/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgeqrf_mgpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double error, work[1];
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_R, *tau, *h_work, tmp[1];
double *d_lA[ MagmaMaxGPUs ];
magma_int_t M, N, n2, lda, ldda, n_local, ngpu;
magma_int_t info, min_mn, nb, lhwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1}, ISEED2[4];
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= (opts.check == 2); // check (-c2) implies lapack (-l)
magma_int_t status = 0;
double tol, eps = lapackf77_dlamch("E");
tol = opts.tolerance * eps;
printf("ngpu %d\n", (int) opts.ngpu );
if ( opts.check == 1 ) {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R-Q'A||_1 / (M*||A||_1) ||I-Q'Q||_1 / M\n");
printf("================================================================================================\n");
} else {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F /(M*||A||_F)\n");
printf("==========================================================================\n");
}
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
nb = magma_get_dgeqrf_nb( M );
gflops = FLOPS_DGEQRF( M, N ) / 1e9;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, int((N+nb-1)/nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// query for workspace size
lhwork = -1;
lapackf77_dgeqrf( &M, &N, h_A, &M, tau, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_D_REAL( tmp[0] );
// Allocate host memory for the matrix
TESTING_MALLOC( tau, double, min_mn );
TESTING_MALLOC( h_A, double, n2 );
TESTING_HOSTALLOC( h_R, double, n2 );
TESTING_MALLOC( h_work, double, lhwork );
// Allocate device memory
for( int dev = 0; dev < ngpu; dev++ ) {
n_local = ((N/nb)/ngpu)*nb;
if (dev < (N/nb) % ngpu)
n_local += nb;
else if (dev == (N/nb) % ngpu)
n_local += N % nb;
magma_setdevice( dev );
TESTING_DEVALLOC( d_lA[dev], double, ldda*n_local );
}
/* Initialize the matrix */
for ( int j=0; j<4; j++ ) ISEED2[j] = ISEED[j]; // saving seeds
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
lapackf77_dlacpy( MagmaUpperLowerStr, &M, &N, h_A, &lda, h_R, &lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
double *tau;
TESTING_MALLOC( tau, double, min_mn );
cpu_time = magma_wtime();
lapackf77_dgeqrf( &M, &N, h_A, &M, tau, h_work, &lhwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapack_dgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
TESTING_FREE( tau );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_dsetmatrix_1D_col_bcyclic( M, N, h_R, lda, d_lA, ldda, ngpu, nb );
gpu_time = magma_wtime();
magma_dgeqrf2_mgpu( ngpu, M, N, d_lA, ldda, tau, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_dgeqrf2 returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_dgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_R, lda, ngpu, nb );
magma_queue_sync( NULL );
if ( opts.check == 1 ) {
/* =====================================================================
Check the result
=================================================================== */
magma_int_t lwork = n2+N;
double *h_W1, *h_W2, *h_W3;
double *h_RW, results[2];
TESTING_MALLOC( h_W1, double, n2 ); // Q
TESTING_MALLOC( h_W2, double, n2 ); // R
TESTING_MALLOC( h_W3, double, lwork ); // WORK
TESTING_MALLOC( h_RW, double, M ); // RWORK
lapackf77_dlarnv( &ione, ISEED2, &n2, h_A );
lapackf77_dqrt02( &M, &N, &min_mn, h_A, h_R, h_W1, h_W2, &lda, tau, h_W3, &lwork,
h_RW, results );
results[0] *= eps;
results[1] *= eps;
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, results[0],results[1] );
printf("%s\n", (results[0] < tol ? "" : " failed"));
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, gpu_perf, gpu_time, results[0],results[1] );
printf("%s\n", (results[0] < tol ? "" : " failed"));
}
status |= ! (results[0] < tol);
TESTING_FREE( h_W1 );
TESTING_FREE( h_W2 );
TESTING_FREE( h_W3 );
TESTING_FREE( h_RW );
} else if ( opts.check == 2 ) {
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
error = lapackf77_dlange("f", &M, &N, h_A, &lda, work );
blasf77_daxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
error = lapackf77_dlange("f", &M, &N, h_R, &lda, work ) / (min_mn*error);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, error );
printf("%s\n", (error < tol ? "" : " failed"));
status |= ! (error < tol);
}
else {
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) ---\n",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) --- \n",
(int) M, (int) N, gpu_perf, gpu_time);
}
}
TESTING_FREE( tau );
TESTING_FREE( h_A );
TESTING_FREE( h_work );
TESTING_HOSTFREE( h_R );
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_DEVFREE( d_lA[dev] );
}
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/***************************************************************************//**
Purpose
-------
DORGQR generates an M-by-N DOUBLE PRECISION matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by DGEQRF.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix Q. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix Q. M >= N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
@param[in,out]
A DOUBLE PRECISION array A, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by DGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau DOUBLE PRECISION array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by DGEQRF_GPU.
@param[in]
T DOUBLE PRECISION array, dimension (NB, min(M,N)).
T contains the T matrices used in blocking the elementary
reflectors H(i), e.g., this can be the 6th argument of
magma_dgeqrf_gpu (except stored on the CPU, not the GPU).
@param[in]
nb INTEGER
This is the block size used in DGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in T.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_ungqr
*******************************************************************************/
extern "C" magma_int_t
magma_dorgqr_m(
magma_int_t m, magma_int_t n, magma_int_t k,
double *A, magma_int_t lda,
double *tau,
double *T, magma_int_t nb,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(d,i,j) (dA[d] + (i) + (j)*ldda)
#define dT(d,i,j) (dT[d] + (i) + (j)*nb)
double c_zero = MAGMA_D_ZERO;
double c_one = MAGMA_D_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldwork;
magma_int_t d, i, ib, j, jb, ki, kk;
double *work=NULL;
*info = 0;
if (m < 0) {
*info = -1;
} else if ((n < 0) || (n > m)) {
*info = -2;
} else if ((k < 0) || (k > n)) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (n <= 0) {
return *info;
}
magma_int_t di, dn;
magma_int_t dpanel;
magma_int_t ngpu = magma_num_gpus();
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
// Allocate memory on GPUs for A and workspaces
magma_int_t ldda = magma_roundup( m, 32 );
magma_int_t lddwork = magma_roundup( n, 32 );
magma_int_t min_lblocks = (n / nb) / ngpu; // min. blocks per gpu
magma_int_t last_dev = (n / nb) % ngpu; // device with last block
magma_int_t nlocal[ MagmaMaxGPUs ] = { 0 };
double *dA[ MagmaMaxGPUs ] = { NULL };
double *dT[ MagmaMaxGPUs ] = { NULL };
double *dV[ MagmaMaxGPUs ] = { NULL };
double *dW[ MagmaMaxGPUs ] = { NULL };
magma_queue_t queues[ MagmaMaxGPUs ] = { NULL };
for( d = 0; d < ngpu; ++d ) {
// example with n = 75, nb = 10, ngpu = 3
// min_lblocks = 2
// last_dev = 1
// gpu 0: 2 blocks, cols: 0- 9, 30-39, 60-69
// gpu 1: 1+ blocks, cols: 10-19, 40-49, 70-74 (partial)
// gpu 2: 1 block, cols: 20-29, 50-59
magma_setdevice( d );
nlocal[d] = min_lblocks*nb;
if ( d < last_dev ) {
nlocal[d] += nb;
}
else if ( d == last_dev ) {
nlocal[d] += (n % nb);
}
ldwork = nlocal[d]*ldda // dA
+ nb*m // dT
+ nb*ldda // dV
+ nb*lddwork; // dW
if ( MAGMA_SUCCESS != magma_dmalloc( &dA[d], ldwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
dT[d] = dA[d] + nlocal[d]*ldda;
dV[d] = dT[d] + nb*m;
dW[d] = dV[d] + nb*ldda;
magma_queue_create( d, &queues[d] );
}
trace_init( 1, ngpu, 1, queues );
// first kk columns are handled by blocked method.
// ki is start of 2nd-to-last block
if ((nb > 1) && (nb < k)) {
ki = (k - nb - 1) / nb * nb;
kk = min(k, ki + nb);
} else {
ki = 0;
kk = 0;
}
// Allocate CPU work space
// n*nb for larfb work
// m*nb for V
// nb*nb for T
lwork = (n + m + nb) * nb;
magma_dmalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto cleanup;
}
double *work_T, *work_V;
work_T = work + n*nb;
work_V = work + n*nb + nb*nb;
// Use unblocked code for the last or only block.
if (kk < n) {
trace_cpu_start( 0, "ungqr", "ungqr last block" );
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
// dorgqr requires less workspace (n*nb), but is slow if k < dorgqr's block size.
