void magma_dmake_hpd( magma_int_t N, double* A, magma_int_t lda )
{
magma_int_t i, j;
for( i=0; i<N; ++i ) {
A(i,i) = MAGMA_D_MAKE( MAGMA_D_REAL( A(i,i) ) + N, 0. );
for( j=0; j<i; ++j ) {
A(j,i) = MAGMA_D_CNJG( A(i,j) );
}
}
}
void init_matrix( magma_int_t N, double *h_A, magma_int_t lda )
{
magma_int_t ione = 1, n2 = N*lda;
magma_int_t ISEED[4] = {0,0,0,1};
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
/* Symmetrize and increase the diagonal */
for (magma_int_t i = 0; i < N; ++i) {
h_A(i,i) = MAGMA_D_MAKE( MAGMA_D_REAL(h_A(i,i)) + N, 0 );
for (magma_int_t j = 0; j < i; ++j)
h_A(i, j) = MAGMA_D_CNJG( h_A(j, i) );
}
}
extern "C" void
magma_dtrdtype2cbHLsym_withQ_v2(magma_int_t n, magma_int_t nb,
double *A, magma_int_t lda,
double *V, magma_int_t ldv,
double *TAU,
magma_int_t st, magma_int_t ed,
magma_int_t sweep, magma_int_t Vblksiz,
double *work)
{
/*
WORK (workspace) double real array, dimension NB
*/
magma_int_t ione = 1;
magma_int_t vpos, taupos;
double conjtmp;
double c_one = MAGMA_D_ONE;
magma_int_t ldx = lda-1;
magma_int_t len = ed - st + 1;
magma_int_t lem = min(ed+nb, n) - ed;
if (lem > 0) {
magma_bulge_findVTAUpos(n, nb, Vblksiz, sweep-1, st-1, ldv, &vpos, &taupos);
/* apply remaining right coming from the top block */
lapackf77_dlarfx("R", &lem, &len, V(vpos), TAU(taupos), A(ed+1, st), &ldx, work);
}
if (lem > 1) {
magma_bulge_findVTAUpos(n, nb, Vblksiz, sweep-1, ed, ldv, &vpos, &taupos);
/* remove the first column of the created bulge */
*V(vpos) = c_one;
//memcpy(V(vpos+1), A(ed+2, st), (lem-1)*sizeof(double));
cblas_dcopy(lem-1, A(ed+2, st), ione, V(vpos+1), ione);
memset(A(ed+2, st),0,(lem-1)*sizeof(double));
/* Eliminate the col at st */
lapackf77_dlarfg( &lem, A(ed+1, st), V(vpos+1), &ione, TAU(taupos) );
/* apply left on A(J1:J2,st+1:ed) */
len = len-1; /* because we start at col st+1 instead of st. col st is the col that has been removed; */
conjtmp = MAGMA_D_CNJG(*TAU(taupos));
lapackf77_dlarfx("L", &lem, &len, V(vpos), &conjtmp, A(ed+1, st+1), &ldx, work);
}
}
extern "C" void
magma_dtrdtype2cbHLsym_withQ(
magma_int_t N, magma_int_t NB,
double *A, magma_int_t LDA,
double *V, double *TAU,
magma_int_t st, magma_int_t ed, magma_int_t sweep, magma_int_t Vblksiz)
{
magma_int_t J1, J2, len, lem, LDX;
//magma_int_t i, j;
magma_int_t IONE=1;
magma_int_t blkid, vpos, taupos, tpos;
double conjtmp;
double Z_ONE = MAGMA_D_ONE;
//double WORK[NB];
double *WORK;
magma_dmalloc_cpu( &WORK, NB );
findVTpos(N,NB,Vblksiz,sweep-1,st-1, &vpos, &taupos, &tpos, &blkid);
LDX = LDA-1;
J1 = ed+1;
J2 = min(ed+NB,N);
len = ed-st+1;
lem = J2-J1+1;
if (lem > 0) {
/* apply remaining right commming from the top block */
lapackf77_dlarfx("R", &lem, &len, V(vpos), TAU(taupos), A(J1, st), &LDX, WORK);
}
if (lem > 1) {
findVTpos(N,NB,Vblksiz,sweep-1,J1-1, &vpos, &taupos, &tpos, &blkid);
/* remove the first column of the created bulge */
*V(vpos) = Z_ONE;
