
op_arg_dat(p_x, 1, pcell, 2, "float", OP_READ),
op_arg_dat(p_x, 2, pcell, 2, "float", OP_READ),
op_arg_dat(p_x, 3, pcell, 2, "float", OP_READ),
op_arg_dat(p_q, -1, OP_ID, 4, "float", OP_READ),
op_arg_dat(p_adt, -1, OP_ID, 1, "float", OP_WRITE));
// calculate flux residual
op_par_loop_res_calc("res_calc", edges,
op_arg_dat(p_x, 0, pedge, 2, "float", OP_READ),
op_arg_dat(p_x, 1, pedge, 2, "float", OP_READ),
op_arg_dat(p_q, 0, pecell, 4, "float", OP_READ),
op_arg_dat(p_q, 1, pecell, 4, "float", OP_READ),
op_arg_dat(p_adt, 0, pecell, 1, "float", OP_READ),
op_arg_dat(p_adt, 1, pecell, 1, "float", OP_READ),
op_arg_dat(p_res, 0, pecell, 4, "float", OP_INC),
op_arg_dat(p_res, 1, pecell, 4, "float", OP_INC));
op_par_loop_bres_calc("bres_calc", bedges,
op_arg_dat(p_x, 0, pbedge, 2, "float", OP_READ),
op_arg_dat(p_x, 1, pbedge, 2, "float", OP_READ),
op_arg_dat(p_q, 0, pbecell, 4, "float", OP_READ),
op_arg_dat(p_adt, 0, pbecell, 1, "float", OP_READ),
op_arg_dat(p_res, 0, pbecell, 4, "float", OP_INC),
op_arg_dat(p_bound, -1, OP_ID, 1, "int", OP_READ));
// update flow field
rms = 0.0;
op_par_loop_update("update", cells,
op_arg_dat(p_qold, -1, OP_ID, 4, "float", OP_READ),
op_arg_dat(p_q, -1, OP_ID, 4, "float", OP_WRITE),
op_arg_dat(p_res, -1, OP_ID, 4, "float", OP_RW),
op_arg_dat(p_adt, -1, OP_ID, 1, "float", OP_READ),
op_arg_gbl(&rms, 1, "float", OP_INC));
}
// print iteration history
rms = sqrtf(rms / (float)g_ncell);
if (iter % 100 == 0)
op_printf(" %d %10.5e \n", iter, rms);
if (iter % 1000 == 0 &&
g_ncell == 720000) { // defailt mesh -- for validation testing
op_printf(" %d %3.16f \n", iter, rms);
float diff = fabsf((100.0 * (rms / 0.000105987)) - 100.0);
op_printf("\n\nTest problem with %d cells is within %3.15E %% of the "
"expected solution\n",
720000, diff);
if (diff < 0.1) {
op_printf("This test is considered PASSED\n");
} else {
op_printf("This test is considered FAILED\n");
}
}
}
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_printf("Max total runtime = %f\n", wall_t2 - wall_t1);
op_exit();
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
int *bnode, *cell;
double *xm;//, *q;
int nnode,ncell,nbnodes,niter;
double rms = 1;
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
double gam = 1.4;
gm1 = gam - 1.0;
gm1i = 1.0/gm1;
wtg1[0] = 0.5;
wtg1[1] = 0.5;
xi1[0] = 0.211324865405187;
xi1[1] = 0.788675134594813;
Ng1[0] = 0.788675134594813;
Ng1[1] = 0.211324865405187;
Ng1[2] = 0.211324865405187;
Ng1[3] = 0.788675134594813;
Ng1_xi[0] = -1;
Ng1_xi[1] = -1;
Ng1_xi[2] = 1;
Ng1_xi[3] = 1;
wtg2[0] = 0.25;
wtg2[1] = 0.25;
wtg2[2] = 0.25;
wtg2[3] = 0.25;
Ng2[0] = 0.622008467928146; Ng2[1] = 0.166666666666667; Ng2[2] = 0.166666666666667; Ng2[3] = 0.044658198738520;
Ng2[4] = 0.166666666666667; Ng2[5] = 0.622008467928146; Ng2[6] = 0.044658198738520; Ng2[7] = 0.166666666666667;
Ng2[8] = 0.166666666666667; Ng2[9] = 0.044658198738520; Ng2[10] = 0.622008467928146; Ng2[11] = 0.166666666666667;
Ng2[12] = 0.044658198738520; Ng2[13] = 0.166666666666667; Ng2[14] = 0.166666666666667; Ng2[15] = 0.622008467928146;
Ng2_xi[0] = -0.788675134594813; Ng2_xi[1] = 0.788675134594813; Ng2_xi[2] = -0.211324865405187;Ng2_xi[3] = 0.211324865405187;
Ng2_xi[4] = -0.788675134594813; Ng2_xi[5] = 0.788675134594813; Ng2_xi[6] = -0.211324865405187; Ng2_xi[7] = 0.211324865405187;
Ng2_xi[8] = -0.211324865405187; Ng2_xi[9] = 0.211324865405187; Ng2_xi[10] = -0.788675134594813; Ng2_xi[11] = 0.788675134594813;
Ng2_xi[12] = -0.211324865405187; Ng2_xi[13] = 0.211324865405187; Ng2_xi[14] = -0.788675134594813; Ng2_xi[15] = 0.788675134594813;
Ng2_xi[16] = -0.788675134594813; Ng2_xi[17] = -0.211324865405187; Ng2_xi[18] = 0.788675134594813; Ng2_xi[19] = 0.211324865405187;
Ng2_xi[20] = -0.211324865405187; Ng2_xi[21] = -0.788675134594813; Ng2_xi[22] = 0.211324865405187; Ng2_xi[23] = 0.788675134594813;
Ng2_xi[24] = -0.788675134594813; Ng2_xi[25] = -0.211324865405187; Ng2_xi[26] = 0.788675134594813; Ng2_xi[27] = 0.211324865405187;
Ng2_xi[28] = -0.211324865405187; Ng2_xi[29] = -0.788675134594813; Ng2_xi[30] = 0.211324865405187; Ng2_xi[31] = 0.788675134594813;
minf = 0.1;
m2 = minf*minf;
freq = 1;
kappa = 1;
nmode = 0;
mfan = 1.0;
char file[] = "FE_grid.h5";
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set_hdf5(file, "nodes");
op_set bnodes = op_decl_set_hdf5(file, "bedges");
op_set cells = op_decl_set_hdf5(file, "cells");
op_map pbnodes = op_decl_map_hdf5(bnodes,nodes,1,file, "pbedge");
op_map pcell = op_decl_map_hdf5(cells, nodes,4,file, "pcell");
op_dat p_xm = op_decl_dat_hdf5(nodes ,2,"double", file, "p_x");
op_dat p_phim = op_decl_dat_hdf5(nodes, 1, "double", file, "p_phim");
op_dat p_resm = op_decl_dat_hdf5(nodes, 1, "double", file, "p_resm");
op_dat p_K = op_decl_dat_hdf5(cells, 16, "double:soa",file, "p_K");
op_dat p_V = op_decl_dat_hdf5(nodes, 1, "double", file, "p_V");
op_dat p_P = op_decl_dat_hdf5(nodes, 1, "double", file, "p_P");
op_dat p_U = op_decl_dat_hdf5(nodes, 1, "double", file, "p_U");
op_decl_const2("gam",1,"double",&gam );
op_decl_const2("gm1",1,"double",&gm1 );
op_decl_const2("gm1i",1,"double",&gm1i );
op_decl_const2("m2",1,"double",&m2 );
op_decl_const2("wtg1",2,"double",wtg1 );
op_decl_const2("xi1",2,"double",xi1 );
op_decl_const2("Ng1",4,"double",Ng1 );
op_decl_const2("Ng1_xi",4,"double",Ng1_xi );
op_decl_const2("wtg2",4,"double",wtg2 );
op_decl_const2("Ng2",16,"double",Ng2 );
op_decl_const2("Ng2_xi",32,"double",Ng2_xi );
op_decl_const2("minf",1,"double",&minf );
op_decl_const2("freq",1,"double",&freq );
op_decl_const2("kappa",1,"double",&kappa );
op_decl_const2("nmode",1,"double",&nmode );
op_decl_const2("mfan",1,"double",&mfan );
op_diagnostic_output();
op_partition("PTSCOTCH", "KWAY", cells, pcell, p_xm);
op_printf("nodes: %d cells: %d bnodes: %d\n", nodes->size, cells->size, bnodes->size);
nnode = op_get_size(nodes);
ncell = op_get_size(cells);
nbnodes = op_get_size(bnodes);
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 20;
for(int iter=1; iter<=niter; iter++) {
op_par_loop_res_calc("res_calc",cells,
op_arg_dat(p_xm,-4,pcell,2,"double",OP_READ),
op_arg_dat(p_phim,-4,pcell,1,"double",OP_READ),
op_arg_dat(p_K,-1,OP_ID,16,"double:soa",OP_WRITE),
op_arg_dat(p_resm,-4,pcell,1,"double",OP_INC));
op_par_loop_dirichlet("dirichlet",bnodes,
op_arg_dat(p_resm,0,pbnodes,1,"double",OP_WRITE));
double c1 = 0;
double c2 = 0;
double c3 = 0;
double alpha = 0;
double beta = 0;
//c1 = R'*R;
op_par_loop_init_cg("init_cg",nodes,
op_arg_dat(p_resm,-1,OP_ID,1,"double",OP_READ),
op_arg_gbl(&c1,1,"double",OP_INC),
op_arg_dat(p_U,-1,OP_ID,1,"double",OP_WRITE),
op_arg_dat(p_V,-1,OP_ID,1,"double",OP_WRITE),
op_arg_dat(p_P,-1,OP_ID,1,"double",OP_WRITE));
//set up stopping conditions
