main(int argc, char **argv)
{
int rank, nprocs, i, j;
int g_A, g_B, g_C, local_C[DIM][DIM], dims[DIM]={5,5};
int val_A=5, val_B=3, ld=DIM, max;
int lo[DIM]={2,2}, hi[DIM]={4,4}, blo[DIM]={0,0}, bhi[DIM]={2,2}, clo[DIM]={1,1}, chi[DIM]={3,3};
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_B = NGA_Create(C_INT, DIM, dims, "array_B", NULL);
g_C = NGA_Create(C_INT, DIM, dims, "array_C", NULL);
GA_Fill(g_A, &val_A);
GA_Fill(g_B, &val_B);
GA_Zero(g_C);
GA_Elem_maximum_patch(g_A, lo, hi, g_B, blo, bhi, g_C, clo, chi);
GA_Print(g_C);
GA_Sync();
NGA_Get(g_C, clo, chi, local_C, &ld);
if(rank==1)
{
for(i=0; i<DIM; i++)
{
for(j=0; j<DIM; j++)printf("%d ", local_C[i][j]);
printf("\n");
}
if(val_A>val_B) max=val_A;
else max=val_B;
for(i=0; i<DIM; i++)
{
for(j=0; j<DIM; j++)
if(local_C[i][j]!=max) printf("GA Error : \n");
}
}
GA_Sync();
if(rank == 0)
printf("Test Completed \n");
GA_Terminate();
MPI_Finalize();
}
main(int argc, char **argv)
{
int rank, nprocs, i, j;
int g_A, g_B;
int dims[MAX_DIM], val=4, ndim, re;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
for(i=1; i<=MAX_DIM; i++)
{
ndim=i;
dims[i]=SIZE;
// for(j=0; j<ndim; j++)
g_A = NGA_Create(C_INT, ndim, dims, "array_A", NULL);
g_B = NGA_Create(C_INT, ndim, dims, "array_B", NULL);
if(!g_A)
GA_Error("GA Error: no global array exists \n", ndim);
if(!g_B)
GA_Error("GA Error: no global array exists \n", ndim);
}
GA_Sync();
GA_Fill(g_A, &val);
re=GA_Solve(g_A, g_B);
if(re==0)
printf("Cholesky Fact is Successful \n");
else if (re >0)
printf("Cholesky Fact couldn't be completed \n");
else
printf("An Error occured\n");
if(rank == 0)
GA_PRINT_MSG();
GA_Destroy(g_A);
GA_Destroy(g_B);
GA_Terminate();
MPI_Finalize();
}
int main(int argc, char **argv)
{
int rank, nprocs, n=1;
int g_A, dims[DIM]={5,5};
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
printf("check %d \n", n);
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
if(GA_Ndim(g_A)!=DIM)
printf("ERROR: GA_Ndim didnt return nDimension after GA_Initialize\n");
printf("%d : %d \n", rank, GA_Ndim(g_A));
GA_Terminate();
if(rank==0)
printf(" GA: Test Completed \n");
MPI_Finalize();
return 0;
}
irregular_array2(int rank)
{
int g_A, g_B;
int dims[DIM]={GSIZE,GSIZE}, dims2[DIM], block[DIM]={3,2}, map[5]={0,2,6,0,4}, val_A=4, val_B=7;
int n_block[DIM], block_dims[DIM], i;
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_B = NGA_Create_irreg(C_INT, DIM, dims, "array_B", block, map);
GA_Fill(g_A, &val_A);
GA_Print(g_A);
GA_Fill(g_B, &val_B);
GA_Print(g_B);
GA_Sync();
/*
GA_Get_block_info(g_B, n_block, block_dims);
for(i=0; i<DIM; i++)
printf(" %d: %d ___ %d --- \n", rank, n_block[i], block_dims[i]);
*/
GA_Destroy(g_A);
GA_Destroy(g_B);
}
irregular_array1(int rank)
{
int g_A, g_B;
int dims[DIM]={5,10}, dims2[DIM], ndim, type, value=5, block[DIM]={2,3}, map[5]={0,2,0,4,6}, val=7;
int n_block[DIM], block_dims[DIM], i;
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_B = NGA_Create_irreg(C_INT, DIM, dims, "array_B", block, map);
GA_Fill(g_A, &value);
GA_Print(g_A);
GA_Fill(g_B, &val);
GA_Print(g_B);
GA_Sync();
NGA_Inquire(g_A, &type, &ndim, dims2);
//printf(" %d -- %d,,\n", type, ndim);
/*
GA_Get_block_info(g_B, n_block, block_dims);
for(i=0; i<DIM; i++)
printf(" %d: %d ___ %d --- \n", rank, n_block[i], block_dims[i]);
*/
GA_Destroy(g_A);
GA_Destroy(g_B);
}
fillandscale(int rank, int nprocs)
{
int g_A, val1=5, val2=5, local_A[SIZE][SIZE], i, j;
int dims[DIM]={SIZE,SIZE}, alo[DIM]={1,1}, ahi[DIM]={2,2}, ld=5;
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
GA_Zero(g_A);
NGA_Fill_patch(g_A, alo, ahi, &val1);
GA_Print(g_A);
GA_Scale(g_A, &val2);
GA_Print(g_A);
NGA_Get(g_A, alo, ahi, local_A, &ld);
if(rank == 1)
{
for(i=0; i<DIM; i++)
{
for(j=0; j<DIM; j++) if(local_A[i][j]!=val1*val2)
printf(" GA ERROR: \n");
}
}
GA_Destroy(g_A);
}
int main(int argc, char** argv)
{
int nprocs,myid,nprocssq;
int dims[2],chunk[2];
int i,j,k;
int stack = 100000, heap = 100000;
MPI_Init(&argc,&argv);
GA_Initialize();
MA_init(C_DBL,stack,heap);
nprocssq = GA_Nnodes();
nprocs = sqrt(nprocssq);
myid = GA_Nodeid();
dims[0] = N; dims[1] = N;
chunk[0] = N/nprocs;
chunk[1] = N/nprocs;
int g_a = NGA_Create(C_DBL,2,dims,"Array A",chunk);
int lo[2],hi[2];
NGA_Distribution(g_a,myid,lo,hi);
int ld[1] = {N/nprocs};
void *ptr;
double *local;
printf("Myid = %d, lo = [%d,%d] , hi = [%d,%d] , ld = %d \n",myid,lo[0],lo[1],hi[0],hi[1],ld[0]);
NGA_Access(g_a,lo,hi,&ptr,ld);
local = (double*) ptr;
printf("Myid = %d , local[0][0] = %f\n",*local);
GA_Sync();
GA_Destroy(g_a);
GA_Terminate();
MPI_Finalize();
return 0;
}
Integer util_gnxtval_(Integer *val) {
if(*val > 0) {
if(!initialized) ga_error("nxtval: not yet initialized", 0L);
return (Integer) NGA_Read_inc(g_T, &subscript, 1);
}
else if(*val==0) {
int n = 1;
initialized=1;
/* create task array */
g_T = NGA_Create(C_LONG, 1, &n,"Atomic Task", NULL);
/* Initialize the task array */
if(GA_Nodeid()==0) {
int lo=0, hi=0;
NGA_Put (g_T, &lo, &hi, &initval, &hi);
initval=0;
}
GA_Sync();
return 0;
}
else if (*val < 0) { GA_Destroy(g_T); initialized=0; initval=0; return 0;}
ga_error("nxtval: invalid value passed", 0L);
return -1;
}
main(int argc, char **argv)
{
int rank, nprocs;
int g_A, g_V, val1=5, val2=5, local_A[SIZE][SIZE], dims_V=SIZE, local_V[dims_V];
int dims[DIM]={SIZE,SIZE}, dims2[DIM], lo[DIM]={1,1}, hi[DIM]={2,2}, ld=5, i, j;
int loV=0, hiV=dims_V-1;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_V = NGA_Create(C_INT, 1, &dims_V, "array_A", NULL);
GA_Fill(g_A, &val1);
GA_Print(g_A);
printf("\n");
GA_Scale(g_A, &val2);
GA_Print(g_A);
GA_Get_diag(g_A, g_V);
GA_Print(g_V);
NGA_Get(g_A, lo, hi, local_A, &ld);
NGA_Get(g_V, &loV, &hiV, local_V, &ld);
if(rank==1)
{
for(i=0; i<dims_V; i++)
if(local_V[i]!=val1*val2) printf(" GA Error: \n");
}
if(rank == 0)
printf("Test Completed \n");
GA_Terminate();
MPI_Finalize();
}
int main(int argc, char **argv)
{
int size_dst = 15;
int g_a = 0;
int I_NEG_ONE = -1;
long L_NEG_ONE = -1;
long long LL_NEG_ONE = -1;
int FIVE = 5;
int TEN = 10;
int lo;
int hi;
int *ptr;
int i;
MP_INIT(argc,argv);
GA_INIT(argc,argv);
for (i=0; i<3; ++i) {
if (0 == i) {
g_a = NGA_Create(C_INT, 1, &size_dst, "dst", NULL);
GA_Fill(g_a, &I_NEG_ONE);
} else if (1 == i) {
g_a = NGA_Create(C_LONG, 1, &size_dst, "dst", NULL);
GA_Fill(g_a, &L_NEG_ONE);
} else if (2 == i) {
g_a = NGA_Create(C_LONGLONG, 1, &size_dst, "dst", NULL);
GA_Fill(g_a, &LL_NEG_ONE);
}
GA_Sync();
GA_Print(g_a);
NGA_Print_patch(g_a, &FIVE, &TEN, 0);
NGA_Print_patch(g_a, &FIVE, &TEN, 1);
NGA_Distribution(g_a, GA_Nodeid(), &lo, &hi);
NGA_Access(g_a, &lo, &hi, &ptr, NULL);
printf("[%d] (%d)=%d\n", GA_Nodeid(), lo, *ptr);
NGA_Release(g_a, &lo, &hi);
}
GA_Terminate();
MP_FINALIZE();
exit(EXIT_SUCCESS);
}
main(int argc, char **argv)
{
int rank, nprocs;
int g_A, dims[D]={SIZE,SIZE}, *local_A=NULL, *local_G=NULL, **sub_array=NULL, **s_array=NULL;
int i, j, value=5;
MPI_Init(&argc, &argv);
GA_Initialize();
MA_init(C_INT, 1000, 1000);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
s_array=(int**)malloc(N*sizeof(int*));
for(i=0; i<N; i++)
{
s_array[i]=(int*)malloc(D*sizeof(int));
for(j=0; j<D; j++) s_array[i][j]=rand()%10;
}
sub_array=(int**)malloc(N*sizeof(int*));
for(i=0; i<N; i++)
{
sub_array[i]=(int*)malloc(D*sizeof(int));
for(j=0; j<D; j++) sub_array[i][j]=rand()%10;
}
for(i=0; i<N; i++) local_A=(int*)malloc(N*sizeof(int));
for(i=0; i<N; i++) local_G=(int*)malloc(N*sizeof(int));
g_A=NGA_Create(C_INT, D, dims, "array_A", NULL);
GA_Fill(g_A, &value);
GA_Sync();
NGA_Scatter(g_A, local_A, s_array, N);
NGA_Gather(g_A, local_G, s_array, N);
GA_Sync();
GA_Print(g_A);
if(rank==0)
{
for(i=0; i<N; i++)
if(local_G[i]!=local_A[i]) printf("GA Error: \n");
}
GA_Sync();
if(rank==0)
GA_PRINT_MSG();
GA_Terminate();
MPI_Finalize();
return 0;
}
/**
* Block Topology (of Force Matrix):
* Say for example: If there are 4 block and 100 atoms, the size of
* the force matrix is 100x100 and each block size is 50x50.