memcpy(V(vpos+1), A(J1+1, st), (lem-1)*sizeof(double));
memset(A(J1+1, st),0,(lem-1)*sizeof(double));
/* Eliminate the col at st */
lapackf77_dlarfg( &lem, A(J1, st), V(vpos+1), &IONE, TAU(taupos) );
/* apply left on A(J1:J2,st+1:ed) */
len = len-1; /* because we start at col st+1 instead of st. col st is the col that has been revomved; */
conjtmp = MAGMA_D_CNJG(*TAU(taupos));
lapackf77_dlarfx("L", &lem, &len, V(vpos), &conjtmp, A(J1, st+1), &LDX, WORK);
}
magma_free_cpu(WORK);
}
inline static void
magma_dlarfxsym_v2(magma_int_t n,
double *A, magma_int_t lda,
double *V, double *TAU,
double *work)
{
/*
WORK (workspace) double real array, dimension N
*/
magma_int_t ione = 1;
double dtmp;
double c_zero = MAGMA_D_ZERO;
double c_neg_one= MAGMA_D_NEG_ONE;
double c_half = MAGMA_D_HALF;
/* X = AVtau */
blasf77_dsymv("L",&n, TAU, A, &lda, V, &ione, &c_zero, work, &ione);
/* compute dtmp= X'*V */
#if defined(PRECISION_z) || defined(PRECISION_c)
dtmp = c_zero;
for (magma_int_t j = 0; j < n; j++)
dtmp = dtmp + MAGMA_D_CNJG(work[j]) * V[j];
//cblas_ddot_sub(n, work, ione, V, ione, &dtmp);
#else
dtmp = cblas_ddot(n, work, ione, V, ione);
#endif
/* compute 1/2 X'*V*t = 1/2*dtmp*tau */
dtmp = -dtmp * c_half * (*TAU);
/* compute W=X-1/2VX'Vt = X - dtmp*V */
blasf77_daxpy(&n, &dtmp, V, &ione, work, &ione);
/* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */
blasf77_dsyr2("L", &n, &c_neg_one, work, &ione, V, &ione, A, &lda);
}
extern "C" void
magma_dlarfxsym(magma_int_t N, double *A, magma_int_t LDA, double *V, double *TAU) {
magma_int_t IONE=1;
double dtmp;
double Z_ZERO = MAGMA_D_ZERO;
//double Z_ONE = MAGMA_D_ONE;
double Z_MONE = MAGMA_D_NEG_ONE;
double Z_HALF = MAGMA_D_HALF;
//double WORK[N];
double *WORK;
magma_dmalloc_cpu( &WORK, N );
/* apply left and right on A(st:ed,st:ed)*/
//magma_dlarfxsym(len,A(st,st),LDX,V(st),TAU(st));
/* X = AVtau */
blasf77_dsymv("L",&N, TAU, A, &LDA, V, &IONE, &Z_ZERO, WORK, &IONE);
/* je calcul dtmp= X'*V */
#if defined(PRECISION_z) || defined(PRECISION_c)
dtmp = Z_ZERO;
for (magma_int_t j = 0; j < N; j++)
dtmp = dtmp + MAGMA_D_CNJG(WORK[j]) * V[j];
//cblas_ddot_sub(N, WORK, IONE, V, IONE, &dtmp);
#else
dtmp = cblas_ddot(N, WORK, IONE, V, IONE);
#endif
/* je calcul 1/2 X'*V*t = 1/2*dtmp*tau */
dtmp = -dtmp * Z_HALF * (*TAU);
/* je calcul W=X-1/2VX'Vt = X - dtmp*V */
/*
for (j = 0; j < N; j++)
WORK[j] = WORK[j] + (dtmp*V[j]); */
blasf77_daxpy(&N, &dtmp, V, &IONE, WORK, &IONE);
/* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */
blasf77_dsyr2("L",&N,&Z_MONE,WORK,&IONE,V,&IONE,A,&LDA);
magma_free_cpu(WORK);
}
void magmablas_dsyr2k_mgpu_spec(
magma_uplo_t uplo, magma_trans_t trans, magma_int_t n, magma_int_t k,
double alpha, double *dA[], magma_int_t lda, magma_int_t aoffset,
double *dB[], magma_int_t ldb, magma_int_t boffset,
double beta, double *dC[], magma_int_t ldc, magma_int_t coffset,
magma_int_t ngpu, magma_int_t nb, magma_queue_t streams[][20], magma_int_t nstream )