double res0 = sqrt(c1);
double res = res0;
int iter = 0;
int maxiter = 200;
while (res > 0.1*res0 && iter < maxiter) {
//V = Stiffness*P
op_par_loop_spMV("spMV",cells,
op_arg_dat(p_V,-4,pcell,1,"double",OP_INC),
op_arg_dat(p_K,-1,OP_ID,16,"double:soa",OP_READ),
op_arg_dat(p_P,-4,pcell,1,"double",OP_READ));
op_par_loop_dirichlet("dirichlet",bnodes,
op_arg_dat(p_V,0,pbnodes,1,"double",OP_WRITE));
c2 = 0;
//c2 = P'*V;
op_par_loop_dotPV("dotPV",nodes,
op_arg_dat(p_P,-1,OP_ID,1,"double",OP_READ),
op_arg_dat(p_V,-1,OP_ID,1,"double",OP_READ),
op_arg_gbl(&c2,1,"double",OP_INC));
alpha = c1/c2;
//U = U + alpha*P;
//resm = resm-alpha*V;
op_par_loop_updateUR("updateUR",nodes,
op_arg_dat(p_U,-1,OP_ID,1,"double",OP_INC),
op_arg_dat(p_resm,-1,OP_ID,1,"double",OP_INC),
op_arg_dat(p_P,-1,OP_ID,1,"double",OP_READ),
op_arg_dat(p_V,-1,OP_ID,1,"double",OP_RW),
op_arg_gbl(&alpha,1,"double",OP_READ));
c3 = 0;
//c3 = resm'*resm;
op_par_loop_dotR("dotR",nodes,
op_arg_dat(p_resm,-1,OP_ID,1,"double",OP_READ),
op_arg_gbl(&c3,1,"double",OP_INC));
beta = c3/c1;
//P = beta*P+resm;
op_par_loop_updateP("updateP",nodes,
op_arg_dat(p_resm,-1,OP_ID,1,"double",OP_READ),
op_arg_dat(p_P,-1,OP_ID,1,"double",OP_RW),
op_arg_gbl(&beta,1,"double",OP_READ));
c1 = c3;
res = sqrt(c1);
iter++;
}
rms = 0;
//phim = phim - Stiffness\Load;
op_par_loop_update("update",nodes,
op_arg_dat(p_phim,-1,OP_ID,1,"double",OP_RW),
op_arg_dat(p_resm,-1,OP_ID,1,"double",OP_WRITE),
op_arg_dat(p_U,-1,OP_ID,1,"double",OP_READ),
op_arg_gbl(&rms,1,"double",OP_INC));
op_printf("rms = %10.5e iter: %d\n", sqrt(rms)/sqrt(nnode), iter);
}
op_timing_output();
op_timers(&cpu_t2, &wall_t2);
op_printf("Max total runtime = %f\n",wall_t2-wall_t1);
op_exit();
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
int niter;
double rms;
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
char file[] = "new_grid.h5";
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set_hdf5(file, "nodes");
op_set edges = op_decl_set_hdf5(file, "edges");
op_set bedges = op_decl_set_hdf5(file, "bedges");
op_set cells = op_decl_set_hdf5(file, "cells");
op_map pedge = op_decl_map_hdf5(edges, nodes, 2, file, "pedge");
op_map pecell = op_decl_map_hdf5(edges, cells,2, file, "pecell");
op_map pbedge = op_decl_map_hdf5(bedges,nodes,2, file, "pbedge");
op_map pbecell = op_decl_map_hdf5(bedges,cells,1, file, "pbecell");
op_map pcell = op_decl_map_hdf5(cells, nodes,4, file, "pcell");
op_dat p_bound = op_decl_dat_hdf5(bedges,1,"int" ,file,"p_bound");
op_dat p_x = op_decl_dat_hdf5(nodes ,2,"double",file,"p_x");
op_dat p_q = op_decl_dat_hdf5(cells ,4,"double",file,"p_q");
op_dat p_qold = op_decl_dat_hdf5(cells ,4,"double",file,"p_qold");
op_dat p_adt = op_decl_dat_hdf5(cells ,1,"double",file,"p_adt");
op_dat p_res = op_decl_dat_hdf5(cells ,4,"double",file,"p_res");
op_get_const_hdf5("gam", 1, "double", (char *)&gam, "new_grid.h5");
op_get_const_hdf5("gm1", 1, "double", (char *)&gm1, "new_grid.h5");
op_get_const_hdf5("cfl", 1, "double", (char *)&cfl, "new_grid.h5");
op_get_const_hdf5("eps", 1, "double", (char *)&eps, "new_grid.h5");
op_get_const_hdf5("mach", 1, "double", (char *)&mach, "new_grid.h5");
op_get_const_hdf5("alpha", 1, "double", (char *)&alpha, "new_grid.h5");
op_get_const_hdf5("qinf", 4, "double", (char *)&qinf, "new_grid.h5");
op_decl_const(1,"double",&gam );
op_decl_const(1,"double",&gm1 );
op_decl_const(1,"double",&cfl );
op_decl_const(1,"double",&eps );
op_decl_const(1,"double",&mach );
op_decl_const(1,"double",&alpha);
op_decl_const(4,"double",qinf );
op_diagnostic_output();
//write back original data just to compare you read the file correctly
//do an h5diff between new_grid_out.h5 and new_grid.h5 to
//compare two hdf5 files
op_write_hdf5("new_grid_out.h5");
op_write_const_hdf5("gam",1,"double",(char *)&gam, "new_grid_out.h5");
op_write_const_hdf5("gm1",1,"double",(char *)&gm1, "new_grid_out.h5");
op_write_const_hdf5("cfl",1,"double",(char *)&cfl, "new_grid_out.h5");
op_write_const_hdf5("eps",1,"double",(char *)&eps, "new_grid_out.h5");
op_write_const_hdf5("mach",1,"double",(char *)&mach, "new_grid_out.h5");
op_write_const_hdf5("alpha",1,"double",(char *)&alpha, "new_grid_out.h5");
op_write_const_hdf5("qinf",4,"double",(char *)qinf, "new_grid_out.h5");
//trigger partitioning and halo creation routines
op_partition("PTSCOTCH", "KWAY", edges, pecell, p_x);
int g_ncell = op_get_size(cells);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
// save old flow solution
op_par_loop(save_soln,"save_soln", cells,
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop(adt_calc,"adt_calc",cells,
op_arg_dat(p_x, -4,pcell, 2,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_adt,-1,OP_ID, 1,"double",OP_WRITE));
// calculate flux residual
op_par_loop(res_calc,"res_calc",edges,
op_arg_dat(p_x, -2,pedge, 2,"double",OP_READ),
op_arg_dat(p_q, -2,pecell,4,"double",OP_READ),
op_arg_dat(p_adt, -2,pecell,1,"double",OP_READ),
op_arg_dat(p_res, -2,pecell,4,"double",OP_INC ));
op_par_loop(bres_calc,"bres_calc",bedges,
op_arg_dat(p_x, -2,pbedge, 2,"double",OP_READ),
op_arg_dat(p_q, 0,pbecell,4,"double",OP_READ),
op_arg_dat(p_adt, 0,pbecell,1,"double",OP_READ),
op_arg_dat(p_res, 0,pbecell,4,"double",OP_INC ),
op_arg_dat(p_bound,-1,OP_ID ,1,"int", OP_READ));
// update flow field
rms = 0.0;
op_par_loop(update,"update",cells,
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_WRITE),
op_arg_dat(p_res, -1,OP_ID, 4,"double",OP_RW ),
op_arg_dat(p_adt, -1,OP_ID, 1,"double",OP_READ ),
op_arg_gbl(&rms,1,"double",OP_INC));
}
// print iteration history
rms = sqrt(rms/(double)g_ncell);
if (iter%100 == 0)
op_printf(" %d %10.5e \n",iter,rms);
}
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_printf("Max total runtime = \n%f\n",wall_t2-wall_t1);
op_exit();
}
//
// main program
//
int main(int argc, char **argv){
int my_rank;
int comm_size;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
double time;
double max_time;
int *becell, *ecell, *bound, *bedge, *edge, *cell;
double *x, *q, *qold, *adt, *res;
int niter;
double rms;
op_timers(&cpu_t1, &wall_t1);
// set constants
if(my_rank == MPI_ROOT )printf("initialising flow field\n");
gam = 1.4f;
gm1 = gam - 1.0f;
cfl = 0.9f;
eps = 0.05f;
double mach = 0.4f;
double alpha = 3.0f*atan(1.0f)/45.0f;
double p = 1.0f;
double r = 1.0f;
double u = sqrt(gam*p/r)*mach;
double e = p/(r*gm1) + 0.5f*u*u;
qinf[0] = r;
qinf[1] = r*u;
qinf[2] = 0.0f;
qinf[3] = r*e;
// OP initialisation
op_init(argc,argv,2);
/**------------------------BEGIN Parallel I/O -------------------**/
char file[] = "new_grid.h5";//"new_grid-26mil.h5";//"new_grid.h5";
// declare sets, pointers, datasets and global constants - reading in from file
op_set nodes = op_decl_set_hdf5(file, "nodes");