*
* -----------
* | | |
* | 0,0 | 0,1 |
* -----------
* | 1,0 | 1,1 |
* | | |
* -----------
*/
int SetupBlocks(AppCtx *user)
{
int i,j,k=0;
int n;
int zero = 0;
int x_space, g_space;
if (user->natoms % user->BlockSize) {
GA_Error("Number of atoms should be a multiple of block size. Choose a different block size.", 0L);
}
n = user->natoms / user->BlockSize;
user->nBlocks = n*n;
if (user->nBlocks > MAX_BLOCKS)
GA_Error("Number of blocks is greater that MAX_BLOCKS: Solution is either to increase the defined MAX_BLOCKS or increase your block size",0L);
if (user->nBlocks < user->nproc)
GA_Error("Number of blocks should be greater than or equal to the number of processors",0L);
for (i=0;i<n;i++)
for (j=0;j<n;j++,k++) {
user->btopo[k].x = i;
user->btopo[k].y = j;
}
/* Create task array */
n = 1;
user->atomicTask = NGA_Create(C_INT, 1, &n, "Atomic Task", NULL);
if (!user->atomicTask)
GA_Error("NGA_Create failed for Atomic Task",0);
if (user->me == 0)
NGA_Put(user->atomicTask, &zero, &zero, &user->nproc, &zero);
/* space for x values from two processors */
x_space = 2 * user->BlockSize * user->ndim;
/* space for ALL gradient value */
g_space = user->natoms * user->ndim;
if (MA_push_stack(C_DBL, x_space + g_space+3, "GA LJ bufs", &user->memHandle))
MA_get_pointer(user->memHandle, &user->x1);
else
GA_Error("ma_alloc_get failed",x_space + g_space);
user->x2 = user->x1 + x_space/2 + 1;
user->grad = user->x2 + x_space/2 + 1;
GA_Sync();
return 0;
}
integer_dot(int rank, int nprocs)
{
int g_A, g_B;
int dims[DIM]={SIZE,SIZE}, val_A=5, val_B=10, op;
MA_init(C_INT, 1000, 1000);
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_B = NGA_Create(C_INT, DIM, dims, "array_B", NULL);
GA_Fill(g_A, &val_A);
GA_Fill(g_B, &val_B);
if(rank==0)
{
op=GA_Idot(g_A, g_B);
printf("%d \n", op);
}
}
verify_ga_dim(int ndim)
{
int g_A, dims[ndim], i;
for(i=0; i<ndim; i++) dims[i]=SIZE;
g_A = NGA_Create(C_INT, ndim, dims, "array_A", NULL);
if(GA_Ndim(g_A) != ndim)
printf("ERROR: GA_Ndim -- %d returned wrong \n", ndim);
GA_Destroy(g_A);
}
main(int argc, char **argv)
{
int rank, nprocs, i, j;
int g_A, **local_A=NULL, **local_B=NULL;
int dims[DIM]={SIZE,SIZE}, dims2[DIM], lo[DIM]={SIZE-SIZE,SIZE-SIZE}, hi[DIM]={SIZE-1,SIZE-1}, ld=5, value=5;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
local_A=(int**)malloc(SIZE*sizeof(int*));
for(i=0; i<SIZE; i++)
{
local_A[i]=(int*)malloc(SIZE*sizeof(int));
for(j=0; j<SIZE; j++) local_A[i][j]=rand()%10;
}
local_B=(int**)malloc(SIZE*sizeof(int*));
for(i=0; i<SIZE; i++)
{
local_B[i]=(int*)malloc(SIZE*sizeof(int));
for(j=0; j<SIZE; j++) local_B[i][j]=rand()%10;
}
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
GA_Zero(g_A);
if(rank==0)
{
NGA_Put(g_A, lo, hi, local_A, &ld);
NGA_Get(g_A, lo, hi, local_B, &ld);
for(i=0; i<SIZE; i++)
{
for(j=0; j<SIZE; j++)
if(local_A[i][j]!=local_B[i][j]) GA_ERROR_MSG();
}
}
GA_Sync();
GA_Destroy(g_A);
if(rank == 0) GA_PRINT_MSG();
GA_Terminate();
MPI_Finalize();
}
main(int argc, char **argv)
{
int rank, nprocs, i, j;
int g_A, g_B, **local_value=NULL;
int dims[DIM]={SIZE,SIZE}, lo[DIM]={SIZE-SIZE,SIZE-SIZE}, hi[DIM]={SIZE-1,SIZE-1}, ld=SIZE;
local_value=(int**)malloc(SIZE*sizeof(int*));
for(i=0; i<SIZE; i++)
local_value[i]=(int*)malloc(SIZE*sizeof(int));
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_B = NGA_Create(C_INT, DIM, dims, "array_B", NULL);
for(i=0; i<SIZE; i++)
for(j=0; j<SIZE; j++)
local_value[i][j]=rand()%10;
if(rank==0) NGA_Put(g_A, lo, hi, local_value, &ld);
GA_Transpose(g_A, g_B);
if(rank==0) validate_transpose(g_A, g_B, lo, hi, ld);
GA_Sync();
if(rank == 1) GA_PRINT_MSG();
GA_Terminate();
MPI_Finalize();
}
main(int argc, char **argv)
{
int rank, nprocs, i, j;
int g_A, g_B, g_C, local_C[DIM][DIM], dims[DIM]={5,5}, val1=5, val2=4, alpha=3, beta=2, ld=5;
int alo[DIM]={2,2}, ahi[DIM]={3,3}, blo[DIM]={2,2}, bhi[DIM]={3,3}, clo[DIM]={1,1}, chi[DIM]={2,2};
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_B = GA_Duplicate(g_A, "array_B");
g_C = GA_Duplicate(g_A, "array_C");
GA_Fill(g_A, &val1);
GA_Fill(g_B, &val2);
GA_Zero(g_C);
NGA_Add_patch(&alpha, g_A, clo, chi, &beta, g_B, blo, bhi, g_C, clo, chi);
GA_Sync();
GA_Print(g_A);
GA_Print(g_B);
GA_Print(g_C);
NGA_Get(g_C, clo, chi, local_C, &ld);
//printf("check 1 \n");
for(i=0; i<DIM; i++)
{
for(j=0; j<DIM; j++)printf("%d ", local_C[i][j]);
printf("\n");
}
if(rank == 0)
{
printf("check 2\n");
for(i=0; i<DIM; i++)
{
for(j=0; j<DIM; j++)
if(local_C[i][j]!=(alpha*val1)+(beta*val2)) printf("GA Error : \n");
}
}
if(rank==0)
GA_PRINT_MSG();
GA_Sync();
/*
GA_Destroy(g_A);
GA_Destroy(g_B);
GA_Destroy(g_C);
*/
//*******************************************************************
/* what would be the possible reason for GA_destroy to get failed ..,
* solve this before consolidate the whole
*/
GA_Terminate();
MPI_Finalize();
}
int main (int argc, char **argv)
{
double startTime;
int info; /* used to check for functions returning nonzeros */
GAVec ga_x; /* solution vector */
TAO_SOLVER tao; /* TAO_SOLVER solver context */
TAO_GA_APPLICATION taoapp; /* TAO application context */
TaoTerminateReason reason;
AppCtx user; /* user-defined application context */
/*initialize GA and MPI */
int heap = 4000, stack = 4000;
MPI_Init (&argc, &argv); /* initialize MPI */
GA_Initialize (); /* initialize GA */
if (!MA_init(MT_F_DBL, stack, heap))
GA_Error((char*)"MA_init failed", stack+heap);
/* Initialize TAO */
TaoInitialize (&argc, &argv, (char *) 0, help);
startTime = MPI_Wtime();
/* Initialize problem parameters */
user.natoms = NATOMS;
user.ndim = NDIM;
user.n = user.natoms*user.ndim;
/* Create working space */
if (MA_push_stack(C_DBL, 2*user.n, "Vector buffers", &user.memHandle) == MA_FALSE)
GA_Error((char*)"MAIN::ma_alloc_get failed",2*user.n);
/* Allocate Global Array vector for the solution */
int dims[2];
dims[0] = user.n;
ga_x = NGA_Create (C_DBL, 1, dims, (char*)"GA_X", NULL);
if (!ga_x) GA_Error ((char*)"lennard-jones.main::NGA_Create ga_x", ga_x);
/* The TAO code begins here */
/* Create TAO solver with desired solution method */
info = TaoCreate (MPI_COMM_WORLD, "tao_cg", &tao); CHKERRQ(info);
info = TaoGAApplicationCreate (MPI_COMM_WORLD, &taoapp); CHKERRQ(info);
/* Set initial vector */
info = InitializeVariables(ga_x, &user); CHKERRQ(info);
info = TaoGAAppSetInitialSolutionVec(taoapp, ga_x); CHKERRQ(info);
/* Set routines for function, gradient */
info = TaoGAAppSetObjectiveAndGradientRoutine (taoapp, FormFunctionGradient, (void *) &user);
CHKERRQ(info);
/* Check for TAO command line options */
info = TaoSetFromOptions (tao); CHKERRQ(info);
/* SOLVE THE APPLICATION */
info = TaoSolveGAApplication (taoapp, tao); CHKERRQ(info);
/* To View TAO solver information use */
info = TaoView(tao); CHKERRQ(info);
/* Get termination information */
info = TaoGetTerminationReason (tao, &reason); CHKERRQ(info);
if (reason <= 0)
printf("Try a different TAO method, adjust some parameters, or check the function evaluation routines\n");
printf("TIME TAKEN = %lf\n", MPI_Wtime()-startTime);
/*output the solutions */
