extern "C" magma_int_t
magma_slarfb2_gpu( magma_int_t m, magma_int_t n, magma_int_t k,
const float *dV, magma_int_t ldv,
const float *dT, magma_int_t ldt,
float *dC, magma_int_t ldc,
float *dwork, magma_int_t ldwork )
{
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
if (m <= 0 || n <= 0)
return MAGMA_SUCCESS;
// W = C^H V
// magma_sgemm( MagmaTrans, MagmaNoTrans,
magmablas_sgemm_reduce(
n, k, m,
c_one, dC, ldc,
dV, ldv,
c_zero, dwork, ldwork);
// W = W T^H = C^H V T^H
magma_strmm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit,
n, k,
c_one, dT, ldt,
dwork, ldwork);
// C = C - V W^H = C - V T V^H C = (I - V T V^H) C = H C
magma_sgemm( MagmaNoTrans, MagmaTrans,
m, n, k,
c_neg_one, dV, ldv,
dwork, ldwork,
c_one, dC, ldc);
return MAGMA_SUCCESS;
}
void gmm_magma(const Tensor_core<float,2>& A, const Tensor_core<float,2>& B, Tensor_core<float,2>& C,
char TRANSA, char TRANSB, float alpha, float beta)
{
int AL0 = A.rank(0); int AL1 = A.rank(1);
int BL0 = B.rank(0); int BL1 = B.rank(1);
int CL0 = C.rank(0); int CL1 = C.rank(1);
magma_int_t M, N, K, LDA, LDB, LDC;
magma_trans_t transA=magma_trans_const(TRANSA), transB=magma_trans_const(TRANSB);
magmaFloat_ptr d_A, d_B, d_C;
//Set LDA, LDB, and LDC, round up to multiple of 32 for best GPU performance
LDA = ((AL0+31)/32)*32; LDB = ((BL0+31)/32)*32; LDC = ((CL0+31)/32)*32;
// Allocate memory for the matrices on GPU
magma_smalloc(&d_A, LDA*AL1 );
magma_smalloc(&d_B, LDB*BL1 );
magma_smalloc(&d_C, LDC*CL1 );
// Copy data from host (CPU) to device (GPU)
magma_ssetmatrix( AL0, AL1, A.data(), AL0, d_A, LDA );
magma_ssetmatrix( BL0, BL1, B.data(), BL0, d_B, LDB );
if( abs(beta)>1e-32 ) magma_ssetmatrix( CL0, CL1, C.data(), CL0, d_C, LDC );
//Call magma_sgemm
M=( TRANSA=='N' || TRANSA=='n' ) ? AL0:AL1;
K=( TRANSA=='N' || TRANSA=='n' ) ? AL1:AL0;
N=( TRANSB=='N' || TRANSB=='n' ) ? BL1:BL0;
magma_sgemm(transA, transB, M, N, K, alpha, d_A, LDA, d_B, LDB, beta,d_C, LDC);
// Copy solution from device (GPU) to host (CPU)
magma_sgetmatrix(CL0, CL1, d_C, LDC, C.data(), CL0);
// Free memory on GPU
magma_free(d_A); magma_free(d_B); magma_free(d_C);
}
/**
Purpose
-------
SLARFB applies a real block reflector H or its transpose H^H to a
REAL m by n matrix C, from the left.
Arguments
---------
@param[in]
side magma_side_t
- = MagmaLeft: apply H or H^H from the Left
- = MagmaRight: apply H or H^H from the Right
@param[in]
trans magma_trans_t
- = MagmaNoTrans: apply H (No transpose)
- = MagmaTrans: apply H^H (Conjugate transpose)
@param[in]
direct magma_direct_t
Indicates how H is formed from a product of elementary
reflectors
- = MagmaForward: H = H(1) H(2) . . . H(k) (Forward)
- = MagmaBackward: H = H(k) . . . H(2) H(1) (Backward)
@param[in]
storev magma_storev_t
Indicates how the vectors which define the elementary
reflectors are stored:
- = MagmaColumnwise: Columnwise
- = MagmaRowwise: Rowwise
@param[in]
m INTEGER
The number of rows of the matrix C.
@param[in]
n INTEGER
The number of columns of the matrix C.
@param[in]
k INTEGER
The order of the matrix T (= the number of elementary
reflectors whose product defines the block reflector).
@param[in]
dV REAL array on the GPU, dimension
(LDDV,K) if STOREV = MagmaColumnwise
(LDDV,M) if STOREV = MagmaRowwise and SIDE = MagmaLeft
(LDDV,N) if STOREV = MagmaRowwise and SIDE = MagmaRight
The matrix V. See further details.
@param[in]
lddv INTEGER
The leading dimension of the array V.
If STOREV = MagmaColumnwise and SIDE = MagmaLeft, LDDV >= max(1,M);
if STOREV = MagmaColumnwise and SIDE = MagmaRight, LDDV >= max(1,N);
if STOREV = MagmaRowwise, LDDV >= K.
@param[in]
dT REAL array on the GPU, dimension (LDDT,K)
The triangular k by k matrix T in the representation of the
block reflector.
@param[in]
lddt INTEGER
The leading dimension of the array T. LDDT >= K.
@param[in,out]
dC REAL array on the GPU, dimension (LDDC,N)
On entry, the m by n matrix C.
On exit, C is overwritten by H*C, or H^H*C, or C*H, or C*H^H.
@param[in]
lddc INTEGER
The leading dimension of the array C. LDA >= max(1,M).
@param
dwork (workspace) REAL array, dimension (LDWORK,K)
@param[in]
ldwork INTEGER
The leading dimension of the array WORK.
If SIDE = MagmaLeft, LDWORK >= max(1,N);
if SIDE = MagmaRight, LDWORK >= max(1,M);
Further Details
---------------
The shape of the matrix V and the storage of the vectors which define
the H(i) is best illustrated by the following example with n = 5 and
k = 3.
All elements including 0's and 1's are stored, unlike LAPACK.
DIRECT = MagmaForward and DIRECT = MagmaForward and
STOREV = MagmaColumnwise: STOREV = MagmaRowwise:
V = ( 1 0 0 ) V = ( 1 v1 v1 v1 v1 )
( v1 1 0 ) ( 0 1 v2 v2 v2 )
( v1 v2 1 ) ( 0 0 1 v3 v3 )
( v1 v2 v3 )
( v1 v2 v3 )
DIRECT = MagmaBackward and DIRECT = MagmaBackward and
STOREV = MagmaColumnwise: STOREV = MagmaRowwise:
V = ( v1 v2 v3 ) V = ( v1 v1 1 0 0 )
( v1 v2 v3 ) ( v2 v2 v2 1 0 )
( 1 v2 v3 ) ( v3 v3 v3 v3 1 )
( 0 1 v3 )
( 0 0 1 )
@ingroup magma_saux3
********************************************************************/
extern "C" magma_int_t
magma_slarfb_gpu(
magma_side_t side, magma_trans_t trans, magma_direct_t direct, magma_storev_t storev,
magma_int_t m, magma_int_t n, magma_int_t k,
magmaFloat_const_ptr dV, magma_int_t lddv,
magmaFloat_const_ptr dT, magma_int_t lddt,
magmaFloat_ptr dC, magma_int_t lddc,
magmaFloat_ptr dwork, magma_int_t ldwork )
{
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
/* Check input arguments */
magma_int_t info = 0;
if (m < 0) {
info = -5;
} else if (n < 0) {
info = -6;
} else if (k < 0) {
info = -7;
} else if ( ((storev == MagmaColumnwise) && (side == MagmaLeft) && lddv < max(1,m)) ||
((storev == MagmaColumnwise) && (side == MagmaRight) && lddv < max(1,n)) ||
((storev == MagmaRowwise) && lddv < k) ) {
info = -9;
} else if (lddt < k) {
info = -11;
} else if (lddc < max(1,m)) {
info = -13;
} else if ( ((side == MagmaLeft) && ldwork < max(1,n)) ||
((side == MagmaRight) && ldwork < max(1,m)) ) {
info = -15;
}
if (info != 0) {
magma_xerbla( __func__, -(info) );
return info;
}
/* Function Body */
if (m <= 0 || n <= 0) {
return info;
}
// opposite of trans
magma_trans_t transt;
if (trans == MagmaNoTrans)
transt = MagmaTrans;
else
transt = MagmaNoTrans;
// whether T is upper or lower triangular
magma_uplo_t uplo;
if (direct == MagmaForward)
uplo = MagmaUpper;
else
uplo = MagmaLower;
// whether V is stored transposed or not
magma_trans_t notransV, transV;
if (storev == MagmaColumnwise) {
notransV = MagmaNoTrans;
transV = MagmaTrans;
}
else {
notransV = MagmaTrans;
transV = MagmaNoTrans;
}
if ( side == MagmaLeft ) {
// Form H C or H^H C
// Comments assume H C. When forming H^H C, T gets transposed via transt.
// W = C^H V
magma_sgemm( MagmaTrans, notransV,
n, k, m,
c_one, dC, lddc,
dV, lddv,
c_zero, dwork, ldwork);
// W = W T^H = C^H V T^H
magma_strmm( MagmaRight, uplo, transt, MagmaNonUnit,
n, k,
c_one, dT, lddt,
dwork, ldwork);
// C = C - V W^H = C - V T V^H C = (I - V T V^H) C = H C
magma_sgemm( notransV, MagmaTrans,
m, n, k,
c_neg_one, dV, lddv,
dwork, ldwork,
c_one, dC, lddc);
}
else {
// Form C H or C H^H
// Comments assume C H. When forming C H^H, T gets transposed via trans.
// W = C V
magma_sgemm( MagmaNoTrans, notransV,
m, k, n,
c_one, dC, lddc,
dV, lddv,
c_zero, dwork, ldwork);
// W = W T = C V T
magma_strmm( MagmaRight, uplo, trans, MagmaNonUnit,
m, k,
c_one, dT, lddt,
dwork, ldwork);
// C = C - W V^H = C - C V T V^H = C (I - V T V^H) = C H
magma_sgemm( MagmaNoTrans, transV,
m, n, k,
c_neg_one, dwork, ldwork,
dV, lddv,
c_one, dC, lddc);
}
return info;
} /* magma_slarfb */
extern "C" magma_int_t
magma_ssytrf_nopiv(magma_uplo_t uplo, magma_int_t n,
float *A, magma_int_t lda,
magma_int_t *info)
{
/* -- MAGMA (version 1.6.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2011
Purpose
=======
SSYTRF_nopiv computes the LDLt factorization of a real symmetric
matrix A. 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 = U\*\*H * D * U, if UPLO = 'U', or
A = L * D * L\*\*H, if UPLO = 'L',
where U is an upper triangular matrix, L is lower triangular, and
D is a diagonal matrix.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
=========
UPLO (input) CHARACTER*1
= 'U': Upper triangle of A is stored;
= 'L': Lower triangle of A is stored.
N (input) INTEGER
The order of the matrix A. N >= 0.
A (input/output) REAL array, dimension (LDA,N)
On entry, the symmetric matrix A. If UPLO = 'U', the leading
N-by-N upper triangular part of A contains the upper
triangular part of the matrix A, and the strictly lower
triangular part of A is not referenced. If UPLO = 'L', the
leading N-by-N lower triangular part of A contains the lower
triangular part of the matrix A, and the strictly upper
triangular part of A is not referenced.
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization A = U\*\*H*U or A = L*L\*\*H.
Higher performance is achieved if A is in pinned memory, e.g.
allocated using cudaMallocHost.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,N).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
if INFO = -6, the GPU memory allocation failed
> 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
===================================================================== */
/* Local variables */
float zone = MAGMA_S_ONE;
float mzone = MAGMA_S_NEG_ONE;
int upper = (uplo == MagmaUpper);
magma_int_t j, k, jb, ldda, nb, ib, iinfo;
magmaFloat_ptr dA;
magmaFloat_ptr dW;
*info = 0;
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,n)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return MAGMA_ERR_ILLEGAL_VALUE;
}
/* Quick return */
if ( n == 0 )
return MAGMA_SUCCESS;
ldda = ((n+31)/32)*32;
nb = magma_get_ssytrf_nopiv_nb(n);
ib = min(32, nb); // inner-block for diagonal factorization
if ((MAGMA_SUCCESS != magma_smalloc(&dA, n *ldda)) ||
(MAGMA_SUCCESS != magma_smalloc(&dW, nb*ldda))) {
/* alloc failed so call the non-GPU-resident version */
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_queue_t stream[2];
magma_event_t event;
magma_queue_create(&stream[0]);
magma_queue_create(&stream[1]);
magma_event_create( &event );
trace_init( 1, 1, 2, (CUstream_st**)stream );
//if (nb <= 1 || nb >= n)
//{
// lapackf77_spotrf(uplo_, &n, a, &lda, info);
//} else
{
/* Use hybrid blocked code. */
if (upper) {
//=========================================================
// Compute the LDLt factorization A = U'*D*U without pivoting.
// copy matrix to GPU
for (j=0; j<n; j+=nb) {
jb = min(nb, (n-j));
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(j+jb, jb, A(0, j), lda, dA(0, j), ldda, stream[0]);
trace_gpu_end( 0, 0 );
}
// main loop
for (j=0; j<n; j += nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), lda, stream[0]);
trace_gpu_end( 0, 0 );
// copy j-th column of U back to CPU
magma_queue_wait_event( stream[1], event );
trace_gpu_start( 0, 1, "get", "get" );
magma_sgetmatrix_async(j, jb, dA(0, j), ldda, A(0, j), lda, stream[1]);
trace_gpu_end( 0, 1 );
// factorize the diagonal block
magma_queue_sync(stream[0]);
trace_cpu_start( 0, "potrf", "potrf" );
ssytrf_nopiv_cpu(MagmaUpper, jb, ib, A(j, j), lda, info);
trace_cpu_end( 0 );
if (*info != 0){
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), lda, dA(j, j), ldda, stream[0]);
trace_gpu_end( 0, 0 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
magmablasSetKernelStream( stream[0] );
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm(MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaUnit,
jb, (n-j-jb),
zone, dA(j, j), ldda,
dA(j, j+jb), ldda);
magma_scopymatrix( jb, n-j-jb, dA( j, j+jb ), ldda, dWt( 0, j+jb ), nb );
// update the trailing submatrix with D
magmablas_slascl_diag(MagmaUpper, jb, n-j-jb,
dA(j, j), ldda,
dA(j, j+jb), ldda,
&iinfo);
magma_event_record( event, stream[0] );
trace_gpu_end( 0, 0 );
// update the trailing submatrix with U and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k<n; k+=nb)
{
magma_int_t kb = min(nb,n-k);
magma_sgemm(MagmaConjTrans, MagmaNoTrans, kb, n-k, jb,
mzone, dWt(0, k), nb,
dA(j, k), ldda,
zone, dA(k, k), ldda);
}
trace_gpu_end( 0, 0 );
}
}
} else {
//=========================================================
// Compute the LDLt factorization A = L*D*L' without pivoting.
// copy the matrix to GPU
for (j=0; j<n; j+=nb) {
jb = min(nb, (n-j));
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async((n-j), jb, A(j, j), lda, dA(j, j), ldda, stream[0]);
trace_gpu_end( 0, 0 );
}
// main loop
for (j=0; j<n; j+=nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), lda, stream[0]);
trace_gpu_end( 0, 0 );
// copy j-th row of L back to CPU
magma_queue_wait_event( stream[1], event );
trace_gpu_start( 0, 1, "get", "get" );
magma_sgetmatrix_async(jb, j, dA(j, 0), ldda, A(j, 0), lda, stream[1]);
trace_gpu_end( 0, 1 );
// factorize the diagonal block
magma_queue_sync(stream[0]);
trace_cpu_start( 0, "potrf", "potrf" );
ssytrf_nopiv_cpu(MagmaLower, jb, ib, A(j, j), lda, info);
trace_cpu_end( 0 );
if (*info != 0){
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), lda, dA(j, j), ldda, stream[0]);
trace_gpu_end( 0, 0 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
magmablasSetKernelStream( stream[0] );
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm(MagmaRight, MagmaLower, MagmaConjTrans, MagmaUnit,
(n-j-jb), jb,
zone, dA(j, j), ldda,
dA(j+jb, j), ldda);
magma_scopymatrix( n-j-jb,jb, dA( j+jb, j ), ldda, dW( j+jb, 0 ), ldda );
// update the trailing submatrix with D
magmablas_slascl_diag(MagmaLower, n-j-jb, jb,
dA(j, j), ldda,
dA(j+jb, j), ldda,
&iinfo);
magma_event_record( event, stream[0] );
trace_gpu_end( 0, 0 );
// update the trailing submatrix with L and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k<n; k+=nb)
{
magma_int_t kb = min(nb,n-k);
magma_sgemm(MagmaNoTrans, MagmaConjTrans, n-k, kb, jb,
mzone, dA(k, j), ldda,
dW(k, 0), ldda,
zone, dA(k, k), ldda);
}
trace_gpu_end( 0, 0 );
}
}
}
}
trace_finalize( "ssytrf.svg","trace.css" );
magma_queue_destroy(stream[0]);
magma_queue_destroy(stream[1]);
magma_event_destroy( event );
magma_free(dW);
magma_free(dA);
return MAGMA_SUCCESS;
} /* magma_ssytrf_nopiv */
/**
Purpose
=======
SSYTRF_nopiv computes the LDLt factorization of a real symmetric
matrix A. 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 = U^H * D * U, if UPLO = MagmaUpper, or
A = L * D * L^H, if UPLO = MagmaLower,
where U is an upper triangular matrix, L is lower triangular, and
D is a diagonal matrix.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangle of A is stored;
- = MagmaLower: Lower triangle of A is stored.
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the symmetric matrix A. If UPLO = MagmaUpper, the leading
N-by-N upper triangular part of A contains the upper
triangular part of the matrix A, and the strictly lower
triangular part of A is not referenced. If UPLO = MagmaLower, the
leading N-by-N lower triangular part of A contains the lower
triangular part of the matrix A, and the strictly upper
triangular part of A is not referenced.
\n
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization A = U^H D U or A = L D L^H.
\n
Higher performance is achieved if A is in pinned memory.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
if INFO = -6, the GPU memory allocation failed
- > 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
@ingroup magma_ssysv_comp
******************************************************************* */
extern "C" magma_int_t
magma_ssytrf_nopiv(
magma_uplo_t uplo, magma_int_t n,
float *A, magma_int_t lda,
magma_int_t *info)
{
#define A(i, j) ( A +(j)*lda + (i))
#define dA(i, j) (dA +(j)*ldda + (i))
#define dW(i, j) (dW +(j)*ldda + (i))
#define dWt(i, j) (dW +(j)*nb + (i))
/* Constants */
const float c_one = MAGMA_S_ONE;
const float c_neg_one = MAGMA_S_NEG_ONE;
/* Local variables */
bool upper = (uplo == MagmaUpper);
magma_int_t j, k, jb, ldda, nb, ib, iinfo;
magmaFloat_ptr dA;
magmaFloat_ptr dW;
*info = 0;
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,n)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return MAGMA_ERR_ILLEGAL_VALUE;
}
/* Quick return */
if ( n == 0 )
return MAGMA_SUCCESS;
ldda = magma_roundup( n, 32 );
nb = magma_get_ssytrf_nopiv_nb(n);
ib = min(32, nb); // inner-block for diagonal factorization
if ((MAGMA_SUCCESS != magma_smalloc(&dA, n *ldda)) ||
(MAGMA_SUCCESS != magma_smalloc(&dW, nb*ldda))) {
/* alloc failed so call the non-GPU-resident version */
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_device_t cdev;
magma_queue_t queues[2];
magma_event_t event;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queues[0] );
magma_queue_create( cdev, &queues[1] );
magma_event_create( &event );
trace_init( 1, 1, 2, queues );
/* Use hybrid blocked code. */
if (upper) {
//=========================================================
// Compute the LDLt factorization A = U'*D*U without pivoting.
// copy matrix to GPU
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(j+jb, jb, A(0, j), lda, dA(0, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
}
// main loop
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
if ( j != 0) {
//magma_event_sync(event);
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), lda, queues[1]);
}
trace_gpu_end( 0, 0 );
// factorize the diagonal block
magma_queue_sync(queues[1]);
trace_cpu_start( 0, "potrf", "potrf" );
magma_ssytrf_nopiv_cpu( MagmaUpper, jb, ib, A(j, j), lda, info );
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), lda, dA(j, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
// copy j-th column of U back to CPU
trace_gpu_start( 0, 1, "get", "get" );
magma_sgetmatrix_async(j, jb, dA(0, j), ldda, A(0, j), lda, queues[1]);
trace_gpu_end( 0, 1 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaUnit,
jb, (n-j-jb),
c_one, dA(j, j), ldda,
dA(j, j+jb), ldda,
queues[0] );
magma_scopymatrix( jb, n-j-jb, dA( j, j+jb ), ldda, dWt( 0, j+jb ), nb, queues[0] );
// update the trailing submatrix with D
magmablas_slascl_diag( MagmaUpper, jb, n-j-jb,
dA(j, j), ldda,
dA(j, j+jb), ldda,
queues[0], &iinfo);
trace_gpu_end( 0, 0 );
// update the trailing submatrix with U and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k < n; k += nb) {
magma_int_t kb = min(nb,n-k);
magma_sgemm( MagmaConjTrans, MagmaNoTrans, kb, n-k, jb,
c_neg_one, dWt(0, k), nb,
dA(j, k), ldda,
c_one, dA(k, k), ldda,
queues[0]);
if (k == j+jb) {
// magma_event_record( event, queues[0] );
magma_queue_sync( queues[0] );
}
}
trace_gpu_end( 0, 0 );
}
}
} else {
//=========================================================
// Compute the LDLt factorization A = L*D*L' without pivoting.
// copy the matrix to GPU
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async((n-j), jb, A(j, j), lda, dA(j, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
}
// main loop
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
if (j != 0) {
//magma_event_sync(event);
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), lda, queues[1]);
}
trace_gpu_end( 0, 0 );
// factorize the diagonal block
magma_queue_sync(queues[1]);
trace_cpu_start( 0, "potrf", "potrf" );
magma_ssytrf_nopiv_cpu( MagmaLower, jb, ib, A(j, j), lda, info );
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), lda, dA(j, j), ldda, queues[0]);
trace_gpu_end( 0, 0 );
// copy j-th row of L back to CPU
trace_gpu_start( 0, 1, "get", "get" );
magma_sgetmatrix_async(jb, j, dA(j, 0), ldda, A(j, 0), lda, queues[1]);
trace_gpu_end( 0, 1 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaUnit,
(n-j-jb), jb,
c_one, dA(j, j), ldda,
dA(j+jb, j), ldda,
queues[0] );
magma_scopymatrix( n-j-jb,jb, dA( j+jb, j ), ldda, dW( j+jb, 0 ), ldda, queues[0] );
// update the trailing submatrix with D
magmablas_slascl_diag( MagmaLower, n-j-jb, jb,
dA(j, j), ldda,
dA(j+jb, j), ldda,
queues[0], &iinfo );
trace_gpu_end( 0, 0 );
// update the trailing submatrix with L and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k < n; k += nb) {
magma_int_t kb = min(nb,n-k);
magma_sgemm( MagmaNoTrans, MagmaConjTrans, n-k, kb, jb,
c_neg_one, dA(k, j), ldda,
dW(k, 0), ldda,
c_one, dA(k, k), ldda,
queues[0] );
if (k == j+jb) {
//magma_event_record( event, queues[0] );
magma_queue_sync(queues[0]);
}
}
trace_gpu_end( 0, 0 );
}
}
}
trace_finalize( "ssytrf.svg","trace.css" );
magma_queue_destroy(queues[0]);
magma_queue_destroy(queues[1]);
magma_event_destroy( event );
magma_free(dW);
magma_free(dA);
return MAGMA_SUCCESS;
} /* magma_ssytrf_nopiv */
/**
Purpose
-------
SLAHRU is an auxiliary MAGMA routine that is used in SGEHRD to update
the trailing sub-matrices after the reductions of the corresponding
panels.
See further details below.
Arguments
---------
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in]
ihi INTEGER
Last row to update. Same as IHI in sgehrd.
@param[in]
k INTEGER
Number of rows of the matrix Am (see details below)
@param[in]
nb INTEGER
Block size
@param[out]
A REAL array, dimension (LDA,N-K)
On entry, the N-by-(N-K) general matrix to be updated. The
computation is done on the GPU. After Am is updated on the GPU
only Am(1:NB) is transferred to the CPU - to update the
corresponding Am matrix. See Further Details below.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[in,out]
dA REAL array on the GPU, dimension (LDDA,N-K).
On entry, the N-by-(N-K) general matrix to be updated.
On exit, the 1st K rows (matrix Am) of A are updated by
applying an orthogonal transformation from the right
Am = Am (I-V T V'), and sub-matrix Ag is updated by
Ag = (I - V T V') Ag (I - V T V(NB+1:)' )
where Q = I - V T V' represent the orthogonal matrix
(as a product of elementary reflectors V) used to reduce
the current panel of A to upper Hessenberg form. After Am
is updated Am(:,1:NB) is sent to the CPU.
See Further Details below.
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,N).
@param[in,out]
dY (workspace) REAL array on the GPU, dimension (LDDY, NB).
On entry the (N-K)-by-NB Y = A V. It is used internally
as workspace, so its value is changed on exit.
@param[in]
lddy INTEGER
The leading dimension of the array dY. LDDY >= max(1,N).
@param[in,out]
dV (workspace) REAL array on the GPU, dimension (LDDV, NB).
On entry the (N-K)-by-NB matrix V of elementary reflectors
used to reduce the current panel of A to upper Hessenberg form.
The rest K-by-NB part is used as workspace. V is unchanged on
exit.
@param[in]
lddv INTEGER
The leading dimension of the array dV. LDDV >= max(1,N).
@param[in]
dT REAL array on the GPU, dimension (NB, NB).
On entry the NB-by-NB upper trinagular matrix defining the
orthogonal Hessenberg reduction transformation matrix for
the current panel. The lower triangular part are 0s.
@param
dwork (workspace) REAL array on the GPU, dimension N*NB.
Further Details
---------------
This implementation follows the algorithm and notations described in:
S. Tomov and J. Dongarra, "Accelerating the reduction to upper Hessenberg
form through hybrid GPU-based computing," University of Tennessee Computer
Science Technical Report, UT-CS-09-642 (also LAPACK Working Note 219),
May 24, 2009.
The difference is that here Am is computed on the GPU.
M is renamed Am, G is renamed Ag.
@ingroup magma_sgeev_aux
********************************************************************/
extern "C" magma_int_t
magma_slahru(
magma_int_t n, magma_int_t ihi, magma_int_t k, magma_int_t nb,
float *A, magma_int_t lda,
float *dA, magma_int_t ldda,
float *dY, magma_int_t lddy,
float *dV, magma_int_t lddv,
float *dT,
float *dwork )
{
#define dA(i_,j_) (dA + (i_) + (j_)*ldda)
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *dYm = dV + ihi - k;
magma_int_t info = 0;
if (n < 0) {
info = -1;
} else if (ihi < 0 || ihi > n) {
info = -2;
} else if (k < 0 || k > n) {
info = -3;
} else if (nb < 1 || nb > n) {
info = -4;
} else if (lda < max(1,n)) {
info = -6;
} else if (ldda < max(1,n)) {
info = -8;
} else if (lddy < max(1,n)) {
info = -10;
} else if (lddv < max(1,n)) {
info = -12;
}
if (info != 0) {
magma_xerbla( __func__, -(info) );
return info;
}
// top part of Y, above panel, hasn't been computed yet, so do that now
// Ym = Am V = A(0:k-1, 0:ihi-k-1) * V(0:ihi-k-1, 0:nb-1)
magma_sgemm( MagmaNoTrans, MagmaNoTrans, k, nb, ihi-k,
c_one, dA, ldda,
dV, lddv,
c_zero, dYm, ldda );
// -----
// on right, A := A Q = A - A V T V'
// Update Am = Am - Am V T V' = Am - Ym W', with W = V T'
// W = V T' = V(0:ihi-k-1, 0:nb-1) * T(0:nb-1, 0:nb-1)'
magma_sgemm( MagmaNoTrans, MagmaConjTrans, ihi-k, nb, nb,
c_one, dV, lddv,
dT, nb,
c_zero, dwork, ldda );
// Am = Am - Ym W' = A(0:k-1, 0:ihi-k-1) - Ym(0:k-1, 0:nb-1) * W(0:ihi-k-1, 0:nb-1)'
magma_sgemm( MagmaNoTrans, MagmaConjTrans, k, ihi-k, nb,
c_neg_one, dYm, ldda,
dwork, ldda,
c_one, dA, ldda );
// copy first nb columns of Am, A(0:k-1, 0:nb-1), to host
magma_sgetmatrix( k, nb, dA, ldda, A, lda );
// -----
// on right, A := A Q = A - A V T V'
// Update Ag = Ag - Ag V T V' = Ag - Y W'
// Ag = Ag - Y W' = A(k:ihi-1, nb:ihi-k-1) - Y(0:ihi-k-1, 0:nb-1) * W(nb:ihi-k-1, 0:nb-1)'
magma_sgemm( MagmaNoTrans, MagmaConjTrans, ihi-k, ihi-k-nb, nb,
c_neg_one, dY, ldda,
dwork + nb, ldda,
c_one, dA(k,nb), ldda );
// -----
// on left, A := Q' A = A - V T' V' A
// Ag2 = Ag2 - V T' V' Ag2 = W Yg, with W = V T' and Yg = V' Ag2
// Note that Ag is A(k:ihi, nb+1:ihi-k)
// while Ag2 is A(k:ihi, nb+1: n -k)
// Z = V(0:ihi-k-1, 0:nb-1)' * A(k:ihi-1, nb:n-k-1); Z is stored over Y
magma_sgemm( MagmaConjTrans, MagmaNoTrans, nb, n-k-nb, ihi-k,
c_one, dV, lddv,
dA(k,nb), ldda,
c_zero, dY, nb );
// Ag2 = Ag2 - W Z = A(k:ihi-1, nb:n-k-1) - W(nb:n-k-1, 0:nb-1) * Z(0:nb-1, nb:n-k-1)
magma_sgemm( MagmaNoTrans, MagmaNoTrans, ihi-k, n-k-nb, nb,
c_neg_one, dwork, ldda,
dY, nb,
c_one, dA(k,nb), ldda );
return info;
}
void magmablas_ssymm_mgpu_spec(
magma_side_t side, magma_uplo_t uplo, magma_int_t m, magma_int_t n,
float alpha,
magmaFloat_ptr dA[], magma_int_t ldda, magma_int_t offset,
magmaFloat_ptr dB[], magma_int_t lddb,
float beta,
magmaFloat_ptr dC[], magma_int_t lddc,
magmaFloat_ptr dwork[], magma_int_t dworksiz,
float *C, magma_int_t ldc,
float *work[], magma_int_t worksiz, // TODO unused
magma_int_t ngpu, magma_int_t nb,
magma_queue_t queues[][20], magma_int_t nqueue,
magma_event_t redevents[][MagmaMaxGPUs*MagmaMaxGPUs+10], magma_int_t nbevents,
magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx )
{
#define dA(dev, i, j) (dA[dev] + (i) + (j)*ldda)
#define dB(dev, i, j) (dB[dev] + (i) + (j)*lddb)
#define dC(dev, i, j) (dC[dev] + (i) + (j)*lddc)
#define dwork(dev, i, j) (dwork[dev] + (i) + (j)*lddwork)
#define C(i, j) (C + (i) + (j)*ldc)
if ( side != MagmaLeft || uplo != MagmaLower ) {
fprintf( stderr, "%s: only Left Lower implemented\n", __func__ );
}
assert( ldda >= m );
assert( lddb >= m );
assert( lddc >= m );
assert( nqueue >= ngpu );
assert( nbevents >= ngpu*ngpu );
magmaFloat_ptr dwork1[MagmaMaxGPUs];
magmaFloat_ptr dwork2[MagmaMaxGPUs];
magma_int_t lddwork = lddc;
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
dwork1[dev] = dwork[dev];
dwork2[dev] = dwork[dev]+n*lddwork;
}
assert( dworksiz >= (2*n*lddwork) );
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_t cstream;
magmablasGetKernelStream(&cstream);
magma_int_t dev, devperm, myblk, mycolsize, myblkoffst;
magma_int_t gdev, gcolsize, gmaster, gngpu;
magma_int_t masterdev, lcdev, lccolsize, myngpu;
magma_int_t stdev = (offset/nb)%ngpu;
magma_int_t blockoffset = offset % nb;
magma_int_t fstblksiz = 0;
if(blockoffset>0){
fstblksiz = min(m, (nb - blockoffset));
}
//magma_int_t nbblk = magma_ceildiv(m, nb);
magma_int_t nbblk = magma_ceildiv((m+blockoffset), nb);
magma_int_t maxgsize = n*nb*magma_ceildiv(nbblk, ngpu);
magma_int_t remm = m- fstblksiz;
magma_int_t nbblkoffst = offset/nb;
magma_int_t nblstblks = -1;
magma_int_t devlstblk = -1;
magma_int_t lstblksiz = remm%nb;
if(lstblksiz>0){
nblstblks = nbblk%ngpu;
devlstblk = (nblstblks-1+ngpu)%ngpu;
}
magma_int_t nbcmplxactive = 0;
magma_int_t cmplxisactive[MagmaMaxGPUs];
magma_int_t gpuisactive[MagmaMaxGPUs];
memset(gpuisactive, 0, MagmaMaxGPUs*sizeof(magma_int_t));
memset(cmplxisactive, 0, MagmaMaxGPUs*sizeof(magma_int_t));
//*******************************
// each GPU make a GEMM with the
// transpose of its blocks to compute
// a final portion of X=A*VT
//*******************************
/* dB = V*T already ==> dB**H = T**H * V**H
* compute T**H * V**H * X is equal to compute locally (VT)**H_i*X_i
* then each GPU broadcast its X_i to assemble the full X which is used
* to compute W = X - 0.5 * V * T**H * V**H * X = X - 0.5 * V *dwork3
*/
if(ngpu ==1){
magma_setdevice( 0 );
magmablasSetKernelStream( queues[ 0 ][ 0 ] );
// compute X[me] = A*VT = A[me]^tr *VT;
magma_sgemm( MagmaConjTrans, MagmaNoTrans, m, n, m,
alpha, dA(0, offset, offset), ldda,
dB[0], lddb,
beta, dC[0], lddc );
return;
}
//ngpu>1
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
masterdev = -1;
gnode[cmplxid][MagmaMaxGPUs+1] = -1;
myngpu = gnode[cmplxid][MagmaMaxGPUs];
for( magma_int_t idev = 0; idev < myngpu; ++idev ) {
dev = gnode[cmplxid][idev];
devperm = (dev-stdev+ngpu)%ngpu;
myblk = (nbblk/ngpu) + (nbblk%ngpu > devperm ? 1:0 );
mycolsize = myblk*nb;
myblkoffst = nb*((nbblkoffst/ngpu)+(nbblkoffst%ngpu > dev?1:0));
if(dev==stdev){
mycolsize -= blockoffset;
myblkoffst += blockoffset; // local index in parent matrix
}
if((devperm==devlstblk)&&(lstblksiz>0)){
mycolsize -= (nb-(remm%nb));
}
mycolsize = min(mycolsize, m);
if(mycolsize>0){
if(masterdev==-1) masterdev = dev;
//printf("dev %d devperm %d on cmplx %d master %d nbblk %d myblk %d m %d n %d mycolsize %d stdev %d fstblksize %d lastdev %d lastsize %d dA(%d, %d, %d) ==> dwork(%d, %d)\n", dev, devperm, cmplxid, masterdev, nbblk, myblk, m, n, mycolsize, stdev, fstblksiz, devlstblk, remm%nb, dev, offset, myblkoffst, dev, maxgsize*dev);
gpuisactive[dev] = mycolsize;
magma_setdevice( dev );
magmablasSetKernelStream( queues[ dev ][ dev ] );
magma_sgemm( MagmaConjTrans, MagmaNoTrans, mycolsize, n, m,
alpha, dA(dev, offset, myblkoffst), ldda,
dB(dev, 0, 0), lddb,
beta, &dwork[dev][maxgsize*dev], mycolsize );
magma_event_record(redevents[dev][dev*ngpu+dev], queues[dev][dev]);
}
if(dev == masterdev){
nbcmplxactive = nbcmplxactive +1;
cmplxisactive[cmplxid] = 1;
gnode[cmplxid][MagmaMaxGPUs+1] = masterdev;
}
}
}
/*
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_queue_sync( queues[ dev ][ dev ] );
}
*/
//*******************************
// each Master GPU has the final
// result either by receiving
// from CPU of by making the add
// by himself, so now it is time
// to broadcast over the GPUs of
// its board.