op_set edges = op_decl_set_hdf5(file, "edges");
op_set bedges = op_decl_set_hdf5(file, "bedges");
op_set cells = op_decl_set_hdf5(file, "cells");
op_map pedge = op_decl_map_hdf5(edges, nodes, 2, file, "pedge");
op_map pecell = op_decl_map_hdf5(edges, cells,2, file, "pecell");
op_map pbedge = op_decl_map_hdf5(bedges,nodes,2, file, "pbedge");
op_map pbecell = op_decl_map_hdf5(bedges,cells,1, file, "pbecell");
op_map pcell = op_decl_map_hdf5(cells, nodes,4, file, "pcell");
op_dat p_bound = op_decl_dat_hdf5(bedges,1,"int" ,file,"p_bound");
op_dat p_x = op_decl_dat_hdf5(nodes ,2,"double",file,"p_x");
op_dat p_q = op_decl_dat_hdf5(cells ,4,"double",file,"p_q");
op_dat p_qold = op_decl_dat_hdf5(cells ,4,"double",file,"p_qold");
op_dat p_adt = op_decl_dat_hdf5(cells ,1,"double",file,"p_adt");
op_dat p_res = op_decl_dat_hdf5(cells ,4,"double",file,"p_res");
/**------------------------END Parallel I/O -----------------------**/
op_timers(&cpu_t2, &wall_t2);
time = wall_t2-wall_t1;
MPI_Reduce(&time,&max_time,1,MPI_DOUBLE, MPI_MAX,MPI_ROOT, MPI_COMM_WORLD);
if(my_rank==MPI_ROOT)printf("Max total file read time = %f\n",max_time);
op_decl_const(1,"double",&gam );
op_decl_const(1,"double",&gm1 );
op_decl_const(1,"double",&cfl );
op_decl_const(1,"double",&eps );
op_decl_const(1,"double",&mach );
op_decl_const(1,"double",&alpha);
op_decl_const(4,"double",qinf );
op_diagnostic_output();
//write back original data just to compare you read the file correctly
//do an h5diff between new_grid_writeback.h5 and new_grid.h5 to
//compare two hdf5 files
op_write_hdf5("new_grid_out.h5");
//partition with ParMetis
//op_partition_geom(p_x);
//op_partition_random(cells);
//op_partition_kway(pecell);
//op_partition_geomkway(p_x, pcell);
//partition with PT-Scotch
op_partition_ptscotch(pecell);
//create halos
op_halo_create();
int g_ncell = 0;
int* sizes = (int *)malloc(sizeof(int)*comm_size);
MPI_Allgather(&cells->size, 1, MPI_INT, sizes, 1, MPI_INT, MPI_COMM_WORLD);
for(int i = 0; i<comm_size; i++)g_ncell = g_ncell + sizes[i];
free(sizes);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
//save old flow solution
op_par_loop(save_soln,"save_soln", cells,
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop(adt_calc,"adt_calc",cells,
op_arg_dat(p_x, 0,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 1,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 2,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 3,pcell, 2,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_adt,-1,OP_ID, 1,"double",OP_WRITE));
// calculate flux residual
op_par_loop(res_calc,"res_calc",edges,
op_arg_dat(p_x, 0,pedge, 2,"double",OP_READ),
op_arg_dat(p_x, 1,pedge, 2,"double",OP_READ),
op_arg_dat(p_q, 0,pecell,4,"double",OP_READ),
op_arg_dat(p_q, 1,pecell,4,"double",OP_READ),
op_arg_dat(p_adt, 0,pecell,1,"double",OP_READ),
op_arg_dat(p_adt, 1,pecell,1,"double",OP_READ),
op_arg_dat(p_res, 0,pecell,4,"double",OP_INC ),
op_arg_dat(p_res, 1,pecell,4,"double",OP_INC ));
op_par_loop(bres_calc,"bres_calc",bedges,
op_arg_dat(p_x, 0,pbedge, 2,"double",OP_READ),
op_arg_dat(p_x, 1,pbedge, 2,"double",OP_READ),
op_arg_dat(p_q, 0,pbecell,4,"double",OP_READ),
op_arg_dat(p_adt, 0,pbecell,1,"double",OP_READ),
op_arg_dat(p_res, 0,pbecell,4,"double",OP_INC ),
op_arg_dat(p_bound,-1,OP_ID ,1,"int", OP_READ));
// update flow field
rms = 0.0;
op_par_loop(update,"update",cells,
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_WRITE),
op_arg_dat(p_res, -1,OP_ID, 4,"double",OP_RW ),
op_arg_dat(p_adt, -1,OP_ID, 1,"double",OP_READ ),
op_arg_gbl(&rms,1,"double",OP_INC));
}
//print iteration history
if(my_rank==MPI_ROOT)
{
rms = sqrt(rms/(double) g_ncell);
if (iter%100 == 0)
printf("%d %10.5e \n",iter,rms);
}
}
op_timers(&cpu_t2, &wall_t2);
//get results data array
op_dat temp = op_mpi_get_data(p_q);
//output the result dat array to files
//op_write_hdf5("new_grid_out.h5");
//compress using
// ~/hdf5/bin/h5repack -f GZIP=9 new_grid.h5 new_grid_pack.h5
//free memory allocated to halos
op_halo_destroy();
//return all op_dats, op_maps back to original element order
op_partition_reverse();
//print each mpi process's timing info for each kernel
op_mpi_timing_output();
//print total time for niter interations
time = wall_t2-wall_t1;
MPI_Reduce(&time,&max_time,1,MPI_DOUBLE, MPI_MAX,MPI_ROOT, MPI_COMM_WORLD);
if(my_rank==MPI_ROOT)printf("Max total runtime = %f\n",max_time);
op_exit();
MPI_Finalize(); //user mpi finalize
}
void op_par_loop_res(char const *name, op_set set,
op_arg arg0,
op_arg arg1,
op_arg arg2,
op_arg arg3 ){
float *arg3h = (float *)arg3.data;
int nargs = 4;
op_arg args[4] = {arg0,arg1,arg2,arg3};
int ninds = 2;
int inds[4] = {-1,0,1,-1};
if (OP_diags>2) {
printf(" kernel routine with indirection: res \n");
}
// get plan
#ifdef OP_PART_SIZE_0
int part_size = OP_PART_SIZE_0;
#else
int part_size = OP_part_size;
#endif
op_plan *Plan = op_plan_get(name,set,part_size,nargs,args,ninds,inds);
// initialise timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
op_timers(&cpu_t1, &wall_t1);
// set number of threads
#ifdef _OPENMP
int nthreads = omp_get_max_threads( );
#else
int nthreads = 1;
#endif
// execute plan
int block_offset = 0;
for (int col=0; col < Plan->ncolors; col++) {
int nblocks = Plan->ncolblk[col];
#pragma omp parallel for
for (int blockIdx=0; blockIdx<nblocks; blockIdx++)
op_x86_res( blockIdx,
(float *)arg1.data, Plan->ind_maps[0],
(float *)arg2.data, Plan->ind_maps[1],
(float *)arg0.data,
Plan->loc_maps[1],
Plan->loc_maps[2],
(float *)arg3.data,
Plan->ind_sizes,
Plan->ind_offs,
block_offset,
Plan->blkmap,
Plan->offset,
Plan->nelems,
Plan->nthrcol,
Plan->thrcol);
block_offset += nblocks;
}
// update kernel record
op_timers(&cpu_t2, &wall_t2);
op_timing_realloc(0);
OP_kernels[0].name = name;
OP_kernels[0].count += 1;
OP_kernels[0].time += wall_t2 - wall_t1;
OP_kernels[0].transfer += Plan->transfer;
OP_kernels[0].transfer2 += Plan->transfer2;
}
int main(int argc, char **argv)
{
MPI_Init(&argc, &argv);
int rank, size;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
int *groups = (int *)malloc(size * sizeof(int));
int *groups2 = (int *)malloc(size * sizeof(int));
int my_type = 1; //This is to be read from a configuration file
MPI_Allgather(&my_type, 1, MPI_INT, groups, 1, MPI_INT, MPI_COMM_WORLD);
int num_groups = 0;
for (int i = 0; i < size; i++) num_groups = num_groups > groups[i] ? num_groups : groups[i];
num_groups++;
//The global group
MPI_Group global_grp;
MPI_Comm_group(MPI_COMM_WORLD, &global_grp);
//Create sub-groups and sub-communicators
MPI_Group mpigroups[num_groups];
MPI_Comm mpicomms[num_groups];
int count = 0;
for (int i = 0; i < num_groups; ++i) {
count = 0;
