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);
}
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();
}
// -------------------------------------------------------------
// MatZeroEntries_DenseGA
// -------------------------------------------------------------
static
PetscErrorCode
MatZeroEntries_DenseGA(Mat A)
{
PetscErrorCode ierr = 0;
struct MatGACtx *ctx;
ierr = MatShellGetContext(A, &ctx); CHKERRQ(ierr);
GA_Zero(ctx->ga);
return ierr;
}
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();
}
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[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();
}
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);
}
int main(int argc, char **argv)
{
int me;
int nproc;
int status;
int g_a;
int dims[NDIM];
int chunk[NDIM];
int pg_world;
size_t num = 10;
double *p1 = NULL;
double *p2 = NULL;
size_t i;
int num_mutex;
int lo[1];
int hi[1];
int ld[1]={1};
MPI_Comm comm;
MP_INIT(argc,argv);
GA_INIT(argc,argv);
me = GA_Nodeid();
nproc = GA_Nnodes();
comm = GA_MPI_Comm_pgroup_default();
printf("%d: Hello world!\n",me);
if (me==0) printf("%d: GA_Initialize\n",me);
/*if (me==0) printf("%d: ARMCI_Init\n",me);*/
/*ARMCI_Init();*/
/*if (me==0) printf("%d: MA_Init\n",me);*/
/*MA_init(MT_DBL, 8*1024*1024, 2*1024*1024);*/
if (me==0) printf("%d: GA_Create_handle\n",me);
g_a = GA_Create_handle();
if (me==0) printf("%d: GA_Set_array_name\n",me);
GA_Set_array_name(g_a,"test array A");
dims[0] = 30;
if (me==0) printf("%d: GA_Set_data\n",me);
GA_Set_data(g_a,NDIM,dims,MT_DBL);
chunk[0] = -1;
if (me==0) printf("%d: GA_Set_chunk\n",me);
GA_Set_chunk(g_a,chunk);
if (me==0) printf("%d: GA_Pgroup_get_world\n",me);
pg_world = GA_Pgroup_get_world();
if (me==0) printf("%d: GA_Set_pgroup\n",me);
GA_Set_pgroup(g_a,pg_world);
if (me==0) printf("%d: GA_Allocate\n",me);
status = GA_Allocate(g_a);
if(0 == status) MPI_Abort(comm,100);
if (me==0) printf("%d: GA_Zero\n",me);
GA_Zero(g_a);
if (me==0) printf("%d: GA_Sync\n",me);
GA_Sync();
num = 10;
p1 = malloc(num*sizeof(double));
/*double* p1 = ARMCI_Malloc_local(num*sizeof(double));*/
if (p1==NULL) MPI_Abort(comm,1000);
p2 = malloc(num*sizeof(double));
/*double* p2 = ARMCI_Malloc_local(num*sizeof(double));*/
if (p2==NULL) MPI_Abort(comm,2000);
for ( i=0 ; i<num ; i++ ) p1[i] = 7.0;
for ( i=0 ; i<num ; i++ ) p2[i] = 3.0;
num_mutex = 17;
status = GA_Create_mutexes(num_mutex);
if (me==0) printf("%d: GA_Create_mutexes = %d\n",me,status);
/***************************************************************/
if (me==0) {
printf("%d: before GA_Lock\n",me);
GA_Lock(0);
lo[0] = 0;
hi[0] = num-1;
GA_Init_fence();
NGA_Put(g_a,lo,hi,p1,ld);
GA_Fence();
GA_Unlock(0);
printf("%d: after GA_Unlock\n",me);
}
GA_Print(g_a);
if (me==1) {
printf("%d: before GA_Lock\n",me);
GA_Lock(0);
lo[0] = 0;
hi[0] = num-1;
GA_Init_fence();
NGA_Get(g_a,lo,hi,p2,ld);
GA_Fence();
GA_Unlock(0);
printf("%d: after GA_Unlock\n",me);
