//
// m, n - dimensions in the source (input) matrix.
// This routine copies the hA matrix from the CPU
// to dAT on the GPU. In addition, the output matrix
// is transposed. The routine uses a buffer of size
// 2*lddw*nb pointed to by dwork (lddw > m) on the GPU.
// Note that lda >= m and lddat >= n.
//
extern "C" void
magmablas_ssetmatrix_transpose_mgpu(
magma_int_t ngpu, magma_queue_t queues[][2],
const float *hA, magma_int_t lda,
magmaFloat_ptr dAT[], magma_int_t ldda,
magmaFloat_ptr dwork[], magma_int_t lddw,
magma_int_t m, magma_int_t n, magma_int_t nb)
{
#define hA(j) (hA + (j)*lda)
#define dwork(d, j) (dwork[(d)] + (j)*nb*lddw)
#define dAT(d, j) (dAT[(d)] + (j)*nb)
magma_int_t nqueues = 2, d, j, j_local, id, ib;
/* Quick return */
if ( (m == 0) || (n == 0) )
return;
if (lda < m || ngpu*ldda < n || lddw < m) {
fprintf( stderr, "%s: wrong arguments (%d<%d), (%d*%d<%d), or (%d<%d).\n",
__func__, (int) lda, (int) m, (int) ngpu, (int) ldda, (int) n, (int) lddw, (int) m );
return;
}
/* Move data from CPU to GPU by block columns and transpose it */
for (j=0; j < n; j += nb) {
d = (j/nb)%ngpu;
j_local = (j/nb)/ngpu;
id = j_local%nqueues;
magma_setdevice(d);
ib = min(n-j, nb);
magma_ssetmatrix_async( m, ib,
hA(j), lda,
dwork(d, id), lddw,
queues[d][id] );
magmablas_stranspose_q( m, ib, dwork(d,id), lddw, dAT(d,j_local), ldda, queues[d][id] );
}
}
/**
Purpose
-------
SGETRF 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 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.
If the current stream is NULL, this version replaces it with a new
stream to overlap computation with communication.
Arguments
---------
@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 REAL 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_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrf(
magma_int_t m, magma_int_t n, float *A, magma_int_t lda,
magma_int_t *ipiv,
magma_int_t *info)
{
#define dAT(i_, j_) (dAT + (i_)*nb*ldda + (j_)*nb)
float *dAT, *dA, *da, *work;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t iinfo, nb;
/* Check arguments */
*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;
/* Function Body */
nb = magma_get_sgetrf_nb(m);
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_sgetrf(&m, &n, A, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, maxdim;
magma_int_t i, j, rows, cols, s = min(m, n)/nb;
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
maxdim = max(maxm, maxn);
/* set number of GPUs */
magma_int_t ngpu = magma_num_gpus();
if ( ngpu > 1 ) {
/* call multi-GPU non-GPU-resident interface */
magma_sgetrf_m(ngpu, m, n, A, lda, ipiv, info);
return *info;
}
/* explicitly checking the memory requirement */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(float);
int h = 1+(2+ngpu), ngpu2 = ngpu;
int 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) ) );
if ( ngpu > ceil((float)NB/nb) ) {
ngpu2 = (int)ceil((float)NB/nb);
h = 1+(2+ngpu2);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
}
if ( ngpu2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
magma_sgetrf_m(ngpu, m, n, A, lda, ipiv, info);
return *info;
}
ldda = maxn;
work = A;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
if (MAGMA_SUCCESS != magma_smalloc( &dA, nb*maxm + maxdim*maxdim )) {
/* alloc failed so call non-GPU-resident version */
magma_sgetrf_m(ngpu, m, n, A, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
ldda = maxdim;
magma_ssetmatrix( m, n, A, lda, da, ldda );
dAT = da;
magmablas_stranspose_inplace( ldda, dAT, ldda );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
if (MAGMA_SUCCESS != magma_smalloc( &dA, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
magma_sgetrf_m(ngpu, m, n, A, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
magma_ssetmatrix( m, n, A, lda, da, maxm );
if (MAGMA_SUCCESS != magma_smalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dA );
magma_sgetrf_m(ngpu, m, n, A, lda, ipiv, info);
return *info;
}
magmablas_stranspose( m, n, da, maxm, dAT, ldda );
}
lapackf77_sgetrf( &m, &nb, work, &lda, ipiv, &iinfo);
/* Define user stream if current stream is NULL */
magma_queue_t stream[2];
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
magma_queue_create( &stream[0] );
if (orig_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else {
stream[1] = orig_stream;
}
for( j = 0; j < s; j++ ) {
// download j-th panel
cols = maxm - j*nb;
if (j > 0) {
magmablas_stranspose( nb, cols, dAT(j,j), ldda, dA, cols );
// make sure that gpu queue is empty
magma_device_sync();
magma_sgetmatrix_async( m-j*nb, nb, dA, cols, work, lda,
stream[0]);
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), ldda,
dAT(j-1,j+1), ldda );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), ldda,
dAT(j, j-1), ldda,
c_one, dAT(j, j+1), ldda );
// do the cpu part
rows = m - j*nb;
magma_queue_sync( stream[0] );
lapackf77_sgetrf( &rows, &nb, work, &lda, ipiv+j*nb, &iinfo);
}
if (*info == 0 && iinfo > 0)
*info = iinfo + j*nb;
// upload j-th panel
magma_ssetmatrix_async( m-j*nb, nb, work, lda, dA, cols,
stream[0]);
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_slaswp( n, dAT, ldda, j*nb + 1, j*nb + nb, ipiv, 1 );
magma_queue_sync( stream[0] );
magmablas_stranspose( cols, nb, dA, cols, dAT(j,j), ldda );
// do the small non-parallel computations (next panel update)
if (s > (j+1)) {
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), ldda,
dAT(j, j+1), ldda);
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), ldda,
dAT(j+1, j ), ldda,
c_one, dAT(j+1, j+1), ldda );
}
else {
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), ldda,
dAT(j, j+1), ldda);
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), ldda,
dAT(j+1, j ), ldda,
c_one, dAT(j+1, j+1), ldda );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_stranspose( nb0, rows, dAT(s,s), ldda, dA, cols );
magma_sgetmatrix( rows, nb0, dA, cols, work, lda );
// make sure that gpu queue is empty
magma_device_sync();
// do the cpu part
lapackf77_sgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo);
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_slaswp( n, dAT, ldda, s*nb + 1, s*nb + nb0, ipiv, 1 );
// upload j-th panel
magma_ssetmatrix( rows, nb0, work, lda, dA, cols );
magmablas_stranspose( rows, nb0, dA, cols, dAT(s,s), ldda );
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s,s), ldda,
dAT(s,s)+nb0, ldda);
}
// undo transpose
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_stranspose_inplace( ldda, dAT, ldda );
magma_sgetmatrix( m, n, da, ldda, A, lda );
}
else {
magmablas_stranspose( n, m, dAT, ldda, da, maxm );
magma_sgetmatrix( m, n, da, maxm, A, lda );
magma_free( dAT );
}
magma_free( dA );
magma_queue_destroy( stream[0] );
if (orig_stream == NULL) {
magma_queue_destroy( stream[1] );
}
magmablasSetKernelStream( orig_stream );
}
return *info;
} /* magma_sgetrf */
/**
Purpose
-------
SGETRF 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 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.
It uses 2 queues to overlap communication and computation.
Arguments
---------
@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 REAL 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_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrf(
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda,
magma_int_t *ipiv,
magma_int_t *info)
{
#ifdef HAVE_clBLAS
#define dA(i_, j_) dA, ((i_)*nb + (j_)*nb*ldda + dA_offset)
#define dAT(i_, j_) dAT, ((i_)*nb*lddat + (j_)*nb + dAT_offset)
#define dwork(i_) dwork, (i_)
#else
#define dA(i_, j_) ( dA + (i_)*nb + (j_)*nb*ldda)
#define dAT(i_, j_) ( dAT + (i_)*nb*lddat + (j_)*nb)
#define dwork(i_) (dwork + (i_))
#endif
// Constants
const float c_one = MAGMA_S_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
// Local variables
float *work;
magmaFloat_ptr dA, dAT, dwork;
magma_int_t iinfo, nb;
/* Check arguments */
*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;
/* Function Body */
nb = magma_get_sgetrf_nb( m, n );
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_sgetrf( &m, &n, A, &lda, ipiv, info );
}
else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, lddat, maxdim;
magma_int_t i, j, rows, cols, s = min(m, n)/nb;
maxm = magma_roundup( m, 32 );
maxn = magma_roundup( n, 32 );
maxdim = max( maxm, maxn );
lddat = maxn;
ldda = maxm;
/* set number of GPUs */
magma_int_t ngpu = magma_num_gpus();
if ( ngpu > 1 ) {
/* call multi-GPU non-GPU-resident interface */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
magma_queue_t queues[2] = { NULL, NULL };
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
/* check the memory requirement */
size_t mem_size = magma_queue_mem_size( queues[0] );
mem_size /= sizeof(float);
magma_int_t h = 1+(2+ngpu);
magma_int_t ngpu2 = ngpu;
magma_int_t NB = (magma_int_t)(0.8*mem_size/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) ) );
if ( ngpu > ceil((float)NB/nb) ) {
ngpu2 = (magma_int_t)ceil((float)NB/nb);
h = 1+(2+ngpu2);
NB = (magma_int_t)(0.8*mem_size/maxm - h*nb);
}
if ( ngpu2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
work = A;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
// dwork is nb*maxm for panel, and maxdim*maxdim for A
if (MAGMA_SUCCESS != magma_smalloc( &dwork, nb*maxm + maxdim*maxdim )) {
/* alloc failed so call non-GPU-resident version */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
dA = dwork + nb*maxm;
ldda = lddat = maxdim;
magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queues[0] );
dAT = dA;
magmablas_stranspose_inplace( maxdim, dAT(0,0), lddat, queues[0] );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
// dwork is nb*maxm for panel, and maxm*maxn for A
if (MAGMA_SUCCESS != magma_smalloc( &dwork, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
dA = dwork + nb*maxm;
magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queues[0] );
if (MAGMA_SUCCESS != magma_smalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dwork );
magma_sgetrf_m( ngpu, m, n, A, lda, ipiv, info );
return *info;
}
magmablas_stranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queues[0] );
}
lapackf77_sgetrf( &m, &nb, work, &lda, ipiv, &iinfo );
for( j = 0; j < s; j++ ) {
// get j-th panel from device
cols = maxm - j*nb;
if (j > 0) {
magmablas_stranspose( nb, cols, dAT(j,j), lddat, dwork(0), cols, queues[0] );
magma_queue_sync( queues[0] );
magma_sgetmatrix_async( m-j*nb, nb, dwork(0), cols, work, lda, queues[1] );
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat, queues[0] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat, queues[0] );
// do the cpu part
rows = m - j*nb;
magma_queue_sync( queues[1] );
lapackf77_sgetrf( &rows, &nb, work, &lda, ipiv+j*nb, &iinfo );
}
if (*info == 0 && iinfo > 0)
*info = iinfo + j*nb;
// put j-th panel onto device
magma_ssetmatrix_async( m-j*nb, nb, work, lda, dwork(0), cols, queues[1] );
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_slaswp( n, dAT(0,0), lddat, j*nb + 1, j*nb + nb, ipiv, 1, queues[0] );
magma_queue_sync( queues[1] );