printf ("The solution is :\n");
GA_Print (ga_x);
/* Free TAO data structures */
info = TaoDestroy (tao); CHKERRQ(info);
info = TaoGAAppDestroy (taoapp); CHKERRQ(info);
/* Free GA data structures */
GA_Destroy (ga_x);
if (!MA_pop_stack(user.memHandle))
GA_Error((char*)"Main::MA_pop_stack failed",0);
/* Finalize TAO, GA, and MPI */
TaoFinalize ();
GA_Terminate ();
MPI_Finalize ();
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;
}
main(int argc, char **argv)
{
int rank, nprocs;
int g_A, dims[D]={5,10}, local_A[N], local_G[N], **sub_array=NULL, **s_array=NULL;
int i, j, value=5;
MPI_Init(&argc, &argv);
GA_Initialize();
MA_init(C_INT, 1000, 1000);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
s_array=(int**)malloc(N*sizeof(int*));
for(i=0; i<N; i++)
{
s_array[i]=(int*)malloc(D*sizeof(int));
for(j=0; j<D; j++) s_array[i][j]=rand()%5;
}
sub_array=(int**)malloc(N*sizeof(int*));
for(i=0; i<N; i++)
{
sub_array[i]=(int*)malloc(D*sizeof(int));
for(j=0; j<D; j++) sub_array[i][j]=rand()%5;
}
for(i=0; i<N; i++)
//local_A=(int*)malloc(N*sizeof(int));
/*
* depends on the value of array ..we can generate the location values in randon
* we can also use the if-condition
*/
// PRINTing all the genrated array for reference
for(i=0; i<N; i++)
{
for(j=0; j<D; j++)printf("%d ",s_array[i][j]);
printf("\n");
}
printf("\n");
for(i=0; i<N; i++)
{
for(j=0; j<D; j++)printf("%d ",sub_array[i][j]);
printf("\n");
}
printf("\n");
for(i=0; i<N; i++)printf("%d \n",local_A[i]=rand()%5+1);
// PRINT done - now creating array
g_A=NGA_Create(C_INT, D, dims, "array_A", NULL);
GA_Fill(g_A, &value);
GA_Sync();
NGA_Scatter(g_A, local_A, s_array, N);
NGA_Gather(g_A, local_G, s_array, N);
GA_Sync();
GA_Print(g_A);
for(i=0; i<N; i++)printf("%d \n",local_G[i]);
printf("\n");
if(rank==0)
{
for(i=0; i<N; i++)
if(local_G[i]!=local_A[i]) printf("GA Error: \n");
}
GA_Sync();
if(rank==0)
GA_PRINT_MSG();
GA_Terminate();
MPI_Finalize();
return 0;
}
void do_work()
{
int ZERO=0; /* useful constants */
int g_a, g_b;
int n=N, ndim=2,type=MT_F_DBL,dims[2]={N,N},coord[2];
int me=GA_Nodeid(), nproc=GA_Nnodes();
int row, i, j;
int lo[2], hi[2];
/* Note: on all current platforms DoublePrecision = double */
DoublePrecision buf[N], *max_row=NULL;
MPI_Comm WORLD_COMM;
MPI_Comm ROW_COMM;
int ilo,ihi, jlo,jhi, ld, prow, pcol;
int root=0, grp_me=-1;
WORLD_COMM = GA_MPI_Comm_pgroup_default();
if(me==0)printf("Creating matrix A\n");
dims[0]=n; dims[1]=n;
g_a = NGA_Create(type, ndim, dims, "A", NULL);
if(!g_a) GA_Error("create failed: A",n);
if(me==0)printf("OK\n");
if(me==0)printf("Creating matrix B\n");
dims[0]=n;
g_b = NGA_Create(type, 1, dims, "B", NULL);
if(!g_b) GA_Error("create failed: B",n);
if(me==0)printf("OK\n");
GA_Zero(g_a); /* zero the matrix */
if(me==0)printf("Initializing matrix A\n");
/* fill in matrix A with values: A(i,j) = (i+j) */
for(row=me; row<n; row+= nproc){
/**
* simple load balancing:
* each process works on a different row in MIMD style
*/
for(i=0; i<n; i++) buf[i]=(DoublePrecision)(i+row+1);
lo[0]=hi[0]=row;
lo[1]=ZERO; hi[1]=n-1;
NGA_Put(g_a, lo, hi, buf, &n);
}
/* GA_print(&g_a);*/
NGA_Distribution(g_a, me, lo, hi);
ilo=lo[0]; ihi=hi[0];
jlo=lo[1]; jhi=hi[1];
GA_Sync();
if(ihi-ilo+1 >0){
max_row=(DoublePrecision*)malloc(sizeof(DoublePrecision)*(ihi-ilo+1));
if (!max_row) GA_Error("malloc 3 failed",(ihi-ilo+1));
for (i=0; i<(ihi-ilo+1); i++) {
max_row[i] = 0.0;
}
}
NGA_Proc_topology(g_a, me, coord); /* block coordinates */
prow = coord[0];
pcol = coord[1];
if(me==0)printf("Splitting comm according to distribution of A\n");
/* GA on SP1 requires synchronization before & after message-passing !!*/
GA_Sync();
if(me==0)printf("Computing max row elements\n");
/* create communicator for processes that 'own' A[:,jlo:jhi] */
MPI_Barrier(WORLD_COMM);
if(pcol < 0 || prow <0)
MPI_Comm_split(WORLD_COMM,MPI_UNDEFINED,MPI_UNDEFINED, &ROW_COMM);
else
MPI_Comm_split(WORLD_COMM, (int)pcol, (int)prow, &ROW_COMM);
if(ROW_COMM != MPI_COMM_NULL){
double *ptr;
MPI_Comm_rank(ROW_COMM, &grp_me);
/* each process computes max elements in the block it 'owns' */
lo[0]=ilo; hi[0]=ihi;
lo[1]=jlo; hi[1]=jhi;
NGA_Access(g_a, lo, hi, &ptr, &ld);
for(i=0; i<ihi-ilo+1; i++){
for(j=0; j<jhi-jlo+1; j++)
if(max_row[i] < ptr[i*ld + j]){
max_row[i] = ptr[i*ld + j];
}
}
MPI_Reduce(max_row, buf, ihi-ilo+1, MPI_DOUBLE, MPI_MAX,
root, ROW_COMM);
}else fprintf(stderr,"process %d not participating\n",me);
GA_Sync();
/* processes with rank=root in ROW_COMM put results into g_b */
ld = 1;
if(grp_me == root) {
lo[0]=ilo; hi[0]=ihi;
NGA_Put(g_b, lo, hi, buf, &ld);
}
GA_Sync();
if(me==0)printf("Checking the result\n");
if(me==0){
lo[0]=ZERO; hi[0]=n-1;
NGA_Get(g_b, lo, hi, buf, &n);
for(i=0; i< n; i++)if(buf[i] != (double)n+i){
fprintf(stderr,"error:%d max=%f should be:%d\n",i,buf[i],n+i);
GA_Error("terminating...",1);
}
}
if(me==0)printf("OK\n");
GA_Destroy(g_a);
GA_Destroy(g_b);
}
/* input is matrix size */
void ga_lu(double *A, int matrix_size)
{
int g_a, g_b, dims[2], type=C_DBL;
int lo[2], hi[2], ld;
int block_size[2], proc_grid[2];
double time, gflops;
/* create a 2-d GA (global matrix) */
dims[0] = matrix_size;
dims[1] = matrix_size;
block_size[0] = BLOCK_SIZE;
block_size[1] = BLOCK_SIZE;
#ifdef USE_SCALAPACK_DISTR
proc_grid[0] = 2;
proc_grid[1] = nprocs/2;
if(nprocs%2) GA_Error("For ScaLAPACK stle distribution, nprocs must be "
" divisible by 2", 0);
#endif
#ifndef BLOCK_CYCLIC
g_a = NGA_Create(type, 2, dims, "A", NULL);
g_b = GA_Duplicate(g_a, "transposed array B");
#else
g_a = GA_Create_handle();
GA_Set_data(g_a, 2, dims, type);
GA_Set_array_name(g_a,"A");
# ifdef USE_SCALAPACK_DISTR
GA_Set_block_cyclic_proc_grid(g_a, block_size, proc_grid);
# else
GA_Set_block_cyclic(g_a, block_size);
# endif
GA_Allocate(g_a);
g_b = GA_Create_handle();
GA_Set_data(g_b, 2, dims, type);
GA_Set_array_name(g_b,"B");
# ifdef USE_SCALAPACK_DISTR
GA_Set_block_cyclic_proc_grid(g_b, block_size, proc_grid);
# else
GA_Set_block_cyclic(g_b, block_size);
# endif
GA_Allocate(g_b);
#endif
/* copy the local matrix into GA */
if(me==0)
{
lo[0] = 0;
hi[0] = matrix_size - 1;
lo[1] = 0;
hi[1] = matrix_size - 1;
ld = matrix_size;
NGA_Put(g_a, lo, hi, A, &ld);
}
GA_Sync();
GA_Transpose(g_a, g_b);
time = CLOCK_();
GA_Lu('n', g_b);
time = CLOCK_() - time;
/* 2/3 N^3 - 1/2 N^2 flops for LU and 2*N^2 for solver */
gflops = ( (((double)matrix_size) * matrix_size)/(time*1.0e+9) *
(2.0/3.0 * (double)matrix_size - 0.5) );
if(me==0) printf("\nGA_Lu: N=%d flops=%2.5e Gflops, time=%2.5e secs\n\n",
matrix_size, gflops, time);
#if DEBUG
GA_Print(g_a);
GA_Print(g_b);
#endif
/* if(me==0) lu(A, matrix_size); */
GA_Destroy(g_a);
GA_Destroy(g_b);
}
int main (int argc, char **argv)
{
double startTime;
int info; /* used to check for functions returning nonzeros */
GAVec ga_x; /* solution vector */
TAO_SOLVER tao; /* TAO_SOLVER solver context */
TAO_GA_APPLICATION taoapp; /* TAO application context */
TaoTerminateReason reason;
AppCtx user; /* user-defined application context */
/*initialize GA and MPI */