for ( i=0 ; i<num ; i++ ) printf("p2[%2lu] = %20.10f\n",
(long unsigned)i,p2[i]);
}
/***************************************************************/
status = GA_Destroy_mutexes();
if (me==0) printf("%d: GA_Destroy_mutexes = %d\n",me,status);
/*ARMCI_Free(p2);*/
/*ARMCI_Free(p1);*/
free(p2);
free(p1);
if (me==0) printf("%d: GA_Destroy\n",me);
GA_Destroy(g_a);
/*if (me==0) printf("%d: ARMCI_Finalize\n",me);*/
/*ARMCI_Finalize();*/
if (me==0) printf("%d: GA_Terminate\n",me);
GA_Terminate();
if (me==0) printf("%d: MPI_Finalize\n",me);
MPI_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);
}
static int test(int shape_idx, int type_idx, int dist_idx)
{
int type = TYPES[type_idx];
int *dims = SHAPES[shape_idx];
int ndim = SHAPES_NDIM[shape_idx];
mock_ga_t *mock_a, *result_a;
int g_a;
int buffer[100];
int lo[GA_MAX_DIM], hi[GA_MAX_DIM], ld[GA_MAX_DIM], shape[GA_MAX_DIM];
int result=0, error_index=-1, error_proc=-1;
mock_a = Mock_Create(type, ndim, dims, "mock", NULL);
result_a = Mock_Create(type, ndim, dims, "mock", NULL);
g_a = create_function[dist_idx](type, ndim, dims);
mock_data(mock_a, g_a);
mock_to_global(mock_a, g_a);
Mock_Zero(mock_a);
GA_Zero(g_a);
global_to_mock(g_a, result_a);
result = neq_mock(mock_a, result_a, &error_index);
if (0 != result) {
error_proc = GA_Nodeid();
}
GA_Igop(&result, 1, "+");
GA_Igop(&error_proc, 1, "max");
if (error_proc != GA_Nodeid()) {
error_index = 0;
}
GA_Igop(&error_index, 1, "+");
if (0 != result) {
if (error_proc == GA_Nodeid()) {
printf("ERROR: local result failed to compare to global result\n");
printf("\terror_proc=%d\n", error_proc);
printf("\terror_index=%d\n", error_index);
printf("***LOCAL RESULT***\n");
Mock_Print(mock_a);
printf("***GLOBAL RESULT***\n");
Mock_Print(result_a);
printf("\tprinting array distribution\n");
}
GA_Sync();
GA_Print(g_a);
GA_Print_distribution(g_a);
return 1;
}
Mock_Destroy(mock_a);
Mock_Destroy(result_a);
GA_Destroy(g_a);
return 0;
}
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);
}
/**
* Evaluate offsets for each network component
*/
void setOffsets(void)
{
// Interleave contributions from buses and branches to match matrices
int i,j,jdx,jdx1,jdx2;
int *i_bus_offsets = new int[p_nBuses];
int *i_branch_offsets = new int[p_nBranches];
for (i=0; i<p_nBuses; i++) {
i_bus_offsets[i] = 0;
}
for (i=0; i<p_nBranches; i++) {
i_branch_offsets[i] = 0;
}
int icnt = 0;
int nsize;
// Evaluate offsets for individual network components
for (i=0; i<p_nBuses; i++) {
if (p_network->getActiveBus(i)) {
i_bus_offsets[i] = icnt;
icnt += p_network->getBus(i)->vectorNumElements();
std::vector<int> nghbrs = p_network->getConnectedBranches(i);
nsize = nghbrs.size();
for (j=0; j<nsize; j++) {
// Need to avoid double counting of branches when evaluating offsets.
// If branch is non-local and it is active, then include it in offsets.
// Otherwise, if branch is local and bus i is equal to the "from" bus,
// then include it in the offsets.