magmablas_stranspose( cols, nb, dwork(0), cols, dAT(j,j), lddat, queues[0] );
// do the small non-parallel computations (next panel update)
if (s > (j+1)) {
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
else {
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
}
magma_int_t nb0 = min( m - s*nb, n - s*nb );
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_stranspose( nb0, rows, dAT(s,s), lddat, dwork(0), cols, queues[0] );
magma_sgetmatrix_async( rows, nb0, dwork(0), cols, work, lda, queues[0] );
magma_queue_sync( queues[0] );
// do the cpu part
lapackf77_sgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo );
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_slaswp( n, dAT(0,0), lddat, s*nb + 1, s*nb + nb0, ipiv, 1, queues[0] );
// put j-th panel onto device
magma_ssetmatrix_async( rows, nb0, work, lda, dwork(0), cols, queues[0] );
magmablas_stranspose( rows, nb0, dwork(0), cols, dAT(s,s), lddat, queues[0] );
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s, s), lddat,
dAT(s, s)+nb0, lddat, queues[0] );
}
// undo transpose
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_stranspose_inplace( maxdim, dAT(0,0), lddat, queues[0] );
magma_sgetmatrix( m, n, dAT(0,0), lddat, A, lda, queues[0] );
}
else {
magmablas_stranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queues[0] );
magma_sgetmatrix( m, n, dA(0,0), ldda, A, lda, queues[0] );
magma_free( dAT );
}
magma_free( dwork );
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
}
return *info;
} /* magma_sgetrf */
/**
Purpose
-------
ZGETRF_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 exceed 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]
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]
A COMPLEX_16 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_zgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_zgetrf_m(
magma_int_t ngpu,
magma_int_t m, magma_int_t n,
magmaDoubleComplex *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;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *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, ngpu0 = ngpu;
magma_int_t ii, jj, h, offset, ib, rows;
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;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/* initialize nb */
nb = magma_get_zgetrf_nb(m);
maxm = ((m + 31)/32)*32;
/* figure out NB */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(magmaDoubleComplex);
/* number of columns in the big panel */
h = 1+(2+ngpu0);
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 ( ngpu0 > ceil((double)NB/nb) ) {
ngpu = (int)ceil((double)NB/nb);
h = 1+(2+ngpu);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
} else {
ngpu = ngpu0;
}
if ( ngpu*NB >= n ) {
#ifdef CHECK_ZGETRF_OOC
printf( " * still fit in GPU memory.\n" );
#endif
NB = n;
} else {
#ifdef CHECK_ZGETRF_OOC
printf( " * don't fit in GPU memory.\n" );
#endif
NB = ngpu*NB;
NB = max( nb, (NB / nb) * nb); /* making sure it's devisable by nb (x64) */
}
#ifdef CHECK_ZGETRF_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_zgetrf(&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)/ngpu;
if ( NB%(nb*ngpu) != 0 )
n_local[0]++;
n_local[0] *= nb;
ldn_local = ((n_local[0]+31)/32)*32;
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
if (MAGMA_SUCCESS != magma_zmalloc( &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 ( ngpu0 > ceil((double)N/nb) ) {
ngpu = (int)ceil((double)N/nb);
} else {
ngpu = ngpu0;
}
for( d=0; d < ngpu; d++ ) {
n_local[d] = ((N/nb)/ngpu)*nb;
if (d < (N/nb)%ngpu)
n_local[d] += nb;
else if (d == (N/nb)%ngpu)
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_zsetmatrix_transpose_mgpu(ngpu, stream, A(0,I), lda,
dAT, ldn_local, dA, maxm, M, N, nb);
for( d=0; d < ngpu; 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 < ngpu; d++ ) {
magma_setdevice(d);
nbi = min( nb, NBk );
magma_zsetmatrix_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_ztranspose( M-offset, nbi, dA[d], maxm-offset, dPT(d,0,0), nb );
}
/* applying the pivot from the previous big-panel */
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magmablas_zlaswp_q( ldn_local, dAT(d,0,0), ldn_local, offset+1, offset+NBk, ipiv, 1, stream[d][1] );
}
/* == 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 < ngpu; 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_zsetmatrix_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_ztranspose( 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_ztrsm( 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_zgemm( 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 < ngpu; 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_zgetrf2_mgpu(ngpu, 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_zgetmatrix_transpose_mgpu(ngpu, stream, dAT, ldn_local, A(0,I), lda, dA, maxm, M, N, nb);
for( d=0; d < ngpu; 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_ZGETRF( 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 < ngpu0; 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] );
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
}
if ( *info >= 0 )
magma_zgetrf_piv(m, n, NB, A, lda, ipiv, info);
return *info;
} /* magma_zgetrf_m */
/**
Purpose
-------
CGETRF 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 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.
If the current stream is NULL, this version replaces it with user defined
stream to overlap computation with communication.
Arguments
---------
@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 COMPLEX 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_cgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_cgetrf(magma_int_t m, magma_int_t n, magmaFloatComplex *A, magma_int_t lda,
magma_int_t *ipiv, magma_int_t *info)
{
#define dAT(i,j) (dAT + (i)*nb*ldda + (j)*nb)
magmaFloatComplex *dAT, *dA, *da, *work;
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t iinfo, nb;
*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;
nb = magma_get_cgetrf_nb(m);
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_cgetrf(&m, &n, A, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, maxdim;
magma_int_t i, rows, cols, s = min(m, n)/nb;
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
maxdim = max(maxm, maxn);
/* set number of GPUs */
magma_int_t num_gpus = magma_num_gpus();
if ( num_gpus > 1 ) {
/* call multi-GPU non-GPU-resident interface */
magma_cgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
return *info;
}
/* explicitly checking the memory requirement */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(magmaFloatComplex);
int h = 1+(2+num_gpus), num_gpus2 = num_gpus;
int 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) ) );
if ( num_gpus > ceil((float)NB/nb) ) {
num_gpus2 = (int)ceil((float)NB/nb);
h = 1+(2+num_gpus2);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
}
if ( num_gpus2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
magma_cgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
return *info;
}
ldda = maxn;
work = A;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
if (MAGMA_SUCCESS != magma_cmalloc( &dA, nb*maxm + maxdim*maxdim )) {
/* alloc failed so call non-GPU-resident version */
magma_cgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
ldda = maxdim;
magma_csetmatrix( m, n, A, lda, da, ldda );
dAT = da;
magmablas_ctranspose_inplace( ldda, dAT, ldda );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
if (MAGMA_SUCCESS != magma_cmalloc( &dA, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
magma_cgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
magma_csetmatrix( m, n, A, lda, da, maxm );
if (MAGMA_SUCCESS != magma_cmalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dA );
magma_cgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
return *info;
}
magmablas_ctranspose( m, n, da, maxm, dAT, ldda );
}
lapackf77_cgetrf( &m, &nb, work, &lda, ipiv, &iinfo);
/* Define user stream if current stream is NULL */
cudaStream_t stream[2], current_stream;
magmablasGetKernelStream(¤t_stream);
magma_queue_create( &stream[0] );
if (current_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else
stream[1] = current_stream;
for( i = 0; i < s; i++ ) {
// download i-th panel
cols = maxm - i*nb;
if (i > 0) {
// download i-th panel
magmablas_ctranspose( nb, cols, dAT(i,i), ldda, dA, cols );
// make sure that gpu queue is empty
magma_device_sync();
magma_cgetmatrix_async( m-i*nb, nb, dA, cols, work, lda,
stream[0]);
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (i+1)*nb, nb,
c_one, dAT(i-1,i-1), ldda,
dAT(i-1,i+1), ldda );
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-i*nb, nb,
c_neg_one, dAT(i-1,i+1), ldda,
dAT(i, i-1), ldda,
c_one, dAT(i, i+1), ldda );
// do the cpu part
rows = m - i*nb;
magma_queue_sync( stream[0] );
lapackf77_cgetrf( &rows, &nb, work, &lda, ipiv+i*nb, &iinfo);
}
if (*info == 0 && iinfo > 0)
*info = iinfo + i*nb;
// upload i-th panel
magma_csetmatrix_async( m-i*nb, nb, work, lda, dA, cols,
stream[0]);
magmablas_cpermute_long2( ldda, dAT, ldda, ipiv, nb, i*nb );
magma_queue_sync( stream[0] );
magmablas_ctranspose( cols, nb, dA, cols, dAT(i,i), ldda );
// do the small non-parallel computations
if (s > (i+1)) {
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(i, i ), ldda,
dAT(i, i+1), ldda);
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), ldda,
dAT(i+1, i ), ldda,
c_one, dAT(i+1, i+1), ldda );
}
else {
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(i, i ), ldda,
dAT(i, i+1), ldda);
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), ldda,
dAT(i+1, i ), ldda,
c_one, dAT(i+1, i+1), ldda );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_ctranspose( nb0, rows, dAT(s,s), ldda, dA, cols );
magma_cgetmatrix( rows, nb0, dA, cols, work, lda );
// make sure that gpu queue is empty
magma_device_sync();
// do the cpu part
lapackf77_cgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo);
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
magmablas_cpermute_long2( ldda, dAT, ldda, ipiv, nb0, s*nb );
magma_csetmatrix( rows, nb0, work, lda, dA, cols );
magmablas_ctranspose( rows, nb0, dA, cols, dAT(s,s), ldda );
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s, s), ldda,
dAT(s, s)+nb0, ldda);
}
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_ctranspose_inplace( ldda, dAT, ldda );
magma_cgetmatrix( m, n, da, ldda, A, lda );
} else {
magmablas_ctranspose( n, m, dAT, ldda, da, maxm );
magma_cgetmatrix( m, n, da, maxm, A, lda );
magma_free( dAT );
}
magma_free( dA );
magma_queue_destroy( stream[0] );
if (current_stream == NULL) {
magma_queue_destroy( stream[1] );
magmablasSetKernelStream(NULL);
}
}
return *info;
} /* magma_cgetrf */
extern "C" magma_int_t
magma_cgetrf_gpu(magma_int_t m, magma_int_t n,
magmaFloatComplex *dA, magma_int_t ldda,
magma_int_t *ipiv, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
CGETRF 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.
If the current stream is NULL, this version replaces it with user defined
stream to overlap computation with communication.