{
#define dA(dev, i, j) (dA[dev] + (i) + (j)*lda + (aoffset) )
#define dB(dev, i, j) (dB[dev] + (i) + (j)*ldb + (boffset) )
#define dC(dev, i, j) (dC[dev] + (i) + (j)*ldc)
/* Check arguments */
magma_int_t info = 0;
if ( uplo != MagmaLower ) {
info = -1; // 'u' not yet handled
} else if ( trans != MagmaNoTrans ) {
info = -2; // 'c' not yet handled
} else if ( n < 0 ) {
info = -3;
} else if ( k < 0 ) {
info = -4;
} else if ( ((trans == MagmaNoTrans) && lda < max(1,n)) ||
((trans == MagmaTrans) && lda < max(1,k)) ) {
info = -7;
} else if ( aoffset < 0 || aoffset > lda ) {
info = -8;
} else if ( ((trans == MagmaNoTrans) && ldb < max(1,n)) ||
((trans == MagmaTrans) && ldb < max(1,k)) ) {
info = -10;
} else if ( boffset < 0 || boffset > ldb ) {
info = -11;
} else if ( ldc < max(1,n) ) {
info = -13;
} else if ( coffset < 0 || coffset > ldc ) {
info = -14;
} else if ( ngpu <= 0 ) {
info = -15;
} else if ( nb <= 0 ) {
info = -16;
} else if ( nstream <= 0 ) {
info = -18;
}
if ( info != 0 ) {
magma_xerbla( __func__, -(info) );
return;
}
const double c_one = MAGMA_D_ONE;
double cbeta = MAGMA_D_MAKE( beta, 0. );
magma_int_t ib, ioff, iblock, idev, di, s;
magma_device_t cdev;
magma_queue_t cqueue;
magma_getdevice( &cdev );
magmablasGetKernelStream( &cqueue );
// loop over all blocks
// Faster to have two loops: first loop does C_hat = alpha*A*B' + beta*C
// blockoffset is offset within first block; for subsequent blocks it is 0
magma_int_t blockoffset = coffset % nb;
for( magma_int_t i = 0; i < n; i += ib ) {
ib = min( nb-blockoffset, n-i ); // block size
ioff = i + coffset; // global index in parent matrix
iblock = (ioff / nb) / ngpu; // local block id
idev = (ioff / nb) % ngpu; // device with this block
di = iblock*nb + blockoffset; // local index in parent matrix
magma_setdevice( idev );
s = iblock % nstream;
magmablasSetKernelStream( streams[ idev ][ s ] );
// C[i:n,i] = alpha * A[i:n,0] * B[i,0]' + beta*C[i:n,i]
//printf( "dgemm n=%4d, ib=%4d, k=%4d, i=%4d\n", n-i, ib, k, i );
magma_dgemm( MagmaNoTrans, MagmaTrans, n, ib, k,
alpha, dA(idev,0,0), lda,
dB(idev,i,0), ldb,
cbeta, dC(idev,coffset,di), ldc );
blockoffset = 0;
}
// second loop does C = (alpha)*B*A' + C_hat
alpha = MAGMA_D_CNJG( alpha );
blockoffset = coffset % nb;
for( magma_int_t i = 0; i < n; i += ib ) {
ib = min( nb-blockoffset, n-i ); // block size
ioff = i + coffset; // global index in parent matrix
iblock = (ioff / nb) / ngpu; // local block id
idev = (ioff / nb) % ngpu; // device with this block
di = iblock*nb + blockoffset; // local index in parent matrix
magma_setdevice( idev );
s = iblock % nstream;
magmablasSetKernelStream( streams[ idev ][ s ] );
// C[i:n,i] += (alpha) * B[i:n,0] * A[i,0]'
//printf( "dgemm n=%4d, ib=%4d, k=%4d, i=%4d\n", n-i, ib, k, i );
magma_dgemm( MagmaNoTrans, MagmaTrans, n, ib, k,
alpha, dB(idev,0,0), ldb,
dA(idev,i,0), lda,
c_one, dC(idev,coffset,di), ldc );
blockoffset = 0;
