extern "C" magma_int_t
magma_dgetrf_mgpu(magma_int_t num_gpus,
magma_int_t m, magma_int_t n,
double **d_lA, magma_int_t ldda,
magma_int_t *ipiv, magma_int_t *info)
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2012
Purpose
=======
DGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges.
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.
===================================================================== */
#define inAT(id,i,j) (d_lAT[(id)] + (i)*nb*lddat + (j)*nb)
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t iinfo, nb, n_local[MagmaMaxGPUs];
magma_int_t maxm, mindim;
magma_int_t i, j, d, rows, cols, s, lddat, lddwork;
magma_int_t id, i_local, i_local2, nb0, nb1;
double *d_lAT[MagmaMaxGPUs];
double *d_panel[MagmaMaxGPUs], *work;
cudaStream_t streaml[4][2];
/* Check arguments */
*info = 0;
if (m < 0)
*info = -2;
else if (n < 0)
*info = -3;
else if (ldda < 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;
/* Function Body */
mindim = min(m, n);
nb = magma_get_dgetrf_nb(m);
if (nb <= 1 || nb >= n) {
/* Use CPU code. */
magma_dmalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_dgetmatrix( m, n, d_lA[0], ldda, work, m );
lapackf77_dgetrf(&m, &n, work, &m, ipiv, info);
magma_dsetmatrix( m, n, work, m, d_lA[0], ldda );
magma_free_cpu(work);
} else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
if( num_gpus > ceil((double)n/nb) ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) num_gpus );
*info = -1;
return *info;
}
/* allocate workspace for each GPU */
lddat = ((((((n+nb-1)/nb)/num_gpus)*nb)+31)/32)*32;
lddat = (n+nb-1)/nb; /* number of block columns */
lddat = (lddat+num_gpus-1)/num_gpus; /* number of block columns per GPU */
lddat = nb*lddat; /* number of columns per GPU */
lddat = ((lddat+31)/32)*32; /* make it a multiple of 32 */
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
/* local-n and local-ld */
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;
/* workspaces */
if (MAGMA_SUCCESS != magma_dmalloc( &d_panel[i], 3*nb*maxm )) {
for( j=0; j<=i; j++ ) {
magma_setdevice(j);
}
for( j=0; j<i; j++ ) {
magma_setdevice(j);
magma_free( d_panel[j] );
magma_free( d_lAT[j] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/* local-matrix storage */
if (MAGMA_SUCCESS != magma_dmalloc( &d_lAT[i], lddat*maxm )) {
for( j=0; j<=i; j++ ) {
magma_setdevice(j);
magma_free( d_panel[j] );
}
for( j=0; j<i; j++ ) {
magma_setdevice(j);
magma_free( d_lAT[j] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/* create the streams */
magma_queue_create( &streaml[i][0] );
magma_queue_create( &streaml[i][1] );
magmablasSetKernelStream(streaml[i][1]);
magmablas_dtranspose2( d_lAT[i], lddat, d_lA[i], ldda, m, n_local[i] );
}
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
cudaStreamSynchronize(streaml[i][0]);
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
/* cpu workspace */
lddwork = maxm;
if (MAGMA_SUCCESS != magma_dmalloc_pinned( &work, lddwork*nb*num_gpus )) {
for(i=0; i<num_gpus; i++ ) {
magma_setdevice(i);
magma_free( d_panel[i] );
magma_free( d_lAT[i] );
}
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
/* calling multi-gpu interface with allocated workspaces and streams */
//magma_dgetrf1_mgpu( num_gpus, m, n, nb, 0, d_lAT, lddat, ipiv, d_panel, work, maxm,
// (cudaStream_t **)streaml, info );
magma_dgetrf2_mgpu(num_gpus, m, n, nb, 0, d_lAT, lddat, ipiv, d_panel, work, maxm,
streaml, info);
/* clean up */
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
/* save on output */
magmablas_dtranspose2( d_lA[d], ldda, d_lAT[d], lddat, n_local[d], m );
magma_device_sync();
magma_free( d_lAT[d] );
magma_free( d_panel[d] );
magma_queue_destroy( streaml[d][0] );
magma_queue_destroy( streaml[d][1] );
magmablasSetKernelStream(NULL);
} /* end of for d=1,..,num_gpus */
magma_setdevice(0);
magma_free_pinned( work );
}
return *info;
/* End of MAGMA_DGETRF_MGPU */
}
/**
Purpose
-------
DGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges.