jdx = nghbrs[j];
if (isLocalBranch(jdx)) {
p_network->getBranchEndpoints(jdx,&jdx1,&jdx2);
if (jdx1 == i) {
i_branch_offsets[jdx] = icnt;
icnt += p_network->getBranch(jdx)->vectorNumElements();
}
} else {
if (p_network->getActiveBranch(jdx)) {
i_branch_offsets[jdx] = icnt;
icnt += p_network->getBranch(jdx)->vectorNumElements();
}
}
}
}
}
// Total number of rows and columns from this processor have been evaluated,
// now create buffers that can scatter individual offsets to global arrays
int **i_bus_index = new int*[p_nBuses];
int **i_branch_index = new int*[p_nBranches];
int *i_bus_index_buf = new int[p_nBuses];
int *i_branch_index_buf = new int[p_nBranches];
int *i_bus_value_buf = new int[p_nBuses];
int *i_branch_value_buf = new int[p_nBranches];
int i_bus_cnt = 0;
int i_branch_cnt = 0;
int row_offset = p_Offsets[p_me];
int nbus = 0;
int nbranch = 0;
for (i=0; i<p_nBuses; i++) {
if (p_network->getActiveBus(i)) {
nbus++;
i_bus_value_buf[i_bus_cnt] = i_bus_offsets[i]+row_offset;
i_bus_index_buf[i_bus_cnt] = p_network->getGlobalBusIndex(i);
i_bus_index[i_bus_cnt] = &i_bus_index_buf[i_bus_cnt];
i_bus_cnt++;
}
}
for (i=0; i<p_nBranches; i++) {
if (p_network->getActiveBranch(i)) {
nbranch++;
i_branch_value_buf[i_branch_cnt] = i_branch_offsets[i]+row_offset;
i_branch_index_buf[i_branch_cnt] = p_network->getGlobalBranchIndex(i);
i_branch_index[i_branch_cnt] = &i_branch_index_buf[i_branch_cnt];
i_branch_cnt++;
}
}
delete [] i_bus_offsets;
delete [] i_branch_offsets;
// Create global arrays that hold column and row offsets for all buses and
// branches in the network. First create map array for global arrays
int *t_busMap = new int[p_nNodes];
int *t_branchMap = new int[p_nNodes];
for (i=0; i<p_nNodes; i++) {
t_busMap[i] = 0;
t_branchMap[i] = 0;
}
t_busMap[p_me] = nbus;
t_branchMap[p_me] = nbranch;
char plus[2];
strcpy(plus,"+");
GA_Pgroup_igop(p_GAgrp, t_busMap, p_nNodes, plus);
GA_Pgroup_igop(p_GAgrp, t_branchMap, p_nNodes, plus);
int *busMap = new int[p_nNodes];
int *branchMap = new int[p_nNodes];
busMap[0] = 0;
branchMap[0] = 0;
int total_buses = t_busMap[0];
int total_branches = t_branchMap[0];
for (i=1; i<p_nNodes; i++) {
busMap[i] = busMap[i-1] + t_busMap[i-1];
total_buses += t_busMap[i];
branchMap[i] = branchMap[i-1] + t_branchMap[i-1];
total_branches += t_branchMap[i];
}
delete [] t_busMap;
delete [] t_branchMap;
int one = 1;
g_bus_offsets = GA_Create_handle();
GA_Set_data(g_bus_offsets, one, &total_buses, C_INT);
GA_Set_irreg_distr(g_bus_offsets, busMap, &p_nNodes);
GA_Set_pgroup(g_bus_offsets, p_GAgrp);
if (!GA_Allocate(g_bus_offsets)) {
char buf[256];
sprintf(buf,"GenVectorMap::setOffsets: Unable to allocate distributed array for bus offsets\n");
printf("%s",buf);
throw gridpack::Exception(buf);
}
GA_Zero(g_bus_offsets);
g_branch_offsets = GA_Create_handle();
GA_Set_data(g_branch_offsets, one, &total_branches, C_INT);
GA_Set_irreg_distr(g_branch_offsets, branchMap, &p_nNodes);
GA_Set_pgroup(g_branch_offsets, p_GAgrp);
if (!GA_Allocate(g_branch_offsets)) {
char buf[256];
sprintf(buf,"GenVectorMap::setOffsets: Unable to allocate distributed array for branch offsets\n");
printf("%s",buf);
throw gridpack::Exception(buf);