// replacing it with the 4 routines below is much faster (e.g., 60x).
//magma_int_t iinfo;
//lapackf77_dorgqr( &m_kk, &n_kk, &k_kk,
// A(kk, kk), &lda,
// &tau[kk], work, &lwork, &iinfo );
lapackf77_dlacpy( MagmaFullStr, &m_kk, &k_kk, A(kk,kk), &lda, work_V, &m_kk);
lapackf77_dlaset( MagmaFullStr, &m_kk, &n_kk, &c_zero, &c_one, A(kk, kk), &lda );
lapackf77_dlarft( MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &k_kk,
work_V, &m_kk, &tau[kk], work_T, &k_kk);
lapackf77_dlarfb( MagmaLeftStr, MagmaNoTransStr, MagmaForwardStr, MagmaColumnwiseStr,
&m_kk, &n_kk, &k_kk,
work_V, &m_kk, work_T, &k_kk, A(kk, kk), &lda, work, &n_kk );
if (kk > 0) {
for( j=kk; j < n; j += nb ) {
jb = min( n-j, nb );
d = (j / nb) % ngpu;
di = ((j / nb) / ngpu) * nb;
magma_setdevice( d );
magma_dsetmatrix( m_kk, jb,
A(kk, j), lda,
dA(d, kk, di), ldda, queues[d] );
// Set A(1:kk,kk+1:n) to zero.
magmablas_dlaset( MagmaFull, kk, jb, c_zero, c_zero, dA(d, 0, di), ldda, queues[d] );
}
}
trace_cpu_end( 0 );
}
if (kk > 0) {
// Use blocked code
// send T to all GPUs
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set T" );
magma_dsetmatrix_async( nb, min(m,n), T, nb, dT[d], nb, queues[d] );
trace_gpu_end( d, 0 );
}
// queue: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
for( i = ki; i >= 0; i -= nb ) {
ib = min(nb, k - i);
mi = m - i;
dpanel = (i / nb) % ngpu;
di = ((i / nb) / ngpu) * nb;
// Send current panel to dV on the GPUs
lapackf77_dlaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set V" );
magma_dsetmatrix_async( mi, ib,
A(i, i), lda,
dV[d], ldda, queues[d] );
trace_gpu_end( d, 0 );
}
// set panel to identity
magma_setdevice( dpanel );
trace_gpu_start( dpanel, 0, "laset", "laset" );
magmablas_dlaset( MagmaFull, i, ib, c_zero, c_zero, dA(dpanel, 0, di), ldda, queues[dpanel] );
magmablas_dlaset( MagmaFull, mi, ib, c_zero, c_one, dA(dpanel, i, di), ldda, queues[dpanel] );
trace_gpu_end( dpanel, 0 );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_indices_1D_bcyclic( nb, ngpu, d, i, n, &di, &dn );
trace_gpu_start( d, 0, "larfb", "larfb" );
magma_dlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, dn-di, ib,
dV[d], ldda, dT(d,0,i), nb,
dA(d, i, di), ldda, dW[d], lddwork, queues[d] );
trace_gpu_end( d, 0 );
}
}
}
// copy result back to CPU
trace_cpu_start( 0, "get", "get A" );
magma_dgetmatrix_1D_col_bcyclic( ngpu, m, n, nb, dA, ldda, A, lda, queues );
trace_cpu_end( 0 );
}
#ifdef TRACING
char name[80];
snprintf( name, sizeof(name), "dorgqr-n%lld-ngpu%lld.svg", (long long) m, (long long) ngpu );
trace_finalize( name, "trace.css" );
#endif
cleanup:
for( d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_free( dA[d] );
magma_queue_destroy( queues[d] );
}
magma_free_cpu( work );
magma_setdevice( orig_dev );
return *info;
} /* magma_dorgqr */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgetrf_mgpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double error;
double *h_A;
double *d_lA[ MagmaMaxGPUs ];
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, n_local, ngpu;
magma_int_t info, min_mn, nb, ldn_local;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("ngpu %d\n", (int) opts.ngpu );
if ( opts.check == 2 ) {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) |Ax-b|/(N*|A|*|x|)\n");
}
else {
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) |PA-LU|/(N*|A|)\n");
}
printf("=========================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
nb = magma_get_dgetrf_nb( M );
gflops = FLOPS_DGETRF( M, N ) / 1e9;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, int((N+nb-1)/nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// Allocate host memory for the matrix
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, double, n2 );
// Allocate device memory
for( int dev=0; dev < ngpu; dev++){
n_local = ((N/nb)/ngpu)*nb;
if (dev < (N/nb) % ngpu)
n_local += nb;
else if (dev == (N/nb) % ngpu)
n_local += N % nb;
ldn_local = ((n_local+31)/32)*32; // TODO why?
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], double, ldda*ldn_local );
}
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
init_matrix( M, N, h_A, lda );
cpu_time = magma_wtime();
lapackf77_dgetrf( &M, &N, h_A, &lda, ipiv, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_dgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
init_matrix( M, N, h_A, lda );
magma_dsetmatrix_1D_col_bcyclic( M, N, h_A, lda, d_lA, ldda, ngpu, nb );
gpu_time = magma_wtime();
magma_dgetrf_mgpu( ngpu, M, N, d_lA, ldda, ipiv, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_dgetrf_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_dgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_A, lda, ngpu, nb );
/* =====================================================================
Check the factorization
=================================================================== */
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f)",
(int) M, (int) N, gpu_perf, gpu_time );
}
if ( opts.check == 2 ) {
error = get_residual( M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else if ( opts.check ) {
error = get_LU_error( M, N, h_A, lda, ipiv );
printf(" %8.2e %s\n", error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
printf( " ---\n" );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
extern "C" magma_int_t
magma_dgetrf_mgpu_work_amc_v3(magma_int_t num_gpus,
magma_int_t m, magma_int_t n,
double **dlA, magma_int_t dlA_LD,
magma_int_t *ipiv, magma_int_t *info,
/*workspace on the cpu side*/
double *AWORK, magma_int_t AWORK_LD, magma_int_t AWORK_n
)
{
/* -- MAGMA (version 1.5.0-beta3) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2011
Purpose
=======
DGETRF_REC_ASYNC computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. The technique used for the panel factorization
is the parallel recursif LU (see lawn 259).
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
Arguments
=========
NUM_GPUS
(input) INTEGER
The number of GPUS to be used for the factorization.