for (int j = 0; j < size; ++j) {
if (groups[j] == i) {
groups2[count++] = j;
}
}
MPI_Group_incl(global_grp, count, groups2, &mpigroups[i]);
MPI_Comm_create(MPI_COMM_WORLD, mpigroups[i], &mpicomms[i]);
}
//coupling procs
for (int i = 0; i < 1; ++i) {
count = 0;
for (int j = 0; j < size; ++j) {
if (groups[j] == i) {
groups2[count++] = j;
}
}
}
// OP initialisation
op_mpi_init(argc,argv,2,MPI_COMM_WORLD, mpicomms[1]);
int niter;
double rms;
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
char file[] = "new_grid.h5";
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set_hdf5(file, "nodes");
op_set edges = op_decl_set_hdf5(file, "edges");
op_set bedges = op_decl_set_hdf5(file, "bedges");
op_set cells = op_decl_set_hdf5(file, "cells");
op_map pedge = op_decl_map_hdf5(edges, nodes, 2, file, "pedge");
op_map pecell = op_decl_map_hdf5(edges, cells,2, file, "pecell");
op_map pbedge = op_decl_map_hdf5(bedges,nodes,2, file, "pbedge");
op_map pbecell = op_decl_map_hdf5(bedges,cells,1, file, "pbecell");
op_map pcell = op_decl_map_hdf5(cells, nodes,4, file, "pcell");
op_map pbndbnd = op_decl_map_hdf5(bedges, bedges,1, file, "pbndbnd");
op_map m_test = op_decl_map_hdf5(cells, nodes,4, file, "m_test");
if (m_test == NULL) printf("m_test not found\n");
op_dat p_bound = op_decl_dat_hdf5(bedges,1,"int" ,file,"p_bound");
op_dat p_x = op_decl_dat_hdf5(nodes ,2,"double",file,"p_x");
op_dat p_q = op_decl_dat_hdf5(cells ,4,"double",file,"p_q");
op_dat p_qold = op_decl_dat_hdf5(cells ,4,"double",file,"p_qold");
op_dat p_adt = op_decl_dat_hdf5(cells ,1,"double",file,"p_adt");
op_dat p_res = op_decl_dat_hdf5(cells ,4,"double",file,"p_res");
op_dat p_test = op_decl_dat_hdf5(cells ,4,"double",file,"p_test");
if (p_test == NULL) printf("p_test not found\n");
op_get_const_hdf5("gam", 1, "double", (char *)&gam, "new_grid.h5");
op_get_const_hdf5("gm1", 1, "double", (char *)&gm1, "new_grid.h5");
op_get_const_hdf5("cfl", 1, "double", (char *)&cfl, "new_grid.h5");
op_get_const_hdf5("eps", 1, "double", (char *)&eps, "new_grid.h5");
op_get_const_hdf5("mach", 1, "double", (char *)&mach, "new_grid.h5");
op_get_const_hdf5("alpha", 1, "double", (char *)&alpha, "new_grid.h5");
op_get_const_hdf5("qinf", 4, "double", (char *)&qinf, "new_grid.h5");
op_decl_const(1,"double",&gam );
op_decl_const(1,"double",&gm1 );
op_decl_const(1,"double",&cfl );
op_decl_const(1,"double",&eps );
op_decl_const(1,"double",&mach );
op_decl_const(1,"double",&alpha);
op_decl_const(4,"double",qinf );
op_diagnostic_output();
//write back original data just to compare you read the file correctly
//do an h5diff between new_grid_out.h5 and new_grid.h5 to
//compare two hdf5 files
op_dump_to_hdf5("new_grid_out.h5");
op_write_const_hdf5("gam",1,"double",(char *)&gam, "new_grid_out.h5");
op_write_const_hdf5("gm1",1,"double",(char *)&gm1, "new_grid_out.h5");
op_write_const_hdf5("cfl",1,"double",(char *)&cfl, "new_grid_out.h5");
op_write_const_hdf5("eps",1,"double",(char *)&eps, "new_grid_out.h5");
op_write_const_hdf5("mach",1,"double",(char *)&mach, "new_grid_out.h5");
op_write_const_hdf5("alpha",1,"double",(char *)&alpha, "new_grid_out.h5");
op_write_const_hdf5("qinf",4,"double",(char *)qinf, "new_grid_out.h5");
//trigger partitioning and halo creation routines
op_partition("PTSCOTCH", "KWAY", edges, pecell, p_x);
//op_partition("PARMETIS", "KWAY", edges, pecell, p_x);
int g_ncell = op_get_size(cells);
//create some temporaries so we can exchange data defined on the boundary
double *ptr = NULL;
op_dat center = op_decl_dat_temp(bedges, 3, "double", ptr, "center");
op_dat pres = op_decl_dat_temp(bedges, 1, "double", ptr, "pres");
int *ptr2 = NULL;
op_dat p_bound2 = op_decl_dat_temp(bedges, 1, "int", ptr2, "p_bound2");
op_dat center2 = op_decl_dat_temp(bedges, 3, "double", ptr, "center2");
op_dat pres2 = op_decl_dat_temp(bedges, 1, "double", ptr, "pres2");
//create import and export handles
op_export_handle handle = op_export_init(count, groups2, pbndbnd);
op_import_handle handle2 = op_import_init(count, groups2, center);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
// main time-marching loop
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
// save old flow solution
op_par_loop(save_soln,"save_soln", cells,
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop(adt_calc,"adt_calc",cells,
op_arg_dat(p_x, 0,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 1,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 2,pcell, 2,"double",OP_READ ),
op_arg_dat(p_x, 3,pcell, 2,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_adt,-1,OP_ID, 1,"double",OP_WRITE));
// calculate flux residual
op_par_loop(res_calc,"res_calc",edges,
op_arg_dat(p_x, 0,pedge, 2,"double",OP_READ),
op_arg_dat(p_x, 1,pedge, 2,"double",OP_READ),
op_arg_dat(p_q, 0,pecell,4,"double",OP_READ),
op_arg_dat(p_q, 1,pecell,4,"double",OP_READ),
op_arg_dat(p_adt, 0,pecell,1,"double",OP_READ),
op_arg_dat(p_adt, 1,pecell,1,"double",OP_READ),
op_arg_dat(p_res, 0,pecell,4,"double",OP_INC ),
op_arg_dat(p_res, 1,pecell,4,"double",OP_INC ));
op_par_loop(bres_calc,"bres_calc",bedges,
op_arg_dat(p_x, 0,pbedge, 2,"double",OP_READ),
op_arg_dat(p_x, 1,pbedge, 2,"double",OP_READ),
op_arg_dat(p_q, 0,pbecell,4,"double",OP_READ),
op_arg_dat(p_adt, 0,pbecell,1,"double",OP_READ),
op_arg_dat(p_res, 0,pbecell,4,"double",OP_INC ),
op_arg_dat(p_bound,-1,OP_ID ,1,"int", OP_READ),
op_arg_dat(center, -1, OP_ID, 3, "double", OP_WRITE),
op_arg_dat(pres, -1, OP_ID, 1, "double", OP_WRITE));
// update flow field
rms = 0.0;
op_par_loop(update,"update",cells,
op_arg_dat(p_qold,-1,OP_ID, 4,"double",OP_READ ),
op_arg_dat(p_q, -1,OP_ID, 4,"double",OP_WRITE),
op_arg_dat(p_res, -1,OP_ID, 4,"double",OP_RW ),
op_arg_dat(p_adt, -1,OP_ID, 1,"double",OP_READ ),
op_arg_gbl(&rms,1,"double",OP_INC));
}
// print iteration history
rms = sqrt(rms/(double)g_ncell);
if (iter%100 == 0) {
op_printf(" %d %10.5e \n",iter,rms);
//Export data
op_dat arr[] = {p_bound, center, pres};
op_export_data(handle, 3, arr);
//Import data
op_dat arr2[] = {p_bound2, center2, pres2};
op_import_data(handle2, 3, arr2);
//check whether the two are the same
op_par_loop(comparethem, "comparethem", bedges,
op_arg_dat(p_bound,-1, OP_ID, 1, "int", OP_READ),
op_arg_dat(p_bound2,-1, OP_ID, 1, "int", OP_READ),
op_arg_dat(center,-1, OP_ID, 3, "double", OP_READ),
op_arg_dat(center2,-1, OP_ID, 3, "double", OP_READ),
op_arg_dat(pres,-1, OP_ID, 1, "double", OP_READ),
op_arg_dat(pres2,-1, OP_ID, 1, "double", OP_READ));
}
}
op_timers(&cpu_t2, &wall_t2);
double* q = (double *)malloc(sizeof(double)*op_get_size(cells)*4);
op_fetch_data_hdf5(p_q, q, 0, op_get_size(cells)-1);
free(q);
op_fetch_data_hdf5_file(p_q, "file_name.h5");
//printf("Root process = %d\n",op_is_root());