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
---------
@param[in]
ngpu INTEGER
Number of GPUs to use. ngpu > 0.
@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]
d_lA DOUBLE PRECISION array of pointers on the GPU, dimension (ngpu).
On entry, the M-by-N matrix A distributed over GPUs
(d_lA[d] points to the local matrix on d-th GPU).
It uses 1D block column cyclic format with the block size of nb,
and each local matrix is stored by column.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
@param[in]
ldda INTEGER
The leading dimension of the array d_lA. LDDA >= max(1,M).
@param[out]
ipiv 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).
@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.
- > 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.
@ingroup magma_dgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_dgetrf_mgpu(
magma_int_t ngpu,
magma_int_t m, magma_int_t n,
magmaDouble_ptr d_lA[], magma_int_t ldda, magma_int_t *ipiv,
magma_int_t *info)
{
magma_int_t nb, n_local[MagmaMaxGPUs];
magma_int_t maxm;
magma_int_t i, j, d, lddat, lddwork;
double *d_lAT[MagmaMaxGPUs];
double *d_panel[MagmaMaxGPUs], *work;
magma_queue_t queues[MagmaMaxGPUs][2];
/* Check arguments */
*info = 0;
if (m < 0)
*info = -2;
else if (n < 0)
*info = -3;
else if (ldda < 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;
/* create the queues */
for( d=0; d < ngpu; d++ ) {
magma_queue_create( d, &queues[d][0] );
magma_queue_create( d, &queues[d][1] );
}
/* Function Body */
nb = magma_get_dgetrf_nb( m, n );
if (nb <= 1 || nb >= n) {
/* Use CPU code. */
magma_dmalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_dgetmatrix( m, n, d_lA[0], ldda, work, m, queues[0][0] );
lapackf77_dgetrf(&m, &n, work, &m, ipiv, info);
magma_dsetmatrix( m, n, work, m, d_lA[0], ldda, queues[0][0] );
magma_free_cpu(work);
} else {
/* Use hybrid blocked code. */
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
maxm = magma_roundup( m, 32 );
if ( ngpu > ceil((double)n/nb) ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) ngpu );
*info = -1;
return *info;
}
/* allocate workspace for each GPU */
lddat = magma_roundup( ((magma_ceildiv( n, nb )/ngpu)*nb), 32 );
lddat = magma_ceildiv( n, nb ); /* number of block columns */
lddat = magma_ceildiv( lddat, ngpu ); /* number of block columns per GPU */
lddat = nb*lddat; /* number of columns per GPU */
lddat = magma_roundup( lddat, 32 ); /* make it a multiple of 32 */
for (i=0; i < ngpu; i++) {
magma_setdevice(i);
/* local-n and local-ld */
n_local[i] = ((n/nb)/ngpu)*nb;
if (i < (n/nb)%ngpu)
n_local[i] += nb;
else if (i == (n/nb)%ngpu)
n_local[i] += n%nb;
/* workspaces */
if (MAGMA_SUCCESS != magma_dmalloc( &d_panel[i], (3+ngpu)*nb*maxm )) {
for( j=0; j <= i; j++ ) {
magma_setdevice(j);
}
for( j=0; j < i; j++ ) {
magma_setdevice(j);
magma_free( d_panel[j] );
magma_free( d_lAT[j] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/* local-matrix storage */
if (MAGMA_SUCCESS != magma_dmalloc( &d_lAT[i], lddat*maxm )) {
for( j=0; j <= i; j++ ) {
magma_setdevice(j);
magma_free( d_panel[j] );
}
for( j=0; j < i; j++ ) {
magma_setdevice(j);
magma_free( d_lAT[j] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_dtranspose( m, n_local[i], d_lA[i], ldda, d_lAT[i], lddat, queues[i][1] );
}
for (i=0; i < ngpu; i++) {
magma_setdevice(i);
magma_queue_sync(queues[i][0]);
}
magma_setdevice(0);
/* cpu workspace */
lddwork = maxm;