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
A (input/output) DOUBLE_PRECISION array on the GPU, dimension (LDDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
LDDA (input) INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
IPIV (output) INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
> 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
===================================================================== */
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
int ONE = 1;
magma_int_t iinfo, nb;
magma_int_t mindim;
magma_int_t nrows, ncols;
//double *work;
magma_int_t dm_max, dn_max;
magma_int_t I, J, K, M, N, U_K, L;
//magma_int_t A_m, A_n, A_N;
//magma_int_t Am_max, An_max;
//magma_int_t A_nb;
//magma_int_t A_K;
double **dlAT;
magma_int_t dlAT_LD;
double *dlAP_get[MagmaMaxGPUs]; //*dlAP_set[MagmaMaxGPUs]
double *dlAP_set[MagmaMaxGPUs];
magma_int_t dlAP_LD;
double *dlpanel[MagmaMaxGPUs];
magma_int_t dlpanel_LD;
int *n_local, *nr_local;
//magma_int_t nrows, ncols;
magma_int_t gpu_nrows, gpu_ncols;
int nbcores; /*Number of cores available for the whole factorization*/
int panel_num_threads; /*Number of threads for the panel*/
double dcpu; /*percentage of the matrix to allocate on the CPUs*/
int B_rows;
double t1;
/*Workspace*/
// magma_int_t AWORK_NMAX;
// magma_int_t AWORK_m, AWORK_n, AWORK_N;
/* Recommanded dimension in the workspace*/
int A_m, A_n, A_N, A_NMAX, A_LD;
int A_NP1;
double *A;
amc_args_t *args;
/*magma_event_t *A_event;*/ /*Control bucket*/
magma_queue_t mstream[MagmaMaxGPUs][3]; /*0: H2D, 1: compute, 2:D2H*/
int dd;
// double *tmpdA;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (dlA_LD < max(1,m))
*info = -4;
else if (AWORK_LD < max(1,m))
*info = -5;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/*Get parameters*/
args = magma_amc_args_get_default();
nb= args->nb;
nbcores = args->P;
panel_num_threads = args->Pr;
dcpu = args->dcpu;
/* Check and fix parameters */
if(nb==0)
nb = magma_get_dgetrf_nb(m) ;/*magma dgetrf block size*/
else
nb = args->nb;
if(nb>n) nb = n;
if(panel_num_threads>nbcores) panel_num_threads = nbcores;
/*check the buffer size*/
if(AWORK_n<nb){
printf("Not enough buffer. Should be greater than the block size: %d\n", nb);
exit(1);
}
/* Compute the number of blocks columns to factorize*/
N = (int) ceil( (double) min(m, n) / nb);
/* Compute the maximum number of panels we can store in the workspace*/
A_NMAX = (int) (AWORK_n/ nb);
/*Compute the recommanded number of panels for the cpu part*/
A_N = NSplit(N, dcpu);
/* Compute the recommanded number of columns for the cpu part*/
A_n = A_N*nb;//(int) ceil(n*dcpu);
//if(A_n<nb)
// A_n = nb;//make sure workspace has at least one block column
/*Make sure we work with multiple of 32*/
/*
if(A_n%32!=0) {
A_n = ((A_n + 31)/32)*32;
}
*/
/* Compute the recommanded number of panels for the cpu part*/
// A_N = (int) (A_n/ nb);
/* Check if there are enough workspace. In case the user gave a workspace lower than the optimal*/
/* NOTE: using small workspace may reduce performance*/
if(A_N>A_NMAX){
#if (dbglevel >=1)
printf("[DBG_WARNING] Resizing buffer to feet user preferences. Recommanded:%d, Max given:%d\n",A_N, A_NMAX);
#endif
A_N = A_NMAX;
/*Make A_n a multiple of nb*/
A_n = A_N*nb;
}
A = AWORK;
A_m = m;
A_LD = AWORK_LD;
#if (dbglevel >=1)
/* Initialize the tracing*/
ca_dbg_trace_init(nbcores,num_gpus); //nbcores + 1 GPU
#endif
#if (dbglevel >=1)
t1 = magma_wtime();
#endif
/* create the streams */
//mstream = (magma_queue_t *) malloc(num_gpus*sizeof(magma_queue_t));
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd); //required
magma_queue_create(&mstream[dd][0]);
magma_queue_create(&mstream[dd][1]);
magma_queue_create(&mstream[dd][2]);
/*Set the stream for internal computations*/
//magmablasSetKernelStream(0); /*Use 0 instead of mstream[dd][1], MagmasetkernelStream is not thread safe*/ /*TODO: mae it safe*/
//task_dev_set_compute_stream(dd, mstream[dd][1]);
magma_task_dev_set_compute_stream(dd, 0); //set to mstream 1 later
}
/* Matrix dimension */
dm_max = m;
dn_max = n;
/*Make sure m and n are multiple of 32*/
if(dm_max%32!=0) dm_max = ((dm_max + 31)/32)*32;
if(dn_max%32!=0) dn_max = ((dn_max + 31)/32)*32;
/* local dimensions of the matrix for each GPU*/
n_local = (int *) malloc(num_gpus*sizeof(int)); /*This do no change during the execution*/
nr_local = (int *) malloc(num_gpus*sizeof(int)); /*Change after each update of the trailing submatrix*/
for(dd=0;dd<num_gpus;dd++){
n_local[dd] = numcols2p(dd, n, nb, num_gpus); //loc2p(dd, N, num_gpus)*nb;
nr_local[dd] = n_local[dd];
}
/*Allocate a workspace for the panels transposition*/
dlAP_LD = dm_max;
//if(dAP_LD%32!=0) dAP_LD = ((dAP_LD + 31)/32)*32;/*Make dAP_LD multiple of 32*/
/// dlAP_set = (double **) malloc(num_gpus*sizeof(double*));
//dlAP_get = (double **) malloc(num_gpus*sizeof(double*));
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
if (MAGMA_SUCCESS != magma_dmalloc( &dlAP_set[dd], dlAP_LD*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/*
if (MAGMA_SUCCESS != magma_dmalloc(&tmpdA, dlAP_LD*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
*/
if ( magma_is_devptr(dlAP_set[dd] ) == 0 ) {
fprintf( stderr, "ERROR: dlAP_set[dd] is host pointer.\n" );
//exit(1);
}
//cudaMemcpy(dlAP_set[dd],&tmpdA,sizeof(double*), cudaMemcpyDeviceToHost);
#if (dbglevel==10)
printf("0.4\n");
//ca_dbg_printMat_gpu(2, 2, dlAP_set[dd], dlAP_LD, "dlAP_set[dd] for testing");
//cudaMemcpy(&tmpdA, &dlAP_set[dd], sizeof(double*), cudaMemcpyHostToDevice);
//ca_dbg_printMat_gpu(2, 2, tmpdA, dlAP_LD, "dlAP_set[dd] for testing");
//printf("0.5: int to continue"); scanf("%d", &I);
#endif
if (MAGMA_SUCCESS != magma_dmalloc(&dlAP_get[dd], dlAP_LD*nb)) {
//magma_free(dlAP_set); //TODO: free all previous buffers
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
/* Workspace for the panels */
// dlpanel = (double **) malloc(num_gpus*sizeof(double*));
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
if (MAGMA_SUCCESS != magma_dmalloc(&dlpanel[dd], nb*dm_max)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
dlpanel_LD = nb;
/*local matrix storage*/
dlAT = (double **) malloc(num_gpus*sizeof(double*));
dlAT_LD = n_local[0];
if(dlAT_LD%32!=0) dlAT_LD = ((dlAT_LD + 31)/32)*32;
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
if (MAGMA_SUCCESS != magma_dmalloc(&dlAT[dd], dlAT_LD*dm_max )) {
for(J=0;J<dd;J++){
magma_setdevice(J);
magma_free( dlAP_set[J]);
magma_free( dlAP_get[J]);
magma_free(dlpanel[J]);
magma_free(dlAT[J]);
}
//free(dlAP_set);
//free(dlAP_get);
//free(dlpanel);
free(dlAT);
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
#if (dbglevel >=1)
printf("[DBG] Time workspace memory alloc (dAP): %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
/*1. Transfer the first column blocks of the matrix from the GPU to the CPUs.*/
