int coefs_ga_create_ghost(int Nx, int Ny, int Nz, int nsplines) {
int data[Nx*Ny*Nz];
int g_a,dims[4]={Nx,Ny,Nz,nsplines},chunk[4]={-1,-1,-1,nsplines};
int width[4] = {3, 3, 3, 0};
int type=C_INT;
//g_a=NGA_Create(type,4,dims,"Coefs",chunk);
g_a=NGA_Create_ghosts(type, 4, dims, width, "Coefs", chunk);
int lo[4],hi[4],ld[3];
//int value=9; GA_Fill(g_a,&value);
GA_Print_distribution(g_a);
return g_a;
}
PetscErrorCode vizGA2DA()
{
PetscErrorCode ierr;
int rank;
MPI_Comm_rank(PETSC_COMM_WORLD,&rank);
int d1 = 40, d2 = 50;
DA da;
Vec vec;
const PetscInt *lx, *ly, *lz;
PetscInt m,n,p;
DALocalInfo info;
ierr = DACreate2d(PETSC_COMM_WORLD,DA_NONPERIODIC,DA_STENCIL_STAR,
d1,d2,PETSC_DECIDE,PETSC_DECIDE,1,1,0,0, &da); CHKERRQ(ierr);
ierr = DACreateGlobalVector(da, &vec); CHKERRQ(ierr);
ierr = DAGetOwnershipRanges(da, &lx, &ly, &lz); CHKERRQ(ierr);
ierr = DAGetLocalInfo(da,&info); CHKERRQ(ierr);
ierr = DAGetInfo(da,0,0,0,0,&m,&n,&p,0,0,0,0); CHKERRQ(ierr);
/**/
ierr = DAView(da, PETSC_VIEWER_STDOUT_WORLD); CHKERRQ(ierr);
for (int i = 0; i < m; ++i) {
PetscPrintf(PETSC_COMM_WORLD,"%d\tlx: %d\n",i,lx[i]);
}
for (int i = 0; i < n; ++i) {
PetscPrintf(PETSC_COMM_WORLD,"%d\tly: %d\n",i,ly[i]);
}
/**/
int ga = GA_Create_handle();
int ndim = 2;
int dims[2] = {d2,d1};
GA_Set_data(ga,2,dims,MT_DBL);
int *map;
PetscMalloc( sizeof(int)*(m+n), &map);
map[0] = 0;
for( int i = 1; i < n; i++ )
{
map[i] = ly[i-1] + map[i-1];
}
map[n] = 0;
for( int i = n+1; i < m+n; i++ )
{
map[i] = lx[i-n-1] + map[i-1];
}
/* correct ordering, but nodeid's dont line up with mpi rank for petsc's da
* DA: +---+---+ GA: +---+---+
* +-2-+-3-+ +-1-+-3-+
* +---+---+ +---+---+
* +-0-+-1-+ +-0-+-2-+
* +---+---+ +---+---+
int *map;
PetscMalloc( sizeof(int)*(m+n), &map);
map[0] = 0;
for( int i = 1; i < m; i++ )
{
map[i] = lx[i] + map[i-1];
}
map[m] = 0;
for( int i = m+1; i < m+n; i++ )
{
map[i] = ly[i-m] + map[i-1];
}
*/
int block[2] = {n,m};
GA_Set_irreg_distr(ga,map,block);
ierr = GA_Allocate( ga );
if( !ierr ) GA_Error("\n\n\nga allocaltion failed\n\n",ierr);
if( !ga ) GA_Error("\n\n\n ga null \n\n",ierr);
if( rank != GA_Nodeid() ) GA_Error("MPI rank does not match GA_Nodeid()",1);
GA_Print_distribution(ga);
int lo[2], hi[2];
NGA_Distribution(ga,rank,lo,hi);
if( lo[1] != info.xs || hi[1] != info.xs+info.xm-1 ||
lo[0] != info.ys || hi[0] != info.ys+info.ym-1 )
{
PetscSynchronizedPrintf(PETSC_COMM_SELF,"[%d] lo:(%2d,%2d) hi:(%2d,%2d) \t DA: (%2d,%2d), (%2d, %2d)\n",
rank, lo[1], lo[0], hi[1], hi[0], info.xs, info.ys, info.xs+info.xm-1, info.ys+info.ym-1);
}
PetscBarrier(0);
PetscSynchronizedFlush(PETSC_COMM_WORLD);
AO ao;
DAGetAO(da,&ao);
if( rank == 0 )
{
int *idx, len = d1*d2;
PetscReal *val;
PetscMalloc(sizeof(PetscReal)*len, &val);
PetscMalloc(sizeof(int)*len, &idx);
for (int j = 0; j < d2; ++j)
{
for (int i = 0; i < d1; ++i)