if (MAGMA_SUCCESS != magma_dmalloc_pinned( &work, lddwork*nb*ngpu )) {
for (i=0; i < ngpu; i++ ) {
magma_setdevice(i);
magma_free( d_panel[i] );
magma_free( d_lAT[i] );
}
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
/* calling multi-gpu interface with allocated workspaces and queues */
magma_dgetrf2_mgpu(ngpu, m, n, nb, 0, d_lAT, lddat, ipiv, d_panel, work, maxm,
queues, info);
/* clean up */
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
/* save on output */
magmablas_dtranspose( n_local[d], m, d_lAT[d], lddat, d_lA[d], ldda, queues[d][0] );
magma_queue_sync(queues[d][0]);
magma_queue_sync(queues[d][1]);
magma_free( d_lAT[d] );
magma_free( d_panel[d] );
} /* end of for d=1,..,ngpu */
magma_setdevice( orig_dev );
magma_free_pinned( work );
}
/* clean up */
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_queue_destroy( queues[d][0] );
magma_queue_destroy( queues[d][1] );
}
return *info;
}
/**
Purpose
-------
DGETRF_m computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. This version does not
require work space on the GPU passed as input. GPU memory is allocated
in the routine. The matrix may not fit entirely in the GPU memory.
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.
Note: The factorization of big panel is done calling multiple-gpu-interface.
Pivots are applied on GPU within the big panel.
Arguments
---------
@param[in]
num_gpus INTEGER
The number of GPUs. num_gpus > 0.
@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 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.
\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]
ipiv 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).
@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.
- > 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.
@ingroup magma_dgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_dgetrf_m(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
double *A, magma_int_t lda,
magma_int_t *ipiv, magma_int_t *info)
{
#define A(i,j) (A + (j)*lda + (i))
#define dAT(d,i,j) (dAT[d] + (i)*nb*ldn_local + (j)*nb)
#define dPT(d,i,j) (dPT[d] + (i)*nb*nb + (j)*nb*maxm)
magma_timer_t time=0, time_total=0, time_alloc=0, time_set=0, time_get=0, time_comp=0;
timer_start( time_total );
real_Double_t flops;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
double *dAT[MagmaMaxGPUs], *dA[MagmaMaxGPUs], *dPT[MagmaMaxGPUs];
magma_int_t iinfo = 0, nb, nbi, maxm, n_local[MagmaMaxGPUs], ldn_local;
magma_int_t N, M, NB, NBk, I, d, num_gpus0 = num_gpus;
magma_int_t ii, jj, h, offset, ib, rows, s;
magma_queue_t stream[MagmaMaxGPUs][2];
magma_event_t event[MagmaMaxGPUs][2];
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* initialize nb */
nb = magma_get_dgetrf_nb(m);
maxm = ((m + 31)/32)*32;
/* figure out NB */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(double);
/* number of columns in the big panel */
h = 1+(2+num_gpus0);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
const char* ngr_nb_char = getenv("MAGMA_NGR_NB");
if ( ngr_nb_char != NULL ) NB = max( nb, min( NB, atoi(ngr_nb_char) ) );
//NB = 5*max(nb,32);
if ( num_gpus0 > ceil((double)NB/nb) ) {
num_gpus = (int)ceil((double)NB/nb);
h = 1+(2+num_gpus);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
} else {
num_gpus = num_gpus0;
}
if ( num_gpus*NB >= n ) {
#ifdef CHECK_DGETRF_OOC
printf( " * still fit in GPU memory.\n" );
#endif
NB = n;
} else {
#ifdef CHECK_DGETRF_OOC