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) COMPLEX 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 dAT(i,j) (dAT + (i)*nb*lddat + (j)*nb)
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, maxn, mindim;
magma_int_t i, rows, cols, s, lddat, lddwork;
magmaFloatComplex *dAT, *dAP, *work;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < 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;
/* Function Body */
mindim = min(m, n);
nb = magma_get_cgetrf_nb(m);
s = mindim / nb;
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
magma_cmalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_cgetmatrix( m, n, dA, ldda, work, m );
lapackf77_cgetrf(&m, &n, work, &m, ipiv, info);
magma_csetmatrix( m, n, work, m, dA, ldda );
magma_free_cpu(work);
}
else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
lddat = maxn;
lddwork = maxm;
dAT = dA;
if (MAGMA_SUCCESS != magma_cmalloc( &dAP, nb*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if ( m == n ) {
lddat = ldda;
magmablas_ctranspose_inplace( m, dAT, ldda );
}
else {
if (MAGMA_SUCCESS != magma_cmalloc( &dAT, maxm*maxn )) {
magma_free( dAP );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_ctranspose2( dAT, lddat, dA, ldda, m, n );
}
if (MAGMA_SUCCESS != magma_cmalloc_pinned( &work, maxm*nb )) {
magma_free( dAP );
if ( ! (m == n))
magma_free( dAT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
/* Define user stream if current stream is NULL */
cudaStream_t stream[2], current_stream;
magmablasGetKernelStream(¤t_stream);
magma_queue_create( &stream[0] );
if (current_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else
stream[1] = current_stream;
for( i=0; i<s; i++ )
{
// download i-th panel
cols = maxm - i*nb;
//magmablas_ctranspose( dAP, cols, dAT(i,i), lddat, nb, cols );
magmablas_ctranspose2( dAP, cols, dAT(i,i), lddat, nb, m-i*nb );
// make sure that that the transpose has completed
magma_queue_sync( stream[1] );
magma_cgetmatrix_async( m-i*nb, nb, dAP, cols, work, lddwork,
stream[0]);
if ( i>0 ){
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (i+1)*nb, nb,
c_one, dAT(i-1,i-1), lddat,
dAT(i-1,i+1), lddat );
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-i*nb, nb,
c_neg_one, dAT(i-1,i+1), lddat,
dAT(i, i-1), lddat,
c_one, dAT(i, i+1), lddat );
}
// do the cpu part
rows = m - i*nb;
magma_queue_sync( stream[0] );
lapackf77_cgetrf( &rows, &nb, work, &lddwork, ipiv+i*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + i*nb;
// upload i-th panel
magma_csetmatrix_async( m-i*nb, nb, work, lddwork, dAP, maxm,
stream[0]);
magmablas_cpermute_long2( n, dAT, lddat, ipiv, nb, i*nb );
magma_queue_sync( stream[0] );
//magmablas_ctranspose(dAT(i,i), lddat, dAP, maxm, cols, nb);
magmablas_ctranspose2(dAT(i,i), lddat, dAP, maxm, m-i*nb, nb);
// do the small non-parallel computations (next panel update)
if ( s > (i+1) ) {
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(i, i ), lddat,
dAT(i, i+1), lddat);
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), lddat,
dAT(i+1, i ), lddat,
c_one, dAT(i+1, i+1), lddat );
}
else {
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(i, i ), lddat,
dAT(i, i+1), lddat);
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), lddat,
dAT(i+1, i ), lddat,
c_one, dAT(i+1, i+1), lddat );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_ctranspose2( dAP, maxm, dAT(s,s), lddat, nb0, rows);
magma_cgetmatrix( rows, nb0, dAP, maxm, work, lddwork );
// do the cpu part
lapackf77_cgetrf( &rows, &nb0, work, &lddwork, ipiv+s*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + s*nb;
magmablas_cpermute_long2( n, dAT, lddat, ipiv, nb0, s*nb );
// upload i-th panel
magma_csetmatrix( rows, nb0, work, lddwork, dAP, maxm );
magmablas_ctranspose2( dAT(s,s), lddat, dAP, maxm, rows, nb0);
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s,s), lddat,
dAT(s,s)+nb0, lddat);
if ( m == n ) {
magmablas_ctranspose_inplace( m, dAT, lddat );
}
else {
magmablas_ctranspose2( dA, ldda, dAT, lddat, n, m );
magma_free( dAT );
}
magma_free( dAP );
magma_free_pinned( work );
magma_queue_destroy( stream[0] );
if (current_stream == NULL) {
magma_queue_destroy( stream[1] );
magmablasSetKernelStream(NULL);
}
}
return *info;
} /* End of MAGMA_CGETRF_GPU */
/**
Purpose
-------
ZGETRF 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.
If the current stream is NULL, this version replaces it with a new
stream to overlap computation with communication.
Arguments
---------
@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]
dA COMPLEX_16 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.
@param[in]
ldda INTEGER
The leading dimension of the array A. 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_zgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_zgetrf_gpu(
magma_int_t m, magma_int_t n,
magmaDoubleComplex_ptr dA, magma_int_t ldda, magma_int_t *ipiv,
magma_int_t *info)
{
#define dAT(i_, j_) (dAT + (i_)*nb*lddat + (j_)*nb)
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, maxn, mindim;
magma_int_t i, j, rows, cols, s, lddat, ldwork;
magmaDoubleComplex *dAT, *dAP, *work;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < 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;
/* Function Body */
mindim = min(m, n);
nb = magma_get_zgetrf_nb(m);
s = mindim / nb;
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
magma_zmalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_zgetmatrix( m, n, dA, ldda, work, m );
lapackf77_zgetrf(&m, &n, work, &m, ipiv, info);
magma_zsetmatrix( m, n, work, m, dA, ldda );
magma_free_cpu(work);
}
else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
if (MAGMA_SUCCESS != magma_zmalloc( &dAP, nb*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
// square matrices can be done in place;
// rectangular requires copy to transpose
if ( m == n ) {
dAT = dA;
lddat = ldda;
magmablas_ztranspose_inplace( m, dAT, ldda );
}
else {
lddat = maxn; // N-by-M
if (MAGMA_SUCCESS != magma_zmalloc( &dAT, lddat*maxm )) {
magma_free( dAP );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_ztranspose( m, n, dA, ldda, dAT, lddat );
}
ldwork = maxm;
if (MAGMA_SUCCESS != magma_zmalloc_pinned( &work, ldwork*nb )) {
magma_free( dAP );
if ( ! (m == n))
magma_free( dAT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
/* Define user stream if current stream is NULL */
magma_queue_t stream[2];
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
magma_queue_create( &stream[0] );
if (orig_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else {
stream[1] = orig_stream;
}
for( j=0; j < s; j++ ) {
// download j-th panel
cols = maxm - j*nb;
magmablas_ztranspose( nb, m-j*nb, dAT(j,j), lddat, dAP, cols );
// make sure that the transpose has completed
magma_queue_sync( stream[1] );
magma_zgetmatrix_async( m-j*nb, nb, dAP, cols, work, ldwork,
stream[0]);
if ( j > 0 ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat );
}
// do the cpu part
rows = m - j*nb;
magma_queue_sync( stream[0] );
lapackf77_zgetrf( &rows, &nb, work, &ldwork, ipiv+j*nb, &iinfo);
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j*nb;
// upload j-th panel
magma_zsetmatrix_async( m-j*nb, nb, work, ldwork, dAP, maxm,
stream[0]);
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_zlaswp( n, dAT, lddat, j*nb + 1, j*nb + nb, ipiv, 1 );
magma_queue_sync( stream[0] );
magmablas_ztranspose( m-j*nb, nb, dAP, maxm, dAT(j,j), lddat );
// do the small non-parallel computations (next panel update)
if ( s > (j+1) ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat );
}
else {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_ztranspose( nb0, rows, dAT(s,s), lddat, dAP, maxm );
magma_zgetmatrix( rows, nb0, dAP, maxm, work, ldwork );
// do the cpu part
lapackf77_zgetrf( &rows, &nb0, work, &ldwork, ipiv+s*nb, &iinfo);
if ( *info == 0 && iinfo > 0 )
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_zlaswp( n, dAT, lddat, s*nb + 1, s*nb + nb0, ipiv, 1 );
// upload j-th panel
magma_zsetmatrix( rows, nb0, work, ldwork, dAP, maxm );
magmablas_ztranspose( rows, nb0, dAP, maxm, dAT(s,s), lddat );
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s,s), lddat,
dAT(s,s)+nb0, lddat);
}
// undo transpose
if ( m == n ) {
magmablas_ztranspose_inplace( m, dAT, lddat );
}
else {
magmablas_ztranspose( n, m, dAT, lddat, dA, ldda );
magma_free( dAT );
}
magma_free( dAP );
magma_free_pinned( work );
magma_queue_destroy( stream[0] );
if (orig_stream == NULL) {
magma_queue_destroy( stream[1] );
}
magmablasSetKernelStream( orig_stream );
}
return *info;
} /* magma_zgetrf_gpu */
/**
Purpose
-------
SGETRF 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.
Use two buffer to send panels.
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 REAL 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.
@param[in]
ldda INTEGER
The leading dimension of the array A. 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_sgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_sgetrf2_mgpu(magma_int_t num_gpus,
magma_int_t m, magma_int_t n, magma_int_t nb, magma_int_t offset,
float *d_lAT[], magma_int_t lddat, magma_int_t *ipiv,
float *d_lAP[], float *w, magma_int_t ldw,
magma_queue_t streaml[][2], magma_int_t *info)
{
#define dAT(id,i,j) (d_lAT[(id)] + ((offset)+(i)*nb)*lddat + (j)*nb)
#define W(j) (w+((j)%num_gpus)*nb*ldw)
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t block_size = 32;
magma_int_t iinfo, n_local[MagmaMaxGPUs];
magma_int_t maxm, mindim;
magma_int_t i, d, dd, rows, cols, s, ldpan[MagmaMaxGPUs];
magma_int_t id, i_local, i_local2, nb0, nb1, h = 2+num_gpus;
float *d_panel[MagmaMaxGPUs], *panel_local[MagmaMaxGPUs];
/* Check arguments */
*info = 0;
if (m < 0)
*info = -2;
else if (n < 0)
*info = -3;
else if (num_gpus*lddat < max(1,n))
*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);
if ( num_gpus > ceil((float)n/nb) ) {
*info = -1;
return *info;
}
/* Use hybrid blocked code. */
maxm = ((m + block_size-1)/block_size)*block_size;
/* some initializations */
for (i=0; i < num_gpus; i++) {
magma_setdevice(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;
/* workspaces */
d_panel[i] = &(d_lAP[i][h*nb*maxm]); /* temporary panel storage */
}
trace_init( 1, num_gpus, 2, (CUstream_st**)streaml );
/* start sending the panel to cpu */
nb0 = min(mindim, nb);
magma_setdevice(0);
magmablasSetKernelStream(streaml[0][1]);
trace_gpu_start( 0, 1, "comm", "get" );
magmablas_stranspose( nb0, m, dAT(0,0,0), lddat, d_lAP[0], maxm );
magma_sgetmatrix_async( m, nb0,
d_lAP[0], maxm,
W(0), ldw, streaml[0][1] );
trace_gpu_end( 0, 1 );
/* ------------------------------------------------------------------------------------- */
magma_timer_t time=0;
timer_start( time );
s = mindim / nb;
for( i=0; i < s; i++ ) {
/* Set the GPU number that holds the current panel */
id = i%num_gpus;
magma_setdevice(id);
/* Set the local index where the current panel is */
i_local = i/num_gpus;
cols = maxm - i*nb;
rows = m - i*nb;
/* synchrnoize i-th panel from id-th gpu into work */
magma_queue_sync( streaml[id][1] );
/* i-th panel factorization */
trace_cpu_start( 0, "getrf", "getrf" );
lapackf77_sgetrf( &rows, &nb, W(i), &ldw, ipiv+i*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) ) {
*info = iinfo + i*nb;
}
trace_cpu_end( 0 );
/* start sending the panel to all the gpus */
d = (i+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
magma_setdevice(d);
trace_gpu_start( 0, 1, "comm", "set" );
magma_ssetmatrix_async( rows, nb,
W(i), ldw,
&d_lAP[d][(i%h)*nb*maxm], cols,
streaml[d][1] );
trace_gpu_end( 0, 1 );
d = (d+1)%num_gpus;
}
/* apply the pivoting */