//output the result dat array to files
//op_write_hdf5("new_grid_out.h5");
//compress using
// ~/hdf5/bin/h5repack -f GZIP=9 new_grid.h5 new_grid_pack.h5
op_timing_output();
op_printf("Max total runtime = \n%f\n",wall_t2-wall_t1);
op_exit();
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
//MPI for user I/O
int my_rank;
int comm_size;
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
int *bnode, *cell, *g_bnode, *g_cell;
double *xm, *g_xm;;
int nnode,ncell,nbnodes,niter, g_nnode, g_ncell, g_nbnodes;
double rms = 1;
// read in grid
op_printf("reading in grid \n");
FILE *fp;
if ( (fp = fopen("FE_grid.dat","r")) == NULL) {
op_printf("can't open file FE_grid.dat\n"); exit(-1);
}
if (fscanf(fp,"%d %d %d \n",&g_nnode, &g_ncell, &g_nbnodes) != 3) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
if (my_rank == MPI_ROOT) {
g_cell = (int *) malloc(4*g_ncell*sizeof(int));
g_bnode = (int *) malloc(g_nbnodes*sizeof(int));
g_xm = (double *) malloc(2*g_nnode*sizeof(double));
for (int n=0; n<g_nnode; n++) {
if (fscanf(fp,"%lf %lf \n",&g_xm[2*n], &g_xm[2*n+1]) != 2) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<g_ncell; n++) {
if (fscanf(fp,"%d %d %d %d \n",&g_cell[4*n ], &g_cell[4*n+1],
&g_cell[4*n+2], &g_cell[4*n+3]) != 4) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
for (int n=0; n<g_nbnodes; n++) {
if (fscanf(fp,"%d \n",&g_bnode[n]) != 1) {
op_printf("error reading from new_grid.dat\n"); exit(-1);
}
}
}
fclose(fp);
nnode = compute_local_size (g_nnode, comm_size, my_rank);
ncell = compute_local_size (g_ncell, comm_size, my_rank);
nbnodes = compute_local_size (g_nbnodes, comm_size, my_rank);
cell = (int *) malloc(4*ncell*sizeof(int));
bnode = (int *) malloc(nbnodes*sizeof(int));
xm = (double *) malloc(2*nnode*sizeof(double));
scatter_int_array(g_cell, cell, comm_size, g_ncell,ncell, 4);
scatter_int_array(g_bnode, bnode, comm_size, g_nbnodes,nbnodes, 1);
scatter_double_array(g_xm, xm, comm_size, g_nnode,nnode, 2);
if(my_rank == MPI_ROOT) {
free(g_cell);
free(g_xm);
free(g_bnode);
}
// set constants and initialise flow field and residual
op_printf("initialising flow field \n");
double gam = 1.4;
gm1 = gam - 1.0;
gm1i = 1.0/gm1;
wtg1[0] = 0.5;
wtg1[1] = 0.5;
xi1[0] = 0.211324865405187;
xi1[1] = 0.788675134594813;
Ng1[0] = 0.788675134594813;
Ng1[1] = 0.211324865405187;
Ng1[2] = 0.211324865405187;
Ng1[3] = 0.788675134594813;
Ng1_xi[0] = -1;
Ng1_xi[1] = -1;
Ng1_xi[2] = 1;
Ng1_xi[3] = 1;
wtg2[0] = 0.25;
wtg2[1] = 0.25;
wtg2[2] = 0.25;
wtg2[3] = 0.25;
Ng2[0] = 0.622008467928146; Ng2[1] = 0.166666666666667; Ng2[2] = 0.166666666666667; Ng2[3] = 0.044658198738520;
Ng2[4] = 0.166666666666667; Ng2[5] = 0.622008467928146; Ng2[6] = 0.044658198738520; Ng2[7] = 0.166666666666667;
Ng2[8] = 0.166666666666667; Ng2[9] = 0.044658198738520; Ng2[10] = 0.622008467928146; Ng2[11] = 0.166666666666667;
Ng2[12] = 0.044658198738520; Ng2[13] = 0.166666666666667; Ng2[14] = 0.166666666666667; Ng2[15] = 0.622008467928146;
Ng2_xi[0] = -0.788675134594813; Ng2_xi[1] = 0.788675134594813; Ng2_xi[2] = -0.211324865405187;Ng2_xi[3] = 0.211324865405187;
Ng2_xi[4] = -0.788675134594813; Ng2_xi[5] = 0.788675134594813; Ng2_xi[6] = -0.211324865405187; Ng2_xi[7] = 0.211324865405187;
Ng2_xi[8] = -0.211324865405187; Ng2_xi[9] = 0.211324865405187; Ng2_xi[10] = -0.788675134594813; Ng2_xi[11] = 0.788675134594813;
Ng2_xi[12] = -0.211324865405187; Ng2_xi[13] = 0.211324865405187; Ng2_xi[14] = -0.788675134594813; Ng2_xi[15] = 0.788675134594813;
Ng2_xi[16] = -0.788675134594813; Ng2_xi[17] = -0.211324865405187; Ng2_xi[18] = 0.788675134594813; Ng2_xi[19] = 0.211324865405187;
Ng2_xi[20] = -0.211324865405187; Ng2_xi[21] = -0.788675134594813; Ng2_xi[22] = 0.211324865405187; Ng2_xi[23] = 0.788675134594813;
Ng2_xi[24] = -0.788675134594813; Ng2_xi[25] = -0.211324865405187; Ng2_xi[26] = 0.788675134594813; Ng2_xi[27] = 0.211324865405187;
Ng2_xi[28] = -0.211324865405187; Ng2_xi[29] = -0.788675134594813; Ng2_xi[30] = 0.211324865405187; Ng2_xi[31] = 0.788675134594813;
minf = 0.1;
m2 = minf*minf;
freq = 1;
kappa = 1;
nmode = 0;
mfan = 1.0;
double *phim = (double *)malloc(nnode*sizeof(double));
memset(phim,0,nnode*sizeof(double));
for (int i = 0;i<nnode;i++) {
phim[i] = minf*xm[2*i];
}
double *K = (double *)malloc(4*4*ncell*sizeof(double));
memset(K,0,4*4*ncell*sizeof(double));
double *resm = (double *)malloc(nnode*sizeof(double));
memset(resm,0,nnode*sizeof(double));
double *V = (double *)malloc(nnode*sizeof(double));
memset(V,0,nnode*sizeof(double));
double *P = (double *)malloc(nnode*sizeof(double));
memset(P,0,nnode*sizeof(double));
double *U = (double *)malloc(nnode*sizeof(double));
memset(U,0,nnode*sizeof(double));
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set(nnode, "nodes");
op_set bnodes = op_decl_set(nbnodes, "bedges");
op_set cells = op_decl_set(ncell, "cells");
op_map pbnodes = op_decl_map(bnodes,nodes,1,bnode, "pbedge");
op_map pcell = op_decl_map(cells, nodes,4,cell, "pcell");
op_dat p_xm = op_decl_dat(nodes ,2,"double",xm ,"p_x");
op_dat p_phim = op_decl_dat(nodes, 1, "double", phim, "p_phim");
op_dat p_resm = op_decl_dat(nodes, 1, "double", resm, "p_resm");
op_dat p_K = op_decl_dat(cells, 16, "double:soa", K, "p_K");
op_dat p_V = op_decl_dat(nodes, 1, "double", V, "p_V");
op_dat p_P = op_decl_dat(nodes, 1, "double", P, "p_P");
op_dat p_U = op_decl_dat(nodes, 1, "double", U, "p_U");
op_decl_const(1,"double",&gam );
op_decl_const(1,"double",&gm1 );
op_decl_const(1,"double",&gm1i );
op_decl_const(1,"double",&m2 );
op_decl_const(2,"double",wtg1 );
op_decl_const(2,"double",xi1 );
op_decl_const(4,"double",Ng1 );
op_decl_const(4,"double",Ng1_xi );
op_decl_const(4,"double",wtg2 );
op_decl_const(16,"double",Ng2 );
op_decl_const(32,"double",Ng2_xi );
op_decl_const(1,"double",&minf );
op_decl_const(1,"double",&freq );
op_decl_const(1,"double",&kappa );
op_decl_const(1,"double",&nmode );
op_decl_const(1,"double",&mfan );
op_diagnostic_output();
op_partition("PTSCOTCH", "KWAY", cells, pcell, NULL);
// main time-marching loop
niter = 20;
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
for(int iter=1; iter<=niter; iter++) {
op_par_loop(res_calc,"res_calc",cells,
op_arg_dat(p_xm, -4, pcell, 2,"double",OP_READ),
op_arg_dat(p_phim, -4, pcell, 1,"double",OP_READ),
op_arg_dat(p_K, -1, OP_ID, 16,"double:soa",OP_WRITE),
op_arg_dat(p_resm, -4, pcell, 1,"double",OP_INC)
);
op_par_loop(dirichlet,"dirichlet",bnodes,
op_arg_dat(p_resm, 0, pbnodes, 1,"double",OP_WRITE));
double c1 = 0;
double c2 = 0;
double c3 = 0;
double alpha = 0;
double beta = 0;
//c1 = R'*R;
op_par_loop(init_cg, "init_cg", nodes,
op_arg_dat(p_resm, -1, OP_ID, 1, "double", OP_READ),
op_arg_gbl(&c1, 1, "double", OP_INC),