int heap = 400000, stack = 400000;
MPI_Init (&argc, &argv); /* initialize MPI */
GA_Initialize (); /* initialize GA */
user.me = GA_Nodeid ();
user.nproc = GA_Nnodes ();
startTime = MPI_Wtime();
if (user.me == 0) {
if (GA_Uses_fapi ())
GA_Error ("Program runs with C array API only", 0);
printf ("Using %ld processes\n", (long) user.nproc);
fflush (stdout);
}
heap /= user.nproc;
stack /= user.nproc;
if (!MA_init (MT_F_DBL, stack, heap))
GA_Error ("MA_init failed", stack + heap); /* initialize memory allocator */
/* Initialize TAO */
TaoInitialize (&argc, &argv, (char *) 0, help);
/* Initialize problem parameters */
user.ndim = NDIM;
user.natoms = NATOMS;
user.BlockSize = BLOCKSIZE;
/* Allocate vectors for the solution and gradient */
int dims[2];
dims[0] = user.ndim*user.natoms;
ga_x = NGA_Create (C_DBL, 1, dims, "GA_X", NULL);
if (!ga_x) GA_Error ("lennard-jones.main::NGA_Create ga_x", ga_x);
/* Set up structures for data distribution */
info = SetupBlocks(&user); CHKERRQ(info);
/* The TAO code begins here */
/* Create TAO solver with desired solution method */
info = TaoCreate (MPI_COMM_WORLD, "tao_lmvm", &tao); CHKERRQ(info);
info = TaoGAApplicationCreate (MPI_COMM_WORLD, &taoapp); CHKERRQ(info);
/* Set the initial solution */
info = InitializeVariables(ga_x, &user); CHKERRQ(info);
info = TaoGAAppSetInitialSolutionVec(taoapp, ga_x); CHKERRQ(info);
/* Set routines for function, gradient */
info = TaoGAAppSetObjectiveAndGradientRoutine (taoapp, FormFunctionGradient,
(void *) &user); CHKERRQ(info);
/* Check for TAO command line options */
info = TaoSetFromOptions (tao); CHKERRQ(info);
/* SOLVE THE APPLICATION */
info = TaoSolveGAApplication (taoapp, tao); CHKERRQ(info);
/* To View TAO solver information use */
info = TaoView(tao); CHKERRQ(info);
/* Get termination information */
info = TaoGetTerminationReason (tao, &reason);
if(info) GA_Error("lennard-jones.main.TaoGetTerminationReason",info);
if (user.me == 0) {
if (reason <= 0)
printf("Try a different TAO method, adjust some parameters, or check the function evaluation routines\n");
printf("WALL TIME TAKEN = %lf\n", MPI_Wtime()-startTime);
/*output the solutions */
printf ("The solution is :\n");
}
GA_Print (ga_x);
/* Free TAO data structures */
info = TaoDestroy (tao); CHKERRQ(info);
info = TaoGAAppDestroy (taoapp); CHKERRQ(info);
/* Free GA data structures */
GA_Destroy (ga_x);
if (!MA_pop_stack(user.memHandle))
ga_error("Main::MA_pop_stack for memHandle failed",0);
/* Finalize TAO, GA, and MPI */
TaoFinalize ();
GA_Terminate ();
MPI_Finalize ();
return 0;
}
int main( int argc, char **argv ) {
int g_a, g_b, i, j, size, size_me;
int icnt, idx, jdx, ld;
int n=N, type=MT_C_INT, one;
int *values, *ptr;
int **indices;
int dims[2]={N,N};
int lo[2], hi[2];
int heap=3000000, stack=2000000;
int me, nproc;
int datatype, elements;
double *prealloc_mem;
MP_INIT(argc,argv);
#if 1
GA_INIT(argc,argv); /* initialize GA */
me=GA_Nodeid();
nproc=GA_Nnodes();
if(me==0) {
if(GA_Uses_fapi())GA_Error("Program runs with C array API only",1);
printf("\nUsing %ld processes\n",(long)nproc);
fflush(stdout);
}
heap /= nproc;
stack /= nproc;
if(! MA_init(MT_F_DBL, stack, heap))
GA_Error("MA_init failed",stack+heap); /* initialize memory allocator*/
/* Create a regular matrix. */
if(me==0)printf("\nCreating matrix A of size %d x %d\n",N,N);
g_a = NGA_Create(type, 2, dims, "A", NULL);
if(!g_a) GA_Error("create failed: A",n);
/* Fill matrix using scatter routines */
size = N*N;
if (size%nproc == 0) {
size_me = size/nproc;
} else {
i = size - size%nproc;
size_me = i/nproc;
if (me < size%nproc) size_me++;
}
/* Check that sizes are all okay */
i = size_me;
GA_Igop(&i,1,"+");
if (i != size) {
GA_Error("Sizes don't add up correctly: ",i);
} else if (me==0) {
printf("\nSizes add up correctly\n");
}
/* Allocate index and value arrays */
indices = (int**)malloc(size_me*sizeof(int*));
values = (int*)malloc(size_me*sizeof(int));
icnt = me;
for (i=0; i<size_me; i++) {
values[i] = icnt;
idx = icnt%N;
jdx = (icnt-idx)/N;
if (idx >= N || idx < 0) {
printf("p[%d] Bogus index i: %d\n",me,idx);
}
if (jdx >= N || jdx < 0) {
printf("p[%d] Bogus index j: %d\n",me,jdx);
}
indices[i] = (int*)malloc(2*sizeof(int));
(indices[i])[0] = idx;
(indices[i])[1] = jdx;
icnt += nproc;
}
/* Scatter values into g_a */
NGA_Scatter(g_a, values, indices, size_me);
GA_Sync();
/* Check to see if contents of g_a are correct */
NGA_Distribution( g_a, me, lo, hi );
NGA_Access(g_a, lo, hi, &ptr, &ld);
for (i=lo[0]; i<hi[0]; i++) {
idx = i-lo[0];
for (j=lo[1]; j<hi[1]; j++) {
jdx = j-lo[1];
if (ptr[idx*ld+jdx] != j*N+i) {
printf("p[%d] (Scatter) expected: %d actual: %d\n",me,j*N+i,ptr[idx*ld+jdx]);
}
}
}
if (me==0) printf("\nCompleted test of NGA_Scatter\n");
for (i=0; i<size_me; i++) {
values[i] = 0;
}
GA_Sync();
NGA_Gather(g_a, values, indices, size_me);
icnt = me;
for (i=0; i<size_me; i++) {
if (icnt != values[i]) {
printf("p[%d] (Gather) expected: %d actual: %d\n",me,icnt,values[i]);
}
icnt += nproc;
}
if (me==0) printf("\nCompleted test of NGA_Gather\n");
GA_Sync();
/* Scatter-accumulate values back into GA*/
one = 1;
NGA_Scatter_acc(g_a, values, indices, size_me, &one);
GA_Sync();
/* Check to see if contents of g_a are correct */
for (i=lo[0]; i<hi[0]; i++) {
idx = i-lo[0];
for (j=lo[1]; j<hi[1]; j++) {
jdx = j-lo[1];
if (ptr[idx*ld+jdx] != 2*(j*N+i)) {
printf("p[%d] (Scatter_acc) expected: %d actual: %d\n",me,2*(j*N+i),ptr[idx*ld+jdx]);
}
}
}
if (me==0) printf("\nCompleted test of NGA_Scatter_acc\n");
NGA_Release(g_a, lo, hi);
/* Test fixed buffer size */
NGA_Alloc_gatscat_buf(size_me);
/* Scatter-accumulate values back into GA*/
GA_Sync();
NGA_Scatter_acc(g_a, values, indices, size_me, &one);
GA_Sync();
/* Check to see if contents of g_a are correct */
for (i=lo[0]; i<hi[0]; i++) {
idx = i-lo[0];
for (j=lo[1]; j<hi[1]; j++) {
jdx = j-lo[1];
if (ptr[idx*ld+jdx] != 3*(j*N+i)) {
printf("p[%d] (Scatter_acc) expected: %d actual: %d\n",me,3*(j*N+i),ptr[idx*ld+jdx]);
}
}
}
if (me==0) printf("\nCompleted test of NGA_Scatter_acc using fixed buffers\n");
NGA_Release(g_a, lo, hi);
NGA_Free_gatscat_buf();
GA_Destroy(g_a);
if(me==0)printf("\nSuccess\n");
GA_Terminate();
#endif
MP_FINALIZE();
return 0;
}
void do_work()
{
int ONE=1 ; /* useful constants */
int g_a, g_b;
int n=N, type=MT_F_DBL;
int me=GA_Nodeid(), nproc=GA_Nnodes();
int i, row;
int dims[2]={N,N};
int lo[2], hi[2], ld;
/* Note: on all current platforms DoublePrecision == double */
double buf[N], err, alpha, beta;
if(me==0)printf("Creating matrix A\n");
g_a = NGA_Create(type, 2, dims, "A", NULL);
if(!g_a) GA_Error("create failed: A",n);
if(me==0)printf("OK\n");
if(me==0)printf("Creating matrix B\n");
/* create matrix B so that it has dims and distribution of A*/
g_b = GA_Duplicate(g_a, "B");
if(! g_b) GA_Error("duplicate failed",n);
if(me==0)printf("OK\n");
GA_Zero(g_a); /* zero the matrix */
if(me==0)printf("Initializing matrix A\n");
/* fill in matrix A with random values in range 0.. 1 */
lo[1]=0; hi[1]=n-1;
for(row=me; row<n; row+= nproc){