{
idx[i + d1*j] = i + d1*j;
val[i + d1*j] = i + d1*j;
}
}
AOApplicationToPetsc(ao,len,idx);
VecSetValues(vec,len,idx,val,INSERT_VALUES);
int a[2], b[2],ld[1]={0};
double c = 0;
for (int j = 0; j < d2; ++j)
{
for (int i = 0; i < d1; ++i)
{
a[0] = j;
a[1] = i;
// printf("%5.0f ",c);
NGA_Put(ga,a,a,&c,ld);
c++;
}
}
}
// GA_Print(ga);
VecAssemblyBegin(vec);
VecAssemblyEnd(vec);
int ld;
double *ptr;
NGA_Access(ga,lo,hi,&ptr,&ld);
PetscReal **d;
int c=0;
ierr = DAVecGetArray(da,vec,&d); CHKERRQ(ierr);
for (int j = info.ys; j < info.ys+info.ym; ++j)
{
for (int i = info.xs; i < info.xs+info.xm; ++i)
{
if( d[j][i] != ptr[(i-info.xs)+ld*(j-info.ys)] )
GA_Error("DA array is not equal to GA array",1);
// printf("%d (%d,%d):\t%3.0f\t%3.0f\n", c, i, j, d[j][i], ptr[(i-info.xs)+ld*(j-info.ys)]);
c++;
}
}
ierr = DAVecRestoreArray(da,vec,&d); CHKERRQ(ierr);
c=0;
PetscReal *v;
int start, end;
VecGetOwnershipRange(vec, &start, &end);
VecGetArray( vec, &v );
for( int i = start; i < end; i++)
{
// printf("%d:\t%3.0f\t%3.0f\t%s\n", start, v[i-start], ptr[i-start], (v[i-start]-ptr[i-start]==0?"":"NO") );
}
VecRestoreArray( vec, &v );
NGA_Release_update(ga,lo,hi);
Vec gada;
VecCreateMPIWithArray(((PetscObject)da)->comm,da->Nlocal,PETSC_DETERMINE,ptr,&gada);
VecView(gada,PETSC_VIEWER_STDOUT_SELF);
GA_Destroy(ga);
ierr = VecDestroy(vec); CHKERRQ(ierr);
ierr = DADestroy(da); CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char**argv)
{
int nprocs, me;
int i,j;
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&me);
GA_Initialize();
const int heap=3000000, stack=300000;
if(! MA_init(C_INT,stack,heap) ) GA_Error((char *) "MA_init failed",stack+heap /*error code*/);
int Nx=97; int Ny=97; int Nz = 97;
Nx+=3; Ny+=3; Nz+=3;
int data[Nx*Ny*Nz];
int num_splines = 32;
int g_a,dims[4]={Nx,Ny,Nz,num_splines},chunk[4]={-1,-1,-1,num_splines};
int width[4] = {3, 3, 3, 0};
int type=C_INT;
//g_a=NGA_Create(type,4,dims,"Coefs",chunk);
g_a=NGA_Create_ghosts(type, 4, dims, width, "Coefs", chunk);
int lo[4],hi[4],ld[3];
//double value=9.0; GA_Fill(g_a,&value);
GA_Print_distribution(g_a);
fflush(stdout);
if(me==0)
{
for (i=0; i<num_splines; i++)
{
int x, y, z;
for (x=0; x<Nx; x++)
for (y=0; y<Ny; y++)
for (z=0; z<Nz; z++)
{ j=x*(Ny*Nz)+y*Nz+z;
data[j] = (x*100*100+y*100+z)*100+i;}
lo[0]=lo[1]=lo[2]=0;
hi[0]=Nx-1;hi[1]=Ny-1;hi[2]=Nz-1;
lo[3]=hi[3]=i%num_splines;
ld[0]=Ny;ld[1]=Nz;ld[2]=1;
NGA_Put(g_a,lo,hi,data,ld);
}
}
GA_Update_ghosts(g_a);
GA_Sync();
printf("done\n"),fflush(stdout);
ga_coefs_t *ga_coefs = malloc(sizeof(ga_coefs_t));
ga_coefs->Mx = Nx;
ga_coefs->My = Ny;
ga_coefs->Mz = Nz;
ga_coefs->nsplines = num_splines;
ga_coefs->g_a=g_a;
int *coefs1 = (int*)malloc((size_t)1*sizeof(int)*4*4*4*num_splines);
int ix,iy,iz;
Nx-=3; Ny-=3; Nz-=3;
ga_coefs->sumt=ga_coefs->amount=0;
NGA_Distribution(g_a,me,lo,hi);
