int main()
{
//Magma initialization
magma_init();
//Declaration of local variables
double *a, *b, *dev_a, results=0;
const int N=4098;
int i,j;
magma_int_t info=0, lda=N, ngpu=2;
//Memory Allocation Segment
magma_malloc_pinned((void**) &a,(N*N)*sizeof(double));
magma_malloc_pinned((void**) &b,(N*N)*sizeof(double));
//Generate two copies of Symmetric Positive Definite Matrix
for(i=0;i<N;i++)
{
for(j=0;j<i;j++)
{
a[i*N+j] = 1e-9* (double)rand();
a[j*N+i] = a[i*N+j];
b[i*N+j] = a[i*N+j];
b[j*N+i] = b[i*N+j];
}
a[i*N+i] = 1e-9*(double)rand() + 1000.0;
b[i*N+i] = a[i*N+i];
}
//Call custom Magma Cholesky for obtaining results
rr_dpotrf_m(ngpu,MagmaUpper,N,a,N,&info);
//Call Standard Magma Cholesky for result validation
magma_dpotrf(MagmaUpper,N,b,N,&info);
if(info != 0)
{
printf("magma_dpotrf original returned error %d: %s. \n",(int) info, magma_strerror(info));
}
//Validate the results; Compute the RMS error value.
for(i=0;i<N;i++)
for(j=0;j<N;j++)
results = results + (a[i*N+j] - b[i*N+j]) * (a[i*N+j] - b[i*N+j]);
//Display the results of the test
if(results < 1e-5)
printf("The two functions have identical results\n");
else
printf("The custom function had significant errors. The RMS value was %g\n",results);
magma_free_pinned(a);
magma_free_pinned(b);
magma_finalize();
return 0;
}
SEXP magQR(SEXP a)
{
SEXP gpu = GET_SLOT(a, install("gpu")),
b = PROTECT(NEW_OBJECT(MAKE_CLASS("magmaQR")));
int *DIMA = INTEGER(GET_DIM(a)), M = DIMA[0], N = DIMA[1],
MIN_MN = (M < N ? M : N), NB = magma_get_dgeqrf_nb(M), *pivot, info;
double *A, *tau;
A = REAL(SET_VECTOR_ELT(b, 0, AS_NUMERIC(duplicate(a))));
SET_VECTOR_ELT(b, 1, ScalarInteger(MIN_MN));
tau = REAL(SET_VECTOR_ELT(b, 2, NEW_NUMERIC(MIN_MN)));
pivot = INTEGER(SET_VECTOR_ELT(b, 3, NEW_INTEGER(N)));
int i;
for(i = 1; i <= N; i++) *pivot++ = i;
if(LOGICAL_VALUE(gpu)) {
int LENT = (2*MIN_MN + (N+31)/32*32)*NB;
double *dA, *dT, *work;
SET_SLOT(b, install("work"), NEW_NUMERIC(LENT));
work = REAL(GET_SLOT(b, install("work")));
magma_malloc((void**)&dA, (M*N)*sizeof(double));
magma_malloc((void**)&dT, LENT*sizeof(double));
magma_dsetmatrix(M, N, A, M, dA, M);
magma_dgeqrf3_gpu(M, N, dA, M, tau, dT, &info);
magma_dgetmatrix(M, N, dA, M, A, M);
magma_dgetvector(LENT, dT, 1, work, 1);
magma_free(dA);
magma_free(dT);
} else {
int LWORK = N * NB;
double *hA, *hwork;
magma_malloc_pinned((void**)&hA, (M*N)*sizeof(double));
magma_malloc_pinned((void**)&hwork, LWORK*sizeof(double));
lapackf77_dlacpy(MagmaUpperLowerStr, &M, &N, A, &M, hA, &M);
magma_dgeqrf_ooc(M, N, hA, M, tau, hwork, LWORK, &info);
lapackf77_dlacpy(MagmaUpperLowerStr, &M, &N, hA, &M, A, &M);
magma_free_pinned(hA);
magma_free_pinned(hwork);
}
if(info < 0) error("illegal argument %d in 'magQR'", -1 * info);
UNPROTECT(1);
return b;
}
K* hostmalloc(size_t size) {
K* ptr;
if(MAGMA_SUCCESS !=
magma_malloc_pinned( (void**)&ptr, size*sizeof(K) )) {
std::cerr << "MAGMA ERROR: magma_malloc_pinned failed\n";
exit(-1);
}
return ptr;
}
SEXP magChol(SEXP a)
{
SEXP gpu = GET_SLOT(a, install("gpu")),
b = PROTECT(NEW_OBJECT(MAKE_CLASS("magma")));
int *DIMA = INTEGER(GET_DIM(a)), N = DIMA[0], N2 = N * N, LDB = N, info;
double *B;
if(DIMA[1] != N) error("non-square matrix");
b = SET_SLOT(b, install(".Data"), AS_NUMERIC(a));
SET_SLOT(b, install("gpu"), duplicate(gpu));
B = REAL(b);
