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
}
/* 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);
}
