Tensor_hao<complex<double>,2> inverse_magma(const LUDecomp<complex<double>>& x)
{
magma_int_t N=x.A.rank(0); magma_int_t info;
magmaDoubleComplex_ptr d_A , dwork;
magma_int_t lda, ldwork;
lda = ((N+31)/32)*32; //round up to multiple of 32 for best GPU performance
ldwork = N*magma_get_zgetri_nb(N); // magma_get_zgetri_nb optimizes the blocksize
magma_zmalloc( &d_A, lda*N ); magma_zmalloc( &dwork, ldwork );
//copy matrix from CPU to GPU
magma_zsetmatrix( N, N, (magmaDoubleComplex* )x.A.data(), N, d_A, lda );
//calculate the inverse matrix with zgetri
magma_zgetri_gpu( N, d_A, lda, (magma_int_t*) x.ipiv.data(), dwork, ldwork, &info );
if(info<0) {cout<<"The "<<info<<"-th parameter is illegal in inverse_magma!"<<endl; exit(1);}
//copy matrix from GPU to CPU
Tensor_hao<complex<double>,2> A(N,N);
magma_zgetmatrix( N, N, d_A, lda, (magmaDoubleComplex* )A.data(), N );
magma_free(d_A); magma_free(dwork);
return A;
}
LUDecomp<complex<double>> LUconstruct_magma(const Tensor_core<complex<double>,2>& x)
{
if( x.rank(0) != x.rank(1) ) {cout<<"Input for LU is not square matrix!"<<endl; exit(1);}
//Create LU object
LUDecomp<complex<double>> y;
y.A = Tensor_hao< complex<double>, 2 > ( x.n_ptr() );
y.ipiv = Tensor_hao<int,1>( x.rank(0) );
//Prepare for zgetrf
magma_int_t M = x.rank(0), N = x.rank(1);
magma_int_t LDA = ((M+31)/32)*32;
magmaDoubleComplex_ptr d_A; magma_zmalloc(&d_A, LDA*N);
magma_int_t info;
//Transfer data and call zgetrf
magma_zsetmatrix(M, N, (magmaDoubleComplex* ) x.data(), M, d_A, LDA );
magma_zgetrf_gpu(M, N, d_A, LDA, (magma_int_t*) y.ipiv.data(), &info);
magma_zgetmatrix(M, N, d_A, LDA, (magmaDoubleComplex* ) y.A.data(), M);
y.info=info;
//Clean
magma_free(d_A);
if(y.info<0) {cout<<"The "<<y.info<<"-th parameter is illegal in LUconstruct_magma!"<<endl; exit(1);}
return y;
}
// ------------------------------------------------------------
// Solve dA * dX = dB, where dA and dX are stored in GPU device memory.
// Internally, MAGMA uses a hybrid CPU + GPU algorithm.
void gpu_interface( magma_int_t n, magma_int_t nrhs )
{
magmaDoubleComplex *dA=NULL, *dX=NULL;
magma_int_t *ipiv=NULL;
magma_int_t ldda = magma_roundup( n, 32 ); // round up to multiple of 32 for best GPU performance
magma_int_t lddx = ldda;
magma_int_t info = 0;
magma_queue_t queue=NULL;
// magma_*malloc routines for GPU memory are type-safe,
// but you can use cudaMalloc if you prefer.
magma_zmalloc( &dA, ldda*n );
magma_zmalloc( &dX, lddx*nrhs );
magma_imalloc_cpu( &ipiv, n ); // ipiv always on CPU
if ( dA == NULL || dX == NULL || ipiv == NULL ) {
fprintf( stderr, "malloc failed\n" );
goto cleanup;
}
magma_int_t dev = 0;
magma_queue_create( dev, &queue );
// Replace these with your code to initialize A and X
zfill_matrix_gpu( n, n, dA, ldda, queue );
zfill_rhs_gpu( n, nrhs, dX, lddx, queue );
magma_zgesv_gpu( n, 1, dA, ldda, ipiv, dX, ldda, &info );
if ( info != 0 ) {
fprintf( stderr, "magma_zgesv_gpu failed with info=%d\n", info );
}
// TODO: use result in dX
cleanup:
magma_queue_destroy( queue );
magma_free( dA );
magma_free( dX );
magma_free_cpu( ipiv );
}
Tensor_hao<complex<double>,2> solve_lineq_magma(const LUDecomp<complex<double>>& x, const Tensor_core<complex<double>,2>& B, char TRANS)
{
if( x.A.rank(0) != B.rank(0) ) {cout<<"Input size for solving linear equation is not consistent!"<<endl; exit(1);}
magma_int_t N=B.rank(0); magma_int_t NRHS=B.rank(1); magma_int_t info;
magma_trans_t Trans = magma_trans_const(TRANS);
magmaDoubleComplex_ptr d_A, d_B;
magma_int_t lda, ldb;
lda = ((N+31)/32)*32;
ldb = ((N+31)/32)*32;
//allocate memory on GPU
magma_zmalloc( &d_A, lda*N );
magma_zmalloc( &d_B, ldb*NRHS );
//copy matrix from CPU to GPU
magma_zsetmatrix( N, N, (magmaDoubleComplex* )x.A.data(), N, d_A, lda );
magma_zsetmatrix( N, NRHS, (magmaDoubleComplex* )B.data(), N, d_B, ldb );
//Solve the equation
magma_zgetrs_gpu( Trans, N, NRHS, d_A, lda, (magma_int_t*)x.ipiv.data(), d_B, ldb, &info );
if(info!=0)
{
cout<<"Solve linear equation is not suceesful: "<<info<<"-th parameter is illegal!"<<endl;
exit(1);
}
//copy matrix from GPU to CPU
Tensor_hao<complex<double>,2> M(N,NRHS);
magma_zgetmatrix( N, NRHS, d_B, ldb, (magmaDoubleComplex* ) M.data(), N );
//free memory
magma_free( d_A ); magma_free( d_B );
return M;
}
void gmm_magma(const Tensor_core<complex<double>,2>& A, const Tensor_core<complex<double>,2>& B, Tensor_core<complex<double>,2>& C,
char TRANSA, char TRANSB, complex<double> alpha, complex<double> 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);
magmaDoubleComplex_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_zmalloc(&d_A, LDA*AL1 );
magma_zmalloc(&d_B, LDB*BL1 );
magma_zmalloc(&d_C, LDC*CL1 );
// Copy data from host (CPU) to device (GPU)
magma_zsetmatrix( AL0, AL1, (magmaDoubleComplex* ) A.data(), AL0, d_A, LDA );
magma_zsetmatrix( BL0, BL1, (magmaDoubleComplex* ) B.data(), BL0, d_B, LDB );
if( abs(beta)>1e-32 ) magma_zsetmatrix( CL0, CL1, (magmaDoubleComplex* ) 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_zgemm(transA, transB, M, N, K, _cast_Z(alpha), d_A, LDA, d_B, LDB, _cast_Z(beta),d_C, LDC);
// Copy solution from device (GPU) to host (CPU)
magma_zgetmatrix(CL0, CL1, d_C, LDC, (magmaDoubleComplex* ) C.data(), CL0);
// Free memory on GPU
magma_free(d_A); magma_free(d_B); magma_free(d_C);
}
extern "C" magma_int_t
magma_zcuspaxpy(
magmaDoubleComplex *alpha, magma_z_sparse_matrix A,
magmaDoubleComplex *beta, magma_z_sparse_matrix B,
magma_z_sparse_matrix *AB,
magma_queue_t queue )
{
if ( A.memory_location == Magma_DEV
&& B.memory_location == Magma_DEV
&& ( A.storage_type == Magma_CSR ||
A.storage_type == Magma_CSRCOO )
&& ( B.storage_type == Magma_CSR ||
B.storage_type == Magma_CSRCOO ) ) {
magma_z_sparse_matrix C;
C.num_rows = A.num_rows;
C.num_cols = A.num_cols;
C.storage_type = A.storage_type;
C.memory_location = A.memory_location;
magma_int_t stat_dev = 0;
C.val = NULL;
C.col = NULL;
C.row = NULL;
C.rowidx = NULL;
C.blockinfo = NULL;
C.diag = NULL;
C.dval = NULL;
C.dcol = NULL;
C.drow = NULL;
C.drowidx = NULL;
C.ddiag = NULL;
// CUSPARSE context //
cusparseHandle_t handle;
cusparseStatus_t cusparseStatus;
cusparseStatus = cusparseCreate(&handle);
cusparseSetStream( handle, queue );
if (cusparseStatus != 0) printf("error in Handle.\n");
cusparseMatDescr_t descrA;
cusparseMatDescr_t descrB;
cusparseMatDescr_t descrC;
cusparseStatus = cusparseCreateMatDescr(&descrA);
cusparseStatus = cusparseCreateMatDescr(&descrB);
cusparseStatus = cusparseCreateMatDescr(&descrC);
if (cusparseStatus != 0) printf("error in MatrDescr.\n");
cusparseStatus =
cusparseSetMatType(descrA,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrB,CUSPARSE_MATRIX_TYPE_GENERAL);
cusparseSetMatType(descrC,CUSPARSE_MATRIX_TYPE_GENERAL);
if (cusparseStatus != 0) printf("error in MatrType.\n");
cusparseStatus =
cusparseSetMatIndexBase(descrA,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrB,CUSPARSE_INDEX_BASE_ZERO);
cusparseSetMatIndexBase(descrC,CUSPARSE_INDEX_BASE_ZERO);
if (cusparseStatus != 0) printf("error in IndexBase.\n");
// multiply A and B on the device
magma_int_t baseC;
// nnzTotalDevHostPtr points to host memory
magma_index_t *nnzTotalDevHostPtr = (magma_index_t*) &C.nnz;
cusparseSetPointerMode(handle, CUSPARSE_POINTER_MODE_HOST);
stat_dev += magma_index_malloc( &C.drow, (A.num_rows + 1) );
cusparseXcsrgeamNnz(handle,A.num_rows, A.num_cols,
descrA, A.nnz, A.drow, A.dcol,
descrB, B.nnz, B.drow, B.dcol,
descrC, C.row, nnzTotalDevHostPtr);
if (NULL != nnzTotalDevHostPtr) {
C.nnz = *nnzTotalDevHostPtr;
} else {
// workaround as nnz and base C are magma_int_t
magma_index_t base_t, nnz_t;
magma_index_getvector( 1, C.drow+C.num_rows, 1, &nnz_t, 1 );
magma_index_getvector( 1, C.drow, 1, &base_t, 1 );
C.nnz = (magma_int_t) nnz_t;
baseC = (magma_int_t) base_t;
C.nnz -= baseC;
}
stat_dev += magma_index_malloc( &C.dcol, C.nnz );
stat_dev += magma_zmalloc( &C.dval, C.nnz );
if( stat_dev != 0 ) {
magma_z_mfree( &C, queue );
return MAGMA_ERR_DEVICE_ALLOC;
}
cusparseZcsrgeam(handle, A.num_rows, A.num_cols,
alpha,
descrA, A.nnz,
A.dval, A.drow, A.dcol,
beta,
descrB, B.nnz,
B.dval, B.drow, B.dcol,
descrC,
C.dval, C.drow, C.dcol);
cusparseDestroyMatDescr( descrA );
cusparseDestroyMatDescr( descrB );
cusparseDestroyMatDescr( descrC );
cusparseDestroy( handle );
// end CUSPARSE context //
magma_z_mtransfer( C, AB, Magma_DEV, Magma_DEV, queue );
magma_z_mfree( &C, queue );
return MAGMA_SUCCESS;
}
else {
printf("error: CSRSPAXPY only supported on device and CSR format.\n");
return MAGMA_SUCCESS;
}
}
/**
Purpose
-------
ZUNMQL overwrites the general complex M-by-N matrix C with
@verbatim
SIDE = MagmaLeft SIDE = MagmaRight
TRANS = MagmaNoTrans: Q * C C * Q
TRANS = Magma_ConjTrans: Q**H * C C * Q**H
@endverbatim
where Q is a complex unitary matrix defined as the product of k
elementary reflectors
Q = H(k) . . . H(2) H(1)
as returned by ZGEQLF.
Q is of order M if SIDE = MagmaLeft
and of order N if SIDE = MagmaRight.
Arguments
---------
@param[in]
side magma_side_t
- = MagmaLeft: apply Q or Q**H from the Left;
- = MagmaRight: apply Q or Q**H from the Right.
@param[in]
trans magma_trans_t
- = MagmaNoTrans: No transpose, apply Q;
- = Magma_ConjTrans: Conjugate transpose, apply Q**H.
@param[in]
m INTEGER
The number of rows of the matrix C. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix C. N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines
the matrix Q.
If SIDE = MagmaLeft, M >= K >= 0;
if SIDE = MagmaRight, N >= K >= 0.
@param[in,out]
dA COMPLEX_16 array on the GPU, dimension (LDDA,K)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,k, as returned by
ZGEQLF in the last k columns of its array argument dA.
The diagonal and the lower part
are destroyed, the reflectors are not modified.
@param[in]
ldda INTEGER
The leading dimension of the array dA.
If SIDE = MagmaLeft, LDDA >= max(1,M);
if SIDE = MagmaRight, LDDA >= max(1,N).
@param[in]
tau COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQLF.
@param[in,out]
dC COMPLEX_16 array on the GPU, dimension (LDDC,N)
On entry, the M-by-N matrix C.
On exit, C is overwritten by (Q*C) or (Q**H * C) or (C * Q**H) or (C*Q).
@param[in]
lddc INTEGER
The leading dimension of the array dC. LDDC >= max(1,M).
@param[in]
wA COMPLEX_16 array, dimension
(LDWA,M) if SIDE = MagmaLeft
(LDWA,N) if SIDE = MagmaRight
The vectors which define the elementary reflectors, as
returned by ZHETRD_GPU.
(A copy of the upper or lower part of dA, on the host.)
@param[in]
ldwa INTEGER
The leading dimension of the array wA.
If SIDE = MagmaLeft, LDWA >= max(1,M);
if SIDE = MagmaRight, LDWA >= max(1,N).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_zgeqlf_comp
********************************************************************/
extern "C" magma_int_t
magma_zunmql2_gpu(
magma_side_t side, magma_trans_t trans,
magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex_ptr dA, magma_int_t ldda,
magmaDoubleComplex *tau,
magmaDoubleComplex_ptr dC, magma_int_t lddc,
const magmaDoubleComplex *wA, magma_int_t ldwa,
magma_int_t *info)
{
#define dA(i_,j_) (dA + (i_) + (j_)*ldda)
#define dC(i_,j_) (dC + (i_) + (j_)*lddc)
#define wA(i_,j_) (wA + (i_) + (j_)*ldwa)
/* Constants */
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const magma_int_t nbmax = 64;
/* Local variables */
magmaDoubleComplex_ptr dwork = NULL, dT = NULL;
magmaDoubleComplex T[ nbmax*nbmax ];
magma_int_t i, i1, i2, step, ib, lddwork, nb, mi, ni, nq, nq_i, nw;
magma_queue_t queue = NULL;
// Parameter adjustments for Fortran indexing
wA -= 1 + ldwa;
dC -= 1 + lddc;
--tau;
*info = 0;
bool left = (side == MagmaLeft);
bool notran = (trans == MagmaNoTrans);
/* NQ is the order of Q and NW is the minimum dimension of WORK */
if (left) {
nq = m;
nw = n;
} else {
nq = n;
nw = m;
}
/* Test the input arguments */
if (! left && side != MagmaRight) {
*info = -1;
} else if (! notran && trans != Magma_ConjTrans) {
*info = -2;
} else if (m < 0) {
*info = -3;
} else if (n < 0) {
*info = -4;
} else if (k < 0 || k > nq) {
*info = -5;
} else if (ldda < max(1,nq)) {
*info = -7;
} else if (lddc < max(1,m)) {
*info = -10;
} else if (ldwa < max(1,nq)) {
*info = -12;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0 || k == 0) {
return *info;
}
// size of the block
nb = nbmax;
lddwork = nw;
/* Use hybrid CPU-GPU code */
if ( ( left && notran) ||
(! left && ! notran) )
{
i1 = 1;
i2 = k;
step = nb;
} else {
i1 = ((k - 1)/nb)*nb + 1;
i2 = 1;
step = -nb;
}
// silence "uninitialized" warnings
mi = 0;
ni = 0;
if (left) {
ni = n;
} else {
mi = m;
}
// dwork is (n or m) x nb + nb x nb, for left or right respectively
if (MAGMA_SUCCESS != magma_zmalloc( &dwork, lddwork*nb + nb*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto cleanup;
}
dT = dwork + lddwork*nb;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
// in bottom k x k portion of dA,
// set nb-1 sub-diagonals to 0, and diagonal to 1, in
// This way we can copy V directly to the GPU,
// with the lower triangle parts already set to identity.
// A is nq x k, either m x k (left) or n x k (right)
magmablas_zlaset_band( MagmaLower, k, k, nb, c_zero, c_one, dA(nq-k,0), ldda, queue );
for (i = i1; (step < 0 ? i >= i2 : i <= i2); i += step) {
ib = min( nb, k - i + 1 );
/* Form the triangular factor of the block reflector
H = H(i+ib-1) . . . H(i+1) H(i) */
nq_i = nq - k + i + ib - 1;
lapackf77_zlarft( "Backward", "Columnwise", &nq_i, &ib,
wA(1,i), &ldwa, &tau[i], T, &ib );
if (left) {
/* H or H^H is applied to C(1:m-k+i+ib-1,1:n) */
mi = m - k + i + ib - 1;
}
else {
/* H or H^H is applied to C(1:m,1:n-k+i+ib-1) */
ni = n - k + i + ib - 1;
}
/* Apply H or H^H; First copy T to the GPU */
magma_zsetmatrix( ib, ib, T, ib, dT, ib, queue );
magma_zlarfb_gpu( side, trans, MagmaBackward, MagmaColumnwise,
mi, ni, ib,
dA(0,i-1), ldda, dT, ib, // dA using 0-based indices here
dC(1,1), lddc,
dwork, lddwork, queue );
}
cleanup:
magma_queue_destroy( queue );
magma_free( dwork );
return *info;
} /* magma_zunmql */
/**
Purpose
-------
ZGETRF_m computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges. This version does not
require work space on the GPU passed as input. GPU memory is allocated
in the routine. The matrix may exceed the GPU memory.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
Note: The factorization of big panel is done calling multiple-gpu-interface.
Pivots are applied on GPU within the big panel.
Arguments
---------
@param[in]
ngpu INTEGER
Number of GPUs to use. ngpu > 0.
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
A COMPLEX_16 array, dimension (LDA,N)
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
\n
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[out]
ipiv INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
- > 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
@ingroup magma_zgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_zgetrf_m(
magma_int_t ngpu,
magma_int_t m, magma_int_t n,
magmaDoubleComplex *A, magma_int_t lda, magma_int_t *ipiv,
magma_int_t *info)
{
#define A(i,j) (A + (j)*lda + (i))
#define dAT(d,i,j) (dAT[d] + (i)*nb*ldn_local + (j)*nb)
#define dPT(d,i,j) (dPT[d] + (i)*nb*nb + (j)*nb*maxm)
magma_timer_t time=0, time_total=0, time_alloc=0, time_set=0, time_get=0, time_comp=0;
timer_start( time_total );
real_Double_t flops;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *dAT[MagmaMaxGPUs], *dA[MagmaMaxGPUs], *dPT[MagmaMaxGPUs];
magma_int_t iinfo = 0, nb, nbi, maxm, n_local[MagmaMaxGPUs], ldn_local;
magma_int_t N, M, NB, NBk, I, d, ngpu0 = ngpu;
magma_int_t ii, jj, h, offset, ib, rows;
magma_queue_t stream[MagmaMaxGPUs][2];
magma_event_t event[MagmaMaxGPUs][2];
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (lda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
/* initialize nb */
nb = magma_get_zgetrf_nb(m);
maxm = ((m + 31)/32)*32;
/* figure out NB */
size_t freeMem, totalMem;
cudaMemGetInfo( &freeMem, &totalMem );
freeMem /= sizeof(magmaDoubleComplex);
/* number of columns in the big panel */
h = 1+(2+ngpu0);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
const char* ngr_nb_char = getenv("MAGMA_NGR_NB");
if ( ngr_nb_char != NULL )
NB = max( nb, min( NB, atoi(ngr_nb_char) ) );
//NB = 5*max(nb,32);
if ( ngpu0 > ceil((double)NB/nb) ) {
ngpu = (int)ceil((double)NB/nb);
h = 1+(2+ngpu);
NB = (magma_int_t)(0.8*freeMem/maxm-h*nb);
} else {
ngpu = ngpu0;
}
if ( ngpu*NB >= n ) {
#ifdef CHECK_ZGETRF_OOC
printf( " * still fit in GPU memory.\n" );
#endif
NB = n;
} else {
#ifdef CHECK_ZGETRF_OOC
printf( " * don't fit in GPU memory.\n" );
#endif
NB = ngpu*NB;
NB = max( nb, (NB / nb) * nb); /* making sure it's devisable by nb (x64) */
}
#ifdef CHECK_ZGETRF_OOC
if ( NB != n ) printf( " * running in out-core mode (n=%d, NB=%d, nb=%d, freeMem=%.2e).\n", n, NB, nb, (double)freeMem );
else printf( " * running in in-core mode (n=%d, NB=%d, nb=%d, freeMem=%.2e).\n", n, NB, nb, (double)freeMem );
#endif
if ( (nb <= 1) || (nb >= min(m,n)) ) {
/* Use CPU code for scalar of one tile. */
lapackf77_zgetrf(&m, &n, A, &lda, ipiv, info);
} else {
/* Use hybrid blocked code. */
/* allocate memory on GPU to store the big panel */
timer_start( time_alloc );
n_local[0] = (NB/nb)/ngpu;
if ( NB%(nb*ngpu) != 0 )
n_local[0]++;
n_local[0] *= nb;
ldn_local = ((n_local[0]+31)/32)*32;
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
if (MAGMA_SUCCESS != magma_zmalloc( &dA[d], (ldn_local+h*nb)*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
dPT[d] = dA[d] + nb*maxm; /* for storing the previous panel from CPU */
dAT[d] = dA[d] + h*nb*maxm; /* for storing the big panel */
magma_queue_create( &stream[d][0] );
magma_queue_create( &stream[d][1] );
magma_event_create( &event[d][0] );
magma_event_create( &event[d][1] );
}
//magma_setdevice(0);
timer_stop( time_alloc );
for( I=0; I < n; I += NB ) {
M = m;
N = min( NB, n-I ); /* number of columns in this big panel */
//s = min( max(m-I,0), N )/nb; /* number of small block-columns in this big panel */
maxm = ((M + 31)/32)*32;
if ( ngpu0 > ceil((double)N/nb) ) {
ngpu = (int)ceil((double)N/nb);
} else {
ngpu = ngpu0;
}
for( d=0; d < ngpu; d++ ) {
n_local[d] = ((N/nb)/ngpu)*nb;
if (d < (N/nb)%ngpu)
n_local[d] += nb;
else if (d == (N/nb)%ngpu)
n_local[d] += N%nb;
}
ldn_local = ((n_local[0]+31)/32)*32;
/* upload the next big panel into GPU, transpose (A->A'), and pivot it */
timer_start( time );
magmablas_zsetmatrix_transpose_mgpu(ngpu, stream, A(0,I), lda,
dAT, ldn_local, dA, maxm, M, N, nb);
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
time_set += timer_stop( time );
timer_start( time );
/* == --------------------------------------------------------------- == */
/* == loop around the previous big-panels to update the new big-panel == */
for( offset = 0; offset < min(m,I); offset += NB ) {
NBk = min( m-offset, NB );
/* start sending the first tile from the previous big-panels to gpus */
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
nbi = min( nb, NBk );
magma_zsetmatrix_async( (M-offset), nbi,
A(offset,offset), lda,
dA[d], (maxm-offset), stream[d][0] );
/* make sure the previous update finished */
magmablasSetKernelStream(stream[d][0]);
//magma_queue_sync( stream[d][1] );
magma_queue_wait_event( stream[d][0], event[d][0] );
/* transpose */
magmablas_ztranspose( M-offset, nbi, dA[d], maxm-offset, dPT(d,0,0), nb );
}
/* applying the pivot from the previous big-panel */
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magmablas_zlaswp_q( ldn_local, dAT(d,0,0), ldn_local, offset+1, offset+NBk, ipiv, 1, stream[d][1] );
}
/* == going through each block-column of previous big-panels == */
for( jj=0, ib=offset/nb; jj < NBk; jj += nb, ib++ ) {
ii = offset+jj;
rows = maxm - ii;
nbi = min( nb, NBk-jj );
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
/* wait for a block-column on GPU */
magma_queue_sync( stream[d][0] );
/* start sending next column */
if ( jj+nb < NBk ) {
magma_zsetmatrix_async( (M-ii-nb), min(nb,NBk-jj-nb),
A(ii+nb,ii+nb), lda,
dA[d], (rows-nb), stream[d][0] );
/* make sure the previous update finished */
magmablasSetKernelStream(stream[d][0]);
//magma_queue_sync( stream[d][1] );
magma_queue_wait_event( stream[d][0], event[d][(1+jj/nb)%2] );
/* transpose next column */
magmablas_ztranspose( M-ii-nb, nb, dA[d], rows-nb, dPT(d,0,(1+jj/nb)%2), nb );
}
/* update with the block column */
magmablasSetKernelStream(stream[d][1]);
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n_local[d], nbi, c_one, dPT(d,0,(jj/nb)%2), nb, dAT(d,ib,0), ldn_local );
if ( M > ii+nb ) {
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n_local[d], M-(ii+nb), nbi, c_neg_one, dAT(d,ib,0), ldn_local,
dPT(d,1,(jj/nb)%2), nb, c_one, dAT(d,ib+1,0), ldn_local );
}
magma_event_record( event[d][(jj/nb)%2], stream[d][1] );
} /* end of for each block-columns in a big-panel */
}
} /* end of for each previous big-panels */
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
/* calling magma-gpu interface to panel-factorize the big panel */
if ( M > I ) {
magma_zgetrf2_mgpu(ngpu, M-I, N, nb, I, dAT, ldn_local, ipiv+I, dA, A(0,I), lda,
stream, &iinfo);
if ( iinfo < 0 ) {
*info = iinfo;
break;
} else if ( iinfo != 0 ) {
*info = iinfo + I * NB;
//break;
}
/* adjust pivots */
for( ii=I; ii < min(I+N,m); ii++ )
ipiv[ii] += I;
}
time_comp += timer_stop( time );
/* download the current big panel to CPU */
timer_start( time );
magmablas_zgetmatrix_transpose_mgpu(ngpu, stream, dAT, ldn_local, A(0,I), lda, dA, maxm, M, N, nb);
for( d=0; d < ngpu; d++ ) {
magma_setdevice(d);
magma_queue_sync( stream[d][0] );
magma_queue_sync( stream[d][1] );
magmablasSetKernelStream(NULL);
}
time_get += timer_stop( time );
} /* end of for */
timer_stop( time_total );
flops = FLOPS_ZGETRF( m, n ) / 1e9;
timer_printf(" memory-allocation time: %e\n", time_alloc );
timer_printf(" NB=%d nb=%d\n", (int) NB, (int) nb );
timer_printf(" memcopy and transpose %e seconds\n", time_set );
timer_printf(" total time %e seconds\n", time_total );
timer_printf(" Performance %f GFlop/s, %f seconds without htod and dtoh\n", flops / (time_comp), time_comp );
timer_printf(" Performance %f GFlop/s, %f seconds with htod\n", flops / (time_comp + time_set), time_comp + time_set );
timer_printf(" Performance %f GFlop/s, %f seconds with dtoh\n", flops / (time_comp + time_get), time_comp + time_get );
timer_printf(" Performance %f GFlop/s, %f seconds without memory-allocation\n", flops / (time_total - time_alloc), time_total - time_alloc );
for( d=0; d < ngpu0; d++ ) {
magma_setdevice(d);
magma_free( dA[d] );
magma_event_destroy( event[d][0] );
magma_event_destroy( event[d][1] );
magma_queue_destroy( stream[d][0] );
magma_queue_destroy( stream[d][1] );
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
}
if ( *info >= 0 )
magma_zgetrf_piv(m, n, NB, A, lda, ipiv, info);
return *info;
} /* magma_zgetrf_m */
/**
Purpose
-------
ZUNGQR generates an M-by-N COMPLEX_16 matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by ZGEQRF_GPU.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix Q. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix Q. M >= N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
@param[in,out]
dA COMPLEX_16 array A on the GPU, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by ZGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
ldda INTEGER
The first dimension of the array A. LDDA >= max(1,M).
