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 */
GA_Mask_sync(0, 1);
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;
}
void verify(int g_a, int g_b, int g_c, int *lo, int *hi, int *ld, int N)
{
double rchk, alpha=1.0, beta=0.0;
int g_chk, me=GA_Nodeid();
g_chk = GA_Duplicate(g_a, "array Check");
if(!g_chk) GA_Error("duplicate failed",NDIMS);
GA_Sync();
GA_Dgemm('n', 'n', N, N, N, 1.0, g_a,
g_b, 0.0, g_chk);
GA_Sync();
alpha=1.0, beta=-1.0;
GA_Add(&alpha, g_c, &beta, g_chk, g_chk);
rchk = GA_Ddot(g_chk, g_chk);
if (me==0) {
printf("Normed difference in matrices: %12.4e\n", rchk);
if(rchk < -TOLERANCE || rchk > TOLERANCE)
GA_Error("Matrix multiply verify failed",0);
else
printf("Matrix Mutiply OK\n");
}
GA_Destroy(g_chk);
}
/**
* 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;
}
int nw_inp_from_string(Integer rtdb, const char *input)
{
char filename[30];
FILE *file;
#if defined(USE_FCD) || defined(CRAY_T3E) || defined(WIN32)
_fcd fstring;
#else
char fstring[255];
#endif
int status;
const char base[] = "temp";
const char ending[] = ".nw";
int number ;
// This is bad, not 100% sure to be unique, since could be subgroup
if (GA_Pgroup_get_world() != GA_Pgroup_get_default()) {
number = (int) util_sgroup_mygroup_() ;
} else {
number = 0 ;
}
sprintf(filename, "%s%d%s", base,number,ending);
if (GA_Nodeid() == 0) {
if (!(file = fopen(filename,"w"))) {
GA_Error("nw_inp_from_string: failed to open temp.nw\n",0);
}
if (fwrite(input, 1, strlen(input), file) != strlen(input)) {
GA_Error("nw_inp_from_string: failed to write to temp.nw\n",0);
}
if (fwrite("\n", 1, 1, file) != 1) {
GA_Error("nw_inp_from_string: failed to write to temp.nw\n",0);
}
(void) fclose(file);
}
#if defined(CRAY_T3E)
fstring = _cptofcd(filename, strlen(filename));
status = nw_inp_from_file_(&rtdb, fstring);
#elif defined(WIN32)
fstring.string = filename;
fstring.len = strlen(filename);
status = nw_inp_from_file_(&rtdb, fstring);
#elif defined(USE_FCD)
#error Do something about _fcd
#else
status = nw_inp_from_file_(&rtdb, filename, strlen(filename));
#endif
if (GA_Nodeid() == 0) (void) unlink(filename);
return status;
}
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();
}
validate_transpose(int g_A, int g_B, int* lo, int* hi, int ld)
{
int i, j;
int local_A[SIZE][SIZE], local_B[SIZE][SIZE];
// int **local_A=NULL, **local_B=NULL;
/*
local_A=(int**)malloc(SIZE*sizeof(int*));
for(i=0; i<SIZE; i++)
local_value[i]=(int*)malloc(SIZE*sizeof(int));
local_B=(int**)malloc(SIZE*sizeof(int*));
for(i=0; i<SIZE; i++)
local_value[i]=(int*)malloc(SIZE*sizeof(int));
*/
NGA_Get(g_A, lo, hi, local_A, &ld);
NGA_Get(g_B, lo, hi, local_B, &ld);
for(i=1; i<SIZE; i++)
{
for(j=1; j<SIZE; j++)
{
if(local_B[j][i]!=local_A[i][j])
GA_Error("ERROR : in passing values", DIM);
}
}
}
void FATR
ga_antisymmetrize_(Integer *g_a) {
DoublePrecision alpha = 0.5;