}
magma_setdevice( cdev );
magmablasSetKernelStream( cqueue );
}
int main( int argc, char** argv)
{
real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time;
double *hA, *hR;
magmaDouble_ptr dA;
magma_int_t N = 0, n2, lda, ldda;
magma_int_t size[10] =
{ 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 8160, 8192 };
magma_int_t i, info;
double mz_one = MAGMA_D_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double work[1], matnorm, diffnorm;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
}
if (N>0) size[0] = size[9] = N;
else exit(1);
}
else {
printf("\nUsage: \n");
printf(" testing_dpotrf2_gpu -N %d\n\n", 1024);
}
/* Initialize */
magma_queue_t queue1, queue2;
magma_device_t device;
int num = 0;
magma_err_t err;
magma_init();
err = magma_get_devices( &device, 2, &num );
if ( err != 0 or num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device, &queue1 );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device, &queue2 );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
magma_queue_t queues[2] = {queue1, queue2};
/* Allocate memory for the largest matrix */
N = size[9];
n2 = N * N;
ldda = ((N+31)/32) * 32;
TESTING_MALLOC_CPU( hA, double, n2 );
TESTING_MALLOC_PIN( hR, double, n2 );
TESTING_MALLOC_DEV( dA, double, ldda*N );
printf("\n\n");
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R_magma-R_lapack||_F / ||R_lapack||_F\n");
printf("========================================================================================\n");
for(i=0; i<10; i++){
N = size[i];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
gflops = FLOPS( (double)N ) * 1e-9;
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, hA );
/* Symmetrize and increase the diagonal */
for( int i = 0; i < N; ++i ) {
MAGMA_D_SET2REAL( hA(i,i), MAGMA_D_REAL(hA(i,i)) + N );
for( int j = 0; j < i; ++j ) {
hA(i, j) = MAGMA_D_CNJG( hA(j,i) );
}
}
lapackf77_dlacpy( MagmaFullStr, &N, &N, hA, &lda, hR, &lda );
/* Warm up to measure the performance */
magma_dsetmatrix( N, N, hA, 0, lda, dA, 0, ldda, queue1);
clFinish(queue1);
magma_dpotrf2_gpu( MagmaLower, N, dA, 0, ldda, &info, queues );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_dsetmatrix( N, N, hA, 0, lda, dA, 0, ldda, queue1 );
clFinish(queue1);
gpu_time = magma_wtime();
magma_dpotrf2_gpu( MagmaLower, N, dA, 0, ldda, &info, queues );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf( "magma_dpotrf2 had error %d.\n", info );
gpu_perf = gflops / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
lapackf77_dpotrf( MagmaLowerStr, &N, hA, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf( "lapackf77_dpotrf had error %d.\n", info );
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result compared to LAPACK
|R_magma - R_lapack| / |R_lapack|
=================================================================== */
magma_dgetmatrix( N, N, dA, 0, ldda, hR, 0, lda, queue1 );
matnorm = lapackf77_dlange("f", &N, &N, hA, &lda, work);