printf( " * don't fit in GPU memory.\n" );
#endif
NB = num_gpus*NB;
NB = max( nb, (NB / nb) * nb); /* making sure it's devisable by nb (x64) */
}
#ifdef CHECK_DGETRF_OOC
if ( NB != n ) printf( " * running in out-core mode (n=%d, NB=%d, nb=%d, freeMem=%.2e).\n", n, NB, nb, (double)freeMem );
else printf( " * running in in-core mode (n=%d, NB=%d, nb=%d, freeMem=%.2e).\n", n, NB, nb, (double)freeMem );
#endif
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code for scalar of one tile. */
lapackf77_dgetrf(&m, &n, A, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
/* allocate memory on GPU to store the big panel */
timer_start( time_alloc );
n_local[0] = (NB/nb)/num_gpus;
if ( NB%(nb*num_gpus) != 0 ) n_local[0] ++;
n_local[0] *= nb;
ldn_local = ((n_local[0]+31)/32)*32;
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
if (MAGMA_SUCCESS != magma_dmalloc( &dA[d], (ldn_local+h*nb)*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dPT[d] = dA[d] + nb*maxm; /* for storing the previous panel from CPU */
dAT[d] = dA[d] + h*nb*maxm; /* for storing the big panel */
magma_queue_create( &stream[d][0] );
magma_queue_create( &stream[d][1] );
magma_event_create( &event[d][0] );
magma_event_create( &event[d][1] );
}
//magma_setdevice(0);
timer_stop( time_alloc );
for( I=0; I < n; I += NB ) {
M = m;
N = min( NB, n-I ); /* number of columns in this big panel */
s = min( max(m-I,0), N )/nb; /* number of small block-columns in this big panel */
maxm = ((M + 31)/32)*32;
if ( num_gpus0 > ceil((double)N/nb) ) {
num_gpus = (int)ceil((double)N/nb);
} else {
num_gpus = num_gpus0;
}
for( d=0; d < num_gpus; d++ ) {
n_local[d] = ((N/nb)/num_gpus)*nb;
if (d < (N/nb)%num_gpus)
n_local[d] += nb;
else if (d == (N/nb)%num_gpus)
n_local[d] += N%nb;
}
ldn_local = ((n_local[0]+31)/32)*32;
/* upload the next big panel into GPU, transpose (A->A'), and pivot it */
timer_start( time );
magmablas_dsetmatrix_transpose_mgpu(num_gpus, stream, A(0,I), lda,
dAT, ldn_local, dA, maxm, M, N, nb);
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
time_set += timer_stop( time );
timer_start( time );
/* == --------------------------------------------------------------- == */
/* == loop around the previous big-panels to update the new big-panel == */
for( offset = 0; offset < min(m,I); offset += NB ) {
NBk = min( m-offset, NB );
/* start sending the first tile from the previous big-panels to gpus */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
nbi = min( nb, NBk );
magma_dsetmatrix_async( (M-offset), nbi,
A(offset,offset), lda,
dA[d], (maxm-offset), stream[d][0] );
/* make sure the previous update finished */
magmablasSetKernelStream(stream[d][0]);
//magma_queue_sync( stream[d][1] );
magma_queue_wait_event( stream[d][0], event[d][0] );
/* transpose */
magmablas_dtranspose( M-offset, nbi, dA[d], maxm-offset, dPT(d,0,0), nb );
}
/* applying the pivot from the previous big-panel */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][1]);
magmablas_dpermute_long3( dAT(d,0,0), ldn_local, ipiv, NBk, offset );
}
/* == going through each block-column of previous big-panels == */
for( jj=0, ib=offset/nb; jj < NBk; jj += nb, ib++ ) {
ii = offset+jj;
rows = maxm - ii;
nbi = min( nb, NBk-jj );
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
/* wait for a block-column on GPU */
magma_queue_sync( stream[d][0] );
/* start sending next column */
if ( jj+nb < NBk ) {