op_arg_dat(p_U, -1, OP_ID, 1, "double", OP_WRITE),
op_arg_dat(p_V, -1, OP_ID, 1, "double", OP_WRITE),
op_arg_dat(p_P, -1, OP_ID, 1, "double", OP_WRITE));
//set up stopping conditions
double res0 = sqrt(c1);
double res = res0;
int inner_iter = 0;
int maxiter = 200;
while (res > 0.1*res0 && inner_iter < maxiter) {
//V = Stiffness*P
op_par_loop(spMV, "spMV", cells,
op_arg_dat(p_V, -4, pcell, 1, "double", OP_INC),
op_arg_dat(p_K, -1, OP_ID, 16, "double:soa", OP_READ),
op_arg_dat(p_P, -4, pcell, 1, "double", OP_READ));
op_par_loop(dirichlet,"dirichlet",bnodes,
op_arg_dat(p_V, 0, pbnodes, 1,"double",OP_WRITE));
c2 = 0;
//c2 = P'*V;
op_par_loop(dotPV, "dotPV", nodes,
op_arg_dat(p_P, -1, OP_ID, 1, "double", OP_READ),
op_arg_dat(p_V, -1, OP_ID, 1, "double", OP_READ),
op_arg_gbl(&c2, 1, "double", OP_INC));
alpha = c1/c2;
//U = U + alpha*P;
//resm = resm-alpha*V;
op_par_loop(updateUR, "updateUR", nodes,
op_arg_dat(p_U, -1, OP_ID, 1, "double", OP_INC),
op_arg_dat(p_resm, -1, OP_ID, 1, "double", OP_INC),
op_arg_dat(p_P, -1, OP_ID, 1, "double", OP_READ),
op_arg_dat(p_V, -1, OP_ID, 1, "double", OP_RW),
op_arg_gbl(&alpha, 1, "double", OP_READ));
c3 = 0;
//c3 = resm'*resm;
op_par_loop(dotR, "dotR", nodes,
op_arg_dat(p_resm, -1, OP_ID, 1, "double", OP_READ),
op_arg_gbl(&c3, 1, "double", OP_INC));
beta = c3/c1;
//P = beta*P+resm;
op_par_loop(updateP, "updateP", nodes,
op_arg_dat(p_resm, -1, OP_ID, 1, "double", OP_READ),
op_arg_dat(p_P, -1, OP_ID, 1, "double", OP_RW),
op_arg_gbl(&beta, 1, "double", OP_READ));
c1 = c3;
res = sqrt(c1);
inner_iter++;
}
rms = 0;
//phim = phim - Stiffness\Load;
op_par_loop(update, "update", nodes,
op_arg_dat(p_phim, -1, OP_ID, 1, "double", OP_RW),
op_arg_dat(p_resm, -1, OP_ID, 1, "double", OP_WRITE),
op_arg_dat(p_U, -1, OP_ID, 1, "double", OP_READ),
op_arg_gbl(&rms, 1, "double", OP_INC));
op_printf("rms = %10.5e iter: %d\n", sqrt(rms)/sqrt(g_nnode), inner_iter);
}
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
op_printf("Max total runtime = %f\n",wall_t2-wall_t1);
op_exit();
/*free(cell);
free(bnode);
free(xm);
free(phim);
free(K);
free(resm);
free(V);
free(P);
free(U);*/
}
int main(int argc, char **argv)
{
// OP initialisation
op_init(argc,argv,2);
//MPI for user I/O
int my_rank;
int comm_size;
MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
//timer
double cpu_t1, cpu_t2, wall_t1, wall_t2;
int *becell, *ecell, *bound, *bedge, *edge, *cell;
double *x, *q, *qold, *adt, *res;
int nnode,ncell,nedge,nbedge,niter;
double rms;
/**------------------------BEGIN I/O and PARTITIONING -------------------**/
op_timers(&cpu_t1, &wall_t1);
/* read in grid from disk on root processor */
FILE *fp;
if ( (fp = fopen("new_grid.dat","r")) == NULL) {
op_printf("can't open file new_grid.dat\n"); exit(-1);
}
int g_nnode,g_ncell,g_nedge,g_nbedge;
check_scan(fscanf(fp,"%d %d %d %d \n",&g_nnode, &g_ncell, &g_nedge, &g_nbedge), 4);
int *g_becell = 0, *g_ecell = 0, *g_bound = 0, *g_bedge = 0, *g_edge = 0, *g_cell = 0;
double *g_x = 0,*g_q = 0, *g_qold = 0, *g_adt = 0, *g_res = 0;
// set constants
op_printf("initialising flow field\n");
gam = 1.4f;
gm1 = gam - 1.0f;
cfl = 0.9f;
eps = 0.05f;
double mach = 0.4f;
double alpha = 3.0f*atan(1.0f)/45.0f;
double p = 1.0f;
double r = 1.0f;
double u = sqrt(gam*p/r)*mach;
double e = p/(r*gm1) + 0.5f*u*u;
qinf[0] = r;
qinf[1] = r*u;
qinf[2] = 0.0f;
qinf[3] = r*e;
op_printf("reading in grid \n");
op_printf("Global number of nodes, cells, edges, bedges = %d, %d, %d, %d\n"
,g_nnode,g_ncell,g_nedge,g_nbedge);
if(my_rank == MPI_ROOT) {
g_cell = (int *) malloc(4*g_ncell*sizeof(int));
g_edge = (int *) malloc(2*g_nedge*sizeof(int));
g_ecell = (int *) malloc(2*g_nedge*sizeof(int));
g_bedge = (int *) malloc(2*g_nbedge*sizeof(int));
g_becell = (int *) malloc( g_nbedge*sizeof(int));
g_bound = (int *) malloc( g_nbedge*sizeof(int));
g_x = (double *) malloc(2*g_nnode*sizeof(double));
g_q = (double *) malloc(4*g_ncell*sizeof(double));
g_qold = (double *) malloc(4*g_ncell*sizeof(double));
g_res = (double *) malloc(4*g_ncell*sizeof(double));
g_adt = (double *) malloc( g_ncell*sizeof(double));
for (int n=0; n<g_nnode; n++){
check_scan(fscanf(fp,"%lf %lf \n",&g_x[2*n], &g_x[2*n+1]), 2);
}
for (int n=0; n<g_ncell; n++) {
check_scan(fscanf(fp,"%d %d %d %d \n",&g_cell[4*n ], &g_cell[4*n+1],
&g_cell[4*n+2], &g_cell[4*n+3]), 4);
}
for (int n=0; n<g_nedge; n++) {
check_scan(fscanf(fp,"%d %d %d %d \n",&g_edge[2*n],&g_edge[2*n+1],
&g_ecell[2*n],&g_ecell[2*n+1]), 4);
}
for (int n=0; n<g_nbedge; n++) {
check_scan(fscanf(fp,"%d %d %d %d \n",&g_bedge[2*n],&g_bedge[2*n+1],
&g_becell[n],&g_bound[n]), 4);
}
//initialise flow field and residual
for (int n=0; n<g_ncell; n++) {
for (int m=0; m<4; m++) {
g_q[4*n+m] = qinf[m];
g_res[4*n+m] = 0.0f;
}
}
}
fclose(fp);
nnode = compute_local_size (g_nnode, comm_size, my_rank);
ncell = compute_local_size (g_ncell, comm_size, my_rank);
nedge = compute_local_size (g_nedge, comm_size, my_rank);
nbedge = compute_local_size (g_nbedge, comm_size, my_rank);
op_printf("Number of nodes, cells, edges, bedges on process %d = %d, %d, %d, %d\n"
,my_rank,nnode,ncell,nedge,nbedge);
/*Allocate memory to hold local sets, mapping tables and data*/
cell = (int *) malloc(4*ncell*sizeof(int));
edge = (int *) malloc(2*nedge*sizeof(int));
ecell = (int *) malloc(2*nedge*sizeof(int));
bedge = (int *) malloc(2*nbedge*sizeof(int));
becell = (int *) malloc( nbedge*sizeof(int));
bound = (int *) malloc( nbedge*sizeof(int));
x = (double *) malloc(2*nnode*sizeof(double));
q = (double *) malloc(4*ncell*sizeof(double));
qold = (double *) malloc(4*ncell*sizeof(double));
res = (double *) malloc(4*ncell*sizeof(double));
adt = (double *) malloc( ncell*sizeof(double));
/* scatter sets, mappings and data on sets*/
scatter_int_array(g_cell, cell, comm_size, g_ncell,ncell, 4);
scatter_int_array(g_edge, edge, comm_size, g_nedge,nedge, 2);
scatter_int_array(g_ecell, ecell, comm_size, g_nedge,nedge, 2);
scatter_int_array(g_bedge, bedge, comm_size, g_nbedge,nbedge, 2);
scatter_int_array(g_becell, becell, comm_size, g_nbedge,nbedge, 1);
scatter_int_array(g_bound, bound, comm_size, g_nbedge,nbedge, 1);
scatter_double_array(g_x, x, comm_size, g_nnode,nnode, 2);
scatter_double_array(g_q, q, comm_size, g_ncell,ncell, 4);
scatter_double_array(g_qold, qold, comm_size, g_ncell,ncell, 4);
scatter_double_array(g_res, res, comm_size, g_ncell,ncell, 4);
scatter_double_array(g_adt, adt, comm_size, g_ncell,ncell, 1);
/*Freeing memory allocated to gloabal arrays on rank 0
after scattering to all processes*/
if(my_rank == MPI_ROOT) {
free(g_cell);
free(g_edge);
free(g_ecell);
free(g_bedge);
free(g_becell);
free(g_bound);
free(g_x );
free(g_q);
free(g_qold);
free(g_adt);
free(g_res);