}
GA_Zero(g_branch_offsets);
delete [] busMap;
delete [] branchMap;
// Scatter offsets to global arrays
NGA_Scatter(g_bus_offsets, i_bus_value_buf, i_bus_index, i_bus_cnt);
NGA_Scatter(g_branch_offsets, i_branch_value_buf, i_branch_index, i_branch_cnt);
NGA_Pgroup_sync(p_GAgrp);
delete [] i_bus_index;
delete [] i_branch_index;
delete [] i_bus_index_buf;
delete [] i_branch_index_buf;
delete [] i_bus_value_buf;
delete [] i_branch_value_buf;
}
// 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;
}
/* Square matrix-matrix multiplication */
void matrix_multiply(int M, int N, int K,
int blockX_len, int blockY_len)
{
/* Local buffers and Global arrays declaration */
double *a=NULL, *b=NULL, *c=NULL;
int dims[NDIMS], ld[NDIMS], chunks[NDIMS];
int lo[NDIMS], hi[NDIMS], cdims[NDIMS]; /* dim of blocks */
int g_a, g_b, g_c, g_cnt, g_cnt2;
int offset;
double alpha = 1.0, beta=0.0;
int count_p = 0, next_p = 0;
int count_gac = 0, next_gac = 0;
double t1,t2,seconds;
ga_nbhdl_t nbh;
int count_acc = 0;
/* Find local processor ID and the number of processes */
int proc=GA_Nodeid(), nprocs=GA_Nnodes();
if ((M % blockX_len) != 0 || (M % blockY_len) != 0 || (N % blockX_len) != 0 || (N % blockY_len) != 0
|| (K % blockX_len) != 0 || (K % blockY_len) != 0)
GA_Error("Dimension size M/N/K is not divisible by X/Y block sizes", 101);
/* Allocate/Set process local buffers */
a = malloc (blockX_len * blockY_len * sizeof(double));
b = malloc (blockX_len * blockY_len * sizeof(double));
c = malloc (blockX_len * blockY_len * sizeof(double));
cdims[0] = blockX_len;
cdims[1] = blockY_len;
/* Configure array dimensions */
for(int i = 0; i < NDIMS; i++) {
dims[i] = N;
chunks[i] = -1;
ld[i] = cdims[i]; /* leading dimension/stride of the local buffer */
}
/* create a global array g_a and duplicate it to get g_b and g_c*/
g_a = NGA_Create(C_DBL, NDIMS, dims, "array A", chunks);
if (!g_a)
GA_Error("NGA_Create failed: A", NDIMS);
#if DEBUG>1
if (proc == 0)
printf(" Created Array A\n");
#endif
/* Ditto for C and B */
g_b = GA_Duplicate(g_a, "array B");
g_c = GA_Duplicate(g_a, "array C");
if (!g_b || !g_c)
GA_Error("GA_Duplicate failed",NDIMS);
if (proc == 0)
printf("Created Arrays B and C\n");
/* Subscript array for read-incr, which is nothing but proc */
int * rdcnt = malloc (nprocs * sizeof(int));
memset (rdcnt, 0, nprocs * sizeof(int));
int * rdcnt2 = malloc (nprocs * sizeof(int));
memset (rdcnt2, 0, nprocs * sizeof(int));
/* Create global array of nprocs elements for nxtval */
int counter_dim[1];
counter_dim[0] = nprocs;
g_cnt = NGA_Create(C_INT, 1, counter_dim, "Shared counter", NULL);
if (!g_cnt)
GA_Error("Shared counter failed",1);
g_cnt2 = GA_Duplicate(g_cnt, "another shared counter");
if (!g_cnt2)
GA_Error("Another shared counter failed",1);
GA_Zero(g_cnt);
GA_Zero(g_cnt2);
#if DEBUG>1
/* initialize data in matrices a and b */
if(proc == 0)
printf("Initializing local buffers - a and b\n");
#endif
int w = 0;
int l = 7;
for(int i = 0; i < cdims[0]; i++) {
for(int j = 0; j < cdims[1]; j++) {
a[i*cdims[1] + j] = (double)(++w%29);
b[i*cdims[1] + j] = (double)(++l%37);
}
}
/* Copy data to global arrays g_a and g_b from local buffers */