//magma_dgetmatrix(A_m, A_n, dA, dA_LD, A, A_LD);
magma_dgetmatrix_1D_col_bcyclic(A_m, A_n, dlA, dlA_LD, A, A_LD, num_gpus, nb);
#if (dbglevel >=1)
printf("[DBG] Time First getmatrix: %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
printf("1.0\n");
ca_dbg_printMat(A_m, A_n, A, A_LD,"A after first getMatrix");
/*
for(dd=0;dd<num_gpus;dd++){
//Fill the matrix with zero for easy visualization of the matrix in debug mode
for(I=0;I<dlAT_LD*dm_max;I++) dlAT[dd][I] = 0.0;
}
*/
// ca_dbg_printMat_mgpu(num_gpus, m, n_local, dlAT, dlAT_LD,"matrix dAlT^T empty");
// ca_dbg_printMat_transpose_mgpu(num_gpus, n_local, m, dlAT, dlAT_LD,"matrix dAT empty");
printf("2.0\n");
#endif
/*Update the remaining number of columns on the GPUs.*/
for(dd=0;dd<num_gpus;dd++){
nr_local[dd] = nr_local[dd] - numcols2p(dd, A_n, nb, num_gpus); //;n_local[dd] - loc2p(dd, A_N, num_gpus)*nb;
}
#if (dbglevel==10)
ca_dbg_printMat_mgpu(num_gpus, m, n_local, dlA, dlA_LD,"matrix dA to factorize");
printf("3.0\n");
#endif
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
//magmablasSetKernelStream(mstream[dd][1]);
magmablas_dtranspose2(dlAT[dd], dlAT_LD, dlA[dd], dlA_LD, m, n_local[dd]);
}
///
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
magma_task_dev_set_compute_stream(dd, mstream[dd][1]);
}
#if (dbglevel >=1)
printf("[DBG] Time First transposition: %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
//ca_dbg_printMat_transpose_mgpu(num_gpus, n_local, m, dlAT, dlAT_LD,"matrix dAT to factorize");
/*
dd = GID(A_N);
magma_setdevice(dd);
ca_dbg_printMat_transpose_gpu(nb, m, dlAT(0, A_N), dlAT_LD,"matrix dAT(0, A_N)");
magma_setdevice(0);
ca_dbg_printMat_transpose_gpu(m, nb, dlA(0, A_N), dlA_LD,"matrix dA(0, A_N)");
*/
printf("4.0\n");
printf("int to continue"); scanf("%d", &I);
#endif
/*
#if (dbglevel==10)
ca_dbg_printMat_transpose_mgpu(num_gpus, m, n_local, dlAT, dlAT_LD,"matrix dAT to factorize");
#endif
*/
/* Compute the maximun number of steps*/
mindim = min(m, n);
M = (int) ceil( (double) m / nb);
N = (int) ceil( (double) mindim / nb); /*N = n/nb*/
/* 3. Let the asynchronous algorithm begin*/
#if (dbglevel >=1)
printf("Starting recursif code ... m:%d, n:%d, nb:%d, nbcores:%d, N:%d, A_N:%d\n", m, n, nb, nbcores, N, A_N); //Summary
#endif
/*Initialize the scheduler*/
magma_schedule_init(nbcores, num_gpus);
K = 0;
/*initialize parallel recursif panel environment*/
CORE_zgetrf_reclap_init();
magma_schedule_set_task_priority(INT_MAX-1);
/*Schedule the first panel factorization*/
magma_insert_core_dgetrf_rec(A_m, nb, A(0,K), A_LD, ipiv(0), &iinfo, panel_num_threads, colptr(K));
//magma_insert_core_dgetrf(A_m, nb, A(0,K), A_LD, ipiv(0), &iinfo, colptr(K));
/*Transfer the factorized panel in the buffer of GPU (dlpanel)*/
for(dd=0;dd<num_gpus;dd++){
///magma_insert_dev_dsetmatrix_transpose(dd, A_m, nb, A(0,K), A_LD, dlpanel(dd,K), dlpanel_LD, dlAP_set[dd], dlAP_LD, colptr(K), dlpanel[dd]);
magma_insert_dev_dsetmatrix_async_transpose(dd, A_m, nb, A(0,K), A_LD, dlpanel(dd,K), dlpanel_LD, mstream[dd][0], dlAP_set[dd], dlAP_LD, colptr(K), dlpanel(dd,K)); //dlpanel[dd]
}
#if (dbglevel==10)
magma_schedule_barrier();
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
ca_dbg_printMat_transpose_gpu(nb, m, dlpanel(dd,K), dlpanel_LD,"dlpanel[dd] after setmatrix_async"); //dlpanel[dd]
}
printf("4.5: int to continue"); scanf("%d", &I);
#endif
/*Transfer also the factorized panel on its right position in the final matrix (transposition included)*/
/*TODO: this may use cudaMemcpyDeviceToDevice and initiate the transfer from dlpanel*/
dd = GID(K);
//magma_insert_dev_dsetmatrix_transpose(dd, A_m, nb, A(0,K), A_LD, dlAT(0,K), dlAT_LD, dlAP_set[dd], dlAP_LD, colptr(K), dlAT(0,K));
magma_insert_dev_dsetmatrix_async_transpose(dd, A_m, nb, A(0,K), A_LD, dlAT(0,K), dlAT_LD, mstream[dd][0], dlAP_set[dd], dlAP_LD, colptr(K), dlAT(0,K));
#if (dbglevel==10)
magma_schedule_barrier();
ca_dbg_printMat(m, nb, A(0,0), A_LD,"A(0,0)");
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
ca_dbg_printMat_transpose_gpu(nb, m, dlpanel[dd], dlpanel_LD,"dlpanel[dd] after setmatrix to dlAT");
}
ca_dbg_printMat_transpose_mgpu(num_gpus, n_local, m, dlAT, dlAT_LD,"dlA");
printf("5.0: int to continue"); scanf("%d", &I);
#endif
for(K=0;K<=N-1;K++){
/*compute the new value of the cpu number of blocks*/
A_N = NSplit(N-K, dcpu);
/*insert the coarse update of the trailing submatrix corresponding to panel K to the GPU, that is submatrix A[K+1:M, K+1+d-1:N]*/
//if(K==0) /*TODO: move outside loop*/
//{
/*NOTE: Here we work on the matrix transpose*/
/*Set the priority max for the GPU computations*/
magma_schedule_set_task_priority(INT_MAX);
//// magma_schedule_set_task_priority(INT_MAX - N*K);
gpu_nrows = m - (K+1)*nb;///
for(J=K+A_N;J<min(K+A_N+num_gpus,N);J++){
/*Determine the device which own the first column of the group of columns to update*/
dd = GID(J);
/*Determine the number of columns to apply the update. */
nr_local[dd] = numcols2p(dd, n - (K+1+A_N-1)*nb, nb, num_gpus);
gpu_ncols = nr_local[dd]; //n - (K+1+A_N-1)*nb;
if(gpu_ncols >0)
{
/*schedule a swap of the trailing submatrix in the gpus using ipiv[K]*/
/*dependency dAT((K+1)-1, (K+A_N)-1) = dAT(K, K+A_N-1) with previous dgemm*/
magma_insert_dev_dlaswp(dd, gpu_ncols, dlAT(K, J), dlAT_LD, ONE, nb, ipiv(K), ONE, dlAT(K, J-1)); /*non blocking*/
//printf("debug barrier\n");
//magma_schedule_barrier();
//&(dlpanel[dd][dlpanel_LD*nb*K])
magma_insert_dev_dtrsm(dd, MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit, gpu_ncols, nb, c_one, dlpanel(dd,K), dlpanel_LD, dlAT(K,J), dlAT_LD);/*non blocking*/
/* aij^T = aij^T - (lik.ukj)^T = aij^T - ukj^T.lik^T*/ //&(dlpanel[dd][dlpanel_LD*nb*(K+1)])
magma_insert_dev_dgemm(dd, MagmaNoTrans,MagmaNoTrans, gpu_ncols, gpu_nrows, nb, c_neg_one, dlAT(K,J), dlAT_LD, dlpanel(dd,K+1), dlpanel_LD, c_one, dlAT(K+1,J), dlAT_LD);/*non blocking*/
/*Transfer asynchronously one column (column K+A_N) from the GPU to the CPU to balance work*/
//// if(K+A_N<N)
//// {
////ncols = min(nb, gpu_ncols);
//////magma_schedule_set_task_priority(INT_MAX);
////magma_insert_dgetmatrix_transpose(gpu_nrows, ncols, dAT(K+1,K+A_N), dAT_LD, A(K+1,K+A_N), A_LD, dAP, dAP_LD, colptr(K+A_N)); //blocking
//// }
}
}
//}
/*iterate over the rest of the columns to update the trailing submatrix on the cpu*/
for(J=K+1;J<=min(K+A_N-1, N-1);J++){
ncols = min(nb, n - J*nb);
/*Set the priority max for column having the next panel (look ahead of deep 1),
and process the rest of the update in a right looking way*/
if(J==K+1)
magma_schedule_set_task_priority(INT_MAX -2 );
//// magma_schedule_set_task_priority(INT_MAX - N*K -1);
else
magma_schedule_set_task_priority(INT_MAX -3 - J );//- N*K
//// magma_schedule_set_task_priority(INT_MAX - N*K -3 -J);
//magma_schedule_set_task_priority(INT_MAX - J);