GA_Print_distribution(g_a);
int low[16][4],high[16][4];
for(i=0;i<nprocs;i++)
NGA_Distribution(g_a,i,low[i],high[i]);
srand ( time(NULL) );
int k=GA_Nodeid();
printf("%d: low[k]=%d high[k]=%d\n", GA_Nodeid(), low[k][2], high[k][2]);
int unequal=0;
for(i=0;i<1000;i++) {
ix=rand_index(low[k][0],high[k][0]);
if(ix+3>=dims[0]) ix=low[k][0];
iy=rand_index(low[k][1],high[k][1]);
if(iy+3>=dims[1]) iy=low[k][1];
iz=rand_index(low[k][2],high[k][2]);
if(iz+3>=dims[2]) iz=low[k][2];
coefs_ga_get_3d(ga_coefs,coefs1,ix,iy,iz);
long get_sum=mini_cube_sum(coefs1, ga_coefs->nsplines);
long ghost_sum=coefs_ghost_access_3d(ga_coefs->g_a, ix, iy, iz, ga_coefs->nsplines);
if(get_sum!=ghost_sum) {
printf("ixyz=\t%d\t%d\t%d\t", ix, iy, iz);
printf("get_sum=%ld ghost_sum=%ld\n", get_sum, ghost_sum);
unequal++;
}
}
printf("unequal count=%d\n", unequal);
free(coefs1);
GA_Terminate();
MPI_Finalize();
return 0;
}
static int test(int shape_idx, int type_idx, int dist_idx)
{
int type = TYPES[type_idx];
int *dims = SHAPES[shape_idx];
int ndim = SHAPES_NDIM[shape_idx];
mock_ga_t *mock_a, *result_a;
int g_a;
int buffer[100];
int lo[GA_MAX_DIM], hi[GA_MAX_DIM], ld[GA_MAX_DIM], shape[GA_MAX_DIM];
int result=0, error_index=-1, error_proc=-1;
mock_a = Mock_Create(type, ndim, dims, "mock", NULL);
result_a = Mock_Create(type, ndim, dims, "mock", NULL);
g_a = create_function[dist_idx](type, ndim, dims);
mock_data(mock_a, g_a);
mock_to_global(mock_a, g_a);
Mock_Zero(mock_a);
GA_Zero(g_a);
global_to_mock(g_a, result_a);
result = neq_mock(mock_a, result_a, &error_index);
if (0 != result) {
error_proc = GA_Nodeid();
}
GA_Igop(&result, 1, "+");
GA_Igop(&error_proc, 1, "max");
if (error_proc != GA_Nodeid()) {
error_index = 0;
}
GA_Igop(&error_index, 1, "+");
if (0 != result) {
if (error_proc == GA_Nodeid()) {
printf("ERROR: local result failed to compare to global result\n");
printf("\terror_proc=%d\n", error_proc);
printf("\terror_index=%d\n", error_index);
printf("***LOCAL RESULT***\n");
Mock_Print(mock_a);
printf("***GLOBAL RESULT***\n");
Mock_Print(result_a);
printf("\tprinting array distribution\n");
}
GA_Sync();
GA_Print(g_a);
GA_Print_distribution(g_a);
return 1;
}
Mock_Destroy(mock_a);
Mock_Destroy(result_a);
GA_Destroy(g_a);
return 0;
}
static int test(int shape_idx, int type_idx, int dist_idx)
{
int type = TYPES[type_idx];
int *dims = SHAPES[shape_idx];
int ndim = SHAPES_NDIM[shape_idx];
mock_ga_t *mock_a, *result_a;
int g_a;
void *alpha = NULL;
int buffer[100];
int lo[GA_MAX_DIM], hi[GA_MAX_DIM], ld[GA_MAX_DIM], shape[GA_MAX_DIM];
int result=0, error_index=-1, error_proc=-1;
int ival = 6;
long lval = 7;
long long llval = 8;
float fval = 9;
double dval = 10;
SingleComplex cval = {11,12};
DoubleComplex zval = {13,14};
/* create the local array and result array */
mock_a = Mock_Create(type, ndim, dims, "mock", NULL);