//*******************************
//printf("=======================================================================\n");
//printf(" sending \n");
//printf("=======================================================================\n");
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
myngpu = gnode[cmplxid][MagmaMaxGPUs];
masterdev = gnode[cmplxid][MagmaMaxGPUs+1];
for( magma_int_t idev = 0; idev < myngpu; ++idev ) {
dev = gnode[cmplxid][idev];
mycolsize = gpuisactive[dev];
if(mycolsize>0){
// I am an active GPU send my portion local
// to all active gpu of my cmplex and global to the
// active master of the other real and they should
// send it out to their actives slaves.
magma_setdevice( dev );
//==============================================
// sending to the master of the active real
//==============================================
//printf ("\n\n**************GPU %d\n ", dev);
//printf (" GPU %d sending to cmplx masters\n", dev);
for( magma_int_t k = 0; k < nbcmplx; ++k ) {
if(k!=cmplxid){
gmaster = gnode[k][MagmaMaxGPUs+1];
if(gmaster!=-1){ //real is active
//printf (" device %d from cmplx %d is sending to master %d on cmplx %d block of size %d event %d\n", dev, cmplxid, gmaster, k, mycolsize, redevents[dev][gmaster*ngpu+dev]);
magma_queue_wait_event(queues[ dev ][ gmaster ], redevents[dev][dev*ngpu+dev]);
magma_scopymatrix_async(
mycolsize, n,
&dwork[dev ][maxgsize*dev], mycolsize,
&dwork[gmaster][maxgsize*dev], mycolsize, queues[dev][gmaster] );
magma_event_record(redevents[dev][gmaster*ngpu+dev], queues[dev][gmaster]);
}
}
}
//==============================================
//
//==============================================
// sending to the active GPUs of my real
//==============================================
//printf (" GPU %d sending internal\n", dev);
for( magma_int_t l = 0; l < myngpu; ++l ) {
lcdev = gnode[cmplxid][l];
lccolsize = gpuisactive[lcdev];
if((lcdev!=dev)&&(lccolsize>0)){
//printf (" device %d from cmplx %d is sending internal to dev %d block of size %d event %d\n", dev, cmplxid, lcdev, mycolsize, redevents[dev][lcdev*ngpu+dev]);
magma_queue_wait_event(queues[ dev ][ lcdev ], redevents[dev][dev*ngpu+dev]);
magma_scopymatrix_async(
mycolsize, n,
&dwork[dev ][maxgsize*dev], mycolsize,
&dwork[lcdev][maxgsize*dev], mycolsize, queues[dev][lcdev] );
magma_event_record(redevents[dev][lcdev*ngpu+dev], queues[dev][lcdev]);
}
}
//==============================================
}// end if mycolsize>0
}// for idev
}// for cmplxid
//printf("=======================================================================\n");
//printf(" master wait and resend internally \n");
//printf("=======================================================================\n");
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
myngpu = gnode[cmplxid][MagmaMaxGPUs];
masterdev = gnode[cmplxid][MagmaMaxGPUs+1];
//==============================================
// if I am active master so wait receiving contribution
// of the GPUs of other real and send it locally
//==============================================
if(masterdev != -1){
mycolsize = gpuisactive[masterdev];
magma_setdevice( masterdev );
//printf(" GPU %d distributing internal\n", masterdev);
for( magma_int_t k = 0; k < nbcmplx; ++k ) {
if(k!=cmplxid){
gngpu = gnode[k][MagmaMaxGPUs];
for( magma_int_t g = 0; g < gngpu; ++g ) {
gdev = gnode[k][g];
gcolsize = gpuisactive[gdev];
// check if I received from this GPU,
// if yes send it to my group
if(gcolsize>0){
magma_queue_wait_event(queues[ masterdev ][ gdev ], redevents[gdev][masterdev*ngpu+gdev]);
for( magma_int_t l = 0; l < myngpu; ++l ) {
lcdev = gnode[cmplxid][l];
lccolsize = gpuisactive[lcdev];
if((lcdev!=masterdev)&&(lccolsize>0)){
//printf(" Master %d on cmplx %d waiting on event %d is distributing internal results of %d to lcdev %d block of size %d event %d\n", masterdev, cmplxid, redevents[gdev][masterdev*ngpu+gdev], gdev, lcdev, gcolsize, redevents[masterdev][lcdev*ngpu+gdev]);
magma_scopymatrix_async(
gcolsize, n,
&dwork[masterdev][maxgsize*gdev], gcolsize,
&dwork[lcdev ][maxgsize*gdev], gcolsize, queues[masterdev][gdev] );
magma_event_record(redevents[masterdev][lcdev*ngpu+gdev], queues[masterdev][gdev]);
}
}
}
}
}
}
}// if active master
//==============================================
}// for cmplxid
/*
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_queue_sync( queues[ dev ][ 0 ] );
for( magma_int_t s = 0; s < ngpu; ++s ) {
magma_queue_sync( queues[ dev ][ s ] );
}
}
*/
//printf("=======================================================================\n");
//printf(" distributing \n");
//printf("=======================================================================\n");
magma_int_t lcblki, gbblki, gblk, ib;
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
myngpu = gnode[cmplxid][MagmaMaxGPUs];
masterdev = gnode[cmplxid][MagmaMaxGPUs+1];
for( magma_int_t idev = 0; idev < myngpu; ++idev ) {
dev = gnode[cmplxid][idev];
mycolsize = gpuisactive[dev];
if(mycolsize>0){ // I am an active GPU
//printf("\n\n==============GPU %d collecting\n", dev);
magma_setdevice( dev );
// collect my results first as tyhere is no need to wait to
// receive nothing, just wait that my gemm are done.
// in theory this should be inside the loop but cuda was not
// able to run it first for all gpu and on gpu>0 it was waiting
// however it was on different stream so it should run. but maybe
// this is because there are too many function call and this make
// cuda not handleit so nice. anyway it coul dbe removed when cuda
// is able to lunch it first without wait.
gdev = dev;
gcolsize = gpuisactive[gdev];
if(gcolsize>0){
devperm = (gdev-stdev+ngpu)%ngpu;
gblk = (nbblk/ngpu) + (nbblk%ngpu > devperm ? 1:0 );
magmablasSetKernelStream( queues[ dev ][ gdev ] );
magma_queue_wait_event(queues[ dev ][ gdev ], redevents[gdev][dev*ngpu+gdev]);
//printf (" GPU %d stream %d doing slacpy\n", dev, queues[ dev ][ gdev ]);
for( magma_int_t blki = 0; blki < gblk; ++blki){
gbblki = (blki*ngpu + devperm)*nb - blockoffset;
lcblki = blki*nb;
ib = nb;//min(nb, m-gbblki);
if(gdev==stdev){
lcblki = blki*nb-blockoffset;
if(blki==0){
gbblki = 0;
lcblki = 0;
ib = nb-blockoffset;
}
}
ib = min(ib, m-gbblki);
//printf(" blockoffset %d nbblk %d stdev %d receiving from gdev %d gblk %d gcolsize %d copying blki %d of size ibxn %dx%d from work[%d] to C[%d]\n", blockoffset, nbblk, stdev, gdev, gblk, gcolsize, blki, ib, n, lcblki, gbblki);
magmablas_slacpy( MagmaFull, ib, n, &dwork[dev][maxgsize*gdev+lcblki], gcolsize, &dC[dev][gbblki], lddc);
}// end blki
}
for( magma_int_t k = 0; k < nbcmplx; ++k ) {
gngpu = gnode[k][MagmaMaxGPUs];
for( magma_int_t g = 0; g < gngpu; ++g ) {
gdev = gnode[k][g];
gcolsize = gpuisactive[gdev];
// if gcolsize>0, ==> gpu gdev was active and so
// I received from him/computed a portion of dwork,
// so go over its gblk and distribute it on dC.
if(gdev!=dev){
if(gcolsize>0){
devperm = (gdev-stdev+ngpu)%ngpu;
gblk = (nbblk/ngpu) + (nbblk%ngpu > devperm ? 1:0 );
magmablasSetKernelStream( queues[ dev ][ gdev ] );
if(k==cmplxid){
//we are on the same group so wait on event issued by gdev for me citing his id
magma_queue_wait_event(queues[ dev ][ gdev ], redevents[gdev][dev*ngpu+gdev]);
//printf (" GPU %d queue %d waiting on event %d to collecte from %d the size of gcolsize %d\n", dev, queues[ dev ][ gdev ], redevents[gdev][dev*ngpu+gdev], gdev, gcolsize);
}else{
//we are on different group so:
//if I am the master wait on the event issued by gdev for me citing his id
//else wait event issued by my master for me on the behalf of gdev
//printf (" GPU %d queue %d waiting on event %d to collecte from %d the size of gcolsize %d\n", dev, queues[ dev ][ gdev ], redevents[masterdev][dev*ngpu+gdev], gdev, gcolsize);
if(dev==masterdev)
magma_queue_wait_event(queues[ dev ][ gdev ], redevents[gdev][dev*ngpu+gdev]);
else
magma_queue_wait_event(queues[ dev ][ gdev ], redevents[masterdev][dev*ngpu+gdev]);
}
//printf (" GPU %d stream %d doing slacpy\n", dev, queues[ dev ][ gdev ]);
for( magma_int_t blki = 0; blki < gblk; ++blki){
gbblki = (blki*ngpu + devperm)*nb - blockoffset;
lcblki = blki*nb;
ib = nb;//min(nb, m-gbblki);
if(gdev==stdev){
lcblki = blki*nb-blockoffset;
if(blki==0){
gbblki = 0;
lcblki = 0;
ib = nb-blockoffset;
}
}
ib = min(ib, m-gbblki);
//printf(" blockoffset %d nbblk %d stdev %d receiving from gdev %d gblk %d gcolsize %d copying blki %d of size ibxn %dx%d from work[%d] to C[%d]\n", blockoffset, nbblk, stdev, gdev, gblk, gcolsize, blki, ib, n, lcblki, gbblki);
magmablas_slacpy( MagmaFull, ib, n, &dwork[dev][maxgsize*gdev+lcblki], gcolsize, &dC[dev][gbblki], lddc);
}// end blki
}// en gcolsize>0 meaning gdev is active
} // end if gdev != dev
}// end loop over the g gpus of the cmplx k
}//end loop over the real k
}// end mycolsize>0 meaning that I am active
}// end loop over idev of cmplxid
}// end loop of the cmplx
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_device_sync();
}
// put back the input gpu and its input stream
magma_setdevice( cdev );
magmablasSetKernelStream( cstream );
}
extern "C" void
magma_ssyr2k_mgpu(
magma_int_t num_gpus, magma_uplo_t uplo, magma_trans_t trans, magma_int_t nb, magma_int_t n, magma_int_t k,
float alpha,
float **db, magma_int_t lddb, magma_int_t offset_b,
float beta,
float **dc, magma_int_t lddc, magma_int_t offset,
magma_int_t num_streams, magma_queue_t stream[][10])
{
#define dB(id, i, j) (db[(id)]+(j)*lddb + (i)+offset_b)
#define dB1(id, i, j) (db[(id)]+(j)*lddb + (i)+offset_b)+k*lddb
#define dC(id, i, j) (dc[(id)]+(j)*lddc + (i))
magma_int_t i, id, ib, ii, kk, n1;
float c_one = MAGMA_S_ONE;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/* diagonal update */
for( i=0; i < n; i += nb ) {
id = ((i+offset)/nb)%num_gpus;
kk = (i/(nb*num_gpus))%num_streams;
magma_setdevice(id);
magmablasSetKernelStream(stream[id][kk]);
ib = min(nb, n-i);
ii = nb*((i+offset)/(nb*num_gpus));
/* ssyr2k on diagonal block */
trace_gpu_start( id, kk, "syr2k", "syr2k" );
magma_ssyr2k(uplo, trans, ib, k,
alpha, dB1(id, i, 0 ), lddb,
dB(id, i, 0 ), lddb,
beta, dC(id, i+offset, ii), lddc);
trace_gpu_end( id, kk );
}
/* off-diagonal update */
if (uplo == MagmaUpper) {
for( i=nb; i < n; i += nb ) {
id = ((i+offset)/nb)%num_gpus;
kk = (i/(nb*num_gpus))%num_streams;
magma_setdevice(id);
magmablasSetKernelStream(stream[id][kk]);
ib = min(nb, n-i);
ii = nb*((i+offset)/(nb*num_gpus));
magma_sgemm(MagmaNoTrans, MagmaConjTrans, i, ib, k,
alpha, dB1(id, 0, 0 ), lddb,
dB(id, i, 0 ), lddb,
c_one, dC(id, 0, ii), lddc);
}
}
else {
for( i=0; i < n-nb; i += nb ) {
id = ((i+offset)/nb)%num_gpus;
kk = (i/(nb*num_gpus))%num_streams;
magma_setdevice(id);
magmablasSetKernelStream(stream[id][kk]);
ib = min(nb, n-i);
ii = nb*((i+offset)/(nb*num_gpus));
n1 = n-i-ib;
// sgemm on off-diagonal blocks
trace_gpu_start( id, kk, "gemm_up", "gemm_up" );
magma_sgemm(MagmaNoTrans, MagmaConjTrans, n1, ib, k,
alpha, dB1(id, i+ib, 0 ), lddb,
dB(id, i, 0 ), lddb,
c_one, dC(id, i+offset+ib, ii), lddc);
trace_gpu_end( id, kk );
}
}
if (uplo == MagmaUpper) {
for( i=nb; i < n; i += nb ) {
id = ((i+offset)/nb)%num_gpus;
kk = (i/(nb*num_gpus))%num_streams;
magma_setdevice(id);
magmablasSetKernelStream(stream[id][kk]);
ib = min(nb, n-i);
ii = nb*((i+offset)/(nb*num_gpus));
magma_sgemm(MagmaNoTrans, MagmaConjTrans, i, ib, k,
alpha, dB( id, 0, 0 ), lddb,
dB1(id, i, 0 ), lddb,
c_one, dC(id, 0, ii), lddc);
}
} else {
for( i=0; i < n-nb; i += nb ) {
id = ((i+offset)/nb)%num_gpus;
kk = (i/(nb*num_gpus))%num_streams;
magma_setdevice(id);
magmablasSetKernelStream(stream[id][kk]);
ib = min(nb, n-i);
ii = nb*((i+offset)/(nb*num_gpus));
n1 = n-i-ib;
/* sgemm on off-diagonal blocks */
trace_gpu_start( id, kk, "gemm_up", "gemm_up" );
magma_sgemm(MagmaNoTrans, MagmaConjTrans, n1, ib, k,
alpha, dB(id, i+ib, 0 ), lddb,
dB1(id, i, 0 ), lddb,
c_one, dC(id, i+offset+ib, ii), lddc);
trace_gpu_end( id, kk );
}
}
for( id=0; id < num_gpus; id++ ) {
magma_setdevice(id);
for( kk=0; kk < num_streams; kk++ ) {
magma_queue_sync(stream[id][kk]);
}
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
}
/**
Purpose
-------
SGEBRD reduces a general real M-by-N matrix A to upper or lower
bidiagonal form B by an orthogonal transformation: Q**H * A * P = B.
If m >= n, B is upper bidiagonal; if m < n, B is lower bidiagonal.
Arguments
---------
@param[in]
m INTEGER
The number of rows in the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns in the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the M-by-N general matrix to be reduced.
On exit,
if m >= n, the diagonal and the first superdiagonal are
overwritten with the upper bidiagonal matrix B; the
elements below the diagonal, with the array TAUQ, represent
the orthogonal matrix Q as a product of elementary
reflectors, and the elements above the first superdiagonal,
with the array TAUP, represent the orthogonal matrix P as
a product of elementary reflectors;
\n
if m < n, the diagonal and the first subdiagonal are
overwritten with the lower bidiagonal matrix B; the
elements below the first subdiagonal, with the array TAUQ,
represent the orthogonal matrix Q as a product of
elementary reflectors, and the elements above the diagonal,
with the array TAUP, represent the orthogonal matrix P as
a product of elementary reflectors.
See Further Details.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
d real array, dimension (min(M,N))
The diagonal elements of the bidiagonal matrix B:
D(i) = A(i,i).
@param[out]
e real array, dimension (min(M,N)-1)
The off-diagonal elements of the bidiagonal matrix B:
if m >= n, E(i) = A(i,i+1) for i = 1,2,...,n-1;
if m < n, E(i) = A(i+1,i) for i = 1,2,...,m-1.
@param[out]
tauq REAL array dimension (min(M,N))
The scalar factors of the elementary reflectors which
represent the orthogonal matrix Q. See Further Details.
@param[out]
taup REAL array, dimension (min(M,N))
The scalar factors of the elementary reflectors which
represent the orthogonal matrix P. 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 length of the array WORK. LWORK >= (M+N)*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 by XERBLA.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value.
Further Details
---------------
The matrices Q and P are represented as products of elementary
reflectors:
If m >= n,
Q = H(1) H(2) . . . H(n) and P = G(1) G(2) . . . G(n-1)
Each H(i) and G(i) has the form:
H(i) = I - tauq * v * v' and G(i) = I - taup * u * u'
where tauq and taup are real scalars, and v and u are real vectors;
v(1:i-1) = 0, v(i) = 1, and v(i+1:m) is stored on exit in A(i+1:m,i);
u(1:i) = 0, u(i+1) = 1, and u(i+2:n) is stored on exit in A(i,i+2:n);
tauq is stored in TAUQ(i) and taup in TAUP(i).
If m < n,
Q = H(1) H(2) . . . H(m-1) and P = G(1) G(2) . . . G(m)
Each H(i) and G(i) has the form:
H(i) = I - tauq * v * v' and G(i) = I - taup * u * u'
where tauq and taup are real scalars, and v and u are real vectors;
v(1:i) = 0, v(i+1) = 1, and v(i+2:m) is stored on exit in A(i+2:m,i);
u(1:i-1) = 0, u(i) = 1, and u(i+1:n) is stored on exit in A(i,i+1:n);
tauq is stored in TAUQ(i) and taup in TAUP(i).
The contents of A on exit are illustrated by the following examples:
@verbatim
m = 6 and n = 5 (m > n): m = 5 and n = 6 (m < n):
( d e u1 u1 u1 ) ( d u1 u1 u1 u1 u1 )
( v1 d e u2 u2 ) ( e d u2 u2 u2 u2 )
( v1 v2 d e u3 ) ( v1 e d u3 u3 u3 )
( v1 v2 v3 d e ) ( v1 v2 e d u4 u4 )
( v1 v2 v3 v4 d ) ( v1 v2 v3 e d u5 )
( v1 v2 v3 v4 v5 )
@endverbatim
where d and e denote diagonal and off-diagonal elements of B, vi
denotes an element of the vector defining H(i), and ui an element of
the vector defining G(i).
@ingroup magma_sgesvd_comp
********************************************************************/
extern "C" magma_int_t
magma_sgebrd(
magma_int_t m, magma_int_t n,
float *A, magma_int_t lda, float *d, float *e,
float *tauq, float *taup,
float *work, magma_int_t lwork,
magma_int_t *info)
{
#define A(i, j) (A + (j)*lda + (i))
#define dA(i, j) (dA + (j)*ldda + (i))
float c_neg_one = MAGMA_S_NEG_ONE;
float c_one = MAGMA_S_ONE;
float *dA, *dwork;
magma_int_t ncol, nrow, jmax, nb, ldda;
magma_int_t i, j, nx;
magma_int_t iinfo;
magma_int_t minmn;
magma_int_t ldwrkx, ldwrky, lwkopt;
magma_int_t lquery;
nb = magma_get_sgebrd_nb(n);
ldda = m;
lwkopt = (m + n) * nb;
work[0] = MAGMA_S_MAKE( lwkopt, 0. );
lquery = (lwork == -1);
/* Check arguments */
*info = 0;
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
} else if (lwork < lwkopt && (! lquery) ) {
*info = -10;
}
if (*info < 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
/* Quick return if possible */
minmn = min(m,n);
if (minmn == 0) {
work[0] = c_one;
return *info;
}
if (MAGMA_SUCCESS != magma_smalloc( &dA, n*ldda + (m + n)*nb )) {
fprintf (stderr, "!!!! device memory allocation error in sgebrd\n" );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dwork = dA + n*ldda;
ldwrkx = m;
ldwrky = n;
/* Set the block/unblock crossover point NX. */
nx = 128;
/* Copy the matrix to the GPU */
if (minmn - nx >= 1) {
magma_ssetmatrix( m, n, A, lda, dA, ldda );
}
for (i=0; i < (minmn - nx); i += nb) {
/* Reduce rows and columns i:i+nb-1 to bidiagonal form and return
the matrices X and Y which are needed to update the unreduced
part of the matrix */
nrow = m - i;
ncol = n - i;
/* Get the current panel (no need for the 1st iteration) */
if ( i > 0 ) {
magma_sgetmatrix( nrow, nb, dA(i, i), ldda, A( i, i), lda );
magma_sgetmatrix( nb, ncol - nb,
dA(i, i+nb), ldda,
A( i, i+nb), lda );
}
magma_slabrd_gpu(nrow, ncol, nb,
A(i, i), lda, dA(i, i), ldda,
d+i, e+i, tauq+i, taup+i,
work, ldwrkx, dwork, ldwrkx, // x, dx
work+(ldwrkx*nb), ldwrky, dwork+(ldwrkx*nb), ldwrky); // y, dy
/* Update the trailing submatrix A(i+nb:m,i+nb:n), using an update
of the form A := A - V*Y' - X*U' */
nrow = m - i - nb;
ncol = n - i - nb;
// Send Y back to the GPU
magma_ssetmatrix( nrow, nb, work + nb, ldwrkx, dwork + nb, ldwrkx );
magma_ssetmatrix( ncol, nb,
work + (ldwrkx+1)*nb, ldwrky,
dwork + (ldwrkx+1)*nb, ldwrky );
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
nrow, ncol, nb,
c_neg_one, dA(i+nb, i ), ldda,
dwork+(ldwrkx+1)*nb, ldwrky,
c_one, dA(i+nb, i+nb), ldda);
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
nrow, ncol, nb,
c_neg_one, dwork+nb, ldwrkx,
dA( i, i+nb ), ldda,
c_one, dA( i+nb, i+nb ), ldda);
/* Copy diagonal and off-diagonal elements of B back into A */
if (m >= n) {
jmax = i + nb;
for (j = i; j < jmax; ++j) {
*A(j, j ) = MAGMA_S_MAKE( d[j], 0. );
*A(j, j+1) = MAGMA_S_MAKE( e[j], 0. );
}
} else {
jmax = i + nb;
for (j = i; j < jmax; ++j) {
*A(j, j ) = MAGMA_S_MAKE( d[j], 0. );
*A(j+1, j ) = MAGMA_S_MAKE( e[j], 0. );
}
}
}
/* Use unblocked code to reduce the remainder of the matrix */
nrow = m - i;
ncol = n - i;
if ( 0 < minmn - nx ) {
magma_sgetmatrix( nrow, ncol, dA(i, i), ldda, A(i, i), lda );
}
lapackf77_sgebrd( &nrow, &ncol,
A(i, i), &lda, d+i, e+i,
tauq+i, taup+i, work, &lwork, &iinfo);
work[0] = MAGMA_S_MAKE( lwkopt, 0. );
magma_free( dA );
return *info;
} /* magma_sgebrd */
/***************************************************************************//**
Purpose
-------
SGERFS improves the computed solution to a system of linear equations.
The iterative refinement process is stopped if
ITER > ITERMAX
or for all the RHS we have:
RNRM < SQRT(n)*XNRM*ANRM*EPS*BWDMAX
where
o ITER is the number of the current iteration in the iterative
refinement process
o RNRM is the infinity-norm of the residual
o XNRM is the infinity-norm of the solution
o ANRM is the infinity-operator-norm of the matrix A
o EPS is the machine epsilon returned by SLAMCH('Epsilon')
The value ITERMAX and BWDMAX are fixed to 30 and 1.0D+00 respectively.
Arguments
---------
@param[in]
trans magma_trans_t
Specifies the form of the system of equations:
- = MagmaNoTrans: A * X = B (No transpose)
- = MagmaTrans: A**T * X = B (Transpose)
- = MagmaConjTrans: A**H * X = B (Conjugate transpose)
@param[in]
n INTEGER
The number of linear equations, i.e., the order of the
matrix A. N >= 0.
@param[in]
nrhs INTEGER
The number of right hand sides, i.e., the number of columns
of the matrix B. NRHS >= 0.
@param[in]
dA REAL array on the GPU, dimension (ldda,N)
the N-by-N coefficient matrix A.
@param[in]
ldda INTEGER
The leading dimension of the array dA. ldda >= max(1,N).
@param[in]
dB REAL array on the GPU, dimension (lddb,NRHS)
The N-by-NRHS right hand side matrix B.
@param[in]
lddb INTEGER
The leading dimension of the array dB. lddb >= max(1,N).
@param[in, out]
dX REAL array on the GPU, dimension (lddx,NRHS)
On entry, the solution matrix X, as computed by
SGETRS_NOPIV. On exit, the improved solution matrix X.
@param[in]
lddx INTEGER
The leading dimension of the array dX. lddx >= max(1,N).
@param
dworkd (workspace) REAL array on the GPU, dimension (N*NRHS)
This array is used to hold the residual vectors.
@param
dAF REAL array on the GPU, dimension (ldda,n)
The factors L and U from the factorization A = L*U
as computed by SGETRF_NOPIV.
@param[out]
iter INTEGER
- < 0: iterative refinement has failed, real
factorization has been performed
+ -1 : the routine fell back to full precision for
implementation- or machine-specific reasons
+ -2 : narrowing the precision induced an overflow,
the routine fell back to full precision
+ -3 : failure of SGETRF
+ -31: stop the iterative refinement after the 30th iteration
- > 0: iterative refinement has been successfully used.
Returns the number of iterations
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if info = -i, the i-th argument had an illegal value
- > 0: if info = i, U(i,i) computed in REAL is
exactly zero. The factorization has been completed,
but the factor U is exactly singular, so the solution
could not be computed.
