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;
}
int
main(int argc, char **argv) {
Integer heap=9000000, stack=9000000;
int me, nproc;
DoublePrecision time;
MP_INIT(argc,argv);
GA_INIT(argc,argv); /* initialize GA */
nproc = GA_Nnodes();
me = GA_Nodeid();
if(me==0) printf("Using %d processes\n\n",nproc);
if (me==0) printf ("Matrix size is %d X %d\n",N,N);
#ifdef USE_REGULAR
if (me == 0) printf("\nUsing regular data distribution\n\n");
#endif
#ifdef USE_SIMPLE_CYCLIC
if (me == 0) printf("\nUsing simple block-cyclic data distribution\n\n");
#endif
#ifdef USE_SCALAPACK
if (me == 0) printf("\nUsing ScaLAPACK data distribution\n\n");
#endif
#ifdef USE_TILED
if (me == 0) printf("\nUsing tiled data distribution\n\n");
#endif
if(!MA_init((Integer)MT_F_DBL, stack/nproc, heap/nproc))
GA_Error("MA_init failed bytes= %d",stack+heap);
#ifdef PERMUTE
{
int i, *list = (int*)malloc(nproc*sizeof(int));
if(!list)GA_Error("malloc failed",nproc);
for(i=0; i<nproc;i++)list[i]=nproc-1-i;
GA_Register_proclist(list, nproc);
free(list);
}
#endif
if(GA_Uses_fapi())GA_Error("Program runs with C API only",1);
time = MP_TIMER();
do_work();
/* printf("%d: Total Time = %lf\n", me, MP_TIMER()-time);
printf("%d: GEMM Total Time = %lf\n", me, gTime);
*/
if(me==0)printf("\nSuccess\n\n");
GA_Terminate();
MP_FINALIZE();
return 0;
}
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;
}