result_a = Mock_Create(type, ndim, dims, "mock", NULL);
/* create the global array */
g_a = create_function[dist_idx](type, ndim, dims);
/* create meaningful data for local array */
mock_data(mock_a, g_a);
/* init global array with same data as local array */
mock_to_global(mock_a, g_a);
switch (type) {
case C_INT: alpha = &ival; break;
case C_LONG: alpha = &lval; break;
case C_LONGLONG: alpha = &llval; break;
case C_FLOAT: alpha = &fval; break;
case C_DBL: alpha = &dval; break;
case C_SCPL: alpha = &cval; break;
case C_DCPL: alpha = &zval; break;
}
/* call the local routine */
Mock_Add_constant(mock_a, alpha);
/* call the global routine */
GA_Add_constant(g_a, alpha);
/* get the results from the global array */
global_to_mock(g_a, result_a);
/* compare the results */
result = neq_mock(mock_a, result_a, &error_index);
if (0 != result) {
error_proc = GA_Nodeid();
}
/* make sure all procs get same result so they can die gracefully */
GA_Igop(&result, 1, "+");
/* if error occured, find the highest failing node ID */
GA_Igop(&error_proc, 1, "max");
/* clear the error index for all but the highest failing node ID */
if (error_proc != GA_Nodeid()) {
error_index = 0;
}
/* make sure all procs get the error index on the highest failing node ID */
GA_Igop(&error_index, 1, "+");
if (0 != result) {
if (error_proc == GA_Nodeid()) {
printf("ERROR: local result failed to compare to global result\n");
printf("\terror_proc=%d\n", error_proc);
printf("\terror_index=%d\n", error_index);
printf("***LOCAL RESULT***\n");
Mock_Print(mock_a);
printf("***GLOBAL RESULT***\n");
Mock_Print(result_a);
printf("\tprinting array distribution\n");
}
GA_Sync();
GA_Print(g_a);
GA_Print_distribution(g_a);
return 1;
}
/* clean up */
Mock_Destroy(mock_a);
Mock_Destroy(result_a);
GA_Destroy(g_a);
return 0;
}
int main(int argc, char **argv) {
int me;
int g_a;
int status;
int i,j;
int dims[] = {n,n};
int proc_group[PROC_LIST_SIZE],proclist[PROC_LIST_SIZE],inode;
int sbuf[1],rbuf[1];
MPI_Comm comm;
MP_INIT(argc,argv);
GA_Initialize();
me = GA_Nodeid();
status = MA_init(MT_DBL, 100000, 100000);
if (!status) GA_Error("ma_init failed",-1);
status = MA_set_auto_verify(1);
status = MA_set_hard_fail(1);
status = MA_set_error_print(1);
inode = GA_Cluster_nodeid();
if (me == 0) {
printf("there are %d nodes, node 0 has %d procs\n",
GA_Cluster_nnodes(), GA_Cluster_nprocs(0));
fflush(stdout);
}
GA_Sync();
for (i=0; i<GA_Cluster_nnodes(); ++i) {
for (j=0; j<GA_Cluster_nprocs(i); ++j) {
proclist[j]=GA_Cluster_procid(i,j);
}
proc_group[i]=GA_Pgroup_create(proclist,GA_Cluster_nprocs(i));
}
GA_Sync();
for (i=0; i<GA_Cluster_nnodes(); ++i) {
if (i == inode) {
printf("%d joining group %d\n", me, proc_group[inode]);
GA_Pgroup_set_default(proc_group[inode]);
g_a = NGA_Create(C_DBL, 2, dims, "a", NULL);
if (!g_a) GA_Error("NGA_Create failed",-1);