/*dependency colptr(J): make sure column J is sent from GPU, and all previous update was done*/
magma_insert_core_dlaswp(ncols, A(K,J), A_LD, ONE, nb, ipiv(K), ONE, colptr(J));
magma_insert_core_dtrsm('L', 'L', 'N', 'U', nb, ncols, c_one, A(K,K), A_LD, A(K,J), A_LD, colptr(J));
/*Compute the number of blocs rows to group together before the update. To avoid scheduling overhead.*/
B_rows = (int) ceil((double) (M-K-1)/panel_num_threads);
B_rows = max(B_rows,4); /*maximun of 4*/
//B_rows = max(B_rows,1);
//printf("B_rows:%d\n",B_rows);
for(I=K+1; I<=M-1; I+=B_rows){
nrows = min(B_rows*nb, m-I*nb);
/*dep colptr(K):make sure the panel is not overwritten or swapped since dgemm use A[I,K]*/
/*dep colptr(J): Gather all dgemm on one column and create dependencies with previous dgemm and the next panel*/
magma_insert_core_dgemm('N','N', nrows, ncols, nb, c_neg_one, A(I,K), A_LD, A(K,J), A_LD, c_one, A(I,J), A_LD, colptr(K), colptr(J));
}
if(J==K+1)
{
/*Look ahead and insert the next panel*/
nrows = m - (K+1)*nb;
ncols = min(nb, n - (K+1)*nb);
/*Schedule the next panel factorization with maximum priority*/
magma_schedule_set_task_priority(INT_MAX -1);
///magma_schedule_set_task_priority(0); //TEST: testing prio_0
//// magma_schedule_set_task_priority(INT_MAX - N*K - 2);
magma_insert_core_dgetrf_rec(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
// magma_insert_core_dgetrf(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, colptr(K+1));
/*Transfer the factorized panel in the buffer of GPU (dlpanel)*/
for(dd=0;dd<num_gpus;dd++){
//&(dlpanel[dd][dlpanel_LD*nb*(K+1)])
///magma_insert_dev_dsetmatrix_transpose(dd, nrows, ncols, A(K+1, K+1), A_LD, dlpanel(dd, K+1), dlpanel_LD, dlAP_set[dd], dlAP_LD, colptr(K+1), dlpanel[dd]);
magma_insert_dev_dsetmatrix_async_transpose(dd, nrows, ncols, A(K+1, K+1), A_LD, dlpanel(dd, K+1), dlpanel_LD, mstream[dd][0], dlAP_set[dd], dlAP_LD, colptr(K+1), dlpanel(dd,K+1));//, dlpanel[dd]
}
/*Determine the upper part of the matrix done by the CPU on that column and send it to the GPU with the panel*/
U_K = max(0, K+1 - A_N +1);
nrows = m - U_K*nb;
///magma_schedule_set_task_priority(INT_MAX);
/*Transfer the upper part of the matrix for that column and the factorized panel to the GPU*/
///magma_insert_dsetmatrix_transpose(nrows, ncols, A(U_K, K+1), A_LD, dAT(U_K, K+1), dAT_LD, dAP, dAP_LD, A(K+1,K+1), dAT(K+1,K+1));
//magma_insert_dev_dsetmatrix_transpose(nrows, ncols, A(U_K, K+1), A_LD, dAT(U_K, K+1), dAT_LD, dAP_set, dAP_LD, colptr(K+1), dAT(K+1,K+1));
/*Transfer also the factorized panel on its right position in the final matrix (transposition included)*/
/*TODO: this may use cudaMemcpyDeviceToDevice and initiate the transfer from dlpanel*/
dd = GID(K+1);
///magma_insert_dev_dsetmatrix_transpose(dd, nrows, ncols, A(U_K, K+1), A_LD, dlAT(U_K,K+1), dlAT_LD, dlAP_set[dd], dlAP_LD, colptr(K+1), dlAT(K+1,K+1));
magma_insert_dev_dsetmatrix_async_transpose(dd, nrows, ncols, A(U_K, K+1), A_LD, dlAT(U_K,K+1), dlAT_LD, mstream[dd][0], dlAP_set[dd], dlAP_LD, colptr(K+1), dlAT(0,K+1));///
}
}
/*compute the next number of blocks colums */
A_NP1 = NSplit(N-(K+1), dcpu) - NSplit(N-K, dcpu) + 1;
/*Transfer asynchronously (A_NP1 - A_N) block column (column K+A_N) from the GPU to the CPU to balance work*/
/*Make sure this is inserted after all dgemm because it schedules to replace a current panel for the case A_N< N*/
for(L=K+A_N;L<K+A_N+A_NP1;L++)
{
if(L<N) {
/*Determine the device which own column K+A_N*/
dd = GID(L);
gpu_ncols = nr_local[dd];
ncols = min(nb, gpu_ncols);
magma_schedule_set_task_priority(INT_MAX);
///magma_insert_dev_dgetmatrix_transpose(dd, gpu_nrows, ncols, dlAT(K+1,K+A_N), dlAT_LD, A(K+1,K+A_N), A_LD, dlAP_get[dd], dlAP_LD, colptr(K+A_N)); //blocking
/*make sure the computations are done on stream 1 and send a block column on stream 2*/
magma_insert_dev_queue_sync(dd, mstream[dd][1], dlAT(K+1,L));
magma_insert_dev_dgetmatrix_async_transpose(dd, gpu_nrows, ncols, dlAT(K+1,L), dlAT_LD, A(K+1,L), A_LD, mstream[dd][2], dlAP_get[dd], dlAP_LD, colptr(L));
/*Update the remaining number of columns*/
//// nr_local[dd]-=nb;
/*if A_N==1, there is no look-ahead, so insert the panel here*/
if((A_N==1) && (L==K+A_N)){
/*Look ahead and insert the next panel*/
nrows = m - (K+1)*nb;
ncols = min(nb, n - (K+1)*nb);
/*Schedule the next panel factorization with maximum priority*/
magma_schedule_set_task_priority(INT_MAX -1);
///magma_schedule_set_task_priority(0); //TEST: testing prio_0
//// magma_schedule_set_task_priority(INT_MAX - N*K - 2);
magma_insert_core_dgetrf_rec(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
//magma_insert_core_dgetrf(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, colptr(K+1));
/*Transfer the factorized panel in the buffer of GPU (dlpanel)*/
for(dd=0;dd<num_gpus;dd++){
//&(dlpanel[dd][dlpanel_LD*nb*(K+1)])
///magma_insert_dev_dsetmatrix_transpose(dd, nrows, ncols, A(K+1, K+1), A_LD, dlpanel(dd, K+1), dlpanel_LD, dlAP_set[dd], dlAP_LD, colptr(K+1), dlpanel[dd]);
magma_insert_dev_dsetmatrix_async_transpose(dd, nrows, ncols, A(K+1, K+1), A_LD, dlpanel(dd, K+1), dlpanel_LD, mstream[dd][0], dlAP_set[dd], dlAP_LD, colptr(K+1), dlpanel(dd,K+1));//dlpanel[dd]
}
/*Determine the upper part of the matrix done by the CPU on that column and send it to the GPU with the panel*/
U_K = max(0, K+1 - A_N +1);
nrows = m - U_K*nb;
///magma_schedule_set_task_priority(INT_MAX);
/*Transfer the upper part of the matrix for that column and the factorized panel to the GPU*/
///magma_insert_dsetmatrix_transpose(nrows, ncols, A(U_K, K+1), A_LD, dAT(U_K, K+1), dAT_LD, dAP, dAP_LD, A(K+1,K+1), dAT(K+1,K+1));
//magma_insert_dev_dsetmatrix_transpose(nrows, ncols, A(U_K, K+1), A_LD, dAT(U_K, K+1), dAT_LD, dAP_set, dAP_LD, colptr(K+1), dAT(K+1,K+1));
/*Transfer also the factorized panel on its right position in the final matrix (transposition included)*/
/*TODO: this may use cudaMemcpyDeviceToDevice and initiate the transfer from dlpanel*/
dd = GID(K+1);
///magma_insert_dev_dsetmatrix_transpose(dd, nrows, ncols, A(U_K, K+1), A_LD, dlAT(U_K,K+1), dlAT_LD, dlAP_set[dd], dlAP_LD, colptr(K+1), dlAT(K+1,K+1));
magma_insert_dev_dsetmatrix_async_transpose(dd, nrows, ncols, A(U_K, K+1), A_LD, dlAT(U_K,K+1), dlAT_LD, mstream[dd][0], dlAP_set[dd], dlAP_LD, colptr(K+1), dlAT(0,K+1));///dlAT(K+1,K+1)
}
}
}
#if (dbglevel==10)
magma_schedule_barrier();
ca_dbg_printMat(m, A_n, A, A_LD,"A");
ca_dbg_printMat_transpose_mgpu(num_gpus, n_local, m, dlAT, dlAT_LD,"dAT (Step K)");
nrows = m - K*nb;
ncols = min(nb, n - K*nb);
dd = GID(K);
magma_setdevice(dd);
ca_dbg_printMat_transpose_gpu(ncols, nrows, dlAT(K,K), dlAT_LD,"dAT(K,K)");
if(K<=5){
printf("Step K:%d done. Int to continue: ",K); scanf("%d", &I);
}
#endif
} //Step K done
/*Wait for all thread termination*/
magma_schedule_barrier();
/*make sure everything arrived*/ ///needed?