if(LOGICAL_VALUE(gpu)) {
double *dB;
magma_malloc((void**)&dB, N2*sizeof(double));
magma_dsetmatrix(N, N, B, LDB, dB, LDB);
magma_dpotrf_gpu(magma_uplo_const('U'), N, dB, LDB, &info);
magma_dgetmatrix(N, N, dB, LDB, B, LDB);
magma_free(dB);
} else {
double *hB;
magma_malloc_pinned((void**)&hB, N2*sizeof(double));
lapackf77_dlacpy(MagmaUpperStr, &N, &N, B, &LDB, hB, &LDB);
magma_dpotrf(magma_uplo_const('U'), N, hB, N, &info);
lapackf77_dlacpy(MagmaUpperStr, &N, &N, hB, &LDB, B, &LDB);
magma_free_pinned(hB);
}
if(info < 0) error("illegal argument %d in 'magChol", -1 * info);
else if(info > 0) error("leading minor of order %d is not positive definite", info);
int i, j;
for(j = 0; j < N; j++) {
for(i = j + 1; i < N; i++) {
B[i + j * N] = 0.0;
}
}
UNPROTECT(1);
return b;
}
SEXP magLU(SEXP a)
{
SEXP gpu = GET_SLOT(a, install("gpu")),
b = PROTECT(NEW_OBJECT(MAKE_CLASS("magmaLU")));
int *DIMA = INTEGER(GET_DIM(a)), M = DIMA[0], N = DIMA[1], LDA = M,
MIN_MN = M < N ? M : N, *ipiv, info;
double *A = REAL(PROTECT(AS_NUMERIC(a)));
b = SET_SLOT(b, install(".Data"), AS_NUMERIC(a));
SET_SLOT(b, install("pivot"), NEW_INTEGER(MIN_MN));
ipiv = INTEGER(GET_SLOT(b, install("pivot")));
SET_SLOT(b, install("gpu"), duplicate(gpu));
if(LOGICAL_VALUE(gpu)) {
double *dA;
magma_malloc((void**)&dA, (M*N)*sizeof(double));
magma_dsetmatrix(M, N, A, LDA, dA, LDA);
magma_dgetrf_gpu(M, N, dA, LDA, ipiv, &info);
magma_dgetmatrix(M, N, dA, LDA, REAL(b), LDA);
magma_free(dA);
} else {
double *hA;
magma_malloc_pinned((void**)&hA, (M*N)*sizeof(double));
lapackf77_dlacpy(MagmaUpperLowerStr, &M, &N, A, &LDA, hA, &LDA);
magma_dgetrf(M, N, hA, LDA, ipiv, &info);
lapackf77_dlacpy(MagmaUpperLowerStr, &M, &N, hA, &LDA, REAL(b), &LDA);
magma_free_pinned(hA);
}
if(info < 0) error("illegal argument %d in 'magLU'", -1 * info);
else if(info > 0) error("factor U is singular");
UNPROTECT(2);
return b;
}
SEXP magmaCholeskyFinal_m(SEXP A, SEXP n, SEXP NB, SEXP zeroTri, SEXP ngpu, SEXP lowerTri)
{
magma_init();
int ndevices;
double *h_R;
ndevices = INTEGER_VALUE(ngpu);
int idevice;
for(idevice=0; idevice < ndevices; idevice++)
{
magma_setdevice(idevice);
if(CUBLAS_STATUS_SUCCESS != cublasInit())
{
printf("Error: gpu %d: cublasInit failed\n", idevice);
magma_finalize();
exit(-1);
}
}
// magma_print_devices();
int In, INB;
In = INTEGER_VALUE(n);
INB = INTEGER_VALUE(NB);
double *PA = NUMERIC_POINTER(A);
int i,j;
//magma_timestr_t start, end;
double gpu_time;
printf("Inside magma_dpotrf_m");
/*for(i = 0; i < 5; i++)
{
for(j = 0; j < 5; j++)
{
printf("%.8f ", PA[i+j*In]);
}
printf("\n");
} */
magma_int_t N, status, info, nGPU, n2, lda;
clock_t t1, t2;
N = In;
status = 0;
int nGPUs = ndevices;
lda = N;
n2 = lda*N;
if ( MAGMA_SUCCESS != magma_malloc_pinned( (void**) &h_R, (n2)*sizeof(double) )) {
fprintf( stderr, "!!!! magma_malloc_pinned failed for: %s\n", h_R );
magma_finalize();
exit(-1);
}
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, PA, &lda, h_R, &lda );
//printf("Modified by Vinay in 2 GPU\n");
//INB = magma_get_dpotrf_nb(N);
// INB = 224;
// printf("INB = %d\n", INB);
//ngpu = ndevices;
// printf("ngpu = %d\n", ngpu);
//max_size = INB*(1+N/(INB*ndevices))*INB*((N+INB-1)/INB);
// printf("max_size = %d\n", max_size);
//int imax_size = max_size;
//double *dA;
//magma_dmalloc_pinned((void**)&dA, In*In*sizeof(double));
