/* input is matrix size */
void ga_lu(double *A, int matrix_size)
{
int g_a, g_b, dims[2], type=C_DBL;
int lo[2], hi[2], ld;
int block_size[2], proc_grid[2];
double time, gflops;
/* create a 2-d GA (global matrix) */
dims[0] = matrix_size;
dims[1] = matrix_size;
block_size[0] = BLOCK_SIZE;
block_size[1] = BLOCK_SIZE;
#ifdef USE_SCALAPACK_DISTR
proc_grid[0] = 2;
proc_grid[1] = nprocs/2;
if(nprocs%2) GA_Error("For ScaLAPACK stle distribution, nprocs must be "
" divisible by 2", 0);
#endif
#ifndef BLOCK_CYCLIC
g_a = NGA_Create(type, 2, dims, "A", NULL);
g_b = GA_Duplicate(g_a, "transposed array B");
#else
g_a = GA_Create_handle();
GA_Set_data(g_a, 2, dims, type);
GA_Set_array_name(g_a,"A");
# ifdef USE_SCALAPACK_DISTR
GA_Set_block_cyclic_proc_grid(g_a, block_size, proc_grid);
# else
GA_Set_block_cyclic(g_a, block_size);
# endif
GA_Allocate(g_a);
g_b = GA_Create_handle();
GA_Set_data(g_b, 2, dims, type);
GA_Set_array_name(g_b,"B");
# ifdef USE_SCALAPACK_DISTR
GA_Set_block_cyclic_proc_grid(g_b, block_size, proc_grid);
# else
GA_Set_block_cyclic(g_b, block_size);
# endif
GA_Allocate(g_b);
#endif
/* copy the local matrix into GA */
if(me==0)
{
lo[0] = 0;
hi[0] = matrix_size - 1;
lo[1] = 0;
hi[1] = matrix_size - 1;
ld = matrix_size;
NGA_Put(g_a, lo, hi, A, &ld);
}
GA_Sync();
GA_Transpose(g_a, g_b);
time = CLOCK_();
GA_Lu('n', g_b);
time = CLOCK_() - time;
/* 2/3 N^3 - 1/2 N^2 flops for LU and 2*N^2 for solver */
gflops = ( (((double)matrix_size) * matrix_size)/(time*1.0e+9) *
(2.0/3.0 * (double)matrix_size - 0.5) );
if(me==0) printf("\nGA_Lu: N=%d flops=%2.5e Gflops, time=%2.5e secs\n\n",
matrix_size, gflops, time);
#if DEBUG
GA_Print(g_a);
GA_Print(g_b);
#endif
/* if(me==0) lu(A, matrix_size); */
GA_Destroy(g_a);
GA_Destroy(g_b);
}
/*
* 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
}