@ingroup magma_gerfs_nopiv
*******************************************************************************/
extern "C" magma_int_t
magma_sgerfs_nopiv_gpu(
magma_trans_t trans, magma_int_t n, magma_int_t nrhs,
magmaFloat_ptr dA, magma_int_t ldda,
magmaFloat_ptr dB, magma_int_t lddb,
magmaFloat_ptr dX, magma_int_t lddx,
magmaFloat_ptr dworkd, magmaFloat_ptr dAF,
magma_int_t *iter,
magma_int_t *info)
{
#define dB(i,j) (dB + (i) + (j)*lddb)
#define dX(i,j) (dX + (i) + (j)*lddx)
#define dR(i,j) (dR + (i) + (j)*lddr)
/* Constants */
const float c_neg_one = MAGMA_S_NEG_ONE;
const float c_one = MAGMA_S_ONE;
const magma_int_t ione = 1;
/* Local variables */
magmaFloat_ptr dR;
float Xnrmv, Rnrmv;
float Anrm, Xnrm, Rnrm, cte, eps;
magma_int_t i, j, iiter, lddsa, lddr;
/* Check arguments */
*iter = 0;
*info = 0;
if ( n < 0 )
*info = -1;
else if ( nrhs < 0 )
*info = -2;
else if ( ldda < max(1,n))
*info = -4;
else if ( lddb < max(1,n))
*info = -8;
else if ( lddx < max(1,n))
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if ( n == 0 || nrhs == 0 )
return *info;
magma_queue_t queue = NULL;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
lddsa = n;
lddr = n;
dR = dworkd;
eps = lapackf77_slamch("Epsilon");
Anrm = magmablas_slange( MagmaInfNorm, n, n, dA, ldda, (magmaFloat_ptr)dworkd, n*nrhs, queue );
cte = Anrm * eps * magma_ssqrt( (float) n ) * BWDMAX;
// residual dR = dB - dA*dX in real
magmablas_slacpy( MagmaFull, n, nrhs, dB, lddb, dR, lddr, queue );
if ( nrhs == 1 ) {
magma_sgemv( trans, n, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1, queue );
}
else {
magma_sgemm( trans, MagmaNoTrans, n, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr, queue );
}
// TODO: use MAGMA_S_ABS( dX(i,j) ) instead of slange?
for( j=0; j < nrhs; j++ ) {
i = magma_isamax( n, dX(0,j), 1, queue ) - 1;
magma_sgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1, queue );
Xnrm = lapackf77_slange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_isamax( n, dR(0,j), 1, queue ) - 1;
magma_sgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1, queue );
Rnrm = lapackf77_slange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
//printf("Rnrm : %e, Xnrm*cte : %e\n", Rnrm, Xnrm*cte);
if ( Rnrm > Xnrm*cte ) {
goto refinement;
}
}
*iter = 0;
goto cleanup;
refinement:
for( iiter=1; iiter < ITERMAX; ) {
*info = 0;
// solve dAF*dX = dR
// it's okay that dR is used for both dB input and dX output.
magma_sgetrs_nopiv_gpu( trans, n, nrhs, dAF, lddsa, dR, lddr, info );
if (*info != 0) {
*iter = -3;
goto fallback;
}
// Add correction and setup residual
// dX += dR --and--
// dR = dB
// This saves going through dR a second time (if done with one more kernel).
// -- not really: first time is read, second time is write.
for( j=0; j < nrhs; j++ ) {
magmablas_saxpycp( n, dR(0,j), dX(0,j), dB(0,j), queue );
}
// residual dR = dB - dA*dX in real
if ( nrhs == 1 ) {
magma_sgemv( trans, n, n,
c_neg_one, dA, ldda,
dX, 1,
c_one, dR, 1, queue );
}
else {
magma_sgemm( trans, MagmaNoTrans, n, nrhs, n,
c_neg_one, dA, ldda,
dX, lddx,
c_one, dR, lddr, queue );
}
/* Check whether the nrhs normwise backward errors satisfy the
* stopping criterion. If yes, set ITER=IITER > 0 and return. */
for( j=0; j < nrhs; j++ ) {
i = magma_isamax( n, dX(0,j), 1, queue ) - 1;
magma_sgetmatrix( 1, 1, dX(i,j), 1, &Xnrmv, 1, queue );
Xnrm = lapackf77_slange( "F", &ione, &ione, &Xnrmv, &ione, NULL );
i = magma_isamax( n, dR(0,j), 1, queue ) - 1;
magma_sgetmatrix( 1, 1, dR(i,j), 1, &Rnrmv, 1, queue );
Rnrm = lapackf77_slange( "F", &ione, &ione, &Rnrmv, &ione, NULL );
if ( Rnrm > Xnrm*cte ) {
goto L20;
}
}
/* If we are here, the nrhs normwise backward errors satisfy
* the stopping criterion, we are good to exit. */
*iter = iiter;
goto cleanup;
L20:
iiter++;
}
/* If we are at this place of the code, this is because we have
* performed ITER=ITERMAX iterations and never satisified the
* stopping criterion. Set up the ITER flag accordingly. */
*iter = -ITERMAX - 1;
fallback:
/* Iterative refinement failed to converge to a
* satisfactory solution. */
cleanup:
magma_queue_destroy( queue );
return *info;
}
/**
Purpose
-------
SLAUUM computes the product U * U' or L' * L, where the triangular
factor U or L is stored in the upper or lower triangular part of
the array A.
If UPLO = MagmaUpper then the upper triangle of the result is stored,
overwriting the factor U in A.
If UPLO = MagmaLower then the lower triangle of the result is stored,
overwriting the factor L in A.
This is the blocked form of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
uplo magma_uplo_t
Specifies whether the triangular factor stored in the array A
is upper or lower triangular:
- = MagmaUpper: Upper triangular
- = MagmaLower: Lower triangular
@param[in]
n INTEGER
The order of the triangular factor U or L. N >= 0.
@param[in,out]
A COPLEX_16 array, dimension (LDA,N)
On entry, the triangular factor U or L.
On exit, if UPLO = MagmaUpper, the upper triangle of A is
overwritten with the upper triangle of the product U * U';
if UPLO = MagmaLower, the lower triangle of A is overwritten with
the lower triangle of the product L' * L.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -k, the k-th argument had an illegal value
@ingroup magma_sposv_aux
***************************************************************************/
extern "C" magma_int_t
magma_slauum(
magma_uplo_t uplo, magma_int_t n,
float *A, magma_int_t lda,
magma_int_t *info)
{
#define A(i, j) (A + (j)*lda + (i))
#define dA(i, j) (dA + (j)*ldda + (i))
/* Local variables */
const char* uplo_ = lapack_uplo_const( uplo );
magma_int_t ldda, nb;
magma_int_t i, ib;
float c_one = MAGMA_S_ONE;
float d_one = MAGMA_D_ONE;
float *dA;
int upper = (uplo == MagmaUpper);
*info = 0;
if (! upper && uplo != MagmaLower)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,n))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return */
if ( n == 0 )
return *info;
ldda = ((n+31)/32)*32;
if (MAGMA_SUCCESS != magma_smalloc( &dA, (n)*ldda )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
nb = magma_get_spotrf_nb(n);
if (nb <= 1 || nb >= n)
lapackf77_slauum(uplo_, &n, A, &lda, info);
else {
if (upper) {
/* Compute the product U * U'. */
for (i=0; i < n; i += nb) {
ib=min(nb,n-i);
magma_ssetmatrix_async( ib, ib,
A(i,i), lda,
dA(i, i), ldda, stream[1] );
magma_ssetmatrix_async( ib, (n-i-ib),
A(i,i+ib), lda,
dA(i,i+ib), ldda, stream[0] );
magma_queue_sync( stream[1] );
magma_strmm( MagmaRight, MagmaUpper,
MagmaConjTrans, MagmaNonUnit, i, ib,
c_one, dA(i,i), ldda, dA(0, i),ldda);
lapackf77_slauum(MagmaUpperStr, &ib, A(i,i), &lda, info);
magma_ssetmatrix_async( ib, ib,
A(i, i), lda,
dA(i, i), ldda, stream[0] );
if (i+ib < n) {
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
i, ib, (n-i-ib), c_one, dA(0,i+ib),
ldda, dA(i, i+ib),ldda, c_one,
dA(0,i), ldda);
magma_queue_sync( stream[0] );
magma_ssyrk( MagmaUpper, MagmaNoTrans, ib,(n-i-ib),
d_one, dA(i, i+ib), ldda,
d_one, dA(i, i), ldda);
}
magma_sgetmatrix( i+ib, ib,
dA(0, i), ldda,
A(0, i), lda );
}
}
else {
/* Compute the product L' * L. */
for (i=0; i < n; i += nb) {
ib=min(nb,n-i);
magma_ssetmatrix_async( ib, ib,
A(i,i), lda,
dA(i, i), ldda, stream[1] );
magma_ssetmatrix_async( (n-i-ib), ib,
A(i+ib, i), lda,
dA(i+ib, i), ldda, stream[0] );
magma_queue_sync( stream[1] );
magma_strmm( MagmaLeft, MagmaLower,
MagmaConjTrans, MagmaNonUnit, ib,
i, c_one, dA(i,i), ldda,
dA(i, 0),ldda);
lapackf77_slauum(MagmaLowerStr, &ib, A(i,i), &lda, info);
magma_ssetmatrix_async( ib, ib,
A(i, i), lda,
dA(i, i), ldda, stream[0] );
if (i+ib < n) {
magma_sgemm(MagmaConjTrans, MagmaNoTrans,
ib, i, (n-i-ib), c_one, dA( i+ib,i),
ldda, dA(i+ib, 0),ldda, c_one,
dA(i,0), ldda);
magma_queue_sync( stream[0] );
magma_ssyrk(MagmaLower, MagmaConjTrans, ib, (n-i-ib),
d_one, dA(i+ib, i), ldda,
d_one, dA(i, i), ldda);
}
magma_sgetmatrix( ib, i+ib,
dA(i, 0), ldda,
A(i, 0), lda );
}
}
}
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_free( dA );
return *info;
}
/**
@deprecated
Purpose
-------
SLAQPS computes a step of QR factorization with column pivoting
of a real M-by-N matrix A by using Blas-3. It tries to factorize
NB columns from A starting from the row OFFSET+1, and updates all
of the matrix with Blas-3 xGEMM.
In some cases, due to catastrophic cancellations, it cannot
factorize NB columns. Hence, the actual number of factorized
columns is returned in KB.
Block A(1:OFFSET,1:N) is accordingly pivoted, but not factorized.
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]
offset INTEGER
The number of rows of A that have been factorized in
previous steps.
@param[in]
nb INTEGER
The number of columns to factorize.
@param[out]
kb INTEGER
The number of columns actually factorized.
@param[in,out]
dA REAL array, dimension (LDDA,N), on the GPU.
On entry, the M-by-N matrix A.
On exit, block A(OFFSET+1:M,1:KB) is the triangular
factor obtained and block A(1:OFFSET,1:N) has been
accordingly pivoted, but no factorized.
The rest of the matrix, block A(OFFSET+1:M,KB+1:N) has
been updated.
@param[in]
ldda INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
@param[in,out]
jpvt INTEGER array, dimension (N)
JPVT(I) = K <==> Column K of the full matrix A has been
permuted into position I in AP.
@param[out]
tau REAL array, dimension (KB)
The scalar factors of the elementary reflectors.
@param[in,out]
vn1 REAL array, dimension (N)
The vector with the partial column norms.
@param[in,out]
vn2 REAL array, dimension (N)
The vector with the exact column norms.
@param[in,out]
dauxv REAL array, dimension (NB), on the GPU
Auxiliary vector.
@param[in,out]
dF REAL array, dimension (LDDF,NB), on the GPU
Matrix F' = L*Y'*A.
@param[in]
lddf INTEGER
The leading dimension of the array F. LDDF >= max(1,N).
@ingroup magma_sgeqp3_aux
********************************************************************/
extern "C" magma_int_t
magma_slaqps_gpu(
magma_int_t m, magma_int_t n, magma_int_t offset,
magma_int_t nb, magma_int_t *kb,
magmaFloat_ptr dA, magma_int_t ldda,
magma_int_t *jpvt, float *tau,
float *vn1, float *vn2,
magmaFloat_ptr dauxv,
magmaFloat_ptr dF, magma_int_t lddf)
{
#define dA(i, j) (dA + (i) + (j)*(ldda))
#define dF(i, j) (dF + (i) + (j)*(lddf))
float c_zero = MAGMA_S_MAKE( 0.,0.);
float c_one = MAGMA_S_MAKE( 1.,0.);
float c_neg_one = MAGMA_S_MAKE(-1.,0.);
magma_int_t ione = 1;
magma_int_t i__1, i__2;
float z__1;
magma_int_t k, rk;
magmaFloat_ptr dAks;
float tauk = MAGMA_S_ZERO;
magma_int_t pvt;
float tol3z;
magma_int_t itemp;
float lsticc;
magmaFloat_ptr dlsticcs;
magma_smalloc( &dlsticcs, 1+256*(n+255)/256 );
tol3z = magma_ssqrt( lapackf77_slamch("Epsilon"));
lsticc = 0;
k = 0;
magma_smalloc( &dAks, nb );
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
while( k < nb && lsticc == 0 ) {
rk = offset + k;
/* Determine ith pivot column and swap if necessary */
// subtract 1 from Fortran/CUBLAS isamax; pvt, k are 0-based.
pvt = k + magma_isamax( n-k, &vn1[k], ione, queue ) - 1;
if (pvt != k) {
/* F gets swapped so F must be sent at the end to GPU */
i__1 = k;
magmablas_sswap( m, dA(0, pvt), ione, dA(0, k), ione, queue );
magmablas_sswap( i__1, dF(pvt, 0), lddf, dF(k, 0), lddf, queue );
itemp = jpvt[pvt];
jpvt[pvt] = jpvt[k];
jpvt[k] = itemp;
magma_sswap( 2, &vn1[pvt], n+offset, &vn1[k], n+offset, queue );
}
/* Apply previous Householder reflectors to column K:
A(RK:M,K) := A(RK:M,K) - A(RK:M,1:K-1)*F(K,1:K-1)'.
Optimization: multiply with beta=0; wait for vector and subtract */
if (k > 0) {
//#define RIGHT_UPDATE
#ifdef RIGHT_UPDATE
i__1 = m - offset - nb;
i__2 = k;
magma_sgemv( MagmaNoTrans, i__1, i__2,
c_neg_one, A(offset+nb, 0), lda,
F(k, 0), ldf,
c_one, A(offset+nb, k), ione, queue );
#else
i__1 = m - rk;
i__2 = k;
magma_sgemv( MagmaNoTrans, i__1, i__2,
c_neg_one, dA(rk, 0), ldda,
dF(k, 0), lddf,
c_one, dA(rk, k), ione, queue );
#endif
}
/* Generate elementary reflector H(k). */
magma_slarfg_gpu( m-rk, dA(rk, k), dA(rk + 1, k), &tau[k], &vn1[k], &dAks[k], queue );
/* needed to avoid the race condition */
if (k == 0) magma_ssetvector( 1, &c_one, 1, dA(rk, k), 1, queue );
else magma_scopymatrix( 1, 1, dA(offset, 0), 1, dA(rk, k), 1, queue );
/* Compute Kth column of F:
Compute F(K+1:N,K) := tau(K)*A(RK:M,K+1:N)'*A(RK:M,K) on the GPU */
if (k < n-1 || k > 0) magma_sgetvector( 1, &tau[k], 1, &tauk, 1, queue );
if (k < n-1) {
i__1 = m - rk;
i__2 = n - k - 1;
/* Multiply on GPU */
magma_sgemv( MagmaConjTrans, m-rk, n-k-1,
tauk, dA( rk, k+1 ), ldda,
dA( rk, k ), 1,
c_zero, dF( k+1, k ), 1, queue );
}
/* Incremental updating of F:
F(1:N,K) := F(1:N,K) - tau(K)*F(1:N,1:K-1)*A(RK:M,1:K-1)'*A(RK:M,K).
F(1:N,K) := tau(K)*A(RK:M,K+1:N)'*A(RK:M,K) - tau(K)*F(1:N,1:K-1)*A(RK:M,1:K-1)'*A(RK:M,K)
:= tau(K)(A(RK:M,K+1:N)' - F(1:N,1:K-1)*A(RK:M,1:K-1)') A(RK:M,K)
so, F is (updated A)*V */
if (k > 0) {
z__1 = MAGMA_S_NEGATE( tauk );
#ifdef RIGHT_UPDATE
i__1 = m - offset - nb;
i__2 = k;
magma_sgemv( MagmaConjTrans, i__1, i__2,
z__1, dA(offset+nb, 0), lda,
dA(offset+nb, k), ione,
c_zero, dauxv, ione, queue );
i__1 = k;
magma_sgemv( MagmaNoTrans, n-k-1, i__1,
c_one, F(k+1,0), ldf,
dauxv, ione,
c_one, F(k+1,k), ione, queue );
#else
i__1 = m - rk;
i__2 = k;
magma_sgemv( MagmaConjTrans, i__1, i__2,
z__1, dA(rk, 0), ldda,
dA(rk, k), ione,
c_zero, dauxv, ione, queue );
/* I think we only need stricly lower-triangular part :) */
magma_sgemv( MagmaNoTrans, n-k-1, i__2,
c_one, dF(k+1,0), lddf,
dauxv, ione,
c_one, dF(k+1,k), ione, queue );
#endif
}
/* Optimization: On the last iteration start sending F back to the GPU */
/* Update the current row of A:
A(RK,K+1:N) := A(RK,K+1:N) - A(RK,1:K)*F(K+1:N,1:K)'. */
if (k < n-1) {
i__1 = n - k - 1;
i__2 = k + 1;
#ifdef RIGHT_UPDATE
/* right-looking update of rows, */
magma_sgemm( MagmaNoTrans, MagmaConjTrans, nb-k, i__1, ione,
c_neg_one, dA(rk, k ), ldda,
dF(k+1, k ), lddf,
c_one, dA(rk, k+1), ldda, queue );
#else
/* left-looking update of rows, *
* since F=A'v with original A, so no right-looking */
magma_sgemm( MagmaNoTrans, MagmaConjTrans, ione, i__1, i__2,
c_neg_one, dA(rk, 0 ), ldda,
dF(k+1,0 ), lddf,
c_one, dA(rk, k+1), ldda, queue );
#endif
}
/* Update partial column norms. */
if (rk < min(m, n+offset)-1 ) {
magmablas_snrm2_row_check_adjust( n-k-1, tol3z, &vn1[k+1], &vn2[k+1],
dA(rk,k+1), ldda, dlsticcs, queue );
//magma_device_sync();
magma_sgetvector( 1, &dlsticcs[0], 1, &lsticc, 1, queue );
}
++k;
}
magma_scopymatrix( 1, k, dAks, 1, dA(offset, 0), ldda+1, queue );
// leave k as the last column done
--k;
*kb = k + 1;
rk = offset + *kb - 1;
/* Apply the block reflector to the rest of the matrix:
A(OFFSET+KB+1:M,KB+1:N) := A(OFFSET+KB+1:M,KB+1:N) - A(OFFSET+KB+1:M,1:KB)*F(KB+1:N,1:KB)' */
if (*kb < min(n, m - offset)) {
i__1 = m - rk - 1;
i__2 = n - *kb;
magma_sgemm( MagmaNoTrans, MagmaConjTrans, i__1, i__2, *kb,
c_neg_one, dA(rk+1, 0 ), ldda,
dF(*kb, 0 ), lddf,
c_one, dA(rk+1, *kb), ldda, queue );
}
/* Recomputation of difficult columns. */
if ( lsticc > 0 ) {
// printf( " -- recompute dnorms --\n" );
magmablas_snrm2_check( m-rk-1, n-*kb, dA(rk+1,*kb), ldda,
&vn1[*kb], dlsticcs, queue );
magma_scopymatrix( n-*kb, 1, &vn1[*kb], *kb, &vn2[*kb], *kb, queue );
}
magma_free( dAks );
magma_free( dlsticcs );
magma_queue_destroy( queue );
return MAGMA_SUCCESS;
} /* magma_slaqps */
/**
Purpose
-------
SLARFB applies a real block reflector H or its transpose H' to a
REAL m by n matrix C, from the left.
__Note that this function assumes__ that the upper part of dV is 0
because it is referenced. Same for upper/lower part of dT.
Arguments
---------
@param[in]
side magma_side_t
- = MagmaLeft: apply H or H' from the Left
- = MagmaRight: apply H or H' from the Right
@param[in]
trans magma_trans_t
- = MagmaNoTrans: apply H (No transpose)
- = MagmaTrans: apply H' (Conjugate transpose)
@param[in]
direct magma_direct_t
Indicates how H is formed from a product of elementary
reflectors
- = MagmaForward: H = H(1) H(2) . . . H(k) (Forward)
- = MagmaBackward: H = H(k) . . . H(2) H(1) (Backward)
@param[in]
storev magma_storev_t
Indicates how the vectors which define the elementary
reflectors are stored:
- = MagmaColumnwise: Columnwise
- = MagmaRowwise: Rowwise
@param[in]
m INTEGER
The number of rows of the matrix C.
@param[in]
n INTEGER
The number of columns of the matrix C.
@param[in]
k INTEGER
The order of the matrix T (= the number of elementary
reflectors whose product defines the block reflector).
@param[in]
dV REAL array on the GPU, dimension
(LDV,K) if STOREV = MagmaColumnwise
(LDV,M) if STOREV = MagmaRowwise and SIDE = MagmaLeft
(LDV,N) if STOREV = MagmaRowwise and SIDE = MagmaRight
The matrix V. See further details.
@param[in]
ldv INTEGER
The leading dimension of the array V.
If STOREV = MagmaColumnwise and SIDE = MagmaLeft, LDV >= max(1,M);
if STOREV = MagmaColumnwise and SIDE = MagmaRight, LDV >= max(1,N);
if STOREV = MagmaRowwise, LDV >= K.
@param[in]
dT REAL array on the GPU, dimension (LDT,K)
The triangular k by k matrix T in the representation of the
block reflector.
@param[in]
ldt INTEGER
The leading dimension of the array T. LDT >= K.
@param[in,out]
dC REAL array on the GPU, dimension (LDC,N)
On entry, the m by n matrix C.
On exit, C is overwritten by H*C, or H'*C, or C*H, or C*H'.
@param[in]
ldc INTEGER
The leading dimension of the array C. LDA >= max(1,M).
@param
dwork (workspace) REAL array, dimension (LDWORK,K)
@param[in]
ldwork INTEGER
The leading dimension of the array WORK.
If SIDE = MagmaLeft, LDWORK >= max(1,N);
if SIDE = MagmaRight, LDWORK >= max(1,M);
@param
dworkvt (workspace) REAL array, dimension (LDWORKT,K)
@param[in]
ldworkvt INTEGER
The leading dimension of the array WORKVT.
LDWORKVT >= max(1,min(M,N));
Further Details
---------------
The shape of the matrix V and the storage of the vectors which define
the H(i) is best illustrated by the following example with n = 5 and
k = 3.
All elements including 0's and 1's are stored, unlike LAPACK.
DIRECT = MagmaForward and DIRECT = MagmaForward and
STOREV = MagmaColumnwise: STOREV = MagmaRowwise:
V = ( 1 0 0 ) V = ( 1 v1 v1 v1 v1 )
( v1 1 0 ) ( 0 1 v2 v2 v2 )
( v1 v2 1 ) ( 0 0 1 v3 v3 )
( v1 v2 v3 )
( v1 v2 v3 )
DIRECT = MagmaBackward and DIRECT = MagmaBackward and
STOREV = MagmaColumnwise: STOREV = MagmaRowwise:
V = ( v1 v2 v3 ) V = ( v1 v1 1 0 0 )
( v1 v2 v3 ) ( v2 v2 v2 1 0 )
( 1 v2 v3 ) ( v3 v3 v3 v3 1 )
( 0 1 v3 )
( 0 0 1 )
@ingroup magma_saux3
********************************************************************/
extern "C" magma_int_t
magma_slarfb_gpu_gemm( magma_side_t side, magma_trans_t trans, magma_direct_t direct, magma_storev_t storev,
magma_int_t m, magma_int_t n, magma_int_t k,
const float *dV, magma_int_t ldv,
const float *dT, magma_int_t ldt,
float *dC, magma_int_t ldc,
float *dwork, magma_int_t ldwork,
float *dworkvt, magma_int_t ldworkvt)
{
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
/* Function Body */
if (m <= 0 || n <= 0) {
return MAGMA_SUCCESS;
}
//internal variable
magma_int_t ldwvt = m > n ? k : m;
magma_int_t ldw;
if ( side == MagmaLeft ) {
ldw = k;
} else {
ldw = m;
}
// opposite of trans
magma_trans_t transt;
if (trans == MagmaNoTrans)
transt = MagmaTrans;
else
transt = MagmaNoTrans;
// whether T is upper or lower triangular
magma_uplo_t uplo;
if (direct == MagmaForward)
uplo = MagmaUpper;
else
uplo = MagmaLower;
// whether V is stored transposed or not
magma_trans_t notransV, transV;
if (storev == MagmaColumnwise) {
notransV = MagmaNoTrans;
transV = MagmaTrans;
}
else {
notransV = MagmaTrans;
transV = MagmaNoTrans;
}
if ( side == MagmaLeft ) {
// Form H C or H' C
// Comments assume H C.
// When forming H' C, T gets transposed via transt for m >= n or by trans for m < n.
// W = V' C
magma_sgemm( MagmaTrans, notransV,
k, n, m,
c_one, dV, ldv,
dC, ldc,
c_zero, dwork, ldw);
if (m <= n) {
// W2 = V T
magma_sgemm( notransV, trans,
m, k, k,
c_one, dV, ldv,
dT, ldt,
c_zero, dworkvt, ldwvt);
// C = C - W2 W = C - V T V' C = (I - V T V') C = H C
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m, n, k,
c_neg_one, dworkvt, ldwvt,
dwork, ldw,
c_one, dC, ldc);
} else {
// W2 = T W = T V' C
magma_sgemm( trans, MagmaNoTrans,
k, n, k,
c_one, dT, ldt,
dwork, ldw,
c_zero, dworkvt, ldwvt);
// C = C - V W2 = C - V T V' C = (I - V T V') C = H C
magma_sgemm( notransV, MagmaNoTrans,
m, n, k,
c_neg_one, dV, ldv,
dworkvt, ldwvt,
c_one, dC, ldc);
}
}
else {
// Form C H or C H'
// Comments assume C H.
// When forming C H', T gets transposed via trans.
// W = C V
magma_sgemm( MagmaNoTrans, notransV,
m, k, n,
c_one, dC, ldc,
dV, ldv,
c_zero, dwork, ldw);
if (m <= n) {
// W2 = W T = C V T
magma_sgemm( MagmaNoTrans, trans,
m, k, k,
c_one, dwork, ldw,
dT, ldt,
c_zero, dworkvt, ldwvt);
// C = C - W2 V' = C - C V T V' = C (I - V T V') = C H
magma_sgemm( MagmaNoTrans, transV,
m, n, k,
c_neg_one, dworkvt, ldwvt,
dV, ldv,
c_one, dC, ldc);
} else {
// W2 = T V'
magma_sgemm( trans, transV,
k, n, k,
c_one, dT, ldt,
dV, ldv,
c_zero, dworkvt, ldwvt);
// C = C - W W2 = C - C V T V' = C (I - V T V') = C H
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m, n, k,
c_neg_one, dwork, ldw,
dworkvt, ldwvt,
c_one, dC, ldc);
}
}
return MAGMA_SUCCESS;
} /* magma_slarfb */
/**
Purpose
-------
SPOTRF_OOC computes the Cholesky factorization of a real symmetric
positive definite matrix A. This version does not require work
space on the GPU passed as input. GPU memory is allocated in the
routine. The matrix A may exceed the GPU memory.
The factorization has the form
A = U**H * U, if UPLO = MagmaUpper, or
A = L * L**H, if UPLO = MagmaLower,
where U is an upper triangular matrix and L is lower triangular.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
num_gpus INTEGER
The number of GPUs. num_gpus > 0.
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangle of A is stored;
- = MagmaLower: Lower triangle of A is stored.
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in,out]
A REAL array, dimension (LDA,N)
On entry, the symmetric matrix A. If UPLO = MagmaUpper, the leading
N-by-N upper triangular part of A contains the upper
triangular part of the matrix A, and the strictly lower
triangular part of A is not referenced. If UPLO = MagmaLower, the
leading N-by-N lower triangular part of A contains the lower
triangular part of the matrix A, and the strictly upper
triangular part of A is not referenced.
\n
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization A = U**H * U or A = L * L**H.
\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,N).
