void magma_cmake_hpd( magma_int_t N, magmaFloatComplex* A, magma_int_t lda )
{
magma_int_t i, j;
for( i=0; i<N; ++i ) {
A(i,i) = MAGMA_C_MAKE( MAGMA_C_REAL( A(i,i) ) + N, 0. );
for( j=0; j<i; ++j ) {
A(j,i) = MAGMA_C_CNJG( A(i,j) );
}
}
}
void init_matrix( int N, magmaFloatComplex *h_A, magma_int_t lda )
{
magma_int_t ione = 1, n2 = N*lda;
magma_int_t ISEED[4] = {0,0,0,1};
lapackf77_clarnv( &ione, ISEED, &n2, h_A );
/* Symmetrize and increase the diagonal */
for (int i = 0; i < N; ++i) {
MAGMA_C_SET2REAL( h_A(i,i), MAGMA_C_REAL(h_A(i,i)) + N );
for (int j = 0; j < i; ++j) h_A(i, j) = MAGMA_C_CNJG( h_A(j, i) );
}
}
extern "C" void
magma_ctrdtype2cbHLsym_withQ_v2(
magma_int_t n, magma_int_t nb,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *V, magma_int_t ldv,
magmaFloatComplex *TAU,
magma_int_t st, magma_int_t ed,
magma_int_t sweep, magma_int_t Vblksiz,
magmaFloatComplex *work)
{
/*
WORK (workspace) float complex array, dimension NB
*/
magma_int_t ione = 1;
magma_int_t vpos, taupos;
magmaFloatComplex conjtmp;
magmaFloatComplex c_one = MAGMA_C_ONE;
magma_int_t ldx = lda-1;
magma_int_t len = ed - st + 1;
magma_int_t lem = min(ed+nb, n) - ed;
magma_int_t lem2;
if (lem > 0) {
magma_bulge_findVTAUpos(n, nb, Vblksiz, sweep-1, st-1, ldv, &vpos, &taupos);
/* apply remaining right coming from the top block */
lapackf77_clarfx("R", &lem, &len, V(vpos), TAU(taupos), A(ed+1, st), &ldx, work);
}
if (lem > 1) {
magma_bulge_findVTAUpos(n, nb, Vblksiz, sweep-1, ed, ldv, &vpos, &taupos);
/* remove the first column of the created bulge */
*V(vpos) = c_one;
//memcpy(V(vpos+1), A(ed+2, st), (lem-1)*sizeof(magmaFloatComplex));
lem2 = lem-1;
blasf77_ccopy( &lem2, A(ed+2, st), &ione, V(vpos+1), &ione );
memset(A(ed+2, st),0,(lem-1)*sizeof(magmaFloatComplex));
/* Eliminate the col at st */
lapackf77_clarfg( &lem, A(ed+1, st), V(vpos+1), &ione, TAU(taupos) );
/* apply left on A(J1:J2,st+1:ed) */
len = len-1; /* because we start at col st+1 instead of st. col st is the col that has been removed; */
conjtmp = MAGMA_C_CNJG(*TAU(taupos));
lapackf77_clarfx("L", &lem, &len, V(vpos), &conjtmp, A(ed+1, st+1), &ldx, work);
}
}
extern "C" void
magma_ctrdtype2cbHLsym_withQ(
magma_int_t N, magma_int_t NB,
magmaFloatComplex *A, magma_int_t LDA,
magmaFloatComplex *V, magmaFloatComplex *TAU,
magma_int_t st, magma_int_t ed, magma_int_t sweep, magma_int_t Vblksiz)
{
magma_int_t J1, J2, len, lem, LDX;
//magma_int_t i, j;
magma_int_t IONE=1;
magma_int_t blkid, vpos, taupos, tpos;
magmaFloatComplex conjtmp;
magmaFloatComplex Z_ONE = MAGMA_C_ONE;
//magmaFloatComplex WORK[NB];
magmaFloatComplex *WORK;
magma_cmalloc_cpu( &WORK, NB );
findVTpos(N,NB,Vblksiz,sweep-1,st-1, &vpos, &taupos, &tpos, &blkid);
LDX = LDA-1;
J1 = ed+1;
J2 = min(ed+NB,N);
len = ed-st+1;
lem = J2-J1+1;
if (lem > 0) {
/* apply remaining right commming from the top block */
lapackf77_clarfx("R", &lem, &len, V(vpos), TAU(taupos), A(J1, st), &LDX, WORK);
}
if (lem > 1) {
findVTpos(N,NB,Vblksiz,sweep-1,J1-1, &vpos, &taupos, &tpos, &blkid);
/* remove the first column of the created bulge */
*V(vpos) = Z_ONE;
memcpy(V(vpos+1), A(J1+1, st), (lem-1)*sizeof(magmaFloatComplex));
memset(A(J1+1, st),0,(lem-1)*sizeof(magmaFloatComplex));
/* Eliminate the col at st */
lapackf77_clarfg( &lem, A(J1, st), V(vpos+1), &IONE, TAU(taupos) );
/* apply left on A(J1:J2,st+1:ed) */
len = len-1; /* because we start at col st+1 instead of st. col st is the col that has been revomved; */
conjtmp = MAGMA_C_CNJG(*TAU(taupos));
lapackf77_clarfx("L", &lem, &len, V(vpos), &conjtmp, A(J1, st+1), &LDX, WORK);
}
magma_free_cpu(WORK);
}
inline static void
magma_clarfxsym_v2(magma_int_t n,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *V, magmaFloatComplex *TAU,
magmaFloatComplex *work)
{
/*
WORK (workspace) float complex array, dimension N
*/
magma_int_t ione = 1;
magmaFloatComplex dtmp;
magmaFloatComplex c_zero = MAGMA_C_ZERO;
magmaFloatComplex c_neg_one= MAGMA_C_NEG_ONE;
magmaFloatComplex c_half = MAGMA_C_HALF;
/* X = AVtau */
blasf77_chemv("L",&n, TAU, A, &lda, V, &ione, &c_zero, work, &ione);
/* compute dtmp= X'*V */