blasf77_daxpy(&n2, &mz_one, hA, &ione, hR, &ione);
diffnorm = lapackf77_dlange("f", &N, &N, hR, &lda, work);
printf( "%5d %6.2f (%6.2f) %6.2f (%6.2f) %e\n",
N, cpu_perf, cpu_time, gpu_perf, gpu_time, diffnorm / matnorm );
if (argc != 1)
break;
}
/* clean up */
TESTING_FREE_CPU( hA );
TESTING_FREE_PIN( hR );
TESTING_FREE_DEV( dA );
magma_queue_destroy( queue1 );
magma_queue_destroy( queue2 );
magma_finalize();
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dsymmetrize
Code is very similar to testing_dtranspose.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
double error, work[1];
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_R;
magmaDouble_ptr d_A;
magma_int_t N, nb, size, lda, ldda, mstride, nstride, ntile;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
nb = (opts.nb == 0 ? 64 : opts.nb);
mstride = 2*nb;
nstride = 3*nb;
printf("uplo = %s, nb = %d, mstride = %d, nstride = %d\n",
lapack_uplo_const(opts.uplo), (int) nb, (int) mstride, (int) nstride );
printf(" N ntile CPU GByte/s (ms) GPU GByte/s (ms) check\n");
printf("===========================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = ((N+31)/32)*32;
size = lda*N;
if ( N < nb ) {
ntile = 0;
} else {
ntile = min( (N - nb)/mstride + 1,
(N - nb)/nstride + 1 );
}
// load each tile, save each tile
gbytes = sizeof(double) * 2.*nb*nb*ntile / 1e9;
TESTING_MALLOC_CPU( h_A, double, size );
TESTING_MALLOC_CPU( h_R, double, size );
TESTING_MALLOC_DEV( d_A, double, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < N; ++i ) {
h_A[i + j*lda] = MAGMA_D_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_dsetmatrix( N, N, h_A, lda, d_A, ldda );
gpu_time = magma_sync_wtime( 0 );
magmablas_dsymmetrize_tiles( opts.uplo, nb, d_A, ldda, ntile, mstride, nstride );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using naive in-place algorithm
(LAPACK doesn't implement symmetrize)
=================================================================== */
cpu_time = magma_wtime();
for( int tile = 0; tile < ntile; ++tile ) {
int offset = tile*mstride + tile*nstride*lda;
for( int j = 0; j < nb; ++j ) {
for( int i = 0; i < j; ++i ) {
if ( opts.uplo == MagmaLower ) {
h_A[offset + i + j*lda] = MAGMA_D_CNJG( h_A[offset + j + i*lda] );
}
else {
h_A[offset + j + i*lda] = MAGMA_D_CNJG( h_A[offset + i + j*lda] );
}
}
}
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_dgetmatrix( N, N, d_A, ldda, h_R, lda );
blasf77_daxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_dlange("f", &N, &N, h_R, &lda, work);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
(int) N, (int) ntile,
cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(error == 0. ? "ok" : "failed") );
status += ! (error == 0.);
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
int main( int argc, char** argv)
{
real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time;
double *h_A, *h_R;