@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, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
@ingroup magma_sposv_comp
********************************************************************/
extern "C" magma_int_t
magma_spotrf_m(magma_int_t num_gpus, magma_uplo_t uplo, magma_int_t n,
float *A, magma_int_t lda, magma_int_t *info)
{
#define A(i, j) ( A + (j)*lda + (i))
#define dA(d, i, j) (dwork[(d)] + (j)*lddla + (i))
#define dT(d, i, j) ( dt[(d)] + (j)*ldda + (i))
#define dAup(d, i, j) (dwork[(d)] + (j)*NB + (i))
#define dTup(d, i, j) ( dt[(d)] + (j)*nb + (i))
/* Local variables */
float d_one = 1.0;
float d_neg_one = -1.0;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
const char* uplo_ = lapack_uplo_const( uplo );
int upper = (uplo == MagmaUpper);
float *dwork[MagmaMaxGPUs], *dt[MagmaMaxGPUs];
magma_int_t ldda, lddla, nb, iinfo, n_local[MagmaMaxGPUs], J2, d, num_gpus0 = num_gpus;
magma_int_t j, jj, jb, J, JB, NB, MB, h;
magma_queue_t stream[MagmaMaxGPUs][3];
magma_event_t event[MagmaMaxGPUs][5];
magma_timer_t time_total=0, time_sum=0, time=0;
*info = 0;
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,n)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return */
if ( n == 0 )
return *info;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
nb = magma_get_dpotrf_nb(n);
if ( num_gpus0 > n/nb ) {
num_gpus = n/nb;
if ( n%nb != 0 ) num_gpus ++;
} else {
num_gpus = num_gpus0;
}
//ldda = ((n+31)/32)*32;
ldda = ((n+nb-1)/nb)*nb;
lddla = ((nb*((n+nb*num_gpus-1)/(nb*num_gpus))+31)/32)*32;
/* figure out NB */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(float);
MB = n; /* number of rows in the big panel */
NB = (magma_int_t)((0.8*freeMem-max(2,num_gpus)*nb*ldda-(n+nb)*nb)/lddla); /* number of columns in the big panel */
//NB = min(5*nb,n);
if ( NB >= n ) {
#ifdef CHECK_SPOTRF_OOC
printf( " * still fits in GPU memory.\n" );
#endif
NB = n;
} else {
#ifdef CHECK_SPOTRF_OOC
printf( " * doesn't fit in GPU memory.\n" );
#endif
NB = (NB/nb) * nb; /* making sure it's devisable by nb */
}
#ifdef CHECK_SPOTRF_OOC
if ( NB != n ) printf( " * running in out-core mode (n=%d, NB=%d, nb=%d, lddla=%d, freeMem=%.2e).\n", n, NB, nb, lddla, (float)freeMem );
else printf( " * running in in-core mode (n=%d, NB=%d, nb=%d, lddla=%d, freeMem=%.2e).\n", n, NB, nb, lddla, (float)freeMem );
fflush(stdout);
#endif
for (d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
if (MAGMA_SUCCESS != magma_smalloc( &dt[d], NB*lddla + max(2,num_gpus)*nb*ldda )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dwork[d] = &dt[d][max(2,num_gpus)*nb*ldda];
for( j=0; j < 3; j++ )
magma_queue_create( &stream[d][j] );
for( j=0; j < 5; j++ )
magma_event_create( &event[d][j] );
magma_device_sync(); // synch the device
}
magma_setdevice(0);
timer_start( time_total );
if (nb <= 1 || nb >= n) {
lapackf77_spotrf(uplo_, &n, A, &lda, info);
} else {
/* Use hybrid blocked code. */
if (upper) {
/* =========================================================== *
* Compute the Cholesky factorization A = U'*U. *
* big panel is divided by block-row and distributed in block *
* column cyclic format */
/* for each big-panel */
for( J=0; J < n; J += NB ) {
JB = min(NB,n-J);
if ( num_gpus0 > (n-J)/nb ) {
num_gpus = (n-J)/nb;
if ( (n-J)%nb != 0 ) num_gpus ++;
} else {
num_gpus = num_gpus0;
}
/* load the new big-panel by block-rows */
magma_shtodpo( num_gpus, uplo, JB, n, J, J, nb, A, lda, dwork, NB, stream, &iinfo);
/* update with the previous big-panels */
timer_start( time );
for( j=0; j < J; j += nb ) {
/* upload the diagonal of the block column (broadcast to all GPUs) */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_ssetmatrix_async( nb, JB,
A(j, J), lda,
dTup(d, 0, J), nb,
stream[d][0] );
n_local[d] = 0;
}
/* distribute off-diagonal blocks to GPUs */
for( jj=J+JB; jj < n; jj += nb ) {
d = ((jj-J)/nb)%num_gpus;
magma_setdevice(d);
jb = min(nb, n-jj);
magma_ssetmatrix_async( nb, jb,
A(j, jj), lda,
dTup(d, 0, J+JB+n_local[d]), nb,
stream[d][0] );
n_local[d] += jb;
}
/* wait for the communication */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
}
/* update the current big-panel using the previous block-row */
/* -- process the big diagonal block of the big panel */
for( jj=0; jj < JB; jj += nb ) { // jj is 'local' column index within the big panel
d = (jj/nb)%num_gpus;
J2 = jj/(nb*num_gpus);
magma_setdevice(d);
magmablasSetKernelStream(stream[d][J2%2]); // the last stream (2) used to process off-diagonal
J2 = nb*J2;
jb = min(nb,JB-jj); // number of columns in this current block-row
magma_sgemm( MagmaConjTrans, MagmaNoTrans,
jj, jb, nb,
c_neg_one, dTup(d, 0, J ), nb,
dTup(d, 0, J+jj), nb,
c_one, dAup(d, 0, J2), NB);
magma_ssyrk(MagmaUpper, MagmaConjTrans, jb, nb,
d_neg_one, dTup(d, 0, J+jj), nb,
d_one, dAup(d, jj, J2), NB);
}
/* -- process the remaining big off-diagonal block of the big panel */
if ( n > J+JB ) {
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][2]);
/* local number of columns in the big panel */
n_local[d] = ((n-J)/(nb*num_gpus))*nb;
if (d < ((n-J)/nb)%num_gpus)
n_local[d] += nb;
else if (d == ((n-J)/nb)%num_gpus)
n_local[d] += (n-J)%nb;
/* subtracting the local number of columns in the diagonal */
J2 = nb*(JB/(nb*num_gpus));
if ( d < (JB/nb)%num_gpus )
J2 += nb;
n_local[d] -= J2;
magma_sgemm( MagmaConjTrans, MagmaNoTrans,
JB, n_local[d], nb,
c_neg_one, dTup(d, 0, J ), nb,
dTup(d, 0, J+JB), nb,
c_one, dAup(d, 0, J2), NB);
}
}
/* wait for the previous updates */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( jj=0; jj < 3; jj++ )
magma_queue_sync( stream[d][jj] );
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
} /* end of updates with previous rows */
/* factor the big panel */
h = (JB+nb-1)/nb; // big diagonal of big panel will be on CPU
// using two streams
//magma_spotrf2_mgpu(num_gpus, uplo, JB, n-J, J, J, nb,
// dwork, NB, dt, ldda, A, lda, h, stream, event, &iinfo);
// using three streams
magma_spotrf3_mgpu(num_gpus, uplo, JB, n-J, J, J, nb,
dwork, NB, dt, ldda, A, lda, h, stream, event, &iinfo);
if ( iinfo != 0 ) {
*info = J+iinfo;
break;
}
time_sum += timer_stop( time );
/* upload the off-diagonal (and diagonal!!!) big panel */
magma_sdtohpo(num_gpus, uplo, JB, n, J, J, nb, NB, A, lda, dwork, NB, stream, &iinfo);
//magma_sdtohpo(num_gpus, uplo, JB, n, J, J, nb, 0, A, lda, dwork, NB, stream, &iinfo);
}
} else {
/* ========================================================= *
* Compute the Cholesky factorization A = L*L'. */
/* for each big-panel */
for( J=0; J < n; J += NB ) {
JB = min(NB,n-J);
if ( num_gpus0 > (n-J)/nb ) {
num_gpus = (n-J)/nb;
if ( (n-J)%nb != 0 ) num_gpus ++;
} else {
num_gpus = num_gpus0;
}
/* load the new big-panel by block-columns */
magma_shtodpo( num_gpus, uplo, n, JB, J, J, nb, A, lda, dwork, lddla, stream, &iinfo);
/* update with the previous big-panels */
timer_start( time );
for( j=0; j < J; j += nb ) {
/* upload the diagonal of big panel */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_ssetmatrix_async( JB, nb,
A(J, j), lda,
dT(d, J, 0), ldda,
stream[d][0] );
n_local[d] = 0;
}
/* upload off-diagonals */
for( jj=J+JB; jj < n; jj += nb ) {
d = ((jj-J)/nb)%num_gpus;
magma_setdevice(d);
jb = min(nb, n-jj);
magma_ssetmatrix_async( jb, nb,
A(jj, j), lda,
dT(d, J+JB+n_local[d], 0), ldda,
stream[d][0] );
n_local[d] += jb;
}
/* wait for the communication */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
}
/* update the current big-panel using the previous block-row */
for( jj=0; jj < JB; jj += nb ) { /* diagonal */
d = (jj/nb)%num_gpus;
J2 = jj/(nb*num_gpus);
magma_setdevice(d);
magmablasSetKernelStream(stream[d][J2%2]);
J2 = nb*J2;
jb = min(nb,JB-jj);
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
jb, jj, nb,
c_neg_one, dT(d, J+jj, 0), ldda,
dT(d, J, 0), ldda,
c_one, dA(d, J2, 0), lddla);
magma_ssyrk(MagmaLower, MagmaNoTrans, jb, nb,
d_neg_one, dT(d, J+jj, 0), ldda,
d_one, dA(d, J2, jj), lddla);
}
if ( n > J+JB ) { /* off-diagonal */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][2]);
/* local number of columns in the big panel */
n_local[d] = (((n-J)/nb)/num_gpus)*nb;
if (d < ((n-J)/nb)%num_gpus)
n_local[d] += nb;
else if (d == ((n-J)/nb)%num_gpus)
n_local[d] += (n-J)%nb;
/* subtracting local number of columns in diagonal */
J2 = nb*(JB/(nb*num_gpus));
if ( d < (JB/nb)%num_gpus )
J2 += nb;
n_local[d] -= J2;
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
n_local[d], JB, nb,
c_neg_one, dT(d, J+JB, 0), ldda,
dT(d, J, 0), ldda,
c_one, dA(d, J2, 0), lddla);
}
}
/* wait for the previous updates */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( jj=0; jj < 3; jj++ )
magma_queue_sync( stream[d][jj] );
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
}
/* factor the big panel */
h = (JB+nb-1)/nb; // big diagonal of big panel will be on CPU
// using two streams
//magma_spotrf2_mgpu(num_gpus, uplo, n-J, JB, J, J, nb,
// dwork, lddla, dt, ldda, A, lda, h, stream, event, &iinfo);
// using three streams
magma_spotrf3_mgpu(num_gpus, uplo, n-J, JB, J, J, nb,
dwork, lddla, dt, ldda, A, lda, h, stream, event, &iinfo);
if ( iinfo != 0 ) {
*info = J+iinfo;
break;
}
time_sum += timer_stop( time );
/* upload the off-diagonal big panel */
magma_sdtohpo( num_gpus, uplo, n, JB, J, J, nb, JB, A, lda, dwork, lddla, stream, &iinfo);
} /* end of for J */
} /* if upper */
} /* if nb */
timer_stop( time_total );
if ( num_gpus0 > n/nb ) {
num_gpus = n/nb;
if ( n%nb != 0 ) num_gpus ++;
} else {
num_gpus = num_gpus0;
}
for (d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j < 3; j++ ) {
magma_queue_destroy( stream[d][j] );
}
magma_free( dt[d] );
for( j=0; j < 5; j++ ) {
magma_event_destroy( event[d][j] );
}
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
timer_printf( "\n n=%d NB=%d nb=%d\n", (int) n, (int) NB, (int) nb );
timer_printf( " Without memory allocation: %f / %f = %f GFlop/s\n",
FLOPS_SPOTRF(n) / 1e9, time_total,
FLOPS_SPOTRF(n) / 1e9 / time_total );
timer_printf( " Performance %f / %f = %f GFlop/s\n",
FLOPS_SPOTRF(n) / 1e9, time_sum,
FLOPS_SPOTRF(n) / 1e9 / time_sum );
return *info;
} /* magma_spotrf_ooc */
/**
Purpose
-------
SLAEX3 finds the roots of the secular equation, as defined by the
values in D, W, and RHO, between 1 and K. It makes the
appropriate calls to SLAED4 and then updates the eigenvectors by
multiplying the matrix of eigenvectors of the pair of eigensystems
being combined by the matrix of eigenvectors of the K-by-K system
which is solved here.
It is used in the last step when only a part of the eigenvectors
is required.
It compute only the required part of the eigenvectors and the rest
is not used.
This code makes very mild assumptions about floating point
arithmetic. It will work on machines with a guard digit in
add/subtract, or on those binary machines without guard digits
which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2.
It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
---------
@param[in]
k INTEGER
The number of terms in the rational function to be solved by
SLAED4. K >= 0.
@param[in]
n INTEGER
The number of rows and columns in the Q matrix.
N >= K (deflation may result in N > K).
@param[in]
n1 INTEGER
The location of the last eigenvalue in the leading submatrix.
min(1,N) <= N1 <= N/2.
@param[out]
d REAL array, dimension (N)
D(I) contains the updated eigenvalues for
1 <= I <= K.
@param[out]
Q REAL array, dimension (LDQ,N)
Initially the first K columns are used as workspace.
On output the columns ??? to ??? contain
the updated eigenvectors.
@param[in]
ldq INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
@param[in]
rho REAL
The value of the parameter in the rank one update equation.
RHO >= 0 required.
@param[in,out]
dlamda REAL array, dimension (K)
The first K elements of this array contain the old roots
of the deflated updating problem. These are the poles
of the secular equation. May be changed on output by
having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
Cray-2, or Cray C-90, as described above.
@param[in]
Q2 REAL array, dimension (LDQ2, N)
The first K columns of this matrix contain the non-deflated
eigenvectors for the split problem.
TODO what is LDQ2?
@param[in]
indx INTEGER array, dimension (N)
The permutation used to arrange the columns of the deflated
Q matrix into three groups (see SLAED2).
The rows of the eigenvectors found by SLAED4 must be likewise
permuted before the matrix multiply can take place.
@param[in]
ctot INTEGER array, dimension (4)
A count of the total number of the various types of columns
in Q, as described in INDX. The fourth column type is any
column which has been deflated.
@param[in,out]
w REAL array, dimension (K)
The first K elements of this array contain the components
of the deflation-adjusted updating vector. Destroyed on
output.
@param
s (workspace) REAL array, dimension (N1 + 1)*K
Will contain the eigenvectors of the repaired matrix which
will be multiplied by the previously accumulated eigenvectors
to update the system.
@param[out]
indxq INTEGER array, dimension (N)
On exit, the permutation which will reintegrate the
subproblems back into sorted order,
i.e. D( INDXQ( I = 1, N ) ) will be in ascending order.
@param
dwork (workspace) REAL array, dimension (3*N*N/2+3*N)
@param[in]
range magma_range_t
- = MagmaRangeAll: all eigenvalues will be found.
- = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU]
will be found.
- = MagmaRangeI: the IL-th through IU-th eigenvalues will be found.
TODO verify range, vl, vu, il, iu -- copied from slaex1.
@param[in]
vl REAL
@param[in]
vu REAL
if RANGE=MagmaRangeV, the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeI.
@param[in]
il INTEGER
@param[in]
iu INTEGER
if RANGE=MagmaRangeI, the indices (in ascending order) of the
smallest and largest eigenvalues to be returned.
1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeV.
@param[out]
info INTEGER
- = 0: successful exit.
- < 0: if INFO = -i, the i-th argument had an illegal value.
- > 0: if INFO = 1, an eigenvalue did not converge
Further Details
---------------
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
@ingroup magma_ssyev_aux
********************************************************************/
extern "C" magma_int_t
magma_slaex3(magma_int_t k, magma_int_t n, magma_int_t n1, float* d,
float* Q, magma_int_t ldq, float rho,
float* dlamda, float* Q2, magma_int_t* indx,
magma_int_t* ctot, float* w, float* s, magma_int_t* indxq,
float* dwork,
magma_range_t range, float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t* info )
{
#define Q(i_,j_) (Q + (i_) + (j_)*ldq)
float d_one = 1.;
float d_zero = 0.;
magma_int_t ione = 1;
magma_int_t ineg_one = -1;
magma_int_t iil, iiu, rk;
float* dq2= dwork;
float* ds = dq2 + n*(n/2+1);
float* dq = ds + n*(n/2+1);
magma_int_t lddq = n/2 + 1;
magma_int_t i, iq2, j, n12, n2, n23, tmp, lq2;
float temp;
magma_int_t alleig, valeig, indeig;
alleig = (range == MagmaRangeAll);
valeig = (range == MagmaRangeV);
indeig = (range == MagmaRangeI);
*info = 0;
if (k < 0)
*info=-1;
else if (n < k)
*info=-2;
else if (ldq < max(1,n))
*info=-6;
else if (! (alleig || valeig || indeig))
*info = -15;
else {
if (valeig) {
if (n > 0 && vu <= vl)
*info = -17;
}
else if (indeig) {
if (il < 1 || il > max(1,n))
*info = -18;
else if (iu < min(n,il) || iu > n)
*info = -19;
}
}
if (*info != 0) {
magma_xerbla(__func__, -(*info));
return *info;
}
// Quick return if possible
if (k == 0)
return *info;
/*
Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
be computed with high relative accuracy (barring over/underflow).
This is a problem on machines without a guard digit in
add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
which on any of these machines zeros out the bottommost
bit of DLAMDA(I) if it is 1; this makes the subsequent
subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
occurs. On binary machines with a guard digit (almost all
machines) it does not change DLAMDA(I) at all. On hexadecimal
and decimal machines with a guard digit, it slightly
changes the bottommost bits of DLAMDA(I). It does not account
for hexadecimal or decimal machines without guard digits
(we know of none). We use a subroutine call to compute
2*DLAMBDA(I) to prevent optimizing compilers from eliminating
this code.*/
n2 = n - n1;
n12 = ctot[0] + ctot[1];
n23 = ctot[1] + ctot[2];
iq2 = n1 * n12;
lq2 = iq2 + n2 * n23;
magma_ssetvector_async( lq2, Q2, 1, dq2, 1, NULL );
#ifdef _OPENMP
/////////////////////////////////////////////////////////////////////////////////
//openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
#pragma omp parallel private(i, j, tmp, temp)
{
magma_int_t id = omp_get_thread_num();
magma_int_t tot = omp_get_num_threads();
magma_int_t ib = ( id * k) / tot; //start index of local loop
magma_int_t ie = ((id+1) * k) / tot; //end index of local loop
magma_int_t ik = ie - ib; //number of local indices
for (i = ib; i < ie; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = ib; j < ie; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0) {
#pragma omp critical (info)
*info=iinfo;
break;
}
}
#pragma omp barrier
if (*info == 0) {
#pragma omp single
{
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
}
if (k == 2) {
#pragma omp single
{
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &ik, &w[ib], &ione, &s[ib], &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &ik, Q(ib,ib), &tmp, &w[ib], &ione);
for (j = 0; j < k; ++j) {
magma_int_t i_tmp = min(j, ie);
for (i = ib; i < i_tmp; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
i_tmp = max(j+1, ib);
for (i = i_tmp; i < ie; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = ib; i < ie; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
#pragma omp barrier
//reduce the number of used threads to have enough S workspace
tot = min(n1, omp_get_num_threads());
if (id < tot) {
ib = ( id * rk) / tot + iil - 1;
ie = ((id+1) * rk) / tot + iil - 1;
ik = ie - ib;
}
else {
ib = -1;
ie = -1;
ik = -1;
}
// Compute eigenvectors of the modified rank-1 modification.
for (j = ib; j < ie; ++j) {
for (i = 0; i < k; ++i)
s[id*k + i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s+id*k, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[id*k + iii] / temp;
}
}
}
}
}
if (*info != 0)
return *info;
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#else
/////////////////////////////////////////////////////////////////////////////////
// Non openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
for (i = 0; i < k; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = 0; j < k; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0)
*info=iinfo;
}
if (*info != 0)
return *info;
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
if (k == 2) {
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &k, w, &ione, s, &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &k, Q, &tmp, w, &ione);
for (j = 0; j < k; ++j) {
for (i = 0; i < j; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
for (i = j+1; i < k; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = 0; i < k; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
// Compute eigenvectors of the modified rank-1 modification.
for (j = iil-1; j < iiu; ++j) {
for (i = 0; i < k; ++i)
s[i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[iii] / temp;
}
}
}
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#endif //_OPENMP
// Compute the updated eigenvectors.
timer_start( time );
magma_queue_sync( NULL );
if (rk != 0) {
if ( n23 != 0 ) {
if (rk < magma_get_slaed3_k()) {
lapackf77_slacpy("A", &n23, &rk, Q(ctot[0],iil-1), &ldq, s, &n23);
blasf77_sgemm("N", "N", &n2, &rk, &n23, &d_one, &Q2[iq2], &n2,
s, &n23, &d_zero, Q(n1,iil-1), &ldq );
} else {
magma_ssetmatrix( n23, rk, Q(ctot[0],iil-1), ldq, ds, n23 );
magma_sgemm( MagmaNoTrans, MagmaNoTrans, n2, rk, n23, d_one, &dq2[iq2], n2, ds, n23, d_zero, dq, lddq);
magma_sgetmatrix( n2, rk, dq, lddq, Q(n1,iil-1), ldq );
}
} else
lapackf77_slaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if ( n12 != 0 ) {
if (rk < magma_get_slaed3_k()) {
lapackf77_slacpy("A", &n12, &rk, Q(0,iil-1), &ldq, s, &n12);
blasf77_sgemm("N", "N", &n1, &rk, &n12, &d_one, Q2, &n1,
s, &n12, &d_zero, Q(0,iil-1), &ldq);
} else {
magma_ssetmatrix( n12, rk, Q(0,iil-1), ldq, ds, n12 );
magma_sgemm( MagmaNoTrans, MagmaNoTrans, n1, rk, n12, d_one, dq2, n1, ds, n12, d_zero, dq, lddq);
magma_sgetmatrix( n1, rk, dq, lddq, Q(0,iil-1), ldq );
}
} else
lapackf77_slaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
timer_stop( time );
timer_printf( "gemms = %6.2f\n", time );
return *info;
} /* magma_slaex3 */
/**
Purpose
-------
SGEGQR orthogonalizes the N vectors given by a real M-by-N matrix A:
A = Q * R.
On exit, if successful, the orthogonal vectors Q overwrite A
and R is given in work (on the CPU memory).
The routine is designed for tall-and-skinny matrices: M >> N, N <= 128.
This version uses normal equations and SVD in an iterative process that
makes the computation numerically accurate.
Arguments
---------
@param[in]
ikind INTEGER
Several versions are implemented indiceted by the ikind value:
1: This version uses normal equations and SVD in an iterative process
that makes the computation numerically accurate.
2: This version uses a standard LAPACK-based orthogonalization through
MAGMA's QR panel factorization (magma_sgeqr2x3_gpu) and magma_sorgqr
3: MGS
4. Cholesky QR [ Note: this method uses the normal equations which
squares the condition number of A, therefore
||I - Q'Q|| < O(eps cond(A)^2) ]
@param[in]
m INTEGER
The number of rows of the matrix A. m >= n >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. 128 >= n >= 0.
@param[in,out]
dA REAL array on the GPU, dimension (ldda,n)
On entry, the m-by-n matrix A.
On exit, the m-by-n matrix Q with orthogonal columns.
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,m).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
@param
dwork (GPU workspace) REAL array, dimension:
n^2 for ikind = 1
3 n^2 + min(m, n) + 2 for ikind = 2
0 (not used) for ikind = 3
n^2 for ikind = 4
@param[out]
work (CPU workspace) REAL array, dimension 3 n^2.
On exit, work(1:n^2) holds the rectangular matrix R.
Preferably, for higher performance, work should be in pinned memory.
@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.
@ingroup magma_sgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_sgegqr_gpu( magma_int_t ikind, magma_int_t m, magma_int_t n,
float *dA, magma_int_t ldda,
float *dwork, float *work,
magma_int_t *info )
{
#define work(i_,j_) (work + (i_) + (j_)*n)
#define dA(i_,j_) (dA + (i_) + (j_)*ldda)
magma_int_t i = 0, j, k, n2 = n*n;
magma_int_t ione = 1;
float c_zero = MAGMA_S_ZERO;
float c_one = MAGMA_S_ONE;
float cn = 200., mins, maxs;
/* check arguments */
*info = 0;
if (ikind < 1 || ikind > 4) {
*info = -1;
} else if (m < 0 || m < n) {
*info = -2;
} else if (n < 0 || n > 128) {
*info = -3;
} else if (ldda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (ikind == 1) {
// === Iterative, based on SVD ============================================================
float *U, *VT, *vt, *R, *G, *hwork, *tau;
float *S;
R = work; // Size n * n
G = R + n*n; // Size n * n
VT = G + n*n; // Size n * n
magma_smalloc_cpu( &hwork, 32 + 2*n*n + 2*n);
if ( hwork == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_int_t lwork=n*n+32; // First part f hwork; used as workspace in svd
U = hwork + n*n + 32; // Size n*n
S = (float *)(U+n*n); // Size n
tau = U + n*n + n; // Size n
#if defined(PRECISION_c) || defined(PRECISION_z)
float *rwork;
magma_smalloc_cpu( &rwork, 5*n);
if ( rwork == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
#endif
do {
i++;
magma_sgemm(MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, dA, ldda, dA, ldda, c_zero, dwork, n );
magma_sgetmatrix(n, n, dwork, n, G, n);
#if defined(PRECISION_s) || defined(PRECISION_d)
lapackf77_sgesvd("n", "a", &n, &n, G, &n, S, U, &n, VT, &n,
hwork, &lwork, info);
#else
lapackf77_sgesvd("n", "a", &n, &n, G, &n, S, U, &n, VT, &n,
hwork, &lwork, rwork, info);
#endif
mins = 100.f, maxs = 0.f;
for (k=0; k < n; k++) {
S[k] = magma_ssqrt( S[k] );
if (S[k] < mins) mins = S[k];
if (S[k] > maxs) maxs = S[k];
}
for (k=0; k < n; k++) {
vt = VT + k*n;
for (j=0; j < n; j++)
vt[j] *= S[j];
}
lapackf77_sgeqrf(&n, &n, VT, &n, tau, hwork, &lwork, info);
if (i == 1)
blasf77_scopy(&n2, VT, &ione, R, &ione);
else
blasf77_strmm("l", "u", "n", "n", &n, &n, &c_one, VT, &n, R, &n);
magma_ssetmatrix(n, n, VT, n, dwork, n);
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, m, n, c_one, dwork, n, dA, ldda);
if (mins > 0.00001f)
cn = maxs/mins;
//fprintf(stderr, "Iteration %d, cond num = %f \n", i, cn);
} while (cn > 10.f);
magma_free_cpu( hwork );
#if defined(PRECISION_c) || defined(PRECISION_z)
magma_free_cpu( rwork );
#endif
// ================== end of ikind == 1 ===================================================
}
else if (ikind == 2) {
// ================== LAPACK based ===================================================
magma_int_t min_mn = min(m, n);
magma_int_t nb = n;
float *dtau = dwork + 2*n*n, *d_T = dwork, *ddA = dwork + n*n;
float *tau = work+n*n;
magmablas_slaset( MagmaFull, n, n, c_zero, c_zero, d_T, n );
magma_sgeqr2x3_gpu(m, n, dA, ldda, dtau, d_T, ddA,
(float *)(dwork+min_mn+2*n*n), info);
magma_sgetmatrix( min_mn, 1, dtau, min_mn, tau, min_mn);
magma_sgetmatrix( n, n, ddA, n, work, n);
magma_sorgqr_gpu( m, n, n, dA, ldda, tau, d_T, nb, info );
// ================== end of ikind == 2 ===================================================
}
else if (ikind == 3) {
// ================== MGS ===================================================
for(magma_int_t j = 0; j<n; j++){
for(magma_int_t i = 0; i<j; i++){
*work(i, j) = magma_sdot(m, dA(0,i), 1, dA(0,j), 1);
magma_saxpy(m, -(*work(i,j)), dA(0,i), 1, dA(0,j), 1);
}
for(magma_int_t i = j; i<n; i++)
*work(i, j) = MAGMA_S_ZERO;
//*work(j,j) = MAGMA_S_MAKE( magma_snrm2(m, dA(0,j), 1), 0. );
*work(j,j) = magma_sdot(m, dA(0,j), 1, dA(0,j), 1);
*work(j,j) = MAGMA_S_MAKE( sqrt(MAGMA_S_REAL( *work(j,j) )), 0.);
magma_sscal(m, 1./ *work(j,j), dA(0,j), 1);
}
// ================== end of ikind == 3 ===================================================
}
else if (ikind == 4) {
// ================== Cholesky QR ===================================================
magma_sgemm(MagmaConjTrans, MagmaNoTrans, n, n, m, c_one, dA, ldda, dA, ldda, c_zero, dwork, n );
magma_sgetmatrix(n, n, dwork, n, work, n);
lapackf77_spotrf("u", &n, work, &n, info);
magma_ssetmatrix(n, n, work, n, dwork, n);
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, m, n, c_one, dwork, n, dA, ldda);
// ================== end of ikind == 4 ===================================================
}
return *info;
} /* magma_sgegqr_gpu */
extern "C" magma_int_t
magma_slobpcg( magma_s_sparse_matrix A, magma_s_solver_par *solver_par ) {
#define residualNorms(i,iter) ( residualNorms + (i) + (iter)*n )
#define magmablas_swap(x, y) { pointer = x; x = y; y = pointer; }
#define hresidualNorms(i,iter) (hresidualNorms + (i) + (iter)*n )
#define gramA( m, n) (gramA + (m) + (n)*ldgram)
#define gramB( m, n) (gramB + (m) + (n)*ldgram)
#define gevectors(m, n) (gevectors + (m) + (n)*ldgram)
#define h_gramB( m, n) (h_gramB + (m) + (n)*ldgram)
#define magma_s_bspmv_tuned(m, n, alpha, A, X, beta, AX) { \
magmablas_stranspose( m, n, X, m, blockW, n ); \
magma_s_vector x, ax; \
x.memory_location = Magma_DEV; x.num_rows = m*n; x.nnz = m*n; x.val = blockW; \
ax.memory_location= Magma_DEV; ax.num_rows = m*n; ax.nnz = m*n; ax.val = AX; \
magma_s_spmv(alpha, A, x, beta, ax ); \
magmablas_stranspose( n, m, blockW, n, X, m ); \
}
//**************************************************************
// Memory allocation for the eigenvectors, eigenvalues, and workspace
solver_par->solver = Magma_LOBPCG;
magma_int_t m = A.num_rows;
magma_int_t n =(solver_par->num_eigenvalues);
float *blockX = solver_par->eigenvectors;
float *evalues = solver_par->eigenvalues;
float *dwork, *hwork;
float *blockP, *blockAP, *blockR, *blockAR, *blockAX, *blockW;
float *gramA, *gramB, *gramM;
float *gevectors, *h_gramB;
float *pointer, *origX = blockX;
float *eval_gpu;
magma_int_t lwork = max( 2*n+n*magma_get_dsytrd_nb(n),
1 + 6*3*n + 2* 3*n* 3*n);
magma_smalloc_pinned( &hwork , lwork );
magma_smalloc( &blockAX , m*n );
magma_smalloc( &blockAR , m*n );
magma_smalloc( &blockAP , m*n );
magma_smalloc( &blockR , m*n );
magma_smalloc( &blockP , m*n );
magma_smalloc( &blockW , m*n );
magma_smalloc( &dwork , m*n );
magma_smalloc( &eval_gpu , 3*n );
//**********************************************************+
magma_int_t verbosity = 1;
magma_int_t *iwork, liwork = 15*n+9;
// === Set solver parameters ===
float residualTolerance = solver_par->epsilon;
magma_int_t maxIterations = solver_par->maxiter;
// === Set some constants & defaults ===
float c_one = MAGMA_S_ONE, c_zero = MAGMA_S_ZERO;
float *residualNorms, *condestGhistory, condestG;
float *gevalues;
magma_int_t *activeMask;
// === Check some parameters for possible quick exit ===
solver_par->info = 0;
if (m < 2)
solver_par->info = -1;
else if (n > m)
solver_par->info = -2;
if (solver_par->info != 0) {
magma_xerbla( __func__, -(solver_par->info) );
return solver_par->info;
}
magma_int_t *info = &(solver_par->info); // local info variable;
// === Allocate GPU memory for the residual norms' history ===
magma_smalloc(&residualNorms, (maxIterations+1) * n);
magma_malloc( (void **)&activeMask, (n+1) * sizeof(magma_int_t) );
// === Allocate CPU work space ===
magma_smalloc_cpu(&condestGhistory, maxIterations+1);
magma_smalloc_cpu(&gevalues, 3 * n);
magma_malloc_cpu((void **)&iwork, liwork * sizeof(magma_int_t));
float *hW;
magma_smalloc_pinned(&hW, n*n);
magma_smalloc_pinned(&gevectors, 9*n*n);
magma_smalloc_pinned(&h_gramB , 9*n*n);
// === Allocate GPU workspace ===
magma_smalloc(&gramM, n * n);
magma_smalloc(&gramA, 9 * n * n);
magma_smalloc(&gramB, 9 * n * n);
#if defined(PRECISION_z) || defined(PRECISION_c)
float *rwork;
magma_int_t lrwork = 1 + 5*(3*n) + 2*(3*n)*(3*n);
magma_smalloc_cpu(&rwork, lrwork);
#endif
// === Set activemask to one ===
for(int k =0; k<n; k++)
iwork[k]=1;
magma_setmatrix(n, 1, sizeof(magma_int_t), iwork, n ,activeMask, n);
magma_int_t gramDim, ldgram = 3*n, ikind = 4;
// === Make the initial vectors orthonormal ===
magma_sgegqr_gpu(ikind, m, n, blockX, m, dwork, hwork, info );
//magma_sorthomgs( m, n, blockX );
magma_s_bspmv_tuned(m, n, c_one, A, blockX, c_zero, blockAX );
// === Compute the Gram matrix = (X, AX) & its eigenstates ===
magma_sgemm(MagmaTrans, MagmaNoTrans, n, n, m,
c_one, blockX, m, blockAX, m, c_zero, gramM, n);
magma_ssyevd_gpu( MagmaVec, MagmaUpper,
n, gramM, n, evalues, hW, n, hwork, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork, info );
// === Update X = X * evectors ===
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, n,
c_one, blockX, m, gramM, n, c_zero, blockW, m);
magmablas_swap(blockW, blockX);
// === Update AX = AX * evectors ===
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, n,
c_one, blockAX, m, gramM, n, c_zero, blockW, m);
magmablas_swap(blockW, blockAX);
condestGhistory[1] = 7.82;
magma_int_t iterationNumber, cBlockSize, restart = 1, iter;
//Chronometry
real_Double_t tempo1, tempo2;
magma_device_sync();
tempo1=magma_wtime();
// === Main LOBPCG loop ============================================================
for(iterationNumber = 1; iterationNumber < maxIterations; iterationNumber++)
{
// === compute the residuals (R = Ax - x evalues )
magmablas_slacpy( MagmaUpperLower, m, n, blockAX, m, blockR, m);
/*
for(int i=0; i<n; i++){
magma_saxpy(m, MAGMA_S_MAKE(-evalues[i],0), blockX+i*m, 1, blockR+i*m, 1);
}
*/
#if defined(PRECISION_z) || defined(PRECISION_d)
magma_dsetmatrix( 3*n, 1, evalues, 3*n, eval_gpu, 3*n );
#else
magma_ssetmatrix( 3*n, 1, evalues, 3*n, eval_gpu, 3*n );
#endif
magma_slobpcg_res( m, n, eval_gpu, blockX, blockR, eval_gpu);
magmablas_snrm2_cols(m, n, blockR, m, residualNorms(0, iterationNumber));
// === remove the residuals corresponding to already converged evectors
magma_scompact(m, n, blockR, m,
residualNorms(0, iterationNumber), residualTolerance,
activeMask, &cBlockSize);
if (cBlockSize == 0)
break;
// === apply a preconditioner P to the active residulas: R_new = P R_old
// === for now set P to be identity (no preconditioner => nothing to be done )
// magmablas_slacpy( MagmaUpperLower, m, cBlockSize, blockR, m, blockW, m);
/*
// === make the preconditioned residuals orthogonal to X
magma_sgemm(MagmaTrans, MagmaNoTrans, n, cBlockSize, m,
c_one, blockX, m, blockR, m, c_zero, gramB(0,0), ldgram);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, cBlockSize, n,
c_mone, blockX, m, gramB(0,0), ldgram, c_one, blockR, m);
*/
// === make the active preconditioned residuals orthonormal
magma_sgegqr_gpu(ikind, m, cBlockSize, blockR, m, dwork, hwork, info );
//magma_sorthomgs( m, cBlockSize, blockR );
// === compute AR
magma_s_bspmv_tuned(m, cBlockSize, c_one, A, blockR, c_zero, blockAR );
if (!restart) {
// === compact P & AP as well
magma_scompactActive(m, n, blockP, m, activeMask);
magma_scompactActive(m, n, blockAP, m, activeMask);
/*
// === make P orthogonal to X ?
magma_sgemm(MagmaTrans, MagmaNoTrans, n, cBlockSize, m,
c_one, blockX, m, blockP, m, c_zero, gramB(0,0), ldgram);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, cBlockSize, n,
c_mone, blockX, m, gramB(0,0), ldgram, c_one, blockP, m);
// === make P orthogonal to R ?