}
op_timers(&cpu_t2, &wall_t2);
op_printf("Max total file read time = %f\n", wall_t2-wall_t1);
/**------------------------END I/O and PARTITIONING -----------------------**/
// declare sets, pointers, datasets and global constants
op_set nodes = op_decl_set(nnode, "nodes");
op_set edges = op_decl_set(nedge, "edges");
op_set bedges = op_decl_set(nbedge, "bedges");
op_set cells = op_decl_set(ncell, "cells");
op_map pedge = op_decl_map(edges, nodes,2,edge, "pedge");
op_map pecell = op_decl_map(edges, cells,2,ecell, "pecell");
op_map pbedge = op_decl_map(bedges,nodes,2,bedge, "pbedge");
op_map pbecell = op_decl_map(bedges,cells,1,becell,"pbecell");
op_map pcell = op_decl_map(cells, nodes,4,cell, "pcell");
op_dat p_bound = op_decl_dat(bedges,1,"int" ,bound,"p_bound");
op_dat p_x = op_decl_dat(nodes ,2,"double",x ,"p_x");
op_dat p_q = op_decl_dat(cells ,4,"double",q ,"p_q");
//op_dat p_qold = op_decl_dat(cells ,4,"double",qold ,"p_qold");
//op_dat p_adt = op_decl_dat(cells ,1,"double",adt ,"p_adt");
//op_dat p_res = op_decl_dat(cells ,4,"double",res ,"p_res");
// p_res, p_adt and p_qold now declared as a temp op_dats during
// the execution of the time-marching loop
op_decl_const2("gam",1,"double",&gam );
op_decl_const2("gm1",1,"double",&gm1 );
op_decl_const2("cfl",1,"double",&cfl );
op_decl_const2("eps",1,"double",&eps );
op_decl_const2("mach",1,"double",&mach );
op_decl_const2("alpha",1,"double",&alpha);
op_decl_const2("qinf",4,"double",qinf );
op_diagnostic_output();
//trigger partitioning and halo creation routines
op_partition("PTSCOTCH", "KWAY", cells, pecell, p_x);
//initialise timers for total execution wall time
op_timers(&cpu_t1, &wall_t1);
niter = 1000;
for(int iter=1; iter<=niter; iter++) {
double* tmp_elem = NULL;
op_dat p_res = op_decl_dat_temp(cells ,4,"double",tmp_elem,"p_res");
op_dat p_adt = op_decl_dat_temp(cells ,1,"double",tmp_elem,"p_adt");
op_dat p_qold = op_decl_dat_temp(cells ,4,"double",qold ,"p_qold");
//save old flow solution
op_par_loop_save_soln("save_soln",cells,
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_qold,-1,OP_ID,4,"double",OP_WRITE));
// predictor/corrector update loop
for(int k=0; k<2; k++) {
// calculate area/timstep
op_par_loop_adt_calc("adt_calc",cells,
op_arg_dat(p_x,0,pcell,2,"double",OP_READ),
op_arg_dat(p_x,1,pcell,2,"double",OP_READ),
op_arg_dat(p_x,2,pcell,2,"double",OP_READ),
op_arg_dat(p_x,3,pcell,2,"double",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_adt,-1,OP_ID,1,"double",OP_WRITE));
// calculate flux residual
op_par_loop_res_calc("res_calc",edges,
op_arg_dat(p_x,0,pedge,2,"double",OP_READ),
op_arg_dat(p_x,1,pedge,2,"double",OP_READ),
op_arg_dat(p_q,0,pecell,4,"double",OP_READ),
op_arg_dat(p_q,1,pecell,4,"double",OP_READ),
op_arg_dat(p_adt,0,pecell,1,"double",OP_READ),
op_arg_dat(p_adt,1,pecell,1,"double",OP_READ),
op_arg_dat(p_res,0,pecell,4,"double",OP_INC),
op_arg_dat(p_res,1,pecell,4,"double",OP_INC));
op_par_loop_bres_calc("bres_calc",bedges,
op_arg_dat(p_x,0,pbedge,2,"double",OP_READ),
op_arg_dat(p_x,1,pbedge,2,"double",OP_READ),
op_arg_dat(p_q,0,pbecell,4,"double",OP_READ),
op_arg_dat(p_adt,0,pbecell,1,"double",OP_READ),
op_arg_dat(p_res,0,pbecell,4,"double",OP_INC),
op_arg_dat(p_bound,-1,OP_ID,1,"int",OP_READ));
// update flow field
rms = 0.0;
op_par_loop_update("update",cells,
op_arg_dat(p_qold,-1,OP_ID,4,"double",OP_READ),
op_arg_dat(p_q,-1,OP_ID,4,"double",OP_WRITE),
op_arg_dat(p_res,-1,OP_ID,4,"double",OP_RW),
op_arg_dat(p_adt,-1,OP_ID,1,"double",OP_READ),
op_arg_gbl(&rms,1,"double",OP_INC));
}
//print iteration history
rms = sqrt(rms/(double) g_ncell);
if (iter%100 == 0)
op_printf("%d %10.5e \n",iter,rms);
if (op_free_dat_temp(p_res) < 0)
op_printf("Error: temporary op_dat %s cannot be removed\n",p_res->name);
if (op_free_dat_temp(p_adt) < 0)
op_printf("Error: temporary op_dat %s cannot be removed\n",p_adt->name);
if (op_free_dat_temp(p_qold) < 0)
op_printf("Error: temporary op_dat %s cannot be removed\n",p_qold->name);
}
op_timers(&cpu_t2, &wall_t2);
op_timing_output();
//print total time for niter interations
op_printf("Max total runtime = %f\n",wall_t2-wall_t1);
op_exit();
free(cell);
free(edge);
free(ecell);
free(bedge);
free(becell);
free(bound);
free(x);
free(q);
free(qold);
free(res);
free(adt);
}
void op_write_hdf5(char const * file_name)
{
printf("Writing to %s\n",file_name);
//declare timers
double cpu_t1, cpu_t2, wall_t1, wall_t2;
double time;
double max_time;
op_timers(&cpu_t1, &wall_t1); //timer start for hdf5 file write
//create new communicator
int my_rank, comm_size;
MPI_Comm_dup(MPI_COMM_WORLD, &OP_MPI_HDF5_WORLD);
MPI_Comm_rank(OP_MPI_HDF5_WORLD, &my_rank);
MPI_Comm_size(OP_MPI_HDF5_WORLD, &comm_size);
//MPI variables
MPI_Info info = MPI_INFO_NULL;
//HDF5 APIs definitions
hid_t file_id; //file identifier
hid_t plist_id; //property list identifier
hid_t dset_id = 0; //dataset identifier
hid_t dataspace; //data space identifier
hid_t memspace; //memory space identifier
hsize_t dimsf[2]; // dataset dimensions
hsize_t count[2]; //hyperslab selection parameters
hsize_t offset[2];
//Set up file access property list with parallel I/O access
plist_id = H5Pcreate(H5P_FILE_ACCESS);
H5Pset_fapl_mpio(plist_id, OP_MPI_HDF5_WORLD, info);
//Create a new file collectively and release property list identifier.
file_id = H5Fcreate(file_name, H5F_ACC_TRUNC, H5P_DEFAULT, plist_id);
H5Pclose(plist_id);
/*loop over all the op_sets and write them to file*/
for(int s=0; s<OP_set_index; s++) {
op_set set=OP_set_list[s];
//Create the dataspace for the dataset.
hsize_t dimsf_set[] = {1};
dataspace = H5Screate_simple(1, dimsf_set, NULL);
//Create the dataset with default properties and close dataspace.
dset_id = H5Dcreate(file_id, set->name, H5T_NATIVE_INT, dataspace,
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
//Create property list for collective dataset write.
plist_id = H5Pcreate(H5P_DATASET_XFER);
H5Pset_dxpl_mpio(plist_id, H5FD_MPIO_COLLECTIVE);
int size = 0;
int* sizes = (int *)xmalloc(sizeof(int)*comm_size);
MPI_Allgather(&set->size, 1, MPI_INT, sizes, 1, MPI_INT, OP_MPI_HDF5_WORLD);
for(int i = 0; i<comm_size; i++)size = size + sizes[i];
//write data
H5Dwrite(dset_id, H5T_NATIVE_INT, H5S_ALL, H5S_ALL, plist_id, &size);
H5Sclose(dataspace);
H5Pclose(plist_id);
H5Dclose(dset_id);
}
/*loop over all the op_maps and write them to file*/
for(int m=0; m<OP_map_index; m++) {
op_map map=OP_map_list[m];
//find total size of map
int* sizes = (int *)xmalloc(sizeof(int)*comm_size);
int g_size = 0;
MPI_Allgather(&map->from->size, 1, MPI_INT, sizes, 1, MPI_INT, OP_MPI_HDF5_WORLD);
for(int i = 0; i<comm_size; i++)g_size = g_size + sizes[i];
//Create the dataspace for the dataset.