magmaDouble_ptr d_lA[MagmaMaxGPUs];
magma_int_t N = 0, n2, lda, ldda;
magma_int_t size[10] =
{ 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10000 };
magma_int_t i, j, k, info;
double mz_one = MAGMA_D_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
double work[1], matnorm, diffnorm;
magma_int_t num_gpus0 = 1, num_gpus, flag = 0;
int nb, mb, n_local, nk;
magma_uplo_t uplo = MagmaLower;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0){
N = atoi(argv[++i]);
if (N>0) {
size[0] = size[9] = N;
flag = 1;
}else exit(1);
}
if(strcmp("-NGPU", argv[i])==0)
num_gpus0 = atoi(argv[++i]);
if(strcmp("-UPLO", argv[i])==0){
if(strcmp("L", argv[++i])==0){
uplo = MagmaLower;
}else{
uplo = MagmaUpper;
}
}
}
}
else {
printf("\nUsage: \n");
printf(" testing_dpotrf_mgpu -N %d -NGPU %d -UPLO -L\n\n", 1024, num_gpus0);
}
/* looking for max. ldda */
ldda = 0;
n2 = 0;
for(i=0;i<10;i++){
N = size[i];
nb = magma_get_dpotrf_nb(N);
mb = nb;
if(num_gpus0 > N/nb){
num_gpus = N/nb;
if(N%nb != 0) num_gpus ++;
}else{
num_gpus = num_gpus0;
}
n_local = nb*(1+N/(nb*num_gpus))*mb*((N+mb-1)/mb);
if(n_local > ldda) ldda = n_local;
if(n2 < N*N) n2 = N*N;
if(flag != 0) break;
}
/* Allocate host memory for the matrix */
TESTING_MALLOC_PIN( h_A, double, n2 );
TESTING_MALLOC_PIN( h_R, double, n2 );
/* Initialize */
magma_queue_t queues[MagmaMaxGPUs * 2];
//magma_queue_t queues[MagmaMaxGPUs];
magma_device_t devices[ MagmaMaxGPUs ];
magma_int_t num = 0;
magma_int_t err;
magma_init();
err = magma_getdevices( devices, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_getdevices failed: %d\n", (int) err );
exit(-1);
}
for(i=0;i<num_gpus;i++){
err = magma_queue_create( devices[i], &queues[2*i] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", (int) err );
exit(-1);
}
err = magma_queue_create( devices[i], &queues[2*i+1] );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", (int) err );
exit(-1);
}
}
printf("each buffer size: %d\n", ldda);
/* allocate local matrix on Buffers */
for(i=0; i<num_gpus0; i++){
TESTING_MALLOC_DEV( d_lA[i], double, ldda );
}
printf("\n\n");
printf("Using GPUs: %d\n", num_gpus0);
if(uplo == MagmaUpper){
printf("\n testing_dpotrf_mgpu -N %d -NGPU %d -UPLO U\n\n", N, num_gpus0);
}else{
printf("\n testing_dpotrf_mgpu -N %d -NGPU %d -UPLO L\n\n", N, num_gpus0);
}
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R_magma-R_lapack||_F / ||R_lapack||_F\n");
printf("========================================================================================\n");
for(i=0; i<10; i++){
N = size[i];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
gflops = FLOPS( (double)N ) * 1e-9;
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
/* Symmetrize and increase the diagonal */
for( int i = 0; i < N; ++i ) {
h_A(i,i) = MAGMA_D_MAKE( MAGMA_D_REAL(h_A(i,i)) + N, 0 );
for( int j = 0; j < i; ++j ) {
h_A(i, j) = MAGMA_D_CNJG( h_A(j,i) );
}
}
lapackf77_dlacpy( MagmaFullStr, &N, &N, h_A, &lda, h_R, &lda );