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, cBlockSize, m,
c_one, blockR, m, blockP, m, c_zero, gramB(0,0), ldgram);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, cBlockSize, cBlockSize,
c_mone, blockR, m, gramB(0,0), ldgram, c_one, blockP, m);
*/
// === Make P orthonormal & properly change AP (without multiplication by A)
magma_sgegqr_gpu(ikind, m, cBlockSize, blockP, m, dwork, hwork, info );
//magma_sorthomgs( m, cBlockSize, blockP );
//magma_s_bspmv_tuned(m, cBlockSize, c_one, A, blockP, c_zero, blockAP );
magma_ssetmatrix( cBlockSize, cBlockSize, hwork, cBlockSize, dwork, cBlockSize);
// magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit,
// m, cBlockSize, c_one, dwork, cBlockSize, blockAP, m);
// replacement according to Stan
#if defined(PRECISION_s) || defined(PRECISION_d)
magmablas_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit,
m, cBlockSize, c_one, dwork, cBlockSize, blockAP, m);
#else
magma_strsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit, m,
cBlockSize, c_one, dwork, cBlockSize, blockAP, m);
#endif
}
iter = max(1,iterationNumber-10- (int)(log(1.*cBlockSize)));
float condestGmean = 0.;
for(int i = 0; i<iterationNumber-iter+1; i++)
condestGmean += condestGhistory[i];
condestGmean = condestGmean / (iterationNumber-iter+1);
if (restart)
gramDim = n+cBlockSize;
else
gramDim = n+2*cBlockSize;
/* --- The Raileight-Ritz method for [X R P] -----------------------
[ X R P ]' [AX AR AP] y = evalues [ X R P ]' [ X R P ], i.e.,
GramA GramB
/ X'AX X'AR X'AP \ / X'X X'R X'P \
| R'AX R'AR R'AP | y = evalues | R'X R'R R'P |
\ P'AX P'AR P'AP / \ P'X P'R P'P /
----------------------------------------------------------------- */
// === assemble GramB; first, set it to I
magmablas_slaset(MagmaFull, ldgram, ldgram, c_zero, c_one, gramB, ldgram); // identity
if (!restart) {
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, n, m,
c_one, blockP, m, blockX, m, c_zero, gramB(n+cBlockSize,0), ldgram);
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, cBlockSize, m,
c_one, blockP, m, blockR, m, c_zero, gramB(n+cBlockSize,n), ldgram);
}
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, n, m,
c_one, blockR, m, blockX, m, c_zero, gramB(n,0), ldgram);
// === get GramB from the GPU to the CPU and compute its eigenvalues only
magma_sgetmatrix(gramDim, gramDim, gramB, ldgram, h_gramB, ldgram);
lapackf77_ssyev("N", "L", &gramDim, h_gramB, &ldgram, gevalues,
hwork, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork,
#endif
info);
// === check stability criteria if we need to restart
condestG = log10( gevalues[gramDim-1]/gevalues[0] ) + 1.;
if ((condestG/condestGmean>2 && condestG>2) || condestG>8) {
// Steepest descent restart for stability
restart=1;
printf("restart at step #%d\n", (int) iterationNumber);
}
// === assemble GramA; first, set it to I
magmablas_slaset(MagmaFull, ldgram, ldgram, c_zero, c_one, gramA, ldgram); // identity
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, n, m,
c_one, blockR, m, blockAX, m, c_zero, gramA(n,0), ldgram);
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, cBlockSize, m,
c_one, blockR, m, blockAR, m, c_zero, gramA(n,n), ldgram);
if (!restart) {
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, n, m,
c_one, blockP, m, blockAX, m, c_zero,
gramA(n+cBlockSize,0), ldgram);
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, cBlockSize, m,
c_one, blockP, m, blockAR, m, c_zero,
gramA(n+cBlockSize,n), ldgram);
magma_sgemm(MagmaTrans, MagmaNoTrans, cBlockSize, cBlockSize, m,
c_one, blockP, m, blockAP, m, c_zero,
gramA(n+cBlockSize,n+cBlockSize), ldgram);
}
/*
// === Compute X' AX or just use the eigenvalues below ?
magma_sgemm(MagmaTrans, MagmaNoTrans, n, n, m,
c_one, blockX, m, blockAX, m, c_zero,
gramA(0,0), ldgram);
*/
if (restart==0) {
magma_sgetmatrix(gramDim, gramDim, gramA, ldgram, gevectors, ldgram);
}
else {
gramDim = n+cBlockSize;
magma_sgetmatrix(gramDim, gramDim, gramA, ldgram, gevectors, ldgram);
}
for(int k=0; k<n; k++)
*gevectors(k,k) = MAGMA_S_MAKE(evalues[k], 0);
// === the previous eigensolver destroyed what is in h_gramB => must copy it again
magma_sgetmatrix(gramDim, gramDim, gramB, ldgram, h_gramB, ldgram);
magma_int_t itype = 1;
lapackf77_ssygvd(&itype, "V", "L", &gramDim,
gevectors, &ldgram, h_gramB, &ldgram,
gevalues, hwork, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork, info);
for(int k =0; k<n; k++)
evalues[k] = gevalues[k];
// === copy back the result to gramA on the GPU and use it for the updates
magma_ssetmatrix(gramDim, gramDim, gevectors, ldgram, gramA, ldgram);
if (restart == 0) {
// === contribution from P to the new X (in new search direction P)
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize,
c_one, blockP, m, gramA(n+cBlockSize,0), ldgram, c_zero, dwork, m);
magmablas_swap(dwork, blockP);
// === contribution from R to the new X (in new search direction P)
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize,
c_one, blockR, m, gramA(n,0), ldgram, c_one, blockP, m);
// === corresponding contribution from AP to the new AX (in AP)
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize,
c_one, blockAP, m, gramA(n+cBlockSize,0), ldgram, c_zero, dwork, m);
magmablas_swap(dwork, blockAP);
// === corresponding contribution from AR to the new AX (in AP)
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize,
c_one, blockAR, m, gramA(n,0), ldgram, c_one, blockAP, m);
}
else {
// === contribution from R (only) to the new X
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, cBlockSize,
c_one, blockR, m, gramA(n,0), ldgram, c_zero, blockP, m);
// === corresponding contribution from AR (only) to the new AX
magma_sgemm(MagmaNoTrans, MagmaNoTrans,m, n, cBlockSize,
c_one, blockAR, m, gramA(n,0), ldgram, c_zero, blockAP, m);
}
// === contribution from old X to the new X + the new search direction P
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, n,
c_one, blockX, m, gramA, ldgram, c_zero, dwork, m);
magmablas_swap(dwork, blockX);
//magma_saxpy(m*n, c_one, blockP, 1, blockX, 1);
magma_slobpcg_maxpy( m, n, blockP, blockX );
// === corresponding contribution from old AX to new AX + AP
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, n,
c_one, blockAX, m, gramA, ldgram, c_zero, dwork, m);
magmablas_swap(dwork, blockAX);
//magma_saxpy(m*n, c_one, blockAP, 1, blockAX, 1);
magma_slobpcg_maxpy( m, n, blockAP, blockAX );
condestGhistory[iterationNumber+1]=condestG;
if (verbosity==1) {
// float res;
// magma_sgetmatrix(1, 1,
// (float*)residualNorms(0, iterationNumber), 1,
// (float*)&res, 1);
//
// printf("Iteration %4d, CBS %4d, Residual: %10.7f\n",
// iterationNumber, cBlockSize, res);
printf("%4d-%2d ", (int) iterationNumber, (int) cBlockSize);
magma_sprint_gpu(1, n, residualNorms(0, iterationNumber), 1);
}
restart = 0;
} // === end for iterationNumber = 1,maxIterations =======================
// fill solver info
magma_device_sync();
tempo2=magma_wtime();
solver_par->runtime = (real_Double_t) tempo2-tempo1;
solver_par->numiter = iterationNumber;
if( solver_par->numiter < solver_par->maxiter) {
solver_par->info = 0;
} else if( solver_par->init_res > solver_par->final_res )
solver_par->info = -2;
else
solver_par->info = -1;
// =============================================================================
// === postprocessing;
// =============================================================================
// === compute the real AX and corresponding eigenvalues
magma_s_bspmv_tuned(m, n, c_one, A, blockX, c_zero, blockAX );
magma_sgemm(MagmaTrans, MagmaNoTrans, n, n, m,
c_one, blockX, m, blockAX, m, c_zero, gramM, n);
magma_ssyevd_gpu( MagmaVec, MagmaUpper,
n, gramM, n, gevalues, dwork, n, hwork, lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, lrwork,
#endif
iwork, liwork, info );
for(int k =0; k<n; k++)
evalues[k] = gevalues[k];
// === update X = X * evectors
magmablas_swap(blockX, dwork);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, n,
c_one, dwork, m, gramM, n, c_zero, blockX, m);
// === update AX = AX * evectors to compute the final residual
magmablas_swap(blockAX, dwork);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, m, n, n,
c_one, dwork, m, gramM, n, c_zero, blockAX, m);
// === compute R = AX - evalues X
magmablas_slacpy( MagmaUpperLower, m, n, blockAX, m, blockR, m);
for(int i=0; i<n; i++)
magma_saxpy(m, MAGMA_S_MAKE(-evalues[i], 0), blockX+i*m, 1, blockR+i*m, 1);
// === residualNorms[iterationNumber] = || R ||
magmablas_snrm2_cols(m, n, blockR, m, residualNorms(0, iterationNumber));
// === restore blockX if needed
if (blockX != origX)
magmablas_slacpy( MagmaUpperLower, m, n, blockX, m, origX, m);
printf("Eigenvalues:\n");
for(int i =0; i<n; i++)
printf("%e ", evalues[i]);
printf("\n\n");
printf("Final residuals:\n");
magma_sprint_gpu(1, n, residualNorms(0, iterationNumber), 1);
printf("\n\n");
//=== Print residual history in a file for plotting ====
float *hresidualNorms;
magma_smalloc_cpu(&hresidualNorms, (iterationNumber+1) * n);
magma_sgetmatrix(n, iterationNumber,
(float*)residualNorms, n,
(float*)hresidualNorms, n);
printf("Residuals are stored in file residualNorms\n");
printf("Plot the residuals using: myplot \n");
FILE *residuals_file;
residuals_file = fopen("residualNorms", "w");
for(int i =1; i<iterationNumber; i++) {
for(int j = 0; j<n; j++)
fprintf(residuals_file, "%f ", *hresidualNorms(j,i));
fprintf(residuals_file, "\n");
}
fclose(residuals_file);
magma_free_cpu(hresidualNorms);
// === free work space
magma_free( residualNorms );
magma_free_cpu( condestGhistory );
magma_free_cpu( gevalues );
magma_free_cpu( iwork );
magma_free_pinned( hW );
magma_free_pinned( gevectors );
magma_free_pinned( h_gramB );
magma_free( gramM );
magma_free( gramA );
magma_free( gramB );
magma_free( activeMask );
magma_free( blockAX );
magma_free( blockAR );
magma_free( blockAP );
magma_free( blockR );
magma_free( blockP );
magma_free( blockW );
magma_free( dwork );
magma_free( eval_gpu );
magma_free_pinned( hwork );
#if defined(PRECISION_z) || defined(PRECISION_c)
magma_free_cpu( rwork );
#endif
return MAGMA_SUCCESS;
}
/**
Purpose
-------
SGETRF_NOPIV_GPU computes an LU factorization of a general M-by-N
matrix A without any pivoting.
The factorization has the form
A = L * U
where 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 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]
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_nopiv_gpu(
magma_int_t m, magma_int_t n,
magmaFloat_ptr dA, magma_int_t ldda,
magma_int_t *info )
{
#ifdef HAVE_clBLAS
#define dA(i_, j_) dA, (dA_offset + (i_)*nb + (j_)*nb*ldda)
#else
#define dA(i_, j_) (dA + (i_)*nb + (j_)*nb*ldda)
#endif
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, mindim;
magma_int_t j, rows, s, ldwork;
float *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_sgetrf_nb( m, n );
s = mindim / nb;
magma_queue_t queues[2];
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_smalloc_cpu( &work, m*n )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_sgetmatrix( m, n, dA(0,0), ldda, work, m, queues[0] );
magma_sgetrf_nopiv( m, n, work, m, info );
magma_ssetmatrix( m, n, work, m, dA(0,0), ldda, queues[0] );
magma_free_cpu( work );
}
else {
/* Use hybrid blocked code. */
maxm = magma_roundup( m, 32 );
ldwork = maxm;
if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, ldwork*nb )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
for( j=0; j < s; j++ ) {
// get j-th panel from device
magma_queue_sync( queues[1] );
magma_sgetmatrix_async( m-j*nb, nb, dA(j,j), ldda, work, ldwork, queues[0] );
if ( j > 0 ) {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, n - (j+1)*nb,
c_one, dA(j-1,j-1), ldda,
dA(j-1,j+1), ldda, queues[1] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-j*nb, n-(j+1)*nb, nb,
c_neg_one, dA(j, j-1), ldda,
dA(j-1,j+1), ldda,
c_one, dA(j, j+1), ldda, queues[1] );
}
// do the cpu part
rows = m - j*nb;
magma_queue_sync( queues[0] );
magma_sgetrf_nopiv( rows, nb, work, ldwork, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j*nb;
// send j-th panel to device
magma_ssetmatrix_async( m-j*nb, nb, work, ldwork, dA(j, j), ldda, queues[0] );
magma_queue_sync( queues[0] );
// do the small non-parallel computations (next panel update)
if ( s > j+1 ) {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dA(j, j ), ldda,
dA(j, j+1), ldda, queues[1] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-(j+1)*nb, nb, nb,
c_neg_one, dA(j+1, j ), ldda,
dA(j, j+1), ldda,
c_one, dA(j+1, j+1), ldda, queues[1] );
}
else {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, n-s*nb,
c_one, dA(j, j ), ldda,
dA(j, j+1), ldda, queues[1] );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-(j+1)*nb, n-(j+1)*nb, nb,
c_neg_one, dA(j+1, j ), ldda,
dA(j, j+1), ldda,
c_one, dA(j+1, j+1), ldda, queues[1] );
}
}
magma_int_t nb0 = min( m - s*nb, n - s*nb );
if ( nb0 > 0 ) {
rows = m - s*nb;
magma_sgetmatrix( rows, nb0, dA(s,s), ldda, work, ldwork, queues[1] );
// do the cpu part
magma_sgetrf_nopiv( rows, nb0, work, ldwork, &iinfo );
if ( *info == 0 && iinfo > 0 )
*info = iinfo + s*nb;
// send j-th panel to device
magma_ssetmatrix( rows, nb0, work, ldwork, dA(s,s), ldda, queues[1] );
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb0, n-s*nb-nb0,
c_one, dA(s,s), ldda,
dA(s,s)+nb0, ldda, queues[1] );
}
magma_free_pinned( work );
}
magma_queue_destroy( queues[0] );
magma_queue_destroy( queues[1] );
return *info;
} /* magma_sgetrf_nopiv_gpu */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing sgemm
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gflops, magma_perf, magma_time, dev_perf, dev_time, cpu_perf, cpu_time;
float magma_error, dev_error, Cnorm, work[1];
magma_int_t M, N, K;
magma_int_t Am, An, Bm, Bn;
magma_int_t sizeA, sizeB, sizeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t status = 0;
float *h_A, *h_B, *h_C, *h_Cmagma, *h_Cdev;
magmaFloat_ptr d_A, d_B, d_C;
float c_neg_one = MAGMA_S_NEG_ONE;
float alpha = MAGMA_S_MAKE( 0.29, -0.86 );
float beta = MAGMA_S_MAKE( -0.48, 0.38 );
magma_opts opts;
parse_opts( argc, argv, &opts );
float tol = opts.tolerance * lapackf77_slamch("E");
#ifdef HAVE_CUBLAS
// for CUDA, we can check MAGMA vs. CUBLAS, without running LAPACK
printf("If running lapack (option --lapack), MAGMA and %s error are both computed\n"
"relative to CPU BLAS result. Else, MAGMA error is computed relative to %s result.\n\n",
g_platform_str, g_platform_str );
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K MAGMA Gflop/s (ms) %s Gflop/s (ms) CPU Gflop/s (ms) MAGMA error %s error\n",
g_platform_str, g_platform_str );
#else
// for others, we need LAPACK for check
opts.lapack |= opts.check; // check (-c) implies lapack (-l)
printf("transA = %s, transB = %s\n",
lapack_trans_const(opts.transA),
lapack_trans_const(opts.transB) );
printf(" M N K %s Gflop/s (ms) CPU Gflop/s (ms) %s error\n",
g_platform_str, g_platform_str );
#endif
printf("=========================================================================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
M = opts.msize[itest];
N = opts.nsize[itest];
K = opts.ksize[itest];
gflops = FLOPS_SGEMM( M, N, K ) / 1e9;
if ( opts.transA == MagmaNoTrans ) {
lda = Am = M;
An = K;
} else {
lda = Am = K;
An = M;
}
if ( opts.transB == MagmaNoTrans ) {
ldb = Bm = K;
Bn = N;
} else {
ldb = Bm = N;
Bn = K;
}
ldc = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
sizeA = lda*An;
sizeB = ldb*Bn;
sizeC = ldc*N;
TESTING_MALLOC_CPU( h_A, float, lda*An );
TESTING_MALLOC_CPU( h_B, float, ldb*Bn );
TESTING_MALLOC_CPU( h_C, float, ldc*N );
TESTING_MALLOC_CPU( h_Cmagma, float, ldc*N );
TESTING_MALLOC_CPU( h_Cdev, float, ldc*N );
TESTING_MALLOC_DEV( d_A, float, ldda*An );
TESTING_MALLOC_DEV( d_B, float, lddb*Bn );
TESTING_MALLOC_DEV( d_C, float, lddc*N );
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &sizeA, h_A );
lapackf77_slarnv( &ione, ISEED, &sizeB, h_B );
lapackf77_slarnv( &ione, ISEED, &sizeC, h_C );
magma_ssetmatrix( Am, An, h_A, lda, d_A, 0, ldda, opts.queue );
magma_ssetmatrix( Bm, Bn, h_B, ldb, d_B, 0, lddb, opts.queue );
/* =====================================================================
Performs operation using MAGMABLAS (currently only with CUDA)
=================================================================== */
#ifdef HAVE_CUBLAS
magma_ssetmatrix( M, N, h_C, ldc, d_C, lddc );
magma_time = magma_sync_wtime( NULL );
magmablas_sgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, ldda,
d_B, lddb,
beta, d_C, lddc );
magma_time = magma_sync_wtime( NULL ) - magma_time;
magma_perf = gflops / magma_time;
magma_sgetmatrix( M, N, d_C, lddc, h_Cmagma, ldc );
#endif
/* =====================================================================
Performs operation using CUBLAS / clBLAS / Xeon Phi MKL
=================================================================== */
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, opts.queue );
#ifdef HAVE_CUBLAS
dev_time = magma_sync_wtime( NULL );
cublasSgemm( opts.handle, cublas_trans_const(opts.transA), cublas_trans_const(opts.transB), M, N, K,
&alpha, d_A, ldda,
d_B, lddb,
&beta, d_C, lddc );
dev_time = magma_sync_wtime( NULL ) - dev_time;
#else
dev_time = magma_sync_wtime( opts.queue );
magma_sgemm( opts.transA, opts.transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, opts.queue );
dev_time = magma_sync_wtime( opts.queue ) - dev_time;
#endif
dev_perf = gflops / dev_time;
magma_sgetmatrix( M, N, d_C, 0, lddc, h_Cdev, ldc, opts.queue );
/* =====================================================================
Performs operation using CPU BLAS
=================================================================== */
if ( opts.lapack ) {
cpu_time = magma_wtime();
blasf77_sgemm( lapack_trans_const(opts.transA), lapack_trans_const(opts.transB), &M, &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Check the result
=================================================================== */
if ( opts.lapack ) {
// compute relative error for both magma & dev, relative to lapack,
// |C_magma - C_lapack| / |C_lapack|
Cnorm = lapackf77_slange( "F", &M, &N, h_C, &ldc, work );
blasf77_saxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cdev, &ione );
dev_error = lapackf77_slange( "F", &M, &N, h_Cdev, &ldc, work ) / Cnorm;
#ifdef HAVE_CUBLAS
blasf77_saxpy( &sizeC, &c_neg_one, h_C, &ione, h_Cmagma, &ione );
magma_error = lapackf77_slange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %8.2e %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
magma_error, dev_error,
(magma_error < tol && dev_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol && dev_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) %8.2e %s\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time,
cpu_perf, 1000.*cpu_time,
dev_error,
(dev_error < tol ? "ok" : "failed"));
status += ! (dev_error < tol);
#endif
}
else {
#ifdef HAVE_CUBLAS
// compute relative error for magma, relative to dev (currently only with CUDA)
Cnorm = lapackf77_slange( "F", &M, &N, h_Cdev, &ldc, work );
blasf77_saxpy( &sizeC, &c_neg_one, h_Cdev, &ione, h_Cmagma, &ione );
magma_error = lapackf77_slange( "F", &M, &N, h_Cmagma, &ldc, work ) / Cnorm;
printf("%5d %5d %5d %7.2f (%7.2f) %7.2f (%7.2f) --- ( --- ) %8.2e --- %s\n",
(int) M, (int) N, (int) K,
magma_perf, 1000.*magma_time,
dev_perf, 1000.*dev_time,
magma_error,
(magma_error < tol ? "ok" : "failed"));
status += ! (magma_error < tol);
#else
printf("%5d %5d %5d %7.2f (%7.2f) --- ( --- ) ---\n",
(int) M, (int) N, (int) K,
dev_perf, 1000.*dev_time );
#endif
}
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_Cmagma );
TESTING_FREE_CPU( h_Cdev );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
/**
Purpose
-------
SGETRF_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]
num_gpus INTEGER
The number of GPUs. num_gpus > 0.