printf("%d Created array of group %d as proc no. %d\n",
me, proc_group[inode], GA_Nodeid());
GA_Print_distribution(g_a);
comm = GA_MPI_Comm_pgroup_default();
if (comm != MPI_COMM_NULL) {
sbuf[0] = GA_Nodeid();
status = MPI_Allreduce(sbuf, rbuf, 1, MPI_INT, MPI_MAX, comm);
printf("%d max nodeid is %d\n", me, rbuf[0]);
if ((rbuf[0]+1) != GA_Cluster_nprocs(i)) {
GA_Error("MPI_Allreduce failed",1);
}
}
else {
printf("MPI_Comm was null!\n");
}
GA_Pgroup_set_default(GA_Pgroup_get_world());
}
GA_Sync();
}
GA_Terminate();
MP_FINALIZE();
return 0;
}
// note: Sayan: brings down memory requirement to about 268 MB
int main(int argc, char **argv)
{
int me, nproc, g_a = -1, i, j;
#if defined(USE_ELEMENTAL)
int ndim=2, dims[2]= {N1,N2};
#else
int ndim=2, type=MT_F_DBL, dims[2]= {N1,N2};
#endif
double *buf;
int lo[2], hi[2], ld[1];
double alpha = 1.0;
#if defined(USE_ELEMENTAL)
// initialize Elemental (which will initialize MPI)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &me );
ElMPICommSize( MPI_COMM_WORLD, &nproc );
ElGlobalArrays_d eldga;
// instantiate el::global array
ElGlobalArraysConstruct_d( &eldga );
// initialize global arrays
ElGlobalArraysInitialize_d( eldga );
printf ("INITIALIZED elemental global array...\n");
#else
MP_INIT(argc,argv);
GA_Initialize_ltd(-1);
me=GA_Nodeid();
nproc=GA_Nnodes();
#endif
if(me==0) printf("Using %ld processes\n",(long)nproc);
if(me==0) printf("memory = %ld bytes\n",((long)N1)*((long)N2)*8);
#if defined(USE_ELEMENTAL)
// create and allocate a global array
printf ("ndim = %d\n", ndim);
printf ("dim[0] = %d and dim[1] = %d\n", dims[0], dims[1]);
ElGlobalArraysCreate_d( eldga, ndim, dims, "A", &g_a);
printf ("CREATED elemental global array...\n");
// print distribution
ElGlobalArraysPrint_d( eldga, g_a );
#else
g_a = NGA_Create(type, ndim, dims, "A", NULL);
GA_Zero(g_a); /* zero the matrix */
GA_Print_distribution(g_a);
#endif
if(me == 0) {
// buf = (double*)(malloc(N1*1024*sizeof(double)));
buf = (double*)(malloc(N1*128*sizeof(double)));
// for(j = 0; j < N1*1024; ++j) buf[j] = 1.0;
// for(i = 0; i < N2/1024; ++i) {
for(j = 0; j < N1*128; ++j) buf[j] = 1.0;
for(i = 0; i < N2/128; ++i) {
lo[0] = 0;
hi[0] = lo[0] + N1 -1;
/*
lo[1] = i*1024;
hi[1] = lo[1] + 1024 -1;
ld[0] = 1024;
*/
lo[1] = i*128;
hi[1] = lo[1] + 128 -1;
ld[0] = 128;
printf("NGA_Acc.%d: %d:%d %d:%d\n",i,lo[0],hi[0],lo[1],hi[1]);
#if defined(USE_ELEMENTAL)
ElGlobalArraysAccumulate_d( eldga, g_a, lo, hi, buf, ld, &alpha );
// there is an explicit flush in NGA_Acc/Put, so when it returns, the buffer
// can be reused and data has reached the destination
#else
NGA_Init_fence();
NGA_Acc(g_a, lo, hi, buf, ld, &alpha);
NGA_Fence();
#endif
}
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysSync_d( eldga );
ElGlobalArraysDestroy_d( eldga, g_a );
ElGlobalArraysTerminate_d( eldga );