/* each process works on a different row in MIMD style */
lo[0]=hi[0]=row;
for(i=0; i<n; i++) buf[i]=sin((double)i + 0.1*(row+1));
NGA_Put(g_a, lo, hi, buf, &n);
}
if(me==0)printf("Symmetrizing matrix A\n");
GA_Symmetrize(g_a); /* symmetrize the matrix A = 0.5*(A+A') */
/* check if A is symmetric */
if(me==0)printf("Checking if matrix A is symmetric\n");
GA_Transpose(g_a, g_b); /* B=A' */
alpha=1.; beta=-1.;
GA_Add(&alpha, g_a, &beta, g_b, g_b); /* B= A - B */
err= GA_Ddot(g_b, g_b);
if(me==0)printf("Error=%f\n",(double)err);
if(me==0)printf("\nChecking atomic accumulate \n");
GA_Zero(g_a); /* zero the matrix */
for(i=0; i<n; i++) buf[i]=(double)i;
/* everybody accumulates to the same location/row */
alpha = 1.0;
row = n/2;
lo[0]=hi[0]=row;
lo[1]=0; hi[1]=n-1;
ld = hi[1]-lo[1]+1;
NGA_Acc(g_a, lo, hi, buf, &ld, &alpha );
GA_Sync();
if(me==0){ /* node 0 is checking the result */
NGA_Get(g_a, lo, hi, buf,&ld);
for(i=0; i<n; i++) if(buf[i] != (double)nproc*i)
GA_Error("failed: column=",i);
printf("OK\n\n");
}
GA_Destroy(g_a);
GA_Destroy(g_b);
}
// 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 rank, nprocs;
int g_A;
int *local_A=NULL, *local_B=NULL, *output_A=NULL;
int dims[DIM]={SIZE,SIZE}, dims2[DIM], lo[DIM]={SIZE-SIZE,SIZE-SIZE}, hi[DIM]={SIZE-1,SIZE-1}, ld=SIZE;
int value=SIZE;
#if defined(USE_ELEMENTAL)
// initialize Elemental (which will initialize MPI)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &rank );
ElMPICommSize( MPI_COMM_WORLD, &nprocs );
// instantiate el::global array
ElGlobalArraysConstruct_i( &eliga );
// initialize global arrays
ElGlobalArraysInitialize_i( eliga );
#else
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
#endif
local_A=(int*)malloc(SIZE*SIZE*sizeof(int));
output_A=(int*)malloc(SIZE*SIZE*sizeof(int));
memset (output_A, 0, SIZE*SIZE*sizeof(int));
for(int j=0; j<SIZE; j++)
for(int i=0; i<SIZE; i++) local_A[i+j*ld]=(i + j);
//for(int i=0; i<SIZE; i++) local_A[i+j*ld]=(rand()%10);
local_B=(int*)malloc(SIZE*SIZE*sizeof(int));
memset (local_B, 0, SIZE*SIZE*sizeof(int));
// nb handle
#if defined(USE_ELEMENTAL)
typedef ElInt ga_nbhdl_t;
#endif
ga_nbhdl_t nbnb;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_i( eliga, DIM, dims, "array_A", NULL, &g_A );
ElGlobalArraysFill_i( eliga, g_A, &value );
#else
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
GA_Fill(g_A, &value);
#endif
if (rank == 0) printf ("Initial global array:\n");
#if defined(USE_ELEMENTAL)
ElGlobalArraysPrint_i( eliga, g_A );
#else
GA_Print(g_A);
#endif
for (int i = 0; i < NITERS; i++)
{
// acc data
#if defined(USE_ELEMENTAL)
ElGlobalArraysNBAccumulate_i( eliga, g_A, lo, hi, local_A, &ld, &value, &nbnb );
#else
NGA_NbAcc(g_A, lo, hi, local_A, &ld, &value, &nbnb);
#endif
// updated output
MPI_Reduce (local_A, output_A, SIZE*SIZE, MPI_INT, MPI_SUM, 0, MPI_COMM_WORLD);
#if defined(USE_ELEMENTAL)
ElGlobalArraysNBWait_i( eliga, &nbnb );
#else
NGA_NbWait (&nbnb);
#endif
// get
if (rank == 0) printf ("Get in iter #%d\n", i);
#if defined(USE_ELEMENTAL)
ElGlobalArraysSync_i( eliga );
ElGlobalArraysGet_i( eliga, g_A, lo, hi, local_B, &ld );
ElGlobalArraysPrint_i( eliga, g_A );
#else
GA_Sync();
NGA_Get(g_A, lo, hi, local_B, &ld);
GA_Print(g_A);
#endif
} // end of iters
if(rank==0)
{
printf(" Alpha (multiplier): %d\n", value);
printf(" Original local buffer (before accumulation): \n");
for(int i=0; i<SIZE; i++)
{
for(int j=0; j<SIZE; j++)
printf("%d ", local_A[i*ld+j]);
printf("\n");
}
printf("\n");
printf(" Get returns: \n");
for(int i=0; i<SIZE; i++)
{
for(int j=0; j<SIZE; j++)
printf("%d ", local_B[i*ld + j]);
printf("\n");
}
printf("\n");
for(int i=0; i<SIZE; i++)
{
for(int j=0; j<SIZE; j++)
{
if(local_B[i*ld+j]!=(value + (NITERS * value * (output_A[i*ld+j]))))
GA_Error("ERROR", -99);
}
}
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysDestroy_i( eliga, g_A );
#else
GA_Destroy(g_A);
#endif
if(rank == 0)
printf ("OK. Test passed\n");
free (local_A);
free (local_B);
free (output_A);
#if defined(USE_ELEMENTAL)
ElGlobalArraysTerminate_i( eliga );
// call el::global arrays destructor
ElGlobalArraysDestruct_i( eliga );
ElFinalize();
#else
GA_Terminate();
MPI_Finalize();
#endif
}
void test_io_dbl()
{
int n, ndim = NDIM;
double err, tt0, tt1, mbytes;
int g_a, g_b, d_a;
int i, itmp, j, req, loop;
int glo[MAXDIM],ghi[MAXDIM];
dra_size_t dlo[MAXDIM],dhi[MAXDIM];
dra_size_t ddims[MAXDIM],reqdims[MAXDIM];
dra_size_t m;
int index[MAXDIM], dims[MAXDIM];
int me, nproc, isize;
double *ptr;
double plus, minus;
int ld[MAXDIM], chunk[MAXDIM];
char filename[80];
FILE *fd;
n = SIZE;
m = ((dra_size_t)NFACTOR)*((dra_size_t)SIZE);
loop = 1;
for (i=0; i<ndim; i++) loop *= NFACTOR;
req = -1;
nproc = GA_Nnodes();
me = GA_Nodeid();
if (me == 0) {
printf("Creating temporary global arrays %d",n);
for (i=1; i<ndim; i++) {
printf(" x %d",n);
}
printf("\n");
}
if (me == 0) fflush(stdout);
GA_Sync();
for (i=0; i<ndim; i++) {
dims[i] = n;
chunk[i] = 1;
}
g_a = NGA_Create(MT_DBL, ndim, dims, "a", chunk);
if (!g_a) GA_Error("NGA_Create failed: a", 0);
g_b = NGA_Create(MT_DBL, ndim, dims, "b", chunk);
if (!g_b) GA_Error("NGA_Create failed: b", 0);
if (me == 0) printf("done\n");
if (me == 0) fflush(stdout);
/* initialize g_a, g_b with random values
... use ga_access to avoid allocating local buffers for ga_put */
GA_Sync();
NGA_Distribution(g_a, me, glo, ghi);
NGA_Access(g_a, glo, ghi, &ptr, ld);
isize = 1;
for (i=0; i<ndim; i++) isize *= (ghi[i]-glo[i]+1);
fill_random(ptr, isize);
GA_Sync();
GA_Zero(g_b);
/*.......................................................................*/
if (me == 0) {
printf("Creating Disk array %ld",m);
for (i=1; i<ndim; i++) {
printf(" x %ld",m);
}
printf("\n");
}
if (me == 0) fflush(stdout);
for (i=0; i<ndim; i++) {
ddims[i] = m;
reqdims[i] = (dra_size_t)n;
}
GA_Sync();
strcpy(filename,FNAME);
if (! (fd = fopen(filename, "w"))) {
strcpy(filename,FNAME_ALT);
if (! (fd = fopen(filename, "w"))) {
GA_Error("open failed",0);
}
}
fclose(fd);
if (NDRA_Create(MT_DBL, ndim, ddims, "A", filename, DRA_RW,
reqdims, &d_a) != 0) {
GA_Error("NDRA_Create failed(d_a): ",0);
}
if (me == 0) printf("testing write\n");
fflush(stdout);
tt1 = 0.0;
for (i=0; i<loop; i++) {
itmp=i;
for (j=0; j<ndim; j++) {
index[j] = itmp%NFACTOR;
itmp = (itmp - index[j])/NFACTOR;
}
for (j=0; j<ndim; j++) {
glo[j] = 0;
ghi[j] = SIZE - 1;
dlo[j] = ((dra_size_t)index[j])*((dra_size_t)SIZE);
dhi[j] = (((dra_size_t)index[j])+(dra_size_t)1)
* ((dra_size_t)SIZE) - (dra_size_t)1;
}
tt0 = MP_TIMER();
if (NDRA_Write_section(FALSE, g_a, glo, ghi,
d_a, dlo, dhi, &req) != 0) {
GA_Error("ndra_write_section failed:",0);
}
if (DRA_Wait(req) != 0) {
GA_Error("DRA_Wait failed(d_a): ",req);
}
tt1 += (MP_TIMER() - tt0);
}
GA_Dgop(&tt1,1,"+");
tt1 = tt1/((double)nproc);
mbytes = 1.e-6 * (double)(pow(m,ndim)*sizeof(double));
if (me == 0) {
printf("%11.2f MB time = %11.2f rate = %11.3f MB/s\n",
mbytes,tt1,mbytes/tt1);
}
if (DRA_Close(d_a) != 0) {
GA_Error("DRA_Close failed(d_a): ",d_a);
}
if (me == 0) printf("\n");
if (me == 0) printf("disk array closed\n");