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A 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_m(magma_int_t num_gpus, magma_int_t m, magma_int_t n,
float *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;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float *dAT[MagmaMaxGPUs], *dA[MagmaMaxGPUs], *dPT[MagmaMaxGPUs];
magma_int_t iinfo = 0, nb, nbi, maxm, n_local[MagmaMaxGPUs], ldn_local;
magma_int_t N, M, NB, NBk, I, d, num_gpus0 = num_gpus;
magma_int_t ii, jj, h, offset, ib, rows, s;
magma_queue_t stream[MagmaMaxGPUs][2];
magma_event_t event[MagmaMaxGPUs][2];
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/* initialize nb */
nb = magma_get_sgetrf_nb(m);
maxm = ((m + 31)/32)*32;
/* figure out NB */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(float);
/* number of columns in the big panel */
h = 1+(2+num_gpus0);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
const char* ngr_nb_char = getenv("MAGMA_NGR_NB");
if ( ngr_nb_char != NULL )
NB = max( nb, min( NB, atoi(ngr_nb_char) ) );
//NB = 5*max(nb,32);
if ( num_gpus0 > ceil((float)NB/nb) ) {
num_gpus = (int)ceil((float)NB/nb);
h = 1+(2+num_gpus);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
} else {
num_gpus = num_gpus0;
}
if ( num_gpus*NB >= n ) {
#ifdef CHECK_SGETRF_OOC
printf( " * still fit in GPU memory.\n" );
#endif
NB = n;
} else {
#ifdef CHECK_SGETRF_OOC
printf( " * don't fit in GPU memory.\n" );
#endif
NB = num_gpus*NB;
NB = max( nb, (NB / nb) * nb); /* making sure it's devisable by nb (x64) */
}
#ifdef CHECK_SGETRF_OOC
if ( NB != n ) printf( " * running in out-core mode (n=%d, NB=%d, nb=%d, freeMem=%.2e).\n", n, NB, nb, (float)freeMem );
else printf( " * running in in-core mode (n=%d, NB=%d, nb=%d, freeMem=%.2e).\n", n, NB, nb, (float)freeMem );
#endif
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code for scalar of one tile. */
lapackf77_sgetrf(&m, &n, A, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
/* allocate memory on GPU to store the big panel */
timer_start( time_alloc );
n_local[0] = (NB/nb)/num_gpus;
if ( NB%(nb*num_gpus) != 0 )
n_local[0]++;
n_local[0] *= nb;
ldn_local = ((n_local[0]+31)/32)*32;
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
if (MAGMA_SUCCESS != magma_smalloc( &dA[d], (ldn_local+h*nb)*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dPT[d] = dA[d] + nb*maxm; /* for storing the previous panel from CPU */
dAT[d] = dA[d] + h*nb*maxm; /* for storing the big panel */
magma_queue_create( &stream[d][0] );
magma_queue_create( &stream[d][1] );
magma_event_create( &event[d][0] );
magma_event_create( &event[d][1] );
}
//magma_setdevice(0);
timer_stop( time_alloc );
for( I=0; I < n; I += NB ) {
M = m;
N = min( NB, n-I ); /* number of columns in this big panel */
s = min( max(m-I,0), N )/nb; /* number of small block-columns in this big panel */
maxm = ((M + 31)/32)*32;
if ( num_gpus0 > ceil((float)N/nb) ) {
num_gpus = (int)ceil((float)N/nb);
} else {
num_gpus = num_gpus0;
}
for( d=0; d < num_gpus; d++ ) {
n_local[d] = ((N/nb)/num_gpus)*nb;
if (d < (N/nb)%num_gpus)
n_local[d] += nb;
else if (d == (N/nb)%num_gpus)
n_local[d] += N%nb;
}
ldn_local = ((n_local[0]+31)/32)*32;
/* upload the next big panel into GPU, transpose (A->A'), and pivot it */
timer_start( time );
magmablas_ssetmatrix_transpose_mgpu(num_gpus, stream, A(0,I), lda,
dAT, ldn_local, dA, maxm, M, N, nb);
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
time_set += timer_stop( time );
timer_start( time );
/* == --------------------------------------------------------------- == */
/* == loop around the previous big-panels to update the new big-panel == */
for( offset = 0; offset < min(m,I); offset += NB ) {
NBk = min( m-offset, NB );
/* start sending the first tile from the previous big-panels to gpus */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
nbi = min( nb, NBk );
magma_ssetmatrix_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_stranspose( M-offset, nbi, dA[d], maxm-offset, dPT(d,0,0), nb );
}
/* applying the pivot from the previous big-panel */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][1]);
magmablas_spermute_long3( dAT(d,0,0), ldn_local, ipiv, NBk, offset );
}
/* == going through each block-column of previous big-panels == */
for( jj=0, ib=offset/nb; jj < NBk; jj += nb, ib++ ) {
ii = offset+jj;
rows = maxm - ii;
nbi = min( nb, NBk-jj );
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
/* wait for a block-column on GPU */
magma_queue_sync( stream[d][0] );
/* start sending next column */
if ( jj+nb < NBk ) {
magma_ssetmatrix_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_stranspose( 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_strsm( 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_sgemm( MagmaNoTrans, MagmaNoTrans,
n_local[d], M-(ii+nb), nbi, c_neg_one, dAT(d,ib,0), ldn_local,
dPT(d,1,(jj/nb)%2), nb, c_one, dAT(d,ib+1,0), ldn_local );
}
magma_event_record( event[d][(jj/nb)%2], stream[d][1] );
} /* end of for each block-columns in a big-panel */
}
} /* end of for each previous big-panels */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
/* calling magma-gpu interface to panel-factorize the big panel */
if ( M > I ) {
//magma_sgetrf1_mgpu(num_gpus, M-I, N, nb, I, dAT, ldn_local, ipiv+I, dA, A(0,I), lda,
// (magma_queue_t **)stream, &iinfo);
magma_sgetrf2_mgpu(num_gpus, M-I, N, nb, I, dAT, ldn_local, ipiv+I, dA, A(0,I), lda,
stream, &iinfo);
if ( iinfo < 0 ) {
*info = iinfo;
break;
} else if ( iinfo != 0 ) {
*info = iinfo + I * NB;
//break;
}
/* adjust pivots */
for( ii=I; ii < min(I+N,m); ii++ )
ipiv[ii] += I;
}
time_comp += timer_stop( time );
/* download the current big panel to CPU */
timer_start( time );
magmablas_sgetmatrix_transpose_mgpu(num_gpus, stream, dAT, ldn_local, A(0,I), lda, dA, maxm, M, N, nb);
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
time_get += timer_stop( time );
} /* end of for */
timer_stop( time_total );
flops = FLOPS_SGETRF( m, n ) / 1e9;
timer_printf(" memory-allocation time: %e\n", time_alloc );
timer_printf(" NB=%d nb=%d\n", (int) NB, (int) nb );
timer_printf(" memcopy and transpose %e seconds\n", time_set );
timer_printf(" total time %e seconds\n", time_total );
timer_printf(" Performance %f GFlop/s, %f seconds without htod and dtoh\n", flops / (time_comp), time_comp );
timer_printf(" Performance %f GFlop/s, %f seconds with htod\n", flops / (time_comp + time_set), time_comp + time_set );
timer_printf(" Performance %f GFlop/s, %f seconds with dtoh\n", flops / (time_comp + time_get), time_comp + time_get );
timer_printf(" Performance %f GFlop/s, %f seconds without memory-allocation\n", flops / (time_total - time_alloc), time_total - time_alloc );
for( d=0; d < num_gpus0; d++ ) {
magma_setdevice(d);
magma_free( dA[d] );
magma_event_destroy( event[d][0] );
magma_event_destroy( event[d][1] );
magma_queue_destroy( stream[d][0] );
magma_queue_destroy( stream[d][1] );
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
}
if ( *info >= 0 )
magma_sgetrf_piv(m, n, NB, A, lda, ipiv, info);
return *info;
} /* magma_sgetrf_m */
int main( int argc, char** argv)
{
real_Double_t gflops, magma_perf, magma_time, clblas_perf, clblas_time, cpu_perf, cpu_time;
float magma_error, clblas_error, work[1];
magma_trans_t transA = MagmaNoTrans;
magma_trans_t transB = MagmaNoTrans;
magma_int_t istart = 1024;
magma_int_t iend = 6240;
magma_int_t M, M0 = 0;
magma_int_t N, N0 = 0;
magma_int_t K, K0 = 0;
magma_int_t i;
magma_int_t Am, An, Bm, Bn;
magma_int_t szeA, szeB, szeC;
magma_int_t lda, ldb, ldc, ldda, lddb, lddc;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float *h_A, *h_B, *h_C, *h_C2, *h_C3;
magmaFloat_ptr d_A, d_B, d_C;
float c_neg_one = MAGMA_S_NEG_ONE;
float alpha = MAGMA_S_MAKE( 0.29, -0.86 );
float beta = MAGMA_S_MAKE( -0.48, 0.38 );
int lapack = getenv("MAGMA_RUN_LAPACK") != NULL;
int count = 1;
printf("\nUsage: testing_sgemm [-NN|NT|TN|TT|NC|CN|TC|CT|CC] -M m -N n -K k -count c -l\n"
" -l or setting $MAGMA_RUN_LAPACK runs CPU BLAS,\n"
" and computes both MAGMA and CLBLAS error using CPU BLAS result.\n"
" Else, MAGMA error is computed using CLBLAS result.\n\n");
for( int i = 1; i < argc; ++i ) {
if ( strcmp("-N", argv[i]) == 0 && i+1 < argc ){
N0 = atoi(argv[++i]);
}
else if ( strcmp("-M", argv[i]) == 0 && i+1 < argc ){
M0 = atoi(argv[++i]);
}
else if ( strcmp("-K", argv[i]) == 0 && i+1 < argc ){
K0 = atoi(argv[++i]);
}
else if (strcmp("-NN", argv[i])==0){
transA = transB = MagmaNoTrans;
}
else if (strcmp("-TT", argv[i])==0){
transA = transB = MagmaTrans;
}
else if (strcmp("-NT", argv[i])==0){
transA = MagmaNoTrans;
transB = MagmaTrans;
}
else if (strcmp("-TN", argv[i])==0){
transA = MagmaTrans;
transB = MagmaNoTrans;
}
else if (strcmp("-NC", argv[i])==0){
transA = MagmaNoTrans;
transB = MagmaConjTrans;
}
else if (strcmp("-TC", argv[i])==0){
transA = MagmaTrans;
transB = MagmaConjTrans;
}
else if (strcmp("-CN", argv[i])==0){
transA = MagmaConjTrans;
transB = MagmaNoTrans;
}
else if (strcmp("-CT", argv[i])==0){
transA = MagmaConjTrans;
transB = MagmaTrans;
}
else if (strcmp("-CC", argv[i])==0){
transA = transB = MagmaConjTrans;
}
else if (strcmp("-l", argv[i])==0) {
lapack = true;
}
else if ( strcmp("-count", argv[i]) == 0 && i+1 < argc ){
count = atoi(argv[++i]);
}
else {
printf( "invalid argument: %s\n", argv[i] );
exit(1);
}
}
if ( (M0 != 0) && (N0 != 0) && (K0 != 0) )
iend = istart + 1;
M = N = K = iend;
if ( M0 != 0 ) M = M0;
if ( N0 != 0 ) N = N0;
if ( K0 != 0 ) K = K0;
if( transA == MagmaNoTrans ) {
Am = M;
An = K;
} else {
Am = K;
An = M;
}
if( transB == MagmaNoTrans ) {
Bm = K;
Bn = N;
} else {
Bm = N;
Bn = K;
}
/* Initialize */
magma_queue_t queue;
magma_device_t device[ MagmaMaxGPUs ];
magma_int_t num = 0;
magma_int_t err;
magma_init();
err = magma_getdevices( device, MagmaMaxGPUs, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_getdevices failed: %d\n", (int) err );
exit(-1);
}
err = magma_queue_create( device[0], &queue );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", (int) err );
exit(-1);
}
lda = ldc = M;
ldb = Bm;
ldda = ((M+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ldda;
K += 32;
M += 32;
N += 32;
TESTING_MALLOC_CPU( h_A, float, lda*K );
TESTING_MALLOC_CPU( h_B, float, ldb*Bn );
TESTING_MALLOC_CPU( h_C, float, ldc*N );
TESTING_MALLOC_CPU( h_C2, float, ldc*N );
TESTING_MALLOC_CPU( h_C3, float, ldc*N );
TESTING_MALLOC_DEV( d_A, float, ldda*K );
TESTING_MALLOC_DEV( d_B, float, lddb*Bn );
TESTING_MALLOC_DEV( d_C, float, lddc*N );
printf("Testing transA = %c transB = %c\n", *lapack_const(transA), *lapack_const(transB));
printf(" M N K MAGMA Gflop/s (sec) CLBLAS Gflop/s (sec) CPU Gflop/s (sec) MAGMA error CLBLAS error\n");
printf("===========================================================================================================\n");
for( i=istart; i<iend; i = (int)(i*1.25) ) {
for( int cnt = 0; cnt < count; ++cnt ) {
M = N = K = i;
if ( M0 != 0 ) M = M0;
if ( N0 != 0 ) N = N0;
if ( K0 != 0 ) K = K0;
if( transA == MagmaNoTrans ) {
lda = Am = M;
An = K;
} else {
lda = Am = K;
An = M;
}
if( transB == MagmaNoTrans ) {
ldb = Bm = K;
Bn = N;
} else {
ldb = Bm = N;
Bn = K;
}
gflops = FLOPS_SGEMM( M, N, K ) / 1e9;
ldc = M;
ldda = ((lda+31)/32)*32;
lddb = ((ldb+31)/32)*32;
lddc = ((ldc+31)/32)*32;
szeA = lda * An;
szeB = ldb * Bn;
szeC = ldc * N;
/* Initialize the matrices */
lapackf77_slarnv( &ione, ISEED, &szeA, h_A );
lapackf77_slarnv( &ione, ISEED, &szeB, h_B );
lapackf77_slarnv( &ione, ISEED, &szeC, h_C );
/* =====================================================================
Performs operation using MAGMA-BLAS
=================================================================== */
magma_ssetmatrix( Am, An, h_A, lda, d_A, 0, ldda, queue );
magma_ssetmatrix( Bm, Bn, h_B, ldb, d_B, 0, lddb, queue );
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magmablas_sgemm_reduce( M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magma_queue_sync(queue);
magma_time = magma_wtime();
magmablas_sgemm_reduce( M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_queue_sync(queue);
magma_time = magma_wtime() - magma_time;
magma_perf = gflops / magma_time;
magma_sgetmatrix( M, N, d_C, 0, lddc, h_C2, ldc, queue );
/* =====================================================================
Performs operation using CUDA-BLAS
=================================================================== */
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magma_sgemm( transA, transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_ssetmatrix( M, N, h_C, ldc, d_C, 0, lddc, queue );
magma_queue_sync(queue);
clblas_time = magma_wtime();
magma_sgemm( transA, transB, M, N, K,
alpha, d_A, 0, ldda,
d_B, 0, lddb,
beta, d_C, 0, lddc, queue );
magma_queue_sync(queue);
clblas_time = magma_wtime() - clblas_time;
clblas_perf = gflops / clblas_time;
magma_sgetmatrix( M, N, d_C, 0, lddc, h_C3, ldc, queue );
/* =====================================================================
Performs operation using BLAS
=================================================================== */
if ( lapack ) {
cpu_time = magma_wtime();
blasf77_sgemm( lapack_const(transA), lapack_const(transB), &M, &N, &K,
&alpha, h_A, &lda,
h_B, &ldb,
&beta, h_C, &ldc );
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gflops / cpu_time;
}
/* =====================================================================
Error Computation and Performance Compariosn
=================================================================== */
if ( lapack ) {
// compare both magma & clblas to lapack
blasf77_saxpy(&szeC, &c_neg_one, h_C, &ione, h_C2, &ione);
magma_error = lapackf77_slange("M", &M, &N, h_C2, &ldc, work);
blasf77_saxpy(&szeC, &c_neg_one, h_C, &ione, h_C3, &ione);
clblas_error = lapackf77_slange("M", &M, &N, h_C3, &ldc, work);
printf("%5d %5d %5d %7.2f (%7.4f) %7.2f (%7.4f) %7.2f (%7.4f) %8.2e %8.2e\n",
(int) M, (int) N, (int) K,
magma_perf, magma_time, clblas_perf, clblas_time, cpu_perf, cpu_time,
magma_error, clblas_error );
}
else {
// compare magma to clblas
blasf77_saxpy(&szeC, &c_neg_one, h_C3, &ione, h_C2, &ione);
magma_error = lapackf77_slange("M", &M, &N, h_C2, &ldc, work);
printf("%5d %5d %5d %7.2f (%7.4f) %7.2f (%7.4f) --- ( --- ) %8.2e ---\n",
(int) M, (int) N, (int) K,
magma_perf, magma_time, clblas_perf, clblas_time,
magma_error );
}
}
if ( count > 1 ) {
printf( "\n" );
}
}
/* Memory clean up */
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_B );
TESTING_FREE_CPU( h_C );
TESTING_FREE_CPU( h_C2 );
TESTING_FREE_CPU( h_C3 );
TESTING_FREE_DEV( d_A );
TESTING_FREE_DEV( d_B );
TESTING_FREE_DEV( d_C );
magma_queue_destroy( queue );
magma_finalize();
}
/**
Purpose
-------
SLAEX3 finds the roots of the secular equation, as defined by the
values in D, W, and RHO, between 1 and K. It makes the
appropriate calls to SLAED4 and then updates the eigenvectors by
multiplying the matrix of eigenvectors of the pair of eigensystems
being combined by the matrix of eigenvectors of the K-by-K system
which is solved here.
It is used in the last step when only a part of the eigenvectors
is required.
It compute only the required part of the eigenvectors and the rest
is not used.
This code makes very mild assumptions about floating point
arithmetic. It will work on machines with a guard digit in
add/subtract, or on those binary machines without guard digits
which subtract like the Cray X-MP, Cray Y-MP, Cray C-90, or Cray-2.
It could conceivably fail on hexadecimal or decimal machines
without guard digits, but we know of none.
Arguments
---------
@param[in]
ngpu INTEGER
Number of GPUs to use. ngpu > 0.
@param[in]
k INTEGER
The number of terms in the rational function to be solved by
SLAED4. K >= 0.
@param[in]
n INTEGER
The number of rows and columns in the Q matrix.
N >= K (deflation may result in N > K).
@param[in]
n1 INTEGER
The location of the last eigenvalue in the leading submatrix.
min(1,N) <= N1 <= N/2.
@param[out]
d REAL array, dimension (N)
D(I) contains the updated eigenvalues for
1 <= I <= K.
@param[out]
Q REAL array, dimension (LDQ,N)
Initially the first K columns are used as workspace.
On output the columns ??? to ??? contain
the updated eigenvectors.
@param[in]
ldq INTEGER
The leading dimension of the array Q. LDQ >= max(1,N).
@param[in]
rho REAL
The value of the parameter in the rank one update equation.
RHO >= 0 required.
@param[in,out]
dlamda REAL array, dimension (K)
The first K elements of this array contain the old roots
of the deflated updating problem. These are the poles
of the secular equation. May be changed on output by
having lowest order bit set to zero on Cray X-MP, Cray Y-MP,
Cray-2, or Cray C-90, as described above.
@param[in]
Q2 REAL array, dimension (LDQ2, N)
The first K columns of this matrix contain the non-deflated
eigenvectors for the split problem.
@param[in]
indx INTEGER array, dimension (N)
The permutation used to arrange the columns of the deflated
Q matrix into three groups (see SLAED2).
The rows of the eigenvectors found by SLAED4 must be likewise
permuted before the matrix multiply can take place.
@param[in]
ctot INTEGER array, dimension (4)
A count of the total number of the various types of columns
in Q, as described in INDX. The fourth column type is any
column which has been deflated.
@param[in,out]
w REAL array, dimension (K)
The first K elements of this array contain the components
of the deflation-adjusted updating vector. Destroyed on
output.
@param
s (workspace) REAL array, dimension (N1 + 1)*K
Will contain the eigenvectors of the repaired matrix which
will be multiplied by the previously accumulated eigenvectors
to update the system.
@param[out]
indxq INTEGER array, dimension (N)
On exit, the permutation which will reintegrate the
subproblems back into sorted order,
i.e. D( INDXQ( I = 1, N ) ) will be in ascending order.
@param
dwork (devices workspaces) REAL array of arrays,
dimension NRGPU.
if NRGPU = 1 the dimension of the first workspace
should be (3*N*N/2+3*N)
otherwise the NRGPU workspaces should have the size
ceil((N-N1) * (N-N1) / floor(ngpu/2)) +
NB * ((N-N1) + (N-N1) / floor(ngpu/2))
@param
queues (device queues) magma_queue_t array,
dimension (MagmaMaxGPUs,2)
@param[in]
range magma_range_t
- = MagmaRangeAll: all eigenvalues will be found.
- = MagmaRangeV: all eigenvalues in the half-open interval (VL,VU]
will be found.
- = MagmaRangeI: the IL-th through IU-th eigenvalues will be found.
TODO verify range, vl, vu, il, iu -- copied from slaex1.
@param[in]
vl REAL
@param[in]
vu REAL
if RANGE=MagmaRangeV, the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeI.
@param[in]
il INTEGER
@param[in]
iu INTEGER
if RANGE=MagmaRangeI, the indices (in ascending order) of the
smallest and largest eigenvalues to be returned.
1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0.
Not referenced if RANGE = MagmaRangeAll or MagmaRangeV.
@param[out]
info INTEGER
- = 0: successful exit.
- < 0: if INFO = -i, the i-th argument had an illegal value.
- > 0: if INFO = 1, an eigenvalue did not converge
Further Details
---------------
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified by Francoise Tisseur, University of Tennessee.
@ingroup magma_ssyev_aux
********************************************************************/
extern "C" magma_int_t
magma_slaex3_m(
magma_int_t ngpu,
magma_int_t k, magma_int_t n, magma_int_t n1, float *d,
float *Q, magma_int_t ldq, float rho,
float *dlamda, float *Q2, magma_int_t *indx,
magma_int_t *ctot, float *w, float *s, magma_int_t *indxq,
magmaFloat_ptr dwork[],
magma_queue_t queues[MagmaMaxGPUs][2],
magma_range_t range, float vl, float vu, magma_int_t il, magma_int_t iu,
magma_int_t *info )
{
#define Q(i_,j_) (Q + (i_) + (j_)*ldq)
#define dQ2(id) (dwork[id])
#define dS(id, ii) (dwork[id] + n2*n2_loc + (ii)*(n2*nb))
#define dQ(id, ii) (dwork[id] + n2*n2_loc + 2*(n2*nb) + (ii)*(n2_loc*nb))
if (ngpu == 1) {
magma_setdevice(0);
magma_slaex3(k, n, n1, d, Q, ldq, rho,
dlamda, Q2, indx, ctot, w, s, indxq,
*dwork, range, vl, vu, il, iu, info );
return *info;
}
float d_one = 1.;
float d_zero = 0.;
magma_int_t ione = 1;
magma_int_t ineg_one = -1;
magma_int_t iil, iiu, rk;
magma_int_t n1_loc, n2_loc, ib, nb, ib2, igpu;
magma_int_t ni_loc[MagmaMaxGPUs];
magma_int_t i, ind, iq2, j, n12, n2, n23, tmp;
float temp;
magma_int_t alleig, valeig, indeig;
alleig = (range == MagmaRangeAll);
valeig = (range == MagmaRangeV);
indeig = (range == MagmaRangeI);
*info = 0;
if (k < 0)
*info=-1;
else if (n < k)
*info=-2;
else if (ldq < max(1,n))
*info=-6;
else if (! (alleig || valeig || indeig))
*info = -15;
else {
if (valeig) {
if (n > 0 && vu <= vl)
*info = -17;
}
else if (indeig) {
if (il < 1 || il > max(1,n))
*info = -18;
else if (iu < min(n,il) || iu > n)
*info = -19;
}
}
if (*info != 0) {
magma_xerbla(__func__, -(*info));
return *info;
}
// Quick return if possible
if (k == 0)
return *info;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/*
Modify values DLAMDA(i) to make sure all DLAMDA(i)-DLAMDA(j) can
be computed with high relative accuracy (barring over/underflow).
This is a problem on machines without a guard digit in
add/subtract (Cray XMP, Cray YMP, Cray C 90 and Cray 2).
The following code replaces DLAMDA(I) by 2*DLAMDA(I)-DLAMDA(I),
which on any of these machines zeros out the bottommost
bit of DLAMDA(I) if it is 1; this makes the subsequent
subtractions DLAMDA(I)-DLAMDA(J) unproblematic when cancellation
occurs. On binary machines with a guard digit (almost all
machines) it does not change DLAMDA(I) at all. On hexadecimal
and decimal machines with a guard digit, it slightly
changes the bottommost bits of DLAMDA(I). It does not account
for hexadecimal or decimal machines without guard digits
(we know of none). We use a subroutine call to compute
2*DLAMBDA(I) to prevent optimizing compilers from eliminating
this code.*/
//#define CHECK_CPU
#ifdef CHECK_CPU
float *hwS[2][MagmaMaxGPUs], *hwQ[2][MagmaMaxGPUs], *hwQ2[MagmaMaxGPUs];
#define hQ2(id) (hwQ2[id])
#define hS(id, ii) (hwS[ii][id])
#define hQ(id, ii) (hwQ[ii][id])
#endif
n2 = n - n1;
n12 = ctot[0] + ctot[1];
n23 = ctot[1] + ctot[2];
iq2 = n1 * n12;
//lq2 = iq2 + n2 * n23;
n1_loc = (n1-1) / (ngpu/2) + 1;
n2_loc = (n2-1) / (ngpu/2) + 1;
nb = magma_get_slaex3_m_nb();
if (n1 >= magma_get_slaex3_m_k()) {
#ifdef CHECK_CPU
for (igpu = 0; igpu < ngpu; ++igpu) {
magma_smalloc_pinned( &(hwS[0][igpu]), n2*nb );
magma_smalloc_pinned( &(hwS[1][igpu]), n2*nb );
magma_smalloc_pinned( &(hwQ2[igpu]), n2*n2_loc );
magma_smalloc_pinned( &(hwQ[0][igpu]), n2_loc*nb );
magma_smalloc_pinned( &(hwQ[1][igpu]), n2_loc*nb );
}
#endif
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
ni_loc[igpu] = min(n1_loc, n1 - igpu/2 * n1_loc);
#ifdef CHECK_CPU
lapackf77_slacpy("A", &ni_loc[igpu], &n12, Q2+n1_loc*(igpu/2), &n1, hQ2(igpu), &n1_loc);
#endif
magma_setdevice(igpu);
magma_ssetmatrix_async( ni_loc[igpu], n12,
Q2+n1_loc*(igpu/2), n1,
dQ2(igpu), n1_loc, queues[igpu][0] );
ni_loc[igpu+1] = min(n2_loc, n2 - igpu/2 * n2_loc);
#ifdef CHECK_CPU
lapackf77_slacpy("A", &ni_loc[igpu+1], &n23, Q2+iq2+n2_loc*(igpu/2), &n2, hQ2(igpu+1), &n2_loc);
#endif
magma_setdevice(igpu+1);
magma_ssetmatrix_async( ni_loc[igpu+1], n23,
Q2+iq2+n2_loc*(igpu/2), n2,
dQ2(igpu+1), n2_loc, queues[igpu+1][0] );
}
}
//
#ifdef _OPENMP
/////////////////////////////////////////////////////////////////////////////////
//openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
#pragma omp parallel private(i, j, tmp, temp)
{
magma_int_t id = omp_get_thread_num();
magma_int_t tot = omp_get_num_threads();
magma_int_t ib = ( id * k) / tot; //start index of local loop
magma_int_t ie = ((id+1) * k) / tot; //end index of local loop
magma_int_t ik = ie - ib; //number of local indices
for (i = ib; i < ie; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = ib; j < ie; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0) {
#pragma omp critical (info)
*info = iinfo;
break;
}
}
#pragma omp barrier
if (*info == 0) {
#pragma omp single
{
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
}
if (k == 2) {
#pragma omp single
{
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &ik, &w[ib], &ione, &s[ib], &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &ik, Q(ib,ib), &tmp, &w[ib], &ione);
for (j = 0; j < k; ++j) {
magma_int_t i_tmp = min(j, ie);
for (i = ib; i < i_tmp; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
i_tmp = max(j+1, ib);
for (i = i_tmp; i < ie; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = ib; i < ie; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
#pragma omp barrier
//reduce the number of used threads to have enough S workspace
tot = min(n1, omp_get_num_threads());
if (id < tot) {
ib = ( id * rk) / tot + iil - 1;
ie = ((id+1) * rk) / tot + iil - 1;
ik = ie - ib;
}
else {
ib = -1;
ie = -1;
ik = -1;
}
// Compute eigenvectors of the modified rank-1 modification.
for (j = ib; j < ie; ++j) {
for (i = 0; i < k; ++i)
s[id*k + i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s+id*k, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[id*k + iii] / temp;
}
}
}
}
}
if (*info != 0)
return *info;
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#else
/////////////////////////////////////////////////////////////////////////////////
// Non openmp implementation
/////////////////////////////////////////////////////////////////////////////////
magma_timer_t time=0;
timer_start( time );
for (i = 0; i < k; ++i)
dlamda[i]=lapackf77_slamc3(&dlamda[i], &dlamda[i]) - dlamda[i];
for (j = 0; j < k; ++j) {
magma_int_t tmpp=j+1;
magma_int_t iinfo = 0;
lapackf77_slaed4(&k, &tmpp, dlamda, w, Q(0,j), &rho, &d[j], &iinfo);
// If the zero finder fails, the computation is terminated.
if (iinfo != 0)
*info=iinfo;
}
if (*info != 0)
return *info;
//Prepare the INDXQ sorting permutation.
magma_int_t nk = n - k;
lapackf77_slamrg( &k, &nk, d, &ione, &ineg_one, indxq);
//compute the lower and upper bound of the non-deflated eigenvectors
if (valeig)
magma_svrange(k, d, &iil, &iiu, vl, vu);
else if (indeig)
magma_sirange(k, indxq, &iil, &iiu, il, iu);
else {
iil = 1;
iiu = k;
}
rk = iiu - iil + 1;
if (k == 2) {
for (j = 0; j < k; ++j) {
w[0] = *Q(0,j);
w[1] = *Q(1,j);
i = indx[0] - 1;
*Q(0,j) = w[i];
i = indx[1] - 1;
*Q(1,j) = w[i];
}
}
else if (k != 1) {
// Compute updated W.
blasf77_scopy( &k, w, &ione, s, &ione);
// Initialize W(I) = Q(I,I)
tmp = ldq + 1;
blasf77_scopy( &k, Q, &tmp, w, &ione);
for (j = 0; j < k; ++j) {
for (i = 0; i < j; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
for (i = j+1; i < k; ++i)
w[i] = w[i] * ( *Q(i, j) / ( dlamda[i] - dlamda[j] ) );
}
for (i = 0; i < k; ++i)
w[i] = copysign( sqrt( -w[i] ), s[i]);
// Compute eigenvectors of the modified rank-1 modification.
for (j = iil-1; j < iiu; ++j) {
for (i = 0; i < k; ++i)
s[i] = w[i] / *Q(i,j);
temp = magma_cblas_snrm2( k, s, 1 );
for (i = 0; i < k; ++i) {
magma_int_t iii = indx[i] - 1;
*Q(i,j) = s[iii] / temp;
}
}
}
timer_stop( time );
timer_printf( "eigenvalues/vector D+zzT = %6.2f\n", time );
#endif //_OPENMP
// Compute the updated eigenvectors.
timer_start( time );
if (rk > 0) {
if (n1 < magma_get_slaex3_m_k()) {
// stay on the CPU
if ( n23 != 0 ) {
lapackf77_slacpy("A", &n23, &rk, Q(ctot[0],iil-1), &ldq, s, &n23);
blasf77_sgemm("N", "N", &n2, &rk, &n23, &d_one, &Q2[iq2], &n2,
s, &n23, &d_zero, Q(n1,iil-1), &ldq );
}
else
lapackf77_slaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if ( n12 != 0 ) {
lapackf77_slacpy("A", &n12, &rk, Q(0,iil-1), &ldq, s, &n12);
blasf77_sgemm("N", "N", &n1, &rk, &n12, &d_one, Q2, &n1,
s, &n12, &d_zero, Q(0,iil-1), &ldq);
}
else
lapackf77_slaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
else {
//use the gpus
ib = min(nb, rk);
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_ssetmatrix_async( n23, ib,
Q(ctot[0],iil-1), ldq,
dS(igpu+1,0), n23, queues[igpu+1][0] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_ssetmatrix_async( n12, ib,
Q(0,iil-1), ldq,
dS(igpu,0), n12, queues[igpu][0] );
}
}
for (i = 0; i < rk; i += nb) {
ib = min(nb, rk - i);
ind = (i/nb)%2;
if (i+nb < rk) {
ib2 = min(nb, rk - i - nb);
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_ssetmatrix_async( n23, ib2,
Q(ctot[0],iil-1+i+nb), ldq,
dS(igpu+1,(ind+1)%2), n23, queues[igpu+1][(ind+1)%2] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_ssetmatrix_async( n12, ib2,
Q(0,iil-1+i+nb), ldq,
dS(igpu,(ind+1)%2), n12, queues[igpu][(ind+1)%2] );
}
}
}
// Ensure that the data is copied on gpu since we will overwrite it.
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
#ifdef CHECK_CPU
lapackf77_slacpy("A", &n23, &ib, Q(ctot[0],iil-1+i), &ldq, hS(igpu+1,ind), &n23);
#endif
magma_setdevice(igpu+1);
magma_queue_sync( queues[igpu+1][ind] );
}
if (n12 != 0) {
#ifdef CHECK_CPU
lapackf77_slacpy("A", &n12, &ib, Q(0,iil-1+i), &ldq, hS(igpu,ind), &n12);
#endif
magma_setdevice(igpu);
magma_queue_sync( queues[igpu][ind] );
}
}
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
#ifdef CHECK_CPU
blasf77_sgemm("N", "N", &ni_loc[igpu+1], &ib, &n23, &d_one, hQ2(igpu+1), &n2_loc,
hS(igpu+1,ind), &n23, &d_zero, hQ(igpu+1, ind), &n2_loc);
#endif
magma_setdevice(igpu+1);
magmablasSetKernelStream(queues[igpu+1][ind]);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, ni_loc[igpu+1], ib, n23, d_one, dQ2(igpu+1), n2_loc,
dS(igpu+1, ind), n23, d_zero, dQ(igpu+1, ind), n2_loc);
#ifdef CHECK_CPU
printf("norm Q %d: %f\n", igpu+1, cpu_gpu_sdiff(ni_loc[igpu+1], ib, hQ(igpu+1, ind), n2_loc, dQ(igpu+1, ind), n2_loc));
#endif
}
if (n12 != 0) {
#ifdef CHECK_CPU
blasf77_sgemm("N", "N", &ni_loc[igpu], &ib, &n12, &d_one, hQ2(igpu), &n1_loc,
hS(igpu,ind%2), &n12, &d_zero, hQ(igpu, ind%2), &n1_loc);
#endif
magma_setdevice(igpu);
magmablasSetKernelStream(queues[igpu][ind]);
magma_sgemm(MagmaNoTrans, MagmaNoTrans, ni_loc[igpu], ib, n12, d_one, dQ2(igpu), n1_loc,
dS(igpu, ind), n12, d_zero, dQ(igpu, ind), n1_loc);
#ifdef CHECK_CPU
printf("norm Q %d: %f\n", igpu, cpu_gpu_sdiff(ni_loc[igpu], ib, hQ(igpu, ind), n1_loc, dQ(igpu, ind), n1_loc));
#endif
}
}
for (igpu = 0; igpu < ngpu-1; igpu += 2) {
if (n23 != 0) {
magma_setdevice(igpu+1);
magma_sgetmatrix( ni_loc[igpu+1], ib, dQ(igpu+1, ind), n2_loc,
Q(n1+n2_loc*(igpu/2),iil-1+i), ldq );
// magma_sgetmatrix_async( ni_loc[igpu+1], ib, dQ(igpu+1, ind), n2_loc,
// Q(n1+n2_loc*(igpu/2),iil-1+i), ldq, queues[igpu+1][ind] );
}
if (n12 != 0) {
magma_setdevice(igpu);
magma_sgetmatrix( ni_loc[igpu], ib, dQ(igpu, ind), n1_loc,
Q(n1_loc*(igpu/2),iil-1+i), ldq );
// magma_sgetmatrix_async( ni_loc[igpu], ib, dQ(igpu, ind), n1_loc,
// Q(n1_loc*(igpu/2),iil-1+i), ldq, queues[igpu][ind] );
}
}
}
for (igpu = 0; igpu < ngpu; ++igpu) {
#ifdef CHECK_CPU
magma_free_pinned( hwS[1][igpu] );
magma_free_pinned( hwS[0][igpu] );
magma_free_pinned( hwQ2[igpu] );
magma_free_pinned( hwQ[1][igpu] );
magma_free_pinned( hwQ[0][igpu] );
#endif
magma_setdevice(igpu);
magma_queue_sync( queues[igpu][0] );
magma_queue_sync( queues[igpu][1] );
}
if ( n23 == 0 )
lapackf77_slaset("A", &n2, &rk, &d_zero, &d_zero, Q(n1,iil-1), &ldq);
if ( n12 == 0 )
lapackf77_slaset("A", &n1, &rk, &d_zero, &d_zero, Q(0,iil-1), &ldq);
}
}
timer_stop( time );
timer_printf( "gemms = %6.2f\n", time );
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_slaed3_m */
/**
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
=======
SSYTRF_nopiv_gpu computes the LDLt factorization of a real symmetric
matrix A.
The factorization has the form
A = U^H * D * U , if UPLO = 'U', or
A = L * D * L^H, if UPLO = 'L',
where U is an upper triangular matrix, L is lower triangular, and
D is a diagonal matrix.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
UPLO CHARACTER*1
- = 'U': Upper triangle of A is stored;
- = 'L': Lower triangle of A is stored.
@param[in]
N INTEGER
The order of the matrix A. N >= 0.
@param[in,out]
dA REAL array on the GPU, dimension (LDA,N)
On entry, the symmetric matrix A. If UPLO = 'U', the leading
N-by-N upper triangular part of A contains the upper
triangular part of the matrix A, and the strictly lower
triangular part of A is not referenced. If UPLO = 'L', the
leading N-by-N lower triangular part of A contains the lower
triangular part of the matrix A, and the strictly upper
triangular part of A is not referenced.
\n
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization A = U^H D U or A = L D L^H.
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using cudaMallocHost.