// call el::global arrays destructor
ElGlobalArraysDestruct_d( eldga );
ElFinalize();
#else
GA_Sync();
GA_Destroy(g_a);
GA_Terminate();
MP_FINALIZE();
#endif
return 0;
}
int main(int argc, char**argv)
{
int nprocs, me;
int i,j;
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&me);
GA_Initialize();
int flag=1;
int Nx=108; int Ny=108; int Nz = 108;
Nx+=3; Ny+=3; Nz+=3;
int data[Nx*Ny*Nz];
int num_splines = 32;
int g_a,dims[4]={Nx,Ny,Nz,num_splines},chunk[4]={-1,-1,-1,num_splines};
int type=C_INT;
g_a=NGA_Create(type,4,dims,"Coefs",chunk);
int lo[4],hi[4],ld[3];
double value=9.0;
GA_Fill(g_a,&value);
GA_Print_distribution(g_a);
fflush(stdout);
if(me==0)
{
for (i=0; i<num_splines; i++)
{
for (j=0; j<Nx*Ny*Nz; j++)
data[j] = rand()%1000;
lo[0]=lo[1]=lo[2]=0;
hi[0]=Nx-1;hi[1]=Ny-1;hi[2]=Nz-1;
lo[3]=hi[3]=i%num_splines;
ld[0]=Ny;ld[1]=Nz;ld[2]=1;
NGA_Put(g_a,lo,hi,data,ld);
}
}
printf("done\n"),fflush(stdout);
GA_Sync();
ga_coefs_t *ga_coefs = malloc(sizeof(ga_coefs_t));
ga_coefs->Mx = Nx;
ga_coefs->My = Ny;
ga_coefs->Mz = Nz;
ga_coefs->nsplines = num_splines;
ga_coefs->g_a=g_a;
int *coefs1 = (int*)malloc((size_t)1*sizeof(int)*4*4*4*num_splines);
int ix,iy,iz;
Nx-=3; Ny-=3; Nz-=3;
ga_coefs->sumt=ga_coefs->amount=0;
NGA_Distribution(g_a,me,lo,hi);
int low[16][4],high[16][4];
for(i=0;i<nprocs;i++)
NGA_Distribution(g_a,i,low[i],high[i]);
srand ( time(NULL) );
int k;
for(k=0;k<nprocs;k++)
{
ga_coefs->sumt=ga_coefs->amount=0;
{
for(i=0;i<1000;i++)
{
ix=rand_index(low[k][0],high[k][0]);
iy=rand_index(low[k][1],high[k][1]);
iz=rand_index(low[k][2],high[k][2]);
coefs_ga_get_3d(ga_coefs,coefs1,ix,iy,iz);
mini_cube_sum(coefs1, ga_coefs->nsplines);
}
}
printf("<%d,%d>\t %lf \t %d \t %lf\n", me,k, ga_coefs->sumt, ga_coefs->amount, ga_coefs->sumt/ga_coefs->amount),fflush(stdout);
}
free(coefs1);
GA_Terminate();
MPI_Finalize();
return 0;
}
void do_work()
{
int g_a, g_b;
int me=GA_Nodeid(), nproc=GA_Nnodes(), proc, loop;
int dims[NDIM], lo[NDIM], hi[NDIM], block[NDIM], ld[NDIM-1];
int i,d,*proclist, offset;
int adims[NDIM], ndim,type;
typedef struct {
int lo[NDIM];
int hi[NDIM];
} patch_t;
patch_t *regions;
int *map;
double *buf;
/***** create array A with default distribution *****/
if(me==0){printf("Creating array A\n"); fflush(stdout);}
for(i = 0; i<NDIM; i++)dims[i] = N*(i+1);
#ifdef NEW_API
g_a = GA_Create_handle();
GA_Set_data(g_a,NDIM,dims,MT_F_DBL);
GA_Set_array_name(g_a,"array A");
(void)GA_Allocate(g_a);
#else
g_a = NGA_Create(MT_F_DBL, NDIM, dims, "array A", NULL);
#endif
if(!g_a) GA_Error("create failed: A",0);
if(me==0)printf("OK\n\n");
/* print info about array we got */
NGA_Inquire(g_a, &type, &ndim, adims);
GA_Print_distribution(g_a);
GA_Sync();
/* duplicate array A with ga_create irreg rather than ga_duplicate
* -- want to show distribution control