next_p = NGA_Read_inc(g_cnt2,&rdcnt[proc],(long)1);
for (int i = 0; i < N; i+=cdims[0])
{
if (next_p == count_p) {
for (int j = 0; j < N; j+=cdims[1])
{
/* Indices of patch */
lo[0] = i;
lo[1] = j;
hi[0] = lo[0] + cdims[0];
hi[1] = lo[1] + cdims[1];
hi[0] = hi[0]-1;
hi[1] = hi[1]-1;
#if DEBUG>1
printf ("%d: PUT_GA_A_B: lo[0,1] = %d,%d and hi[0,1] = %d,%d\n",proc,lo[0],lo[1],hi[0],hi[1]);
#endif
NGA_Put(g_a, lo, hi, a, ld);
NGA_Put(g_b, lo, hi, b, ld);
}
next_p = NGA_Read_inc(g_cnt2,&rdcnt[proc],(long)1);
}
count_p++;
}
#if DEBUG>1
printf ("After NGA_PUT to global - A and B arrays\n");
#endif
/* Synchronize all processors to make sure puts from
nprocs has finished before proceeding with dgemm */
GA_Sync();
t1 = GA_Wtime();
next_gac = NGA_Read_inc(g_cnt,&rdcnt2[proc],(long)1);
for (int m = 0; m < N; m+=cdims[0])
{
for (int k = 0; k < N; k+=cdims[0])
{
if (next_gac == count_gac)
{
/* A = m x k */
lo[0] = m; lo[1] = k;
hi[0] = cdims[0] + lo[0]; hi[1] = cdims[1] + lo[1];
hi[0] = hi[0]-1; hi[1] = hi[1]-1;
#if DEBUG>3
printf ("%d: GET GA_A: lo[0,1] = %d,%d and hi[0,1] = %d,%d\n",proc,lo[0],lo[1],hi[0],hi[1]);
#endif
NGA_Get(g_a, lo, hi, a, ld);
for (int n = 0; n < N; n+=cdims[1])
{
memset (c, 0, sizeof(double) * cdims[0] * cdims[1]);
/* B = k x n */
lo[0] = k; lo[1] = n;
hi[0] = cdims[0] + lo[0]; hi[1] = cdims[1] + lo[1];
hi[0] = hi[0]-1; hi[1] = hi[1]-1;
#if DEBUG>3
printf ("%d: GET_GA_B: lo[0,1] = %d,%d and hi[0,1] = %d,%d\n",proc,lo[0],lo[1],hi[0],hi[1]);
#endif
NGA_Get(g_b, lo, hi, b, ld);
//_my_dgemm_ (a, local_N, b, local_N, c, local_N, local_N, local_N, local_N, alpha, beta=1.0);
/* TODO I am assuming square matrix blocks, further testing/work
required for rectangular matrices */
cblas_dgemm ( CblasRowMajor, CblasNoTrans, /* TransA */CblasNoTrans, /* TransB */
cdims[0] /* M */, cdims[1] /* N */, cdims[0] /* K */, alpha,
a, cdims[0], /* lda */ b, cdims[1], /* ldb */
beta=1.0, c, cdims[0] /* ldc */);
NGA_NbWait(&nbh);
/* C = m x n */
lo[0] = m; lo[1] = n;
hi[0] = cdims[0] + lo[0]; hi[1] = cdims[1] + lo[1];
hi[0] = hi[0]-1; hi[1] = hi[1]-1;
#if DEBUG>3
printf ("%d: ACC_GA_C: lo[0,1] = %d,%d and hi[0,1] = %d,%d\n",proc,lo[0],lo[1],hi[0],hi[1]);
#endif
NGA_NbAcc(g_c, lo, hi, c, ld, &alpha, &nbh);
count_acc += 1;
} /* END LOOP N */
next_gac = NGA_Read_inc(g_cnt,&rdcnt2[proc],(long)1);
} /* ENDIF if count == next */
count_gac++;
} /* END LOOP K */
} /* END LOOP M */
GA_Sync();
t2 = GA_Wtime();
seconds = t2 - t1;
if (proc == 0)
printf("Time taken for MM (secs):%lf \n", seconds);
printf("Number of ACC: %d\n", count_acc);
/* Correctness test - modify data again before this function */
for (int i = 0; i < NDIMS; i++) {
lo[i] = 0;
hi[i] = dims[i]-1;
ld[i] = dims[i];
}
verify(g_a, g_b, g_c, lo, hi, ld, N);
/* Clear local buffers */
free(a);
free(b);
free(c);
free(rdcnt);
free(rdcnt2);
GA_Sync();
/* Deallocate arrays */
GA_Destroy(g_a);
GA_Destroy(g_b);
GA_Destroy(g_c);
GA_Destroy(g_cnt);
GA_Destroy(g_cnt2);
}
/*
* test ga_dgemm
* Note: - change nummax for large arrays
* - turn off "dgemm_verify" for large arrays due to memory
* limitations, as dgemm_verify=1 for large arrays produces
* segfault, dumps core,or any crap.