if (me == 0) fflush(stdout);
/*..........................................................*/
if (me == 0) printf("\n");
if (me == 0) printf("opening disk array\n");
if (DRA_Open(filename, DRA_R, &d_a) != 0) {
GA_Error("DRA_Open failed",0);
}
if (me == 0) printf("testing read\n");
/* printf("testing read on proc %d\n",me); */
if (me == 0) fflush(stdout);
tt1 = 0.0;
for (i=0; i<loop; i++) {
itmp=i;
for (j=0; j<ndim; j++) {
index[j] = itmp%NFACTOR;
itmp = (itmp - index[j])/NFACTOR;
}
for (j=0; j<ndim; j++) {
glo[j] = 0;
ghi[j] = SIZE - 1;
dlo[j] = ((dra_size_t)index[j])*((dra_size_t)SIZE);
dhi[j] = (((dra_size_t)index[j])+(dra_size_t)1)
* ((dra_size_t)SIZE) - (dra_size_t)1;
}
tt0 = MP_TIMER();
if (NDRA_Read_section(FALSE, g_b, glo, ghi,
d_a, dlo, dhi, &req) != 0) {
GA_Error("ndra_read_section failed:",0);
}
if (DRA_Wait(req) != 0) {
GA_Error("DRA_Wait failed(d_a): ",req);
}
tt1 += (MP_TIMER() - tt0);
plus = 1.0;
minus = -1.0;
GA_Add(&plus, g_a, &minus, g_b, g_b);
err = GA_Ddot(g_b, g_b);
if (err != 0) {
if (me == 0) {
printf("BTW, we have error = %f on loop value %d\n",
err,i);
}
GA_Error(" bye",0);
}
}
GA_Dgop(&tt1,1,"+");
tt1 = tt1/((double)nproc);
if (me == 0) {
printf("%11.2f MB time = %11.2f rate = %11.3f MB/s\n",
mbytes,tt1,mbytes/tt1);
}
if (DRA_Delete(d_a) != 0) GA_Error("DRA_Delete failed",0);
/*.......................................................................*/
GA_Destroy(g_a);
GA_Destroy(g_b);
}
void
test(int data_type) {
int me=GA_Nodeid();
int nproc = GA_Nnodes();
int g_a, g_b, g_c;
int ndim = 2;
int dims[2]={N,N};
int lo[2]={0,0};
int hi[2]={N-1,N-1};
int block_size[2]={NB,NB-1};
int proc_grid[2];
int i,j,l,k,m,n, ld;
double alpha_dbl = 1.0, beta_dbl = 0.0;
double dzero = 0.0;
double ddiff;
float alpha_flt = 1.0, beta_flt = 0.0;
float fzero = 0.0;
float fdiff;
float ftmp;
double dtmp;
SingleComplex ctmp;
DoubleComplex ztmp;
DoubleComplex alpha_dcpl = {1.0, 0.0} , beta_dcpl = {0.0, 0.0};
DoubleComplex zzero = {0.0,0.0};
DoubleComplex zdiff;
SingleComplex alpha_scpl = {1.0, 0.0} , beta_scpl = {0.0, 0.0};
SingleComplex czero = {0.0,0.0};
SingleComplex cdiff;
void *alpha=NULL, *beta=NULL;
void *abuf=NULL, *bbuf=NULL, *cbuf=NULL, *c_ptr=NULL;
switch (data_type) {
case C_FLOAT:
alpha = (void *)&alpha_flt;
beta = (void *)&beta_flt;
abuf = (void*)malloc(N*N*sizeof(float));
bbuf = (void*)malloc(N*N*sizeof(float));
cbuf = (void*)malloc(N*N*sizeof(float));
if(me==0) printf("Single Precision: Testing GA_Sgemm,NGA_Matmul_patch for %d-Dimension", ndim);
break;
case C_DBL:
alpha = (void *)&alpha_dbl;
beta = (void *)&beta_dbl;
abuf = (void*)malloc(N*N*sizeof(double));
bbuf = (void*)malloc(N*N*sizeof(double));
cbuf = (void*)malloc(N*N*sizeof(double));
if(me==0) printf("Double Precision: Testing GA_Dgemm,NGA_Matmul_patch for %d-Dimension", ndim);
break;
case C_DCPL:
alpha = (void *)&alpha_dcpl;
beta = (void *)&beta_dcpl;
abuf = (void*)malloc(N*N*sizeof(DoubleComplex));
bbuf = (void*)malloc(N*N*sizeof(DoubleComplex));
cbuf = (void*)malloc(N*N*sizeof(DoubleComplex));
if(me==0) printf("Double Complex: Testing GA_Zgemm,NGA_Matmul_patch for %d-Dimension", ndim);
break;
case C_SCPL:
alpha = (void *)&alpha_scpl;
beta = (void *)&beta_scpl;
abuf = (void*)malloc(N*N*sizeof(SingleComplex));
bbuf = (void*)malloc(N*N*sizeof(SingleComplex));
cbuf = (void*)malloc(N*N*sizeof(SingleComplex));
if(me==0) printf("Single Complex: Testing GA_Cgemm,NGA_Matmul_patch for %d-Dimension", ndim);
break;
default:
GA_Error("wrong data type", data_type);
}
if (me==0) printf("\nCreate A, B, C\n");
#ifdef USE_REGULAR
g_a = NGA_Create(data_type, ndim, dims, "array A", NULL);
#endif
#ifdef USE_SIMPLE_CYCLIC
g_a = NGA_Create_handle();
NGA_Set_data(g_a,ndim,dims,data_type);
NGA_Set_array_name(g_a,"array A");
NGA_Set_block_cyclic(g_a,block_size);
if (!GA_Allocate(g_a)) {
GA_Error("Failed: create: g_a",40);
}
#endif
#ifdef USE_SCALAPACK
g_a = NGA_Create_handle();
NGA_Set_data(g_a,ndim,dims,data_type);
NGA_Set_array_name(g_a,"array A");
grid_factor(nproc,&i,&j);
proc_grid[0] = i;
proc_grid[1] = j;
NGA_Set_block_cyclic_proc_grid(g_a,block_size,proc_grid);
if (!GA_Allocate(g_a)) {
GA_Error("Failed: create: g_a",40);
}
#endif
#ifdef USE_TILED
g_a = NGA_Create_handle();
NGA_Set_data(g_a,ndim,dims,data_type);
NGA_Set_array_name(g_a,"array A");
grid_factor(nproc,&i,&j);
proc_grid[0] = i;
proc_grid[1] = j;
NGA_Set_tiled_proc_grid(g_a,block_size,proc_grid);
if (!GA_Allocate(g_a)) {
GA_Error("Failed: create: g_a",40);
}
#endif
g_b = GA_Duplicate(g_a, "array B");
g_c = GA_Duplicate(g_a, "array C");
if(!g_a || !g_b || !g_c) GA_Error("Create failed: a, b or c",1);
ld = N;
if (me==0) printf("\nInitialize A\n");
/* Set up matrix A */
if (me == 0) {
for (i=0; i<N; i++) {
for (j=0; j<N; j++) {
switch (data_type) {
case C_FLOAT:
((float*)abuf)[i*N+j] = (float)(i*N+j);
break;
case C_DBL:
((double*)abuf)[i*N+j] = (double)(i*N+j);
break;
case C_DCPL:
((DoubleComplex*)abuf)[i*N+j].real = (double)(i*N+j);
((DoubleComplex*)abuf)[i*N+j].imag = 1.0;
break;
case C_SCPL:
((SingleComplex*)abuf)[i*N+j].real = (float)(i*N+j);
((SingleComplex*)abuf)[i*N+j].imag = 1.0;
break;
default:
GA_Error("wrong data type", data_type);
}
}
}
NGA_Put(g_a,lo,hi,abuf,&ld);
}
GA_Sync();
if (me==0) printf("\nInitialize B\n");
/* Set up matrix B */
if (me == 0) {
for (i=0; i<N; i++) {
for (j=0; j<N; j++) {
switch (data_type) {
case C_FLOAT:
((float*)bbuf)[i*N+j] = (float)(j*N+i);
break;
case C_DBL:
((double*)bbuf)[i*N+j] = (double)(j*N+i);
break;
case C_DCPL:
((DoubleComplex*)bbuf)[i*N+j].real = (double)(j*N+i);
((DoubleComplex*)bbuf)[i*N+j].imag = 1.0;
break;
case C_SCPL:
((SingleComplex*)bbuf)[i*N+j].real = (float)(j*N+i);
((SingleComplex*)bbuf)[i*N+j].imag = 1.0;
break;
default:
GA_Error("wrong data type", data_type);
}
}
}
NGA_Put(g_b,lo,hi,bbuf,&ld);
}
GA_Sync();
if (me==0) printf("\nPerform matrix multiply\n");
switch (data_type) {
case C_FLOAT:
NGA_Matmul_patch('N','N',&alpha_flt,&beta_flt,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
case C_DBL:
NGA_Matmul_patch('N','N',&alpha_dbl,&beta_dbl,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
case C_SCPL:
NGA_Matmul_patch('N','N',&alpha_scpl,&beta_scpl,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
case C_DCPL:
NGA_Matmul_patch('N','N',&alpha_dcpl,&beta_dcpl,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
default:
GA_Error("wrong data type", data_type);
}
GA_Sync();
#if 0
if (me==0) printf("\nCheck answer\n");
/*
GA_Print(g_a);
if (me == 0) printf("\n\n\n\n");
GA_Print(g_b);
if (me == 0) printf("\n\n\n\n");
GA_Print(g_c);
*/
/* Check answer */
NGA_Get(g_a,lo,hi,abuf,&ld);
NGA_Get(g_b,lo,hi,bbuf,&ld);
for (i=0; i<N; i++) {
for (j=0; j<N; j++) {
switch (data_type) {
case C_FLOAT:
((float*)cbuf)[i*N+j] = fzero;
break;
case C_DBL:
((double*)cbuf)[i*N+j] = dzero;
break;
case C_DCPL:
((DoubleComplex*)cbuf)[i*N+j] = zzero;
break;
case C_SCPL:
((SingleComplex*)cbuf)[i*N+j] = czero;
break;
default:
GA_Error("wrong data type", data_type);
}
for (k=0; k<N; k++) {
switch (data_type) {
case C_FLOAT:
((float*)cbuf)[i*N+j] += ((float*)abuf)[i*N+k]
*((float*)bbuf)[k*N+j];
break;
case C_DBL:
((double*)cbuf)[i*N+j] += ((double*)abuf)[i*N+k]
*((double*)bbuf)[k*N+j];
break;
case C_DCPL:
((DoubleComplex*)cbuf)[i*N+j].real +=
(((DoubleComplex*)abuf)[i*N+k].real
*((DoubleComplex*)bbuf)[k*N+j].real
-(((DoubleComplex*)abuf)[i*N+k].imag
*((DoubleComplex*)bbuf)[k*N+j].imag));
((DoubleComplex*)cbuf)[i*N+j].imag +=
(((DoubleComplex*)abuf)[i*N+k].real
*((DoubleComplex*)bbuf)[k*N+j].imag
+(((DoubleComplex*)abuf)[i*N+k].imag
*((DoubleComplex*)bbuf)[k*N+j].real));
break;
case C_SCPL:
((SingleComplex*)cbuf)[i*N+j].real +=
(((SingleComplex*)abuf)[i*N+k].real
*((SingleComplex*)bbuf)[k*N+j].real
-(((SingleComplex*)abuf)[i*N+k].imag
*((SingleComplex*)bbuf)[k*N+j].imag));
((SingleComplex*)cbuf)[i*N+j].imag +=
(((SingleComplex*)abuf)[i*N+k].real
*((SingleComplex*)bbuf)[k*N+j].imag
+(((SingleComplex*)abuf)[i*N+k].imag
*((SingleComplex*)bbuf)[k*N+j].real));
break;
default:
GA_Error("wrong data type", data_type);
}
}
}
}
GA_Sync();
if (me == 0) {
NGA_Get(g_c,lo,hi,abuf,&ld);
for (i=0; i<N; i++) {
for (j=0; j<N; j++) {
switch (data_type) {
case C_FLOAT:
fdiff = ((float*)abuf)[i*N+j]-((float*)cbuf)[i*N+j];
if (((float*)abuf)[i*N+j] != 0.0) {
fdiff /= ((float*)abuf)[i*N+j];
}
if (fabs(fdiff) > TOLERANCE) {
printf("p[%d] [%d,%d] Actual: %f Expected: %f\n",me,i,j,
((float*)abuf)[i*N+j],((float*)cbuf)[i*N+j]);
}
break;
case C_DBL:
ddiff = ((double*)abuf)[i*N+j]-((double*)cbuf)[i*N+j];
if (((double*)abuf)[i*N+j] != 0.0) {
ddiff /= ((double*)abuf)[i*N+j];
}
if (fabs(ddiff) > TOLERANCE) {
printf("p[%d] [%d,%d] Actual: %f Expected: %f\n",me,i,j,
((double*)abuf)[i*N+j],((double*)cbuf)[i*N+j]);
}
break;
case C_DCPL:
zdiff.real = ((DoubleComplex*)abuf)[i*N+j].real
-((DoubleComplex*)cbuf)[i*N+j].real;
zdiff.imag = ((DoubleComplex*)abuf)[i*N+j].imag
-((DoubleComplex*)cbuf)[i*N+j].imag;
if (((DoubleComplex*)abuf)[i*N+j].real != 0.0 ||
((DoubleComplex*)abuf)[i*N+j].imag != 0.0) {
ztmp = ((DoubleComplex*)abuf)[i*N+j];
ddiff = sqrt((zdiff.real*zdiff.real+zdiff.imag*zdiff.imag)
/(ztmp.real*ztmp.real+ztmp.imag*ztmp.imag));
} else {
ddiff = sqrt(zdiff.real*zdiff.real+zdiff.imag*zdiff.imag);
}
if (fabs(ddiff) > TOLERANCE) {
printf("p[%d] [%d,%d] Actual: (%f,%f) Expected: (%f,%f)\n",me,i,j,
((DoubleComplex*)abuf)[i*N+j].real,
((DoubleComplex*)abuf)[i*N+j].imag,
((DoubleComplex*)cbuf)[i*N+j].real,
((DoubleComplex*)cbuf)[i*N+j].imag);
}
break;
case C_SCPL:
cdiff.real = ((SingleComplex*)abuf)[i*N+j].real
-((SingleComplex*)cbuf)[i*N+j].real;
cdiff.imag = ((SingleComplex*)abuf)[i*N+j].imag
-((SingleComplex*)cbuf)[i*N+j].imag;
if (((SingleComplex*)abuf)[i*N+j].real != 0.0 ||
((SingleComplex*)abuf)[i*N+j].imag != 0.0) {
ctmp = ((SingleComplex*)abuf)[i*N+j];
fdiff = sqrt((cdiff.real*cdiff.real+cdiff.imag*cdiff.imag)
/(ctmp.real*ctmp.real+ctmp.imag*ctmp.imag));
} else {
fdiff = sqrt(cdiff.real*cdiff.real+cdiff.imag*cdiff.imag);
}
if (fabs(fdiff) > TOLERANCE) {
printf("p[%d] [%d,%d] Actual: (%f,%f) Expected: (%f,%f)\n",me,i,j,
((SingleComplex*)abuf)[i*N+j].real,
((SingleComplex*)abuf)[i*N+j].imag,
((SingleComplex*)cbuf)[i*N+j].real,
((SingleComplex*)cbuf)[i*N+j].imag);
}
break;
default:
GA_Error("wrong data type", data_type);
}
}
}
}
GA_Sync();
/* copy cbuf back to g_a */
if (me == 0) {
NGA_Put(g_a,lo,hi,cbuf,&ld);
}
GA_Sync();
/* Get norm of g_a */
switch (data_type) {
case C_FLOAT:
ftmp = GA_Fdot(g_a,g_a);
break;
case C_DBL:
dtmp = GA_Ddot(g_a,g_a);
break;
case C_DCPL:
ztmp = GA_Zdot(g_a,g_a);
break;
case C_SCPL:
ctmp = GA_Cdot(g_a,g_a);
break;
default:
GA_Error("wrong data type", data_type);
}
/* subtract C from A and put the results in B */
beta_flt = -1.0;
beta_dbl = -1.0;
beta_scpl.real = -1.0;
beta_dcpl.real = -1.0;
GA_Zero(g_b);
GA_Add(alpha,g_a,beta,g_c,g_b);
/* evaluate the norm of the difference between the two matrices */
switch (data_type) {
case C_FLOAT:
fdiff = GA_Fdot(g_b, g_b);
if (ftmp != 0.0) {
fdiff /= ftmp;
}
if(fabs(fdiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(fdiff), TOLERANCE);
GA_Error("GA_Sgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Sgemm OK\n\n");
}
break;
case C_DBL:
ddiff = GA_Ddot(g_b, g_b);
if (dtmp != 0.0) {
ddiff /= dtmp;
}
if(fabs(ddiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(ddiff), TOLERANCE);
GA_Error("GA_Dgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Dgemm OK\n\n");
}
break;
case C_DCPL:
zdiff = GA_Zdot(g_b, g_b);
if (ztmp.real != 0.0 || ztmp.imag != 0.0) {
ddiff = sqrt((zdiff.real*zdiff.real+zdiff.imag*zdiff.imag)
/(ztmp.real*ztmp.real+ztmp.imag*ztmp.imag));
} else {
ddiff = sqrt(zdiff.real*zdiff.real+zdiff.imag*zdiff.imag);
}
if(fabs(ddiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(zdiff.real), TOLERANCE);
GA_Error("GA_Zgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Zgemm OK\n\n");
}
break;
case C_SCPL:
cdiff = GA_Cdot(g_b, g_b);
if (ctmp.real != 0.0 || ctmp.imag != 0.0) {
fdiff = sqrt((cdiff.real*cdiff.real+cdiff.imag*cdiff.imag)
/(ctmp.real*ctmp.real+ctmp.imag*ctmp.imag));
} else {
fdiff = sqrt(cdiff.real*cdiff.real+cdiff.imag*cdiff.imag);
}
if(fabs(fdiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(cdiff.real), TOLERANCE);
GA_Error("GA_Cgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Cgemm OK\n\n");
}
break;
default:
GA_Error("wrong data type", data_type);
}
#endif
free(abuf);
free(bbuf);
free(cbuf);
switch (data_type) {
case C_FLOAT:
abuf = (void*)malloc(N*N*sizeof(float)/4);
bbuf = (void*)malloc(N*N*sizeof(float)/4);
cbuf = (void*)malloc(N*N*sizeof(float)/4);
break;
case C_DBL:
abuf = (void*)malloc(N*N*sizeof(double)/4);
bbuf = (void*)malloc(N*N*sizeof(double)/4);
cbuf = (void*)malloc(N*N*sizeof(double)/4);
break;
case C_DCPL:
abuf = (void*)malloc(N*N*sizeof(DoubleComplex)/4);
bbuf = (void*)malloc(N*N*sizeof(DoubleComplex)/4);
cbuf = (void*)malloc(N*N*sizeof(DoubleComplex)/4);
break;
case C_SCPL:
abuf = (void*)malloc(N*N*sizeof(SingleComplex)/4);
bbuf = (void*)malloc(N*N*sizeof(SingleComplex)/4);
cbuf = (void*)malloc(N*N*sizeof(SingleComplex)/4);
break;
default:
GA_Error("wrong data type", data_type);
}
/* Test multiply on a fraction of matrix. Start by reinitializing
* A and B */
GA_Zero(g_a);
GA_Zero(g_b);
GA_Zero(g_c);
if (me==0) printf("\nTest patch multiply\n");
lo[0] = N/4;
lo[1] = N/4;
hi[0] = 3*N/4-1;
hi[1] = 3*N/4-1;
ld = N/2;
/* Set up matrix A */
if (me==0) printf("\nInitialize A\n");
if (me == 0) {
for (i=N/4; i<3*N/4; i++) {
for (j=N/4; j<3*N/4; j++) {
switch (data_type) {
case C_FLOAT:
((float*)abuf)[(i-N/4)*N/2+(j-N/4)] = (float)(i*N+j);
break;
case C_DBL:
((double*)abuf)[(i-N/4)*N/2+(j-N/4)] = (double)(i*N+j);
break;
case C_DCPL:
((DoubleComplex*)abuf)[(i-N/4)*N/2+(j-N/4)].real = (double)(i*N+j);
((DoubleComplex*)abuf)[(i-N/4)*N/2+(j-N/4)].imag = 1.0;
break;
case C_SCPL:
((SingleComplex*)abuf)[(i-N/4)*N/2+(j-N/4)].real = (float)(i*N+j);
((SingleComplex*)abuf)[(i-N/4)*N/2+(j-N/4)].imag = 1.0;
break;
default:
GA_Error("wrong data type", data_type);
}
}
}
NGA_Put(g_a,lo,hi,abuf,&ld);
}
GA_Sync();
if (me==0) printf("\nInitialize B\n");