d = (i+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
magma_setdevice(d);
magmablasSetKernelStream(streaml[d][0]);
trace_gpu_start( d, 1, "pivot", "pivot" );
if ( dd == 0 )
magmablas_spermute_long2( lddat, dAT(d,0,0), lddat, ipiv, nb, i*nb );
else
magmablas_spermute_long3( dAT(d,0,0), lddat, ipiv, nb, i*nb );
trace_gpu_end( d, 1 );
d = (d+1)%num_gpus;
}
/* update the trailing-matrix/look-ahead */
d = (i+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
magma_setdevice(d);
/* storage for panel */
if ( d == id ) {
/* the panel belond to this gpu */
panel_local[d] = dAT(d,i,i_local);
ldpan[d] = lddat;
/* next column */
i_local2 = i_local+1;
} else {
/* the panel belong to another gpu */
panel_local[d] = d_panel[d];
ldpan[d] = nb;
/* next column */
i_local2 = i_local;
if ( d < id ) i_local2 ++;
}
/* the size of the next column */
if ( s > (i+1) ) {
nb0 = nb;
} else {
nb0 = n_local[d]-nb*(s/num_gpus);
if ( d < s%num_gpus ) nb0 -= nb;
}
if ( d == (i+1)%num_gpus) {
/* owns the next column, look-ahead the column */
nb1 = nb0;
magmablasSetKernelStream(streaml[d][1]);
/* make sure all the pivoting has been applied */
magma_queue_sync(streaml[d][0]);
trace_gpu_start( d, 1, "gemm", "gemm" );
/* transpose panel on GPU */
magmablas_stranspose( rows, nb, &d_lAP[d][(i%h)*nb*maxm], cols, panel_local[d], ldpan[d] );
/* synch for remaining update */
magma_queue_sync(streaml[d][1]);
} else {
/* update the entire trailing matrix */
nb1 = n_local[d] - i_local2*nb;
magmablasSetKernelStream(streaml[d][0]);
/* synchronization to make sure panel arrived on gpu */
magma_queue_sync(streaml[d][1]);
trace_gpu_start( d, 0, "gemm", "gemm" );
/* transpose panel on GPU */
magmablas_stranspose( rows, nb, &d_lAP[d][(i%h)*nb*maxm], cols, panel_local[d], ldpan[d] );
}
/* gpu updating the trailing matrix */
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb1, nb, c_one,
panel_local[d], ldpan[d],
dAT(d, i, i_local2), lddat);
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
nb1, m-(i+1)*nb, nb,
c_neg_one, dAT(d, i, i_local2), lddat,
&(panel_local[d][nb*ldpan[d]]), ldpan[d],
c_one, dAT(d, i+1, i_local2), lddat );
if ( d == (i+1)%num_gpus ) {
/* Set the local index where the current panel is */
int loff = i+1;
int i_local = (i+1)/num_gpus;
int ldda = maxm - (i+1)*nb;
int cols = m - (i+1)*nb;
nb0 = min(nb, mindim - (i+1)*nb); /* size of the diagonal block */
trace_gpu_end( d, 1 );
if ( nb0 > 0 ) {
/* transpose the panel for sending it to cpu */
trace_gpu_start( d, 1, "comm", "get" );
magmablas_stranspose( nb0, m-(i+1)*nb, dAT(d,loff,i_local), lddat, &d_lAP[d][((i+1)%h)*nb*maxm], ldda );
/* send the panel to cpu */
magma_sgetmatrix_async( cols, nb0,
&d_lAP[d][((i+1)%h)*nb*maxm], ldda,
W(i+1), ldw, streaml[d][1] );
trace_gpu_end( d, 1 );
}
} else {
trace_gpu_end( d, 0 );
}
d = (d+1)%num_gpus;
}
/* update the remaining matrix by gpu owning the next panel */
if ( (i+1) < s ) {
int i_local = (i+1)/num_gpus;
int rows = m - (i+1)*nb;
d = (i+1)%num_gpus;
magma_setdevice(d);
magmablasSetKernelStream(streaml[d][0]);
trace_gpu_start( d, 0, "gemm", "gemm" );
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n_local[d] - (i_local+1)*nb, nb,
c_one, panel_local[d], ldpan[d],
dAT(d,i,i_local+1), lddat );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
n_local[d]-(i_local+1)*nb, rows, nb,
c_neg_one, dAT(d,i,i_local+1), lddat,
&(panel_local[d][nb*ldpan[d]]), ldpan[d],
c_one, dAT(d,i+1, i_local+1), lddat );
trace_gpu_end( d, 0 );
}
} /* end of for i=1..s */
/* ------------------------------------------------------------------------------ */
/* Set the GPU number that holds the last panel */
id = s%num_gpus;
/* Set the local index where the last panel is */
i_local = s/num_gpus;
/* size of the last diagonal-block */
nb0 = min(m - s*nb, n - s*nb);
rows = m - s*nb;
cols = maxm - s*nb;
if ( nb0 > 0 ) {
magma_setdevice(id);
/* wait for the last panel on cpu */
magma_queue_sync( streaml[id][1] );
/* factor on cpu */
lapackf77_sgetrf( &rows, &nb0, W(s), &ldw, ipiv+s*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + s*nb;
/* send the factor to gpus */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
i_local2 = i_local;
if ( d < id ) i_local2 ++;
if ( d == id || n_local[d] > i_local2*nb ) {
magma_ssetmatrix_async( rows, nb0,
W(s), ldw,
&d_lAP[d][(s%h)*nb*maxm],
cols, streaml[d][1] );
}
}
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(streaml[d][0]);
if ( d == 0 )
magmablas_spermute_long2( lddat, dAT(d,0,0), lddat, ipiv, nb0, s*nb );
else
magmablas_spermute_long3( dAT(d,0,0), lddat, ipiv, nb0, s*nb );
}
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(streaml[d][1]);
/* wait for the pivoting to be done */
magma_queue_sync( streaml[d][0] );
i_local2 = i_local;
if ( d < id ) i_local2++;
if ( d == id ) {
/* the panel belond to this gpu */
panel_local[d] = dAT(d,s,i_local);
/* next column */
nb1 = n_local[d] - i_local*nb-nb0;
magmablas_stranspose( rows, nb0, &d_lAP[d][(s%h)*nb*maxm], cols, panel_local[d], lddat );
if ( nb1 > 0 ) {
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb1, nb0, c_one,
panel_local[d], lddat,
dAT(d,s,i_local)+nb0, lddat);
}
} else if ( n_local[d] > i_local2*nb ) {
/* the panel belong to another gpu */
panel_local[d] = d_panel[d];
/* next column */
nb1 = n_local[d] - i_local2*nb;
magmablas_stranspose( rows, nb0, &d_lAP[d][(s%h)*nb*maxm], cols, panel_local[d], nb );
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb1, nb0, c_one,
panel_local[d], nb,
dAT(d,s,i_local2), lddat);
}
}
} /* if ( nb0 > 0 ) */
/* clean up */
trace_finalize( "sgetrf_mgpu.svg","trace.css" );
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( streaml[d][0] );
magma_queue_sync( streaml[d][1] );
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
timer_start( time );
timer_printf("\n Performance %f GFlop/s\n", FLOPS_SGETRF(m,n) / 1e9 / time );
return *info;
} /* magma_sgetrf2_mgpu */
extern "C" magma_int_t
magma_zgetrf2_gpu(
magma_int_t m, magma_int_t n,
magmaDoubleComplex_ptr dA, size_t dA_offset, magma_int_t ldda,
magma_int_t *ipiv,
magma_queue_t queues[2],
magma_int_t *info )
{
/* -- clMAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date November 2014
Purpose
=======
ZGETRF 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
=========
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) COMPLEX_16 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 dA(i_, j_) dA, dA_offset + (i_)*nb + (j_)*nb*ldda
#define dAT(i_, j_) dAT, dAT_offset + (i_)*nb*lddat + (j_)*nb
#define dAP(i_, j_) dAP, (i_) + (j_)*maxm
#define work(i_) (work + (i_))
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, maxn, mindim;
magma_int_t i, j, rows, s, lddat, ldwork;
magmaDoubleComplex_ptr dAT, dAP;
magmaDoubleComplex *work;
size_t dAT_offset;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < 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;
/* Function Body */
mindim = min(m, n);
nb = magma_get_zgetrf_nb(m);
s = mindim / nb;
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
if ( MAGMA_SUCCESS != magma_zmalloc_cpu( &work, m*n )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_zgetmatrix( m, n, dA(0,0), ldda, work(0), m, queues[0] );
lapackf77_zgetrf( &m, &n, work, &m, ipiv, info );
magma_zsetmatrix( m, n, work(0), m, dA(0,0), ldda, queues[0] );
magma_free_cpu( work );
}
else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
if ( MAGMA_SUCCESS != magma_zmalloc( &dAP, nb*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
// square matrices can be done in place;
// rectangular requires copy to transpose
if ( m == n ) {
dAT = dA;
dAT_offset = dA_offset;
lddat = ldda;
magmablas_ztranspose_inplace( m, dAT(0,0), lddat, queues[0] );
}
else {
lddat = maxn; // N-by-M
dAT_offset = 0;
if ( MAGMA_SUCCESS != magma_zmalloc( &dAT, lddat*maxm )) {
magma_free( dAP );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_ztranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queues[0] );
}
ldwork = maxm;
/*
if ( MAGMA_SUCCESS != magma_zmalloc_cpu( &work, ldwork*nb ) ) {
magma_free( dAP );
if ( dA != dAT )
magma_free( dAT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
*/
cl_mem work_mapped = clCreateBuffer( gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, ldwork*nb * sizeof(magmaDoubleComplex), NULL, NULL );
work = (magmaDoubleComplex*) clEnqueueMapBuffer( queues[0], work_mapped, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, ldwork*nb * sizeof(magmaDoubleComplex), 0, NULL, NULL, NULL );
for( j=0; j < s; j++ ) {
// download j-th panel
magmablas_ztranspose( nb, m-j*nb, dAT(j,j), lddat, dAP(0,0), maxm, queues[0] );
clFlush( queues[0] );
magma_queue_sync( queues[0] );
magma_zgetmatrix_async( m-j*nb, nb, dAP(0,0), maxm, work(0), ldwork, queues[1], NULL );
clFlush( queues[1] );
if ( j > 0 ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat, queues[0] );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat, queues[0] );
}
magma_queue_sync( queues[1] );
// do the cpu part
rows = m - j*nb;
lapackf77_zgetrf( &rows, &nb, work, &ldwork, ipiv+j*nb, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j*nb;
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_zlaswp( n, dAT(0,0), lddat, j*nb + 1, j*nb + nb, ipiv, 1, queues[0] );
clFlush( queues[0] );
// upload j-th panel
magma_zsetmatrix_async( m-j*nb, nb, work(0), ldwork, dAP(0,0), maxm, queues[1], NULL );
magma_queue_sync( queues[1] );
magmablas_ztranspose( m-j*nb, nb, dAP(0,0), maxm, dAT(j,j), lddat, queues[0] );
clFlush( queues[0] );
// do the small non-parallel computations (next panel update)
if ( s > (j+1) ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
else {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queues[0] );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queues[0] );
}
}
magma_int_t nb0 = min( m - s*nb, n - s*nb );
if ( nb0 > 0 ) {
rows = m - s*nb;
magmablas_ztranspose( nb0, rows, dAT(s,s), lddat, dAP(0,0), maxm, queues[0] );
clFlush( queues[0] );
magma_queue_sync( queues[0] );
magma_zgetmatrix_async( rows, nb0, dAP(0,0), maxm, work(0), ldwork, queues[1], NULL );
magma_queue_sync( queues[1] );
// do the cpu part
lapackf77_zgetrf( &rows, &nb0, work, &ldwork, ipiv+s*nb, &iinfo );
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_zlaswp( n, dAT(0,0), lddat, s*nb + 1, s*nb + nb0, ipiv, 1, queues[0] );
clFlush( queues[0] );
// upload j-th panel
magma_zsetmatrix_async( rows, nb0, work(0), ldwork, dAP(0,0), maxm, queues[1], NULL );
magma_queue_sync( queues[1] );
magmablas_ztranspose( rows, nb0, dAP(0,0), maxm, dAT(s,s), lddat, queues[0] );
clFlush( queues[0] );
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s,s), lddat,
dAT(s,s)+nb0, lddat, queues[0] );
}
// undo transpose
if ( dA == dAT ) {
magmablas_ztranspose_inplace( m, dAT(0,0), lddat, queues[0] );
}
else {
magmablas_ztranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queues[0] );
magma_free( dAT );
}
magma_queue_sync( queues[0] );
magma_queue_sync( queues[1] );
magma_free( dAP );
// magma_free_cpu( work );
clEnqueueUnmapMemObject( queues[0], work_mapped, work, 0, NULL, NULL );
clReleaseMemObject( work_mapped );
}
return *info;
} /* magma_zgetrf_gpu */
extern "C" magma_int_t
magma_zgetrf(
magma_int_t m, magma_int_t n,
magmaDoubleComplex *A, magma_int_t lda, magma_int_t *ipiv,
magma_queue_t queue[2],
magma_int_t *info)
{
/* -- clMAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date November 2014
Purpose
=======
ZGETRF 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 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.
If the current stream is NULL, this version replaces it with user defined
stream to overlap computation with communication.