int i, me = GA_Nodeid();
extern void * FATR ga_malloc(Integer nelem, int type, char *name);
extern void FATR ga_free(void *ptr);
void FATR gai_subtr(int *lo, int *hi, void *a, void *b, DoublePrecision alpha,
int type, Integer nelem, int ndim);
int alo[GA_MAX_DIM], ahi[GA_MAX_DIM], lda[GA_MAX_DIM];
int blo[GA_MAX_DIM], bhi[GA_MAX_DIM], ldb[GA_MAX_DIM];
int ndim, dims[GA_MAX_DIM], type;
Integer nelem=1;
Logical have_data;
void *a_ptr, *b_ptr;
GA_Sync();
NGA_Inquire((int)(*g_a), &type, &ndim, dims);
if (dims[0] != dims[1])
GA_Error("ga_sym: can only sym square matrix", 0L);
/* Find the local distribution */
NGA_Distribution((int)(*g_a), me, alo, ahi);
have_data = ahi[0]>=0;
for(i=1; i<ndim; i++) have_data = have_data && ahi[i]>=0;
if(have_data) {
NGA_Access((int)(*g_a), alo, ahi, &a_ptr, lda);
for(i=0; i<ndim; i++) nelem *= ahi[i]-alo[i] +1;
b_ptr = (void *) ga_malloc(nelem, MT_C_DBL, "v");
for(i=2; i<ndim; i++) {bhi[i]=ahi[i]; blo[i]=alo[i]; }
/* switch rows and cols */
blo[1]=alo[0];
bhi[1]=ahi[0];
blo[0]=alo[1];
bhi[0]=ahi[1];
for (i=0; i < ndim-1; i++)
ldb[i] = bhi[i+1] - blo[i+1] + 1;
NGA_Get((int)(*g_a), blo, bhi, b_ptr, ldb);
}
GA_Sync();
if(have_data) {
gai_subtr(alo, ahi, a_ptr, b_ptr, alpha, type, nelem, ndim);
NGA_Release_update((int)(*g_a), alo, ahi);
ga_free(b_ptr);
}
GA_Sync();
}
/*
* Return the no. of bytes that n ints=Integers occupy
*/
Integer FATR util_mitob_(Integer *n)
{
if (*n < 0)
GA_Error("util_MITOB_: negative argument",*n);
return (Integer) (*n * sizeof(Integer));
}
/*
* Check to see if inversion is correct. Start by copying g_a into local
* buffer a, and g_b into local buffer b.
*/
void verify(int g_a, int g_b) {
int i, type, ndim, dims[MAXDIM], lo[MAXDIM], hi[MAXDIM], ld[MAXDIM];
int a[MAXPROC*TOTALELEMS],b[MAXPROC*TOTALELEMS];
/* Get dimensions of GA */
NGA_Inquire(g_a, &type, &ndim, dims);
lo[0] = 0;
hi[0] = dims[0]-1;
/* ### copy the block of data described by the arrays "lo" and "hi" from
* ### the global array "g_a" into the local array "a". Copy the same block
* ### of data from "g_b" into the local array "b". Use the array of strides
* ### "ld" to describe the physical layout of "a" and "b". */
NGA_Get(g_a,lo,hi,a,ld);
NGA_Get(g_b,lo,hi,b,ld);
for(i=0; i<dims[0]; i++)
if (a[i] != b[dims[0]-i-1])
{
printf("Mismatch: a[%d]=%d is not equal to b[%d]=%d\n",
i, a[i], dims[0]-i-1, b[dims[0]-i-1]);
GA_Error("verify failed",0);
}
printf(" Transpose OK\n");
}
int main(int argc, char **argv) {
int heap=300000, stack=300000;
int me, nprocs, i;
double start, end;
/* Initialize Message Passing library */
MPI_Init(&argc, &argv);
/* Initialize GA */
GA_Initialize();
/* Initialize Memory Allocator (MA) */
if(! MA_init(C_DBL, stack, heap) ) GA_Error("MA_init failed",stack+heap);
me = GA_Nodeid();
nprocs = GA_Nnodes();
if(me==0) {
printf("\nUsing %d processes\n\n", nprocs); fflush(stdout);