@param[in]
LDA INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
INFO INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
if INFO = -6, the GPU memory allocation failed
- > 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
@ingroup magma_ssytrf_comp
******************************************************************* */
extern "C" magma_int_t
magma_ssytrf_nopiv_gpu(
magma_uplo_t uplo, magma_int_t n,
magmaFloat_ptr dA, magma_int_t ldda,
magma_int_t *info)
{
#define A(i, j) (A)
#define dA(i, j) (dA +(j)*ldda + (i))
#define dW(i, j) (dW +(j)*ldda + (i))
#define dWt(i, j) (dW +(j)*nb + (i))
/* Local variables */
float zone = MAGMA_S_ONE;
float mzone = MAGMA_S_NEG_ONE;
int upper = (uplo == MagmaUpper);
magma_int_t j, k, jb, nb, ib, iinfo;
*info = 0;
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,n)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return MAGMA_ERR_ILLEGAL_VALUE;
}
/* Quick return */
if ( n == 0 )
return MAGMA_SUCCESS;
nb = magma_get_ssytrf_nopiv_nb(n);
ib = min(32, nb); // inner-block for diagonal factorization
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
magma_queue_t stream[2];
magma_event_t event;
magma_queue_create(&stream[0]);
magma_queue_create(&stream[1]);
magma_event_create( &event );
trace_init( 1, 1, 2, stream );
// CPU workspace
float *A;
if (MAGMA_SUCCESS != magma_smalloc_pinned( &A, nb*nb )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
// GPU workspace
magmaFloat_ptr dW;
if (MAGMA_SUCCESS != magma_smalloc( &dW, (1+nb)*ldda )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/* Use hybrid blocked code. */
if (upper) {
//=========================================================
// Compute the LDLt factorization A = U'*D*U without pivoting.
// main loop
for (j=0; j<n; j += nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
//magma_queue_wait_event( stream[1], event );
magma_event_sync(event);
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), nb, stream[1]);
trace_gpu_end( 0, 0 );
// factorize the diagonal block
magma_queue_sync(stream[1]);
trace_cpu_start( 0, "potrf", "potrf" );
ssytrf_nopiv_cpu(MagmaUpper, jb, ib, A(j, j), nb, info);
trace_cpu_end( 0 );
if (*info != 0){
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), nb, dA(j, j), ldda, stream[0]);
trace_gpu_end( 0, 0 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
magmablasSetKernelStream( stream[0] );
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm(MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaUnit,
jb, (n-j-jb),
zone, dA(j, j), ldda,
dA(j, j+jb), ldda);
magma_scopymatrix( jb, n-j-jb, dA( j, j+jb ), ldda, dWt( 0, j+jb ), nb );
// update the trailing submatrix with D
magmablas_slascl_diag(MagmaUpper, jb, n-j-jb,
dA(j, j), ldda,
dA(j, j+jb), ldda,
&iinfo);
trace_gpu_end( 0, 0 );
// update the trailing submatrix with U and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k<n; k+=nb) {
magma_int_t kb = min(nb,n-k);
magma_sgemm(MagmaConjTrans, MagmaNoTrans, kb, n-k, jb,
mzone, dWt(0, k), nb,
dA(j, k), ldda,
zone, dA(k, k), ldda);
if (k==j+jb)
magma_event_record( event, stream[0] );
}
trace_gpu_end( 0, 0 );
}
}
} else {
//=========================================================
// Compute the LDLt factorization A = L*D*L' without pivoting.
// main loop
for (j=0; j<n; j+=nb) {
jb = min(nb, (n-j));
// copy A(j,j) back to CPU
trace_gpu_start( 0, 0, "get", "get" );
//magma_queue_wait_event( stream[0], event );
magma_event_sync(event);
magma_sgetmatrix_async(jb, jb, dA(j, j), ldda, A(j,j), nb, stream[1]);
trace_gpu_end( 0, 0 );
// factorize the diagonal block
magma_queue_sync(stream[1]);
trace_cpu_start( 0, "potrf", "potrf" );
ssytrf_nopiv_cpu(MagmaLower, jb, ib, A(j, j), nb, info);
trace_cpu_end( 0 );
if (*info != 0){
*info = *info + j;
break;
}
// copy A(j,j) back to GPU
trace_gpu_start( 0, 0, "set", "set" );
magma_ssetmatrix_async(jb, jb, A(j, j), nb, dA(j, j), ldda, stream[0]);
trace_gpu_end( 0, 0 );
if ( (j+jb) < n) {
// compute the off-diagonal blocks of current block column
magmablasSetKernelStream( stream[0] );
trace_gpu_start( 0, 0, "trsm", "trsm" );
magma_strsm(MagmaRight, MagmaLower, MagmaConjTrans, MagmaUnit,
(n-j-jb), jb,
zone, dA(j, j), ldda,
dA(j+jb, j), ldda);
magma_scopymatrix( n-j-jb,jb, dA( j+jb, j ), ldda, dW( j+jb, 0 ), ldda );
// update the trailing submatrix with D
magmablas_slascl_diag(MagmaLower, n-j-jb, jb,
dA(j, j), ldda,
dA(j+jb, j), ldda,
&iinfo);
trace_gpu_end( 0, 0 );
// update the trailing submatrix with L and W
trace_gpu_start( 0, 0, "gemm", "gemm" );
for (k=j+jb; k<n; k+=nb) {
magma_int_t kb = min(nb,n-k);
magma_sgemm(MagmaNoTrans, MagmaConjTrans, n-k, kb, jb,
mzone, dA(k, j), ldda,
dW(k, 0), ldda,
zone, dA(k, k), ldda);
if (k==j+jb)
magma_event_record( event, stream[0] );
}
trace_gpu_end( 0, 0 );
}
}
}
trace_finalize( "ssytrf.svg","trace.css" );
magma_queue_destroy(stream[0]);
magma_queue_destroy(stream[1]);
magma_event_destroy( event );
magma_free( dW );
magma_free_pinned( A );
magmablasSetKernelStream( orig_stream );
return MAGMA_SUCCESS;
} /* magma_ssytrf_nopiv */
void magmablas_ssymm_mgpu_com(
magma_side_t side, magma_uplo_t uplo, magma_int_t m, magma_int_t n,
float alpha,
float *dA[], magma_int_t ldda, magma_int_t offset,
float *dB[], magma_int_t lddb,
float beta, float *dC[], magma_int_t lddc,
float *dwork[], magma_int_t dworksiz,
float *C, magma_int_t ldc,
float *work[], magma_int_t worksiz,
magma_int_t ngpu, magma_int_t nb,
magma_queue_t streams[][20], magma_int_t nstream,
magma_event_t redevents[][MagmaMaxGPUs*MagmaMaxGPUs+10], magma_int_t nbevents,
magma_int_t gnode[MagmaMaxGPUs][MagmaMaxGPUs+2], magma_int_t nbcmplx )
{
#define dA(dev, i, j) (dA[dev] + (i) + (j)*ldda)
#define dB(dev, i, j) (dB[dev] + (i) + (j)*lddb)
#define dC(dev, i, j) (dC[dev] + (i) + (j)*lddc)
#define dwork(dev, i, j) (dwork[dev] + (i) + (j)*lddwork)
#define C(i, j) (C + (i) + (j)*ldc)
//printf("####################################################\n");
//printf(" start ssymm \n");
//printf("####################################################\n");
if ( side != MagmaLeft || uplo != MagmaLower ) {
fprintf( stderr, "%s: only Left Lower implemented\n", __func__ );
}
assert( ldda >= m );
assert( lddb >= m );
assert( lddc >= m );
assert( nstream >= ngpu );
assert( nbevents >= ngpu*ngpu );
float c_one = MAGMA_S_ONE;
float *dwork1[MagmaMaxGPUs];
float *dwork2[MagmaMaxGPUs];
magma_int_t maxgsize = n*m;
magma_int_t lddwork = lddc;
magma_int_t ldwork = m;
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
dwork1[dev] = dwork[dev]; // size of dwork1 is n*lddwork
dwork2[dev] = dwork[dev]+n*lddwork; // size of dwork2 is maxgsize*ngpu
}
assert( dworksiz >= (n*lddwork+maxgsize*ngpu) );
assert( worksiz >= (n*ldwork) );
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_t cstream;
magmablasGetKernelStream(&cstream);
magma_int_t dev, devperm, myblk, mycolsize, myblkoffst;
magma_int_t gmaster;
magma_int_t masterdev, lcdev, lccolsize, myngpu;
magma_int_t stdev = (offset/nb)%ngpu;
magma_int_t blockoffset = offset % nb;
magma_int_t fstblksiz = 0;
if(blockoffset>0){
fstblksiz = min(m, (nb - blockoffset));
}
//magma_int_t nbblk = magma_ceildiv(m, nb);
magma_int_t nbblk = magma_ceildiv((m+blockoffset), nb);
magma_int_t remm = m- fstblksiz;
magma_int_t nbblkoffst = offset/nb;
magma_int_t nblstblks = -1;
magma_int_t devlstblk = -1;
magma_int_t lstblksiz = remm%nb;
if(lstblksiz>0){
nblstblks = nbblk%ngpu;
devlstblk = (nblstblks-1+ngpu)%ngpu;
}
magma_int_t nbcmplxactive = 0;
magma_int_t cmplxisactive[MagmaMaxGPUs];
magma_int_t gpuisactive[MagmaMaxGPUs];
memset(gpuisactive, 0, MagmaMaxGPUs*sizeof(magma_int_t));
memset(cmplxisactive, 0, MagmaMaxGPUs*sizeof(magma_int_t));
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 0 ] );
cudaMemset(dwork(dev,0,0), 0, (lddwork)*(n)*sizeof(float) );
// put all dC on all dev to 0 except the one which
// hold i==0 because this one has to multiply by beta.
if(dev!=stdev){
cudaMemset(dC(dev,0,0), 0, (lddc)*(n)*sizeof(float) );
}
}
magma_int_t newoffset = offset;
// 1. symmetrize
if(blockoffset>0){
newoffset = offset+fstblksiz; // newoffset is adjusted over nb
magma_int_t myblkoffst = (nbblkoffst/ngpu)+(nbblkoffst%ngpu > stdev?1:0);
//printf("STDEV %d voici offset %d remm %d myblockoffset %d siz %d \n", stdev, offset, remm, myblkoffst, fstblksiz);
magma_setdevice( stdev );
magmablasSetKernelStream( streams[ stdev ][ 0 ] );
magmablas_ssymmetrize_tiles( MagmaLower, fstblksiz, dA(stdev, offset, myblkoffst*nb+blockoffset), ldda, 1, ngpu*nb, nb );
}
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_int_t newstdev = (newoffset/nb)%ngpu;
magma_int_t nbblk = remm/nb; // number of block of size nb. if m%nb>0 then a last block exist and is of size ib=m%nb
magma_int_t myblk = (nbblk/ngpu) + (nbblk%ngpu > ((dev-newstdev+ngpu)%ngpu) ? 1:0 );
magma_int_t devperm = (dev-newstdev+ngpu)%ngpu;
magma_int_t nbblkoffst = newoffset/nb;
magma_int_t myblkoffst = (nbblkoffst/ngpu)+(nbblkoffst%ngpu > dev?1:0);
//printf("dev %d devperm %d newoffset %d rowoff %d coloff %d myblk %d \n", dev, devperm, newoffset, newoffset+devperm*nb, myblkoffst*nb, myblk);
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 0 ] );
magmablas_ssymmetrize_tiles( MagmaLower, nb, dA(dev, newoffset+devperm*nb, myblkoffst*nb), ldda, myblk, ngpu*nb, nb );
if(remm%nb>0){
magma_int_t nblstblks = (nbblk+1)%ngpu;
magma_int_t devlstblk = (nblstblks-1+ngpu)%ngpu;
//printf("==> siz %d devperm %d, devlstblk %d, newoffset+nbblk*nb %d, myblkoffst*nb+ myblk*nb %d\n", remm % nb, devperm, devlstblk, newoffset+nbblk*nb, myblkoffst*nb+ myblk*nb);
if(devperm==devlstblk)
magmablas_ssymmetrize( MagmaLower, remm % nb, dA(dev, newoffset+nbblk*nb, myblkoffst*nb+ myblk*nb), ldda ); // last partial tile
}
}
/*
magma_int_t siz = m+offset;
float *R;
magma_smalloc_cpu( &R, siz*siz );
// collecte back A
magmablas_sgetmatrix_1D_bcyclic( siz, siz, dA, ldda, R, siz, ngpu, nb );
magma_setdevice( 0 );
magmablasSetKernelStream( streams[ dev ][ 0 ] );
//magma_sgetmatrix( siz, siz, dA[0], ldda, R, siz );
FILE *trace_file;
trace_file = fopen("AJETE/Aafter", "w");
for (int j = 0; j < siz ; j++)
for (int i = 0; i < siz ; i++)
fprintf(trace_file, "%10d%10d%40.30e\n", i+1, j+1, R[j*siz+i]);
fclose(trace_file);
return;
*/
// ROW GEMM transpose a row and make a gemm with a block
// if only 1 GPU used the ROW GEMM is integrated with the
// COL GEMM (better accuracy observed) and better perf
if(ngpu>1){
for( magma_int_t i = fstblksiz; i < m; i += nb ) {
magma_int_t ib = min( nb, m-i ); // block size
magma_int_t ioff = i + offset; // start global index in parent matrix
//magma_int_t dev = (ioff / nb) % ngpu;
magma_int_t nbblkoffst = offset/nb;
magma_int_t nbblk = magma_ceildiv(i, nb);
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_int_t myblk = (nbblk/ngpu) + (nbblk%ngpu > ((dev-stdev+ngpu)%ngpu) ? 1:0 );
magma_int_t myblkoffst = (nbblkoffst/ngpu)+(nbblkoffst%ngpu > dev?1:0);
magma_int_t myrowsize = myblk * nb;
magma_int_t coloffset = myblkoffst*nb;
if(dev==stdev) {
myrowsize = myrowsize -blockoffset;
coloffset = myblkoffst*nb+blockoffset;
}
//printf("ROW GEMM: voici i %d ib %d ioff %d nbblkoffst %d stdev %d dev %d myblk %d myblkoffset %d coloffset %d rowsize %d\n", i, ib, ioff, nbblkoffst, stdev, dev, myblk, myblkoffst, coloffset, myrowsize);
if(myrowsize>0){
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 1 ] );
magma_sgemm( MagmaConjTrans, MagmaNoTrans, myrowsize, n, ib,
alpha, dA(dev,ioff,coloffset), ldda,
dB(dev,i,0), lddb,
c_one, dwork(dev,0,0), lddwork );
}
}
}
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_event_record(redevents[dev][1], streams[dev][1]);
}
}
// COL GEMM
// blockoffset is offset within first block; for subsequent blocks it is 0
if(blockoffset>0){
magma_int_t ib = min( nb-blockoffset, m ); // block size
magma_int_t iblock = (offset / nb) / ngpu; // local block id
magma_int_t di = iblock*nb+blockoffset; // local index in parent matrix
magma_setdevice( stdev );
magmablasSetKernelStream( streams[ stdev ][ 0 ] );
//printf("DEV %d COL GEMM first ioff %d di %d m %d n %d ib %d \n", stdev, offset, di, m, n, ib);
magma_sgemm( MagmaNoTrans, MagmaNoTrans, m, n, ib,
alpha, dA(stdev,offset,di), ldda,
dB(stdev,0,0), lddb,
beta, dC(stdev,0,0), lddc );
}
// COL GEMM
for( magma_int_t i = fstblksiz; i < m; i += nb ) {
magma_int_t ib = min( nb, m-i ); // block size
magma_int_t ioff = i + offset; // start global index in parent matrix
magma_int_t iblock = (ioff / nb) / ngpu; // local block id
magma_int_t dev = (ioff / nb) % ngpu;
magma_int_t di = iblock*nb; // local index in parent matrix
//printf("DEV %d COL GEMM i %d ioff %d di %d m-i %d n %d ib %d \n", dev, i, ioff, di, m-i, n, ib);
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 0 ] );
if(i==0){
magma_sgemm( MagmaNoTrans, MagmaNoTrans, m-i, n, ib,
alpha, dA(dev,ioff,di), ldda,
dB(dev,i,0), lddb,
beta, dC(dev,i,0), lddc );
}else{
magma_sgemm( MagmaNoTrans, MagmaNoTrans, m-i, n, ib,
alpha, dA(dev,ioff,di), ldda,
dB(dev,i,0), lddb,
c_one, dC(dev,i,0), lddc );
}
magma_event_record(redevents[dev][0], streams[dev][0]);
// if only 1 GPU is used, do the ROW GEMM
if(ngpu==1){
// NOTE THAT because the COL gemm write dC below the diagonal (i)
// and the ROW GEMM write dC from 0 to diag-1, so they could
// run in parallel on different streams.
//
// NO NO NO because
// it might happen that col finished i and strated i+1 while row still at i
// magmablasSetKernelStream( streams[ dev ][ 0 ] );
magma_sgemm( MagmaConjTrans, MagmaNoTrans, i, n, ib,
alpha, dA(dev,ioff,offset), ldda,
dB(dev,i,0), lddb,
c_one, dC(dev,0,0), lddc );
}
}
if(ngpu>1){
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_int_t nbblk = magma_ceildiv((m+blockoffset), nb);
magma_int_t nbblkrow = nbblk-1;
magma_int_t devperm = (dev-stdev+ngpu)%ngpu;
magma_int_t myblk = (nbblkrow/ngpu) + (nbblkrow%ngpu > devperm ? 1:0 );
magma_int_t myrowsize = myblk * nb;
if(dev==stdev) {
myrowsize = myrowsize - blockoffset;
}
//printf("blockoffset %d nbblkrow %d devperm %d DEV %d RECEIVING myblk %d myrowsize %d\n", blockoffset, nbblkrow, devperm, dev, myblk, myrowsize);
if(myrowsize>0){
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 0 ] );
magma_queue_wait_event(streams[ dev ][ 0 ], redevents[dev][1]);
//magma_queue_sync( streams[ dev ][ 1 ] );
// for each dev add the computed ROW block each on its placment with dC
for( magma_int_t blki = 0; blki < myblk; ++blki){
magma_int_t gbblki = (blki*ngpu + devperm)*nb - blockoffset;
magma_int_t lcblki = blki*nb;
magma_int_t ib = nb;// min(nb, m-gbblki);
if(dev==stdev){
lcblki = blki*nb-blockoffset;
if(blki==0){
gbblki = 0;
lcblki = 0;
ib = nb-blockoffset;
}
}
magmablas_sgeadd(ib, n, c_one,
&dwork[dev][lcblki], lddwork,
&dC[dev][gbblki] , lddc );
}
magma_event_record(redevents[dev][0], streams[dev][0]);
}
}
}
// ===========================================================
// COMMUNICATION ALL_REDUCE_SUM
// ===========================================================
if(ngpu==1){
return;
}
// INITIALIZE COMM
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
masterdev = -1;
gnode[cmplxid][MagmaMaxGPUs+1] = -1;
myngpu = gnode[cmplxid][MagmaMaxGPUs];
for( magma_int_t idev = 0; idev < myngpu; ++idev ) {
dev = gnode[cmplxid][idev];
devperm = (dev-stdev+ngpu)%ngpu;
myblk = (nbblk/ngpu) + (nbblk%ngpu > devperm ? 1:0 );
mycolsize = myblk*nb;
myblkoffst = nb*((nbblkoffst/ngpu)+(nbblkoffst%ngpu > dev?1:0));
if(dev==stdev){
mycolsize -= blockoffset;
myblkoffst += blockoffset; // local index in parent matrix
}
if((devperm==devlstblk)&&(lstblksiz>0)){
mycolsize -= (nb-(remm%nb));
}
mycolsize = min(mycolsize, m);
if(mycolsize>0){
gpuisactive[dev] = mycolsize;
if(masterdev==-1) {
masterdev = dev;
nbcmplxactive = nbcmplxactive +1;
cmplxisactive[cmplxid] = 1;
gnode[cmplxid][MagmaMaxGPUs+1] = masterdev;
}
}
}
}
/*
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_device_sync();
}
*/
//*******************************
// each GPU send its result
// to its master. The master make
// the addition and then send to
// to the masters of other real
// and receive from the masters of
// other real make the addition
// and broadcast locally the final
// result.
//*******************************
//printf("=======================================================================\n");
//printf(" sending to my master \n");
//printf("=======================================================================\n");
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
myngpu = gnode[cmplxid][MagmaMaxGPUs];
masterdev = gnode[cmplxid][MagmaMaxGPUs+1];
//check if real is active
if(masterdev!=-1){
for( magma_int_t idev = 0; idev < myngpu; ++idev ) {
dev = gnode[cmplxid][idev];
mycolsize = gpuisactive[dev];
if(mycolsize>0){
// I am an active GPU. if I am not the master, then send my result to my master.
// store result on dwork[masterdev][dev*maxgsize]
if(dev!=masterdev){
magma_setdevice( dev );
//printf(" GPU %d sending to my master %d\n", dev, masterdev);
// wait the geadd of my ROW and COL GEMM is done
magma_queue_wait_event(streams[ dev ][ 0 ], redevents[dev][0]);
// sending to the master of my real
magma_scopymatrix_async(
m, n,
&dC[dev][0], lddc,
&dwork2[masterdev][maxgsize*dev], m, streams[dev][0] );
magma_event_record(redevents[dev][masterdev], streams[dev][0]);
} // end I am not the masterdev
}// end if mycolsize>0
}// for idev
}// end of if masterdev!=-1 maening real is active
}// for cmplxid
/*
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_device_sync();
}
*/
//printf("=======================================================================\n");
//printf(" each master do addition of local result and broadcast to other masters \n");
//printf("=======================================================================\n");
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
myngpu = gnode[cmplxid][MagmaMaxGPUs];
masterdev = gnode[cmplxid][MagmaMaxGPUs+1];
//check if real is active
if(masterdev!=-1){
magma_setdevice( masterdev );
// addition is done on stream 0 sequentially
magmablasSetKernelStream( streams[ masterdev ][ 0 ] );
// wait the geadd of my ROW and COL GEMM is done
magma_queue_wait_event(streams[ masterdev ][ 0 ], redevents[masterdev][0]);
// ========================================
// local addition
// ========================================
for( magma_int_t l = 0; l < myngpu; ++l ) {
lcdev = gnode[cmplxid][l];
lccolsize = gpuisactive[lcdev];
if((lcdev!=masterdev)&&(lccolsize>0)){
//printf(" master %d receiving from %d and adding \n", masterdev, lcdev);
// this is an active GPU of my real.
// wait I received what he send it to me and then do addition.
magma_queue_wait_event(streams[ masterdev ][ 0 ], redevents[lcdev][masterdev]);
magmablas_sgeadd(m, n, c_one,
&dwork2[masterdev][maxgsize*lcdev], m,
&dC[masterdev][0] , lddc );
}
}// for l=1:myngpu
// because addition is done sequentially on stream 0,
// I have to record this to be able to synch using it
magma_event_record(redevents[masterdev][masterdev], streams[masterdev][0]);
// ========================================
//
// ========================================
// send to other masters
// ========================================
for( magma_int_t k = 0; k < nbcmplx; ++k ) {
if(k!=cmplxid){
gmaster = gnode[k][MagmaMaxGPUs+1];
if(gmaster!=-1){ //real is active
//Master has to wait until finish the local addition then send using gmaster stream.
//use stream 0 to make it sequential or stream gmaster to make it parallel.
//Now both re the same.
//printf(" master %d from cmplx %d sending to other master %d on cmplx %d \n", masterdev, cmplxid, gmaster, k);
magma_queue_wait_event(streams[ masterdev ][ gmaster ], redevents[masterdev][masterdev]);
magma_scopymatrix_async(
m, n,
&dC[masterdev][0], lddc,
&dwork2[gmaster][maxgsize*masterdev], m, streams[masterdev][gmaster] );
magma_event_record(redevents[masterdev][gmaster], streams[masterdev][gmaster]);
magma_event_record(redevents[masterdev][masterdev], streams[masterdev][gmaster]);
} // end of gmaster!=-1
} // end of k!=cmplxid
}// for k = 0: nbcmplx
// ========================================
}// end of if masterdev!=-1 maening real is active
}// for cmplxid
/*
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_device_sync();
}
*/
//printf("=======================================================================\n");
//printf(" each master wait receiving other masters results, do the addition and broadcast locally \n");
//printf("=======================================================================\n");
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
myngpu = gnode[cmplxid][MagmaMaxGPUs];
masterdev = gnode[cmplxid][MagmaMaxGPUs+1];
//check if real is active
if(masterdev!=-1){
magma_setdevice( masterdev );
// addition is done on stream 0 sequentially
magmablasSetKernelStream( streams[ masterdev ][ 0 ] );
// master has to wait until finishing all the send to other masters.
magma_queue_wait_event(streams[ masterdev ][ 0 ], redevents[masterdev][masterdev]);
// ========================================
// addition of results from other masters
// ========================================
for( magma_int_t k = 0; k < nbcmplx; ++k ) {
if(k!=cmplxid){
gmaster = gnode[k][MagmaMaxGPUs+1];
if(gmaster!=-1){ //real is active
//Master has to wait until receiving from gmaster, then do addition using stream 0
//printf(" master %d from cmplx %d receiving from other master %d on cmplx %d and adding \n", masterdev, cmplxid, gmaster, k);
magma_queue_wait_event(streams[ masterdev ][ 0 ], redevents[gmaster][masterdev]);
magmablas_sgeadd(m, n, c_one,
&dwork2[masterdev][maxgsize*gmaster], m,
&dC[masterdev][0] , lddc );
} // end of gmaster!=-1
} // end of k!=cmplxid
}// for k = 0: nbcmplx
// because addition is done sequentially on stream 0,
// I have to record this to be able to synch using it
magma_event_record(redevents[masterdev][masterdev], streams[masterdev][0]);
// ========================================
// ========================================
// local broadcast of final results
// ========================================
for( magma_int_t l = 0; l < myngpu; ++l ) {
lcdev = gnode[cmplxid][l];
lccolsize = gpuisactive[lcdev];
if((lcdev!=masterdev)&&(lccolsize>0)){
// this is an active GPU of my real.
// wait the previous addition is done maening stream 0 is finished and broadcast sequentially for now.
// to make it parallel put stream lcdev instead of stream 0
//printf(" master %d broadcasting local to %d \n", masterdev, lcdev);
magma_queue_wait_event(streams[ masterdev ][ 0 ], redevents[masterdev][masterdev]);
magma_scopymatrix_async(
m, n,
&dC[masterdev][0], lddc,
&dC[lcdev][0], lddc, streams[masterdev][0] );
magma_event_record(redevents[masterdev][lcdev], streams[masterdev][0]);
}
}// for l=1:myngpu
// ========================================
}// end of if masterdev!=-1 maening real is active
}// for cmplxid
/*
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magma_device_sync();
}
*/
for( magma_int_t cmplxid = 0; cmplxid < nbcmplx; ++cmplxid ) {
myngpu = gnode[cmplxid][MagmaMaxGPUs];
masterdev = gnode[cmplxid][MagmaMaxGPUs+1];
//check if real is active
if(masterdev!=-1){
for( magma_int_t l = 0; l < myngpu; ++l ) {
lcdev = gnode[cmplxid][l];
lccolsize = gpuisactive[lcdev];
if(lccolsize>0){
magma_setdevice( lcdev );
magma_queue_wait_event(streams[ lcdev ][ 0 ], redevents[lcdev][0]);
magma_queue_wait_event(streams[ lcdev ][ 0 ], redevents[masterdev][lcdev]);
}
}// for l=1:myngpu
}// end of if masterdev!=-1 maening real is active
}// for cmplxid
//printf("****************************************************\n");
//printf(" finish ssymm \n");
//printf("****************************************************\n");
magma_setdevice( cdev );
magmablasSetKernelStream( cstream );
}
/**
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 */
/**
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_NOPIV_GPU computes an LU factorization of a general M-by-N
matrix A without any pivoting.
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 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]
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_nopiv_gpu(
magma_int_t m, magma_int_t n,
magmaFloat_ptr dA, magma_int_t ldda,
magma_int_t *info)
{
#define dA(i,j) (dA + (i)*nb + (j)*nb*ldda)
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, mindim;
magma_int_t i, rows, s, lddwork;
float *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_sgetrf_nb(m);
s = mindim / nb;
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
magma_smalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_sgetmatrix( m, n, dA, ldda, work, m );
magma_sgetrf_nopiv( m, n, work, m, info);
magma_ssetmatrix( m, n, work, m, dA, ldda );
magma_free_cpu(work);
}
else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
lddwork = maxm;
if (MAGMA_SUCCESS != magma_smalloc_pinned( &work, maxm*nb )) {
*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( i=0; i < s; i++ ) {
// download i-th panel
magma_queue_sync( stream[1] );
magma_sgetmatrix_async( m-i*nb, nb, dA(i,i), ldda, work, lddwork, stream[0] );
if ( i > 0 ) {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, n - (i+1)*nb,
c_one, dA(i-1,i-1), ldda,
dA(i-1,i+1), ldda );
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-i*nb, n-(i+1)*nb, nb,
c_neg_one, dA(i, i-1), ldda, dA(i-1,i+1), ldda,
c_one, dA(i, i+1), ldda );
}
// do the cpu part
rows = m - i*nb;
magma_queue_sync( stream[0] );
magma_sgetrf_nopiv( rows, nb, work, lddwork, &iinfo );
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + i*nb;
// upload i-th panel
magma_ssetmatrix_async( m-i*nb, nb, work, lddwork, dA(i, i), ldda, stream[0] );
magma_queue_sync( stream[0] );
// do the small non-parallel computations
if ( s > (i+1) ) {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dA(i, i ), ldda,
dA(i, i+1), ldda);
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-(i+1)*nb, nb, nb,
c_neg_one, dA(i+1, i ), ldda, dA(i, i+1), ldda,
c_one, dA(i+1, i+1), ldda );
}
else {
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb, n-s*nb,
c_one, dA(i, i ), ldda,
dA(i, i+1), ldda);
magma_sgemm( MagmaNoTrans, MagmaNoTrans,
m-(i+1)*nb, n-(i+1)*nb, nb,
c_neg_one, dA(i+1, i ), ldda, dA(i, i+1), ldda,
c_one, dA(i+1, i+1), ldda );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
rows = m - s*nb;
magma_sgetmatrix( rows, nb0, dA(s,s), ldda, work, lddwork );
// make sure that gpu queue is empty
magma_device_sync();
// do the cpu part
magma_sgetrf_nopiv( rows, nb0, work, lddwork, &iinfo );
if ( (*info == 0) && (iinfo > 0) )
*info = iinfo + s*nb;
// upload i-th panel
magma_ssetmatrix( rows, nb0, work, lddwork, dA(s,s), ldda );
magma_strsm( MagmaLeft, MagmaLower, MagmaNoTrans, MagmaUnit,
nb0, n-s*nb-nb0,
c_one, dA(s,s), ldda,
dA(s,s)+nb0, ldda);
magma_free_pinned( work );
magma_queue_destroy( stream[0] );
if (orig_stream == NULL) {
magma_queue_destroy( stream[1] );
}
magmablasSetKernelStream( orig_stream );
}
return *info;
} /* magma_sgetrf_nopiv_gpu */
/**
Purpose
-------
SPOTRF computes the Cholesky factorization of a real symmetric
positive definite matrix dA.
The factorization has the form
dA = U**T * U, if UPLO = MagmaUpper, or
dA = L * L**T, if UPLO = MagmaLower,
where U is an upper triangular matrix and L is lower triangular.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangle of dA is stored;
- = MagmaLower: Lower triangle of dA is stored.
@param[in]
n INTEGER
The order of the matrix dA. N >= 0.
@param[in,out]
dA REAL array on the GPU, dimension (LDDA,N)
On entry, the symmetric matrix dA. If UPLO = MagmaUpper, the leading
N-by-N upper triangular part of dA contains the upper
triangular part of the matrix dA, and the strictly lower
triangular part of dA is not referenced. If UPLO = MagmaLower, the
leading N-by-N lower triangular part of dA contains the lower
triangular part of the matrix dA, and the strictly upper
triangular part of dA is not referenced.