magma_dsetmatrix_async( (M-ii-nb), min(nb,NBk-jj-nb),
A(ii+nb,ii+nb), lda,
dA[d], (rows-nb), stream[d][0] );
/* make sure the previous update finished */
magmablasSetKernelStream(stream[d][0]);
//magma_queue_sync( stream[d][1] );
magma_queue_wait_event( stream[d][0], event[d][(1+jj/nb)%2] );
/* transpose next column */
magmablas_dtranspose( M-ii-nb, nb, dA[d], rows-nb, dPT(d,0,(1+jj/nb)%2), nb );
}
/* update with the block column */
magmablasSetKernelStream(stream[d][1]);
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n_local[d], nbi, c_one, dPT(d,0,(jj/nb)%2), nb, dAT(d,ib,0), ldn_local );
if ( M > ii+nb ) {
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n_local[d], M-(ii+nb), nbi, c_neg_one, dAT(d,ib,0), ldn_local,
dPT(d,1,(jj/nb)%2), nb, c_one, dAT(d,ib+1,0), ldn_local );
}
magma_event_record( event[d][(jj/nb)%2], stream[d][1] );
} /* end of for each block-columns in a big-panel */
}
} /* end of for each previous big-panels */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
/* calling magma-gpu interface to panel-factorize the big panel */
if ( M > I ) {
//magma_dgetrf1_mgpu(num_gpus, M-I, N, nb, I, dAT, ldn_local, ipiv+I, dA, A(0,I), lda,
// (magma_queue_t **)stream, &iinfo);
magma_dgetrf2_mgpu(num_gpus, M-I, N, nb, I, dAT, ldn_local, ipiv+I, dA, A(0,I), lda,
stream, &iinfo);
if ( iinfo < 0 ) {
*info = iinfo;
break;
} else if ( iinfo != 0 ) {
*info = iinfo + I * NB;
//break;
}
/* adjust pivots */
for( ii=I; ii < min(I+N,m); ii++ )
ipiv[ii] += I;
}
time_comp += timer_stop( time );
/* download the current big panel to CPU */
timer_start( time );
magmablas_dgetmatrix_transpose_mgpu(num_gpus, stream, dAT, ldn_local, A(0,I), lda, dA, maxm, M, N, nb);
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
time_get += timer_stop( time );
} /* end of for */
timer_stop( time_total );
flops = FLOPS_DGETRF( m, n ) / 1e9;
timer_printf(" memory-allocation time: %e\n", time_alloc );
timer_printf(" NB=%d nb=%d\n", (int) NB, (int) nb );
timer_printf(" memcopy and transpose %e seconds\n", time_set );
timer_printf(" total time %e seconds\n", time_total );
timer_printf(" Performance %f GFlop/s, %f seconds without htod and dtoh\n", flops / (time_comp), time_comp );
timer_printf(" Performance %f GFlop/s, %f seconds with htod\n", flops / (time_comp + time_set), time_comp + time_set );
timer_printf(" Performance %f GFlop/s, %f seconds with dtoh\n", flops / (time_comp + time_get), time_comp + time_get );
timer_printf(" Performance %f GFlop/s, %f seconds without memory-allocation\n", flops / (time_total - time_alloc), time_total - time_alloc );
for( d=0; d < num_gpus0; d++ ) {
magma_setdevice(d);
magma_free( dA[d] );
magma_event_destroy( event[d][0] );
magma_event_destroy( event[d][1] );
magma_queue_destroy( stream[d][0] );
magma_queue_destroy( stream[d][1] );
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
}
if ( *info >= 0 ) magma_dgetrf_piv(m, n, NB, A, lda, ipiv, info);
return *info;
} /* magma_dgetrf_m */
extern "C" magma_int_t
magma_dgetrf_m(magma_int_t num_gpus0, magma_int_t m, magma_int_t n, double *a, magma_int_t lda,
magma_int_t *ipiv, magma_int_t *info)
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2010
Purpose
=======
DGETRF_m computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. This version does not
require work space on the GPU passed as input. GPU memory is allocated
in the routine. The matrix may not fit entirely in the GPU memory.