dimsf[0] = g_size;
dimsf[1] = map->dim;
dataspace = H5Screate_simple(2, dimsf, NULL);
//Create the dataset with default properties and close dataspace.
if(sizeof(map->map[0]) == sizeof(int))
dset_id = H5Dcreate(file_id, map->name, H5T_NATIVE_INT, dataspace,
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
else if(sizeof(map->map[0]) == sizeof(long))
dset_id = H5Dcreate(file_id, map->name, H5T_NATIVE_LONG, dataspace,
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
else if(sizeof(map->map[0]) == sizeof(long long))
dset_id = H5Dcreate(file_id, map->name, H5T_NATIVE_LLONG, dataspace,
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
H5Sclose(dataspace);
//Each process defines dataset in memory and writes it to a hyperslab
//in the file.
int disp = 0;
for(int i = 0; i<my_rank; i++)disp = disp + sizes[i];
count[0] = map->from->size;
count[1] = dimsf[1];
offset[0] = disp;
offset[1] = 0;
memspace = H5Screate_simple(2, count, NULL);
//Select hyperslab in the file.
dataspace = H5Dget_space(dset_id);
H5Sselect_hyperslab(dataspace, H5S_SELECT_SET, offset, NULL, count, NULL);
//Create property list for collective dataset write.
plist_id = H5Pcreate(H5P_DATASET_XFER);
H5Pset_dxpl_mpio(plist_id, H5FD_MPIO_COLLECTIVE);
//write data
if(sizeof(map->map[0]) == sizeof(int))
H5Dwrite(dset_id, H5T_NATIVE_INT, memspace, dataspace, plist_id, map->map);
else if(sizeof(map->map[0]) == sizeof(long))
H5Dwrite(dset_id, H5T_NATIVE_LONG, memspace, dataspace, plist_id, map->map);
else if(sizeof(map->map[0]) == sizeof(long long))
H5Dwrite(dset_id, H5T_NATIVE_LLONG, memspace, dataspace, plist_id, map->map);
H5Pclose(plist_id);
H5Sclose(memspace);
H5Sclose(dataspace);
H5Dclose(dset_id);
free(sizes);
/*attach attributes to map*/
//open existing data set
dset_id = H5Dopen(file_id, map->name, H5P_DEFAULT);
//create the data space for the attribute
hsize_t dims = 1;
dataspace = H5Screate_simple(1, &dims, NULL);
//Create an int attribute - size
hid_t attribute = H5Acreate(dset_id, "size", H5T_NATIVE_INT, dataspace,
H5P_DEFAULT, H5P_DEFAULT);
//Write the attribute data.
H5Awrite(attribute, H5T_NATIVE_INT, &g_size);
//Close the attribute.
H5Aclose(attribute);
//Create an int attribute - dimension
attribute = H5Acreate(dset_id, "dim", H5T_NATIVE_INT, dataspace,
H5P_DEFAULT, H5P_DEFAULT);
//Write the attribute data.
H5Awrite(attribute, H5T_NATIVE_INT, &map->dim);
//Close the attribute.
H5Aclose(attribute);
H5Sclose(dataspace);
//Create an string attribute - type
dataspace= H5Screate(H5S_SCALAR);
hid_t atype = H5Tcopy(H5T_C_S1);
H5Tset_size(atype, 10);
attribute = H5Acreate(dset_id, "type", atype, dataspace,
H5P_DEFAULT, H5P_DEFAULT);
if(sizeof(map->map[0]) == sizeof(int))
H5Awrite(attribute, atype, "int");
if(sizeof(map->map[0]) == sizeof(long))
H5Awrite(attribute, atype, "long");
if(sizeof(map->map[0]) == sizeof(long long))
H5Awrite(attribute, atype, "long long");
H5Aclose(attribute);
//Close the dataspace
H5Sclose(dataspace);
//Close to the dataset.
H5Dclose(dset_id);
}
/*loop over all the op_dats and write them to file*/
for(int d=0; d<OP_dat_index; d++) {
op_dat dat=OP_dat_list[d];
//find total size of map
int* sizes = (int *)xmalloc(sizeof(int)*comm_size);
int g_size = 0;
MPI_Allgather(&dat->set->size, 1, MPI_INT, sizes, 1, MPI_INT, OP_MPI_HDF5_WORLD);
for(int i = 0; i<comm_size; i++)g_size = g_size + sizes[i];
//Create the dataspace for the dataset.
dimsf[0] = g_size;
dimsf[1] = dat->dim;
dataspace = H5Screate_simple(2, dimsf, NULL);
//Create the dataset with default properties and close dataspace.
if(strcmp(dat->type,"double")==0)
dset_id = H5Dcreate(file_id, dat->name, H5T_NATIVE_DOUBLE, dataspace,
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
else if(strcmp(dat->type,"float")==0)
dset_id = H5Dcreate(file_id, dat->name, H5T_NATIVE_FLOAT, dataspace,
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
else if(strcmp(dat->type,"int")==0)
dset_id = H5Dcreate(file_id, dat->name, H5T_NATIVE_INT, dataspace,
H5P_DEFAULT, H5P_DEFAULT, H5P_DEFAULT);
else printf("Unknown type\n");
H5Sclose(dataspace);
//Each process defines dataset in memory and writes it to a hyperslab
//in the file.
int disp = 0;
for(int i = 0; i<my_rank; i++)disp = disp + sizes[i];
count[0] = dat->set->size;
count[1] = dimsf[1];
offset[0] = disp;
offset[1] = 0;
memspace = H5Screate_simple(2, count, NULL);
//Select hyperslab in the file.
dataspace = H5Dget_space(dset_id);
H5Sselect_hyperslab(dataspace, H5S_SELECT_SET, offset, NULL, count, NULL);
//Create property list for collective dataset write.
plist_id = H5Pcreate(H5P_DATASET_XFER);
H5Pset_dxpl_mpio(plist_id, H5FD_MPIO_COLLECTIVE);
//write data
if(strcmp(dat->type,"double") == 0)
H5Dwrite(dset_id, H5T_NATIVE_DOUBLE, memspace, dataspace, plist_id, dat->data);
else if(strcmp(dat->type,"float") == 0)
H5Dwrite(dset_id, H5T_NATIVE_FLOAT, memspace, dataspace, plist_id, dat->data);
else if(strcmp(dat->type,"int") == 0)
H5Dwrite(dset_id, H5T_NATIVE_INT, memspace, dataspace, plist_id, dat->data);
else printf("Unknown type\n");
H5Pclose(plist_id);
H5Sclose(memspace);
H5Sclose(dataspace);
H5Dclose(dset_id);
free(sizes);
/*attach attributes to dat*/
//open existing data set
dset_id = H5Dopen(file_id, dat->name, H5P_DEFAULT);
//create the data space for the attribute
hsize_t dims = 1;
dataspace = H5Screate_simple(1, &dims, NULL);
//Create an int attribute - size
hid_t attribute = H5Acreate(dset_id, "size", H5T_NATIVE_INT, dataspace,
H5P_DEFAULT, H5P_DEFAULT);
//Write the attribute data.
H5Awrite(attribute, H5T_NATIVE_INT, &dat->size);
//Close the attribute.
H5Aclose(attribute);
//Create an int attribute - dimension
attribute = H5Acreate(dset_id, "dim", H5T_NATIVE_INT, dataspace,
H5P_DEFAULT, H5P_DEFAULT);
//Write the attribute data.
H5Awrite(attribute, H5T_NATIVE_INT, &dat->dim);
H5Aclose(attribute);
H5Sclose(dataspace);
//Create an string attribute - type
dataspace= H5Screate(H5S_SCALAR);
hid_t atype = H5Tcopy(H5T_C_S1);
H5Tset_size(atype, 10);
attribute = H5Acreate(dset_id, "type", atype, dataspace,
H5P_DEFAULT, H5P_DEFAULT);
H5Awrite(attribute, atype, dat->type);
H5Aclose(attribute);
//Close the dataspace.
H5Sclose(dataspace);
H5Dclose(dset_id);
}
H5Fclose(file_id);
op_timers(&cpu_t2, &wall_t2); //timer stop for hdf5 file write
//compute import/export lists creation time
time = wall_t2-wall_t1;
MPI_Reduce(&time, &max_time, 1, MPI_DOUBLE, MPI_MAX, MPI_ROOT, OP_MPI_HDF5_WORLD);
//print performance results
if(my_rank == MPI_ROOT)
{
printf("Max hdf5 file write time = %lf\n\n",max_time);
}
MPI_Comm_free(&OP_MPI_HDF5_WORLD);
}