#if defined(PRECISION_z) || defined(PRECISION_c)
dtmp = c_zero;
for (magma_int_t j = 0; j < n; j++)
dtmp = dtmp + MAGMA_C_CNJG(work[j]) * V[j];
//cblas_cdotc_sub(n, work, ione, V, ione, &dtmp);
#else
dtmp = cblas_cdotc(n, work, ione, V, ione);
#endif
/* compute 1/2 X'*V*t = 1/2*dtmp*tau */
dtmp = -dtmp * c_half * (*TAU);
/* compute W=X-1/2VX'Vt = X - dtmp*V */
blasf77_caxpy(&n, &dtmp, V, &ione, work, &ione);
/* performs the symmetric rank 2 operation A := alpha*x*y' + alpha*y*x' + A */
blasf77_cher2("L", &n, &c_neg_one, work, &ione, V, &ione, A, &lda);
}
magma_int_t magma_ctrevc3_mt(
magma_side_t side, magma_vec_t howmany,
magma_int_t *select, // logical in Fortran
magma_int_t n,
magmaFloatComplex *T, magma_int_t ldt,
magmaFloatComplex *VL, magma_int_t ldvl,
magmaFloatComplex *VR, magma_int_t ldvr,
magma_int_t mm, magma_int_t *mout,
magmaFloatComplex *work, magma_int_t lwork,
float *rwork, magma_int_t *info )
{
#define T(i,j) ( T + (i) + (j)*ldt )
#define VL(i,j) (VL + (i) + (j)*ldvl)
#define VR(i,j) (VR + (i) + (j)*ldvr)
#define work(i,j) (work + (i) + (j)*n)
// .. Parameters ..
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magmaFloatComplex c_one = MAGMA_C_ONE;
const magma_int_t nbmin = 16, nbmax = 128;
const magma_int_t ione = 1;
// .. Local Scalars ..
magma_int_t allv, bothv, leftv, over, rightv, somev;
magma_int_t i, ii, is, j, k, ki, iv, n2, nb, nb2, version;
float ovfl, remax, smin, smlnum, ulp, unfl;
// Decode and test the input parameters
bothv = (side == MagmaBothSides);
rightv = (side == MagmaRight) || bothv;
leftv = (side == MagmaLeft ) || bothv;
allv = (howmany == MagmaAllVec);
over = (howmany == MagmaBacktransVec);
somev = (howmany == MagmaSomeVec);
// Set mout to the number of columns required to store the selected
// eigenvectors.
if ( somev ) {
*mout = 0;
for( j=0; j < n; ++j ) {
if ( select[j] ) {
*mout += 1;
}
}
}
else {
*mout = n;
}
*info = 0;
if ( ! rightv && ! leftv )
*info = -1;
else if ( ! allv && ! over && ! somev )
*info = -2;
else if ( n < 0 )
*info = -4;
else if ( ldt < max( 1, n ) )
*info = -6;
else if ( ldvl < 1 || ( leftv && ldvl < n ) )
*info = -8;
else if ( ldvr < 1 || ( rightv && ldvr < n ) )
*info = -10;
else if ( mm < *mout )
*info = -11;
else if ( lwork < max( 1, 2*n ) )
*info = -14;
if ( *info != 0 ) {
magma_xerbla( __func__, -(*info) );
return *info;
}
// Quick return if possible.
if ( n == 0 ) {
return *info;
}
// Use blocked version (2) if sufficient workspace.
// Requires 1 vector to save diagonal elements, and 2*nb vectors for x and Q*x.
// (Compared to dtrevc3, rwork stores 1-norms.)
// Zero-out the workspace to avoid potential NaN propagation.
nb = 2;
if ( lwork >= n + 2*n*nbmin ) {
version = 2;
nb = (lwork - n) / (2*n);
nb = min( nb, nbmax );
nb2 = 1 + 2*nb;
lapackf77_claset( "F", &n, &nb2, &c_zero, &c_zero, work, &n );
}
else {
version = 1;
}
// Set the constants to control overflow.
unfl = lapackf77_slamch( "Safe minimum" );
ovfl = 1. / unfl;
lapackf77_slabad( &unfl, &ovfl );
ulp = lapackf77_slamch( "Precision" );
smlnum = unfl*( n / ulp );
// Store the diagonal elements of T in working array work.
for( i=0; i < n; ++i ) {
*work(i,0) = *T(i,i);
}
// Compute 1-norm of each column of strictly upper triangular
// part of T to control overflow in triangular solver.
rwork[0] = 0.;
for( j=1; j < n; ++j ) {
rwork[j] = cblas_scasum( j, T(0,j), ione );
}
// launch threads -- each single-threaded MKL
magma_int_t nthread = magma_get_parallel_numthreads();
magma_int_t lapack_nthread = magma_get_lapack_numthreads();
magma_set_lapack_numthreads( 1 );
magma_queue queue;
queue.launch( nthread );
//printf( "nthread %d, %d\n", nthread, lapack_nthread );
// NB = N/thread, rounded up to multiple of 16,
// but avoid multiples of page size, e.g., 512*8 bytes = 4096.
magma_int_t NB = magma_int_t( ceil( ceil( ((float)n) / nthread ) / 16. ) * 16. );
if ( NB % 512 == 0 ) {
NB += 32;
}
magma_timer_t time_total=0, time_trsv=0, time_gemm=0, time_gemv=0, time_trsv_sum=0, time_gemm_sum=0, time_gemv_sum=0;
timer_start( time_total );
if ( rightv ) {
// ============================================================
// Compute right eigenvectors.