/* Set up matrix B */
if (me == 0) {
for (i=N/4; i<3*N/4; i++) {
for (j=N/4; j<3*N/4; j++) {
switch (data_type) {
case C_FLOAT:
((float*)bbuf)[(i-N/4)*N/2+(j-N/4)] = (float)(j*N+i);
break;
case C_DBL:
((double*)bbuf)[(i-N/4)*N/2+(j-N/4)] = (double)(j*N+i);
break;
case C_DCPL:
((DoubleComplex*)bbuf)[(i-N/4)*N/2+(j-N/4)].real = (double)(j*N+i);
((DoubleComplex*)bbuf)[(i-N/4)*N/2+(j-N/4)].imag = 1.0;
break;
case C_SCPL:
((SingleComplex*)bbuf)[(i-N/4)*N/2+(j-N/4)].real = (float)(j*N+i);
((SingleComplex*)bbuf)[(i-N/4)*N/2+(j-N/4)].imag = 1.0;
break;
default:
GA_Error("wrong data type", data_type);
}
}
}
NGA_Put(g_b,lo,hi,bbuf,&ld);
}
GA_Sync();
beta_flt = 0.0;
beta_dbl = 0.0;
beta_scpl.real = 0.0;
beta_dcpl.real = 0.0;
if (me==0) printf("\nPerform matrix multiply on sub-blocks\n");
switch (data_type) {
case C_FLOAT:
NGA_Matmul_patch('N','N',&alpha_flt,&beta_flt,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
case C_DBL:
NGA_Matmul_patch('N','N',&alpha_dbl,&beta_dbl,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
case C_SCPL:
NGA_Matmul_patch('N','N',&alpha_scpl,&beta_scpl,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
case C_DCPL:
NGA_Matmul_patch('N','N',&alpha_dcpl,&beta_dcpl,g_a,lo,hi,
g_b,lo,hi,g_c,lo,hi);
break;
default:
GA_Error("wrong data type", data_type);
}
GA_Sync();
#if 0
if (0) {
/*
if (data_type != C_SCPL && data_type != C_DCPL) {
*/
if (me==0) printf("\nCheck answer\n");
/* Multiply buffers by hand */
if (me == 0) {
for (i=0; i<N/2; i++) {
for (j=0; j<N/2; j++) {
switch (data_type) {
case C_FLOAT:
((float*)cbuf)[i*N/2+j] = fzero;
break;
case C_DBL:
((double*)cbuf)[i*N/2+j] = dzero;
break;
case C_DCPL:
((DoubleComplex*)cbuf)[i*N/2+j] = zzero;
break;
case C_SCPL:
((SingleComplex*)cbuf)[i*N/2+j] = czero;
break;
default:
GA_Error("wrong data type", data_type);
}
for (k=0; k<N/2; k++) {
switch (data_type) {
case C_FLOAT:
((float*)cbuf)[i*N/2+j] += ((float*)abuf)[i*N/2+k]
*((float*)bbuf)[k*N/2+j];
break;
case C_DBL:
((double*)cbuf)[i*N/2+j] += ((double*)abuf)[i*N/2+k]
*((double*)bbuf)[k*N/2+j];
break;
case C_DCPL:
((DoubleComplex*)cbuf)[i*N/2+j].real +=
(((DoubleComplex*)abuf)[i*N/2+k].real
*((DoubleComplex*)bbuf)[k*N/2+j].real
-(((DoubleComplex*)abuf)[i*N/2+k].imag
*((DoubleComplex*)bbuf)[k*N/2+j].imag));
((DoubleComplex*)cbuf)[i*N/2+j].imag +=
(((DoubleComplex*)abuf)[i*N/2+k].real
*((DoubleComplex*)bbuf)[k*N/2+j].imag
+(((DoubleComplex*)abuf)[i*N/2+k].imag
*((DoubleComplex*)bbuf)[k*N/2+j].real));
break;
case C_SCPL:
((SingleComplex*)cbuf)[i*N/2+j].real +=
(((SingleComplex*)abuf)[i*N/2+k].real
*((SingleComplex*)bbuf)[k*N/2+j].real
-(((SingleComplex*)abuf)[i*N/2+k].imag
*((SingleComplex*)bbuf)[k*N/2+j].imag));
((SingleComplex*)cbuf)[i*N/2+j].imag +=
(((SingleComplex*)abuf)[i*N/2+k].real
*((SingleComplex*)bbuf)[k*N/2+j].imag
+(((SingleComplex*)abuf)[i*N/2+k].imag
*((SingleComplex*)bbuf)[k*N/2+j].real));
break;
default:
GA_Error("wrong data type", data_type);
}
}
}
}
NGA_Put(g_a,lo,hi,cbuf,&ld);
}
if (me == 0) printf("\n\n\n\n");
/* Get norm of g_a */
switch (data_type) {
case C_FLOAT:
ftmp = NGA_Fdot_patch(g_a,'N',lo,hi,g_a,'N',lo,hi);
break;
case C_DBL:
dtmp = NGA_Ddot_patch(g_a,'N',lo,hi,g_a,'N',lo,hi);
break;
case C_DCPL:
ztmp = NGA_Zdot_patch(g_a,'N',lo,hi,g_a,'N',lo,hi);
break;
case C_SCPL:
ctmp = NGA_Cdot_patch(g_a,'N',lo,hi,g_a,'N',lo,hi);
break;
default:
GA_Error("wrong data type", data_type);
}
/* subtract C from A and put the results in B */
beta_flt = -1.0;
beta_dbl = -1.0;
beta_scpl.real = -1.0;
beta_dcpl.real = -1.0;
NGA_Zero_patch(g_b,lo,hi);
NGA_Add_patch(alpha,g_a,lo,hi,beta,g_c,lo,hi,g_b,lo,hi);
/* evaluate the norm of the difference between the two matrices */
switch (data_type) {
case C_FLOAT:
fdiff = NGA_Fdot_patch(g_b,'N',lo,hi,g_b,'N',lo,hi);
if (ftmp != 0.0) {
fdiff /= ftmp;
}
if(fabs(fdiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(fdiff), TOLERANCE);
GA_Error("GA_Sgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Sgemm OK\n\n");
}
break;
case C_DBL:
ddiff = NGA_Ddot_patch(g_b,'N',lo,hi,g_b,'N',lo,hi);
if (dtmp != 0.0) {
ddiff /= dtmp;
}
if(fabs(ddiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(ddiff), TOLERANCE);
GA_Error("GA_Dgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Dgemm OK\n\n");
}
break;
case C_DCPL:
zdiff = NGA_Zdot_patch(g_b,'N',lo,hi,g_b,'N',lo,hi);
if (ztmp.real != 0.0 || ztmp.imag != 0.0) {
ddiff = sqrt((zdiff.real*zdiff.real+zdiff.imag*zdiff.imag)
/(ztmp.real*ztmp.real+ztmp.imag*ztmp.imag));
} else {
ddiff = sqrt(zdiff.real*zdiff.real+zdiff.imag*zdiff.imag);
}
if(fabs(ddiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(zdiff.real), TOLERANCE);
GA_Error("GA_Zgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Zgemm OK\n\n");
}
break;
case C_SCPL:
cdiff = NGA_Cdot_patch(g_b,'N',lo,hi,g_b,'N',lo,hi);
if (ctmp.real != 0.0 || ctmp.imag != 0.0) {
fdiff = sqrt((cdiff.real*cdiff.real+cdiff.imag*cdiff.imag)
/(ctmp.real*ctmp.real+ctmp.imag*ctmp.imag));
} else {
fdiff = sqrt(cdiff.real*cdiff.real+cdiff.imag*cdiff.imag);
}
if(fabs(fdiff) > TOLERANCE) {
printf("\nabs(result) = %f > %f\n", fabsf(cdiff.real), TOLERANCE);
GA_Error("GA_Cgemm Failed", 1);
} else if (me == 0) {
printf("\nGA_Cgemm OK\n\n");
}
break;
default:
GA_Error("wrong data type", data_type);
}
}
#endif
free(abuf);
free(bbuf);
free(cbuf);
GA_Destroy(g_a);
GA_Destroy(g_b);
GA_Destroy(g_c);
}
main(int argc, char **argv)
{
int rank, nprocs, i, j;
int g_A, g_B, g_C, **local_C=NULL, dims[DIM]={SIZE,SIZE}, val1=5, val2=4, alpha=3, beta=2;
int clo[DIM]={SIZE-SIZE,SIZE-SIZE}, chi[DIM]={SIZE-1,SIZE-1}, ld=SIZE;
int **local_tm=NULL;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MA_init(C_INT, 1000, 1000);
GA_Initialize();
local_C=(int**)malloc(SIZE*sizeof(int*));
for(i=0; i<SIZE; i++)
local_C[i]=(int*)malloc(SIZE*sizeof(int));
local_tm=(int**)malloc(SIZE*sizeof(int*));
for(i=0; i<SIZE; i++)
local_tm[i]=(int*)malloc(SIZE*sizeof(int));
g_A = NGA_Create(C_INT, DIM, dims, "array_A", NULL);
g_B = GA_Duplicate(g_A, "array_B");
g_C = GA_Duplicate(g_A, "array_C");
GA_Fill(g_A, &val1);
GA_Fill(g_B, &val2);
GA_Add(&alpha, g_A, &beta, g_B, g_C);
GA_Sync();
GA_Print(g_A);
GA_Print(g_B);
GA_Print(g_C);
//printf("check 1\n");
NGA_Get(g_C, clo, chi, local_tm, &ld);
//printf("check 2\n");
// GA_Sync();
if(rank==0)
{
for(i=0; i<SIZE; i++)
{
for(j=0; j<SIZE; j++)printf("%d ", local_tm[i][j]);
printf("\n");
}
}
/*
if(rank==0)
{
NGA_Get(g_C, clo, chi, local_C, &ld);
printf("check 1 \n");
for(i=0; i<SIZE; i++)
{
for(j=0; j<SIZE; j++)printf("%d ", local_C[i][j]);
printf("\n");
}
printf("check 2\n");
for(i=0; i<SIZE; i++)
{
for(j=0; j<SIZE; j++)
if(local_C[i][j]!=(alpha*val1)+(beta*val2)) printf("GA Error : \n");
}
}
*/
//GA_Sync();
if(rank==0)
printf("Test Completed \n");
// GA_Sync();
/*
GA_Destroy(g_A);
GA_Destroy(g_B);
GA_Destroy(g_C);
*/
//*******************************************************************
/* what would be the possible reason for GA_destroy to get failed ..,
* solve this before consolidate the whole
*/
GA_Terminate();
MPI_Finalize();
}