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
magma_queue_sync(mstream[dd][0]);
magma_queue_sync(mstream[dd][1]);
magma_queue_sync(mstream[dd][2]);
}
/*TODO: don't need quark here*/
/*Perform a sequence of left swap on the matrix corresponding to the different panel*/
for(K=1;K<=N-1;K++){
#if (dbglevel >=1)
ca_trace_start();
#endif
nrows = min(nb,m - K*nb);
ncols = min(K*nb,n);
for(dd=0;dd<=min(num_gpus-1, K-1);dd++){
gpu_ncols = numcols2p(dd, ncols, nb, num_gpus);
J = dd;
if(gpu_ncols>0){
magma_setdevice(dd);
//pthread_mutex_lock(&mutex_compute_stream);
magmablasSetKernelStream(mstream[dd][1]);
magmablas_dlaswp(gpu_ncols, dlAT(K, J), dlAT_LD, ONE, nrows, ipiv(K), ONE);
//pthread_mutex_lock(&mutex_compute_stream);
}
}
#if (dbglevel >=1)
ca_trace_end_1gpu('W');
#endif
}
#if (dbglevel==10)
ca_dbg_printMat_transpose_mgpu(num_gpus, n_local, m, dlAT, dlAT_LD,"dAT after lswap");
#endif
/*Shutdown the scheduler*/
magma_schedule_delete();
/*update permutation vector indexes*/
for(K=1;K<=N-1;K++){
nrows = min(nb, n-K*nb);
for(J=0;J<=nrows-1;J++){
ipiv[K*nb+J] += K*nb;
}
}
#if dbglevel>=1
printf("[DBG] Time Factorization:%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
/* 4. Transpose back the matrix in/out of place*/
for(dd=0;dd<num_gpus;dd++){
//n_local[dd] = numcols2p(dd, n, nb, num_gpus); //loc2p(dd, N, num_gpus)*nb;
magma_setdevice(dd);
magmablasSetKernelStream(mstream[dd][1]);
magmablas_dtranspose2(dlA[dd], dlA_LD, dlAT[dd], dlAT_LD, n_local[dd], m);
}
for(dd=0;dd<num_gpus;dd++){ //needed
magma_setdevice(dd);
magmablasSetKernelStream(NULL);
}
#if dbglevel>=1
printf("[DBG] Time Final in/out of place transpose:%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
ca_dbg_printMat_mgpu(num_gpus, m, n_local, dlA, dlA_LD,"dA = LU");
#endif
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
magma_queue_destroy(mstream[dd][0]);
magma_queue_destroy(mstream[dd][1]);
magma_queue_destroy(mstream[dd][2]);
}
//free(mstream);
// printf("Step 4: time:%f\n",magma_wtime()-t1);
// t1 = magma_wtime();
free(n_local);
free(nr_local);
// free(k_local);
for(dd=0;dd<num_gpus;dd++){
magma_setdevice(dd);
magma_free( dlAP_set[dd]);
magma_free( dlAP_get[dd]);
magma_free(dlpanel[dd]);
magma_free(dlAT[dd]);
}
//free(dlAP_set);
//free(dlAP_get);
//free(dlpanel);
free(dlAT);
#if dbglevel>=1
printf("[DBG] Time memory free (dAP):%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if dbglevel>=1
/*Finalize the tracing*/
ca_dbg_trace_finalize();
printf("[DBG] Time llog:%f\n",magma_wtime()-t1);
#endif
return *info;
} /* End of MAGMA_DGETRF_REC_ASYNC_WORK_GPU */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgeqrf_mgpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double error, work[1];
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_R, *tau, *h_work, tmp[1];
magmaDouble_ptr d_lA[ MagmaMaxGPUs ];
magma_int_t M, N, n2, lda, ldda, n_local, ngpu;
magma_int_t info, min_mn, nb, lhwork;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1}, ISEED2[4];
magma_opts opts;
opts.parse_opts( argc, argv );
opts.ngpu = abs( opts.ngpu ); // always uses multi-GPU code
opts.lapack |= (opts.check == 2); // check (-c2) implies lapack (-l)
magma_int_t status = 0;
double eps = lapackf77_dlamch("E");
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("%% ngpu %d\n", (int) opts.ngpu );
if ( opts.check == 1 ) {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) ||R-Q'A||_1 / (M*||A||_1) ||I-Q'Q||_1 / M\n");
printf("%%===============================================================================================\n");
}
else {
printf("%% M N CPU Gflop/s (sec) GPU Gflop/s (sec) ||R||_F /(M*||A||_F)\n");
printf("%%=========================================================================\n");
}
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = magma_roundup( M, opts.align ); // multiple of 32 by default
nb = magma_get_dgeqrf_nb( M, N );
gflops = FLOPS_DGEQRF( M, N ) / 1e9;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, magma_ceildiv(N,nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// query for workspace size
lhwork = -1;
lapackf77_dgeqrf( &M, &N, NULL, &M, NULL, tmp, &lhwork, &info );
lhwork = (magma_int_t) MAGMA_D_REAL( tmp[0] );
// Allocate host memory for the matrix
TESTING_MALLOC_CPU( tau, double, min_mn );
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_CPU( h_work, double, lhwork );
TESTING_MALLOC_PIN( h_R, double, n2 );
// Allocate device memory
for( int dev = 0; dev < ngpu; dev++ ) {
n_local = ((N/nb)/ngpu)*nb;
if (dev < (N/nb) % ngpu)
n_local += nb;
else if (dev == (N/nb) % ngpu)
n_local += N % nb;
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], double, ldda*n_local );
}
/* Initialize the matrix */
for( int j=0; j < 4; j++ )
ISEED2[j] = ISEED[j]; // save seeds
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
lapackf77_dlacpy( MagmaFullStr, &M, &N, h_A, &lda, h_R, &lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
double *tau2;
TESTING_MALLOC_CPU( tau2, double, min_mn );
cpu_time = magma_wtime();
lapackf77_dgeqrf( &M, &N, h_A, &lda, tau2, h_work, &lhwork, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0) {
printf("lapack_dgeqrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
TESTING_FREE_CPU( tau2 );
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_dsetmatrix_1D_col_bcyclic( M, N, h_R, lda, d_lA, ldda, ngpu, nb );
gpu_time = magma_wtime();
magma_dgeqrf2_mgpu( ngpu, M, N, d_lA, ldda, tau, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0) {
printf("magma_dgeqrf2 returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
magma_dgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_R, lda, ngpu, nb );
if ( opts.check == 1 && M >= N ) {
/* =====================================================================
Check the result -- dqrt02 requires M >= N
=================================================================== */
magma_int_t lwork = n2+N;
double *h_W1, *h_W2, *h_W3;
double *h_RW, results[2];
TESTING_MALLOC_CPU( h_W1, double, n2 ); // Q
TESTING_MALLOC_CPU( h_W2, double, n2 ); // R
TESTING_MALLOC_CPU( h_W3, double, lwork ); // WORK
TESTING_MALLOC_CPU( h_RW, double, M ); // RWORK
lapackf77_dlarnv( &ione, ISEED2, &n2, h_A );
lapackf77_dqrt02( &M, &N, &min_mn, h_A, h_R, h_W1, h_W2, &lda, tau, h_W3, &lwork,
h_RW, results );
results[0] *= eps;
results[1] *= eps;
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time, results[0], results[1] );
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) %8.2e %8.2e",
(int) M, (int) N, gpu_perf, gpu_time, results[0], results[1] );
}
// todo also check results[1] < tol?