// iv is index of column in current block.
// Non-blocked version always uses iv=1;
// blocked version starts with iv=nb, goes down to 1.
// (Note the "0-th" column is used to store the original diagonal.)
iv = 1;
if ( version == 2 ) {
iv = nb;
}
timer_start( time_trsv );
is = *mout - 1;
for( ki=n-1; ki >= 0; --ki ) {
if ( somev ) {
if ( ! select[ki] ) {
continue;
}
}
smin = max( ulp*( MAGMA_C_ABS1( *T(ki,ki) ) ), smlnum );
// --------------------------------------------------------
// Complex right eigenvector
*work(ki,iv) = c_one;
// Form right-hand side.
for( k=0; k < ki; ++k ) {
*work(k,iv) = -(*T(k,ki));
}
// Solve upper triangular system:
// [ T(1:ki-1,1:ki-1) - T(ki,ki) ]*X = scale*work.
if ( ki > 0 ) {
queue.push_task( new magma_clatrsd_task(
MagmaUpper, MagmaNoTrans, MagmaNonUnit, MagmaTrue,
ki, T, ldt, *T(ki,ki),
work(0,iv), work(ki,iv), rwork ));
}
// Copy the vector x or Q*x to VR and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VR and normalize
queue.sync();
n2 = ki+1;
blasf77_ccopy( &n2, work(0,iv), &ione, VR(0,is), &ione );
ii = blasf77_icamax( &n2, VR(0,is), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *VR(ii,is) );
blasf77_csscal( &n2, &remax, VR(0,is), &ione );
for( k=ki+1; k < n; ++k ) {
*VR(k,is) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
queue.sync();
time_trsv_sum += timer_stop( time_trsv );
timer_start( time_gemv );
if ( ki > 0 ) {
blasf77_cgemv( "n", &n, &ki, &c_one,
VR, &ldvr,
work(0, iv), &ione,
work(ki,iv), VR(0,ki), &ione );
}
time_gemv_sum += timer_stop( time_gemv );
ii = blasf77_icamax( &n, VR(0,ki), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *VR(ii,ki) );
blasf77_csscal( &n, &remax, VR(0,ki), &ione );
timer_start( time_trsv );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out below vector
for( k=ki+1; k < n; ++k ) {
*work(k,iv) = c_zero;
}
// Columns iv:nb of work are valid vectors.
// When the number of vectors stored reaches nb,
// or if this was last vector, do the GEMM
if ( (iv == 1) || (ki == 0) ) {
queue.sync();
time_trsv_sum += timer_stop( time_trsv );
timer_start( time_gemm );
nb2 = nb-iv+1;
n2 = ki+nb-iv+1;
// split gemm into multiple tasks, each doing one block row
for( i=0; i < n; i += NB ) {
magma_int_t ib = min( NB, n-i );
queue.push_task( new cgemm_task(
MagmaNoTrans, MagmaNoTrans, ib, nb2, n2, c_one,
VR(i,0), ldvr,
work(0,iv ), n, c_zero,
work(i,nb+iv), n ));
}
queue.sync();
time_gemm_sum += timer_stop( time_gemm );
// normalize vectors
// TODO if somev, should copy vectors individually to correct location.
for( k = iv; k <= nb; ++k ) {
ii = blasf77_icamax( &n, work(0,nb+k), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *work(ii,nb+k) );
blasf77_csscal( &n, &remax, work(0,nb+k), &ione );
}
lapackf77_clacpy( "F", &n, &nb2, work(0,nb+iv), &n, VR(0,ki), &ldvr );
iv = nb;
timer_start( time_trsv );
}
else {
iv -= 1;
}
} // blocked back-transform
is -= 1;
}
}
timer_stop( time_trsv );
timer_stop( time_total );
timer_printf( "trevc trsv %.4f, gemm %.4f, gemv %.4f, total %.4f\n",
time_trsv_sum, time_gemm_sum, time_gemv_sum, time_total );
if ( leftv ) {
// ============================================================
// Compute left eigenvectors.
// iv is index of column in current block.
// Non-blocked version always uses iv=1;
// blocked version starts with iv=1, goes up to nb.
// (Note the "0-th" column is used to store the original diagonal.)
iv = 1;
is = 0;
for( ki=0; ki < n; ++ki ) {
if ( somev ) {
if ( ! select[ki] ) {
continue;
}
}
smin = max( ulp*MAGMA_C_ABS1( *T(ki,ki) ), smlnum );
// --------------------------------------------------------
// Complex left eigenvector
*work(ki,iv) = c_one;
// Form right-hand side.
for( k = ki + 1; k < n; ++k ) {
*work(k,iv) = -MAGMA_C_CNJG( *T(ki,k) );
}
// Solve conjugate-transposed triangular system:
// [ T(ki+1:n,ki+1:n) - T(ki,ki) ]**H * X = scale*work.