@param[in]
tau COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQRF_GPU.
@param[in]
dT (workspace) COMPLEX_16 work space array on the GPU,
dimension (2*MIN(M, N) + ceil(N/32)*32 )*NB.
This must be the 6th argument of magma_zgeqrf_gpu
[ note that if N here is bigger than N in magma_zgeqrf_gpu,
the workspace requirement DT in magma_zgeqrf_gpu must be
as specified in this routine ].
@param[in]
nb INTEGER
This is the block size used in ZGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in DT.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_zgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_zungqr_gpu(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex_ptr dA, magma_int_t ldda,
magmaDoubleComplex *tau,
magmaDoubleComplex_ptr dT, magma_int_t nb,
magma_int_t *info)
{
#define dA(i,j) (dA + (i) + (j)*ldda)
#define dT(j) (dT + (j)*nb)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, lpanel;
magma_int_t i, ib, ki, kk, iinfo;
magma_int_t lddwork;
magmaDoubleComplex_ptr dV, dW;
magmaDoubleComplex *work, *panel;
*info = 0;
if (m < 0) {
*info = -1;
} else if ((n < 0) || (n > m)) {
*info = -2;
} else if ((k < 0) || (k > n)) {
*info = -3;
} else if (ldda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (n <= 0) {
return *info;
}
// first kk columns are handled by blocked method.
// ki is start of 2nd-to-last block
if ((nb > 1) && (nb < k)) {
ki = (k - nb - 1) / nb * nb;
kk = min( k, ki+nb );
} else {
ki = 0;
kk = 0;
}
// Allocate CPU work space
// n*nb for zungqr workspace
// (m - kk)*(n - kk) for last block's panel
lwork = n*nb;
lpanel = (m - kk)*(n - kk);
magma_zmalloc_cpu( &work, lwork + lpanel );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
panel = work + lwork;
// Allocate work space on GPU
if (MAGMA_SUCCESS != magma_zmalloc( &dV, ldda*nb )) {
magma_free_cpu( work );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
// dT workspace has:
// 2*min(m,n)*nb for T and R^{-1} matrices from geqrf
// roundup(n,32) * nb for dW larfb workspace.
lddwork = min(m,n);
dW = dT + 2*lddwork*nb;
magma_queue_t queue;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
// Use unblocked code for the last or only block.
if (kk < n) {
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
magma_zgetmatrix( m_kk, k_kk,
dA(kk, kk), ldda, panel, m_kk, queue );
lapackf77_zungqr( &m_kk, &n_kk, &k_kk,
panel, &m_kk,
&tau[kk], work, &lwork, &iinfo );
magma_zsetmatrix( m_kk, n_kk,
panel, m_kk, dA(kk, kk), ldda, queue );
// Set A(1:kk,kk+1:n) to zero.
magmablas_zlaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(0, kk), ldda, queue );
}
if (kk > 0) {
// Use blocked code
// queue: copy Aii to V --> laset --> laset --> larfb --> [next]
// CPU has no computation
for (i = ki; i >= 0; i -= nb) {
ib = min( nb, k-i );
mi = m - i;
// Copy current panel on the GPU from dA to dV
magma_zcopymatrix_async( mi, ib,
dA(i,i), ldda,
dV, ldda, queue );
// set panel to identity
magmablas_zlaset( MagmaFull, i, ib, c_zero, c_zero, dA(0, i), ldda, queue );
magmablas_zlaset( MagmaFull, mi, ib, c_zero, c_one, dA(i, i), ldda, queue );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, n-i, ib,
dV, ldda, dT(i), nb,
dA(i, i), ldda, dW, lddwork, queue );
}
}
}
magma_queue_sync( queue );
magma_free( dV );
magma_free_cpu( work );
magma_queue_destroy( queue );
return *info;
} /* magma_zungqr_gpu */
/**
Purpose
-------
ZGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
If the current stream is NULL, this version replaces it with a new
stream to overlap computation with communication.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0.
@param[in,out]
dA COMPLEX_16 array on the GPU, dimension (LDDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
@param[in]
ldda INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
@param[out]
ipiv INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
- > 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
@ingroup magma_zgesv_comp
********************************************************************/
extern "C" magma_int_t
magma_zgetrf_gpu(
magma_int_t m, magma_int_t n,
magmaDoubleComplex_ptr dA, magma_int_t ldda, magma_int_t *ipiv,
magma_int_t *info)
{
#define dAT(i_, j_) (dAT + (i_)*nb*lddat + (j_)*nb)
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t iinfo, nb;
magma_int_t maxm, maxn, mindim;
magma_int_t i, j, rows, cols, s, lddat, ldwork;
magmaDoubleComplex *dAT, *dAP, *work;
/* Check arguments */
*info = 0;
if (m < 0)
*info = -1;
else if (n < 0)
*info = -2;
else if (ldda < max(1,m))
*info = -4;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* Function Body */
mindim = min(m, n);
nb = magma_get_zgetrf_nb(m);
s = mindim / nb;
if (nb <= 1 || nb >= min(m,n)) {
/* Use CPU code. */
magma_zmalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_zgetmatrix( m, n, dA, ldda, work, m );
lapackf77_zgetrf(&m, &n, work, &m, ipiv, info);
magma_zsetmatrix( m, n, work, m, dA, ldda );
magma_free_cpu(work);
}
else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
maxn = ((n + 31)/32)*32;
if (MAGMA_SUCCESS != magma_zmalloc( &dAP, nb*maxm )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
// square matrices can be done in place;
// rectangular requires copy to transpose
if ( m == n ) {
dAT = dA;
lddat = ldda;
magmablas_ztranspose_inplace( m, dAT, ldda );
}
else {
lddat = maxn; // N-by-M
if (MAGMA_SUCCESS != magma_zmalloc( &dAT, lddat*maxm )) {
magma_free( dAP );
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmablas_ztranspose( m, n, dA, ldda, dAT, lddat );
}
ldwork = maxm;
if (MAGMA_SUCCESS != magma_zmalloc_pinned( &work, ldwork*nb )) {
magma_free( dAP );
if ( ! (m == n))
magma_free( dAT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
/* Define user stream if current stream is NULL */
magma_queue_t stream[2];
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
magma_queue_create( &stream[0] );
if (orig_stream == NULL) {
magma_queue_create( &stream[1] );
magmablasSetKernelStream(stream[1]);
}
else {
stream[1] = orig_stream;
}
for( j=0; j < s; j++ ) {
// download j-th panel
cols = maxm - j*nb;
magmablas_ztranspose( nb, m-j*nb, dAT(j,j), lddat, dAP, cols );
// make sure that the transpose has completed
magma_queue_sync( stream[1] );
magma_zgetmatrix_async( m-j*nb, nb, dAP, cols, work, ldwork,
stream[0]);
if ( j > 0 ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n - (j+1)*nb, nb,
c_one, dAT(j-1,j-1), lddat,
dAT(j-1,j+1), lddat );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-j*nb, nb,
c_neg_one, dAT(j-1,j+1), lddat,
dAT(j, j-1), lddat,
c_one, dAT(j, j+1), lddat );
}
// do the cpu part
rows = m - j*nb;
magma_queue_sync( stream[0] );
lapackf77_zgetrf( &rows, &nb, work, &ldwork, ipiv+j*nb, &iinfo);
if ( *info == 0 && iinfo > 0 )
*info = iinfo + j*nb;
// upload j-th panel
magma_zsetmatrix_async( m-j*nb, nb, work, ldwork, dAP, maxm,
stream[0]);
for( i=j*nb; i < j*nb + nb; ++i ) {
ipiv[i] += j*nb;
}
magmablas_zlaswp( n, dAT, lddat, j*nb + 1, j*nb + nb, ipiv, 1 );
magma_queue_sync( stream[0] );
magmablas_ztranspose( m-j*nb, nb, dAP, maxm, dAT(j,j), lddat );
// do the small non-parallel computations (next panel update)
if ( s > (j+1) ) {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat );
}
else {
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb, nb,
c_one, dAT(j, j ), lddat,
dAT(j, j+1), lddat);
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
n-(j+1)*nb, m-(j+1)*nb, nb,
c_neg_one, dAT(j, j+1), lddat,
dAT(j+1, j ), lddat,
c_one, dAT(j+1, j+1), lddat );
}
}
magma_int_t nb0 = min(m - s*nb, n - s*nb);
if ( nb0 > 0 ) {
rows = m - s*nb;
cols = maxm - s*nb;
magmablas_ztranspose( nb0, rows, dAT(s,s), lddat, dAP, maxm );
magma_zgetmatrix( rows, nb0, dAP, maxm, work, ldwork );
// do the cpu part
lapackf77_zgetrf( &rows, &nb0, work, &ldwork, ipiv+s*nb, &iinfo);
if ( *info == 0 && iinfo > 0 )
*info = iinfo + s*nb;
for( i=s*nb; i < s*nb + nb0; ++i ) {
ipiv[i] += s*nb;
}
magmablas_zlaswp( n, dAT, lddat, s*nb + 1, s*nb + nb0, ipiv, 1 );
// upload j-th panel
magma_zsetmatrix( rows, nb0, work, ldwork, dAP, maxm );
magmablas_ztranspose( rows, nb0, dAP, maxm, dAT(s,s), lddat );
magma_ztrsm( MagmaRight, MagmaUpper, MagmaNoTrans, MagmaUnit,
n-s*nb-nb0, nb0,
c_one, dAT(s,s), lddat,
dAT(s,s)+nb0, lddat);
}
// undo transpose
if ( m == n ) {
magmablas_ztranspose_inplace( m, dAT, lddat );
}
else {
magmablas_ztranspose( n, m, dAT, lddat, dA, ldda );
magma_free( dAT );
}
magma_free( dAP );
magma_free_pinned( work );
magma_queue_destroy( stream[0] );
if (orig_stream == NULL) {
magma_queue_destroy( stream[1] );
}
magmablasSetKernelStream( orig_stream );
}
return *info;
} /* magma_zgetrf_gpu */
/**
Purpose
-------
ZHETRD reduces a complex Hermitian matrix A to real symmetric
tridiagonal form T by an orthogonal similarity transformation:
Q**H * A * Q = T.
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 COMPLEX_16 array, dimension (LDA,N)
On entry, the Hermitian 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.
On exit, if UPLO = MagmaUpper, the diagonal and first superdiagonal
of A are overwritten by the corresponding elements of the
tridiagonal matrix T, and the elements above the first
superdiagonal, with the array TAU, represent the orthogonal
matrix Q as a product of elementary reflectors; if UPLO
= MagmaLower, the diagonal and first subdiagonal of A are over-
written by the corresponding elements of the tridiagonal
matrix T, and the elements below the first subdiagonal, with
the array TAU, represent the orthogonal matrix Q as a product
of elementary reflectors. See Further Details.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
d COMPLEX_16 array, dimension (N)
The diagonal elements of the tridiagonal matrix T:
D(i) = A(i,i).
@param[out]
e COMPLEX_16 array, dimension (N-1)
The off-diagonal elements of the tridiagonal matrix T:
E(i) = A(i,i+1) if UPLO = MagmaUpper, E(i) = A(i+1,i) if UPLO = MagmaLower.
@param[out]
tau COMPLEX_16 array, dimension (N-1)
The scalar factors of the elementary reflectors (see Further
Details).
@param[out]
work (workspace) COMPLEX_16 array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= N*NB, where NB is the
optimal blocksize given by magma_get_zhetrd_nb().
\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
---------------
If UPLO = MagmaUpper, the matrix Q is represented as a product of elementary
reflectors
Q = H(n-1) . . . H(2) H(1).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(i+1:n) = 0 and v(i) = 1; v(1:i-1) is stored on exit in
A(1:i-1,i+1), and tau in TAU(i).
If UPLO = MagmaLower, the matrix Q is represented as a product of elementary
reflectors
Q = H(1) H(2) . . . H(n-1).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(1:i) = 0 and v(i+1) = 1; v(i+2:n) is stored on exit in A(i+2:n,i),
and tau in TAU(i).
The contents of A on exit are illustrated by the following examples
with n = 5:
if UPLO = MagmaUpper: if UPLO = MagmaLower:
( d e v2 v3 v4 ) ( d )
( d e v3 v4 ) ( e d )
( d e v4 ) ( v1 e d )
( d e ) ( v1 v2 e d )
( d ) ( v1 v2 v3 e d )
where d and e denote diagonal and off-diagonal elements of T, and vi
denotes an element of the vector defining H(i).
@ingroup magma_zheev_comp
********************************************************************/
extern "C" magma_int_t
magma_zhetrd(
magma_uplo_t uplo, magma_int_t n,
magmaDoubleComplex *A, magma_int_t lda,
double *d, double *e, magmaDoubleComplex *tau,
magmaDoubleComplex *work, magma_int_t lwork,
magma_int_t *info)
{
#define A(i_, j_) ( A + (i_) + (j_)*lda )
#define dA(i_, j_) (dA + (i_) + (j_)*ldda)
const char* uplo_ = lapack_uplo_const( uplo );
magma_int_t ldda = roundup( n, 32 );
magma_int_t nb = magma_get_zhetrd_nb( n );
const magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
const magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
const magmaDoubleComplex c_one = MAGMA_Z_ONE;
const double d_one = MAGMA_D_ONE;
magma_int_t kk, nx;
magma_int_t i, j, i_n;
magma_int_t iinfo;
magma_int_t ldw, lddw, lwkopt;
magma_int_t lquery;
*info = 0;
int upper = (uplo == MagmaUpper);
lquery = (lwork == -1);
if (! upper && uplo != MagmaLower) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,n)) {
*info = -4;
} else if (lwork < nb*n && ! lquery) {
*info = -9;
}
/* Determine the block size. */
ldw = n;
lddw = ldda;
lwkopt = n * nb;
if (*info == 0) {
work[0] = MAGMA_Z_MAKE( lwkopt, 0 );
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
/* Quick return if possible */
if (n == 0) {
work[0] = c_one;
return *info;
}
magmaDoubleComplex *dA;
#ifdef FAST_HEMV
magma_int_t ldwork2 = ldda*ceildiv(n,64);
#else
magma_int_t ldwork2 = 0;
#endif
if (MAGMA_SUCCESS != magma_zmalloc( &dA, ldda*n + 2*lddw*nb + ldwork2 )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magmaDoubleComplex *dwork = dA + ldda*n;
#ifdef FAST_HEMV
magmaDoubleComplex *dwork2 = dwork + 2*lddw*nb;
#endif
//if (n < 2048)
// nx = n;
//else
// nx = 512;
nx = min( 128, n ); // nx <= n is required
// clear out dwork in case it has NANs (used as y in zhemv)
// rest of dwork (used as work in magmablas_zhemv) doesn't need to be cleared
magmablas_zlaset( MagmaFull, n, nb, c_zero, c_zero, dwork, lddw );
if (upper) {
/* Copy the matrix to the GPU */
magma_zsetmatrix( n, n, A(0, 0), lda, dA(0, 0), ldda );
/* Reduce the upper triangle of A.
Columns 1:kk are handled by the unblocked method. */
kk = n - (n - nx + nb - 1) / nb * nb;
for (i = n - nb; i >= kk; i -= nb) {
/* Reduce columns i:i+nb-1 to tridiagonal form and form the
matrix W which is needed to update the unreduced part of
the matrix */
/* Get the current panel (no need for the 1st iteration) */
if (i != n-nb)
magma_zgetmatrix( i+nb, nb, dA(0, i), ldda, A(0, i), lda );
#ifdef FAST_HEMV
magma_zlatrd2( uplo, i+nb, nb, A(0, 0), lda, e, tau,
work, ldw, dA(0, 0), ldda, dwork, lddw,
dwork2, ldwork2 );
#else
magma_zlatrd( uplo, i+nb, nb, A(0, 0), lda, e, tau,
work, ldw, dA(0, 0), ldda, dwork, lddw );
#endif
/* Update the unreduced submatrix A(0:i-2,0:i-2), using an
update of the form: A := A - V*W' - W*V' */
magma_zsetmatrix( i + nb, nb, work, ldw, dwork, lddw );
magma_zher2k( uplo, MagmaNoTrans, i, nb, c_neg_one,
dA(0, i), ldda, dwork, lddw,
d_one, dA(0, 0), ldda );
/* Copy superdiagonal elements back into A, and diagonal
elements into D */
for (j = i; j < i+nb; ++j) {
*A(j-1,j) = MAGMA_Z_MAKE( e[j - 1], 0 );
d[j] = MAGMA_Z_REAL( *A(j, j) );
}
}
magma_zgetmatrix( kk, kk, dA(0, 0), ldda, A(0, 0), lda );
/* Use CPU code to reduce the last or only block */
lapackf77_zhetrd( uplo_, &kk, A(0, 0), &lda, d, e, tau, work, &lwork, &iinfo );
}
else {
/* Copy the matrix to the GPU */
if (1 <= n-nx)
magma_zsetmatrix( n, n, A(0,0), lda, dA(0,0), ldda );
/* Reduce the lower triangle of A */
for (i = 0; i < n-nx; i += nb) {
/* Reduce columns i:i+nb-1 to tridiagonal form and form the
matrix W which is needed to update the unreduced part of
the matrix */
/* Get the current panel (no need for the 1st iteration) */
if (i != 0)
magma_zgetmatrix( n-i, nb, dA(i, i), ldda, A(i, i), lda );
#ifdef FAST_HEMV
magma_zlatrd2( uplo, n-i, nb, A(i, i), lda, &e[i], &tau[i],
work, ldw, dA(i, i), ldda, dwork, lddw,
dwork2, ldwork2 );
#else
magma_zlatrd( uplo, n-i, nb, A(i, i), lda, &e[i], &tau[i],
work, ldw, dA(i, i), ldda, dwork, lddw );
#endif
/* Update the unreduced submatrix A(i+ib:n,i+ib:n), using
an update of the form: A := A - V*W' - W*V' */
magma_zsetmatrix( n-i, nb, work, ldw, dwork, lddw );
magma_zher2k( MagmaLower, MagmaNoTrans, n-i-nb, nb, c_neg_one,
dA(i+nb, i), ldda, &dwork[nb], lddw,
d_one, dA(i+nb, i+nb), ldda );
/* Copy subdiagonal elements back into A, and diagonal
elements into D */
for (j = i; j < i+nb; ++j) {
*A(j+1,j) = MAGMA_Z_MAKE( e[j], 0 );
d[j] = MAGMA_Z_REAL( *A(j, j) );
}
}
/* Use CPU code to reduce the last or only block */
if (1 <= n-nx)
magma_zgetmatrix( n-i, n-i, dA(i, i), ldda, A(i, i), lda );
i_n = n-i;
lapackf77_zhetrd( uplo_, &i_n, A(i, i), &lda, &d[i], &e[i],
&tau[i], work, &lwork, &iinfo );
}
magma_free( dA );
work[0] = MAGMA_Z_MAKE( lwkopt, 0 );
return *info;
} /* magma_zhetrd */
extern "C" magma_int_t
magma_zunmqr(const char side, const char trans,
magma_int_t m, magma_int_t n, magma_int_t k,
cuDoubleComplex *A, magma_int_t lda,
cuDoubleComplex *tau,
cuDoubleComplex *C, magma_int_t ldc,
cuDoubleComplex *work, magma_int_t lwork,
magma_int_t *info)
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2012
Purpose
=======
ZUNMQR overwrites the general complex M-by-N matrix C with
SIDE = 'L' SIDE = 'R'
TRANS = 'N': Q * C C * Q
TRANS = 'T': Q**H * C C * Q**H
where Q is a complex orthogonal matrix defined as the product of k
elementary reflectors
Q = H(1) H(2) . . . H(k)
as returned by ZGEQRF. Q is of order M if SIDE = 'L' and of order N
if SIDE = 'R'.
Arguments
=========
SIDE (input) CHARACTER*1
= 'L': apply Q or Q**H from the Left;
= 'R': apply Q or Q**H from the Right.
TRANS (input) CHARACTER*1
= 'N': No transpose, apply Q;
= 'T': Transpose, apply Q**H.
M (input) INTEGER
The number of rows of the matrix C. M >= 0.
N (input) INTEGER
The number of columns of the matrix C. N >= 0.
K (input) INTEGER
The number of elementary reflectors whose product defines
the matrix Q.
If SIDE = 'L', M >= K >= 0;
if SIDE = 'R', N >= K >= 0.
A (input) COMPLEX_16 array, dimension (LDA,K)
The i-th column must contain the vector which defines the
elementary reflector H(i), for i = 1,2,...,k, as returned by
ZGEQRF in the first k columns of its array argument A.
A is modified by the routine but restored on exit.
LDA (input) INTEGER
The leading dimension of the array A.
If SIDE = 'L', LDA >= max(1,M);
if SIDE = 'R', LDA >= max(1,N).
TAU (input) COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQRF.
C (input/output) COMPLEX_16 array, dimension (LDC,N)
On entry, the M-by-N matrix C.
On exit, C is overwritten by Q*C or Q**H * C or C * Q**H or C*Q.
LDC (input) INTEGER
The leading dimension of the array C. LDC >= max(1,M).
WORK (workspace/output) COMPLEX_16 array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(0) returns the optimal LWORK.
LWORK (input) INTEGER
The dimension of the array WORK.
If SIDE = 'L', LWORK >= max(1,N);
if SIDE = 'R', LWORK >= max(1,M).
For optimum performance
LWORK >= N*NB if SIDE = 'L', and
LWORK >= M*NB if SIDE = 'R',
where NB is the optimal blocksize.
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.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
===================================================================== */
#define A(a_1,a_2) ( A + (a_1) + (a_2)*lda)
#define dC(a_1,a_2) (dC + (a_1) + (a_2)*lddc)
magma_int_t nb = magma_get_zgeqrf_nb( min( m, n ));
cuDoubleComplex c_one = MAGMA_Z_ONE;
char side_[2] = {side, 0};
char trans_[2] = {trans, 0};
magma_int_t nq_i, lddwork;
magma_int_t i;
cuDoubleComplex T[ 2*nb*nb ];
magma_int_t i1, i2, step, ib, ic, jc, mi, ni, nq, nw;
int left, notran, lquery;
magma_int_t iinfo, lwkopt;
*info = 0;
left = lapackf77_lsame(side_, "L");
notran = lapackf77_lsame(trans_, "N");
lquery = (lwork == -1);
/* NQ is the order of Q and NW is the minimum dimension of WORK */
if (left) {
nq = m;
nw = n;
} else {
nq = n;
nw = m;
}
lwkopt = max(1,nw) * nb;
work[0] = MAGMA_Z_MAKE( lwkopt, 0 );
if (! left && ! lapackf77_lsame(side_, "R")) {
*info = -1;
} else if (! notran && ! lapackf77_lsame(trans_, MagmaConjTransStr)) {
*info = -2;
} else if (m < 0) {
*info = -3;
} else if (n < 0) {
*info = -4;
} else if (k < 0 || k > nq) {
*info = -5;
} else if (lda < max(1,nq)) {
*info = -7;
} else if (ldc < max(1,m)) {
*info = -10;
} else if (lwork < max(1,nw) && ! lquery) {
*info = -12;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0 || k == 0) {
work[0] = c_one;
return *info;
}
/* Allocate work space on the GPU */
magma_int_t lddc = m;
cuDoubleComplex *dwork, *dC;
magma_zmalloc( &dC, lddc*n );
magma_zmalloc( &dwork, (m + n + nb)*nb );
/* Copy matrix C from the CPU to the GPU */
magma_zsetmatrix( m, n, C, ldc, dC, lddc );
if (nb >= k) {
/* Use CPU code */
lapackf77_zunmqr(side_, trans_, &m, &n, &k, A, &lda, &tau[1],
C, &ldc, work, &lwork, &iinfo);
}
else {
/* Use hybrid CPU-GPU code */
if ( (left && (! notran)) || ((! left) && notran) ) {
i1 = 0;
i2 = k;
step = nb;
} else {
i1 = ((k - 1) / nb) * nb;
i2 = 0;
step = -nb;
}
if (left) {
ni = n;
jc = 0;
} else {
mi = m;
ic = 0;
}
for( i=i1; (step<0 ? i>=i2 : i<i2); i += step ) {
ib = min(nb, k - i);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
nq_i = nq - i;
lapackf77_zlarft("F", "C", &nq_i, &ib, A(i,i), &lda,
&tau[i], T, &ib);
/* 1) Put 0s in the upper triangular part of A;
2) copy the panel from A to the GPU, and
3) restore A */
zpanel_to_q('U', ib, A(i,i), lda, T+ib*ib);
magma_zsetmatrix( nq_i, ib, A(i,i), lda, dwork, nq_i );
zq_to_panel('U', ib, A(i,i), lda, T+ib*ib);
if (left) {
/* H or H' is applied to C(i:m,1:n) */
mi = m - i;
ic = i;
}
else {
/* H or H' is applied to C(1:m,i:n) */
ni = n - i;
jc = i;
}
if (left)
lddwork = ni;
else
lddwork = mi;
/* Apply H or H'; First copy T to the GPU */
magma_zsetmatrix( ib, ib, T, ib, dwork+nq_i*ib, ib );
magma_zlarfb_gpu( side, trans, MagmaForward, MagmaColumnwise,
mi, ni, ib,
dwork, nq_i, dwork+nq_i*ib, ib,
dC(ic,jc), lddc,
dwork+nq_i*ib + ib*ib, lddwork);
}
magma_zgetmatrix( m, n, dC, lddc, C, ldc );
}
work[0] = MAGMA_Z_MAKE( lwkopt, 0 );
magma_free( dC );
magma_free( dwork );
return *info;
} /* magma_zunmqr */
/**
Purpose
-------
ZTRTRI computes the inverse of a real upper or lower triangular
matrix A.