//ldda = (1+N/(INB*ndevices))*INB;
// printf("ldda = %d\n", ldda);
//magma_dsetmatrix_1D_row_bcyclic(N, N, PA, N, dA, ldda, ngpu, INB);
//magma_dpotrf_mgpu(ngpu, MagmaLower, N, dA, ldda, &info);
int lTri;
lTri = INTEGER_VALUE(lowerTri);
if(lTri){
t1 = clock();
magma_dpotrf_m(nGPUs, MagmaLower, N, h_R, N, &info);
t2 = clock ();
}
else{
t1 = clock();
magma_dpotrf_m(nGPUs, MagmaUpper, N, h_R, N, &info);
t2 = clock ();
}
gpu_time = (double) (t2-t1)/(CLOCKS_PER_SEC) ; // Magma time
printf (" magma_dpotrf_m time : %f sec. \n", gpu_time );
if(info != 0)
{
printf("magma_dpotrf returned error %d: %s.\n", (int) info, magma_strerror(info));
}
//magma_dgetmatrix_1D_row_bcyclic(N, N, dA, ldda, PA, N, ngpu, INB);
//for(dev = 0; dev < ndevices; dev++)
//{
//magma_setdevice(dev);
//cudaFree(dA[dev]);
//}
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_R, &lda, PA, &lda );
magma_free_pinned(h_R);
magma_finalize();
cublasShutdown();
int IZeroTri;
IZeroTri = INTEGER_VALUE(zeroTri);
if(IZeroTri & lTri) {
for(i = 1; i < In; i++)
{
for(j=0; j< i; j++)
{
PA[i*In+j] = 0.0;
}
}
}
else if(IZeroTri){
for(i = 0; i < In; i++)
{
for(j=i+1; j < In; j++)
{
PA[i*In+j] = 0.0;
}
}
}
return(R_NilValue);
}
SEXP magmaCholeskyFinal(SEXP A, SEXP n, SEXP NB, SEXP id, SEXP zeroTri, SEXP lowerTri)
{
magma_init();
// magma_print_devices();
double *h_R;
int In, INB, ID;
In = INTEGER_VALUE(n);
INB = INTEGER_VALUE(NB);
ID = INTEGER_VALUE(id);
double *PA = NUMERIC_POINTER(A);
int i,j;
magma_int_t N, n2, lda, status, info, max_size;
N=In;
lda = N;
n2 = lda*N;
/* for(i = 0; i < In; i++)
{
for(j = 0; j < In; j++)
{
printf("%.8f ", PA[i+j*In]);
}
printf("\n");
} */
if ( MAGMA_SUCCESS != magma_malloc_pinned( (void**) &h_R, (n2)*sizeof(double) )) {
fprintf( stderr, "!!!! magma_malloc_pinned failed for: %s\n", h_R );
magma_finalize();
exit(-1);
}
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, PA, &lda, h_R, &lda );
N = In;
status = 0;
magma_setdevice(ID);
//printf("Modified by Vinay in one GPU\n");
//INB = magma_get_dpotrf_nb(N);
// INB = 224;
// printf("INB = %d\n", INB);
//ngpu = ndevices;
// printf("ngpu = %d\n", ngpu);
//max_size = INB*(1+N/(INB*ndevices))*INB*((N+INB-1)/INB);
// printf("max_size = %d\n", max_size);
//int imax_size = max_size;
//double *dA;
//magma_dmalloc_pinned((void**)&dA, In*In*sizeof(double));
//ldda = (1+N/(INB*ndevices))*INB;
// printf("ldda = %d\n", ldda);
//magma_dsetmatrix_1D_row_bcyclic(N, N, PA, N, dA, ldda, ngpu, INB);
//magma_dpotrf_mgpu(ngpu, MagmaLower, N, dA, ldda, &info);
int lTri;
lTri = INTEGER_VALUE(lowerTri);
if(lTri)
magma_dpotrf(MagmaLower, N, h_R, N, &info);
else
magma_dpotrf(MagmaUpper, N, h_R, N, &info);
if(info != 0)
{
printf("magma_dpotrf returned error %d: %s.\n", (int) info, magma_strerror(info));
}
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_R, &lda, PA, &lda );
//magma_dgetmatrix_1D_row_bcyclic(N, N, dA, ldda, PA, N, ngpu, INB);
//for(dev = 0; dev < ndevices; dev++)
//{
//magma_setdevice(dev);
//cudaFree(dA[dev]);
//}
magma_free_pinned(h_R);
magma_finalize();
cublasShutdown();
/*
int IZeroTri;
IZeroTri = INTEGER_VALUE(zeroTri);
if(IZeroTri & lTri) {
for(i = 1; i < In; i++)
{
for(j=0; j< i; j++)
{
PA[i*In+j] = 0.0;
}
}
}
else if(IZeroTri)
for(i = 0; i < In; i++)
{
for(j=i+1; j < In; j++)
{
PA[i*In+j] = 0.0;
}
}*/
return(R_NilValue);
}