// TODO what happens with T(k,k) - lambda is small? Used to have < smin test.
if ( ki < n-1 ) {
n2 = n-ki-1;
queue.push_task( new magma_clatrsd_task(
MagmaUpper, MagmaConjTrans, MagmaNonUnit, MagmaTrue,
n2, T(ki+1,ki+1), ldt, *T(ki,ki),
work(ki+1,iv), work(ki,iv), rwork ));
}
// Copy the vector x or Q*x to VL and normalize.
if ( ! over ) {
// ------------------------------
// no back-transform: copy x to VL and normalize
queue.sync();
n2 = n-ki;
blasf77_ccopy( &n2, work(ki,iv), &ione, VL(ki,is), &ione );
ii = blasf77_icamax( &n2, VL(ki,is), &ione ) + ki - 1;
remax = 1. / MAGMA_C_ABS1( *VL(ii,is) );
blasf77_csscal( &n2, &remax, VL(ki,is), &ione );
for( k=0; k < ki; ++k ) {
*VL(k,is) = c_zero;
}
}
else if ( version == 1 ) {
// ------------------------------
// version 1: back-transform each vector with GEMV, Q*x.
queue.sync();
if ( ki < n-1 ) {
n2 = n-ki-1;
blasf77_cgemv( "n", &n, &n2, &c_one,
VL(0,ki+1), &ldvl,
work(ki+1,iv), &ione,
work(ki, iv), VL(0,ki), &ione );
}
ii = blasf77_icamax( &n, VL(0,ki), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *VL(ii,ki) );
blasf77_csscal( &n, &remax, VL(0,ki), &ione );
}
else if ( version == 2 ) {
// ------------------------------
// version 2: back-transform block of vectors with GEMM
// zero out above vector
// could go from (ki+1)-NV+1 to ki
for( k=0; k < ki; ++k ) {
*work(k,iv) = c_zero;
}
// Columns 1:iv of work are valid vectors.
// When the number of vectors stored reaches nb,
// or if this was last vector, do the GEMM
if ( (iv == nb) || (ki == n-1) ) {
queue.sync();
n2 = n-(ki+1)+iv;
// split gemm into multiple tasks, each doing one block row
for( i=0; i < n; i += NB ) {
magma_int_t ib = min( NB, n-i );
queue.push_task( new cgemm_task(
MagmaNoTrans, MagmaNoTrans, ib, iv, n2, c_one,
VL(i,ki-iv+1), ldvl,
work(ki-iv+1,1), n, c_zero,
work(i,nb+1), n ));
}
queue.sync();
// normalize vectors
for( k=1; k <= iv; ++k ) {
ii = blasf77_icamax( &n, work(0,nb+k), &ione ) - 1;
remax = 1. / MAGMA_C_ABS1( *work(ii,nb+k) );
blasf77_csscal( &n, &remax, work(0,nb+k), &ione );
}
lapackf77_clacpy( "F", &n, &iv, work(0,nb+1), &n, VL(0,ki-iv+1), &ldvl );
iv = 1;
}
else {
iv += 1;
}
} // blocked back-transform
is += 1;
}
}
// close down threads
queue.quit();
magma_set_lapack_numthreads( lapack_nthread );
return *info;
} // End of CTREVC
/**
Purpose
-------
CGEEV computes for an N-by-N complex nonsymmetric matrix A, the
eigenvalues and, optionally, the left and/or right eigenvectors.
The right eigenvector v(j) of A satisfies
A * v(j) = lambda(j) * v(j)
where lambda(j) is its eigenvalue.
The left eigenvector u(j) of A satisfies
u(j)**H * A = lambda(j) * u(j)**H
where u(j)**H denotes the conjugate transpose of u(j).
The computed eigenvectors are normalized to have Euclidean norm
equal to 1 and largest component real.
Arguments
---------
@param[in]
jobvl magma_vec_t
- = MagmaNoVec: left eigenvectors of A are not computed;
- = MagmaVec: left eigenvectors of are computed.
@param[in]
jobvr magma_vec_t
- = MagmaNoVec: right eigenvectors of A are not computed;
- = MagmaVec: right eigenvectors of A are computed.
@param[in]
n INTEGER
The order of the matrix A. N >= 0.
@param[in,out]
A COMPLEX array, dimension (LDA,N)
On entry, the N-by-N matrix A.
On exit, A has been overwritten.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,N).
@param[out]
W COMPLEX array, dimension (N)
W contains the computed eigenvalues.
@param[out]
VL COMPLEX array, dimension (LDVL,N)
If JOBVL = MagmaVec, the left eigenvectors u(j) are stored one
after another in the columns of VL, in the same order
as their eigenvalues.
If JOBVL = MagmaNoVec, VL is not referenced.
u(j) = VL(:,j), the j-th column of VL.
@param[in]
ldvl INTEGER
The leading dimension of the array VL. LDVL >= 1; if
JOBVL = MagmaVec, LDVL >= N.
@param[out]
VR COMPLEX array, dimension (LDVR,N)
If JOBVR = MagmaVec, the right eigenvectors v(j) are stored one
after another in the columns of VR, in the same order
as their eigenvalues.
If JOBVR = MagmaNoVec, VR is not referenced.
v(j) = VR(:,j), the j-th column of VR.
@param[in]
ldvr INTEGER
The leading dimension of the array VR. LDVR >= 1; if
JOBVR = MagmaVec, LDVR >= N.
@param[out]
work (workspace) COMPLEX array, dimension (MAX(1,LWORK))
On exit, if INFO = 0, WORK(1) returns the optimal LWORK.