This is the Level 3 BLAS version of the algorithm.
Arguments
---------
@param[in]
uplo magma_uplo_t
- = MagmaUpper: A is upper triangular;
- = MagmaLower: A is lower triangular.
@param[in]
diag magma_diag_t
- = MagmaNonUnit: A is non-unit triangular;
- = MagmaUnit: A is unit triangular.
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in,out]
A COMPLEX_16 array, dimension (LDA,N)
On entry, the triangular matrix A. If UPLO = MagmaUpper, the
leading N-by-N upper triangular part of the array A contains
the upper triangular matrix, and the strictly lower
triangular part of A is not referenced. If UPLO = MagmaLower, the
leading N-by-N lower triangular part of the array A contains
the lower triangular matrix, and the strictly upper
triangular part of A is not referenced. If DIAG = MagmaUnit, the
diagonal elements of A are also not referenced and are
assumed to be 1.
On exit, the (triangular) inverse of the original matrix, in
the same storage format.
@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
- > 0: if INFO = i, A(i,i) is exactly zero. The triangular
matrix is singular and its inverse cannot be computed.
@ingroup magma_zgesv_aux
********************************************************************/
extern "C" magma_int_t
magma_ztrtri(
magma_uplo_t uplo, magma_diag_t diag, magma_int_t n,
magmaDoubleComplex *A, magma_int_t lda,
magma_int_t *info)
{
#define A(i, j) ( A + (i) + (j)*lda )
#define dA(i, j) (dA + (i) + (j)*ldda)
/* Local variables */
const char* uplo_ = lapack_uplo_const( uplo );
const char* diag_ = lapack_diag_const( diag );
magma_int_t ldda, nb, nn, j, jb;
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex *dA;
int upper = (uplo == MagmaUpper);
int nounit = (diag == MagmaNonUnit);
*info = 0;
if (! upper && uplo != MagmaLower)
*info = -1;
else if (! nounit && diag != MagmaUnit)
*info = -2;
else if (n < 0)
*info = -3;
else if (lda < max(1,n))
*info = -5;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return */
if ( n == 0 )
return *info;
/* Check for singularity if non-unit */
if (nounit) {
for (j=0; j < n; ++j) {
if ( MAGMA_Z_EQUAL( *A(j,j), c_zero )) {
*info = j+1; // Fortran index
return *info;
}
}
}
/* Determine the block size for this environment */
nb = magma_get_zpotrf_nb(n);
ldda = ((n+31)/32)*32;
if (MAGMA_SUCCESS != magma_zmalloc( &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] );
if (nb <= 1 || nb >= n)
lapackf77_ztrtri(uplo_, diag_, &n, A, &lda, info);
else {
if (upper) {
/* Compute inverse of upper triangular matrix */
for (j=0; j < n; j += nb) {
jb = min(nb, (n-j));
magma_zsetmatrix( jb, (n-j),
A(j, j), lda,
dA(j, j), ldda );
/* Compute rows 1:j-1 of current block column */
magma_ztrmm( MagmaLeft, MagmaUpper,
MagmaNoTrans, MagmaNonUnit, j, jb,
c_one, dA(0,0), ldda, dA(0, j),ldda);
magma_ztrsm( MagmaRight, MagmaUpper,
MagmaNoTrans, MagmaNonUnit, j, jb,
c_neg_one, dA(j,j), ldda, dA(0, j),ldda);
magma_zgetmatrix_async( jb, jb,
dA(j, j), ldda,
A(j, j), lda, stream[1] );
magma_zgetmatrix_async( j, jb,
dA(0, j), ldda,
A(0, j), lda, stream[0] );
magma_queue_sync( stream[1] );
/* Compute inverse of current diagonal block */
lapackf77_ztrtri(MagmaUpperStr, diag_, &jb, A(j,j), &lda, info);
magma_zsetmatrix( jb, jb,
A(j, j), lda,
dA(j, j), ldda );
}
}
else {
/* Compute inverse of lower triangular matrix */
nn=((n-1)/nb)*nb+1;
for (j=nn-1; j >= 0; j -= nb) {
jb=min(nb,(n-j));
if ((j+jb) < n) {
magma_zsetmatrix( (n-j), jb,
A(j, j), lda,
dA(j, j), ldda );
/* Compute rows j+jb:n of current block column */
magma_ztrmm( MagmaLeft, MagmaLower,
MagmaNoTrans, MagmaNonUnit, (n-j-jb), jb,
c_one, dA(j+jb,j+jb), ldda, dA(j+jb, j), ldda );
magma_ztrsm( MagmaRight, MagmaLower,
MagmaNoTrans, MagmaNonUnit, (n-j-jb), jb,
c_neg_one, dA(j,j), ldda, dA(j+jb, j), ldda );
magma_zgetmatrix_async( n-j-jb, jb,
dA(j+jb, j), ldda,
A(j+jb, j), lda, stream[1] );
magma_zgetmatrix_async( jb, jb,
dA(j,j), ldda,
A(j,j), lda, stream[0] );
magma_queue_sync( stream[0] );
}
/* Compute inverse of current diagonal block */
lapackf77_ztrtri(MagmaLowerStr, diag_, &jb, A(j,j), &lda, info);
magma_zsetmatrix( jb, jb,
A(j, j), lda,
dA(j, j), ldda );
}
}
}
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_free( dA );
return *info;
}
extern "C" magma_int_t
magma_zheevdx_2stage(char jobz, char range, char uplo,
magma_int_t n,
magmaDoubleComplex *a, magma_int_t lda,
double vl, double vu, magma_int_t il, magma_int_t iu,
magma_int_t *m, double *w,
magmaDoubleComplex *work, magma_int_t lwork,
double *rwork, magma_int_t lrwork,
magma_int_t *iwork, magma_int_t liwork,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
ZHEEVD_2STAGE computes all eigenvalues and, optionally, eigenvectors of a
complex Hermitian matrix A. It uses a two-stage algorithm for the tridiagonalization.
If eigenvectors are desired, it uses a divide and conquer algorithm.
The divide and conquer algorithm 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
=========
JOBZ (input) CHARACTER*1
= 'N': Compute eigenvalues only;
= 'V': Compute eigenvalues and eigenvectors.
RANGE (input) CHARACTER*1
= 'A': all eigenvalues will be found.
= 'V': all eigenvalues in the half-open interval (VL,VU]
will be found.
= 'I': the IL-th through IU-th eigenvalues will be found.
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) COMPLEX_16 array, dimension (LDA, N)
On entry, the Hermitian matrix A. If UPLO = 'U', the
leading N-by-N upper triangular part of A contains the
upper triangular part of the matrix A. If UPLO = 'L',
the leading N-by-N lower triangular part of A contains
the lower triangular part of the matrix A.
On exit, if JOBZ = 'V', then if INFO = 0, the first m columns
of A contains the required
orthonormal eigenvectors of the matrix A.
If JOBZ = 'N', then on exit the lower triangle (if UPLO='L')
or the upper triangle (if UPLO='U') of A, including the
diagonal, is destroyed.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,N).
VL (input) DOUBLE PRECISION
VU (input) DOUBLE PRECISION
If RANGE='V', the lower and upper bounds of the interval to
be searched for eigenvalues. VL < VU.
Not referenced if RANGE = 'A' or 'I'.
IL (input) INTEGER
IU (input) INTEGER
If RANGE='I', 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 = 'A' or 'V'.
M (output) INTEGER
The total number of eigenvalues found. 0 <= M <= N.
If RANGE = 'A', M = N, and if RANGE = 'I', M = IU-IL+1.
W (output) DOUBLE PRECISION array, dimension (N)
If INFO = 0, the required m eigenvalues in ascending order.
WORK (workspace/output) COMPLEX_16 array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
LWORK (input) INTEGER
The length of the array WORK.
If N <= 1, LWORK >= 1.
If JOBZ = 'N' and N > 1, LWORK >= LQ2 + N * (NB + 1).
If JOBZ = 'V' and N > 1, LWORK >= LQ2 + 2*N + N**2.
where LQ2 is the size needed to store
the Q2 matrix and is returned by
MAGMA_BULGE_GET_LQ2.
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal sizes of the WORK, RWORK and
IWORK arrays, returns these values as the first entries of
the WORK, RWORK and IWORK arrays, and no error message
related to LWORK or LRWORK or LIWORK is issued by XERBLA.
RWORK (workspace/output) DOUBLE PRECISION array,
dimension (LRWORK)
On exit, if INFO = 0, RWORK(1) returns the optimal LRWORK.
LRWORK (input) INTEGER
The dimension of the array RWORK.
If N <= 1, LRWORK >= 1.
If JOBZ = 'N' and N > 1, LRWORK >= N.
If JOBZ = 'V' and N > 1, LRWORK >= 1 + 5*N + 2*N**2.
If LRWORK = -1, then a workspace query is assumed; the
routine only calculates the optimal sizes of the WORK, RWORK
and IWORK arrays, returns these values as the first entries
of the WORK, RWORK and IWORK arrays, and no error message
related to LWORK or LRWORK or LIWORK is issued by XERBLA.
IWORK (workspace/output) INTEGER array, dimension (MAX(1,LIWORK))
On exit, if INFO = 0, IWORK(1) returns the optimal LIWORK.
LIWORK (input) INTEGER
The dimension of the array IWORK.
If N <= 1, LIWORK >= 1.
If JOBZ = 'N' and N > 1, LIWORK >= 1.
If JOBZ = 'V' and N > 1, LIWORK >= 3 + 5*N.
If LIWORK = -1, then a workspace query is assumed; the
routine only calculates the optimal sizes of the WORK, RWORK
and IWORK arrays, returns these values as the first entries
of the WORK, RWORK and IWORK arrays, and no error message
related to LWORK or LRWORK or LIWORK is issued by XERBLA.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
> 0: if INFO = i and JOBZ = 'N', then the algorithm failed
to converge; i off-diagonal elements of an intermediate
tridiagonal form did not converge to zero;
if INFO = i and JOBZ = 'V', then the algorithm failed
to compute an eigenvalue while working on the submatrix
lying in rows and columns INFO/(N+1) through
mod(INFO,N+1).
Further Details
===============
Based on contributions by
Jeff Rutter, Computer Science Division, University of California
at Berkeley, USA
Modified description of INFO. Sven, 16 Feb 05.
===================================================================== */
char uplo_[2] = {uplo, 0};
char jobz_[2] = {jobz, 0};
char range_[2] = {range, 0};
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t ione = 1;
magma_int_t izero = 0;
double d_one = 1.;
double d__1;
double eps;
double anrm;
magma_int_t imax;
double rmin, rmax;
double sigma;
//magma_int_t iinfo;
magma_int_t lwmin, lrwmin, liwmin;
magma_int_t lower;
magma_int_t wantz;
magma_int_t iscale;
double safmin;
double bignum;
double smlnum;
magma_int_t lquery;
magma_int_t alleig, valeig, indeig;
double* dwork;
/* determine the number of threads */
magma_int_t threads = magma_get_numthreads();
magma_setlapack_numthreads(threads);
wantz = lapackf77_lsame(jobz_, MagmaVecStr);
lower = lapackf77_lsame(uplo_, MagmaLowerStr);
alleig = lapackf77_lsame( range_, "A" );
valeig = lapackf77_lsame( range_, "V" );
indeig = lapackf77_lsame( range_, "I" );
lquery = lwork == -1 || lrwork == -1 || liwork == -1;
*info = 0;
if (! (wantz || lapackf77_lsame(jobz_, MagmaNoVecStr))) {
*info = -1;
} else if (! (alleig || valeig || indeig)) {
*info = -2;
} else if (! (lower || lapackf77_lsame(uplo_, MagmaUpperStr))) {
*info = -3;
} else if (n < 0) {
*info = -4;
} else if (lda < max(1,n)) {
*info = -6;
} else {
if (valeig) {
if (n > 0 && vu <= vl) {
*info = -8;
}
} else if (indeig) {
if (il < 1 || il > max(1,n)) {
*info = -9;
} else if (iu < min(n,il) || iu > n) {
*info = -10;
}
}
}
magma_int_t nb = magma_get_zbulge_nb(n,threads);
magma_int_t Vblksiz = magma_zbulge_get_Vblksiz(n, nb, threads);
magma_int_t ldt = Vblksiz;
magma_int_t ldv = nb + Vblksiz;
magma_int_t blkcnt = magma_bulge_get_blkcnt(n, nb, Vblksiz);
magma_int_t lq2 = magma_zbulge_get_lq2(n, threads);
if (wantz) {
lwmin = lq2 + 2 * n + n * n;
lrwmin = 1 + 5 * n + 2 * n * n;
liwmin = 5 * n + 3;
} else {
lwmin = lq2 + n * (nb + 1);
lrwmin = n;
liwmin = 1;
}
work[0] = MAGMA_Z_MAKE( lwmin * (1. + lapackf77_dlamch("Epsilon")), 0.); // round up
rwork[0] = lrwmin * (1. + lapackf77_dlamch("Epsilon"));
iwork[0] = liwmin;
if ((lwork < lwmin) && !lquery) {
*info = -14;
} else if ((lrwork < lrwmin) && ! lquery) {
*info = -16;
} else if ((liwork < liwmin) && ! lquery) {
*info = -18;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (n == 0) {
return *info;
}
if (n == 1) {
w[0] = MAGMA_Z_REAL(a[0]);
if (wantz) {
a[0] = MAGMA_Z_ONE;
}
return *info;
}
#ifdef ENABLE_TIMER
printf("using %d threads\n", threads);
#endif
/* Check if matrix is very small then just call LAPACK on CPU, no need for GPU */
magma_int_t ntiles = n/nb;
if( ( ntiles < 2 ) || ( n <= 128 ) ){
#ifdef ENABLE_DEBUG
printf("--------------------------------------------------------------\n");
printf(" warning matrix too small N=%d NB=%d, calling lapack on CPU \n", (int) n, (int) nb);
printf("--------------------------------------------------------------\n");
#endif
lapackf77_zheevd(jobz_, &uplo, &n,
a, &lda, w,
work, &lwork,
#if defined(PRECISION_z) || defined(PRECISION_c)
rwork, &lrwork,
#endif
iwork, &liwork,
info);
*m = n;
return *info;
}
/* Get machine constants. */
safmin = lapackf77_dlamch("Safe minimum");
eps = lapackf77_dlamch("Precision");
smlnum = safmin / eps;
bignum = 1. / smlnum;
rmin = magma_dsqrt(smlnum);
rmax = magma_dsqrt(bignum);
/* Scale matrix to allowable range, if necessary. */
anrm = lapackf77_zlanhe("M", uplo_, &n, a, &lda, rwork);
iscale = 0;
if (anrm > 0. && anrm < rmin) {
iscale = 1;
sigma = rmin / anrm;
} else if (anrm > rmax) {
iscale = 1;
sigma = rmax / anrm;
}
if (iscale == 1) {
lapackf77_zlascl(uplo_, &izero, &izero, &d_one, &sigma, &n, &n, a,
&lda, info);
}
magma_int_t indT2 = 0;
magma_int_t indTAU2 = indT2 + blkcnt*ldt*Vblksiz;
magma_int_t indV2 = indTAU2+ blkcnt*Vblksiz;
magma_int_t indtau1 = indV2 + blkcnt*ldv*Vblksiz;
magma_int_t indwrk = indtau1+ n;
//magma_int_t indwk2 = indwrk + n * n;
magma_int_t llwork = lwork - indwrk;
//magma_int_t llwrk2 = lwork - indwk2;
magma_int_t inde = 0;
magma_int_t indrwk = inde + n;
magma_int_t llrwk = lrwork - indrwk;
#ifdef ENABLE_TIMER
magma_timestr_t start, st1, st2, end;
start = get_current_time();
#endif
magmaDoubleComplex *dT1;
if (MAGMA_SUCCESS != magma_zmalloc( &dT1, n*nb)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_zhetrd_he2hb(uplo, n, nb, a, lda, &work[indtau1], &work[indwrk], llwork, dT1, threads, info);
#ifdef ENABLE_TIMER
st1 = get_current_time();
printf(" time zhetrd_he2hb = %6.2f\n" , GetTimerValue(start,st1)/1000.);
#endif
/* copy the input matrix into WORK(INDWRK) with band storage */
/* PAY ATTENTION THAT work[indwrk] should be able to be of size lda2*n which it should be checked in any future modification of lwork.*/
magma_int_t lda2 = 2*nb; //nb+1+(nb-1);
magmaDoubleComplex* A2 = &work[indwrk];
memset(A2 , 0, n*lda2*sizeof(magmaDoubleComplex));
for (magma_int_t j = 0; j < n-nb; j++)
{
cblas_zcopy(nb+1, &a[j*(lda+1)], 1, &A2[j*lda2], 1);
memset(&a[j*(lda+1)], 0, (nb+1)*sizeof(magmaDoubleComplex));
a[nb + j*(lda+1)] = c_one;
}
for (magma_int_t j = 0; j < nb; j++)
{
cblas_zcopy(nb-j, &a[(j+n-nb)*(lda+1)], 1, &A2[(j+n-nb)*lda2], 1);
memset(&a[(j+n-nb)*(lda+1)], 0, (nb-j)*sizeof(magmaDoubleComplex));
}
#ifdef ENABLE_TIMER
st2 = get_current_time();
printf(" time zhetrd_convert = %6.2f\n" , GetTimerValue(st1,st2)/1000.);
#endif
magma_zhetrd_hb2st(threads, uplo, n, nb, Vblksiz, A2, lda2, w, &rwork[inde], &work[indV2], ldv, &work[indTAU2], wantz, &work[indT2], ldt);
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time zhetrd_hb2st = %6.2f\n" , GetTimerValue(st2,end)/1000.);
printf(" time zhetrd = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
/* For eigenvalues only, call DSTERF. For eigenvectors, first call
ZSTEDC to generate the eigenvector matrix, WORK(INDWRK), of the
tridiagonal matrix, then call ZUNMTR to multiply it to the Householder
transformations represented as Householder vectors in A. */
if (! wantz) {
#ifdef ENABLE_TIMER
start = get_current_time();
#endif
lapackf77_dsterf(&n, w, &rwork[inde], info);
magma_dmove_eig(range, n, w, &il, &iu, vl, vu, m);
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time dstedc = %6.2f\n", GetTimerValue(start,end)/1000.);
#endif
} else {
#ifdef ENABLE_TIMER
start = get_current_time();
#endif
if (MAGMA_SUCCESS != magma_dmalloc( &dwork, 3*n*(n/2 + 1) )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_zstedx(range, n, vl, vu, il, iu, w, &rwork[inde],
&work[indwrk], n, &rwork[indrwk],
llrwk, iwork, liwork, dwork, info);
magma_free( dwork );
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time zstedx = %6.2f\n", GetTimerValue(start,end)/1000.);
start = get_current_time();
#endif
magmaDoubleComplex *dZ;
magma_int_t lddz = n;
magmaDoubleComplex *da;
magma_int_t ldda = n;
magma_dmove_eig(range, n, w, &il, &iu, vl, vu, m);
if (MAGMA_SUCCESS != magma_zmalloc( &dZ, *m*lddz)) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if (MAGMA_SUCCESS != magma_zmalloc( &da, n*ldda )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_zbulge_back(threads, uplo, n, nb, *m, Vblksiz, &work[indwrk + n * (il-1)], n, dZ, lddz,
&work[indV2], ldv, &work[indTAU2], &work[indT2], ldt, info);
#ifdef ENABLE_TIMER
st1 = get_current_time();
printf(" time zbulge_back = %6.2f\n" , GetTimerValue(start,st1)/1000.);
#endif
magma_zsetmatrix( n, n, a, lda, da, ldda );
magma_zunmqr_gpu_2stages(MagmaLeft, MagmaNoTrans, n-nb, *m, n-nb, da+nb, ldda,
dZ+nb, n, dT1, nb, info);
magma_zgetmatrix( n, *m, dZ, lddz, a, lda );
magma_free(dT1);
magma_free(dZ);
magma_free(da);
#ifdef ENABLE_TIMER
end = get_current_time();
printf(" time zunmqr + copy = %6.2f\n", GetTimerValue(st1,end)/1000.);
printf(" time eigenvectors backtransf. = %6.2f\n" , GetTimerValue(start,end)/1000.);
#endif
}
/* If matrix was scaled, then rescale eigenvalues appropriately. */
if (iscale == 1) {
if (*info == 0) {
imax = n;
} else {
imax = *info - 1;
}
d__1 = 1. / sigma;
blasf77_dscal(&imax, &d__1, w, &ione);
}
work[0] = MAGMA_Z_MAKE( lwmin * (1. + lapackf77_dlamch("Epsilon")), 0.); // round up
rwork[0] = lrwmin * (1. + lapackf77_dlamch("Epsilon"));
iwork[0] = liwmin;
return *info;
} /* magma_zheevdx_2stage */
extern "C" magma_int_t
magma_zgeqrf(magma_int_t m, magma_int_t n,
cuDoubleComplex *a, magma_int_t lda, cuDoubleComplex *tau,
cuDoubleComplex *work, magma_int_t lwork,
magma_int_t *info )
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2012
Purpose
=======
ZGEQRF computes a QR factorization of a COMPLEX_16 M-by-N matrix A:
A = Q * R. This version does not require work space on the GPU
passed as input. GPU memory is allocated in the routine.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
A (input/output) COMPLEX_16 array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
Higher performance is achieved if A is in pinned memory, e.g.
allocated using magma_malloc_pinned.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,M).
TAU (output) COMPLEX_16 array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
WORK (workspace/output) COMPLEX_16 array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
Higher performance is achieved if WORK is in pinned memory, e.g.
allocated using magma_malloc_pinned.
LWORK (input) INTEGER
The dimension of the array WORK. LWORK >= N*NB,
where NB can be obtained through magma_get_zgeqrf_nb(M).
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.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define a_ref(a_1,a_2) ( a+(a_2)*(lda) + (a_1))
#define da_ref(a_1,a_2) (da+(a_2)*ldda + (a_1))
cuDoubleComplex *da, *dwork;
cuDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t i, k, lddwork, old_i, old_ib;
magma_int_t ib, ldda;
/* Function Body */
*info = 0;
magma_int_t nb = magma_get_zgeqrf_nb(min(m, n));
magma_int_t lwkopt = n * nb;
work[0] = MAGMA_Z_MAKE( (double)lwkopt, 0 );
int lquery = (lwork == -1);
if (m < 0) {
*info = -1;
} else if (n < 0) {
*info = -2;
} else if (lda < max(1,m)) {
*info = -4;
} else if (lwork < max(1,n) && ! lquery) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
work[0] = c_one;
return *info;
}
lddwork = ((n+31)/32)*32;
ldda = ((m+31)/32)*32;
magma_int_t num_gpus = magma_num_gpus();
if( num_gpus > 1 ) {
/* call multiple-GPU interface */
return magma_zgeqrf4(num_gpus, m, n, a, lda, tau, work, lwork, info);
}
if (MAGMA_SUCCESS != magma_zmalloc( &da, (n)*ldda + nb*lddwork )) {
/* Switch to the "out-of-core" (out of GPU-memory) version */
return magma_zgeqrf_ooc(m, n, a, lda, tau, work, lwork, info);
}
cudaStream_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
dwork = da + ldda*(n);
if ( (nb > 1) && (nb < k) ) {
/* Use blocked code initially */
magma_zsetmatrix_async( (m), (n-nb),
a_ref(0,nb), lda,
da_ref(0,nb), ldda, stream[0] );
old_i = 0; old_ib = nb;
for (i = 0; i < k-nb; i += nb) {
ib = min(k-i, nb);
if (i>0){
magma_zgetmatrix_async( (m-i), ib,
da_ref(i,i), ldda,
a_ref(i,i), lda, stream[1] );
magma_zgetmatrix_async( i, ib,
da_ref(0,i), ldda,
a_ref(0,i), lda, stream[0] );
/* Apply H' to A(i:m,i+2*ib:n) from the left */
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, n-old_i-2*old_ib, old_ib,
da_ref(old_i, old_i), ldda, dwork, lddwork,
da_ref(old_i, old_i+2*old_ib), ldda, dwork+old_ib, lddwork);
}
magma_queue_sync( stream[1] );
magma_int_t rows = m-i;
lapackf77_zgeqrf(&rows, &ib, a_ref(i,i), &lda, tau+i, work, &lwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_zlarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib, a_ref(i,i), &lda, tau+i, work, &ib);
zpanel_to_q(MagmaUpper, ib, a_ref(i,i), lda, work+ib*ib);
magma_zsetmatrix( rows, ib, a_ref(i,i), lda, da_ref(i,i), ldda );
zq_to_panel(MagmaUpper, ib, a_ref(i,i), lda, work+ib*ib);
if (i + ib < n) {
magma_zsetmatrix( ib, ib, work, ib, dwork, lddwork );
if (i+ib < k-nb)
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left */
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
da_ref(i, i ), ldda, dwork, lddwork,
da_ref(i, i+ib), ldda, dwork+ib, lddwork);
else
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n-i-ib, ib,
da_ref(i, i ), ldda, dwork, lddwork,
da_ref(i, i+ib), ldda, dwork+ib, lddwork);
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
if (i!=0)
magma_zgetmatrix( m, ib, da_ref(0,i), ldda, a_ref(0,i), lda );
magma_int_t rows = m-i;
lapackf77_zgeqrf(&rows, &ib, a_ref(i,i), &lda, tau+i, work, &lwork, info);
}
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_free( da );
return *info;
} /* magma_zgeqrf */
/**
Purpose
-------
ZUNGQR generates an M-by-N COMPLEX_16 matrix Q with orthonormal columns,
which is defined as the first N columns of a product of K elementary
reflectors of order M
Q = H(1) H(2) . . . H(k)
as returned by ZGEQRF.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix Q. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix Q. M >= N >= 0.