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) COMPLEX_16 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 dAT(i,j) dAT, dAT_offset + ((i)*nb*lddat + (j)*nb)
magmaDoubleComplex *work;
magmaDoubleComplex_ptr dAT, dA, dwork, dAP;
size_t dA_offset, dAT_offset;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t iinfo, nb;
*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;
nb = magma_get_zgetrf_nb(m);
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_zgetrf(&m, &n, A, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, maxdim, lddat;
magma_int_t i, j, rows, cols, s = min(m, n)/nb;
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
lddat = maxn;
ldda = maxm;
maxdim = max(maxm, maxn);
/* set number of GPUs */
magma_int_t num_gpus = magma_num_gpus();
if ( num_gpus > 1 ) {
/* call multi-GPU non-GPU-resident interface */
printf("multiple-GPU verison not implemented\n");
return MAGMA_ERR_NOT_IMPLEMENTED;
// magma_zgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
// return *info;
}
/* explicitly checking the memory requirement */
magma_int_t totalMem = magma_queue_meminfo( queue[0] );
totalMem /= sizeof(magmaDoubleComplex);
int h = 1+(2+num_gpus), num_gpus2 = num_gpus;
int NB = (magma_int_t)(0.8*totalMem/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) ) );
if( num_gpus > ceil((double)NB/nb) ) {
num_gpus2 = (int)ceil((double)NB/nb);
h = 1+(2+num_gpus2);
NB = (magma_int_t)(0.8*totalMem/maxm-h*nb);
}
if( num_gpus2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
printf("non-GPU-resident version not implemented\n");
return MAGMA_ERR_NOT_IMPLEMENTED;
//magma_zgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
//return *info;
}
work = A;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
if (MAGMA_SUCCESS !=
magma_zmalloc( &dwork, (nb*maxm + maxdim*maxdim) ) ) {
/* alloc failed so call non-GPU-resident version */
printf("non-GPU-resident version not implemented\n");
return MAGMA_ERR_NOT_IMPLEMENTED;
//magma_zgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
//return *info;
}
dAP = dwork;
dA = dwork;
dA_offset = nb*maxm;
ldda = lddat = maxdim;
magma_zsetmatrix( m, n, A, lda, dA, dA_offset, ldda, queue[0] );
dAT = dA;
dAT_offset = dA_offset;
magmablas_ztranspose_inplace( m, dAT, dAT_offset, ldda, queue[0] );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
if (MAGMA_SUCCESS !=
magma_zmalloc( &dwork, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
printf("non-GPU-resident version not implemented\n");
return MAGMA_ERR_NOT_IMPLEMENTED;
//magma_zgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
//return *info;
}
dAP = dwork;
dA = dwork;
dA_offset = nb*maxm;
magma_zsetmatrix( m, n, A, lda, dA, dA_offset, ldda, queue[0] );
if (MAGMA_SUCCESS != magma_zmalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dwork );
printf("non-GPU-resident version not implemented\n");
return MAGMA_ERR_NOT_IMPLEMENTED;
//magma_zgetrf_m(num_gpus, m, n, A, lda, ipiv, info);
//return *info;
}
dAT_offset = 0;
magmablas_ztranspose( m, n, dA, dA_offset, ldda, dAT, dAT_offset, lddat, queue[0] );
}
lapackf77_zgetrf( &m, &nb, work, &lda, ipiv, &iinfo);
for( j = 0; j < s; j++ )
{
// download j-th panel
cols = maxm - j*nb;
if (j>0){
// download j-th panel
magmablas_ztranspose( nb, cols, dAT(j,j), lddat, dAP, 0, cols, queue[0] );
magma_queue_sync(queue[0]);
magma_zgetmatrix_async( m-j*nb, nb, dAP, 0, cols, work, lda,
queue[1], NULL);
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat, queue[0] );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat, queue[0] );
// do the cpu part
rows = m - j*nb;
magma_queue_sync( queue[1] );
lapackf77_zgetrf( &rows, &nb, work, &lda, ipiv+j*nb, &iinfo);
}
if (*info == 0 && iinfo > 0)
*info = iinfo + j*nb;
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_zlaswp( n, dAT, dAT_offset, lddat, j*nb + 1, j*nb + nb, ipiv, 1, queue[0] );
// upload j-th panel
magma_zsetmatrix_async( m-j*nb, nb, work, lda, dAP, 0, maxm,
queue[1], NULL);
magma_queue_sync( queue[1] );
magmablas_ztranspose( cols, nb, dAP, 0, maxm, dAT(j,j), lddat, queue[0] );
// do the small non-parallel computations
if (s > (j+1)){
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queue[0]);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queue[0] );
}
else{
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat, queue[0] );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat, queue[0] );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_ztranspose( nb0, rows, dAT(s,s), lddat, dAP, 0, maxm, queue[0]);
magma_queue_sync(queue[0]);
magma_zgetmatrix_async( rows, nb0, dAP, 0, maxm, work, lda, queue[1], NULL );
magma_queue_sync(queue[1]);
// do the cpu part
lapackf77_zgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo);
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_zlaswp( n, dAT, dAT_offset, lddat, s*nb + 1, s*nb + nb0, ipiv, 1, queue[0] );
magma_zsetmatrix_async( rows, nb0, work, lda, dAP, 0, maxm, queue[1], NULL );
magma_queue_sync(queue[1]);
magmablas_ztranspose( rows, nb0, dAP, 0, maxm, dAT(s,s), lddat, queue[0]);
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s, s), lddat,
dAT(s, s)+nb0, lddat, queue[0] );
}
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_ztranspose_inplace( m, dAT, dAT_offset, lddat, queue[0] );
magma_zgetmatrix( m, n, dA, dA_offset, ldda, A, lda, queue[0] );
} else {
magmablas_ztranspose( n, m, dAT, dAT_offset, lddat, dA, dA_offset, ldda, queue[0] );
magma_zgetmatrix( m, n, dA, dA_offset, ldda, A, lda, queue[0] );
magma_queue_sync(queue[0]);
magma_free( dAT );
}
magma_queue_sync(queue[0]);
magma_free( dwork );
}
return *info;
} /* magma_zgetrf */
/**
Purpose
-------
ZGETRF 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.
This version assumes the computation runs through the NULL stream
and therefore is not overlapping computation with communication.
Arguments
---------
@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]
dA COMPLEX_16 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.
@param[in]
ldda INTEGER
The leading dimension of the array A. 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_zgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_zgetrf2_gpu(magma_int_t m, magma_int_t n,
magmaDoubleComplex *dA, magma_int_t ldda,
magma_int_t *ipiv, magma_int_t *info)
{
#define dAT(i,j) (dAT + (i)*nb*lddat + (j)*nb)
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, maxn, mindim;
magma_int_t i, rows, cols, s, lddat, lddwork;
magmaDoubleComplex *dAT, *dAP, *work;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < 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;
/* Function Body */
mindim = min(m, n);
nb = magma_get_zgetrf_nb(m);
s = mindim / nb;
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
magma_zmalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_zgetmatrix( m, n, dA, ldda, work, m );
lapackf77_zgetrf(&m, &n, work, &m, ipiv, info);
magma_zsetmatrix( m, n, work, m, dA, ldda );
magma_free_cpu(work);
}
else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
lddat = maxn;
lddwork = maxm;
dAT = dA;
if (MAGMA_SUCCESS != magma_zmalloc( &dAP, nb*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if ( m == n ) {
lddat = ldda;
magmablas_ztranspose_inplace( m, dAT, ldda );
}
else {
if (MAGMA_SUCCESS != magma_zmalloc( &dAT, maxm*maxn )) {
magma_free( dAP );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_ztranspose( m, n, dA, ldda, dAT, lddat );
}
if (MAGMA_SUCCESS != magma_zmalloc_pinned( &work, maxm*nb )) {
magma_free( dAP );
if ( ! (m == n))
magma_free( dAT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
for( i=0; i < s; i++ ) {
// download i-th panel
cols = maxm - i*nb;
//magmablas_ztranspose( nb, cols, dAT(i,i), lddat, dAP, cols );
magmablas_ztranspose( nb, m-i*nb, dAT(i,i), lddat, dAP, cols );
magma_zgetmatrix( m-i*nb, nb, dAP, cols, work, lddwork );
// make sure that gpu queue is empty
magma_device_sync();
if ( i > 0 ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (i+1)*nb, nb,
c_one, dAT(i-1,i-1), lddat,
dAT(i-1,i+1), lddat );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-i*nb, nb,
c_neg_one, dAT(i-1,i+1), lddat,
dAT(i, i-1), lddat,
c_one, dAT(i, i+1), lddat );
}
// do the cpu part
rows = m - i*nb;
lapackf77_zgetrf( &rows, &nb, work, &lddwork, ipiv+i*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + i*nb;
magmablas_zpermute_long2( n, dAT, lddat, ipiv, nb, i*nb );
// upload i-th panel
magma_zsetmatrix( m-i*nb, nb, work, lddwork, dAP, maxm );
//magmablas_ztranspose( cols, nb, dAP, maxm, dAT(i,i), lddat );
magmablas_ztranspose( m-i*nb, nb, dAP, maxm, dAT(i,i), lddat );
// do the small non-parallel computations (next panel update)
if ( s > (i+1) ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(i, i ), lddat,
dAT(i, i+1), lddat);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), lddat,
dAT(i+1, i ), lddat,
c_one, dAT(i+1, i+1), lddat );
}
else {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(i, i ), lddat,
dAT(i, i+1), lddat);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), lddat,
dAT(i+1, i ), lddat,
c_one, dAT(i+1, i+1), lddat );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_ztranspose( nb0, rows, dAT(s,s), lddat, dAP, maxm );
magma_zgetmatrix( rows, nb0, dAP, maxm, work, lddwork );
// make sure that gpu queue is empty
magma_device_sync();
// do the cpu part
lapackf77_zgetrf( &rows, &nb0, work, &lddwork, ipiv+s*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + s*nb;
magmablas_zpermute_long2( n, dAT, lddat, ipiv, nb0, s*nb );
// upload i-th panel
magma_zsetmatrix( rows, nb0, work, lddwork, dAP, maxm );
magmablas_ztranspose( rows, nb0, dAP, maxm, dAT(s,s), lddat );
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s,s), lddat,
dAT(s,s)+nb0, lddat);
if ( m == n ) {
magmablas_ztranspose_inplace( m, dAT, lddat );
}
else {
magmablas_ztranspose( n, m, dAT, lddat, dA, ldda );
magma_free( dAT );
}
magma_free( dAP );
magma_free_pinned( work );
}
return *info;
} /* magma_zgetrf2_gpu */
/**
Purpose
-------
ZGETRF 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.
Use two buffer to send panels.
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_lAT COMPLEX_16 array of pointers on the GPU, dimension (ngpu).
On entry, the M-by-N matrix A distributed over GPUs
(d_lAT[d] points to the local matrix on d-th GPU).
It uses a 1D block column cyclic format (with the block size
nb), and each local matrix is stored by row.
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]
lddat INTEGER
The leading dimension of the array d_lAT[d]. 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 (workspace) on device
d_lAP COMPLEX_16 array of pointers on the GPU, dimension (ngpu).
d_lAP[d] is the workspace on d-th GPU. Each local workspace
must be of size (3+ngpu)*nb*maxm, where maxm is m rounded
up to a multiple of 32 and nb is the block size.
@param (workspace)
W COMPLEX_16 array, dimension (ngpu*nb*maxm).
It is used to store panel on CPU.
@param[in]
ldw INTEGER
The leading dimension of the workspace w.
@param[in]
queues magma_queue_t
queues[d] points to the streams for the d-th GPU to execute
in. Each GPU require two streams.