\n
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization dA = U**T * U or dA = L * L**T.
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,N).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
- > 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
@ingroup magma_sposv_comp
********************************************************************/
extern "C" magma_int_t
magma_spotrf2_mgpu(int num_gpus, magma_uplo_t uplo, magma_int_t m, magma_int_t n,
magma_int_t off_i, magma_int_t off_j, magma_int_t nb,
float **d_lA, magma_int_t ldda,
float **d_lP, magma_int_t lddp,
float *A, magma_int_t lda, magma_int_t h,
magma_queue_t stream[][3], magma_event_t event[][5],
magma_int_t *info )
{
#define Alo(i, j) (A + ((j)+off_j)*lda + (nb*(((i)/nb)%h)+off_i))
#define Aup(i, j) (A + (nb*(((j)/nb)%h)+off_j)*lda + (i+off_i))
#define dlA(id, i, j) (d_lA[(id)] + (j)*ldda + (i))
#define dlP(id, i, j, k) (d_lP[(id)] + (k)*nb*lddp + (j)*lddp + (i))
#define dlPT(id, i, j, k) (d_lP[(id)] + (k)*nb*lddp + (j)*nb + (i))
magma_int_t j, jb, nb0, nb2, dd, d, id, j_local, j_local2, buf;
float c_one = MAGMA_S_ONE;
float c_neg_one = MAGMA_S_NEG_ONE;
float d_one = 1.0;
float d_neg_one = -1.0;
int upper = (uplo == MagmaUpper);
float *dlpanel;
//magma_event_t event0[MagmaMaxGPUs], // syrk
// event1[MagmaMaxGPUs], // send off-diagonal
// event2[MagmaMaxGPUs], // send diagonal
// event3[MagmaMaxGPUs]; // trsm
magma_int_t n_local[MagmaMaxGPUs], ldpanel;
int stream0 = 0, stream1 = 1;
#ifdef STRSM_WORK
float *d_dinvA[MagmaMaxGPUs][2], *d_x[MagmaMaxGPUs][2]; /* used by strsm_work */
#endif
*info = 0;
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (!upper && num_gpus*ldda < max(1,n)) {
*info = -4;
} else if (upper && ldda < max(1,m)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
for( d=0; d < num_gpus; d++ ) {
/* local-n and local-ld */
if (upper) {
n_local[d] = ((n/nb)/num_gpus)*nb;
if (d < (n/nb)%num_gpus)
n_local[d] += nb;
else if (d == (n/nb)%num_gpus)
n_local[d] += n%nb;
} else {
n_local[d] = ((m/nb)/num_gpus)*nb;
if (d < (m/nb)%num_gpus)
n_local[d] += nb;
else if (d == (m/nb)%num_gpus)
n_local[d] += m%nb;
}
//magma_setdevice(d);
//magma_event_create( &event0[d] );
//magma_event_create( &event1[d] );
//magma_event_create( &event2[d] );
//magma_event_create( &event3[d] );
}
magma_setdevice(0);
/* == initialize the trace */
trace_init( 1, num_gpus, 3, (magma_queue_t*)stream );
/* Use blocked code. */
if (upper) {
/* ---------------------------------------------- */
/* Upper-triangular case */
/* > Compute the Cholesky factorization A = U'*U. */
/* ---------------------------------------------- */
#if defined(PRECISION_d) && defined(STRSM_WORK)
/* invert the diagonals
* Allocate device memory for the inversed diagonal blocks, size=m*NB
*/
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j < 2; j++ ) {
magma_smalloc( &d_dinvA[d][j], nb*nb );
magma_smalloc( &d_x[d][j], n*nb );
cudaMemset(d_dinvA[d][j], 0, nb*nb*sizeof(float));
cudaMemset(d_x[d][j], 0, n*nb*sizeof(float));
}
}
magma_setdevice(0);
#endif
for (j=0; j < m; j += nb) {
/* Set the GPU number that holds the current panel */
id = (j/nb)%num_gpus;
buf = (j/nb)%num_gpus;
/* Set the local index where the current panel is */
j_local = j/(nb*num_gpus);
jb = min(nb, (m-j));
if ( j > 0 ) {
/* needed on pluto... */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] ); // wait for the column on CPU
/* broadcast off-diagonal column to all gpus */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
if ( d != id ) {
magma_setdevice(d);
/* wait for it on CPU */
magma_queue_wait_event( stream[d][stream0], event[id][1] );
/* send it to GPU */
trace_gpu_start( d, stream0, "comm", "rows to GPUs" );
magma_ssetmatrix_async( j, jb,
Aup(0,j), lda,
dlP(d,jb,0,buf), lddp,
stream[d][stream0] );
trace_gpu_end( d, stream0 );
magma_event_record( event[d][1], stream[d][stream0] );
}
d = (d+1)%num_gpus;
}
}
/* Update the current diagonal block */
magma_setdevice(id);
if ( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
trace_gpu_start( id, stream1, "syrk", "syrk" );
magma_ssyrk(MagmaUpper, MagmaTrans, jb, j,
d_neg_one, dlA(id, 0, nb*j_local), ldda,
d_one, dlA(id, j, nb*j_local), ldda);
trace_gpu_end( id, stream1 );
magma_event_record( event[id][0], stream[id][stream1] );
}
/* send the diagonal to cpu */
magma_queue_wait_event( stream[id][stream0], event[id][0] ); // wait for syrk
trace_gpu_start( id, stream0, "comm", "D to CPU" );
magma_sgetmatrix_async( jb, jb,
dlA(id, j, nb*j_local), ldda,
Aup(j,j), lda,
stream[id][stream0] );
trace_gpu_end( id, stream0 );
if ( j > 0 ) {
/* Compute the local block column of the panel. */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
j_local2 = j_local+1;
if ( d > id ) j_local2 --;
nb0 = nb*j_local2;
if ( n_local[d] > nb0 ) {
/* wait for the off-diagonal */
if ( d != id ) {
//magma_queue_sync( stream[id][3] );
dlpanel = dlP(d, jb, 0, buf);
ldpanel = lddp;
/* wait for the offdiagonal column */
magma_queue_wait_event( stream[d][stream1], event[d][1] );
} else {
dlpanel = dlA(d, 0, nb*j_local);
ldpanel = ldda;
}
/* update the panel */
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
trace_gpu_start( d, stream1, "gemm", "gemm" );
magma_sgemm(MagmaTrans, MagmaNoTrans,
jb, n_local[d]-nb0, j,
c_neg_one, dlpanel, ldpanel,
dlA(d, 0, nb0), ldda,
c_one, dlA(d, j, nb0), ldda);
trace_gpu_end( d, stream1 );
}
d = (d+1)%num_gpus;
}
}
/* factor the diagonal */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] ); // wait for the diagonal
trace_cpu_start( 0, "getrf", "getrf" );
lapackf77_spotrf(MagmaUpperStr, &jb, Aup(j,j), &lda, info);
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus */
if ( (j+jb) < n) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
magma_setdevice(d);
if ( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d, 0, 0, buf);
ldpanel = lddp;
}
trace_gpu_start( d, stream0, "comm", "D to GPUs" );
magma_ssetmatrix_async( jb, jb,
Aup(j,j), lda,
dlpanel, ldpanel,
stream[d][stream0] );
trace_gpu_end( d, stream0 );
magma_event_record( event[d][2], stream[d][stream0] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
trace_gpu_start( id, stream0, "comm", "D to GPUs" );
magma_ssetmatrix_async( jb, jb,
Aup(j,j), lda,
dlA(id, j, nb*j_local), ldda,
stream[id][stream0] );
trace_gpu_end( id, stream0 );
}
/* panel-factorize the off-diagonal */
if ( (j+jb) < n) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
/* next column */
j_local2 = j_local+1;
if ( d > id ) j_local2--;
if ( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d, 0, 0, buf);
ldpanel = lddp;
}
nb2 = n_local[d]-nb*j_local2;
nb0 = min(nb, nb2 );
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for the diagonal
if ( j+jb < m && d == (j/nb+1)%num_gpus ) {
/* owns the next column, look-ahead the column */
trace_gpu_start( d, stream1, "trsm", "trsm" );
#if defined(PRECISION_d) && defined(STRSM_WORK)
magmablas_strsm_work( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
d_dinvA[d][0], d_x[d][0] );
/*nb2 = n_local[d] - j_local2*nb;
magmablas_strsm_work( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
d_dinvA[d], d_x[d] ); */
#else
/*nb2 = n_local[d] - j_local2*nb;
magma_strsm( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldda,
dlA(d, j, nb*j_local2), ldda);
*/
magma_strsm( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
trace_gpu_end( d, stream1 );
magma_event_record( event[d][3], stream[d][stream1] );
/* send the column to cpu */
if ( j+jb < m ) {
trace_gpu_start( d, stream0, "comm", "rows to CPU" );
magma_queue_wait_event( stream[d][stream0], event[d][3] ); // wait for lookahead
magma_sgetmatrix_async( (j+jb), nb0,
dlA(d, 0, nb*j_local2), ldda,
Aup(0,j+jb), lda,
stream[d][stream0] );
trace_gpu_end( d, stream0 );
magma_event_record( event[d][1], stream[d][stream0] );
}
/* update the remaining blocks */
nb2 = nb2 - nb0;
#if defined(PRECISION_d) && defined(STRSM_WORK)
magmablas_strsm_work( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_strsm( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda);
#endif
} else if ( nb2 > 0 ) {
/* update the entire trailing matrix */
trace_gpu_start( d, stream1, "trsm", "trsm" );
#if defined(PRECISION_d) && defined(STRSM_WORK)
magmablas_strsm_work( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_strsm( MagmaLeft, MagmaUpper, MagmaTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
trace_gpu_end( d, stream1 );
}
d = (d+1)%num_gpus;
}
} /* end of strsm */
} /* end of for j=1, .., n */
} else {
/* -------------------------------------------- */
/* Lower-triangular case */
/* Compute the Cholesky factorization A = L*L'. */
/* -------------------------------------------- */
#if defined(PRECISION_d) && defined(STRSM_WORK)
/*
* Allocate device memory for the inversed diagonal blocks, size=N*BLOCK_SIZE
*/
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j < 2; j++ ) {
magma_smalloc( &d_dinvA[d][j], nb*nb );
magma_smalloc( &d_x[d][j], nb*m );
cudaMemset(d_dinvA[d][j], 0, nb*nb*sizeof(float));
cudaMemset(d_x[d][j], 0, nb* m*sizeof(float));
}
}
magma_setdevice(0);
#endif
for (j=0; j < n; j += nb) {
/* Set the GPU number that holds the current panel */
id = (j/nb)%num_gpus;
buf = (j/nb)%num_gpus;
/* Set the local index where the current panel is */
j_local = j/(nb*num_gpus);
jb = min(nb, (n-j));
if ( j > 0 ) {
/* needed on pluto... */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] ); // wait for the column on CPU
/* broadcast offdiagonal row to all gpus */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
if ( d != id ) {
magma_setdevice(d);
/* wait for it on CPU */
magma_queue_wait_event( stream[d][stream0], event[id][1] );
/* send it to GPU */
magma_ssetmatrix_async( jb, j,
Alo(j,0), lda,
dlPT(d,0,jb,buf), nb,
stream[d][stream0] );
magma_event_record( event[d][1], stream[d][stream0] );
}
d = (d+1)%num_gpus;
}
}
/* Update the current diagonal block */
magma_setdevice(id);
if ( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
magma_ssyrk(MagmaLower, MagmaNoTrans, jb, j,
d_neg_one, dlA(id, nb*j_local, 0), ldda,
d_one, dlA(id, nb*j_local, j), ldda);
magma_event_record( event[id][0], stream[id][stream1] );
}
/* send the diagonal to cpu */
magma_queue_wait_event( stream[id][stream0], event[id][0] ); // wait for syrk
magma_sgetmatrix_async( jb, jb,
dlA(id, nb*j_local, j), ldda,
Alo(j,j), lda,
stream[id][stream0] );
/* update the offdiagonal blocks */
if ( j > 0 ) {
/* compute the block-rows of the panel */
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
j_local2 = j_local+1;
if ( d > id ) j_local2 --;
nb0 = nb*j_local2;
if ( nb0 < n_local[d] ) {
if ( d != id ) {
dlpanel = dlPT(d, 0, jb, buf);
ldpanel = nb;
/* wait for offdiagonal row */
magma_queue_wait_event( stream[d][stream1], event[d][1] );
} else {
dlpanel = dlA(d, nb*j_local, 0);
ldpanel = ldda;
}
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
magma_sgemm( MagmaNoTrans, MagmaTrans,
n_local[d]-nb0, jb, j,
c_neg_one, dlA(d, nb0, 0), ldda,
dlpanel, ldpanel,
c_one, dlA(d, nb0, j), ldda);
}
d = (d+1)%num_gpus;
}
}
/* factor the diagonal */
magma_setdevice(id);
magma_queue_sync( stream[id][stream0] );
lapackf77_spotrf(MagmaLowerStr, &jb, Alo(j,j), &lda, info);
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus */
if ( (j+jb) < m ) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
magma_setdevice(d);
if ( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d, 0, 0, buf);
ldpanel = nb;
}
magma_ssetmatrix_async( jb, jb,
Alo(j,j), lda,
dlpanel, ldpanel,
stream[d][stream0] );
magma_event_record( event[d][2], stream[d][stream0] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
magma_ssetmatrix_async( jb, jb,
Alo(j,j), lda,
dlA(id, nb*j_local, j), ldda,
stream[id][stream0] );
}
/* factorize off-diagonal blocks */
if ( (j+jb) < m ) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd < num_gpus; dd++ ) {
/* next column */
j_local2 = j_local+1;
if ( d > id ) j_local2--;
if ( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d, 0, 0, buf);
ldpanel = nb;
}
nb2 = n_local[d] - j_local2*nb;
nb0 = min(nb, nb2 );
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream1]);
magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for the diagonal
if ( j+jb < n && d == (j/nb+1)%num_gpus ) {
/* owns the next column, look-ahead the column */
#if defined(PRECISION_d) && defined(STRSM_WORK)
magmablas_strsm_work( MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
d_dinvA[d][0], d_x[d][0]);
#else
magma_strsm( MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
magma_event_record( event[d][3], stream[d][stream1] );
/* send the column to cpu */
if ( j+jb < n ) {
magma_queue_wait_event( stream[d][stream0], event[d][3] ); // wait for lookahead
magma_sgetmatrix_async( nb0, j+jb,
dlA(d, nb*j_local2, 0), ldda,
Alo(j+jb,0), lda,
stream[d][stream0] );
magma_event_record( event[d][1], stream[d][stream0] );
}
/* update the remaining blocks */
nb2 = nb2 - nb0;
#if defined(PRECISION_d) && defined(STRSM_WORK)
magmablas_strsm_work( MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_strsm( MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda);
#endif
} else if ( nb2 > 0 ) {
/* update the entire trailing matrix */
#if defined(PRECISION_d) && defined(STRSM_WORK)
magmablas_strsm_work( MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_strsm( MagmaRight, MagmaLower, MagmaTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
}
d = (d+1)%num_gpus;
}
}
}
} /* end of else not upper */
/* == finalize the trace == */
trace_finalize( "spotrf.svg", "trace.css" );
/* clean up */
for( d=0; d < num_gpus; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
//magma_event_destroy( event0[d] );
//magma_event_destroy( event1[d] );
//magma_event_destroy( event2[d] );
//magma_event_destroy( event3[d] );
}
magma_setdevice(0);
return *info;
} /* magma_spotrf_mgpu */
/**
Purpose
-------
SPOTRF computes the Cholesky factorization of a real symmetric
positive definite matrix dA.
The factorization has the form
dA = U**H * U, if UPLO = MagmaUpper, or
dA = L * L**H, if UPLO = MagmaLower,
where U is an upper triangular matrix and L is lower triangular.
This is the block version of the algorithm, calling Level 3 BLAS.
Arguments
---------
@param[in]
uplo magma_uplo_t
- = MagmaUpper: Upper triangle of dA is stored;
- = MagmaLower: Lower triangle of dA is stored.
@param[in]
n INTEGER
The order of the matrix dA. N >= 0.
@param[in,out]
d_lA REAL array of pointers on the GPU, dimension (ngpu)
On entry, the symmetric matrix dA distributed over GPUs
(dl_A[d] points to the local matrix on the d-th GPU).
It is distributed in 1D block column or row cyclic (with the
block size of nb) if UPLO = MagmaUpper or MagmaLower, respectively.
If UPLO = MagmaUpper, the leading N-by-N upper triangular
part of dA contains the upper triangular part of the matrix dA,
and the strictly lower triangular part of dA is not referenced.
If UPLO = MagmaLower, the leading N-by-N lower triangular part
of dA contains the lower triangular part of the matrix dA, and
the strictly upper triangular part of dA is not referenced.
\n
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization dA = U**H * U or dA = L * L**H.
@param[in]
ldda INTEGER
The leading dimension of the array dA. LDDA >= max(1,N).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
- > 0: if INFO = i, the leading minor of order i is not
positive definite, and the factorization could not be
completed.
@ingroup magma_sposv_comp
********************************************************************/
extern "C" magma_int_t
magma_spotrf_mgpu_right(
magma_int_t ngpu,
magma_uplo_t uplo, magma_int_t n,
magmaFloat_ptr d_lA[], magma_int_t ldda,
magma_int_t *info )
{
#define dlA(id, i, j) (d_lA[(id)] + (j) * ldda + (i))
#define dlP(id, i, j) (d_lP[(id)] + (j) * ldda + (i))
#define panel(j) (panel + (j))
#define tmppanel(j) (tmppanel + (j))
#define tmpprevpanel(j) (tmpprevpanel + (j))
#define STREAM_ID(i) (nqueue > 1 ? 1+((i)/nb)%(nqueue-1) : 0)
float z_one = MAGMA_S_MAKE( 1.0, 0.0 );
float mz_one = MAGMA_S_MAKE( -1.0, 0.0 );
float one = 1.0;
float m_one = -1.0;
const char* uplo_ = lapack_uplo_const( uplo );
magma_int_t j, nb, d, id, j_local, blkid, crosspoint, prevtrsmrows=0, nqueue = 5;
float *panel, *tmppanel0, *tmppanel1, *tmppanel, *tmpprevpanel;
float *d_lP[MagmaMaxGPUs], *dlpanel, *dlpanels[MagmaMaxGPUs];
magma_int_t rows, trsmrows, igpu, n_local[MagmaMaxGPUs], ldpanel;
magma_queue_t queues[MagmaMaxGPUs][10];
*info = 0;
if ( uplo != MagmaUpper && uplo != MagmaLower ) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (ldda < max(1,n)) {
*info = -4;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
nb = magma_get_spotrf_nb(n);
ldpanel = ldda;
magma_setdevice(0);
if (MAGMA_SUCCESS != magma_smalloc_pinned( &panel, 2 * nb * ldpanel )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
tmppanel0 = panel;
tmppanel1 = tmppanel0 + nb * ldpanel;
if ((nb <= 1) || (nb >= n)) {
// Use unblocked code.
magma_sgetmatrix( n, n, dlA(0, 0, 0), ldda, panel, ldpanel);
lapackf77_spotrf( uplo_, &n, panel, &ldpanel, info);
magma_ssetmatrix( n, n, panel, ldpanel, dlA(0, 0, 0), ldda );
} else {
for( d = 0; d < ngpu; d++ ) {
// local-n and local-ld
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;
magma_setdevice(d);
magma_device_sync();
if (MAGMA_SUCCESS != magma_smalloc( &d_lP[d], nb * ldda )) {
for( j = 0; j < d; j++ ) {
magma_setdevice(j);
magma_free( d_lP[d] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
for( j=0; j < nqueue; j++ ) {
magma_queue_create( &queues[d][j] );
}
}
//#define ENABLE_TIMER
#if defined (ENABLE_TIMER)
real_Double_t therk[4], tmtc, tcchol, tctrsm, tctm, tmnp, tcnp;
real_Double_t ttot_herk[4] = {0,0,0,0}, ttot_mtc = 0, ttot_cchol = 0, ttot_ctrsm = 0, ttot_ctm = 0, ttot_mnp = 0, ttot_cnp = 0;
printf("\n\n %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s %10s\n",
"j", "nb", "row", "mtc", "CPU_np", "panel", "ctrsm", "CH+TRSM", "CPU", "dsyrk[0]", "dsyrk[1]", "dsyrk[2]", "dsyrk[3]", "ctm P", "gpu_np");
printf(" ====================================================================================================\n");
#endif
// Use blocked code.
if (uplo == MagmaUpper) {
printf( " === not supported, yet ===\n" );
} else {
blkid = -1;
if (ngpu == 4)
crosspoint = n;
else if (ngpu == 3)
crosspoint = n;
else if (ngpu == 2)
crosspoint = 20160;
else
crosspoint = 0;
crosspoint = 0; //n; //n -- > gpu always does next panel, 0 --> cpu always does next panel
crosspoint = n;
#if defined (ENABLE_TIMER)
real_Double_t tget = magma_wtime(), tset = 0.0, ttot = 0.0;
#endif
if ( n > nb ) {
// send first panel to cpu
magma_setdevice(0);
tmppanel = tmppanel0;
magma_sgetmatrix_async(n, nb,
dlA(0, 0, 0), ldda,
tmppanel(0), ldpanel,
queues[0][0] );
}
#if defined (ENABLE_TIMER)
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_device_sync();
}
tget = magma_wtime()-tget;
#endif
// Compute the Cholesky factorization A = L*L'
for (j = 0; (j + nb) < n; j += nb) {
#if defined (ENABLE_TIMER)
therk[0] = therk[1] = therk[2] = therk[3] = tmtc = tcchol = tctrsm = tctm = tmnp = tcnp = 0.0;
#endif
blkid += 1;
tmppanel = (blkid % 2 == 0) ? tmppanel0 : tmppanel1;
// Set the gpu number that holds the current panel
id = (j / nb) % ngpu;
magma_setdevice(id);
// Set the local index where the current panel is
j_local = j / (nb * ngpu) * nb;
rows = n - j;
// Wait for the panel on cpu
magma_queue_sync( queues[id][0] );
if (j > 0 && prevtrsmrows > crosspoint) {
#if defined (ENABLE_TIMER)
tcnp = magma_wtime();
#endif
tmpprevpanel = ((blkid - 1) % 2) == 0 ? tmppanel0 : tmppanel1;
blasf77_sgemm( MagmaNoTransStr, MagmaConjTransStr,
&rows, &nb, &nb,
&mz_one, tmpprevpanel(j), &ldpanel,
tmpprevpanel(j), &ldpanel,
&z_one, tmppanel(j), &ldpanel );
#if defined (ENABLE_TIMER)
tcnp = magma_wtime() - tcnp;
ttot_cnp += tcnp;
#endif
}
#if defined (ENABLE_TIMER)
tcchol = magma_wtime();
#endif
lapackf77_spotrf(MagmaLowerStr, &nb, tmppanel(j), &ldpanel, info);
if (*info != 0) {
*info = *info + j;
break;
}
#if defined (ENABLE_TIMER)
tcchol = magma_wtime() - tcchol;
ttot_cchol += tcchol;
tctrsm = magma_wtime();
#endif
trsmrows = rows - nb;
if (trsmrows > 0) {
blasf77_strsm(MagmaRightStr, MagmaLowerStr, MagmaConjTransStr, MagmaNonUnitStr,
&trsmrows, &nb,
&z_one, tmppanel(j), &ldpanel,
tmppanel(j + nb), &ldpanel);
}
#if defined (ENABLE_TIMER)
tctrsm = magma_wtime() - tctrsm;
ttot_ctrsm += tctrsm;
tctm = magma_wtime();
#endif
d = (id + 1) % ngpu;
// send current panel to gpus
for (igpu = 0; igpu < ngpu; igpu++, d = (d + 1) % ngpu ) {
magma_int_t myrows = 0;
magma_int_t row_offset = 0;
if ( d == id ) {
dlpanel = dlA(d, j, j_local);
myrows = rows;
row_offset = 0;
} else {
dlpanel = dlP(d, 0, 0);
myrows = trsmrows;
row_offset = nb;
}
if (myrows > 0) {
magma_setdevice(d);
magma_ssetmatrix_async(myrows, nb,
tmppanel(j + row_offset), ldpanel,
dlpanel, ldda, queues[d][0] );
}
}
/* make sure panel is on GPUs */
d = (id + 1) % ngpu;
for (igpu = 0; igpu < ngpu; igpu++, d = (d + 1) % ngpu ) {
magma_setdevice(d);
magma_queue_sync( queues[d][0] );
}
#if defined (ENABLE_TIMER)
tctm = magma_wtime() - tctm;
ttot_ctm += tctm;
#endif
if ( (j + nb) < n) {
magma_int_t offset = 0;
magma_int_t row_offset = 0;
if (j + nb + nb < n) {
d = (id + 1) % ngpu;
magma_setdevice(d);
magma_int_t j_local2 = (j + nb) / (nb * ngpu) * nb;
if (trsmrows <= crosspoint) {
#if defined (ENABLE_TIMER)
tmnp = magma_wtime();
#endif
// do gemm on look ahead panel
if ( d == id ) {
dlpanel = dlA(d, j + nb, j_local);
} else {
dlpanel = dlP(d, 0, 0);
}
magmablasSetKernelStream( queues[d][STREAM_ID(j_local2)] );
#define SSYRK_ON_DIAG
#ifdef SSYRK_ON_DIAG
magma_ssyrk( MagmaLower, MagmaNoTrans,
nb, nb,
m_one, dlpanel, ldda,
one, dlA(d, j + nb, j_local2), ldda);
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
trsmrows-nb, nb, nb,
mz_one, dlpanel+nb, ldda,
dlpanel, ldda,
z_one, dlA(d, j + nb +nb, j_local2), ldda);
#else
magma_sgemm( MagmaNoTrans, MagmaConjTrans,
trsmrows, nb, nb,
mz_one, dlpanel, ldda,
dlpanel, ldda,
z_one, dlA(d, j + nb, j_local2), ldda);
#endif
#if defined (ENABLE_TIMER)
magma_device_sync();
tmnp = magma_wtime() - tmnp;
ttot_mnp += tmnp;
#endif
}
// send next panel to cpu
magma_queue_sync( queues[d][STREAM_ID(j_local2)] ); // make sure lookahead is done
tmppanel = ((blkid+1) % 2 == 0) ? tmppanel0 : tmppanel1;
magma_sgetmatrix_async(rows-nb, nb,
dlA(d, j+nb, j_local2), ldda,
tmppanel(j+nb), ldpanel,
queues[d][0] );
tmppanel = (blkid % 2 == 0) ? tmppanel0 : tmppanel1;
offset = j + nb + nb;
row_offset = nb;
} else {
offset = j + nb;
row_offset = 0;
}
if (n - offset > 0) {
// syrk on multiple gpu
for (d = 0; d < ngpu; d++ ) {
if ( d == id ) {
dlpanels[d] = dlA(d, j + nb + row_offset, j_local);
} else {
dlpanels[d] = dlP(d, row_offset, 0);
}
}
#if defined (ENABLE_TIMER)
for( d=0; d < ngpu; d++ ) therk[d] = magma_wtime();
#endif
//magmablasSetKernelStream( queues[d] );
//magma_ssyrk(MagmaLower, MagmaNoTrans, n - offset, nb,
// m_one, dlpanel, ldda,
// one, &d_lA[d][offset + offset*ldda], ldda );
#ifdef SSYRK_ON_DIAG
magma_ssyrk_mgpu
#else
magma_ssyrk_mgpu2
#endif
(ngpu, MagmaLower, MagmaNoTrans,
nb, n - offset, nb,
m_one, dlpanels, ldda, 0,
one, d_lA, ldda, offset,
nqueue, queues );
#if defined (ENABLE_TIMER)
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_device_sync();
therk[d] = magma_wtime() - therk[d];
ttot_herk[d] += therk[d];
}
#endif
}
prevtrsmrows = trsmrows;
#if defined (ENABLE_TIMER)
ttot += (tcnp+tcchol+tctrsm+therk[0]+therk[1]+therk[2]+tctm+tmnp);
printf("%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(%d) %10.3lf\n",
j, nb, rows, tmtc,
tcnp, // gemm
tcchol, // potrf
tctrsm, // trsm
(tcchol + tctrsm),
(tmtc+tcnp+tcchol+tctrsm),
therk[0], therk[1], therk[2], therk[3], // syrk
tctm, // copy panel to GPU
tmnp, // lookahead on GPU
(id + 1) % ngpu,
(tcnp+tcchol+tctrsm+therk[0]+therk[1]+therk[2]+tctm+tmnp));
fflush(0);
#endif
}
}
for( d = 0; d < ngpu; d++ ) {
magma_setdevice(d);
for( id=0; id < nqueue; id++ ) {
magma_queue_sync( queues[d][id] );
}
}
#if defined (ENABLE_TIMER)
printf("\n%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(-) %10.3lf\n",
n, n, 0, ttot_mtc,
ttot_cnp, // gemm
ttot_cchol, // potrf
ttot_ctrsm, // trsm
(ttot_cchol + ttot_ctrsm),
(ttot_mtc+ttot_cnp+ttot_cchol+ttot_ctrsm),
ttot_herk[0], ttot_herk[1], ttot_herk[2], ttot_herk[3], // syrk
ttot_ctm, // copy panel to GPU
ttot_mnp, // lookahead on GPU
(ttot_cnp+ttot_cchol+ttot_ctrsm+ttot_herk[0]+ttot_herk[1]+ttot_herk[2]+ttot_ctm+ttot_mnp));
printf("%10d %10d %10d %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf %10.3lf(-) %10.3lf (ratio)\n",
n, n, 0, ttot_mtc/ttot,
ttot_cnp/ttot, // gemm
ttot_cchol/ttot, // potrf
ttot_ctrsm/ttot, // trsm
(ttot_cchol + ttot_ctrsm)/ttot,
(ttot_mtc+ttot_cnp+ttot_cchol+ttot_ctrsm)/ttot,
ttot_herk[0]/ttot, ttot_herk[1]/ttot, ttot_herk[2]/ttot, ttot_herk[3]/ttot, // syrk
ttot_ctm/ttot, // copy panel to GPU
ttot_mnp/ttot, // lookahead on GPU
(ttot_cnp+ttot_cchol+ttot_ctrsm+ttot_herk[0]+ttot_herk[1]+ttot_herk[2]+ttot_ctm+ttot_mnp)/ttot);
#endif
// cholesky for the last block
if (j < n && *info == 0) {
rows = n - j;
id = (j / nb) % ngpu;
// Set the local index where the current panel is
j_local = j / (nb * ngpu) * nb;
magma_setdevice(id);
#if defined (ENABLE_TIMER)
tset = magma_wtime();
#endif
magma_sgetmatrix(rows, rows, dlA(id, j, j_local), ldda, panel(j), ldpanel);
lapackf77_spotrf(MagmaLowerStr, &rows, panel(j), &ldpanel, info);
magma_ssetmatrix(rows, rows, panel(j), ldpanel, dlA(id, j, j_local), ldda);
#if defined (ENABLE_TIMER)
tset = magma_wtime() - tset;
#endif
}
#if defined (ENABLE_TIMER)
printf( " matrix_get,set: %10.3lf %10.3lf -> %10.3lf\n",tget,tset,ttot+tget+tset );
#endif
} // end of else not upper
// clean up
for( d = 0; d < ngpu; d++ ) {
magma_setdevice(d);
for( j=0; j < nqueue; j++ ) {
magma_queue_destroy( queues[d][j] );
}
magma_free( d_lP[d] );
}
} // end of not lapack
// free workspace
magma_free_pinned( panel );
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_spotrf_mgpu_right */