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.
Note: The factorization of big panel is done calling multiple-gpu-interface.
Pivots are applied on GPU within the big panel.
Arguments
=========
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, dimension (LDA,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.
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= 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.
===================================================================== */
#define A(i,j) (a + (j)*lda + (i))
#define inAT(d,i,j) (dAT[d] + (i)*nb*ldn_local + (j)*nb)
#define inPT(d,i,j) (dPT[d] + (i)*nb*nb + (j)*nb*maxm)
/* Flops formula */
//#define PROFILE
#ifdef PROFILE
double flops, time_rmajor = 0, time_rmajor2 = 0, time_rmajor3 = 0, time_mem = 0;
magma_timestr_t start, start1, start2, end1, end, start0 = get_current_time();
#define FMULS_GETRF(__m, __n) ( ((__m) < (__n)) ? (0.5 * (__m) * ((__m) * ((__n) - (1./3.) * (__m) - 1. ) + (__n)) + (2. / 3.) * (__m)) \
: (0.5 * (__n) * ((__n) * ((__m) - (1./3.) * (__n) - 1. ) + (__m)) + (2. / 3.) * (__n)) )
#define FADDS_GETRF(__m, __n) ( ((__m) < (__n)) ? (0.5 * (__m) * ((__m) * ((__n) - (1./3.) * (__m) ) - (__n)) + (1. / 6.) * (__m)) \
: (0.5 * (__n) * ((__n) * ((__m) - (1./3.) * (__n) ) - (__m)) + (1. / 6.) * (__n)) )
#define PRECISION_d
#if defined(PRECISION_z) || defined(PRECISION_c)
#define FLOPS(m, n) ( 6. * FMULS_GETRF(m, n) + 2. * FADDS_GETRF(m, n) )
#else
#define FLOPS(m, n) ( FMULS_GETRF(m, n) + FADDS_GETRF(m, n) )
#endif
#endif
double *dAT[4], *dA[4], *dPT[4];
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t iinfo = 0, nb, nbi, maxm, n_local[4], ldn_local;
magma_int_t N, M, NB, NBk, I, d, num_gpus;
magma_int_t i, ii, jj, h = 3, offset, ib, rows, s;
cudaStream_t stream[4][2];
cudaEvent_t event[4][2];
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* initialize nb */
nb = magma_get_dgetrf_nb(m);
maxm = ((m + 31)/32)*32;
/* figure out NB */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(double);
/* number of columns in the big panel */
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
char * ngr_nb_char = getenv("MAGMA_NGR_NB");
if( ngr_nb_char != NULL ) NB = max( nb, min( NB, atoi(ngr_nb_char) ) );
//NB = 5*max(nb,32);
if( num_gpus0 > ceil((double)NB/nb) ) {
num_gpus = (int)ceil((double)NB/nb);
} else {
num_gpus = num_gpus0;
}
if( num_gpus*NB >= n ) {
#ifdef CHECK_DGETRF_OOC
printf( " * still fit in GPU memory.\n" );
#endif
NB = n;
} else {
#ifdef CHECK_DGETRF_OOC
printf( " * don't fit in GPU memory.\n" );
#endif
NB = num_gpus*NB;
NB = max(nb,(NB / nb) * nb); /* making sure it's devisable by nb (x64) */
}
#ifdef CHECK_DGETRF_OOC
if( NB != n ) printf( " * running in out-core mode (n=%d, NB=%d, nb=%d).\n",n,NB,nb );
else printf( " * running in in-core mode (n=%d, NB=%d, nb=%d).\n",n,NB,nb );
fflush(stdout);
#endif
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code for scalar of one tile. */
lapackf77_dgetrf(&m, &n, a, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