@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_zgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_zgetrf2_mgpu(
magma_int_t ngpu,
magma_int_t m, magma_int_t n, magma_int_t nb, magma_int_t offset,
magmaDoubleComplex_ptr d_lAT[], magma_int_t lddat, magma_int_t *ipiv,
magmaDoubleComplex_ptr d_lAP[],
magmaDoubleComplex *W, magma_int_t ldw,
magma_queue_t queues[][2],
magma_int_t *info)
{
#define dAT(id,i,j) (d_lAT[(id)] + ((offset)+(i)*nb)*lddat + (j)*nb)
#define W(j) (W + ((j)%ngpu)*nb*ldw)
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t block_size = 32;
magma_int_t iinfo, n_local[MagmaMaxGPUs];
magma_int_t maxm, mindim;
magma_int_t i, j, d, dd, rows, cols, s, ldpan[MagmaMaxGPUs];
magma_int_t id, j_local, j_local2, nb0, nb1, h = 2+ngpu;
magmaDoubleComplex *d_panel[MagmaMaxGPUs], *panel_local[MagmaMaxGPUs];
/* Check arguments */
*info = 0;
if (m < 0)
*info = -2;
else if (n < 0)
*info = -3;
else if (ngpu*lddat < max(1,n))
*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);
if ( ngpu > ceil((double)n/nb) ) {
*info = -1;
return *info;
}
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/* Use hybrid blocked code. */
maxm = ((m + block_size-1)/block_size)*block_size;
/* some initializations */
for (d=0; d < ngpu; d++) {
magma_setdevice(d);
n_local[d] = ((n/nb)/ngpu)*nb;
if (d < (n/nb) % ngpu)
n_local[d] += nb;
else if (d == (n/nb) % ngpu)
n_local[d] += n % nb;
/* workspaces */
d_panel[d] = &(d_lAP[d][h*nb*maxm]); /* temporary panel storage */
}
trace_init( 1, ngpu, 2, (CUstream_st**)queues );
/* start sending the panel to cpu */
nb0 = min(mindim, nb);
magma_setdevice(0);
magmablasSetKernelStream(queues[0][1]);
trace_gpu_start( 0, 1, "comm", "get" );
magmablas_ztranspose( nb0, m, dAT(0,0,0), lddat, d_lAP[0], maxm );
magma_zgetmatrix_async( m, nb0,
d_lAP[0], maxm,
W(0), ldw, queues[0][1] );
trace_gpu_end( 0, 1 );
/* ------------------------------------------------------------------------------------- */
magma_timer_t time=0;
timer_start( time );
s = mindim / nb;
for( j=0; j < s; j++ ) {
/* Set the GPU number that holds the current panel */
id = j % ngpu;
magma_setdevice(id);
/* Set the local index where the current panel is */
j_local = j/ngpu;
cols = maxm - j*nb;
rows = m - j*nb;
/* synchronize j-th panel from id-th gpu into work */
magma_queue_sync( queues[id][1] );
/* j-th panel factorization */
trace_cpu_start( 0, "getrf", "getrf" );
lapackf77_zgetrf( &rows, &nb, W(j), &ldw, ipiv+j*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) ) {
*info = iinfo + j*nb;
}
trace_cpu_end( 0 );
/* start sending the panel to all the gpus */
d = (j+1) % ngpu;
for( dd=0; dd < ngpu; dd++ ) {
magma_setdevice(d);
trace_gpu_start( 0, 1, "comm", "set" );
magma_zsetmatrix_async( rows, nb,
W(j), ldw,
&d_lAP[d][(j%h)*nb*maxm], cols,
queues[d][1] );
trace_gpu_end( 0, 1 );
d = (d+1) % ngpu;
}
/* apply the pivoting */
d = (j+1) % ngpu;
for( dd=0; dd < ngpu; dd++ ) {
magma_setdevice(d);
trace_gpu_start( d, 1, "pivot", "pivot" );
if ( dd == 0 ) {
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
}
magmablas_zlaswp_q( lddat, dAT(d,0,0), lddat, j*nb + 1, j*nb + nb, ipiv, 1, queues[d][0] );
trace_gpu_end( d, 1 );
d = (d+1) % ngpu;
}
/* update the trailing-matrix/look-ahead */
d = (j+1) % ngpu;
for( dd=0; dd < ngpu; dd++ ) {
magma_setdevice(d);
/* storage for panel */
if ( d == id ) {
/* the panel belond to this gpu */
panel_local[d] = dAT(d,j,j_local);
ldpan[d] = lddat;
/* next column */
j_local2 = j_local+1;
} else {
/* the panel belong to another gpu */
panel_local[d] = d_panel[d];
ldpan[d] = nb;
/* next column */
j_local2 = j_local;
if ( d < id ) j_local2 ++;
}
/* the size of the next column */
if ( s > (j+1) ) {
nb0 = nb;
} else {
nb0 = n_local[d]-nb*(s/ngpu);
if ( d < s % ngpu ) nb0 -= nb;
}
if ( d == (j+1) % ngpu) {
/* owns the next column, look-ahead the column */
nb1 = nb0;
magmablasSetKernelStream(queues[d][1]);
/* make sure all the pivoting has been applied */
magma_queue_sync(queues[d][0]);
trace_gpu_start( d, 1, "gemm", "gemm" );
/* transpose panel on GPU */
magmablas_ztranspose( rows, nb, &d_lAP[d][(j%h)*nb*maxm], cols, panel_local[d], ldpan[d] );
/* synch for remaining update */
magma_queue_sync(queues[d][1]);
} else {
/* update the entire trailing matrix */
nb1 = n_local[d] - j_local2*nb;
magmablasSetKernelStream(queues[d][0]);
/* synchronization to make sure panel arrived on gpu */
magma_queue_sync(queues[d][1]);
trace_gpu_start( d, 0, "gemm", "gemm" );
/* transpose panel on GPU */
magmablas_ztranspose( rows, nb, &d_lAP[d][(j%h)*nb*maxm], cols, panel_local[d], ldpan[d] );
}
/* gpu updating the trailing matrix */
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb1, nb, c_one,
panel_local[d], ldpan[d],
dAT(d, j, j_local2), lddat);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
nb1, m-(j+1)*nb, nb,
c_neg_one, dAT(d, j, j_local2), lddat,
&(panel_local[d][nb*ldpan[d]]), ldpan[d],
c_one, dAT(d, j+1, j_local2), lddat );
if ( d == (j+1) % ngpu ) {
/* Set the local index where the current panel is */
int loff = j+1;
int j_local = (j+1)/ngpu;
int ldda = maxm - (j+1)*nb;
int cols = m - (j+1)*nb;
nb0 = min(nb, mindim - (j+1)*nb); /* size of the diagonal block */
trace_gpu_end( d, 1 );
if ( nb0 > 0 ) {
/* transpose the panel for sending it to cpu */
trace_gpu_start( d, 1, "comm", "get" );
magmablas_ztranspose( nb0, m-(j+1)*nb, dAT(d,loff,j_local), lddat, &d_lAP[d][((j+1)%h)*nb*maxm], ldda );
/* send the panel to cpu */
magma_zgetmatrix_async( cols, nb0,
&d_lAP[d][((j+1)%h)*nb*maxm], ldda,
W(j+1), ldw, queues[d][1] );
trace_gpu_end( d, 1 );
}
} else {
trace_gpu_end( d, 0 );
}
d = (d+1) % ngpu;
}
/* update the remaining matrix by gpu owning the next panel */
if ( (j+1) < s ) {
int j_local = (j+1)/ngpu;
int rows = m - (j+1)*nb;
d = (j+1) % ngpu;
magma_setdevice(d);
magmablasSetKernelStream(queues[d][0]);
trace_gpu_start( d, 0, "gemm", "gemm" );
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n_local[d] - (j_local+1)*nb, nb,
c_one, panel_local[d], ldpan[d],
dAT(d,j,j_local+1), lddat );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n_local[d]-(j_local+1)*nb, rows, nb,
c_neg_one, dAT(d,j,j_local+1), lddat,
&(panel_local[d][nb*ldpan[d]]), ldpan[d],
c_one, dAT(d,j+1, j_local+1), lddat );
trace_gpu_end( d, 0 );
}
} /* end of for j=1..s */
/* ------------------------------------------------------------------------------ */
/* Set the GPU number that holds the last panel */
id = s % ngpu;
/* Set the local index where the last panel is */
j_local = s/ngpu;
/* size of the last diagonal-block */
nb0 = min(m - s*nb, n - s*nb);
rows = m - s*nb;
cols = maxm - s*nb;
if ( nb0 > 0 ) {
magma_setdevice(id);
/* wait for the last panel on cpu */
magma_queue_sync( queues[id][1] );
/* factor on cpu */
lapackf77_zgetrf( &rows, &nb0, W(s), &ldw, ipiv+s*nb, &iinfo);
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + s*nb;
/* send the factor to gpus */
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
j_local2 = j_local;
if ( d < id ) j_local2 ++;
if ( d == id || n_local[d] > j_local2*nb ) {
magma_zsetmatrix_async( rows, nb0,
W(s), ldw,
&d_lAP[d][(s%h)*nb*maxm],
cols, queues[d][1] );
}
}
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
if ( d == 0 ) {
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
}
magmablas_zlaswp_q( lddat, dAT(d,0,0), lddat, s*nb + 1, s*nb + nb0, ipiv, 1, queues[d][0] );
}
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(queues[d][1]);
/* wait for the pivoting to be done */
magma_queue_sync( queues[d][0] );
j_local2 = j_local;
if ( d < id ) j_local2++;
if ( d == id ) {
/* the panel belond to this gpu */
panel_local[d] = dAT(d,s,j_local);
/* next column */
nb1 = n_local[d] - j_local*nb-nb0;
magmablas_ztranspose( rows, nb0, &d_lAP[d][(s%h)*nb*maxm], cols, panel_local[d], lddat );
if ( nb1 > 0 ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb1, nb0, c_one,
panel_local[d], lddat,
dAT(d,s,j_local)+nb0, lddat);
}
} else if ( n_local[d] > j_local2*nb ) {
/* the panel belong to another gpu */
panel_local[d] = d_panel[d];
/* next column */
nb1 = n_local[d] - j_local2*nb;
magmablas_ztranspose( rows, nb0, &d_lAP[d][(s%h)*nb*maxm], cols, panel_local[d], nb );
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb1, nb0, c_one,
panel_local[d], nb,
dAT(d,s,j_local2), lddat);
}
}
} /* if ( nb0 > 0 ) */
/* clean up */
trace_finalize( "zgetrf_mgpu.svg","trace.css" );
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_queue_sync( queues[d][0] );
magma_queue_sync( queues[d][1] );
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
timer_start( time );
timer_printf("\n Performance %f GFlop/s\n", FLOPS_ZGETRF(m,n) / 1e9 / time );
return *info;
} /* magma_zgetrf2_mgpu */
/***************************************************************************//**
Purpose
-------
SGELQF computes an LQ factorization of a REAL M-by-N matrix A:
A = L * Q.
Arguments
---------
@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 REAL array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and below the diagonal of the array
contain the m-by-min(m,n) lower trapezoidal matrix L (L is
lower triangular if m <= n); the elements above the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of 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 REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) REAL array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= max(1,M).
For optimum performance LWORK >= M*NB, where NB is the
optimal blocksize.
\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.
\n
TODO: work is currently unused. sgeqrf2 allocates its own work of (m + n)*nb.
@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(k) . . . H(2) H(1), 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:n) is stored on exit in A(i,i+1:n),
and tau in TAU(i).