printf(" %s\n", (results[0] < tol ? "ok" : "failed"));
status += ! (results[0] < tol);
TESTING_FREE_CPU( h_W1 );
TESTING_FREE_CPU( h_W2 );
TESTING_FREE_CPU( h_W3 );
TESTING_FREE_CPU( h_RW );
}
else if ( opts.check == 2 ) {
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
error = lapackf77_dlange("f", &M, &N, h_A, &lda, work );
blasf77_daxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
error = lapackf77_dlange("f", &M, &N, h_R, &lda, work ) / (min_mn*error);
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed"));
status += ! (error < tol);
}
else {
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f) ---",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
} else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f) ---",
(int) M, (int) N, gpu_perf, gpu_time);
}
printf("%s\n", (opts.check != 0 ? " (error check only for M >= N)" : ""));
}
TESTING_FREE_CPU( tau );
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_work );
TESTING_FREE_PIN( h_R );
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
opts.cleanup();
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dgetrf_mgpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
real_Double_t gpu_perf1, gpu_time1, gpu_perf2, gpu_time2, gpu_perf3, gpu_time3, alloc_time, free_time;
double error;
double *h_A;
double *d_lA[ MagmaMaxGPUs ];
magma_int_t *ipiv;
magma_int_t M, N, n2, lda, ldda, n_local, ngpu, NB;
magma_int_t info, min_mn, nb, ldn_local;
magma_int_t status = 0;
magma_int_t P=-1; /*Number of threads in the CPU part*/
double d_cpu=-1; /*pourcentgae of the matrix to allocate in the cpu part*/
magma_int_t Pr=-1; /*Number of threads for the panel*/
magma_int_t async_nb; /*Block size*/
double *WORK;
magma_int_t WORK_LD, WORK_n;
double **dlpanelT;
magma_int_t dlpanelT_m, dlpanelT_n;
magma_opts opts;
parse_opts( argc, argv, &opts );
double tol = opts.tolerance * lapackf77_dlamch("E");
P = opts.nthread;
async_nb = opts.nb;
Pr = opts.panel_nthread;
d_cpu = 0.0;
#if defined(CPU_PEAK) && defined(GPU_PEAK)
d_cpu = magma_amc_recommanded_dcpu(opts.nthread, CPU_PEAK, opts.ngpu, GPU_PEAK);
#endif
if(opts.fraction_dcpu!=0){ /*Overwrite the one computed with the model*/
d_cpu = opts.fraction_dcpu;
}
magma_assert(d_cpu > 0 && d_cpu<=1.0,
"error: The cpu fraction is invalid. Ensure you use --fraction_dcpu with fraction_dcpu in [0.0, 1.0] or compile with both -DCPU_PEAK=<cpu peak performance> and -DGPU_PEAK=<gpu peak performance> set.\n");
printf("Asynchronous recursif LU... nb:%d, nbcores:%d, dcpu:%f, panel_nbcores:%d, ngpu: %d\n", async_nb, P, d_cpu, Pr, opts.ngpu);
printf(" M N CPU GFlop/s (sec) GPU GFlop/s (sec) GPU_Async_v2 GFlop/s (sec) GPU_Async_work_v2 GFlop/s (sec)");
if ( opts.check == 2 ) {
printf(" |Ax-b|/(N*|A|*|x|)\n");
}
else {
printf(" |PA-LU|/(N*|A|)\n");
}
printf("=========================================================================\n");
for( int i = 0; i < opts.ntest; ++i ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[i];
N = opts.nsize[i];
min_mn = min(M, N);
lda = M;
n2 = lda*N;
ldda = ((M+31)/32)*32;
//nb = magma_get_dgetrf_nb( M );
gflops = FLOPS_DGETRF( M, N ) / 1e9;
// Allocate host memory for the matrix
TESTING_MALLOC_CPU( ipiv, magma_int_t, min_mn );
TESTING_MALLOC_CPU( h_A, double, n2 );
/*set default number of threads for lapack*/
magma_setlapack_numthreads(P);
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
init_matrix( M, N, h_A, lda );
cpu_time = magma_wtime();
lapackf77_dgetrf( &M, &N, h_A, &lda, ipiv, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_dgetrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
nb = magma_get_dgetrf_nb( M );
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, int((N+nb-1)/nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// Allocate device memory
for( int dev=0; dev < ngpu; dev++){
n_local = ((N/nb)/ngpu)*nb;
if (dev < (N/nb) % ngpu)
n_local += nb;
else if (dev == (N/nb) % ngpu)
n_local += N % nb;
ldn_local = ((n_local+31)/32)*32; // TODO why?
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], double, ldda*ldn_local );
}
init_matrix( M, N, h_A, lda );
magma_dsetmatrix_1D_col_bcyclic( M, N, h_A, lda, d_lA, ldda, ngpu, nb );
gpu_time1 = magma_wtime();
magma_dgetrf_mgpu( ngpu, M, N, d_lA, ldda, ipiv, &info );
gpu_time1 = magma_wtime() - gpu_time1;
gpu_perf1 = gflops / gpu_time1;
if (info != 0)
printf("magma_dgetrf_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_dgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_A, lda, ngpu, nb );
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
/* ====================================================================
Performs operation using MAGMA_Async: This interface allocate workspace internally
=================================================================== */
/*For the benchmark we have 2 approaches*/
/*1. use directly magma_amc */
/*2. use magma_amc_work and add pinned memory time*/
/*We choose approach 2*/
/*
nb = async_nb;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, int((N+nb-1)/nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// Allocate device memory
n_local = numcols2p(0, N, nb, ngpu);
ldn_local = n_local;
//ldn_local = ((n_local+31)/32)*32;
for( int dev=0; dev < ngpu; dev++){
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], double, ldda*ldn_local );
}
init_matrix( M, N, h_A, lda );
magma_dsetmatrix_1D_col_bcyclic( M, N, h_A, lda, d_lA, ldda, ngpu, nb );
// Switch to the sequential version of BLAS
magma_setlapack_numthreads(1);
magma_amc_init(P, d_cpu, Pr, nb);
gpu_time2 = magma_wtime();
magma_dgetrf_async_mgpu( ngpu, M, N, d_lA, ldda, ipiv, &info );
gpu_time2 = magma_wtime() - gpu_time2;
gpu_perf2 = gflops / gpu_time2;
magma_amc_finalize();
if (info != 0)
printf("magma_dgetrf_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_dgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_A, lda, ngpu, nb );
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
*/
/* ====================================================================
Performs operation using MAGMA_Async_Work
=================================================================== */
nb = async_nb;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, int((N+nb-1)/nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// Allocate device memory
n_local = numcols2p(0, N, nb, ngpu);
ldn_local = n_local;
//ldn_local = ((n_local+31)/32)*32;
for( int dev=0; dev < ngpu; dev++){
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], double, ldda*ldn_local );
}
init_matrix( M, N, h_A, lda );
magma_dsetmatrix_1D_col_bcyclic( M, N, h_A, lda, d_lA, ldda, ngpu, nb );
// Switch to the sequential version of BLAS
magma_setlapack_numthreads(1);
//Compute workspace dimension
WORK_LD = M;
NB = (int) ceil( (double) N / nb);
WORK_n = (int) ceil(N*d_cpu)+nb; /*TODO:remove +nb replace with A_N*/
//WORK_n = NSplit(NB, d_cpu)*nb;
if(WORK_n<nb) WORK_n = nb;//make sure workspace has at least one block column
//Make LD and n multiple of 32
//if(WORK_LD%32!=0) WORK_LD = ((WORK_LD + 31)/32)*32;
//if(WORK_n%32!=0) WORK_n = ((WORK_n + 31)/32)*32;
//Allocate workspace
alloc_time = magma_wtime();
if (MAGMA_SUCCESS != magma_dmalloc_pinned(&WORK, WORK_LD*WORK_n)) {
//if (MAGMA_SUCCESS != magma_dmalloc_cpu(&WORK, WORK_LD*WORK_n)) {
info = MAGMA_ERR_HOST_ALLOC;
printf("magma_dmalloc_pinned returned error %d: %s.\n ", (int) info);
}
/* Workspace for the panels on the GPU*/
dlpanelT_m = WORK_n; /*assume that the cpu and gpu use the same buffer size*/
dlpanelT_n = M;
dlpanelT = (double **) malloc(ngpu*sizeof(double*));
for(int dev=0;dev<ngpu;dev++){
magma_setdevice(dev);
if (MAGMA_SUCCESS != magma_dmalloc(&dlpanelT[dev], dlpanelT_m*dlpanelT_n)) {
info = MAGMA_ERR_DEVICE_ALLOC;
printf("magma_dmalloc returned error %d: %s.\n ", (int) info);
}
}
alloc_time = magma_wtime() - alloc_time;
//First touch the workspace with each thread. This may be needed to avoid using numactl --interleave
//magma_amc_dmemset(WORK, 0.0, WORK_LD*WORK_n, 256, P); //nb
//#pragma omp parallel for private(info) schedule(static,nb)
//for(info=0;info<WORK_LD*WORK_n;info++) WORK[info] = 0.0; //alternative first touch by the thread
magma_amc_init(P, d_cpu, Pr, nb);
gpu_time3 = magma_wtime();
magma_dgetrf_mgpu_work_amc_v3(ngpu, M, N, d_lA, ldda, ipiv, &info, WORK, WORK_LD, WORK_n);
gpu_time3 = magma_wtime() - gpu_time3;