/* allocate memory on GPU to store the big panel */
#ifdef PROFILE
start = get_current_time();
#endif
n_local[0] = (NB/nb)/num_gpus;
if( NB%(nb*num_gpus) != 0 ) n_local[0] ++;
n_local[0] *= nb;
ldn_local = ((n_local[0]+31)/32)*32;
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
if (MAGMA_SUCCESS != magma_dmalloc( &dA[d], (h*nb + ldn_local)*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dPT[d] = dA[d] + nb*maxm; /* for storing the previous panel from CPU */
dAT[d] = dA[d] + h*nb*maxm; /* for storing the big panel */
magma_queue_create( &stream[d][0] );
magma_queue_create( &stream[d][1] );
magma_event_create( &event[d][0] );
magma_event_create( &event[d][1] );
}
//magma_setdevice(0);
#ifdef PROFILE
end = get_current_time();
printf( " memory-allocation time: %e\n",GetTimerValue(start, end)/1000.0 );
start = get_current_time();
#endif
for( I=0; I<n; I+=NB ) {
M = m;
N = min( NB, n-I ); /* number of columns in this big panel */
s = min(max(m-I,0),N)/nb; /* number of small block-columns in this big panel */
maxm = ((M + 31)/32)*32;
if( num_gpus0 > ceil((double)N/nb) ) {
num_gpus = (int)ceil((double)N/nb);
} else {
num_gpus = num_gpus0;
}
for( d=0; d<num_gpus; d++ ) {
n_local[d] = ((N/nb)/num_gpus)*nb;
if (d < (N/nb)%num_gpus)
n_local[d] += nb;
else if (d == (N/nb)%num_gpus)
n_local[d] += N%nb;
}
ldn_local = ((n_local[0]+31)/32)*32;
#ifdef PROFILE
start2 = get_current_time();
#endif
/* upload the next big panel into GPU, transpose (A->A'), and pivot it */
magmablas_dsetmatrix_transpose_mgpu(num_gpus, stream, A(0,I), lda,
dAT, ldn_local, dA, maxm, M, N, nb);
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
#ifdef PROFILE
start1 = get_current_time();
#endif
/* == --------------------------------------------------------------- == */
/* == loop around the previous big-panels to update the new big-panel == */
for( offset = 0; offset<min(m,I); offset+=NB )
{
NBk = min( m-offset, NB );
/* start sending the first tile from the previous big-panels to gpus */
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
nbi = min( nb, NBk );
magma_dsetmatrix_async( (M-offset), nbi,
A(offset,offset), lda,
dA[d], (maxm-offset), stream[d][0] );
/* make sure the previous update finished */
magmablasSetKernelStream(stream[d][0]);
//magma_queue_sync( stream[d][1] );
magma_queue_wait_event( stream[d][0], event[d][0] );
/* transpose */
magmablas_dtranspose2( inPT(d,0,0), nb, dA[d], maxm-offset, M-offset, nbi);
}
/* applying the pivot from the previous big-panel */
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][1]);
magmablas_dpermute_long3( inAT(d,0,0), ldn_local, ipiv, NBk, offset );
}
/* == going through each block-column of previous big-panels == */
for( jj=0, ib=offset/nb; jj<NBk; jj+=nb, ib++ )
{
ii = offset+jj;
rows = maxm - ii;
nbi = min( nb, NBk-jj );
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
/* wait for a block-column on GPU */
magma_queue_sync( stream[d][0] );
/* start sending next column */
if( jj+nb < NBk ) {
magma_dsetmatrix_async( (M-ii-nb), min(nb,NBk-jj-nb),