@param[in]
lwork INTEGER
The dimension of the array WORK. LWORK >= (1+nb)*N.
\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
rwork (workspace) REAL array, dimension (2*N)
@param[out]
info INTEGER
- = 0: successful exit
- < 0: if INFO = -i, the i-th argument had an illegal value.
- > 0: if INFO = i, the QR algorithm failed to compute all the
eigenvalues, and no eigenvectors have been computed;
elements and i+1:N of W contain eigenvalues which have
converged.
@ingroup magma_cgeev_driver
********************************************************************/
extern "C" magma_int_t
magma_cgeev(
magma_vec_t jobvl, magma_vec_t jobvr, magma_int_t n,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *W,
magmaFloatComplex *VL, magma_int_t ldvl,
magmaFloatComplex *VR, magma_int_t ldvr,
magmaFloatComplex *work, magma_int_t lwork,
float *rwork, magma_int_t *info )
{
#define VL(i,j) (VL + (i) + (j)*ldvl)
#define VR(i,j) (VR + (i) + (j)*ldvr)
magma_int_t c_one = 1;
magma_int_t c_zero = 0;
float d__1, d__2;
magmaFloatComplex tmp;
float scl;
float dum[1], eps;
float anrm, cscale, bignum, smlnum;
magma_int_t i, k, ilo, ihi;
magma_int_t ibal, ierr, itau, iwrk, nout, liwrk, nb;
magma_int_t scalea, minwrk, irwork, lquery, wantvl, wantvr, select[1];
magma_side_t side = MagmaRight;
irwork = 0;
*info = 0;
lquery = (lwork == -1);
wantvl = (jobvl == MagmaVec);
wantvr = (jobvr == MagmaVec);
if (! wantvl && jobvl != MagmaNoVec) {
*info = -1;
} else if (! wantvr && jobvr != MagmaNoVec) {
*info = -2;
} else if (n < 0) {
*info = -3;
} else if (lda < max(1,n)) {
*info = -5;
} else if ( (ldvl < 1) || (wantvl && (ldvl < n))) {
*info = -8;
} else if ( (ldvr < 1) || (wantvr && (ldvr < n))) {
*info = -10;
}
/* Compute workspace */
nb = magma_get_cgehrd_nb( n );
if (*info == 0) {
minwrk = (1+nb)*n;
work[0] = MAGMA_C_MAKE( minwrk, 0 );
if (lwork < minwrk && ! lquery) {
*info = -12;
}
}
if (*info != 0) {
magma_xerbla( __func__, -(*info) );
return *info;
}
else if (lquery) {
return *info;
}
/* Quick return if possible */
if (n == 0) {
return *info;
}
#if defined(VERSION3)
magmaFloatComplex *dT;
if (MAGMA_SUCCESS != magma_cmalloc( &dT, nb*n )) {
*info = MAGMA_ERR_DEVICE_ALLOC;
return *info;
}
#endif
/* Get machine constants */
eps = lapackf77_slamch( "P" );
smlnum = lapackf77_slamch( "S" );
bignum = 1. / smlnum;
lapackf77_slabad( &smlnum, &bignum );
smlnum = magma_ssqrt( smlnum ) / eps;
bignum = 1. / smlnum;
/* Scale A if max element outside range [SMLNUM,BIGNUM] */
anrm = lapackf77_clange( "M", &n, &n, A, &lda, dum );
scalea = 0;
if (anrm > 0. && anrm < smlnum) {
scalea = 1;
cscale = smlnum;
} else if (anrm > bignum) {
scalea = 1;
cscale = bignum;
}
if (scalea) {
lapackf77_clascl( "G", &c_zero, &c_zero, &anrm, &cscale, &n, &n, A, &lda, &ierr );
}
/* Balance the matrix
* (CWorkspace: none)
* (RWorkspace: need N)
* - this space is reserved until after gebak */
ibal = 0;
lapackf77_cgebal( "B", &n, A, &lda, &ilo, &ihi, &rwork[ibal], &ierr );
/* Reduce to upper Hessenberg form
* (CWorkspace: need 2*N, prefer N + N*NB)
* (RWorkspace: N)
* - including N reserved for gebal/gebak, unused by cgehrd */
itau = 0;
iwrk = itau + n;
liwrk = lwork - iwrk;
#if defined(VERSION1)
// Version 1 - LAPACK
lapackf77_cgehrd( &n, &ilo, &ihi, A, &lda,
&work[itau], &work[iwrk], &liwrk, &ierr );
#elif defined(VERSION2)
// Version 2 - LAPACK consistent HRD
magma_cgehrd2( n, ilo, ihi, A, lda,
&work[itau], &work[iwrk], liwrk, &ierr );
#elif defined(VERSION3)
// Version 3 - LAPACK consistent MAGMA HRD + T matrices stored,
magma_cgehrd( n, ilo, ihi, A, lda,
&work[itau], &work[iwrk], liwrk, dT, &ierr );
#endif
if (wantvl) {
/* Want left eigenvectors
* Copy Householder vectors to VL */
side = MagmaLeft;
lapackf77_clacpy( MagmaLowerStr, &n, &n, A, &lda, VL, &ldvl );
/* Generate unitary matrix in VL
* (CWorkspace: need 2*N-1, prefer N + (N-1)*NB)
* (RWorkspace: N)
* - including N reserved for gebal/gebak, unused by cunghr */
#if defined(VERSION1) || defined(VERSION2)