*/
int main(int argc, char **argv)
{
int num_m;
int num_n;
int num_k;
int i;
int ii;
double *h0;
int g_c;
int g_b;
int g_a;
double a;
double t1;
double mf;
double avg_t[ntrans];
double avg_mf[ntrans];
int itime;
int ntimes;
int nums_m[/*howmany*/] = {512,1024};
int nums_n[/*howmany*/] = {512,1024};
int nums_k[/*howmany*/] = {512,1024};
char transa[/*ntrans*/] = "ntnt";
char transb[/*ntrans*/] = "nntt";
char ta;
char tb;
double *tmpa;
double *tmpb;
double *tmpc;
int ndim;
int dims[2];
#ifdef BLOCK_CYCLIC
int block_size[2];
#endif
#if defined(USE_ELEMENTAL)
// initialize Elemental (which will initialize MPI)
ElInitialize( &argc, &argv );
ElMPICommRank( MPI_COMM_WORLD, &me );
ElMPICommSize( MPI_COMM_WORLD, &nproc );
// instantiate el::global array
ElGlobalArraysConstruct_d( &eldga );
// initialize global arrays
ElGlobalArraysInitialize_d( eldga );
#else
MP_INIT(argc,argv);
if (!MA_init(MT_DBL,1,20000000)) {
GA_Error("failed: ma_init(MT_DBL,1,20000000)",10);
}
GA_INIT(argc,argv);
me = GA_Nodeid();
#endif
h0 = (double*)malloc(sizeof(double) * nummax*nummax);
tmpa = (double*)malloc(sizeof(double) * nummax*nummax);
tmpb = (double*)malloc(sizeof(double) * nummax*nummax);
tmpc = (double*)malloc(sizeof(double) * nummax*nummax);
ii = 0;
for (i=0; i<nummax*nummax; i++) {
ii = ii + 1;
if (ii > nummax) {
ii = 0;
}
h0[i] = ii;
}
/* Compute times assuming 500 mflops and 5 second target time */
/* ntimes = max(3.0d0,5.0d0/(4.0d-9*num**3)); */
ntimes = 5;
for (ii=0; ii<howmany; ii++) {
num_m = nums_m[ii];
num_n = nums_n[ii];
num_k = nums_k[ii];
a = 0.5/(num_m*num_n);
if (num_m > nummax || num_n > nummax || num_k > nummax) {
GA_Error("Insufficient memory: check nummax", 1);
}
#ifndef BLOCK_CYCLIC
ndim = 2;
/*
dims[0] = num_m;
dims[1] = num_n;
*/
dims[1] = num_m;
dims[0] = num_n;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_c", NULL, &g_c );
#else
if (!((g_c = NGA_Create(MT_DBL,ndim,dims,"g_c",NULL)))) {
GA_Error("failed: create g_c",20);
}
#endif
/*
dims[0] = num_k;
dims[1] = num_n;
*/
dims[1] = num_k;
dims[0] = num_n;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_b", NULL, &g_b );
#else
if (!((g_b = NGA_Create(MT_DBL,ndim,dims,"g_b",NULL)))) {
GA_Error("failed: create g_b",30);
}
#endif
/*
dims[0] = num_m;
dims[1] = num_k;
*/
dims[1] = num_m;
dims[0] = num_k;
#if defined(USE_ELEMENTAL)
ElGlobalArraysCreate_d( eldga, ndim, dims, "g_a", NULL, &g_a );
#else
if (!((g_a = NGA_Create(MT_DBL,ndim,dims,"g_a",NULL)))) {
GA_Error("failed: create g_a",40);
}
#endif
#else
ndim = 2;
block_size[0] = 128;
block_size[1] = 128;
dims[0] = num_m;
dims[1] = num_n;
g_c = GA_Create_handle();
GA_Set_data(g_c,ndim,dims,MT_DBL);