}
start = MPI_Wtime();
for(i =0; i< 1000; i++) {
TRANSPOSE1D();
}
end = MPI_Wtime();
if(me==0) printf(" Time=%2.5e secs\n\n",end-start);
if(me==0)printf("\nTerminating ..\n");
GA_Terminate();
MPI_Finalize();
}
void FATR util_fadvise_noreuse_(const char *fort_fname, int flen){
char buf[1024];
if (!fortchar_to_string(fort_fname, flen, buf, sizeof(buf)))
GA_Error("util_fadvise: fortchar_to_string failed for fname",0);
(void) FATR util_fadvise(buf, POSIX_FADV_NOREUSE);
}
int main(int argc, char **argv)
{
int status, me;
int max_arrays = 10;
double max_sz = 1e8, max_disk = 1e10, max_mem = 1e6;
int stack = 120000, heap = 3200000;
int numfiles, numioprocs;
int total, nproc;
MP_INIT(argc,argv);
GA_Initialize();
me = GA_Nodeid();
nproc = GA_Nnodes();
total = pow(SIZE,NDIM)*sizeof(double);
if (!GA_Uses_ma()) {
if (GA_Nodeid() == 0) {
printf("GA is not using MA\n");
}
stack = 100000;
heap = (int)(2.2*(float)(total));
}
if (MA_init(MT_F_DBL, stack, heap) ) {
if (DRA_Init(max_arrays, max_sz, max_disk, max_mem) != 0)
GA_Error("DRA_Init failed: ",0);
if (USER_CONFIG == 0) {
numfiles = -1;
numioprocs = -1;
} else if (USER_CONFIG == 1) {
numfiles = 1;
numioprocs = GA_Cluster_nnodes();
} else if (USER_CONFIG == 2) {
numfiles = GA_Cluster_nnodes();
numioprocs = GA_Cluster_nnodes();
} else {
numfiles = 1;
numioprocs = 1;
}
if (me==0) {
printf("Disk resident arrays configured as:\n");
printf(" Number of files: %d\n",numfiles);
printf(" Number of I/O processors: %d\n",numioprocs);
}
DRA_Set_default_config(numfiles,numioprocs);
if (me == 0) printf("\n");
if (me == 0) printf("TESTING PERFORMANCE OF DISK ARRAYS\n");
if (me == 0) printf("\n");
test_io_dbl();
status = DRA_Terminate();
GA_Terminate();
} else {
printf("MA_init failed\n");
}
if(me == 0) printf("all done ...\n");
MP_FINALIZE();
return 0;
}
static void Error(char *string, int integer)
{
/*
(void) fflush(stdout);
(void) fprintf(stderr,"\n\nError was called.\n");
(void) fprintf(stderr,string);
(void) fprintf(stderr," %d (%#x).\n",integer,integer);
exit(1); */
GA_Error(string, (long) integer);
}
void FATR UTIL_FILE_UNLINK(_fcd input)
{
int lin = _fcdlen(input);
#else
void util_file_unlink_(const char *input, int lin)
{
#endif
char in[255];
if (!fortchar_to_string(input, lin, in, sizeof(in)))
GA_Error("util_file_unlink: fortchar_to_string failed for in",0);
util_file_unlink(in);
}
int util_mic_get_device_() {
int count;
if (!offload_master_()) {
fprintf(stdout, "%02d: need to be offload master\n", GA_Nodeid());
GA_Error("util_mic_get_device error", 0L);
}
count = util_cgetppn()/util_mic_get_num_devices_();
if (count > 0) {
count=util_my_smp_index()/util_nwc_ranks_per_device_();
}
return count;
}
void util_file_unlink(const char *filename)
/*
Delete the file. If the file does not exist, quietly return.
If the file exists and the unlink fails then abort.