* -- with ga_duplicate it would be g_b=GA_Duplicate(g_a,name)
*/
if(me==0)printf("\nReconstructing distribution description for A\n");
/* get memory for arrays describing distribution */
proclist = (int*)malloc(nproc*sizeof(int));
if(!proclist)GA_Error("malloc failed for proclist",0);
regions = (patch_t*)malloc(nproc*sizeof(patch_t));
if(!regions)GA_Error("malloc failed for regions",0);
map = (int*)malloc((nproc+ndim)*sizeof(int)); /* ubound= nproc+mdim */
if(!map)GA_Error("malloc failed for map",0);
/* first find out how array g_a is distributed */
for(i=0;i<ndim;i++)lo[i]=BASE;
for(i=0;i<ndim;i++)hi[i]=adims[i] -1 + BASE;
proc = NGA_Locate_region(g_a, lo, hi, (int*)regions, proclist);
if(proc<1) GA_Error("error in NGA_Locate_region",proc);
/* determine blocking for each dimension */
for(i=0;i<ndim;i++)block[i]=0;
for(i=0;i<ndim;i++)adims[i]=0;
offset =0;
for(d=0; d<ndim; d++)
for(i=0;i<proc;i++)
if( regions[i].hi[d]>adims[d] ){
map[offset] = regions[i].lo[d];
offset++;
block[d]++;
adims[d]= regions[i].hi[d];
}
if(me==0){
printf("Distribution map contains %d elements\n",offset);
print_subscript("number of blocks for each dimension",ndim,block,"\n");
print_subscript("distribution map",offset,map,"\n\n");
fflush(stdout);
}
if(me==0)printf("Creating array B applying distribution of A\n");
# ifdef USE_DUPLICATE
g_b = GA_Duplicate(g_a,"array B");
# else
g_b = NGA_Create_irreg(MT_F_DBL, NDIM, dims, "array B", block,map);
# endif
if(!g_b) GA_Error("create failed: B",0);
if(me==0)printf("OK\n\n");
free(proclist); free(regions); free(map);
GA_Print_distribution(g_b);
GA_Sync();
if(me==0){
printf("\nCompare distributions of A and B\n");
if(GA_Compare_distr(g_a,g_b))
printf("Failure: distributions NOT identical\n");
else
printf("Success: distributions identical\n");
fflush(stdout);
}
if(me==0){
printf("\nAccessing local elements of A: set them to the owner process id\n");
fflush(stdout);
}
GA_Sync();
NGA_Distribution(g_a,me,lo,hi);
if(hi[0]>=0){/* -1 means no elements stored on this processor */
double *ptr;
int locdim[NDIM];
NGA_Access(g_a, lo,hi, &ptr, ld);
for(i=0;i<ndim;i++)locdim[i]=hi[i]-lo[i]+1;
fill_patch(ptr, locdim, ld, ndim,(double)me);
}
for(i=0;i<nproc; i++){
if(me==i && hi[0]>=0){
char msg[100];
sprintf(msg,"%d: leading dimensions",me);
print_subscript(msg,ndim-1,ld,"\n");
fflush(stdout);
}
GA_Sync();
}
GA_Sync();
if(me==0)printf("\nRandomly checking the update using ga_get on array sections\n");
GA_Sync();
/* show ga_get working and verify array updates
* every process does N random gets
* for simplicity get only a single row at a time
*/
srand(me); /* different seed for every process */
hi[ndim-1]=adims[ndim-1] -1 + BASE;
for(i=1;i<ndim-1; i++)ld[i]=1; ld[ndim-2]=adims[ndim-1] -1 + BASE;
/* get buffer memory */
buf = (double*)malloc(adims[ndim-1]*sizeof(double));