@ingroup magma_gelqf
*******************************************************************************/
extern "C" magma_int_t
magma_sgelqf(
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda, float *tau,
float *work, magma_int_t lwork,
magma_int_t *info)
{
#define dA(i_, j_) (dA + (i_) + (j_)*ldda)
#define dAT(i_, j_) (dAT + (i_) + (j_)*ldda)
/* Constants */
const float c_one = MAGMA_S_ONE;
const magma_int_t ione = 1;
MAGMA_UNUSED( ione ); // used only for real
/* Local variables */
magmaFloat_ptr dA=NULL, dAT=NULL;
magma_int_t min_mn, maxm, maxn, maxdim, nb;
magma_int_t iinfo, ldda, lddat;
/* Function Body */
*info = 0;
nb = magma_get_sgelqf_nb( m, n );
min_mn = min( m, n );
work[0] = magma_smake_lwork( m*nb );
bool lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
} else if (lwork < max(1,m) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (min_mn == 0) {
work[0] = c_one;
return *info;
}
maxm = magma_roundup( m, 32 );
maxn = magma_roundup( n, 32 );
maxdim = max( maxm, maxn );
magma_queue_t queue = NULL;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
// copy to GPU and transpose
if (maxdim*maxdim < 2*maxm*maxn) {
// close to square, do everything in-place
ldda = maxdim;
lddat = maxdim;
if (MAGMA_SUCCESS != magma_smalloc( &dA, maxdim*maxdim )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queue );
dAT = dA;
magmablas_stranspose_inplace( lddat, dAT(0,0), lddat, queue );
}
else {
// rectangular, do everything out-of-place
ldda = maxm;
lddat = maxn;
if (MAGMA_SUCCESS != magma_smalloc( &dA, 2*maxn*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
magma_ssetmatrix( m, n, A, lda, dA(0,0), ldda, queue );
dAT = dA + maxn * maxm;
magmablas_stranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queue );
}
// factor QR
magma_sgeqrf2_gpu( n, m, dAT(0,0), lddat, tau, &iinfo );
assert( iinfo >= 0 );
if ( iinfo > 0 ) {
*info = iinfo;
}
// conjugate tau
#ifdef COMPLEX
lapackf77_slacgv( &min_mn, tau, &ione );
#endif
// undo transpose
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_stranspose_inplace( lddat, dAT(0,0), lddat, queue );
magma_sgetmatrix( m, n, dA(0,0), ldda, A, lda, queue );
} else {
magmablas_stranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queue );
magma_sgetmatrix( m, n, dA(0,0), ldda, A, lda, queue );
}
cleanup:
magma_queue_destroy( queue );
magma_free( dA );
return *info;
} /* magma_sgelqf */
/**
Purpose
-------
CGETRF 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]
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]
dA COMPLEX 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.
@param[in]
ldda INTEGER
The leading dimension of the array A. 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_cgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_cgetrf_gpu(
magma_int_t m, magma_int_t n,
magmaFloatComplex_ptr dA, magma_int_t ldda,
magma_int_t *ipiv,
magma_int_t *info )
{
#ifdef HAVE_clBLAS
#define dA(i_, j_) dA, (dA_offset + (i_) + (j_)*ldda)
#define dAT(i_, j_) dAT, (dAT_offset + (i_)*lddat + (j_))
#define dAP(i_, j_) dAP, ( (i_) + (j_)*maxm)
#else
#define dA(i_, j_) (dA + (i_) + (j_)*ldda)
#define dAT(i_, j_) (dAT + (i_)*lddat + (j_))
#define dAP(i_, j_) (dAP + (i_) + (j_)*maxm)
#endif
magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, maxn, minmn;
magma_int_t i, j, jb, rows, lddat, ldwork;
magmaFloatComplex_ptr dAT=NULL, dAP=NULL;
magmaFloatComplex *work=NULL;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < 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;
/* Function Body */
minmn = min( m, n );
nb = magma_get_cgetrf_nb( m, n );
magma_queue_t queues[2] = { NULL };
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
if ( MAGMA_SUCCESS != magma_cmalloc_cpu( &work, m*n )) {
*info = MAGMA_ERR_HOST_ALLOC;
goto cleanup;
}
magma_cgetmatrix( m, n, dA(0,0), ldda, work, m, queues[0] );
lapackf77_cgetrf( &m, &n, work, &m, ipiv, info );
magma_csetmatrix( m, n, work, m, dA(0,0), ldda, queues[0] );
magma_free_cpu( work ); work=NULL;
}
else {
/* Use hybrid blocked code. */
maxm = magma_roundup( m, 32 );
maxn = magma_roundup( n, 32 );
if (MAGMA_SUCCESS != magma_cmalloc( &dAP, nb*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
// square matrices can be done in place;
// rectangular requires copy to transpose
if ( m == n ) {
dAT = dA;
lddat = ldda;
magmablas_ctranspose_inplace( m, dAT(0,0), lddat, queues[0] );
}
else {
lddat = maxn; // N-by-M
if (MAGMA_SUCCESS != magma_cmalloc( &dAT, lddat*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
magmablas_ctranspose( m, n, dA(0,0), ldda, dAT(0,0), lddat, queues[0] );
}
magma_queue_sync( queues[0] ); // finish transpose
ldwork = maxm;
if (MAGMA_SUCCESS != magma_cmalloc_pinned( &work, ldwork*nb )) {
*info = MAGMA_ERR_HOST_ALLOC;
goto cleanup;
}
for( j=0; j < minmn-nb; j += nb ) {
// get j-th panel from device
magmablas_ctranspose( nb, m-j, dAT(j,j), lddat, dAP(0,0), maxm, queues[1] );
magma_queue_sync( queues[1] ); // wait for transpose
magma_cgetmatrix_async( m-j, nb, dAP(0,0), maxm, work, ldwork, queues[0] );
if ( j > 0 ) {
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-(j+nb), nb,
c_one, dAT(j-nb, j-nb), lddat,
dAT(j-nb, j+nb), lddat, queues[1] );
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+nb), m-j, nb,
c_neg_one, dAT(j-nb, j+nb), lddat,
dAT(j, j-nb), lddat,
c_one, dAT(j, j+nb), lddat, queues[1] );
}
// do the cpu part
rows = m - j;
magma_queue_sync( queues[0] ); // wait to get work
lapackf77_cgetrf( &rows, &nb, work, &ldwork, ipiv+j, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j;
// send j-th panel to device
magma_csetmatrix_async( m-j, nb, work, ldwork, dAP, maxm, queues[0] );
for( i=j; i < j + nb; ++i ) {
ipiv[i] += j;
}
magmablas_claswp( n, dAT(0,0), lddat, j + 1, j + nb, ipiv, 1, queues[1] );
magma_queue_sync( queues[0] ); // wait to set dAP
magmablas_ctranspose( m-j, nb, dAP(0,0), maxm, dAT(j,j), lddat, queues[1] );
// do the small non-parallel computations (next panel update)
if ( j + nb < minmn - nb ) {
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+nb), lddat, queues[1] );
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+nb), nb,
c_neg_one, dAT(j, j+nb), lddat,
dAT(j+nb, j ), lddat,
c_one, dAT(j+nb, j+nb), lddat, queues[1] );
}
else {
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-(j+nb), nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+nb), lddat, queues[1] );
magma_cgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+nb), m-(j+nb), nb,
c_neg_one, dAT(j, j+nb), lddat,
dAT(j+nb, j ), lddat,
c_one, dAT(j+nb, j+nb), lddat, queues[1] );
}
}
jb = min( m-j, n-j );
if ( jb > 0 ) {
rows = m - j;
magmablas_ctranspose( jb, rows, dAT(j,j), lddat, dAP(0,0), maxm, queues[1] );
magma_cgetmatrix( rows, jb, dAP(0,0), maxm, work, ldwork, queues[1] );
// do the cpu part
lapackf77_cgetrf( &rows, &jb, work, &ldwork, ipiv+j, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j;
for( i=j; i < j + jb; ++i ) {
ipiv[i] += j;
}
magmablas_claswp( n, dAT(0,0), lddat, j + 1, j + jb, ipiv, 1, queues[1] );
// send j-th panel to device
magma_csetmatrix( rows, jb, work, ldwork, dAP(0,0), maxm, queues[1] );
magmablas_ctranspose( rows, jb, dAP(0,0), maxm, dAT(j,j), lddat, queues[1] );
magma_ctrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-j-jb, jb,
c_one, dAT(j,j), lddat,
dAT(j,j+jb), lddat, queues[1] );
}
// undo transpose
if ( m == n ) {
magmablas_ctranspose_inplace( m, dAT(0,0), lddat, queues[1] );
}
else {
magmablas_ctranspose( n, m, dAT(0,0), lddat, dA(0,0), ldda, queues[1] );
}
}
cleanup:
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
magma_free( dAP );
if (m != n) {
magma_free( dAT );
}
magma_free_pinned( work );
return *info;
} /* magma_cgetrf_gpu */
extern "C" magma_int_t
magma_dgetrf_gpu_work_amc(
magma_int_t m, magma_int_t n,
double *dA, magma_int_t dA_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_GPU_WORK_AMC 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
=========
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;
//magma_int_t A_K;
double *dAT;
magma_int_t dAT_LD;
double *dAP_set,*dAP_get;
magma_int_t dAP_LD;
//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;
/* Recommanded dimension in the workspace*/
int A_m, A_n, A_N, A_NMAX, A_LD;
double *A;
#ifdef USE_CALU
int i_nrows;
#endif
amc_args_t *args;
/*magma_event_t *A_event;*/ /*Control bucket*/
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (dA_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;
/* Compute the maximum number of panels we can store in the workspace*/
A_NMAX = (int) (AWORK_n/ nb);
/* Compute the recommanded number of columns for the cpu part*/
A_n = (int) ceil(n*dcpu);
/*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;
}
A = AWORK;
A_m = m;
A_LD = AWORK_LD;
#if (dbglevel >=1)
/* Initialize the tracing*/
ca_dbg_trace_init(nbcores,1); //nbcores + 1 GPU
#endif
#if (dbglevel >=1)
t1 = magma_wtime();
#endif
/*Transfer the first column block of the matrix from the GPU to the CPUs*/
magma_dgetmatrix(A_m, A_n, dA, dA_LD, A, A_LD);
#if (dbglevel >=1)
printf("[DBG] Time First getmatrix: %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
ca_dbg_printMat(m, A_n, A, A_LD,"A after first getMatrix");
#endif
/*Allocate a workspace for the panels transposition*/
dAP_LD = m;
if(dAP_LD%32!=0) dAP_LD = ((dAP_LD + 31)/32)*32;/*Make dAP_LD multiple of 32*/
if (MAGMA_SUCCESS != magma_dmalloc(&dAP_set, dAP_LD*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if (MAGMA_SUCCESS != magma_dmalloc(&dAP_get, dAP_LD*nb)) {
magma_free(dAP_set);
*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
/*Transpose the gpu part of the matrix in/out of place*/
if ((m == n) ){ //&& (m % 32 == 0) && (dA_LD%32 == 0)
dAT = dA;
dAT_LD= dA_LD;
magmablas_dtranspose_inplace(m, dAT, dAT_LD);
}
else {
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;
if (MAGMA_SUCCESS != magma_dmalloc(&dAT, dm_max*dn_max )) {
magma_free(dAP_set);
magma_free(dAP_get);
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dAT_LD = dn_max;
magmablas_dtranspose2( dAT, dAT_LD, dA, dA_LD, m, n );
}
#if (dbglevel >=1)
printf("[DBG] Time First transposition: %f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
ca_dbg_printMat_transpose_gpu(m, n, dAT, dAT_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*/
/*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, 1);
K = 0;
#ifdef USE_CALU
/*initialize calu environment*/
core_dtslu_alloc(panel_num_threads, A_m, nb);
core_dtslu_init(panel_num_threads);
/*Initialize rows indice: required*/
for(I=0;I<A_m;I++) ipiv[I]=I;
#else
/*initialize parallel recursif panel environment*/
CORE_zgetrf_reclap_init();
#endif
magma_schedule_set_task_priority(INT_MAX-1);
/*Schedule the first panel factorization*/