A(ii+nb,ii+nb), lda,
dA[d], (rows-nb), stream[d][0] );
/* make sure the previous update finished */
magmablasSetKernelStream(stream[d][0]);
//magma_queue_sync( stream[d][1] );
magma_queue_wait_event( stream[d][0], event[d][(1+jj/nb)%2] );
/* transpose next column */
magmablas_dtranspose2( inPT(d,0,(1+jj/nb)%2), nb, dA[d], rows-nb, M-ii-nb, nb);
}
/* update with the block column */
magmablasSetKernelStream(stream[d][1]);
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n_local[d], nbi, c_one, inPT(d,0,(jj/nb)%2), nb, inAT(d,ib,0), ldn_local );
if( M > ii+nb ) {
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n_local[d], M-(ii+nb), nbi, c_neg_one, inAT(d,ib,0), ldn_local,
inPT(d,1,(jj/nb)%2), nb, c_one, inAT(d,ib+1,0), ldn_local );
}
magma_event_record( event[d][(jj/nb)%2], stream[d][1] );
} /* end of for each block-columns in a big-panel */
}
} /* end of for each previous big-panels */
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
/* calling magma-gpu interface to panel-factorize the big panel */
if( M > I ) {
//magma_dgetrf1_mgpu(num_gpus, M-I, N, nb, I, dAT, ldn_local, ipiv+I, dA, &a[I*lda], lda,
// (cudaStream_t **)stream, &iinfo);
magma_dgetrf2_mgpu(num_gpus, M-I, N, nb, I, dAT, ldn_local, ipiv+I, dA, &a[I*lda], lda,
stream, &iinfo);
if( iinfo < 0 ) {
*info = iinfo;
break;
} else if( iinfo != 0 ) {
*info = iinfo + I * NB;
//break;
}
/* adjust pivots */
for( ii=I; ii<min(I+N,m); ii++ ) ipiv[ii] += I;
}
#ifdef PROFILE
end1 = get_current_time();
time_rmajor += GetTimerValue(start1, end1);
time_rmajor3 += GetTimerValue(start2, end1);
time_mem += (GetTimerValue(start2, end1)-GetTimerValue(start1, end1))/1000.0;
#endif
/* download the current big panel to CPU */
magmablas_dgetmatrix_transpose_mgpu(num_gpus, stream, dAT, ldn_local, A(0,I), lda, dA, maxm, M, N, nb);
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
#ifdef PROFILE
end1 = get_current_time();
time_rmajor2 += GetTimerValue(start1, end1);
#endif
} /* end of for */
#ifdef PROFILE
end = get_current_time();
flops = FLOPS( (double)m, (double)n ) / 1000000;
printf(" NB=%d nb=%d\n",NB,nb);
printf(" memcopy and transpose %e seconds\n",time_mem );
printf(" total time %e seconds\n",GetTimerValue(start0,end)/1000.0);
printf(" Performance %f GFlop/s, %f seconds without htod and dtoh\n", flops / time_rmajor, time_rmajor /1000.0);
printf(" Performance %f GFlop/s, %f seconds with htod\n", flops / time_rmajor3, time_rmajor3/1000.0);
printf(" Performance %f GFlop/s, %f seconds with dtoh\n", flops / time_rmajor2, time_rmajor2/1000.0);
printf(" Performance %f GFlop/s, %f seconds without memory-allocation\n", flops / GetTimerValue(start, end), GetTimerValue(start,end)/1000.0);
#endif
for( d=0; d<num_gpus0; d++ ) {
magma_setdevice(d);
magma_free( dA[d] );
magma_event_destroy( event[d][0] );
magma_event_destroy( event[d][1] );
magma_queue_destroy( stream[d][0] );
magma_queue_destroy( stream[d][1] );
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
}
return *info;
} /* magma_dgetrf_m */