@param[in]
k INTEGER
The number of elementary reflectors whose product defines the
matrix Q. N >= K >= 0.
@param[in,out]
A COMPLEX_16 array A, dimension (LDDA,N).
On entry, the i-th column must contain the vector
which defines the elementary reflector H(i), for
i = 1,2,...,k, as returned by ZGEQRF_GPU in the
first k columns of its array argument A.
On exit, the M-by-N matrix Q.
@param[in]
lda INTEGER
The first dimension of the array A. LDA >= max(1,M).
@param[in]
tau COMPLEX_16 array, dimension (K)
TAU(i) must contain the scalar factor of the elementary
reflector H(i), as returned by ZGEQRF_GPU.
@param[in]
T COMPLEX_16 array, dimension (NB, min(M,N)).
T contains the T matrices used in blocking the elementary
reflectors H(i), e.g., this can be the 6th argument of
magma_zgeqrf_gpu (except stored on the CPU, not the GPU).
@param[in]
nb INTEGER
This is the block size used in ZGEQRF_GPU, and correspondingly
the size of the T matrices, used in the factorization, and
stored in T.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument has an illegal value
@ingroup magma_zgeqrf_comp
********************************************************************/
extern "C" magma_int_t
magma_zungqr_m(
magma_int_t m, magma_int_t n, magma_int_t k,
magmaDoubleComplex *A, magma_int_t lda,
magmaDoubleComplex *tau,
magmaDoubleComplex *T, magma_int_t nb,
magma_int_t *info)
{
#define A(i,j) ( A + (i) + (j)*lda )
#define dA(d,i,j) (dA[d] + (i) + (j)*ldda)
#define dT(d,i,j) (dT[d] + (i) + (j)*nb)
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t m_kk, n_kk, k_kk, mi;
magma_int_t lwork, ldwork;
magma_int_t i, ib, ki, kk, iinfo;
magmaDoubleComplex *work;
*info = 0;
if (m < 0) {
*info = -1;
} else if ((n < 0) || (n > m)) {
*info = -2;
} else if ((k < 0) || (k > n)) {
*info = -3;
} else if (lda < max(1,m)) {
*info = -5;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
if (n <= 0) {
return *info;
}
magma_int_t di, dn;
magma_int_t dpanel;
magma_int_t ngpu = magma_num_gpus();
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
// Allocate memory on GPUs for A and workspaces
magma_int_t ldda = ((m + 31) / 32) * 32;
magma_int_t lddwork = ((n + 31) / 32) * 32;
magma_int_t min_lblocks = (n / nb) / ngpu; // min. blocks per gpu
magma_int_t last_dev = (n / nb) % ngpu; // device with last block
magma_int_t nlocal[ MagmaMaxGPUs ] = { 0 };
magmaDoubleComplex *dA[ MagmaMaxGPUs ] = { NULL };
magmaDoubleComplex *dT[ MagmaMaxGPUs ] = { NULL };
magmaDoubleComplex *dV[ MagmaMaxGPUs ] = { NULL };
magmaDoubleComplex *dW[ MagmaMaxGPUs ] = { NULL };
magma_queue_t stream[ MagmaMaxGPUs ] = { NULL };
for( int d = 0; d < ngpu; ++d ) {
// example with n = 75, nb = 10, ngpu = 3
// min_lblocks = 2
// last_dev = 1
// gpu 0: 2 blocks, cols: 0- 9, 30-39, 60-69
// gpu 1: 1+ blocks, cols: 10-19, 40-49, 70-74 (partial)
// gpu 2: 1 block, cols: 20-29, 50-59
magma_setdevice( d );
nlocal[d] = min_lblocks*nb;
if ( d < last_dev ) {
nlocal[d] += nb;
}
else if ( d == last_dev ) {
nlocal[d] += (n % nb);
}
ldwork = nlocal[d]*ldda // dA
+ nb*m // dT
+ nb*ldda // dV
+ nb*lddwork; // dW
if ( MAGMA_SUCCESS != magma_zmalloc( &dA[d], ldwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
goto CLEANUP;
}
dT[d] = dA[d] + nlocal[d]*ldda;
dV[d] = dT[d] + nb*m;
dW[d] = dV[d] + nb*ldda;
magma_queue_create( &stream[d] );
}
trace_init( 1, ngpu, 1, stream );
// first kk columns are handled by blocked method.
// ki is start of 2nd-to-last block
if ((nb > 1) && (nb < k)) {
ki = (k - nb - 1) / nb * nb;
kk = min(k, ki + nb);
} else {
ki = 0;
kk = 0;
}
// Allocate CPU work space
// n*nb for zungqr workspace
lwork = n * nb;
magma_zmalloc_cpu( &work, lwork );
if (work == NULL) {
*info = MAGMA_ERR_HOST_ALLOC;
goto CLEANUP;
}
// Use unblocked code for the last or only block.
if (kk < n) {
trace_cpu_start( 0, "ungqr", "ungqr last block" );
m_kk = m - kk;
n_kk = n - kk;
k_kk = k - kk;
dpanel = (kk / nb) % ngpu;
di = ((kk / nb) / ngpu) * nb;
magma_setdevice( dpanel );
lapackf77_zungqr( &m_kk, &n_kk, &k_kk,
A(kk, kk), &lda,
&tau[kk], work, &lwork, &iinfo );
magma_zsetmatrix( m_kk, n_kk,
A(kk, kk), lda,
dA(dpanel, kk, di), ldda );
// Set A(1:kk,kk+1:n) to zero.
magmablas_zlaset( MagmaFull, kk, n - kk, c_zero, c_zero, dA(dpanel, 0, di), ldda );
trace_cpu_end( 0 );
}
if (kk > 0) {
// Use blocked code
// send T to all GPUs
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set T" );
magma_zsetmatrix_async( nb, min(m,n), T, nb, dT[d], nb, stream[d] );
trace_gpu_end( d, 0 );
}
// stream: set Aii (V) --> laset --> laset --> larfb --> [next]
// CPU has no computation
for( i = ki; i >= 0; i -= nb ) {
ib = min(nb, k - i);
mi = m - i;
dpanel = (i / nb) % ngpu;
di = ((i / nb) / ngpu) * nb;
// Send current panel to the GPUs
lapackf77_zlaset( "Upper", &ib, &ib, &c_zero, &c_one, A(i, i), &lda );
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
trace_gpu_start( d, 0, "set", "set V" );
magma_zsetmatrix_async( mi, ib,
A(i, i), lda,
dV[d], ldda, stream[d] );
trace_gpu_end( d, 0 );
}
// set panel to identity
magma_setdevice( dpanel );
magmablasSetKernelStream( stream[dpanel] );
trace_gpu_start( dpanel, 0, "laset", "laset" );
magmablas_zlaset( MagmaFull, i, ib, c_zero, c_zero, dA(dpanel, 0, di), ldda );
magmablas_zlaset( MagmaFull, mi, ib, c_zero, c_one, dA(dpanel, i, di), ldda );
trace_gpu_end( dpanel, 0 );
if (i < n) {
// Apply H to A(i:m,i:n) from the left
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magmablasSetKernelStream( stream[d] );
magma_indices_1D_bcyclic( nb, ngpu, d, i, n, &di, &dn );
trace_gpu_start( d, 0, "larfb", "larfb" );
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise,
mi, dn-di, ib,
dV[d], ldda, dT(d,0,i), nb,
dA(d, i, di), ldda, dW[d], lddwork );
trace_gpu_end( d, 0 );
}
}
}
}
// copy result back to CPU
trace_cpu_start( 0, "get", "get A" );
magma_zgetmatrix_1D_col_bcyclic( m, n, dA, ldda, A, lda, ngpu, nb );
trace_cpu_end( 0 );
#ifdef TRACING
char name[80];
snprintf( name, sizeof(name), "zungqr-n%d-ngpu%d.svg", m, ngpu );
trace_finalize( name, "trace.css" );
#endif
CLEANUP:
for( int d = 0; d < ngpu; ++d ) {
magma_setdevice( d );
magma_free( dA[d] );
magma_queue_destroy( stream[d] );
}
magma_free_cpu( work );
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
} /* magma_zungqr */
extern "C" magma_int_t
magma_zhegst(magma_int_t itype, char uplo, magma_int_t n,
magmaDoubleComplex *a, magma_int_t lda,
magmaDoubleComplex *b, magma_int_t ldb, magma_int_t *info)
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
ZHEGST reduces a complex Hermitian-definite generalized
eigenproblem to standard form.
If ITYPE = 1, the problem is A*x = lambda*B*x,
and A is overwritten by inv(U**H)*A*inv(U) or inv(L)*A*inv(L**H)
If ITYPE = 2 or 3, the problem is A*B*x = lambda*x or
B*A*x = lambda*x, and A is overwritten by U*A*U**H or L**H*A*L.
B must have been previously factorized as U**H*U or L*L**H by ZPOTRF.
Arguments
=========
ITYPE (input) INTEGER
= 1: compute inv(U**H)*A*inv(U) or inv(L)*A*inv(L**H);
= 2 or 3: compute U*A*U**H or L**H*A*L.
UPLO (input) CHARACTER*1
= 'U': Upper triangle of A is stored and B is factored as
U**H*U;
= 'L': Lower triangle of A is stored and B is factored as
L*L**H.
N (input) INTEGER
The order of the matrices A and B. N >= 0.
A (input/output) COMPLEX_16 array, dimension (LDA,N)
On entry, the Hermitian 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 transformed matrix, stored in the
same format as A.
LDA (input) INTEGER
The leading dimension of the array A. LDA >= max(1,N).
B (input) COMPLEX_16 array, dimension (LDB,N)
The triangular factor from the Cholesky factorization of B,
as returned by ZPOTRF.
LDB (input) INTEGER
The leading dimension of the array B. LDB >= max(1,N).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
=====================================================================*/
char uplo_[2] = {uplo, 0};
magma_int_t nb;
magma_int_t k, kb, kb2;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex c_half = MAGMA_Z_HALF;
magmaDoubleComplex c_neg_half = MAGMA_Z_NEG_HALF;
magmaDoubleComplex *dw;
magma_int_t ldda = n;
magma_int_t lddb = n;
double d_one = 1.0;
int upper = lapackf77_lsame(uplo_, "U");
/* Test the input parameters. */
*info = 0;
if (itype<1 || itype>3){
*info = -1;
}else if ((! upper) && (! lapackf77_lsame(uplo_, "L"))) {
*info = -2;
} else if (n < 0) {
*info = -3;
} else if (lda < max(1,n)) {
*info = -5;
}else if (ldb < max(1,n)) {
*info = -7;
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return */
if ( n == 0 )
return *info;
if (MAGMA_SUCCESS != magma_zmalloc( &dw, 2*n*n )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
nb = magma_get_zhegst_nb(n);
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
magma_zsetmatrix( n, n, A(0, 0), lda, dA(0, 0), ldda );
magma_zsetmatrix( n, n, B(0, 0), ldb, dB(0, 0), lddb );
/* Use hybrid blocked code */
if (itype==1) {
if (upper) {
/* Compute inv(U')*A*inv(U) */
for(k = 0; k<n; k+=nb){
kb = min(n-k,nb);
kb2= min(n-k-nb,nb);
/* Update the upper triangle of A(k:n,k:n) */
lapackf77_zhegst( &itype, uplo_, &kb, A(k,k), &lda, B(k,k), &ldb, info);
magma_zsetmatrix_async( kb, kb,
A(k, k), lda,
dA(k, k), ldda, stream[0] );
if(k+kb<n){
magma_ztrsm(MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
kb, n-k-kb,
c_one, dB(k,k), lddb,
dA(k,k+kb), ldda);
magma_queue_sync( stream[0] );
magma_zhemm(MagmaLeft, MagmaUpper,
kb, n-k-kb,
c_neg_half, dA(k,k), ldda,
dB(k,k+kb), lddb,
c_one, dA(k, k+kb), ldda);
magma_zher2k(MagmaUpper, MagmaConjTrans,
n-k-kb, kb,
c_neg_one, dA(k,k+kb), ldda,
dB(k,k+kb), lddb,
d_one, dA(k+kb,k+kb), ldda);
magma_zgetmatrix_async( kb2, kb2,
dA(k+kb, k+kb), ldda,
A(k+kb, k+kb), lda, stream[1] );
magma_zhemm(MagmaLeft, MagmaUpper,
kb, n-k-kb,
c_neg_half, dA(k,k), ldda,
dB(k,k+kb), lddb,
c_one, dA(k, k+kb), ldda);
magma_ztrsm(MagmaRight, MagmaUpper, MagmaNoTrans, MagmaNonUnit,
kb, n-k-kb,
c_one ,dB(k+kb,k+kb), lddb,
dA(k,k+kb), ldda);
magma_queue_sync( stream[1] );
}
}
magma_queue_sync( stream[0] );
}
else {
/* Compute inv(L)*A*inv(L') */
for(k = 0; k<n; k+=nb){
kb= min(n-k,nb);
kb2= min(n-k-nb,nb);
/* Update the lower triangle of A(k:n,k:n) */
lapackf77_zhegst( &itype, uplo_, &kb, A(k,k), &lda, B(k,k), &ldb, info);
magma_zsetmatrix_async( kb, kb,
A(k, k), lda,
dA(k, k), ldda, stream[0] );
if(k+kb<n){
magma_ztrsm(MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
n-k-kb, kb,
c_one, dB(k,k), lddb,
dA(k+kb,k), ldda);
magma_queue_sync( stream[0] );
magma_zhemm(MagmaRight, MagmaLower,
n-k-kb, kb,
c_neg_half, dA(k,k), ldda,
dB(k+kb,k), lddb,
c_one, dA(k+kb, k), ldda);
magma_zher2k(MagmaLower, MagmaNoTrans,
n-k-kb, kb,
c_neg_one, dA(k+kb,k), ldda,
dB(k+kb,k), lddb,
d_one, dA(k+kb,k+kb), ldda);
magma_zgetmatrix_async( kb2, kb2,
dA(k+kb, k+kb), ldda,
A(k+kb, k+kb), lda, stream[1] );
magma_zhemm(MagmaRight, MagmaLower,
n-k-kb, kb,
c_neg_half, dA(k,k), ldda,
dB(k+kb,k), lddb,
c_one, dA(k+kb, k), ldda);
magma_ztrsm(MagmaLeft, MagmaLower, MagmaNoTrans, MagmaNonUnit,
n-k-kb, kb,
c_one, dB(k+kb,k+kb), lddb,
dA(k+kb,k), ldda);
}
magma_queue_sync( stream[1] );
}
}
magma_queue_sync( stream[0] );
}
else {
if (upper) {
/* Compute U*A*U' */
for(k = 0; k<n; k+=nb){
kb= min(n-k,nb);
magma_zgetmatrix_async( kb, kb,
dA(k, k), ldda,
A(k, k), lda, stream[0] );
/* Update the upper triangle of A(1:k+kb-1,1:k+kb-1) */
if(k>0){
magma_ztrmm(MagmaLeft, MagmaUpper, MagmaNoTrans, MagmaNonUnit,
k, kb,
c_one ,dB(0,0), lddb,
dA(0,k), ldda);
magma_zhemm(MagmaRight, MagmaUpper,
k, kb,
c_half, dA(k,k), ldda,
dB(0,k), lddb,
c_one, dA(0, k), ldda);
magma_queue_sync( stream[1] );
magma_zher2k(MagmaUpper, MagmaNoTrans,
k, kb,
c_one, dA(0,k), ldda,
dB(0,k), lddb,
d_one, dA(0,0), ldda);
magma_zhemm(MagmaRight, MagmaUpper,
k, kb,
c_half, dA(k,k), ldda,
dB(0,k), lddb,
c_one, dA(0, k), ldda);
magma_ztrmm(MagmaRight, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
k, kb,
c_one, dB(k,k), lddb,
dA(0,k), ldda);
}
magma_queue_sync( stream[0] );
lapackf77_zhegst( &itype, uplo_, &kb, A(k, k), &lda, B(k, k), &ldb, info);
magma_zsetmatrix_async( kb, kb,
A(k, k), lda,
dA(k, k), ldda, stream[1] );
}
magma_queue_sync( stream[1] );
}
else {
/* Compute L'*A*L */
for(k = 0; k<n; k+=nb){
kb= min(n-k,nb);
magma_zgetmatrix_async( kb, kb,
dA(k, k), ldda,
A(k, k), lda, stream[0] );
/* Update the lower triangle of A(1:k+kb-1,1:k+kb-1) */
if(k>0){
magma_ztrmm(MagmaRight, MagmaLower, MagmaNoTrans, MagmaNonUnit,
kb, k,
c_one ,dB(0,0), lddb,
dA(k,0), ldda);
magma_zhemm(MagmaLeft, MagmaLower,
kb, k,
c_half, dA(k,k), ldda,
dB(k,0), lddb,
c_one, dA(k, 0), ldda);
magma_queue_sync( stream[1] );
magma_zher2k(MagmaLower, MagmaConjTrans,
k, kb,
c_one, dA(k,0), ldda,
dB(k,0), lddb,
d_one, dA(0,0), ldda);
magma_zhemm(MagmaLeft, MagmaLower,
kb, k,
c_half, dA(k,k), ldda,
dB(k,0), lddb,
c_one, dA(k, 0), ldda);
magma_ztrmm(MagmaLeft, MagmaLower, MagmaConjTrans, MagmaNonUnit,
kb, k,
c_one, dB(k,k), lddb,
dA(k,0), ldda);
}
magma_queue_sync( stream[0] );
lapackf77_zhegst( &itype, uplo_, &kb, A(k,k), &lda, B(k,k), &ldb, info);
magma_zsetmatrix_async( kb, kb,
A(k, k), lda,
dA(k, k), ldda, stream[1] );
}
magma_queue_sync( stream[1] );
}
}
magma_zgetmatrix( n, n, dA(0, 0), ldda, A(0, 0), lda );
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
magma_free( dw );
return *info;
} /* magma_zhegst_gpu */
extern "C" magma_int_t
magma_zgeqrf2_gpu( magma_int_t m, magma_int_t n,
magmaDoubleComplex *dA, magma_int_t ldda,
magmaDoubleComplex *tau,
magma_int_t *info )
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
ZGEQRF computes a QR factorization of a complex M-by-N matrix A:
A = Q * R.
This version has LAPACK-complaint arguments.
This version assumes the computation runs through the NULL stream
and therefore is not overlapping some computation with communication.
Other versions (magma_zgeqrf_gpu and magma_zgeqrf3_gpu) store the
intermediate T matrices.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
dA (input/output) COMPLEX_16 array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix A.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
dividable by 16.
TAU (output) COMPLEX_16 array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define dA(a_1,a_2) ( dA+(a_2)*(ldda) + (a_1))
#define work_ref(a_1) ( work + (a_1))
#define hwork ( work + (nb)*(m))
magmaDoubleComplex *dwork;
magmaDoubleComplex *work;
magma_int_t i, k, ldwork, lddwork, old_i, old_ib, rows;
magma_int_t nbmin, nx, ib, nb;
magma_int_t lhwork, lwork;
/* Function Body */
*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;
}
k = min(m,n);
if (k == 0)
return *info;
nb = magma_get_zgeqrf_nb(m);
lwork = (m+n) * nb;
lhwork = lwork - (m)*nb;
if (MAGMA_SUCCESS != magma_zmalloc( &dwork, (n)*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
if (MAGMA_SUCCESS != magma_zmalloc_pinned( &work, lwork )) {
magma_free( dwork );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
nbmin = 2;
nx = nb;
ldwork = m;
lddwork= n;
if (nb >= nbmin && nb < k && nx < k) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nx; i += nb) {
ib = min(k-i, nb);
rows = m -i;
magma_zgetmatrix_async( rows, ib,
dA(i,i), ldda,
work_ref(i), ldwork, stream[1] );
if (i>0){
/* Apply H' to A(i:m,i+2*ib:n) from the left */
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, n-old_i-2*old_ib, old_ib,
dA(old_i, old_i ), ldda, dwork, lddwork,
dA(old_i, old_i+2*old_ib), ldda, dwork+old_ib, lddwork);
magma_zsetmatrix_async( old_ib, old_ib,
work_ref(old_i), ldwork,
dA(old_i, old_i), ldda, stream[0] );
}
magma_queue_sync( stream[1] );
lapackf77_zgeqrf(&rows, &ib, work_ref(i), &ldwork, tau+i, hwork, &lhwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_zlarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
work_ref(i), &ldwork, tau+i, hwork, &ib);
zpanel_to_q( MagmaUpper, ib, work_ref(i), ldwork, hwork+ib*ib );
magma_zsetmatrix( rows, ib, work_ref(i), ldwork, dA(i,i), ldda );
zq_to_panel( MagmaUpper, ib, work_ref(i), ldwork, hwork+ib*ib );
if (i + ib < n) {
magma_zsetmatrix( ib, ib, hwork, ib, dwork, lddwork );
if (i+nb < k-nx) {
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left */
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dA(i, i ), ldda, dwork, lddwork,
dA(i, i+ib), ldda, dwork+ib, lddwork);
}
else {
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n-i-ib, ib,
dA(i, i ), ldda, dwork, lddwork,
dA(i, i+ib), ldda, dwork+ib, lddwork);
magma_zsetmatrix( ib, ib,
work_ref(i), ldwork,
dA(i,i), ldda );
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
magma_free( dwork );
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
magma_zgetmatrix( rows, ib, dA(i, i), ldda, work, rows );
lhwork = lwork - rows*ib;
lapackf77_zgeqrf(&rows, &ib, work, &rows, tau+i, work+ib*rows, &lhwork, info);
magma_zsetmatrix( rows, ib, work, rows, dA(i, i), ldda );
}
magma_free_pinned( work );
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
return *info;
} /* magma_zgeqrf2_gpu */
extern "C" magma_int_t
magma_zgels3_gpu( char trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaDoubleComplex *dA, magma_int_t ldda,
magmaDoubleComplex *dB, magma_int_t lddb,
magmaDoubleComplex *hwork, magma_int_t lwork,
magma_int_t *info)
{
/* -- MAGMA (version 1.4.1) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
December 2013
Purpose
=======
Solves the overdetermined, least squares problem
min || A*X - C ||
using the QR factorization A.
The underdetermined problem (m < n) is not currently handled.
Arguments
=========
TRANS (input) CHARACTER*1
= 'N': the linear system involves A.
Only trans='N' is currently handled.
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. M >= N >= 0.
NRHS (input) INTEGER
The number of columns of the matrix C. NRHS >= 0.
A (input/output) COMPLEX_16 array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, A is overwritten by details of its QR
factorization as returned by ZGEQRF3.
LDDA (input) INTEGER
The leading dimension of the array A, LDDA >= M.
DB (input/output) COMPLEX_16 array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
LDDB (input) INTEGER
The leading dimension of the array DB. LDDB >= M.
HWORK (workspace/output) COMPLEX_16 array, dimension MAX(1,LWORK).
On exit, if INFO = 0, HWORK(1) returns the optimal LWORK.
LWORK (input) INTEGER
The dimension of the array HWORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_zgeqrf_nb( M ).
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the HWORK array.
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
===================================================================== */
#define a_ref(a_1,a_2) (dA + (a_2)*(ldda) + (a_1))
magmaDoubleComplex *dT, *tau;
magma_int_t k;
magma_int_t nb = magma_get_zgeqrf_nb(m);
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
int lquery = (lwork == -1);
hwork[0] = MAGMA_Z_MAKE( (double)lwkopt, 0. );
*info = 0;
/* For now, N is the only case working */
if ( (trans != 'N') && (trans != 'n' ) )
*info = -1;
else if (m < 0)
*info = -2;
else if (n < 0 || m < n) /* LQ is not handle for now*/
*info = -3;
else if (nrhs < 0)
*info = -4;
else if (ldda < max(1,m))
*info = -6;
else if (lddb < max(1,m))
*info = -8;
else if (lwork < lwkopt && ! lquery)
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = MAGMA_Z_ONE;
return *info;
}
/*
* Allocate temporary buffers
*/
int ldtwork = ( 2*k + ((n+31)/32)*32 )*nb;
if (nb < nrhs)
ldtwork = ( 2*k + ((n+31)/32)*32 )*nrhs;
if (MAGMA_SUCCESS != magma_zmalloc( &dT, ldtwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_zmalloc_cpu( &tau, k );
if ( tau == NULL ) {
magma_free( dT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_zgeqrf3_gpu( m, n, dA, ldda, tau, dT, info );
if ( *info == 0 ) {
magma_zgeqrs3_gpu( m, n, nrhs,
dA, ldda, tau, dT,
dB, lddb, hwork, lwork, info );
}
magma_free( dT );
magma_free_cpu(tau);
return *info;
}
/**
Purpose
-------
ZGEBRD reduces a general complex 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 COMPLEX_16 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 double precision array, dimension (min(M,N))
The diagonal elements of the bidiagonal matrix B:
D(i) = A(i,i).