if(!buf)GA_Error("malloc failed for buf",0);
/* half of the processes check the result */
if(me<=nproc/2)
for(loop = 0; loop< N; loop++){ /* task parallel loop */
lo[ndim-1]=BASE;
for (i= 0; i < ndim -1; i ++){
lo[i] = hi[i] = rand()%adims[i]+BASE;
}
/* print_subscript("getting",ndim,lo,"\n");*/
NGA_Get(g_a,lo,hi,buf,ld);
/* check values */
for(i=0;i<adims[ndim-1]; i++){
int p = NGA_Locate(g_a, lo);
if((double)p != buf[i]) {
char msg[100];
sprintf(msg,"%d: wrong value: %d != %lf a",me, p, buf[i]);
print_subscript(msg,ndim,lo,"\n");
GA_Error("Error - bye",i);
}
lo[ndim-1]++;
}
}
free(buf);
GA_Sync();
if(me==0)printf("OK\n");
GA_Destroy(g_a);
GA_Destroy(g_b);
}
static int test(int shape_idx, int type_idx, int dist_idx)
{
int type = TYPES[type_idx];
int *dims = SHAPES[shape_idx];
int ndim = SHAPES_NDIM[shape_idx];
mock_ga_t *mock_a, *mock_b, *mock_c, *mock_r;
int g_a, g_b, g_c;
int buffer[100];
int lo[GA_MAX_DIM], hi[GA_MAX_DIM], ld[GA_MAX_DIM], shape[GA_MAX_DIM];
int result=0, error_index=-1, error_proc=-1;
mock_a = Mock_Create(type, ndim, dims, "mock", NULL);
mock_b = Mock_Create(type, ndim, dims, "mock", NULL);
mock_c = Mock_Create(type, ndim, dims, "mock", NULL);
mock_r = Mock_Create(type, ndim, dims, "mock", NULL);
g_a = create_function[dist_idx](type, ndim, dims);
g_b = create_function[dist_idx](type, ndim, dims);
g_c = create_function[dist_idx](type, ndim, dims);
/* create meaningful data for local array */
mock_data(mock_a, g_a);
mock_data(mock_b, g_b);
/* init global array with same data as local array */
mock_to_global(mock_a, g_a);
mock_to_global(mock_b, g_b);
/* call the local routine */
Mock_Elem_multiply(mock_a, mock_b, mock_c);
/* call the global routine */
GA_Elem_multiply(g_a, g_b, g_c);
/* get the results from the global array */
global_to_mock(g_c, mock_r);
/* compare the results */
result = neq_mock(mock_c, mock_r, &error_index);
if (0 != result) {
error_proc = GA_Nodeid();
}
/* make sure all procs get same result so they can die gracefully */
GA_Igop(&result, 1, "+");
/* if error occured, find the highest failing node ID */
GA_Igop(&error_proc, 1, "max");
/* clear the error index for all but the highest failing node ID */
if (error_proc != GA_Nodeid()) {
error_index = 0;
}
/* make sure all procs get the error index on the highest failing node ID */
GA_Igop(&error_index, 1, "+");
if (0 != result) {
if (error_proc == GA_Nodeid()) {
printf("ERROR: local result failed to compare to global result\n");
printf("\terror_proc=%d\n", error_proc);
printf("\terror_index=%d\n", error_index);
printf("***LOCAL RESULT***\n");
Mock_Print(mock_a);
printf("***GLOBAL RESULT***\n");
Mock_Print(mock_r);
printf("\tprinting array distribution\n");
}
GA_Sync();
GA_Print(g_a);
GA_Print_distribution(g_a);
return 1;
}
/* clean up */
Mock_Destroy(mock_a);
Mock_Destroy(mock_r);
GA_Destroy(g_a);
return 0;
}