GA_Set_array_name(g_c,"g_c");
GA_Set_block_cyclic(g_c,block_size);
if (!GA_Allocate(g_c)) {
GA_Error("failed: create g_c",40);
}
dims[0] = num_k;
dims[1] = num_n;
g_b = GA_Create_handle();
GA_Set_data(g_b,ndim,dims,MT_DBL);
GA_Set_array_name(g_b,"g_b");
GA_Set_block_cyclic(g_b,block_size);
if (!ga_allocate(g_b)) {
GA_Error("failed: create g_b",40);
}
dims[0] = num_m;
dims[1] = num_k;
g_a = GA_Create_handle();
GA_Set_data(g_a,ndim,dims,MT_DBL);
GA_Set_array_name(g_a,"g_a");
GA_Set_block_cyclic(g_a,block_size);
if (!ga_allocate(g_a)) {
GA_Error('failed: create g_a',40);
}
#endif
/* Initialize matrices A and B */
if (me == 0) {
load_ga(g_a, h0, num_m, num_k);
load_ga(g_b, h0, num_k, num_n);
}
#if defined(USE_ELEMENTAL)
double zero = 0.0;
ElGlobalArraysFill_d( eldga, g_c, &zero );
ElGlobalArraysSync_d( eldga );
#else
GA_Zero(g_c);
GA_Sync();
#endif
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("\nMatrix Multiplication on C = A[%ld,%ld]xB[%ld,%ld]\n",
(long)num_m, (long)num_k, (long)num_k, (long)num_n);
fflush(stdout);
}
for (i=0; i<ntrans; i++) {
avg_t[i] = 0.0;
avg_mf[i] = 0.0;
}
for (itime=0; itime<ntimes; itime++) {
for (i=0; i<ntrans; i++) {
#if defined(USE_ELEMENTAL)
ElGlobalArraysSync_d( eldga );
#else
GA_Sync();
#endif
ta = transa[i];
tb = transb[i];
t1 = MP_TIMER();
#if defined(USE_ELEMENTAL)
ElGlobalArraysDgemm_d( eldga, ta, tb, num_m, num_n, num_k, 1.0, g_a, g_b, 0.0, g_c );
#else
GA_Dgemm(ta,tb,num_m,num_n,num_k,1.0, g_a, g_b, 0.0, g_c);
#endif
t1 = MP_TIMER() - t1;
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
#if defined(USE_ELEMENTAL)
mf = 2e0*num_m*num_n*num_k/t1*1e-6/nproc;
#else
mf = 2e0*num_m*num_n*num_k/t1*1e-6/GA_Nnodes();
#endif
avg_t[i] = avg_t[i]+t1;
avg_mf[i] = avg_mf[i] + mf;
printf("%15s%2d: %12.4f seconds %12.1f mflops/proc %c %c\n",
"Run#", itime, t1, mf, ta, tb);
fflush(stdout);
if (dgemm_verify && itime == 0) {
/* recall the C API swaps the matrix order */
/* we swap it here for the Fortran-based verify */
verify_ga_dgemm(tb, ta, num_n, num_m, num_k, 1.0,
g_b, g_a, 0.0, g_c, tmpb, tmpa, tmpc);
}
}
}
}
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("\n");
for (i=0; i<ntrans; i++) {
printf("%17s: %12.4f seconds %12.1f mflops/proc %c %c\n",
"Average", avg_t[i]/ntimes, avg_mf[i]/ntimes,
transa[i], transb[i]);
}
if(dgemm_verify) {
printf("All GA_Dgemms are verified...O.K.\n");
}
fflush(stdout);
}
/*
GA_Print(g_a);
GA_Print(g_b);
GA_Print(g_c);
*/
#if defined(USE_ELEMENTAL)
ElGlobalArraysDestroy_d( eldga, g_a );
ElGlobalArraysDestroy_d( eldga, g_b );
ElGlobalArraysDestroy_d( eldga, g_c );
#else
GA_Destroy(g_c);
GA_Destroy(g_b);
GA_Destroy(g_a);
#endif
}
/* ???