gpu_perf3 = gflops / gpu_time3;
magma_amc_finalize();
if (info != 0)
printf("magma_dgetrf_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
magma_dgetmatrix_1D_col_bcyclic( M, N, d_lA, ldda, h_A, lda, ngpu, nb );
//Free workspace
free_time = magma_wtime();
magma_free_pinned(WORK);
for(int dev=0;dev<ngpu;dev++){
magma_setdevice(dev);
magma_free(dlpanelT[dev]);
}
free(dlpanelT);
free_time = magma_wtime() - free_time;
/*DEDUCE t2, JUST FOR THE BENCHMARK*/
gpu_time2 = gpu_time3 + alloc_time + free_time;
gpu_perf2 = gflops / gpu_time2;
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
/* =====================================================================
Check the factorization
=================================================================== */
/*
if ( opts.lapack ) {
printf("%5d %5d %7.2f (%7.2f) %7.2f (%7.2f)",
(int) M, (int) N, cpu_perf, cpu_time, gpu_perf, gpu_time );
}
else {
printf("%5d %5d --- ( --- ) %7.2f (%7.2f)",
(int) M, (int) N, gpu_perf, gpu_time );
}
*/
printf("%5d %5d", (int) M, (int) N);
if(cpu_perf!=0.0){
printf(" %7.2f (%7.2f)", cpu_perf, cpu_time);
}
else{
printf(" --- ( --- )");
}
if(gpu_perf1!=0.0){
printf(" %7.2f (%7.2f)", gpu_perf1, gpu_time1);
}
else{
printf(" --- ( --- )");
}
if(gpu_perf2!=0.0){
printf(" %7.2f (%7.2f)", gpu_perf2, gpu_time2);
}
else{
printf(" --- ( --- )");
}
if(gpu_perf3!=0.0){
printf(" %7.2f (%7.2f)", gpu_perf3, gpu_time3);
}
else{
printf(" --- ( --- )");
}
if ( opts.check == 2 ) {
error = get_residual( M, N, h_A, lda, ipiv );
printf(" %8.2e%s\n", error, (error < tol ? "" : " failed"));
status |= ! (error < tol);
}
else if ( opts.check ) {
error = get_LU_error( M, N, h_A, lda, ipiv );
printf(" %8.2e%s\n", error, (error < tol ? "" : " failed"));
status |= ! (error < tol);
}
else {
printf( " ---\n" );
}
TESTING_FREE_CPU( ipiv );
TESTING_FREE_CPU( h_A );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/* ////////////////////////////////////////////////////////////////////////////
-- Testing dpotrf_mgpu
*/
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, gpu_perf, gpu_time, cpu_perf=0, cpu_time=0;
double error, work[1];
double c_neg_one = MAGMA_D_NEG_ONE;
double *h_A, *h_R;
double *d_lA[ MagmaMaxGPUs ];
magma_int_t N, n2, lda, ldda, max_size, ngpu;
magma_int_t info, nb;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
double tol = opts.tolerance * lapackf77_dlamch("E");
printf("ngpu = %d, uplo = %s\n", (int) opts.ngpu, lapack_uplo_const(opts.uplo) );
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R||_F / ||A||_F\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;
n2 = lda*N;
nb = magma_get_dpotrf_nb( N );
gflops = FLOPS_DPOTRF( N ) / 1e9;
// ngpu must be at least the number of blocks
ngpu = min( opts.ngpu, int((N+nb-1)/nb) );
if ( ngpu < opts.ngpu ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
}
// Allocate host memory for the matrix
TESTING_MALLOC_CPU( h_A, double, n2 );
TESTING_MALLOC_PIN( h_R, double, n2 );
// Allocate device memory
// matrix is distributed by block-rows or block-columns
// this is maximum size that any GPU stores;
// size is rounded up to full blocks in both rows and columns
max_size = nb*(1+N/(nb*ngpu)) * nb*((N+nb-1)/nb);
for( int dev=0; dev < ngpu; dev++ ) {
magma_setdevice( dev );
TESTING_MALLOC_DEV( d_lA[dev], double, max_size );
}
/* Initialize the matrix */
lapackf77_dlarnv( &ione, ISEED, &n2, h_A );
magma_dmake_hpd( N, h_A, lda );
lapackf77_dlacpy( MagmaUpperLowerStr, &N, &N, h_A, &lda, h_R, &lda );
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
lapackf77_dpotrf( lapack_uplo_const(opts.uplo), &N, h_A, &lda, &info );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
if (info != 0)
printf("lapackf77_dpotrf returned error %d: %s.\n",
(int) info, magma_strerror( info ));
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
if ( opts.uplo == MagmaUpper ) {
ldda = ((N+nb-1)/nb)*nb;
magma_dsetmatrix_1D_col_bcyclic( N, N, h_R, lda, d_lA, ldda, ngpu, nb );
} else {
ldda = (1+N/(nb*ngpu))*nb;
magma_dsetmatrix_1D_row_bcyclic( N, N, h_R, lda, d_lA, ldda, ngpu, nb );
}
gpu_time = magma_wtime();
magma_dpotrf_mgpu( ngpu, opts.uplo, N, d_lA, ldda, &info );
gpu_time = magma_wtime() - gpu_time;
gpu_perf = gflops / gpu_time;
if (info != 0)
printf("magma_dpotrf_mgpu returned error %d: %s.\n",
(int) info, magma_strerror( info ));
if ( opts.uplo == MagmaUpper ) {
magma_dgetmatrix_1D_col_bcyclic( N, N, d_lA, ldda, h_R, lda, ngpu, nb );
} else {
magma_dgetmatrix_1D_row_bcyclic( N, N, d_lA, ldda, h_R, lda, ngpu, nb );
}
/* =====================================================================
Check the result compared to LAPACK
=================================================================== */
for( int dev=0; dev < ngpu; dev++ ){
magma_setdevice( dev );
magma_device_sync();
}
if ( opts.lapack ) {
error = lapackf77_dlange("f", &N, &N, h_A, &lda, work );
blasf77_daxpy( &n2, &c_neg_one, h_A, &ione, h_R, &ione );
error = lapackf77_dlange("f", &N, &N, h_R, &lda, work ) / error;
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) N, cpu_perf, cpu_time, gpu_perf, gpu_time,
error, (error < tol ? "ok" : "failed") );
status += ! (error < tol);
}
else {
printf("%5d --- ( --- ) %7.2f (%7.2f) ---\n",
(int) N, gpu_perf, gpu_time );
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_PIN( h_R );
for( int dev=0; dev < ngpu; dev++ ){
magma_setdevice( dev );
TESTING_FREE_DEV( d_lA[dev] );
}
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
DGEQRF4 computes a QR factorization of a DOUBLE_PRECISION M-by-N matrix A:
A = Q * R using multiple GPUs. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
Arguments
---------
@param[in]
num_gpus INTEGER
The number of GPUs to be used for the factorization.
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A DOUBLE_PRECISION array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
tau DOUBLE_PRECISION array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) DOUBLE_PRECISION array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
\n
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= N*NB,
where NB can be obtained through magma_get_dgeqrf_nb(M).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the WORK array, returns
this value as the first entry of the WORK array, and no error
message related to LWORK is issued.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
---------------
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a real scalar, and v is a real vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
@ingroup magma_dgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_dgeqrf4(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
double *A, magma_int_t lda, double *tau,
double *work, magma_int_t lwork,
magma_int_t *info )
{
double *da[MagmaMaxGPUs];
double c_one = MAGMA_D_ONE;
int i, k, ldda;
*info = 0;
int nb = magma_get_dgeqrf_nb(min(m, n));
int lwkopt = n * nb;
work[0] = MAGMA_D_MAKE( (double)lwkopt, 0 );
int lquery = (lwork == -1);
if (num_gpus < 0 || num_gpus > MagmaMaxGPUs) {
*info = -1;
} else if (m < 0) {
*info = -2;
} else if (n < 0) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
} else if (lwork < max(1,n) && ! lquery) {
*info = -8;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
work[0] = c_one;
return *info;
}
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
ldda = ((m+31)/32)*32;
magma_int_t n_local[MagmaMaxGPUs];
for (i=0; i < num_gpus; i++) {
n_local[i] = ((n/nb)/num_gpus)*nb;
if (i < (n/nb)%num_gpus)
n_local[i] += nb;
else if (i == (n/nb)%num_gpus)
n_local[i] += n%nb;
magma_setdevice(i);
// TODO on failure, free previously allocated memory
if (MAGMA_SUCCESS != magma_dmalloc( &da[i], ldda*n_local[i] )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
if (m > nb && n > nb) {
/* Copy the matrix to the GPUs in 1D block cyclic distribution */
magma_dsetmatrix_1D_col_bcyclic(m, n, A, lda, da, ldda, num_gpus, nb);
/* Factor using the GPU interface */
magma_dgeqrf2_mgpu( num_gpus, m, n, da, ldda, tau, info);
/* Copy the matrix back from the GPUs to the CPU */
magma_dgetmatrix_1D_col_bcyclic(m, n, da, ldda, A, lda, num_gpus, nb);
}
else {
lapackf77_dgeqrf(&m, &n, A, &lda, tau, work, &lwork, info);
}
/* Free the allocated GPU memory */
for (i=0; i < num_gpus; i++) {
magma_setdevice(i);
magma_free( da[i] );
}
magma_setdevice( orig_dev );
return *info;
} /* magma_dgeqrf4 */