@param[out]
e double precision 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 COMPLEX_16 array dimension (min(M,N))
The scalar factors of the elementary reflectors which
represent the orthogonal matrix Q. See Further Details.
@param[out]
taup COMPLEX_16 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) COMPLEX_16 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 complex scalars, and v and u are complex 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 complex scalars, and v and u are complex 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_zgesvd_comp
********************************************************************/
extern "C" magma_int_t
magma_zgebrd(
magma_int_t m, magma_int_t n,
magmaDoubleComplex *A, magma_int_t lda, double *d, double *e,
magmaDoubleComplex *tauq, magmaDoubleComplex *taup,
magmaDoubleComplex *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))
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex *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_zgebrd_nb( m, n );
ldda = m;
lwkopt = (m + n) * nb;
work[0] = magma_zmake_lwork( lwkopt );
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;
}
magma_queue_t queue = NULL;
magma_device_t cdev;
magma_getdevice( &cdev );
magma_queue_create( cdev, &queue );
magmaDoubleComplex *work2;
magma_int_t lwork2 = max(m,n);
if (MAGMA_SUCCESS != magma_zmalloc_cpu( &work2, lwork2 )) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
if (MAGMA_SUCCESS != magma_zmalloc( &dA, n*ldda + (m + n)*nb )) {
magma_free_cpu( work2 );
*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_zsetmatrix( m, n, A, lda, dA, ldda, queue );
}
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_zgetmatrix( nrow, nb,
dA(i, i), ldda,
A( i, i), lda, queue );
magma_zgetmatrix( nb, ncol - nb,
dA(i, i+nb), ldda,
A( i, i+nb), lda, queue );
}
magma_zlabrd_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,
work2, lwork2, queue ); // 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_zsetmatrix( nrow, nb,
work + nb, ldwrkx,
dwork + nb, ldwrkx, queue );
magma_zsetmatrix( ncol, nb,
work + (ldwrkx+1)*nb, ldwrky,
dwork + (ldwrkx+1)*nb, ldwrky, queue );
magma_zgemm( 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, queue );
magma_zgemm( MagmaNoTrans, MagmaNoTrans,
nrow, ncol, nb,
c_neg_one, dwork+nb, ldwrkx,
dA( i, i+nb ), ldda,
c_one, dA( i+nb, i+nb ), ldda, queue );
/* 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_Z_MAKE( d[j], 0. );
*A(j, j+1) = MAGMA_Z_MAKE( e[j], 0. );
}
} else {
jmax = i + nb;
for (j = i; j < jmax; ++j) {
*A(j, j ) = MAGMA_Z_MAKE( d[j], 0. );
*A(j+1, j ) = MAGMA_Z_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_zgetmatrix( nrow, ncol,
dA(i, i), ldda,
A( i, i), lda, queue );
}
lapackf77_zgebrd( &nrow, &ncol,
A(i, i), &lda, d+i, e+i,
tauq+i, taup+i, work, &lwork, &iinfo);
work[0] = magma_zmake_lwork( lwkopt );
magma_free_cpu( work2 );
magma_free( dA );
magma_queue_destroy( queue );
return *info;
} /* magma_zgebrd */
extern "C" void magma_zbulge_applyQ(
magma_int_t WANTZ, magma_side_t SIDE, magma_int_t NE, magma_int_t N, magma_int_t NB,
magma_int_t Vblksiz, magmaDoubleComplex *E, magma_int_t LDE,
magmaDoubleComplex *V, magmaDoubleComplex *TAU, magmaDoubleComplex *T,
magma_int_t *INFO, magmaDoubleComplex *dV, magmaDoubleComplex *dT,
magmaDoubleComplex *dE, magma_int_t copytype )
{
//%===========================
//% local variables
//%===========================
magmaDoubleComplex c_zero = MAGMA_Z_ZERO;
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magma_int_t LDT, LDV, firstcolj;
magma_int_t bg, nbGblk, rownbm, k, m, n;
magma_int_t st, ed, fst, vlen, vnb, colj, len;
magma_int_t blkid, vpos, taupos, tpos;
//magmaDoubleComplex *WORK;
magma_int_t LWORK;
magma_int_t cur_blksiz, avai_blksiz, ncolinvolvd;
magma_int_t nbgr, colst, coled, versionL, versionR;
magma_int_t blkcnt=-1;
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
*INFO=0;
versionL = 113;
versionR = 92;
LDT = Vblksiz;
LDV = NB+Vblksiz-1;
//blklen = LDV*Vblksiz;
nbGblk = plasma_ceildiv((N-1), Vblksiz);
//magma_zmalloc_cpu( &WORK, LWORK );
/* find the size of the matrix T V*/
findVTsiz(N, NB, Vblksiz, &blkcnt, &LDV);
/* Copy E & V & T to the GPU in dE and dV and dT
* depending on copytype:
* 1: mean copy only V
* 2: mean copy V and T
* 3: mean copy V, T and E
* */
if (copytype > 0) magma_zsetmatrix( LDV, blkcnt*Vblksiz, V, LDV, dV, LDV );
if (copytype > 1) magma_zsetmatrix( LDT, blkcnt*Vblksiz, T, LDT, dT, LDT );
if (copytype > 2) magma_zsetmatrix( N, NE, E, N, dE, N );
magmaDoubleComplex *dwork;
//ldwork = NE;
LWORK = 2*N*max(Vblksiz, 64);
if (MAGMA_SUCCESS != magma_zmalloc( &dwork, LWORK )) {
printf ("!!!! magma_zbulge_applyQ magma_alloc failed for: dwork\n" );
exit(-1);
}
/* SIDE LEFT meaning apply E = Q*E = (q_1*q_2*.....*q_n) * E ==> so traverse Vs in reverse order (forward) from q_n to q_1
* Also E is splitten by row meaning each apply consist in a block of row (horizontal block) */
/* SIDE RIGHT meaning apply E = E*Q = E * (q_1*q_2*.....*q_n) ==> so tarverse Vs in normal order (forward) from q_1 to q_n
* Also E is splitten by col meaning each apply consist in a block of col (vertical block) */
/* WANTZ = 1 meaning E is IDENTITY so form Q using optimized update.
* So we use the reverse order from small q to large one,
* so from q_n to q_1 so Left update to Identity.
* Use versionL 113 because in 114 we need to update the whole matrix and not in icreasing order.
* WANTZ = 2 meaning E is a full matrix and need to be updated from Left or Right so use normal update
* */
if (WANTZ == 1) {
versionL=113;
SIDE = MagmaLeft;
//set the matrix to Identity here to avoid copying it from the CPU
magmablas_zlaset( MagmaFull, N, N, c_zero, c_one, dE, N );
}
printf(" APPLY Q_v115 GPU with N %d NB %d Vblksiz %d SIDE %c versionL %d versionR %d WANTZ %d \n",
(int) N, (int) NB, (int) Vblksiz, SIDE, (int) versionL, (int) versionR, (int) WANTZ);
#if defined(USESTREAM)
magma_int_t N2=N/2;
magma_int_t N1=N-N2;
printf("using stream\n");
magma_queue_t stream[2];
magma_queue_create( &stream[0] );
magma_queue_create( &stream[1] );
#endif
if (SIDE == MagmaLeft) {
if (versionL == 113) {
for (bg = nbGblk; bg > 0; bg--) {
firstcolj = (bg-1)*Vblksiz + 1;
if (bg == nbGblk)
rownbm = plasma_ceildiv((N-(firstcolj)), NB); // last blk has size=1 used for complex to handle A(N,N-1)
else
rownbm = plasma_ceildiv((N-(firstcolj+1)), NB);
for (m = rownbm; m > 0; m--) {
vlen = 0;
vnb = 0;
colj = (bg-1)*Vblksiz; // for k=0; I compute the fst and then can remove it from the loop
fst = (rownbm -m)*NB+colj +1;
for (k=0; k < Vblksiz; k++) {
colj = (bg-1)*Vblksiz + k;
st = (rownbm -m)*NB+colj +1;
ed = min(st+NB-1, N-1);
if (st > ed) break;
if ((st == ed) && (colj != N-2)) break;
vlen=ed-fst+1;
vnb=k+1;
}
colst = (bg-1)*Vblksiz;
findVTpos(N, NB, Vblksiz, colst, fst, &vpos, &taupos, &tpos, &blkid);
printf("voici bg %d m %d vlen %d vnb %d fcolj %d vpos %d taupos %d \n", (int) bg, (int) m, (int) vlen, (int) vnb, (int) colst+1, (int) vpos+1, (int) taupos+1);
if ((vlen > 0) && (vnb > 0)) {
if (WANTZ == 1) {
len = N-colst;
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, vlen, len, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(fst,colst), LDE, dwork, len);
} else {
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, vlen, NE, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(fst,0), LDE, dwork, NE);
}
}
}
}
} else if (versionL == 114) {
rownbm = plasma_ceildiv((N-1), NB);
for (m = rownbm; m > 0; m--) {
ncolinvolvd = min(N-1, m*NB);
avai_blksiz=min(Vblksiz, ncolinvolvd);
nbgr = plasma_ceildiv(ncolinvolvd, avai_blksiz);
for (n = nbgr; n > 0; n--) {
vlen = 0;
vnb = 0;
cur_blksiz = min(ncolinvolvd-(n-1)*avai_blksiz, avai_blksiz);
colst = (n-1)*avai_blksiz;
coled = colst + cur_blksiz -1;
fst = (rownbm -m)*NB+colst +1;
for (colj=colst; colj <= coled; colj++) {
st = (rownbm -m)*NB+colj +1;
ed = min(st+NB-1, N-1);
if (st > ed) break;
if ((st == ed) && (colj != N-2)) break;
vlen=ed-fst+1;
vnb=vnb+1;
}
findVTpos(N, NB, Vblksiz, colst, fst, &vpos, &taupos, &tpos, &blkid);
//printf("voici bg %d m %d vlen %d vnb %d fcolj %d vpos %d taupos %d \n", bg, m, vlen, vnb, colst+1, vpos+1, taupos+1);
if ((vlen > 0) && (vnb > 0)) {
#if defined(USESTREAM)
magmablasSetKernelStream(stream[0]);
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, vlen, N1, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(fst,0), LDE, dwork, N1);
magmablasSetKernelStream(stream[1]);
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, vlen, N2, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(fst,N1), LDE, &dwork[N1*Vblksiz], N2);
#else
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, vlen, NE, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(fst,0), LDE, dwork, NE);
#endif
}
}
}
}
} else if (SIDE == MagmaRight) {
if (versionR == 91) {
for (bg =1; bg <= nbGblk; bg++) {
firstcolj = (bg-1)*Vblksiz + 1;
rownbm = plasma_ceildiv((N-(firstcolj+1)), NB);
if (bg == nbGblk) rownbm = plasma_ceildiv((N-(firstcolj)), NB); // last blk has size=1 used for complex to handle A(N,N-1)
for (m = 1; m <= rownbm; m++) {
vlen = 0;
vnb = 0;
// for k=0; I compute the fst and then can remove it from the loop
colj = (bg-1)*Vblksiz;
fst = (rownbm -m)*NB+colj +1;
for (k=0; k < Vblksiz; k++) {
colj = (bg-1)*Vblksiz + k;
st = (rownbm -m)*NB+colj +1;
ed = min(st+NB-1, N-1);
if (st > ed) break;
if ((st == ed) && (colj != N-2)) break;
vlen=ed-fst+1;
vnb=k+1;
}
colj = (bg-1)*Vblksiz;
findVTpos(N, NB, Vblksiz, colj, fst, &vpos, &taupos, &tpos, &blkid);
//printf("voici bg %d m %d vlen %d vnb %d fcolj %d vpos %d taupos %d \n", bg, m, vlen, vnb, colj, vpos, taupos);
if ((vlen > 0) && (vnb > 0)) {
#if defined(USESTREAM)
magmablasSetKernelStream(stream[0]);
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, N1, vlen, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(0, fst), LDE, dwork, N1);
magmablasSetKernelStream(stream[1]);
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, N2, vlen, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(N1, fst), LDE, &dwork[N1*Vblksiz], N2);
#else
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, NE, vlen, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(0, fst), LDE, dwork, NE);
#endif
}
}
}
} else if (versionR == 92) {
rownbm = plasma_ceildiv((N-1), NB);
for (m = 1; m <= rownbm; m++) {
ncolinvolvd = min(N-1, m*NB);
avai_blksiz=min(Vblksiz, ncolinvolvd);
nbgr = plasma_ceildiv(ncolinvolvd, avai_blksiz);
for (n = 1; n <= nbgr; n++) {
vlen = 0;
vnb = 0;
cur_blksiz = min(ncolinvolvd-(n-1)*avai_blksiz, avai_blksiz);
colst = (n-1)*avai_blksiz;
coled = colst + cur_blksiz -1;
fst = (rownbm -m)*NB+colst +1;
for (colj=colst; colj <= coled; colj++) {
st = (rownbm -m)*NB+colj +1;
ed = min(st+NB-1, N-1);
if (st > ed) break;
if ((st == ed) && (colj != N-2)) break;
vlen=ed-fst+1;
vnb=vnb+1;
}
findVTpos(N, NB, Vblksiz, colst, fst, &vpos, &taupos, &tpos, &blkid);
if ((vlen > 0) && (vnb > 0)) {
#if defined(USESTREAM)
magmablasSetKernelStream(stream[0]);
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, N1, vlen, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(0, fst), LDE, dwork, N1);
magmablasSetKernelStream(stream[1]);
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, N2, vlen, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(N1, fst), LDE, &dwork[N1*Vblksiz], N2);
#else
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, NE, vlen, vnb, dV(vpos), LDV, dT(tpos), LDT, dE(0, fst), LDE, dwork, NE);
#endif
}
}
}
}
} else {
printf("ERROR SIDE %d\n", SIDE);
}
#if defined(USESTREAM)
magma_queue_destroy( stream[0] );
magma_queue_destroy( stream[1] );
#endif
magmablasSetKernelStream( orig_stream );
}
extern "C" magma_int_t
magma_zpotrf2_mgpu(int num_gpus, char uplo, magma_int_t m, magma_int_t n,
magma_int_t off_i, magma_int_t off_j, magma_int_t nb,
magmaDoubleComplex **d_lA, magma_int_t ldda,
magmaDoubleComplex **d_lP, magma_int_t lddp,
magmaDoubleComplex *a, magma_int_t lda, magma_int_t h,
magma_queue_t stream[][3], magma_event_t event[][5],
magma_int_t *info )
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
ZPOTRF computes the Cholesky factorization of a complex Hermitian
positive definite matrix dA.
The factorization has the form
dA = U**H * U, if UPLO = 'U', or
dA = L * L**H, if UPLO = 'L',
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
=========
UPLO (input) CHARACTER*1
= 'U': Upper triangle of dA is stored;
= 'L': Lower triangle of dA is stored.
N (input) INTEGER
The order of the matrix dA. N >= 0.
dA (input/output) COMPLEX_16 array on the GPU, dimension (LDDA,N)
On entry, the Hermitian matrix dA. If UPLO = 'U', 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 = 'L', 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.
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization dA = U**H * U or dA = L * L**H.
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,N).
To benefit from coalescent memory accesses LDDA must be
dividable by 16.
INFO (output) 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.
===================================================================== */
magma_int_t j, jb, nb0, nb2, dd, d, id, j_local, j_local2, buf;
char uplo_[2] = {uplo, 0};
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
double d_one = 1.0;
double d_neg_one = -1.0;
int upper = lapackf77_lsame(uplo_, "U");
magmaDoubleComplex *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 ZTRSM_WORK
magmaDoubleComplex *d_dinvA[MagmaMaxGPUs][2], *d_x[MagmaMaxGPUs][2]; /* used by ztrsm_work */
#endif
*info = 0;
if ( (! upper) && (! lapackf77_lsame(uplo_, "L")) ) {
*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, stream );
/* Use blocked code. */
if (upper) {
/* ---------------------------------------------- */
/* Upper-triangular case */
/* > Compute the Cholesky factorization A = U'*U. */
/* ---------------------------------------------- */
#if defined(PRECISION_d) && defined(ZTRSM_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_zmalloc( &d_dinvA[d][j], nb*nb );
magma_zmalloc( &d_x[d][j], n*nb );
cudaMemset(d_dinvA[d][j], 0, nb*nb*sizeof(magmaDoubleComplex));
cudaMemset(d_x[d][j], 0, n*nb*sizeof(magmaDoubleComplex));
}
}
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_zsetmatrix_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_zherk(MagmaUpper, MagmaConjTrans, 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_zgetmatrix_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_zgemm(MagmaConjTrans, 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_zpotrf(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_zsetmatrix_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_zsetmatrix_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(ZTRSM_WORK)
magmablas_ztrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, 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_ztrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, 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_ztrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldda,
dlA(d, j, nb*j_local2), ldda);
*/
magma_ztrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, 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_zgetmatrix_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(ZTRSM_WORK)
magmablas_ztrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, 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_ztrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, 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(ZTRSM_WORK)
magmablas_ztrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_ztrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, 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 ztrsm */
} /* end of for j=1, .., n */
} else {
/* -------------------------------------------- */
/* Lower-triangular case */
/* Compute the Cholesky factorization A = L*L'. */
/* -------------------------------------------- */
#if defined(PRECISION_d) && defined(ZTRSM_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_zmalloc( &d_dinvA[d][j], nb*nb );
magma_zmalloc( &d_x[d][j], nb*m );
cudaMemset(d_dinvA[d][j], 0, nb*nb*sizeof(magmaDoubleComplex));
cudaMemset(d_x[d][j], 0, nb* m*sizeof(magmaDoubleComplex));
}
}
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_zsetmatrix_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_zherk(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_zgetmatrix_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_zgemm( MagmaNoTrans, MagmaConjTrans,
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_zpotrf(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_zsetmatrix_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_zsetmatrix_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(ZTRSM_WORK)
magmablas_ztrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
d_dinvA[d][0], d_x[d][0]);
#else
magma_ztrsm( MagmaRight, MagmaLower, MagmaConjTrans, 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_zgetmatrix_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(ZTRSM_WORK)
magmablas_ztrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, 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_ztrsm( MagmaRight, MagmaLower, MagmaConjTrans, 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(ZTRSM_WORK)
magmablas_ztrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
d_dinvA[d][1], d_x[d][1] );
#else
magma_ztrsm( MagmaRight, MagmaLower, MagmaConjTrans, 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( "zpotrf.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_zpotrf_mgpu */
extern "C" magma_int_t
magma_zgetrf_mgpu(magma_int_t num_gpus,
magma_int_t m, magma_int_t n,
cuDoubleComplex **d_lA, magma_int_t ldda,
magma_int_t *ipiv, magma_int_t *info)
{
/* -- MAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
November 2012
Purpose
=======
ZGETRF computes an LU factorization of a general M-by-N matrix A
using partial pivoting with row interchanges.
The factorization has the form
A = P * L * U
where P is a permutation matrix, L is lower triangular with unit
diagonal elements (lower trapezoidal if m > n), and U is upper
triangular (upper trapezoidal if m < n).
This is the right-looking Level 3 BLAS version of the algorithm.
Arguments
=========
NUM_GPUS
(input) INTEGER
The number of GPUS to be used for the factorization.
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
A (input/output) COMPLEX_16 array on the GPU, dimension (LDDA,N).
On entry, the M-by-N matrix to be factored.
On exit, the factors L and U from the factorization
A = P*L*U; the unit diagonal elements of L are not stored.
LDDA (input) INTEGER
The leading dimension of the array A. LDDA >= max(1,M).
IPIV (output) INTEGER array, dimension (min(M,N))
The pivot indices; for 1 <= i <= min(M,N), row i of the
matrix was interchanged with row IPIV(i).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
> 0: if INFO = i, U(i,i) is exactly zero. The factorization
has been completed, but the factor U is exactly
singular, and division by zero will occur if it is used
to solve a system of equations.
===================================================================== */
#define inAT(id,i,j) (d_lAT[(id)] + (i)*nb*lddat + (j)*nb)
cuDoubleComplex c_one = MAGMA_Z_ONE;
cuDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t iinfo, nb, n_local[MagmaMaxGPUs];
magma_int_t maxm, mindim;
magma_int_t i, j, d, rows, cols, s, lddat, lddwork;
magma_int_t id, i_local, i_local2, nb0, nb1;
cuDoubleComplex *d_lAT[MagmaMaxGPUs];
cuDoubleComplex *d_panel[MagmaMaxGPUs], *work;
cudaStream_t streaml[4][2];
/* Check arguments */
*info = 0;
if (m < 0)
*info = -2;
else if (n < 0)
*info = -3;
else if (ldda < max(1,m))
*info = -5;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if (m == 0 || n == 0)
return *info;
/* Function Body */
mindim = min(m, n);
nb = magma_get_zgetrf_nb(m);
if (nb <= 1 || nb >= n) {
/* Use CPU code. */
magma_zmalloc_cpu( &work, m * n );
if ( work == NULL ) {
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_zgetmatrix( m, n, d_lA[0], ldda, work, m );
lapackf77_zgetrf(&m, &n, work, &m, ipiv, info);
magma_zsetmatrix( m, n, work, m, d_lA[0], ldda );
magma_free_cpu(work);
} else {
/* Use hybrid blocked code. */
maxm = ((m + 31)/32)*32;
if( num_gpus > ceil((double)n/nb) ) {
printf( " * too many GPUs for the matrix size, using %d GPUs\n", (int) num_gpus );
*info = -1;
return *info;
}
/* allocate workspace for each GPU */
lddat = ((((((n+nb-1)/nb)/num_gpus)*nb)+31)/32)*32;
lddat = (n+nb-1)/nb; /* number of block columns */
lddat = (lddat+num_gpus-1)/num_gpus; /* number of block columns per GPU */
lddat = nb*lddat; /* number of columns per GPU */
lddat = ((lddat+31)/32)*32; /* make it a multiple of 32 */
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
/* local-n and local-ld */
n_local[i] = ((n/nb)/num_gpus)*nb;
if (i < (n/nb)%num_gpus)
n_local[i] += nb;
else if (i == (n/nb)%num_gpus)
n_local[i] += n%nb;
/* workspaces */
if (MAGMA_SUCCESS != magma_zmalloc( &d_panel[i], 3*nb*maxm )) {
for( j=0; j<=i; j++ ) {
magma_setdevice(j);
}
for( j=0; j<i; j++ ) {
magma_setdevice(j);
magma_free( d_panel[j] );
magma_free( d_lAT[j] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/* local-matrix storage */
if (MAGMA_SUCCESS != magma_zmalloc( &d_lAT[i], lddat*maxm )) {
for( j=0; j<=i; j++ ) {
magma_setdevice(j);
magma_free( d_panel[j] );
}
for( j=0; j<i; j++ ) {
magma_setdevice(j);
magma_free( d_lAT[j] );
}
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/* create the streams */
magma_queue_create( &streaml[i][0] );
magma_queue_create( &streaml[i][1] );
magmablasSetKernelStream(streaml[i][1]);
magmablas_ztranspose2( d_lAT[i], lddat, d_lA[i], ldda, m, n_local[i] );
}
for(i=0; i<num_gpus; i++){
magma_setdevice(i);
cudaStreamSynchronize(streaml[i][0]);
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
/* cpu workspace */
lddwork = maxm;
if (MAGMA_SUCCESS != magma_zmalloc_pinned( &work, lddwork*nb*num_gpus )) {
for(i=0; i<num_gpus; i++ ) {
magma_setdevice(i);
magma_free( d_panel[i] );
magma_free( d_lAT[i] );
}
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
/* calling multi-gpu interface with allocated workspaces and streams */
//magma_zgetrf1_mgpu( num_gpus, m, n, nb, 0, d_lAT, lddat, ipiv, d_panel, work, maxm,
// (cudaStream_t **)streaml, info );
magma_zgetrf2_mgpu(num_gpus, m, n, nb, 0, d_lAT, lddat, ipiv, d_panel, work, maxm,
streaml, info);
/* clean up */
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
/* save on output */
magmablas_ztranspose2( d_lA[d], ldda, d_lAT[d], lddat, n_local[d], m );
magma_device_sync();
magma_free( d_lAT[d] );
magma_free( d_panel[d] );
magma_queue_destroy( streaml[d][0] );
magma_queue_destroy( streaml[d][1] );
magmablasSetKernelStream(NULL);
} /* end of for d=1,..,num_gpus */
magma_setdevice(0);
magma_free_pinned( work );
}
return *info;
/* End of MAGMA_ZGETRF_MGPU */
}
/**
Purpose
-------
Solves the overdetermined, least squares problem
min || A*X - C ||
using the QR factorization A.
The underdetermined problem (m < n) is not currently handled.
Arguments
---------
@param[in]
trans magma_trans_t
- = MagmaNoTrans: the linear system involves A.
Only TRANS=MagmaNoTrans is currently handled.
@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. M >= N >= 0.
@param[in]
nrhs INTEGER
The number of columns of the matrix C. NRHS >= 0.
@param[in,out]
dA COMPLEX_16 array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, A is overwritten by details of its QR
factorization as returned by ZGEQRF3.
@param[in]
ldda INTEGER
The leading dimension of the array A, LDDA >= M.
@param[in,out]
dB COMPLEX_16 array on the GPU, dimension (LDDB,NRHS)
On entry, the M-by-NRHS matrix C.
On exit, the N-by-NRHS solution matrix X.
@param[in]
lddb INTEGER
The leading dimension of the array dB. LDDB >= M.
@param[out]
hwork (workspace) COMPLEX_16 array, dimension MAX(1,LWORK).
On exit, if INFO = 0, HWORK[0] returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array HWORK,
LWORK >= (M - N + NB)*(NRHS + NB) + NRHS*NB,
where NB is the blocksize given by magma_get_zgeqrf_nb( M ).
\n
If LWORK = -1, then a workspace query is assumed; the routine
only calculates the optimal size of the HWORK array, returns
this value as the first entry of the HWORK array.