/* Warm up to measure the performance */
nb = magma_get_dpotrf_nb(N);
if(num_gpus0 > N/nb){
num_gpus = N/nb;
if(N%nb != 0) num_gpus ++;
printf("too many GPUs for the matrix size, using %d GPUs\n", (int)num_gpus);
}else{
num_gpus = num_gpus0;
}
/* distribute matrix to gpus */
if(uplo == MagmaUpper){
// Upper
ldda = ((N+mb-1)/mb)*mb;
for(j=0;j<N;j+=nb){
k = (j/nb)%num_gpus;
nk = min(nb, N-j);
magma_dsetmatrix( N, nk,
&h_A[j*lda], lda,
d_lA[k], j/(nb*num_gpus)*nb*ldda, ldda,
queues[2*k]);
}
}else{
// Lower
ldda = (1+N/(nb*num_gpus))*nb;
for(j=0;j<N;j+=nb){
k = (j/nb)%num_gpus;
nk = min(nb, N-j);
magma_dsetmatrix( nk, N, &h_A[j], lda,
d_lA[k], (j/(nb*num_gpus)*nb), ldda,
queues[2*k]);
}
}
magma_dpotrf_mgpu( num_gpus, uplo, N, d_lA, 0, ldda, queues, &info );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
/* distribute matrix to gpus */
if(uplo == MagmaUpper){
// Upper
ldda = ((N+mb-1)/mb)*mb;
for(j=0;j<N;j+=nb){
k = (j/nb)%num_gpus;
nk = min(nb, N-j);
magma_dsetmatrix( N, nk,
&h_A[j*lda], lda,
d_lA[k], j/(nb*num_gpus)*nb*ldda, ldda,
queues[2*k]);
}
}else{
// Lower
ldda = (1+N/(nb*num_gpus))*nb;
for(j=0;j<N;j+=nb){
k = (j/nb)%num_gpus;
nk = min(nb, N-j);
magma_dsetmatrix( nk, N, &h_A[j], lda,
d_lA[k], (j/(nb*num_gpus)*nb), ldda,
queues[2*k]);
}
}
gpu_time = magma_wtime();
magma_dpotrf_mgpu( num_gpus, uplo, N, d_lA, 0, ldda, queues, &info );
gpu_time = magma_wtime() - gpu_time;
if (info != 0)
printf( "magma_dpotrf had error %d.\n", info );
gpu_perf = gflops / gpu_time;
/* gather matrix from gpus */
if(uplo==MagmaUpper){
// Upper
for(j=0;j<N;j+=nb){
k = (j/nb)%num_gpus;
nk = min(nb, N-j);
magma_dgetmatrix( N, nk,
d_lA[k], j/(nb*num_gpus)*nb*ldda, ldda,
&h_R[j*lda], lda, queues[2*k]);
}
}else{
// Lower
for(j=0; j<N; j+=nb){
k = (j/nb)%num_gpus;
nk = min(nb, N-j);
magma_dgetmatrix( nk, N,
d_lA[k], (j/(nb*num_gpus)*nb), ldda,
&h_R[j], lda, queues[2*k] );
}
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = magma_wtime();
if(uplo == MagmaLower){
lapackf77_dpotrf( MagmaLowerStr, &N, h_A, &lda, &info );
}else{
lapackf77_dpotrf( MagmaUpperStr, &N, h_A, &lda, &info );
}
cpu_time = magma_wtime() - cpu_time;
if (info != 0)
printf( "lapackf77_dpotrf had error %d.\n", info );
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result compared to LAPACK
|R_magma - R_lapack| / |R_lapack|
=================================================================== */
matnorm = lapackf77_dlange("f", &N, &N, h_A, &lda, work);
blasf77_daxpy(&n2, &mz_one, h_A, &ione, h_R, &ione);
diffnorm = lapackf77_dlange("f", &N, &N, h_R, &lda, work);
printf( "%5d %6.2f (%6.2f) %6.2f (%6.2f) %e\n",
N, cpu_perf, cpu_time, gpu_perf, gpu_time, diffnorm / matnorm );
if (flag != 0)
break;
}
/* clean up */
TESTING_FREE_PIN( h_A );
TESTING_FREE_PIN( h_R );
for(i=0;i<num_gpus;i++){
TESTING_FREE_DEV( d_lA[i] );
magma_queue_destroy( queues[2*i] );
magma_queue_destroy( queues[2*i+1] );
}
magma_finalize();
}