// Version 1 & 2 - LAPACK
lapackf77_cunghr( &n, &ilo, &ihi, VL, &ldvl, &work[itau],
&work[iwrk], &liwrk, &ierr );
#elif defined(VERSION3)
// Version 3 - LAPACK consistent MAGMA HRD + T matrices stored
magma_cunghr( n, ilo, ihi, VL, ldvl, &work[itau], dT, nb, &ierr );
#endif
/* Perform QR iteration, accumulating Schur vectors in VL
* (CWorkspace: need 1, prefer HSWORK (see comments) )
* (RWorkspace: N)
* - including N reserved for gebal/gebak, unused by chseqr */
iwrk = itau;
liwrk = lwork - iwrk;
lapackf77_chseqr( "S", "V", &n, &ilo, &ihi, A, &lda, W,
VL, &ldvl, &work[iwrk], &liwrk, info );
if (wantvr) {
/* Want left and right eigenvectors
* Copy Schur vectors to VR */
side = MagmaBothSides;
lapackf77_clacpy( "F", &n, &n, VL, &ldvl, VR, &ldvr );
}
}
else if (wantvr) {
/* Want right eigenvectors
* Copy Householder vectors to VR */
side = MagmaRight;
lapackf77_clacpy( "L", &n, &n, A, &lda, VR, &ldvr );
/* Generate unitary matrix in VR
* (CWorkspace: need 2*N-1, prefer N + (N-1)*NB)
* (RWorkspace: N)
* - including N reserved for gebal/gebak, unused by cunghr */
#if defined(VERSION1) || defined(VERSION2)
// Version 1 & 2 - LAPACK
lapackf77_cunghr( &n, &ilo, &ihi, VR, &ldvr, &work[itau],
&work[iwrk], &liwrk, &ierr );
#elif defined(VERSION3)
// Version 3 - LAPACK consistent MAGMA HRD + T matrices stored
magma_cunghr( n, ilo, ihi, VR, ldvr, &work[itau], dT, nb, &ierr );
#endif
/* Perform QR iteration, accumulating Schur vectors in VR
* (CWorkspace: need 1, prefer HSWORK (see comments) )
* (RWorkspace: N)
* - including N reserved for gebal/gebak, unused by chseqr */
iwrk = itau;
liwrk = lwork - iwrk;
lapackf77_chseqr( "S", "V", &n, &ilo, &ihi, A, &lda, W,
VR, &ldvr, &work[iwrk], &liwrk, info );
}
else {
/* Compute eigenvalues only
* (CWorkspace: need 1, prefer HSWORK (see comments) )
* (RWorkspace: N)
* - including N reserved for gebal/gebak, unused by chseqr */
iwrk = itau;
liwrk = lwork - iwrk;
lapackf77_chseqr( "E", "N", &n, &ilo, &ihi, A, &lda, W,
VR, &ldvr, &work[iwrk], &liwrk, info );
}
/* If INFO > 0 from CHSEQR, then quit */
if (*info > 0) {
goto CLEANUP;
}
if (wantvl || wantvr) {
/* Compute left and/or right eigenvectors
* (CWorkspace: need 2*N)
* (RWorkspace: need 2*N)
* - including N reserved for gebal/gebak, unused by ctrevc */
irwork = ibal + n;
#if TREVC_VERSION == 1
lapackf77_ctrevc( lapack_side_const(side), "B", select, &n, A, &lda, VL, &ldvl,
VR, &ldvr, &n, &nout, &work[iwrk], &rwork[irwork], &ierr );
#elif TREVC_VERSION == 2
liwrk = lwork - iwrk;
lapackf77_ctrevc3( lapack_side_const(side), "B", select, &n, A, &lda, VL, &ldvl,
VR, &ldvr, &n, &nout, &work[iwrk], &liwrk, &rwork[irwork], &ierr );
#elif TREVC_VERSION == 3
magma_ctrevc3( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl,
VR, ldvr, n, &nout, &work[iwrk], liwrk, &rwork[irwork], &ierr );
#elif TREVC_VERSION == 4
magma_ctrevc3_mt( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl,
VR, ldvr, n, &nout, &work[iwrk], liwrk, &rwork[irwork], &ierr );
#elif TREVC_VERSION == 5
magma_ctrevc3_mt_gpu( side, MagmaBacktransVec, select, n, A, lda, VL, ldvl,
VR, ldvr, n, &nout, &work[iwrk], liwrk, &rwork[irwork], &ierr );
#else
#error Unknown TREVC_VERSION
#endif
}
if (wantvl) {
/* Undo balancing of left eigenvectors
* (CWorkspace: none)
* (RWorkspace: need N) */
lapackf77_cgebak( "B", "L", &n, &ilo, &ihi, &rwork[ibal], &n,
VL, &ldvl, &ierr );
/* Normalize left eigenvectors and make largest component real */
for (i = 0; i < n; ++i) {
scl = 1. / cblas_scnrm2( n, VL(0,i), 1 );
cblas_csscal( n, scl, VL(0,i), 1 );
for (k = 0; k < n; ++k) {
/* Computing 2nd power */
d__1 = MAGMA_C_REAL( *VL(k,i) );
d__2 = MAGMA_C_IMAG( *VL(k,i) );
rwork[irwork + k] = d__1*d__1 + d__2*d__2;
}
k = cblas_isamax( n, &rwork[irwork], 1 );
tmp = MAGMA_C_CNJG( *VL(k,i) ) / magma_ssqrt( rwork[irwork + k] );
cblas_cscal( n, CBLAS_SADDR(tmp), VL(0,i), 1 );
*VL(k,i) = MAGMA_C_MAKE( MAGMA_C_REAL( *VL(k,i) ), 0. );
}
}
if (wantvr) {