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value
@ingroup magma_zgels_driver
********************************************************************/
extern "C" magma_int_t
magma_zgels3_gpu( magma_trans_t trans, magma_int_t m, magma_int_t n, magma_int_t nrhs,
magmaDoubleComplex *dA, magma_int_t ldda,
magmaDoubleComplex *dB, magma_int_t lddb,
magmaDoubleComplex *hwork, magma_int_t lwork,
magma_int_t *info)
{
magmaDoubleComplex *dT, *tau;
magma_int_t k;
magma_int_t nb = magma_get_zgeqrf_nb(m);
magma_int_t lwkopt = (m - n + nb)*(nrhs + nb) + nrhs*nb;
int lquery = (lwork == -1);
hwork[0] = MAGMA_Z_MAKE( (double)lwkopt, 0. );
*info = 0;
/* For now, N is the only case working */
if ( trans != MagmaNoTrans )
*info = -1;
else if (m < 0)
*info = -2;
else if (n < 0 || m < n) /* LQ is not handle for now*/
*info = -3;
else if (nrhs < 0)
*info = -4;
else if (ldda < max(1,m))
*info = -6;
else if (lddb < max(1,m))
*info = -8;
else if (lwork < lwkopt && ! lquery)
*info = -10;
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery)
return *info;
k = min(m,n);
if (k == 0) {
hwork[0] = MAGMA_Z_ONE;
return *info;
}
/*
* Allocate temporary buffers
*/
int ldtwork = ( 2*k + ((n+31)/32)*32 )*nb;
if (nb < nrhs)
ldtwork = ( 2*k + ((n+31)/32)*32 )*nrhs;
if (MAGMA_SUCCESS != magma_zmalloc( &dT, ldtwork )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
magma_zmalloc_cpu( &tau, k );
if ( tau == NULL ) {
magma_free( dT );
*info = MAGMA_ERR_HOST_ALLOC;
return *info;
}
magma_zgeqrf3_gpu( m, n, dA, ldda, tau, dT, info );
if ( *info == 0 ) {
magma_zgeqrs3_gpu( m, n, nrhs,
dA, ldda, tau, dT,
dB, lddb, hwork, lwork, info );
}
magma_free( dT );
magma_free_cpu(tau);
return *info;
}
extern "C" magma_int_t
magma_zgeqrf2_2q_gpu(
magma_int_t m, magma_int_t n,
magmaDoubleComplex_ptr dA, size_t dA_offset, magma_int_t ldda,
magmaDoubleComplex *tau,
magma_queue_t* queues,
magma_int_t *info)
{
/* -- clMAGMA (version 1.3.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date November 2014
Purpose
=======
ZGEQRF computes a QR factorization of a complex M-by-N matrix A:
A = Q * R.
Arguments
=========
M (input) INTEGER
The number of rows of the matrix A. M >= 0.
N (input) INTEGER
The number of columns of the matrix A. N >= 0.
dA (input/output) COMPLEX_16 array on the GPU, dimension (LDDA,N)
On entry, the M-by-N matrix dA.
On exit, the elements on and above the diagonal of the array
contain the min(M,N)-by-N upper trapezoidal matrix R (R is
upper triangular if m >= n); the elements below the diagonal,
with the array TAU, represent the orthogonal matrix Q as a
product of min(m,n) elementary reflectors (see Further
Details).
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,M).
To benefit from coalescent memory accesses LDDA must be
divisible by 16.
TAU (output) COMPLEX_16 array, dimension (min(M,N))
The scalar factors of the elementary reflectors (see Further
Details).
INFO (output) INTEGER
= 0: successful exit
< 0: if INFO = -i, the i-th argument had an illegal value
or another error occured, such as memory allocation failed.
Further Details
===============
The matrix Q is represented as a product of elementary reflectors
Q = H(1) H(2) . . . H(k), where k = min(m,n).
Each H(i) has the form
H(i) = I - tau * v * v'
where tau is a complex scalar, and v is a complex vector with
v(1:i-1) = 0 and v(i) = 1; v(i+1:m) is stored on exit in A(i+1:m,i),
and tau in TAU(i).
===================================================================== */
#define dA(a_1,a_2) dA, (dA_offset + (a_1) + (a_2)*(ldda))
#define work_ref(a_1) ( work + (a_1))
#define hwork ( work + (nb)*(m))
magmaDoubleComplex_ptr dwork;
magmaDoubleComplex *work;
magma_int_t i, k, ldwork, lddwork, old_i, old_ib, rows;
magma_int_t nbmin, nx, ib, nb;
magma_int_t lhwork, lwork;
*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;
}
k = min(m,n);
if (k == 0)
return MAGMA_SUCCESS;
nb = magma_get_zgeqrf_nb(m);
lwork = (m+n) * nb;
lhwork = lwork - (m)*nb;
if ( MAGMA_SUCCESS != magma_zmalloc( &dwork, n*nb )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
/*
if ( MAGMA_SUCCESS != magma_zmalloc_cpu( &work, lwork ) ) {
*info = MAGMA_ERR_HOST_ALLOC;
magma_free( dwork );
return *info;
}
*/
cl_mem buffer = clCreateBuffer(gContext, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, sizeof(magmaDoubleComplex)*lwork, NULL, NULL);
work = (magmaDoubleComplex*)clEnqueueMapBuffer(queues[0], buffer, CL_TRUE, CL_MAP_READ | CL_MAP_WRITE, 0, lwork*sizeof(magmaDoubleComplex), 0, NULL, NULL, NULL);
nbmin = 2;
nx = 2*nb;
ldwork = m;
lddwork= n;
if (nb >= nbmin && nb < k && nx < k) {
/* Use blocked code initially */
old_i = 0; old_ib = nb;
for (i = 0; i < k-nx; i += nb) {
ib = min(k-i, nb);
rows = m -i;
magma_zgetmatrix_async(rows, ib, dA(i, i), ldda, work_ref(i), ldwork, queues[0], NULL);
clFlush(queues[0]);
if (i>0){
/* Apply H' to A(i:m,i+2*ib:n) from the left */
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-old_i, n-old_i-2*old_ib, old_ib,
dA(old_i, old_i ), ldda, dwork,0, lddwork,
dA(old_i, old_i+2*old_ib), ldda, dwork,old_ib, lddwork, queues[1]);
magma_zsetmatrix_async( old_ib, old_ib, work_ref(old_i), ldwork,
dA(old_i, old_i), ldda, queues[1], NULL);
clFlush(queues[1]);
}
magma_queue_sync(queues[0]);
lapackf77_zgeqrf(&rows, &ib, work_ref(i), &ldwork, tau+i, hwork, &lhwork, info);
/* Form the triangular factor of the block reflector
H = H(i) H(i+1) . . . H(i+ib-1) */
lapackf77_zlarft( MagmaForwardStr, MagmaColumnwiseStr,
&rows, &ib,
work_ref(i), &ldwork, tau+i, hwork, &ib);
zpanel_to_q( MagmaUpper, ib, work_ref(i), ldwork, hwork+ib*ib );
magma_zsetmatrix( rows, ib, work_ref(i), ldwork, dA(i,i), ldda, queues[0]);
zq_to_panel( MagmaUpper, ib, work_ref(i), ldwork, hwork+ib*ib );
if (i + ib < n)
{
magma_zsetmatrix( ib, ib, hwork, ib, dwork, 0, lddwork, queues[1]);
if (i+nb < k-nx){
/* Apply H' to A(i:m,i+ib:i+2*ib) from the left */
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, ib, ib,
dA(i, i ), ldda, dwork,0, lddwork,
dA(i, i+ib), ldda, dwork,ib, lddwork, queues[1]);
magma_queue_sync(queues[1]);
}else {
magma_zlarfb_gpu( MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
rows, n-i-ib, ib,
dA(i, i ), ldda, dwork,0, lddwork,
dA(i, i+ib), ldda, dwork,ib, lddwork, queues[1]);
magma_zsetmatrix( ib, ib, work_ref(i), ldwork, dA(i,i), ldda, queues[1]);
clFlush(queues[1]);
}
old_i = i;
old_ib = ib;
}
}
} else {
i = 0;
}
magma_free(dwork);
/* Use unblocked code to factor the last or only block. */
if (i < k) {
ib = n-i;
rows = m-i;
magma_zgetmatrix( rows, ib, dA(i, i), ldda, work, rows, queues[0]);
lhwork = lwork - rows*ib;
lapackf77_zgeqrf(&rows, &ib, work, &rows, tau+i, work+ib*rows, &lhwork, info);
magma_zsetmatrix( rows, ib, work, rows, dA(i, i), ldda, queues[0]);
}
clEnqueueUnmapMemObject(queues[0], buffer, work, 0, NULL, NULL);
clReleaseMemObject(buffer);
// magma_free_cpu(work);
return *info;
} /* magma_zgeqrf2_gpu */
extern "C" magma_int_t
magma_zpotrf3_mgpu(magma_int_t num_gpus, char uplo, magma_int_t m, magma_int_t n,
magma_int_t off_i, magma_int_t off_j, magma_int_t nb,
magmaDoubleComplex *d_lA[], magma_int_t ldda,
magmaDoubleComplex *d_lP[], magma_int_t lddp,
magmaDoubleComplex *a, magma_int_t lda, magma_int_t h,
magma_queue_t stream[][3], magma_event_t event[][5],
magma_int_t *info )
{
/* -- MAGMA (version 1.4.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
August 2013
Purpose
=======
ZPOTRF computes the Cholesky factorization of a complex Hermitian
positive definite matrix dA.
Auxiliary subroutine for zpotrf2_ooc. It is multiple gpu interface to compute
Cholesky of a "rectangular" matrix.
The factorization has the form
dA = U**H * U, if UPLO = 'U', or
dA = L * L**H, if UPLO = 'L',
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
=========
UPLO (input) CHARACTER*1
= 'U': Upper triangle of dA is stored;
= 'L': Lower triangle of dA is stored.
N (input) INTEGER
The order of the matrix dA. N >= 0.
dA (input/output) COMPLEX_16 array on the GPU, dimension (LDDA,N)
On entry, the Hermitian matrix dA. If UPLO = 'U', 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 = 'L', 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.
On exit, if INFO = 0, the factor U or L from the Cholesky
factorization dA = U**H * U or dA = L * L**H.
LDDA (input) INTEGER
The leading dimension of the array dA. LDDA >= max(1,N).
To benefit from coalescent memory accesses LDDA must be
dividable by 16.
INFO (output) 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.
===================================================================== */
magma_int_t j, jb, nb0, nb2, d, dd, id, j_local, j_local2, buf;
char uplo_[2] = {uplo, 0};
magmaDoubleComplex c_one = MAGMA_Z_ONE;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
double d_one = 1.0;
double d_neg_one = -1.0;
int upper = lapackf77_lsame(uplo_, "U");
magmaDoubleComplex *dlpanel;
magma_int_t n_local[MagmaMaxGPUs], ldpanel;
const magma_int_t stream1 = 0, stream2 = 1, stream3 = 2;
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
/* used by ztrsm_work */
int trsm_nb = 128;
int trsm_n = trsm_nb*((nb+trsm_nb-1)/trsm_nb);
magmaDoubleComplex *d_dinvA[MagmaMaxGPUs];
magmaDoubleComplex *d_x[MagmaMaxGPUs];
#define dinvA(d,j) &(d_dinvA[(d)][(j)*trsm_nb*trsm_n])
#define dx(d,j) &(d_x[(d)][(j)*nb*m])
/*
* Allocate device memory for the inversed diagonal blocks, size=N*BLOCK_SIZE
*/
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
if ( (MAGMA_SUCCESS != magma_zmalloc( &d_dinvA[d], 2*trsm_nb*trsm_n )) ||
(MAGMA_SUCCESS != magma_zmalloc( &d_x[d], 2*nb*(upper ? n : m) )) ) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
}
magma_setdevice(0);
#endif
*info = 0;
if ( (! upper) && (! lapackf77_lsame(uplo_, "L")) ) {
*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;
}
/* initialization */
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;
}
}
/* == initialize the trace */
trace_init( 1, num_gpus, 3, (CUstream_st**)stream );
if (upper)
{
/* ---------------------------------------------- */
/* Upper-triangular case */
/* > Compute the Cholesky factorization A = U'*U. */
/* ---------------------------------------------- */
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; // right now, we have num_gpu buffers, so id and buf are the same..
/* Set the local index where the current panel is */
j_local = j/(nb*num_gpus);
jb = min(nb, (m-j));
/* Update the current diagonal block on stream1 */
magma_setdevice(id);
if( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
trace_gpu_start( id, stream1, "syrk", "syrk" );
magma_zherk(MagmaUpper, MagmaConjTrans, 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 );
}
/* send the diagonal to cpu on stream1 */
trace_gpu_start( id, stream1, "comm", "D to CPU" );
magma_zgetmatrix_async( jb, jb,
dlA(id, j, nb*j_local), ldda,
Aup(j,j), lda,
stream[id][stream1] );
trace_gpu_end( id, stream1 );
/* update off-diagonal blocks in the panel */
if( j > 0 ) {
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; // number of local columns in the panel, while jb is panel-size (number of rows)
if( n_local[d] > nb0 ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
if( d == id ) {
dlpanel = dlA(d,0,nb*j_local);
ldpanel = ldda;
// the GPU owns the row from start, and no need of synch.
//magma_queue_wait_event( stream[d][stream2], event[d][0] ); // rows arrived at gpu
} else {
dlpanel = dlP(d,nb,0,buf);
ldpanel = lddp;
magma_queue_wait_event( stream[d][stream2], event[d][0] ); // rows arrived at gpu
}
trace_gpu_start( d, stream2, "gemm", "gemm" );
magma_zgemm(MagmaConjTrans, 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, stream2 );
magma_event_record( event[d][2], stream[d][stream2] );
}
d = (d+1)%num_gpus;
}
}
/* wait for panel and factorize it on cpu */
magma_setdevice(id);
magma_queue_sync( stream[id][stream1] );
trace_cpu_start( 0, "getrf", "getrf" );
lapackf77_zpotrf(MagmaUpperStr, &jb, Aup(j,j), &lda, info);
trace_cpu_end( 0 );
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus on stream1 */
if ( (j+jb) < n) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
if( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d,0,0,buf);
ldpanel = lddp;
}
magma_setdevice(d);
trace_gpu_start( d, stream1, "comm", "comm" );
magma_zsetmatrix_async( jb, jb,
Aup(j,j), lda,
dlpanel, ldpanel,
stream[d][stream1] );
trace_gpu_end( d, stream1 );
magma_event_record( event[d][1], stream[d][stream1] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
trace_gpu_start( id, stream1, "comm", "comm" );
magma_zsetmatrix_async( jb, jb,
Aup(j,j), lda,
dlA(id, j, nb*j_local), ldda,
stream[id][stream1] );
trace_gpu_end( id, stream1 );
}
/* 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] - j_local2*nb;
magma_setdevice(d);
if( j+jb < m && d == (j/nb+1)%num_gpus ) {
/* owns the next column, look-ahead next block on stream1 */
nb0 = min(nb, nb2);
magmablasSetKernelStream(stream[d][stream1]);
magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for gemm update
trace_gpu_start( d, stream1, "trsm", "trsm" );
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
magmablas_zlaset( MagmaUpperLower, trsm_nb, trsm_n, dinvA(d,0),trsm_nb );
magmablas_zlaset( MagmaUpperLower, nb0,jb, dx(d,0),nb0 );
magmablas_ztrsm_work( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ztrsm( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb0, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
magma_event_record( event[d][4], stream[d][stream1] );
trace_gpu_end( d, stream1 );
} else if( nb2 > 0 ) {
/* update all the blocks on stream2 */
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // wait for cholesky factor
trace_gpu_start( d, stream2, "trsm", "trsm" );
magmablasSetKernelStream(stream[d][stream2]);
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
magmablas_zlaset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,0),trsm_nb );
magmablas_zlaset( MagmaUpperLower, nb2,jb, dx(d,0),nb2 );
magmablas_ztrsm_work( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ztrsm( MagmaLeft, MagmaUpper,
MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2), ldda);
#endif
trace_gpu_end( d, stream2 );
}
d = (d+1)%num_gpus;
} /* end of for */
/* ========================================================== */
if( j+jb < m ) {
d = (j/nb+1)%num_gpus;
/* next column */
j_local2 = j_local+1;
if( d > id ) j_local2--;
nb0 = min(nb, n_local[d]-nb*j_local2 );
/* even on 1 gpu, off-diagonals are copied to cpu (synchronize at the end). *
* so we have the Cholesky factor, but only diagonal submatrix of the big panel, *
* on cpu at the end. */
int d2, buf2;
magma_setdevice(d);
/* lookahead done */
magma_queue_wait_event( stream[d][stream3], event[d][4] );
trace_gpu_start( d, stream3, "comm", "row to CPU" );
magma_zgetmatrix_async( (j+jb), nb0,
dlA(d, 0, nb*j_local2), ldda,
Aup(0,j+jb), lda,
stream[d][stream3] );
trace_gpu_end( d, stream3 );
magma_event_record( event[d][3], stream[d][stream3] );
/* needed on pluto */
//magma_queue_sync( stream[d][stream3] );
/* broadcast rows to gpus on stream2 */
buf2 = ((j+jb)/nb)%num_gpus;
for( d2=0; d2<num_gpus; d2++ ) {
if( d2 != d )
{
magma_setdevice(d2);
trace_gpu_start( d2, stream3, "comm", "row to GPUs" );
magma_queue_wait_event( stream[d2][stream3], event[d][3] ); // rows arrived at cpu on stream3
magma_zsetmatrix_async( j+jb, nb0,
Aup(0,j+jb), lda,
dlP(d2,nb,0,buf2), lddp,
stream[d2][stream3] );
trace_gpu_end( d2, stream3 );
magma_event_record( event[d2][0], stream[d2][stream3] );
}
}
/* =========================== */
/* update the remaining blocks */
nb2 = n_local[d]-(nb*j_local2 + nb0);
if( nb2 > 0 ) {
if( d == id ) {
dlpanel = dlA(d, j, nb*j_local);
ldpanel = ldda;
} else {
dlpanel = dlP(d,0,0,buf);
ldpanel = lddp;
}
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
trace_gpu_start( d, stream2, "trsm", "trsm" );
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
int flag = 0;
if (flag == 0) {
magma_queue_wait_event( stream[d][stream2], event[d][4] ); // lookahead -> diagonal inversion
} else {
magmablas_zlaset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,flag),trsm_nb );
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // panel received
}
magmablas_zlaset( MagmaUpperLower, nb2,jb, dx(d,1),nb2 );
magmablas_ztrsm_work( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda,
flag, dinvA(d,flag), dx(d,1) );
#else
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // wait for cholesky factor
magma_ztrsm( MagmaLeft, MagmaUpper, MagmaConjTrans, MagmaNonUnit,
jb, nb2, c_one,
dlpanel, ldpanel,
dlA(d, j, nb*j_local2+nb0), ldda);
#endif
trace_gpu_end( d, stream2 );
}
}
} /* end of ztrsm */
} /* end of for j=1, .., n */
} else {
/* ---------------------------------------------- */
/* Lower-triangular case */
/* > Compute the Cholesky factorization A = L*L'. */
/* ---------------------------------------------- */
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));
/* Update the current diagonal block on stream1 */
magma_setdevice(id);
if( j > 0 ) {
magmablasSetKernelStream(stream[id][stream1]);
magma_zherk(MagmaLower, MagmaNoTrans, jb, j,
d_neg_one, dlA(id, nb*j_local, 0), ldda,
d_one, dlA(id, nb*j_local, j), ldda);
}
/* send the diagonal to cpu on stream1 */
magma_zgetmatrix_async( jb, jb,
dlA(id, nb*j_local, j), ldda,
Alo(j,j), lda,
stream[id][stream1] );
/* update off-diagonal blocks of the panel */
if( j > 0 ) {
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] ) {
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
if( d == id ) {
dlpanel = dlA(d, nb*j_local, 0);
ldpanel = ldda;
} else {
dlpanel = dlPT(d,0,nb,buf);
ldpanel = nb;
magma_queue_wait_event( stream[d][stream2], event[d][0] ); // rows arrived at gpu
}
magma_zgemm( MagmaNoTrans, MagmaConjTrans,
n_local[d]-nb0, jb, j,
c_neg_one, dlA(d, nb0, 0), ldda,
dlpanel, ldpanel,
c_one, dlA(d, nb0, j), ldda);
magma_event_record( event[d][2], stream[d][stream2] );
}
d = (d+1)%num_gpus;
}
}
/* wait for the panel and factorized it on cpu */
magma_setdevice(id);
magma_queue_sync( stream[id][stream1] );
lapackf77_zpotrf(MagmaLowerStr, &jb, Alo(j,j), &lda, info);
if (*info != 0) {
*info = *info + j;
break;
}
/* send the diagonal to gpus on stream1 */
if ( (j+jb) < m) {
d = (j/nb+1)%num_gpus;
for( dd=0; dd<num_gpus; dd++ ) {
if( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d, 0, 0, buf);
ldpanel = nb;
}
magma_setdevice(d);
magma_zsetmatrix_async( jb, jb,
Alo(j,j), lda,
dlpanel, ldpanel,
stream[d][stream1] );
magma_event_record( event[d][1], stream[d][stream1] );
d = (d+1)%num_gpus;
}
} else {
magma_setdevice(id);
magma_zsetmatrix_async( jb, jb,
Alo(j,j), lda,
dlA(id, nb*j_local, j), ldda,
stream[id][stream1] );
}
/* panel factorize the off-diagonal */
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);
if( j+nb < n && d == (j/nb+1)%num_gpus ) { /* owns next column, look-ahead next block on stream1 */
if ( j > 0 ) magma_queue_wait_event( stream[d][stream1], event[d][2] ); // wait for gemm update
magmablasSetKernelStream(stream[d][stream1]);
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
magmablas_zlaset( MagmaUpperLower, trsm_nb, trsm_n, dinvA(d,0),trsm_nb );
magmablas_zlaset( MagmaUpperLower, nb0,jb, dx(d,0),nb0 );
magmablas_ztrsm_work( MagmaRight, MagmaLower,
MagmaConjTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ztrsm( MagmaRight, MagmaLower,
MagmaConjTrans, MagmaNonUnit,
nb0, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
magma_event_record( event[d][4], stream[d][stream1] );
} else if( nb2 > 0 ) { /* other gpus updating all the blocks on stream2 */
/* update the entire column */
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // wait for the cholesky factor
magmablasSetKernelStream(stream[d][stream2]);
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
magmablas_zlaset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,0),trsm_nb );
magmablas_zlaset( MagmaUpperLower, nb2,jb, dx(d,0),nb2 );
magmablas_ztrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda,
1, dinvA(d,0), dx(d,0) );
#else
magma_ztrsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2, j), ldda);
#endif
}
d = (d+1)%num_gpus;
} /* end for d */
/* ========================================================== */
if( j+jb < n ) {
d = (j/nb+1)%num_gpus;
/* next column */
j_local2 = j_local+1;
if( d > id ) j_local2--;
nb0 = min(nb, n_local[d]-nb*j_local2 );
/* even on 1 gpu, we copy off-diagonal to cpu (but don't synchronize). */
/* so we have the Cholesky factor on cpu at the end. */
int d2, buf2;
//#define ZPOTRF_DEVICE_TO_DEVICE
#ifdef ZPOTRF_DEVICE_TO_DEVICE
// lookahead done
/* broadcast the rows to gpus */
buf2 = ((j+jb)/nb)%num_gpus;
for( d2=0; d2<num_gpus; d2++ ) {
magma_setdevice(d2);
magma_queue_wait_event( stream[d2][stream3], event[d][4] );
if( d2 != d ) {
magma_zcopymatrix_async( nb0, j+jb,
dlPT(d2,0,nb,buf2), nb, // first nbxnb reserved for diagonal block
dlA(d, nb*j_local2, 0), ldda,
stream[d2][stream3] );
magma_event_record( event[d2][0], stream[d2][stream3] );
} else {
magma_zgetmatrix_async( nb0, j+jb,
dlA(d, nb*j_local2, 0), ldda,
Alo(j+jb,0), lda,
stream[d][stream3] );
}
}
#else
// lookahead done
magma_setdevice(d);
magma_queue_wait_event( stream[d][stream3], event[d][4] );
magma_zgetmatrix_async( nb0, j+jb,
dlA(d, nb*j_local2, 0), ldda,
Alo(j+jb,0), lda,
stream[d][stream3] );
magma_event_record( event[d][3], stream[d][stream3] );
/* syn on rows on CPU, seem to be needed on Pluto */
//magma_queue_sync( stream[d][stream3] );
/* broadcast the rows to gpus */
buf2 = ((j+jb)/nb)%num_gpus;