#ifdef USE_CALU
magma_insert_core_dtslu(A_m, nb, A(0,K), A_LD, ipiv(0), &iinfo, panel_num_threads, colptr(K));
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);
for(I=K+1; I<=M-1; I+=B_rows){
i_nrows = min(B_rows*nb, m-I*nb);
magma_insert_core_dtrsm_gatherv('R', 'U', 'N', 'N', i_nrows, nb, c_one, A(0,K), A_LD, A(I,K), A_LD, colptr(K));
}
#else
magma_insert_core_dgetrf_rec(A_m, nb, A(0,K), A_LD, ipiv(0), &iinfo, panel_num_threads, colptr(K));
#endif
/*Transfer the factorized panel to the GPU (transposition included)*/
magma_insert_dsetmatrix_transpose(A_m, nb, A(0,K), A_LD, dAT(0,K), dAT_LD, dAP_set, dAP_LD, colptr(K), dAT(K,K));
#if (dbglevel==10)
magma_schedule_barrier();
ca_dbg_printMat(m, nb, A(0,0), A_LD,"A(0,0)");
ca_dbg_printMat_transpose_gpu(m, n, dAT, dAT_LD,"dA");
#endif
for(K=0;K<=N-1;K++){
/*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]*/
gpu_nrows = m - (K+1)*nb;
gpu_ncols = n - (K+1+A_N-1)*nb;
if(gpu_ncols >0)
{
/*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);
/*schedule a swap of the trailing submatrix in the gpu using ipiv[K]*/
/*dependency dAT((K+1)-1, (K+A_N)-1) = dAT(K, K+A_N-1) with previous dgemm*/
magma_insert_dlaswp(gpu_ncols, dAT(K, K+A_N), dAT_LD, ONE, nb, ipiv(K), ONE, dAT(K, K+A_N-1)); /*non blocking*/
//printf("debug barrier\n");
//magma_schedule_barrier();
magma_insert_dtrsm(MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit, gpu_ncols, nb, c_one, dAT(K,K), dAT_LD, dAT(K,K+A_N), dAT_LD);/*non blocking*/
/* aij^T = aij^T - (lik.ukj)^T = aij^T - ukj^T.lik^T*/
magma_insert_dgemm(MagmaNoTrans,MagmaNoTrans, gpu_ncols, gpu_nrows, nb, c_neg_one, dAT(K,K+A_N), dAT_LD, dAT(K+1,K), dAT_LD, c_one, dAT(K+1,K+A_N), dAT_LD);/*non 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
/*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);
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);
#ifdef USE_CALU
magma_insert_core_dtslu(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
B_rows = (int) ceil((double) (M-(K+1)-1)/panel_num_threads);
B_rows = max(B_rows,4); /*maximun of 4*/
//B_rows = max(B_rows,1);
for(I=K+2; I<=M-1; I+=B_rows){
i_nrows = min(B_rows*nb, m-I*nb);
magma_insert_core_dtrsm_gatherv('R', 'U', 'N', 'N', i_nrows, ncols, c_one, A(K+1,K+1), A_LD, A(I,K+1), A_LD, colptr(K+1));
//dtrsm("R", "U", "N", "N", &nrowPblock, &panel_NB, &dONE, &(A[M*pos+pos]), &LDA, &(A[lpos]), &LDA); //
}
#else
magma_insert_core_dgetrf_rec(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
#endif
/*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;
/*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_set, dAP_LD, colptr(K+1), dAT(K+1,K+1));
}
}
/*Transfer asynchronously one column (column K+A_N) from the GPU to the CPU to balance work*/
/*Make sure this is inserted after all dgemm before it schedules to replace a current panel in case A_N< N*/
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_get, dAP_LD, colptr(K+A_N)); //blocking
/*if A_N==1 there is no look-ahead, so insert the panel here*/
if(A_N==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);
#ifdef USE_CALU
magma_insert_core_dtslu(nrows, ncols, A(K+1,K+1), A_LD, ipiv(K+1), &iinfo, panel_num_threads, colptr(K+1));
B_rows = (int) ceil((double) (M-(K+1)-1)/panel_num_threads);
B_rows = max(B_rows,4); /*maximun of 4*/
//B_rows = max(B_rows,1);
for(I=K+2; I<=M-1; I+=B_rows){
i_nrows = min(B_rows*nb, m-I*nb);
magma_insert_core_dtrsm_gatherv('R', 'U', 'N', 'N', i_nrows, ncols, c_one, A(K+1,K+1), A_LD, A(I,K+1), A_LD, colptr(K+1));
//dtrsm("R", "U", "N", "N", &nrowPblock, &panel_NB, &dONE, &(A[M*pos+pos]), &LDA, &(A[lpos]), &LDA); //
}
#else
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));
#endif
/*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_set, dAP_LD, colptr(K+1), dAT(K+1,K+1));
}
}
#if (dbglevel==10)
magma_schedule_barrier();
ca_dbg_printMat(m, A_n, A, A_LD,"A");
ca_dbg_printMat_transpose_gpu(m, n, dAT, dAT_LD,"dA");
#endif
} //Step K done
/*Wait for all thread termination*/
magma_schedule_barrier();
/*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);
/*dep dAT(K-1): Make sure the last swap is completed, and also the dgemm using the panel*/
// magma_insert_dlaswp(ncols, dAT(K, 0), dAT_LD, ONE, nrows, ipiv(K), ONE, dAT(K-1,0));
magmablas_dlaswp(ncols, dAT(K, 0), dAT_LD, ONE, nrows, ipiv(K), ONE);
#if (dbglevel >=1)
ca_trace_end_1gpu('W');
#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
/*No need for synchro, since dtranspose is blocking*/
if (m == n) {
magmablas_dtranspose_inplace(m, dAT, dAT_LD); //( m, dAT, dAT_LD );
dA = dAT;
}
else {
magmablas_dtranspose2( dA, dA_LD, dAT, dAT_LD, n, m );
magma_free( dAT );
}
#if dbglevel>=1
printf("[DBG] Time Final in/out of place transpose:%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#ifdef USE_CALU
core_dtslu_free();
#endif
magma_free( dAP_set );
magma_free( dAP_get );
#if dbglevel>=1
printf("[DBG] Time memory free (dAP):%f\n",magma_wtime()-t1);
t1 = magma_wtime();
#endif
#if (dbglevel==10)
ca_dbg_printMat_transpose_gpu(m, n, dA, dA_LD,"dA = LU");
#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 */
extern "C" magma_int_t
magma_dgetrf(magma_int_t m, magma_int_t n, double *a, magma_int_t lda,
magma_int_t *ipiv, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
DGETRF 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 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.
If the current stream is NULL, this version replaces it with user defined
stream to overlap computation with communication.
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 dAT(i,j) (dAT + (i)*nb*ldda + (j)*nb)
double *dAT, *dA, *da, *work;
double c_one = MAGMA_D_ONE;
double c_neg_one = MAGMA_D_NEG_ONE;
magma_int_t iinfo, nb;
*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;
nb = magma_get_dgetrf_nb(m);
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code. */
lapackf77_dgetrf(&m, &n, a, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
magma_int_t maxm, maxn, ldda, maxdim;
magma_int_t i, rows, cols, s = min(m, n)/nb;
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
maxdim = max(maxm, maxn);
/* set number of GPUs */
magma_int_t num_gpus = magma_num_gpus();
if ( num_gpus > 1 ) {
/* call multi-GPU non-GPU-resident interface */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
/* explicitly checking the memory requirement */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(double);
int h = 1+(2+num_gpus), num_gpus2 = num_gpus;
int 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) ) );
if( num_gpus > ceil((double)NB/nb) ) {
num_gpus2 = (int)ceil((double)NB/nb);
h = 1+(2+num_gpus2);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
}
if( num_gpus2*NB < n ) {
/* require too much memory, so call non-GPU-resident version */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
ldda = maxn;
work = a;
if (maxdim*maxdim < 2*maxm*maxn) {
// if close to square, allocate square matrix and transpose in-place
if (MAGMA_SUCCESS != magma_dmalloc( &dA, nb*maxm + maxdim*maxdim )) {
/* alloc failed so call non-GPU-resident version */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
ldda = maxdim;
magma_dsetmatrix( m, n, a, lda, da, ldda );
dAT = da;
magmablas_dtranspose_inplace( ldda, dAT, ldda );
}
else {
// if very rectangular, allocate dA and dAT and transpose out-of-place
if (MAGMA_SUCCESS != magma_dmalloc( &dA, (nb + maxn)*maxm )) {
/* alloc failed so call non-GPU-resident version */
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
da = dA + nb*maxm;
magma_dsetmatrix( m, n, a, lda, da, maxm );
if (MAGMA_SUCCESS != magma_dmalloc( &dAT, maxm*maxn )) {
/* alloc failed so call non-GPU-resident version */
magma_free( dA );
magma_dgetrf_m(num_gpus, m, n, a, lda, ipiv, info);
return *info;
}
magmablas_dtranspose2( dAT, ldda, da, maxm, m, n );
}
lapackf77_dgetrf( &m, &nb, work, &lda, ipiv, &iinfo);
/* Define user stream if current stream is NULL */
cudaStream_t stream[2], current_stream;
magmablasGetKernelStream(¤t_stream);
magma_queue_create( &stream[0] );
if (current_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else
stream[1] = current_stream;
for( i = 0; i < s; i++ )
{
// download i-th panel
cols = maxm - i*nb;
if (i>0){
// download i-th panel
magmablas_dtranspose( dA, cols, dAT(i,i), ldda, nb, cols );
// make sure that gpu queue is empty
magma_device_sync();
magma_dgetmatrix_async( m-i*nb, nb, dA, cols, work, lda,
stream[0]);
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (i+1)*nb, nb,
c_one, dAT(i-1,i-1), ldda,
dAT(i-1,i+1), ldda );
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-i*nb, nb,
c_neg_one, dAT(i-1,i+1), ldda,
dAT(i, i-1), ldda,
c_one, dAT(i, i+1), ldda );
// do the cpu part
rows = m - i*nb;
magma_queue_sync( stream[0] );
lapackf77_dgetrf( &rows, &nb, work, &lda, ipiv+i*nb, &iinfo);
}
if (*info == 0 && iinfo > 0)
*info = iinfo + i*nb;
// upload i-th panel
magma_dsetmatrix_async( m-i*nb, nb, work, lda, dA, cols,
stream[0]);
magmablas_dpermute_long2( ldda, dAT, ldda, ipiv, nb, i*nb );
magma_queue_sync( stream[0] );
magmablas_dtranspose( dAT(i,i), ldda, dA, cols, cols, nb);
// do the small non-parallel computations
if (s > (i+1)){
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(i, i ), ldda,
dAT(i, i+1), ldda);
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), ldda,
dAT(i+1, i ), ldda,
c_one, dAT(i+1, i+1), ldda );
}
else{
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(i, i ), ldda,
dAT(i, i+1), ldda);
magma_dgemm( MagmaNoTrans, MagmaNoTrans,
n-(i+1)*nb, m-(i+1)*nb, nb,
c_neg_one, dAT(i, i+1), ldda,
dAT(i+1, i ), ldda,
c_one, dAT(i+1, i+1), ldda );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_dtranspose2( dA, cols, dAT(s,s), ldda, nb0, rows);
magma_dgetmatrix( rows, nb0, dA, cols, work, lda );
// make sure that gpu queue is empty
magma_device_sync();
// do the cpu part
lapackf77_dgetrf( &rows, &nb0, work, &lda, ipiv+s*nb, &iinfo);
if (*info == 0 && iinfo > 0)
*info = iinfo + s*nb;
magmablas_dpermute_long2( ldda, dAT, ldda, ipiv, nb0, s*nb );
magma_dsetmatrix( rows, nb0, work, lda, dA, cols );
magmablas_dtranspose2( dAT(s,s), ldda, dA, cols, rows, nb0);
magma_dtrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s, s), ldda,
dAT(s, s)+nb0, ldda);
}
if (maxdim*maxdim < 2*maxm*maxn) {
magmablas_dtranspose_inplace( ldda, dAT, ldda );
magma_dgetmatrix( m, n, da, ldda, a, lda );
} else {
magmablas_dtranspose2( da, maxm, dAT, ldda, n, m );
magma_dgetmatrix( m, n, da, maxm, a, lda );
magma_free( dAT );
}
magma_free( dA );
magma_queue_destroy( stream[0] );
if (current_stream == NULL) {
magma_queue_destroy( stream[1] );
magmablasSetKernelStream(NULL);
}
}
return *info;
} /* magma_dgetrf */