/* Undo balancing of right eigenvectors
* (CWorkspace: none)
* (RWorkspace: need N) */
lapackf77_cgebak( "B", "R", &n, &ilo, &ihi, &rwork[ibal], &n,
VR, &ldvr, &ierr );
/* Normalize right eigenvectors and make largest component real */
for (i = 0; i < n; ++i) {
scl = 1. / cblas_scnrm2( n, VR(0,i), 1 );
cblas_csscal( n, scl, VR(0,i), 1 );
for (k = 0; k < n; ++k) {
/* Computing 2nd power */
d__1 = MAGMA_C_REAL( *VR(k,i) );
d__2 = MAGMA_C_IMAG( *VR(k,i) );
rwork[irwork + k] = d__1*d__1 + d__2*d__2;
}
k = cblas_isamax( n, &rwork[irwork], 1 );
tmp = MAGMA_C_CNJG( *VR(k,i) ) / magma_ssqrt( rwork[irwork + k] );
cblas_cscal( n, CBLAS_SADDR(tmp), VR(0,i), 1 );
*VR(k,i) = MAGMA_C_MAKE( MAGMA_C_REAL( *VR(k,i) ), 0. );
}
}
CLEANUP:
/* Undo scaling if necessary */
if (scalea) {
// converged eigenvalues, stored in WR[i+1:n] and WI[i+1:n] for i = INFO
magma_int_t nval = n - (*info);
magma_int_t ld = max( nval, 1 );
lapackf77_clascl( "G", &c_zero, &c_zero, &cscale, &anrm, &nval, &c_one, W + (*info), &ld, &ierr );
if (*info > 0) {
// first ilo columns were already upper triangular,
// so the corresponding eigenvalues are also valid.
nval = ilo - 1;
lapackf77_clascl( "G", &c_zero, &c_zero, &cscale, &anrm, &nval, &c_one, W, &n, &ierr );
}
}
#if defined(VERSION3)
magma_free( dT );
#endif
work[0] = MAGMA_C_MAKE( (float) minwrk, 0. ); // TODO use optwrk as in dgeev
return *info;
} /* magma_cgeev */
/* ////////////////////////////////////////////////////////////////////////////
-- Testing csymmetrize
Code is very similar to testing_ctranspose.cpp
*/
int main( int argc, char** argv)
{
TESTING_INIT();
real_Double_t gbytes, gpu_perf, gpu_time, cpu_perf, cpu_time;
float error, work[1];
magmaFloatComplex c_neg_one = MAGMA_C_NEG_ONE;
magmaFloatComplex *h_A, *h_R;
magmaFloatComplex_ptr d_A;
magma_int_t N, size, lda, ldda;
magma_int_t ione = 1;
magma_int_t status = 0;
magma_opts opts;
parse_opts( argc, argv, &opts );
printf("uplo = %s\n", lapack_uplo_const(opts.uplo) );
printf(" N CPU GByte/s (ms) GPU GByte/s (ms) check\n");
printf("=====================================================\n");
for( int itest = 0; itest < opts.ntest; ++itest ) {
for( int iter = 0; iter < opts.niter; ++iter ) {
N = opts.nsize[itest];
lda = N;
ldda = ((N+31)/32)*32;
size = lda*N;
// load strictly lower triangle, save strictly upper triangle
gbytes = sizeof(magmaFloatComplex) * 1.*N*(N-1) / 1e9;
TESTING_MALLOC_CPU( h_A, magmaFloatComplex, size );
TESTING_MALLOC_CPU( h_R, magmaFloatComplex, size );
TESTING_MALLOC_DEV( d_A, magmaFloatComplex, ldda*N );
/* Initialize the matrix */
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < N; ++i ) {
h_A[i + j*lda] = MAGMA_C_MAKE( i + j/10000., j );
}
}
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( N, N, h_A, lda, d_A, ldda );
gpu_time = magma_sync_wtime( 0 );
//magmablas_csymmetrize( opts.uplo, N-2, d_A+1+ldda, ldda ); // inset by 1 row & col
magmablas_csymmetrize( opts.uplo, N, d_A, ldda );
gpu_time = magma_sync_wtime( 0 ) - gpu_time;
gpu_perf = gbytes / gpu_time;
/* =====================================================================
Performs operation using naive in-place algorithm
(LAPACK doesn't implement symmetrize)
=================================================================== */
cpu_time = magma_wtime();
//for( int j = 1; j < N-1; ++j ) { // inset by 1 row & col
// for( int i = 1; i < j; ++i ) {
for( int j = 0; j < N; ++j ) {
for( int i = 0; i < j; ++i ) {
if ( opts.uplo == MagmaLower ) {
h_A[i + j*lda] = MAGMA_C_CNJG( h_A[j + i*lda] );
}
else {
h_A[j + i*lda] = MAGMA_C_CNJG( h_A[i + j*lda] );
}
}
}
cpu_time = magma_wtime() - cpu_time;
cpu_perf = gbytes / cpu_time;
/* =====================================================================
Check the result
=================================================================== */
magma_cgetmatrix( N, N, d_A, ldda, h_R, lda );
blasf77_caxpy(&size, &c_neg_one, h_A, &ione, h_R, &ione);
error = lapackf77_clange("f", &N, &N, h_R, &lda, work);