for( d2=0; d2<num_gpus; d2++ ) {
if( d2 != d )
{
magma_setdevice(d2);
magma_queue_wait_event( stream[d2][stream3], event[d][3] ); // getmatrix done
magma_zsetmatrix_async( nb0, j+jb,
Alo(j+jb,0), lda,
dlPT(d2,0,nb,buf2), nb, // first nbxnb reserved for diagonal block
stream[d2][stream3] );
magma_event_record( event[d2][0], stream[d2][stream3] );
}
}
#endif
/* =================================== */
/* updates remaining blocks on stream2 */
nb2 = n_local[d] - (j_local2*nb + nb0);
if( nb2 > 0 ) {
if( d == id ) {
dlpanel = dlA(d, nb*j_local, j);
ldpanel = ldda;
} else {
dlpanel = dlPT(d,0,0,buf);
ldpanel = nb;
}
magma_setdevice(d);
magmablasSetKernelStream(stream[d][stream2]);
/* update the remaining blocks in the column */
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
int flag = 0;
if (flag == 0) {
magma_queue_wait_event( stream[d][stream2], event[d][4] ); // lookahead -> diagonal inversion
} else {
magmablas_zlaset( MagmaUpperLower, trsm_nb,trsm_n, dinvA(d,flag),trsm_nb );
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // panel received
}
magmablas_zlaset( MagmaUpperLower, nb2,jb, dx(d,1),nb2 );
magmablas_ztrsm_work( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda,
flag, dinvA(d,flag), dx(d,1) );
#else
magma_queue_wait_event( stream[d][stream2], event[d][1] ); // panel received
magma_ztrsm( MagmaRight, MagmaLower, MagmaConjTrans, MagmaNonUnit,
nb2, jb, c_one,
dlpanel, ldpanel,
dlA(d, nb*j_local2+nb0, j), ldda);
#endif
}
}
}
}
} /* end of else not upper */
/* == finalize the trace == */
trace_finalize( "zpotrf.svg","trace.css" );
for( d=0; d<num_gpus; d++ ) {
magma_setdevice(d);
for( j=0; j<3; j++ ) {
magma_queue_sync( stream[d][j] );
}
#if (defined(PRECISION_d) || defined(PRECISION_s)) && defined(ZTRSM_WORK)
magma_free( d_dinvA[d] );
magma_free( d_x[d] );
#endif
magmablasSetKernelStream(NULL);
}
magma_setdevice(0);
return *info;
} /* magma_zpotrf_mgpu */
extern "C" magma_int_t
magma_zpcg_merge(
magma_z_matrix A, magma_z_matrix b, magma_z_matrix *x,
magma_z_solver_par *solver_par,
magma_z_preconditioner *precond_par,
magma_queue_t queue )
{
magma_int_t info = MAGMA_NOTCONVERGED;
// prepare solver feedback
solver_par->solver = Magma_PCGMERGE;
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// solver variables
magmaDoubleComplex alpha, beta, gamma, rho, tmp1, *skp_h={0};
double nom, nom0, r0, res, nomb;
magmaDoubleComplex den;
// some useful variables
magmaDoubleComplex c_zero = MAGMA_Z_ZERO, c_one = MAGMA_Z_ONE;
magma_int_t dofs = A.num_rows*b.num_cols;
magma_z_matrix r={Magma_CSR}, d={Magma_CSR}, z={Magma_CSR}, h={Magma_CSR},
rt={Magma_CSR};
magmaDoubleComplex *d1=NULL, *d2=NULL, *skp=NULL;
// GPU workspace
CHECK( magma_zvinit( &r, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &d, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &z, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &rt, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zvinit( &h, Magma_DEV, A.num_rows, b.num_cols, c_zero, queue ));
CHECK( magma_zmalloc( &d1, dofs*(2) ));
CHECK( magma_zmalloc( &d2, dofs*(2) ));
// array for the parameters
CHECK( magma_zmalloc( &skp, 7 ));
// skp = [alpha|beta|gamma|rho|tmp1|tmp2|res]
// solver setup
CHECK( magma_zresidualvec( A, b, *x, &r, &nom0, queue));
// preconditioner
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, r, &rt, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, rt, &h, precond_par, queue ));
magma_zcopy( dofs, h.dval, 1, d.dval, 1, queue );
nom = MAGMA_Z_ABS( magma_zdotc( dofs, r.dval, 1, h.dval, 1, queue ));
CHECK( magma_z_spmv( c_one, A, d, c_zero, z, queue )); // z = A d
den = magma_zdotc( dofs, d.dval, 1, z.dval, 1, queue ); // den = d'* z
solver_par->init_res = nom0;
nomb = magma_dznrm2( dofs, b.dval, 1, queue );
if ( nomb == 0.0 ){
nomb=1.0;
}
if ( (r0 = nomb * solver_par->rtol) < ATOLERANCE ){
r0 = ATOLERANCE;
}
solver_par->final_res = solver_par->init_res;
solver_par->iter_res = solver_par->init_res;
if ( solver_par->verbose > 0 ) {
solver_par->res_vec[0] = (real_Double_t)nom0;
solver_par->timing[0] = 0.0;
}
if ( nom < r0 ) {
info = MAGMA_SUCCESS;
goto cleanup;
}
// check positive definite
if ( MAGMA_Z_ABS(den) <= 0.0 ) {
info = MAGMA_NONSPD;
goto cleanup;
}
// array on host for the parameters
CHECK( magma_zmalloc_cpu( &skp_h, 7 ));
alpha = rho = gamma = tmp1 = c_one;
beta = magma_zdotc( dofs, h.dval, 1, r.dval, 1, queue );
skp_h[0]=alpha;
skp_h[1]=beta;
skp_h[2]=gamma;
skp_h[3]=rho;
skp_h[4]=tmp1;
skp_h[5]=MAGMA_Z_MAKE(nom, 0.0);
skp_h[6]=MAGMA_Z_MAKE(nom, 0.0);
magma_zsetvector( 7, skp_h, 1, skp, 1, queue );
//Chronometry
real_Double_t tempo1, tempo2, tempop1, tempop2;
tempo1 = magma_sync_wtime( queue );
solver_par->numiter = 0;
solver_par->spmv_count = 0;
// start iteration
do
{
solver_par->numiter++;
// computes SpMV and dot product
CHECK( magma_zcgmerge_spmv1( A, d1, d2, d.dval, z.dval, skp, queue ));
solver_par->spmv_count++;
if( precond_par->solver == Magma_JACOBI ){
CHECK( magma_zjcgmerge_xrbeta( dofs, d1, d2, precond_par->d.dval, x->dval, r.dval, d.dval, z.dval, h.dval, skp, queue ));
}
else if( precond_par->solver == Magma_NONE ){
// updates x, r
CHECK( magma_zpcgmerge_xrbeta1( dofs, x->dval, r.dval, d.dval, z.dval, skp, queue ));
// computes scalars and updates d
CHECK( magma_zpcgmerge_xrbeta2( dofs, d1, d2, r.dval, r.dval, d.dval, skp, queue ));
} else {
// updates x, r
CHECK( magma_zpcgmerge_xrbeta1( dofs, x->dval, r.dval, d.dval, z.dval, skp, queue ));
// preconditioner in between
tempop1 = magma_sync_wtime( queue );
CHECK( magma_z_applyprecond_left( MagmaNoTrans, A, r, &rt, precond_par, queue ));
CHECK( magma_z_applyprecond_right( MagmaNoTrans, A, rt, &h, precond_par, queue ));
// magma_zcopy( dofs, r.dval, 1, h.dval, 1 );
tempop2 = magma_sync_wtime( queue );
precond_par->runtime += tempop2-tempop1;
// computes scalars and updates d
CHECK( magma_zpcgmerge_xrbeta2( dofs, d1, d2, h.dval, r.dval, d.dval, skp, queue ));
}
//if( solver_par->numiter==1){
// magma_zcopy( dofs, h.dval, 1, d.dval, 1 );
//}
// updates x, r, computes scalars and updates d
//CHECK( magma_zcgmerge_xrbeta( dofs, d1, d2, x->dval, r.dval, d.dval, z.dval, skp, queue ));
// check stopping criterion (asynchronous copy)
magma_zgetvector( 1 , skp+6, 1, skp_h+6, 1, queue );
res = sqrt(MAGMA_Z_ABS(skp_h[6]));
if ( solver_par->verbose > 0 ) {
tempo2 = magma_sync_wtime( queue );
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
if ( res/nomb <= solver_par->rtol || res <= solver_par->atol ){
break;
}
}
while ( solver_par->numiter+1 <= solver_par->maxiter );
tempo2 = magma_sync_wtime( queue );
solver_par->runtime = (real_Double_t) tempo2-tempo1;
double residual;
CHECK( magma_zresidualvec( A, b, *x, &r, &residual, queue));
solver_par->iter_res = res;
solver_par->final_res = residual;
if ( solver_par->numiter < solver_par->maxiter ) {
info = MAGMA_SUCCESS;
} else if ( solver_par->init_res > solver_par->final_res ) {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
info = MAGMA_SLOW_CONVERGENCE;
if( solver_par->iter_res < solver_par->atol ||
solver_par->iter_res/solver_par->init_res < solver_par->rtol ){
info = MAGMA_SUCCESS;
}
}
else {
if ( solver_par->verbose > 0 ) {
if ( (solver_par->numiter)%solver_par->verbose==0 ) {
solver_par->res_vec[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) res;
solver_par->timing[(solver_par->numiter)/solver_par->verbose]
= (real_Double_t) tempo2-tempo1;
}
}
solver_par->info = MAGMA_DIVERGENCE;
}
cleanup:
magma_zmfree(&r, queue );
magma_zmfree(&z, queue );
magma_zmfree(&d, queue );
magma_zmfree(&rt, queue );
magma_zmfree(&h, queue );
magma_free( d1 );
magma_free( d2 );
magma_free( skp );
magma_free_cpu( skp_h );
solver_par->info = info;
return info;
} /* magma_zpcg_merge */
extern "C" magma_int_t
magma_zgeqrf_batched(
magma_int_t m, magma_int_t n,
magmaDoubleComplex **dA_array,
magma_int_t ldda,
magmaDoubleComplex **tau_array,
magma_int_t *info_array, magma_int_t batchCount, magma_queue_t queue)
{
#define dA(i, j) (dA + (i) + (j)*ldda) // A(i, j) means at i row, j column
magma_int_t min_mn = min(m, n);
cudaMemset(info_array, 0, batchCount*sizeof(magma_int_t));
/* Check arguments */
magma_int_t arginfo = 0;
if (m < 0)
arginfo = -1;
else if (n < 0)
arginfo = -2;
else if (ldda < max(1,m))
arginfo = -4;
if (arginfo != 0) {
magma_xerbla( __func__, -(arginfo) );
return arginfo;
}
/* Quick return if possible */
if (m == 0 || n == 0)
if(min_mn == 0 ) return arginfo;
if( m > 2048 || n > 2048 ) {
printf("=========================================================================================\n");
printf(" WARNING batched routines are designed for small sizes it might be better to use the\n Native/Hybrid classical routines if you want performance\n");
printf("=========================================================================================\n");
}
magma_int_t nb = 32;
magma_int_t nnb = 8;
magma_int_t i, k, ib=nb, jb=nnb;
magma_int_t ldw, ldt, ldr, offset;
cublasHandle_t myhandle;
cublasCreate_v2(&myhandle);
magmaDoubleComplex **dW0_displ = NULL;
magmaDoubleComplex **dW1_displ = NULL;
magmaDoubleComplex **dW2_displ = NULL;
magmaDoubleComplex **dW3_displ = NULL;
magmaDoubleComplex **dW4_displ = NULL;
magmaDoubleComplex **dW5_displ = NULL;
magmaDoubleComplex *dwork = NULL;
magmaDoubleComplex *dT = NULL;
magmaDoubleComplex *dR = NULL;
magmaDoubleComplex **dR_array = NULL;
magmaDoubleComplex **dT_array = NULL;
magmaDoubleComplex **cpuAarray = NULL;
magmaDoubleComplex **cpuTarray = NULL;
magma_malloc((void**)&dW0_displ, batchCount * sizeof(*dW0_displ));
magma_malloc((void**)&dW1_displ, batchCount * sizeof(*dW1_displ));
magma_malloc((void**)&dW2_displ, batchCount * sizeof(*dW2_displ));
magma_malloc((void**)&dW3_displ, batchCount * sizeof(*dW3_displ));
magma_malloc((void**)&dW4_displ, batchCount * sizeof(*dW4_displ)); // used in zlarfb
magma_malloc((void**)&dW5_displ, batchCount * sizeof(*dW5_displ));
magma_malloc((void**)&dR_array, batchCount * sizeof(*dR_array));
magma_malloc((void**)&dT_array, batchCount * sizeof(*dT_array));
ldt = ldr = min(nb, min_mn);
magma_zmalloc(&dwork, (2 * nb * n) * batchCount);
magma_zmalloc(&dR, ldr * n * batchCount);
magma_zmalloc(&dT, ldt * ldt * batchCount);
magma_malloc_cpu((void**) &cpuAarray, batchCount*sizeof(magmaDoubleComplex*));
magma_malloc_cpu((void**) &cpuTarray, batchCount*sizeof(magmaDoubleComplex*));
/* check allocation */
if ( dW0_displ == NULL || dW1_displ == NULL || dW2_displ == NULL || dW3_displ == NULL ||
dW4_displ == NULL || dW5_displ == NULL || dR_array == NULL || dT_array == NULL ||
dR == NULL || dT == NULL || dwork == NULL || cpuAarray == NULL || cpuTarray == NULL ) {
magma_free(dW0_displ);
magma_free(dW1_displ);
magma_free(dW2_displ);
magma_free(dW3_displ);
magma_free(dW4_displ);
magma_free(dW5_displ);
magma_free(dR_array);
magma_free(dT_array);
magma_free(dR);
magma_free(dT);
magma_free(dwork);
free(cpuAarray);
free(cpuTarray);
magma_int_t info = MAGMA_ERR_DEVICE_ALLOC;
magma_xerbla( __func__, -(info) );
return info;
}
magmablas_zlaset_q(MagmaFull, ldr, n*batchCount , MAGMA_Z_ZERO, MAGMA_Z_ZERO, dR, ldr, queue);
magmablas_zlaset_q(MagmaFull, ldt, ldt*batchCount, MAGMA_Z_ZERO, MAGMA_Z_ZERO, dT, ldt, queue);
zset_pointer(dR_array, dR, 1, 0, 0, ldr*min(nb, min_mn), batchCount, queue);
zset_pointer(dT_array, dT, 1, 0, 0, ldt*min(nb, min_mn), batchCount, queue);
magma_queue_t cstream;
magmablasGetKernelStream(&cstream);
magma_int_t streamid;
const magma_int_t nbstreams=32;
magma_queue_t stream[nbstreams];
for(i=0; i<nbstreams; i++) {
magma_queue_create( &stream[i] );
}
magma_getvector( batchCount, sizeof(magmaDoubleComplex*), dA_array, 1, cpuAarray, 1);
magma_getvector( batchCount, sizeof(magmaDoubleComplex*), dT_array, 1, cpuTarray, 1);
magmablasSetKernelStream(NULL);
for(i=0; i<min_mn; i+=nb)
{
ib = min(nb, min_mn-i);
//===============================================
// panel factorization
//===============================================
magma_zdisplace_pointers(dW0_displ, dA_array, ldda, i, i, batchCount, queue);
magma_zdisplace_pointers(dW2_displ, tau_array, 1, i, 0, batchCount, queue);
//dwork is used in panel factorization and trailing matrix update
//dW4_displ, dW5_displ are used as workspace and configured inside
magma_zgeqrf_panel_batched(m-i, ib, jb,
dW0_displ, ldda,
dW2_displ,
dT_array, ldt,
dR_array, ldr,
dW1_displ,
dW3_displ,
dwork,
dW4_displ, dW5_displ,
info_array,
batchCount, myhandle, queue);
//===============================================
// end of panel
//===============================================
//direct panel matrix V in dW0_displ,
magma_zdisplace_pointers(dW0_displ, dA_array, ldda, i, i, batchCount, queue);
// copy the upper part of V into dR
zgeqrf_copy_upper_batched(ib, jb, dW0_displ, ldda, dR_array, ldr, batchCount, queue);
//===============================================
// update trailing matrix
//===============================================
//dwork is used in panel factorization and trailing matrix update
//reset dW4_displ
ldw = nb;
zset_pointer(dW4_displ, dwork, 1, 0, 0, ldw*n, batchCount, queue );
offset = ldw*n*batchCount;
zset_pointer(dW5_displ, dwork + offset, 1, 0, 0, ldw*n, batchCount, queue );
if( (n-ib-i) > 0)
{
// set the diagonal of v as one and the upper triangular part as zero
magmablas_zlaset_batched(MagmaUpper, ib, ib, MAGMA_Z_ZERO, MAGMA_Z_ONE, dW0_displ, ldda, batchCount, queue);
magma_zdisplace_pointers(dW2_displ, tau_array, 1, i, 0, batchCount, queue);
// it is faster since it is using BLAS-3 GEMM routines, different from lapack implementation
magma_zlarft_batched(m-i, ib, 0,
dW0_displ, ldda,
dW2_displ,
dT_array, ldt,
dW4_displ, nb*ldt,
batchCount, myhandle, queue);
// perform C = (I-V T^H V^H) * C, C is the trailing matrix
//-------------------------------------------
// USE STREAM GEMM
//-------------------------------------------
if( (m-i) > 100 && (n-i-ib) > 100)
{
// But since the code use the NULL stream everywhere,
// so I don't need it, because the NULL stream do the sync by itself
//magma_device_sync();
for(k=0; k<batchCount; k++)
{
streamid = k%nbstreams;
magmablasSetKernelStream(stream[streamid]);
// the stream gemm must take cpu pointer
magma_zlarfb_gpu_gemm(MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-i, n-i-ib, ib,
cpuAarray[k] + i + i * ldda, ldda,
cpuTarray[k], ldt,
cpuAarray[k] + i + (i+ib) * ldda, ldda,
dwork + nb * n * k, -1,
dwork + nb * n * batchCount + nb * n * k, -1);
}
// need to synchronise to be sure that panel does not start before
// finishing the update at least of the next panel
// BUT no need for it as soon as the other portion of the code
// use the NULL stream which do the sync by itself
//magma_device_sync();
magmablasSetKernelStream(NULL);
}
//-------------------------------------------
// USE BATCHED GEMM
//-------------------------------------------
else
{
//direct trailing matrix in dW1_displ
magma_zdisplace_pointers(dW1_displ, dA_array, ldda, i, i+ib, batchCount, queue);
magma_zlarfb_gemm_batched(
MagmaLeft, MagmaConjTrans, MagmaForward, MagmaColumnwise,
m-i, n-i-ib, ib,
(const magmaDoubleComplex**)dW0_displ, ldda,
(const magmaDoubleComplex**)dT_array, ldt,
dW1_displ, ldda,
dW4_displ, ldw,
dW5_displ, ldw,
batchCount, myhandle, queue);
}
}// update the trailing matrix
//===============================================
// copy dR back to V after the trailing matrix update
magmablas_zlacpy_batched(MagmaUpper, ib, ib, dR_array, ldr, dW0_displ, ldda, batchCount, queue);
}
for(k=0; k<nbstreams; k++) {
magma_queue_destroy( stream[k] );
}
magmablasSetKernelStream(cstream);
cublasDestroy_v2(myhandle);
magma_free(dW0_displ);
magma_free(dW1_displ);
magma_free(dW2_displ);
magma_free(dW3_displ);
magma_free(dW4_displ);
magma_free(dW5_displ);
magma_free(dR_array);
magma_free(dT_array);
magma_free(dR);
magma_free(dT);
magma_free(dwork);
free(cpuAarray);
free(cpuTarray);
return arginfo;
}
extern "C" magma_int_t
magma_zbulge_applyQ_v2_m(
magma_int_t ngpu,
magma_side_t side,
magma_int_t NE, magma_int_t N,
magma_int_t NB, magma_int_t Vblksiz,
magmaDoubleComplex *E, magma_int_t lde,
magmaDoubleComplex *V, magma_int_t ldv,
magmaDoubleComplex *T, magma_int_t ldt,
magma_int_t *info)
{
//%===========================
//% local variables
//%===========================
magma_int_t Vm, Vn, mt, nt;
magma_int_t myrow, mycol, blkj, blki;
magma_int_t blkid,vpos,tpos;
magma_int_t firstrow, nbcolinvolvd;
magma_int_t versionL = 113;
magma_int_t versionR = 92;
magma_int_t Vchunksiz = 10;
*info=0;
/* Quick return */
if ( NE == 0 ) {
return *info;
}
if ( N == 0 ) {
return *info;
}
if ( NB == 0 ) {
return *info;
}
/* ==========================================
* some infos for developer
* Initialisation and checking nb of cores
* ==========================================*/
/* we have 2 algo for left (113 114) and 2 algo for right (91 92)
* which correspond to versionL versionR.
* They are very similar (detail explained in tech report and matlab code)
* however version 114 and 92 improve locality.
* while version 113 is used in case WNATZ=1 (construct Q2) which allow
* the construction to be done in an optimized way taking into
* consideration that the matrix is Identity so making less flops.
*
*/
// Initialize streaming and events
magma_device_sync();
magma_device_t orig_dev;
magma_getdevice( &orig_dev );
magma_queue_t orig_stream;
magmablasGetKernelStream( &orig_stream );
magma_int_t nbevents =2, nstream=2;
magma_queue_t streams[MagmaMaxGPUs][20];
magma_event_t myevent[MagmaMaxGPUs][20];
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_create( &streams[dev][i] );
}
for( magma_int_t i = 0; i < nbevents; ++i ) {
cudaEventCreateWithFlags(&myevent[dev][i],cudaEventDisableTiming);
}
}
// Azzam 21/11/2012
// NOTE THAT dwork was of size 2*NE*Vblksiz+...
// but I am thinking why not modifing it to NE*Vblksiz+...
// BUT NO because the 2* is used because of making 2 streams working and so
// they might be using dwork in parallel
magmaDoubleComplex *dE[MagmaMaxGPUs];
magmaDoubleComplex *dwork[MagmaMaxGPUs], *dwork0[MagmaMaxGPUs], *dwork1[MagmaMaxGPUs];
//magmaDoubleComplex *dwvt[MagmaMaxGPUs];
magmaDoubleComplex *dwvt0[MagmaMaxGPUs], *dwvt1[MagmaMaxGPUs];
magmaDoubleComplex *dT0[MagmaMaxGPUs], *dV0[MagmaMaxGPUs], *dT1[MagmaMaxGPUs], *dV1[MagmaMaxGPUs];
magma_int_t dev;
magma_int_t ldde = N;
magma_int_t lddv = ldv;
magma_int_t lddt = ldt;
magma_int_t ne_loc = magma_ceildiv(NE, ngpu);
if (ne_loc < 256)
ne_loc=256;
magma_int_t dwVTsiz = lddv*Vblksiz; // lddv*lddv + lddv*NE; // lddv*Vblksiz;
magma_int_t dworksiz = ne_loc*Vblksiz; // lddv*Vblksiz; // NE*Vblksiz;
ngpu = min(ngpu, magma_ceildiv(NE,ne_loc)); // Don't use GPU that will not have data.
// copy dE to GPUs
for (dev=0; dev < ngpu; ++dev) {
magma_setdevice( dev );
if (MAGMA_SUCCESS != magma_zmalloc( &dE[dev], ldde * ne_loc)) {
printf ("!!!! magma_zbulge_applyQ magma_alloc failed for: dE\n" );
exit(-1);
}
if (MAGMA_SUCCESS != magma_zmalloc( &dwork[dev], 2*dworksiz + 2*dwVTsiz + 2*Vchunksiz* (Vblksiz* (lddv+lddt)) )) {
printf ("!!!! magma_zbulge_applyQ magma_alloc failed for: dwork\n" );
exit(-1);
}
dwork0[dev] = dwork[dev]; // size = dworksiz;
dwork1[dev] = dwork0[dev] + dworksiz; // size = dworksiz;
dwvt0[dev] = dwork[dev] + 2*dworksiz; // size = dwVTsiz;
dwvt1[dev] = dwvt0[dev] + dwVTsiz; // size = dwVTsiz;
dV0[dev] = dwork[dev] + 2*dworksiz + 2*dwVTsiz;
dT0[dev] = dV0[dev] + Vchunksiz*Vblksiz*lddv;
dV1[dev] = dT0[dev] + Vchunksiz*Vblksiz*lddt;
dT1[dev] = dV1[dev] + Vchunksiz*Vblksiz*lddv;
magma_int_t ie_loc = min(ne_loc, NE - ne_loc*dev);
magma_zsetmatrix_async( N, ie_loc, E+lde*ne_loc*dev, lde, dE(dev, 0, 0), ldde, streams[dev][1] );
}
// make overlapped copy
magma_int_t ncpy = 0;
magma_int_t copyed=0, copyst=0;
magma_int_t blkcnt,nothing, mysiz, flip, vld,tld, locpos;
findVTsiz(N, NB, Vblksiz, &blkcnt, ¬hing);
flip = 0;
/* SIDE LEFT meaning apply E = Q*E = (q_1*q_2*.....*q_n) * E ==> so traverse Vs in reverse order (forward) from q_n to q_1
* Also E is splitten by row meaning each apply consist in a block of row (horizontal block) */
/* SIDE RIGHT meaning apply E = E*Q = E * (q_1*q_2*.....*q_n) ==> so tarverse Vs in normal order (forward) from q_1 to q_n
* Also E is splitten by col meaning each apply consist in a block of col (vertical block) */
#ifdef ENABLE_DEBUG
printf(" APPLY Q_v22_m GPU with NGPU %d N %d, NE %d, NB %d, Vblksiz %d, versionL %d versionR %d SIDE %c \n",
ngpu, N, NE, NB, Vblksiz, versionL, versionR, side);
#endif
/*
* MagmamaLeft
*/
if (side == MagmaLeft) {
/*
* Version 113:
* loop over the block_col (nt) and for each find the
* number of tiles (mt) in this block_col. then loop over mt, find
* the size of the V's(Vm,Vn) and apply it to the corresponding
* portion of E.
*/
if ( versionL == 113 ) {
nt = magma_ceildiv((N-1),Vblksiz);
for (blkj=nt-1; blkj >= 0; blkj--) {
/* the index of the first row on the top of block (blkj) */
firstrow = blkj * Vblksiz + 1;
/*find the number of tile for this block */
if ( blkj == nt-1 )
mt = magma_ceildiv( N - firstrow, NB);
else
mt = magma_ceildiv( N - (firstrow+1), NB);
/*loop over the tiles find the size of the Vs and apply it */
for (blki=mt; blki > 0; blki--) {
/*calculate the size of each losange of Vs= (Vm,Vn)*/
myrow = firstrow + (mt-blki)*NB;
mycol = blkj*Vblksiz;
Vm = min( NB+Vblksiz-1, N-myrow);
if ( ( blkj == nt-1 ) && ( blki == mt ) ) {
Vn = min (Vblksiz, Vm);
} else {
Vn = min (Vblksiz, Vm-1);
}
/*calculate the pointer to the Vs and the Ts.