*/
{
/*
if (access(filename, F_OK) == 0) {
if (unlink(filename)) {
fprintf(stderr,"util_file_unlink: failed unlinking %s\n",
filename);
GA_Error("util_file_unlink",0);
}
}
*/
if (EAF_Delete(filename) != 0)
GA_Error("util_file_unlink",0);
}
int main(int argc, char **argv)
{
int proc, nprocs;
int M, N, K; /* */
int blockX_len, blockY_len;
int heap=3000000, stack=3000000;
if (argc == 6) {
M = atoi(argv[1]);
N = atoi(argv[2]);
K = atoi(argv[3]);
blockX_len = atoi(argv[4]);
blockY_len = atoi(argv[5]);
}
else {
printf("Please enter ./a.out <M> <N> <K> <BLOCK-X-LEN> <BLOCK-Y-LEN>");
exit(-1);
}
MPI_Init(&argc, &argv);
GA_Initialize();
if(! MA_init(C_DBL, stack, heap) ) GA_Error("MA_init failed",stack+heap);
proc = GA_Nodeid();
nprocs = GA_Nnodes();
if(proc == 0) {
printf("Using %d processes\n", nprocs);
fflush(stdout);
}
matrix_multiply(M, N, K, blockX_len, blockY_len);
if(proc == 0)
printf("\nTerminating ..\n");
GA_Terminate();
MPI_Finalize();
return 0;
}
void lu(double *A, int matrix_size)
{
double *a=NULL, *a_verify=NULL;
int info;
a = (double*)malloc(matrix_size*matrix_size*sizeof(double));
a_verify = (double*)malloc(matrix_size*matrix_size*sizeof(double));
if(a == NULL || a_verify == NULL)
{
GA_Error("lu(): malloc failed", matrix_size*matrix_size*sizeof(double));
}
copy_array(A, a, matrix_size);
copy_array(A, a_verify, matrix_size);
#if 0
printf("\nDoing LU Factorization\n\n");
lu_basic(a, matrix_size);
printf("LU = \n");
print_array(a, matrix_size);
printf("\nDoing LAPACK's LU Factorization\n\n");
#endif
info = lu_lapack(a_verify, matrix_size);
#if DEBUG
printf("LU = ");
print_array(a_verify, matrix_size);
#endif
if(info!=0)
{
printf("\nError: dgetrf() of lapack is NOT successful (INFO=%d)\n\n",
info);
printf("NOTE:\n INFO=0: successful exit\n INFO<0: if INFO = -i, the i-th argument had an illegal value\n INFO>0: if INFO = i, U(i,i) is exactly zero. The factorization\n has been completed, but the factor U is exactly singular, and\n division by zero will occur if it is used to solve a system of\n equations.\n");
exit(0);
}
free(a_verify);
free(a);
}
int main(int argc, char **args)
{
PetscErrorCode ierr;
ierr = PetscInitialize(&argc, &args, (char *) 0, ""); CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD, "Start %s\n", __FILE__); CHKERRQ(ierr);
GA_Initialize();
ierr = MA_init(MT_DBL, 1000, 1000 );
if( !ierr ) GA_Error("\n\n\nga allocaltion failed\n\n",ierr);
size_t len = 1024;
char name[1024];
PetscGetHostName(name,len);
PetscPrintf(PETSC_COMM_SELF, "Hello from %s\n", name);
ierr = testCreate3D(); CHKERRQ(ierr);
GA_Terminate();
ierr = PetscPrintf(PETSC_COMM_WORLD, "End %s\n", __FILE__); CHKERRQ(ierr);
ierr = PetscFinalize(); CHKERRQ(ierr);
}
int main(int argc, char **argv) {
int heap=300000, stack=300000;
int me, nprocs;
/* Step1: Initialize Message Passing library */
#ifdef MPI
MPI_Init(&argc, &argv); /* initialize MPI */
#else
PBEGIN_(argc, argv); /* initialize TCGMSG */
#endif
/* Step2: Initialize GA */
/* ### intialize the GA library */
GA_Initialize();
/* Step3: Initialize Memory Allocator (MA) */
if(! MA_init(C_DBL, stack, heap) ) GA_Error("MA_init failed",stack+heap);
/* ### assign the local processor ID to the int variable "me"
* ### and the total number of processors to the int variable
* ### "nprocs" */
me = GA_Nodeid();
nprocs = GA_Nnodes();
if(me==0) {
printf("\nUsing %d processes\n\n", nprocs); fflush(stdout);
}
TRANSPOSE1D();
if(me==0)printf("\nTerminating ..\n");
/* ### terminate the GA library */
GA_Terminate();
#ifdef MPI
MPI_Finalize();
#else
PEND_();
#endif
}
/*\ fill n-dimensional array section with value
\*/
void fill_patch(double *ptr, int dim[], int ld[], int ndim, double val)
{
int i, j, stride=1;
switch (ndim){
case 0: GA_Error("fill_patch: error",ndim);
case 1: for(i=0;i <dim[0];i++)ptr[i]=val;
break;
case 2: for(i=0; i< dim[0]; i++){
double *arr = ptr + i*ld[0];
for(j=0; j< dim[1]; j++)arr[j]=val;
}
break;
default:
for(i=0; i<ndim-1; i++)stride *=ld[i];
for(i=0; i<dim[0]; i++){
double *arr = ptr + stride*i;
fill_patch(arr, dim+1, ld+1, ndim-1, val);
}
}
}
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;
}
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);
}
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(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);
}
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);
}
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;
}
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;
}
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);
}