format(a15, i2, ': ', e12.4, ' seconds ',f12.1,
. ' mflops/proc ', 3a2)
*/
#if defined(USE_ELEMENTAL)
if (me == 0) {
#else
if (GA_Nodeid() == 0) {
#endif
printf("All tests successful\n");
}
free(h0);
free(tmpa);
free(tmpb);
free(tmpc);
#if defined(USE_ELEMENTAL)
// call el::global arrays destructor
ElGlobalArraysTerminate_d( eldga );
ElGlobalArraysDestruct_d( eldga );
ElFinalize();
#else
GA_Terminate();
MP_FINALIZE();
#endif
return 0;
}
/*
* Verify for correctness. Process 0 computes BLAS dgemm
* locally. For larger arrays, disbale this test as memory
* might not be sufficient
*/
void verify_ga_dgemm(char xt1, char xt2, int num_m, int num_n, int num_k,
double alpha, int g_a, int g_b, double beta, int g_c,
double *tmpa, double *tmpb, double *tmpc)
{
int i,j,type,ndim,dims[2],lo[2],hi[2];
double abs_value;
for (i=0; i<num_n; i++) {
for (j=0; j<num_m; j++) {
tmpc[j+i*num_m] = -1.0;
tmpa[j+i*num_m] = -2.0;
}
}
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_a, lo, hi, tmpa, &dims[1] );
#else
NGA_Get(g_a, lo, hi, tmpa, &dims[1]);
#endif
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_b, lo, hi, tmpb, &dims[1] );
#else
NGA_Get(g_b, lo, hi, tmpb, &dims[1]);
#endif
/* compute dgemm sequentially */
#if defined(USE_ELEMENTAL)
cblas_dgemm ( CblasRowMajor, ( xt1 == 'n'? CblasNoTrans: CblasTrans ),
( xt2 == 'n'? CblasNoTrans: CblasTrans ),
num_m /* M */, num_n /* N */, num_k /* K */,
alpha, tmpa, num_m, /* lda */
tmpb, num_k, /* ldb */ beta,
tmpc, num_m /* ldc */);
#else
xb_dgemm(&xt1, &xt2, &num_m, &num_n, &num_k,
&alpha, tmpa, &num_m,
tmpb, &num_k, &beta,
tmpc, &num_m);
#endif
/* after computing c locally, verify it with the values in g_c */
#if defined(USE_ELEMENTAL)
ElGlobalArraysInquire_d( eldga, g_a, &ndim, dims );
#else
NGA_Inquire(g_a, &type, &ndim, dims);
#endif
lo[0] = 0;
lo[1] = 0;
hi[0] = dims[0]-1;
hi[1] = dims[1]-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysGet_d( eldga, g_c, lo, hi, tmpa, &dims[1] );
#else
NGA_Get(g_c, lo, hi, tmpa, &dims[1]);
#endif
for (i=0; i<num_n; i++) {
for (j=0; j<num_m; j++) {
abs_value = fabs(tmpc[j+i*num_m]-tmpa[j+i*num_m]);
if(abs_value > 1.0 || abs_value < -1.0) {
printf("Values are = %f %f\n",
tmpc[j+i*num_m], tmpa[j+i*num_m]);
printf("Values are = %f %f\n",
fabs(tmpc[j+i*num_m]-tmpa[j*i*num_m]), abs_value);
fflush(stdout);
GA_Error("verify ga_dgemm failed", 1);
}
}
}
}
/**
* called by process '0' (or your master process )
*/
void load_ga(int handle, double *f, int dim1, int dim2)
{
int lo[2], hi[2];
if (dim1 < 0 || dim2 < 0) {
return;
}
lo[0] = 0;
lo[1] = 0;
hi[0] = dim1-1;
hi[1] = dim2-1;
#if defined(USE_ELEMENTAL)
ElGlobalArraysPut_d( eldga, handle, lo, hi, f, &dim1 );
#else
NGA_Put(handle, lo, hi, f, &dim1);
#endif
}