* Note that Vs and Ts have special storage done
* by the bulgechasing function*/
//printf("voici blkj %d blki %d Vm %d Vn %d mycol %d vpos %d \n",blkj,blki,Vm, Vn,mycol,vpos);
magma_bulge_findpos113(N, NB, Vblksiz, mycol, myrow, &blkid);
// COPY Vchunksiz Vs and Vchunksiz Ts to GPU and store it in dV0/dV1 and dT0/dT1
if (ncpy == 0) {
// flip = 1 for this.
copyst = 0; // meaning that copy will start copying from blkid =copyst
copyed = min(copyst+Vchunksiz, blkcnt); // meaning that copy will end copying at blkid =copyed-1==> next copy had to start at copyed
mysiz = copyed-copyst; // the size of the chunk to be copied
if (mysiz > 0) {
ncpy = 1;
flip = 1;
vpos = copyst*Vblksiz*ldv;
tpos = copyst*Vblksiz*ldt;
vld = mysiz * ldv;
tld = mysiz * ldt;
for( dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 1 ] );
magma_zsetmatrix_async(vld, Vblksiz, V(vpos), vld, dV1[dev], vld, streams[dev][1]);
magma_zsetmatrix_async(tld, Vblksiz, T(tpos), tld, dT1[dev], tld, streams[dev][1]);
}
//printf("doing the first copy of mysiz %2d copyst %2d copyed %2d vpos %8d tpos %8d into dV1 dT1\n",mysiz,copyst,copyed,vpos,tpos);
}
}
if (blkid == copyst) {
flip = ncpy % 2;
copyst = copyed; // meaning that copy will start copying from blkid =copyst
copyed = min(copyst+Vchunksiz, blkcnt); // meaning that copy will end copying at blkid =copyed-1==> next copy had to start at copyed
mysiz = copyed-copyst; // the size of the chunk to be copied
//printf(" get to copy blkid %d blkid+(2*Vchunksiz) %d copyst %d copyed %d\n",blkid,blkid+(Vchunksiz),copyst,copyed);
if (mysiz > 0) {
ncpy = ncpy + 1;
vpos = copyst*Vblksiz*ldv;
tpos = copyst*Vblksiz*ldt;
vld = mysiz * ldv;
tld = mysiz * ldt;
if (flip == 0) { // now I am working on dV0 so copy the next and put it on dV1
//printf("doing overlapping copy of mysiz %2d copyst %2d copyed %2d vpos %8d tpos %8d into dV1 dT1\n",mysiz,copyst,copyed,vpos,tpos);
for( dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 1 ] );
magma_zsetmatrix_async(vld, Vblksiz, V(vpos), vld, dV1[dev], vld, streams[dev][1]);
magma_zsetmatrix_async(tld, Vblksiz, T(tpos), tld, dT1[dev], tld, streams[dev][1]);
}
} else { // now I am working on dV1 so copy the next and put it on dV0
//printf("doing overlapping copy of mysiz %2d copyst %2d copyed %2d vpos %8d tpos %8d into dV0 dT0\n",mysiz,copyst,copyed,vpos,tpos);
for( dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magmablasSetKernelStream( streams[ dev ][ 0 ] );
magma_zsetmatrix_async(vld, Vblksiz, V(vpos), vld, dV0[dev], vld, streams[dev][0]);
magma_zsetmatrix_async(tld, Vblksiz, T(tpos), tld, dT0[dev], tld, streams[dev][0]);
}
}
}
}
if ((Vm > 0) && (Vn > 0)) {
locpos = blkid%Vchunksiz;
magma_int_t lcvpos = locpos*Vblksiz*lddv;
magma_int_t lctpos = locpos*Vblksiz*lddt;
//printf("voici blkj %d blki %d Vm %d Vn %d mycol %d locvpos %5d loctpos %5d blkid %2d using data in dV%1d dT%1d \n",blkj,blki,Vm, Vn,mycol,lcvpos,lctpos, blkid,flip,flip);
if (flip == 0) {
for( dev = 0; dev < ngpu; ++dev ) {
magma_int_t ie_loc = min(ne_loc, NE - ne_loc*dev);
magma_int_t nr_bl = magma_ceildiv(ie_loc,10000); //nr of blocks
magma_int_t sz_bl = magma_ceildiv(ie_loc,nr_bl*64)*64; //maximum size of blocks (to have blocks of around the same size and multiple of 64)
magma_int_t ib; //size of current block
magma_setdevice( dev );
magmablasSetKernelStream(streams[dev][0]);
magma_queue_wait_event( streams[dev][0], myevent[dev][1] );
for (magma_int_t i=0; i < ie_loc; i += sz_bl) {
ib = min(sz_bl, ie_loc-i);
//magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV0[dev]+lcvpos, lddv, dT0[dev]+lctpos, lddt, dE(dev,myrow,i), ldde, dwork0[dev], ib);
magma_zlarfb_gpu_gemm( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV0[dev]+lcvpos, lddv, dT0[dev]+lctpos, lddt, dE(dev,myrow,i), ldde, dwork0[dev], ib, dwvt0[dev], Vm);
}
magma_event_record( myevent[dev][0], streams[dev][0] );
}
} else {
for( dev = 0; dev < ngpu; ++dev ) {
magma_int_t ie_loc = min(ne_loc, NE - ne_loc*dev);
magma_int_t nr_bl = magma_ceildiv(ie_loc,10000); //nr of blocks
magma_int_t sz_bl = magma_ceildiv(ie_loc,nr_bl*64)*64; //maximum size of blocks (to have blocks of around the same size and multiple of 64)
magma_int_t ib; //size of current block
magma_setdevice( dev );
magmablasSetKernelStream(streams[dev][1]);
magma_queue_wait_event( streams[dev][1], myevent[dev][0] );
for (magma_int_t i=0; i < ie_loc; i += sz_bl) {
ib = min(sz_bl, ie_loc-i);
//magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV1[dev]+lcvpos, lddv, dT1[dev]+lctpos, lddt, dE(dev,myrow,i), ldde, dwork1[dev], ib);
magma_zlarfb_gpu_gemm( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV1[dev]+lcvpos, lddv, dT1[dev]+lctpos, lddt, dE(dev,myrow,i), ldde, dwork1[dev], ib, dwvt1[dev], Vm);
}
magma_event_record( myevent[dev][1], streams[dev][1] );
}
}
} // end for (Vm &Vn) > 0
} // end for blki
} // end for blkj
} // end if version=113
/*
* Version 114:
* loop over the block_row (mt) and for each find diagonally the
* number of tiles (nt) in this block_row. then loop over nt, find
* the size of the V's(Vm,Vn) and apply it to the corresponding
* portion of E.
*/
else {
printf("versionL 114 not implemented in zbulge_applyQ_v2_m\n");
exit(-1);
mt = magma_ceildiv((N-1),NB);
for (blki = mt; blki > 0; blki--) {
/* nbcolinvolvd = number of column corresponding to this block_row (blki) */
nbcolinvolvd = min(N-1, blki*NB);
/*find the number of tile for this block (diagonal row of tiles) */
nt = magma_ceildiv(nbcolinvolvd,Vblksiz);
/*loop over the tiles find the size of the Vs and apply it */
for (blkj = nt-1; blkj >= 0; blkj--) {
/* the index of the first row of the first col meaning
* the block on the top left (blki) */
firstrow = (mt-blki)*NB+1;
/*calculate the size of each losange of Vs= (Vm,Vn)*/
myrow = firstrow + blkj*Vblksiz;
mycol = blkj*Vblksiz;
Vm = min( NB+Vblksiz-1, N-myrow);
if ( ( blkj == nt-1 ) && ( blki == mt ) ) {
Vn = min (Vblksiz, Vm);
} else {
Vn = min (Vblksiz, Vm-1);
}
if ((Vm > 0) && (Vn > 0)) {
/*calculate the pointer to the Vs and the Ts.
* Note that Vs and Ts have special storage done
* by the bulgechasing function*/
/*
magma_bulge_findVTpos(N, NB, Vblksiz, mycol, myrow, ldv, ldt, &vpos, &tpos);
magma_zsetmatrix_async(Vm, Vn, V(vpos), ldv, dV0, lddv, NULL);
magma_zsetmatrix_async(Vn, Vn, T(tpos), ldt, dT0, lddt, NULL);
//printf("voici blki %d rownbm %d mycol %d coled %d blkid %d vpos %d tpos %d\n", blki, rownbm, mycol, coled, blkid, vpos, tpos);
for (magma_int_t i=0; i < NE; i += sz_bl) {
ib = min(sz_bl, NE-i);
magma_zlarfb_gpu( MagmaLeft, MagmaNoTrans, MagmaForward, MagmaColumnwise, Vm, ib, Vn, dV0, lddv, dT0, lddt, dE(myrow,i), ldde, dwork, NE);
}
*/
} // end for (Vm &Vn) > 0
} // end for blkj
} // end for blki
} // end version 114
} // end LEFT
/*
* MagmaRight
*/
else {
printf("Side 'R' not implemented in zbulge_applyQ_v2_m\n");
exit(-1);
/*
* Version 91:
*/
if ( versionR == 91 ) {
nt = magma_ceildiv((N-1),Vblksiz);
for (blkj=0; blkj < nt; blkj++) {
/* the index of the first myrow on the top of block (blkj) */
firstrow = blkj * Vblksiz + 1;
/*find the number of tile for this block */
if ( blkj == nt-1 )
mt = magma_ceildiv( N - firstrow, NB);
else
mt = magma_ceildiv( N - (firstrow+1), NB);
/*loop over the tiles find the size of the Vs and apply it */
for (blki=1; blki <= mt; blki++) {
/*calculate the size of each losange of Vs= (Vm,Vn)*/
myrow = firstrow + (mt-blki)*NB;
Vm = min( NB+Vblksiz-1, N-myrow);
if ( (blkj == nt-1) && (blki == mt) ) {
Vn = min (Vblksiz, Vm);
} else {
Vn = min (Vblksiz, Vm-1);
}
mycol = blkj*Vblksiz;
if ((Vm > 0) && (Vn > 0)) {
/*calculate the pointer to the Vs and the Ts.
* Note that Vs and Ts have special storage done
* by the bulgechasing function*/
/*
magma_bulge_findVTpos(N, NB, Vblksiz, mycol, myrow, ldv, ldt, &vpos, &tpos);
magma_zsetmatrix_async(Vm, Vn, V(vpos), ldv, dV0, lddv, NULL);
magma_zsetmatrix_async(Vn, Vn, T(tpos), ldt, dT0, lddt, NULL);
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, NE, Vm, Vn, dV0, lddv, dT0, lddt, dE(0, myrow), ldde, dwork, NE);
*/
} // end for (Vm &Vn) > 0
} // end for blki
} // end fo blkj
} // end of version 91
/*
* Version 92:
*/
else {
mt = magma_ceildiv((N-1),NB);
for (blki = 1; blki <= mt; blki++) {
/* nbcolinvolvd = number of column corresponding to this block_row (blki) */
nbcolinvolvd = min(N-1, blki*NB);
/*find the number of tile for this block (diagonal row of tiles) */
nt = magma_ceildiv(nbcolinvolvd,Vblksiz);
/*loop over the tiles find the size of the Vs and apply it */
for (blkj = 0; blkj < nt; blkj++) {
/* the index of the first row of the first col meaning
* the block on the top left (blki) */
firstrow = (mt-blki)*NB+1;
/*calculate the size of each losange of Vs= (Vm,Vn)*/
myrow = firstrow + blkj*Vblksiz;
mycol = blkj*Vblksiz;
Vm = min( NB+Vblksiz-1, N-myrow);
if ( ( blkj == nt-1 ) && ( blki == mt ) ) {
Vn = min (Vblksiz, Vm);
} else {
Vn = min (Vblksiz, Vm-1);
}
if ((Vm > 0) && (Vn > 0)) {
/*calculate the pointer to the Vs and the Ts.
* Note that Vs and Ts have special storage done
* by the bulgechasing function*/
/*
magma_bulge_findVTpos(N, NB, Vblksiz, mycol, myrow, ldv, ldt, &vpos, &tpos);
magma_zsetmatrix_async(Vm, Vn, V(vpos), ldv, dV0, lddv, NULL);
magma_zsetmatrix_async(Vn, Vn, T(tpos), ldt, dT0, lddt, NULL);
magma_zlarfb_gpu( MagmaRight, MagmaNoTrans, MagmaForward, MagmaColumnwise, NE, Vm, Vn, dV0, lddv, dT0, lddt, dE(0, myrow), ldde, dwork, NE);
*/
} // end for (Vm &Vn) > 0
} //end for blkj
} // end for blki
} //end of version 92
} // end RIGHT
// copy back the dE form each GPU
for( dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magmablasSetKernelStream(streams[dev][0]);
magma_queue_wait_event( streams[dev][0], myevent[dev][1] );
magma_queue_wait_event( streams[dev][0], myevent[dev][0] );
magma_int_t ie_loc = min(ne_loc, NE - ne_loc*dev);
magma_zgetmatrix_async( N, ie_loc, dE(dev, 0, 0), ldde, E+lde*ne_loc*dev, lde, streams[dev][0] );
magma_event_record( myevent[dev][0], streams[dev][0] );
}
for( magma_int_t dev = 0; dev < ngpu; ++dev ) {
magma_setdevice( dev );
magmablasSetKernelStream(streams[dev][0]);
magma_queue_wait_event( streams[dev][0], myevent[dev][0] );
magma_device_sync(); // no need for synchronize
magma_free(dwork[dev]);
magma_free(dE[dev]);
for( magma_int_t i = 0; i < nbevents; ++i ) {
magma_event_destroy( myevent[dev][i] );
}
for( magma_int_t i = 0; i < nstream; ++i ) {
magma_queue_destroy( streams[dev][i] );
}
}
magma_setdevice( orig_dev );
magmablasSetKernelStream( orig_stream );
return *info;
}
int main( int argc, char** argv )
{
TESTING_INIT();
real_Double_t gflops, t1, t2;
magmaDoubleComplex c_neg_one = MAGMA_Z_NEG_ONE;
magma_int_t ione = 1;
magma_trans_t trans[] = { MagmaNoTrans, MagmaConjTrans, MagmaTrans };
magma_uplo_t uplo [] = { MagmaLower, MagmaUpper };
magma_diag_t diag [] = { MagmaUnit, MagmaNonUnit };
magma_side_t side [] = { MagmaLeft, MagmaRight };
magmaDoubleComplex *A, *B, *C, *C2, *LU;
magmaDoubleComplex *dA, *dB, *dC1, *dC2;
magmaDoubleComplex alpha = MAGMA_Z_MAKE( 0.5, 0.1 );
magmaDoubleComplex beta = MAGMA_Z_MAKE( 0.7, 0.2 );
double dalpha = 0.6;
double dbeta = 0.8;
double work[1], error, total_error;
magma_int_t ISEED[4] = {0,0,0,1};
magma_int_t m, n, k, size, maxn, ld, info;
magma_int_t *piv;
magma_int_t err;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf( "Compares magma wrapper function to cublas function; all diffs should be exactly 0.\n\n" );
total_error = 0.;
for( int itest = 0; itest < opts.ntest; ++itest ) {
m = opts.msize[itest];
n = opts.nsize[itest];
k = opts.ksize[itest];
printf("=========================================================================\n");
printf( "m=%d, n=%d, k=%d\n", (int) m, (int) n, (int) k );
// allocate matrices
// over-allocate so they can be any combination of {m,n,k} x {m,n,k}.
maxn = max( max( m, n ), k );
ld = max( 1, maxn );
size = ld*maxn;
err = magma_malloc_cpu( (void**) &piv, maxn*sizeof(magma_int_t) ); assert( err == 0 );
err = magma_zmalloc_pinned( &A, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &B, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &C, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &C2, size ); assert( err == 0 );
err = magma_zmalloc_pinned( &LU, size ); assert( err == 0 );
err = magma_zmalloc( &dA, size ); assert( err == 0 );
err = magma_zmalloc( &dB, size ); assert( err == 0 );
err = magma_zmalloc( &dC1, size ); assert( err == 0 );
err = magma_zmalloc( &dC2, size ); assert( err == 0 );
// initialize matrices
size = maxn*maxn;
lapackf77_zlarnv( &ione, ISEED, &size, A );
lapackf77_zlarnv( &ione, ISEED, &size, B );
lapackf77_zlarnv( &ione, ISEED, &size, C );
printf( "========== Level 1 BLAS ==========\n" );
// ----- test ZSWAP
// swap columns 2 and 3 of dA, then copy to C2 and compare with A
if ( n >= 3 ) {
magma_zsetmatrix( m, n, A, ld, dA, ld );
magma_zsetmatrix( m, n, A, ld, dB, ld );
magma_zswap( m, dA(0,1), 1, dA(0,2), 1 );
magma_zswap( m, dB(0,1), 1, dB(0,2), 1 );
// check results, storing diff between magma and cuda calls in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dA, 1, dB, 1 );
magma_zgetmatrix( m, n, dB, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &k, C2, &ld, work );
total_error += error;
printf( "zswap diff %.2g\n", error );
}
else {
printf( "zswap skipped for n < 3\n" );
}
// ----- test IZAMAX
// get argmax of column of A
magma_zsetmatrix( m, k, A, ld, dA, ld );
error = 0;
for( int j = 0; j < k; ++j ) {
magma_int_t i1 = magma_izamax( m, dA(0,j), 1 );
int i2; // NOT magma_int_t, for cublas
cublasIzamax( handle, m, dA(0,j), 1, &i2 );
// todo need sync here?
assert( i1 == i2 );
error += abs( i1 - i2 );
}
total_error += error;
gflops = (double)m * k / 1e9;
printf( "izamax diff %.2g\n", error );
printf( "\n" );
printf( "========== Level 2 BLAS ==========\n" );
// ----- test ZGEMV
// c = alpha*A*b + beta*c, with A m*n; b,c m or n-vectors
// try no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
magma_zsetmatrix( m, n, A, ld, dA, ld );
magma_zsetvector( maxn, B, 1, dB, 1 );
magma_zsetvector( maxn, C, 1, dC1, 1 );
magma_zsetvector( maxn, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_zgemv( trans[ia], m, n, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZgemv( handle, cublas_trans_const(trans[ia]),
m, n, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
size = (trans[ia] == MagmaNoTrans ? m : n);
cublasZaxpy( handle, size, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( size, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &size, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZGEMV( m, n ) / 1e9;
printf( "zgemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test ZHEMV
// c = alpha*A*b + beta*c, with A m*m symmetric; b,c m-vectors
// try upper/lower
for( int iu = 0; iu < 2; ++iu ) {
magma_zsetmatrix( m, m, A, ld, dA, ld );
magma_zsetvector( m, B, 1, dB, 1 );
magma_zsetvector( m, C, 1, dC1, 1 );
magma_zsetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_zhemv( uplo[iu], m, alpha, dA, ld, dB, 1, beta, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZhemv( handle, cublas_uplo_const(uplo[iu]),
m, &alpha, dA, ld, dB, 1, &beta, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHEMV( m ) / 1e9;
printf( "zhemv( %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), error, gflops/t1, gflops/t2 );
}
printf( "\n" );
// ----- test ZTRSV
// solve A*c = c, with A m*m triangular; c m-vector
// try upper/lower, no-trans/trans, unit/non-unit diag
// Factor A into LU to get well-conditioned triangles, else solve yields garbage.
// Still can give garbage if solves aren't consistent with LU factors,
// e.g., using unit diag for U, so copy lower triangle to upper triangle.
// Also used for trsm later.
lapackf77_zlacpy( "Full", &maxn, &maxn, A, &ld, LU, &ld );
lapackf77_zgetrf( &maxn, &maxn, LU, &ld, piv, &info );
for( int j = 0; j < maxn; ++j ) {
for( int i = 0; i < j; ++i ) {
*LU(i,j) = *LU(j,i);
}
}
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
magma_zsetmatrix( m, m, LU, ld, dA, ld );
magma_zsetvector( m, C, 1, dC1, 1 );
magma_zsetvector( m, C, 1, dC2, 1 );
t1 = magma_sync_wtime( 0 );
magma_ztrsv( uplo[iu], trans[it], diag[id], m, dA, ld, dC1, 1 );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZtrsv( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
cublas_diag_const(diag[id]), m, dA, ld, dC2, 1 );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, m, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetvector( m, dC2, 1, C2, 1 );
error = lapackf77_zlange( "F", &m, &ione, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRSM( MagmaLeft, m, 1 ) / 1e9;
printf( "ztrsv( %c, %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]), lapacke_diag_const(diag[id]),
error, gflops/t1, gflops/t2 );
}}}
printf( "\n" );
printf( "========== Level 3 BLAS ==========\n" );
// ----- test ZGEMM
// C = alpha*A*B + beta*C, with A m*k or k*m; B k*n or n*k; C m*n
// try combinations of no-trans/trans
for( int ia = 0; ia < 3; ++ia ) {
for( int ib = 0; ib < 3; ++ib ) {
bool nta = (trans[ia] == MagmaNoTrans);
bool ntb = (trans[ib] == MagmaNoTrans);
magma_zsetmatrix( (nta ? m : k), (nta ? m : k), A, ld, dA, ld );
magma_zsetmatrix( (ntb ? k : n), (ntb ? n : k), B, ld, dB, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zgemm( trans[ia], trans[ib], m, n, k, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZgemm( handle, cublas_trans_const(trans[ia]), cublas_trans_const(trans[ib]),
m, n, k, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZGEMM( m, n, k ) / 1e9;
printf( "zgemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_trans_const(trans[ia]), lapacke_trans_const(trans[ib]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHEMM
// C = alpha*A*B + beta*C (left) with A m*m symmetric; B,C m*n; or
// C = alpha*B*A + beta*C (right) with A n*n symmetric; B,C m*n
// try left/right, upper/lower
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
magma_zsetmatrix( m, m, A, ld, dA, ld );
magma_zsetmatrix( m, n, B, ld, dB, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zhemm( side[is], uplo[iu], m, n, alpha, dA, ld, dB, ld, beta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZhemm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
m, n, &alpha, dA, ld, dB, ld, &beta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &m, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHEMM( side[is], m, n ) / 1e9;
printf( "zhemm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_side_const(side[is]), lapacke_uplo_const(uplo[iu]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHERK
// C = alpha*A*A^H + beta*C (no-trans) with A m*k and C m*m symmetric; or
// C = alpha*A^H*A + beta*C (trans) with A k*m and C m*m symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
magma_zsetmatrix( n, k, A, ld, dA, ld );
magma_zsetmatrix( n, n, C, ld, dC1, ld );
magma_zsetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zherk( uplo[iu], trans[it], n, k, dalpha, dA, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZherk( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &dalpha, dA, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHERK( k, n ) / 1e9;
printf( "zherk( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZHER2K
// C = alpha*A*B^H + ^alpha*B*A^H + beta*C (no-trans) with A,B n*k; C n*n symmetric; or
// C = alpha*A^H*B + ^alpha*B^H*A + beta*C (trans) with A,B k*n; C n*n symmetric
// try upper/lower, no-trans/trans
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
bool nt = (trans[it] == MagmaNoTrans);
magma_zsetmatrix( (nt ? n : k), (nt ? n : k), A, ld, dA, ld );
magma_zsetmatrix( n, n, C, ld, dC1, ld );
magma_zsetmatrix( n, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_zher2k( uplo[iu], trans[it], n, k, alpha, dA, ld, dB, ld, dbeta, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZher2k( handle, cublas_uplo_const(uplo[iu]), cublas_trans_const(trans[it]),
n, k, &alpha, dA, ld, dB, ld, &dbeta, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( n, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZHER2K( k, n ) / 1e9;
printf( "zher2k( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}
printf( "\n" );
// ----- test ZTRMM
// C = alpha*A*C (left) with A m*m triangular; C m*n; or
// C = alpha*C*A (right) with A n*n triangular; C m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_zsetmatrix( (left ? m : n), (left ? m : n), A, ld, dA, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ztrmm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
// note cublas does trmm out-of-place (i.e., adds output matrix C),
// but allows C=B to do in-place.
t2 = magma_sync_wtime( 0 );
cublasZtrmm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRMM( side[is], m, n ) / 1e9;
printf( "ztrmm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// ----- test ZTRSM
// solve A*X = alpha*B (left) with A m*m triangular; B m*n; or
// solve X*A = alpha*B (right) with A n*n triangular; B m*n
// try left/right, upper/lower, no-trans/trans, unit/non-unit
for( int is = 0; is < 2; ++is ) {
for( int iu = 0; iu < 2; ++iu ) {
for( int it = 0; it < 3; ++it ) {
for( int id = 0; id < 2; ++id ) {
bool left = (side[is] == MagmaLeft);
magma_zsetmatrix( (left ? m : n), (left ? m : n), LU, ld, dA, ld );
magma_zsetmatrix( m, n, C, ld, dC1, ld );
magma_zsetmatrix( m, n, C, ld, dC2, ld );
t1 = magma_sync_wtime( 0 );
magma_ztrsm( side[is], uplo[iu], trans[it], diag[id], m, n, alpha, dA, ld, dC1, ld );
t1 = magma_sync_wtime( 0 ) - t1;
t2 = magma_sync_wtime( 0 );
cublasZtrsm( handle, cublas_side_const(side[is]), cublas_uplo_const(uplo[iu]),
cublas_trans_const(trans[it]), cublas_diag_const(diag[id]),
m, n, &alpha, dA, ld, dC2, ld );
t2 = magma_sync_wtime( 0 ) - t2;
// check results, storing diff between magma and cuda call in C2
cublasZaxpy( handle, ld*n, &c_neg_one, dC1, 1, dC2, 1 );
magma_zgetmatrix( m, n, dC2, ld, C2, ld );
error = lapackf77_zlange( "F", &n, &n, C2, &ld, work );
total_error += error;
gflops = FLOPS_ZTRSM( side[is], m, n ) / 1e9;
printf( "ztrsm( %c, %c ) diff %.2g, Gflop/s %7.2f, %7.2f\n",
lapacke_uplo_const(uplo[iu]), lapacke_trans_const(trans[it]),
error, gflops/t1, gflops/t2 );
}}}}
printf( "\n" );
// cleanup
magma_free_cpu( piv );
magma_free_pinned( A );
magma_free_pinned( B );
magma_free_pinned( C );
magma_free_pinned( C2 );
magma_free_pinned( LU );
magma_free( dA );
magma_free( dB );
magma_free( dC1 );
magma_free( dC2 );
fflush( stdout );
}
if ( total_error != 0. ) {
printf( "total error %.2g -- ought to be 0 -- some test failed (see above).\n",
total_error );
}
else {
printf( "all tests passed\n" );
}
TESTING_FINALIZE();
int status = (total_error != 0.);
return status;
}