printf("%5d %7.2f (%7.2f) %7.2f (%7.2f) %s\n",
(int) N, cpu_perf, cpu_time*1000., gpu_perf, gpu_time*1000.,
(error == 0. ? "ok" : "failed") );
status += ! (error == 0.);
TESTING_FREE_CPU( h_A );
TESTING_FREE_CPU( h_R );
TESTING_FREE_DEV( d_A );
fflush( stdout );
}
if ( opts.niter > 1 ) {
printf( "\n" );
}
}
TESTING_FINALIZE();
return status;
}
void magmablas_cher2k_mgpu_spec(
char uplo, char trans, magma_int_t n, magma_int_t k,
magmaFloatComplex alpha, magmaFloatComplex *dA[], magma_int_t lda, magma_int_t aoffset,
magmaFloatComplex *dB[], magma_int_t ldb, magma_int_t boffset,
float beta, magmaFloatComplex *dC[], magma_int_t ldc, magma_int_t coffset,
magma_int_t ngpu, magma_int_t nb, magma_queue_t streams[][20], magma_int_t nstream )
{
#define dA(dev, i, j) (dA[dev] + (i) + (j)*lda + (aoffset) )
#define dB(dev, i, j) (dB[dev] + (i) + (j)*ldb + (boffset) )
#define dC(dev, i, j) (dC[dev] + (i) + (j)*ldc)
/* Check arguments */
magma_int_t info = 0;
if ( ! (uplo == 'l' || uplo == 'L')) {
info = -1; // 'u' not yet handled
} else if ( ! (trans == 'n' || trans == 'N')) {
info = -2; // 'c' not yet handled
} else if ( n < 0 ) {
info = -3;
} else if ( k < 0 ) {
info = -4;
} else if ( ((trans == 'n' || trans == 'N') && lda < max(1,n)) ||
((trans == 'c' || trans == 'C') && lda < max(1,k)) ) {
info = -7;
} else if ( aoffset < 0 || aoffset > lda ) {
info = -8;
} else if ( ((trans == 'n' || trans == 'N') && ldb < max(1,n)) ||
((trans == 'c' || trans == 'C') && ldb < max(1,k)) ) {
info = -10;
} else if ( boffset < 0 || boffset > ldb ) {
info = -11;
} else if ( ldc < max(1,n) ) {
info = -13;
} else if ( coffset < 0 || coffset > ldc ) {
info = -14;
} else if ( ngpu <= 0 ) {
info = -15;
} else if ( nb <= 0 ) {
info = -16;
} else if ( nstream <= 0 ) {
info = -18;
}
if ( info != 0 ) {
magma_xerbla( __func__, -(info) );
return;
}
const magmaFloatComplex c_one = MAGMA_C_ONE;
magmaFloatComplex cbeta = MAGMA_C_MAKE( beta, 0. );
magma_int_t ib, ioff, iblock, idev, di, s;
magma_device_t cdev;
magma_queue_t cqueue;
magma_getdevice( &cdev );
magmablasGetKernelStream( &cqueue );
// loop over all blocks
// Faster to have two loops: first loop does C_hat = alpha*A*B' + beta*C
// blockoffset is offset within first block; for subsequent blocks it is 0
magma_int_t blockoffset = coffset % nb;
for( magma_int_t i = 0; i < n; i += ib ) {
ib = min( nb-blockoffset, n-i ); // block size
ioff = i + coffset; // global index in parent matrix
iblock = (ioff / nb) / ngpu; // local block id
idev = (ioff / nb) % ngpu; // device with this block
di = iblock*nb + blockoffset; // local index in parent matrix
magma_setdevice( idev );
s = iblock % nstream;
magmablasSetKernelStream( streams[ idev ][ s ] );
// C[i:n,i] = alpha * A[i:n,0] * B[i,0]' + beta*C[i:n,i]
//printf( "cgemm n=%4d, ib=%4d, k=%4d, i=%4d\n", n-i, ib, k, i );
magma_cgemm( MagmaNoTrans, MagmaConjTrans, n, ib, k,
alpha, dA(idev,0,0), lda,
dB(idev,i,0), ldb,
cbeta, dC(idev,coffset,di), ldc );
blockoffset = 0;
}
// second loop does C = conjf(alpha)*B*A' + C_hat
alpha = MAGMA_C_CNJG( alpha );
blockoffset = coffset % nb;
for( magma_int_t i = 0; i < n; i += ib ) {
ib = min( nb-blockoffset, n-i ); // block size
ioff = i + coffset; // global index in parent matrix
iblock = (ioff / nb) / ngpu; // local block id
idev = (ioff / nb) % ngpu; // device with this block
di = iblock*nb + blockoffset; // local index in parent matrix
magma_setdevice( idev );
s = iblock % nstream;
magmablasSetKernelStream( streams[ idev ][ s ] );
// C[i:n,i] += conjf(alpha) * B[i:n,0] * A[i,0]'
//printf( "cgemm n=%4d, ib=%4d, k=%4d, i=%4d\n", n-i, ib, k, i );
magma_cgemm( MagmaNoTrans, MagmaConjTrans, n, ib, k,
alpha, dB(idev,0,0), ldb,
dA(idev,i,0), lda,
c_one, dC(idev,coffset,di), ldc );
blockoffset = 0;
}
magma_setdevice( cdev );
magmablasSetKernelStream( cqueue );
}
int main( int argc, char** argv)
{
real_Double_t gflops, gpu_perf, cpu_perf, gpu_time, cpu_time;
magmaFloatComplex *hA, *hR;
magmaFloatComplex_ptr dA;
magma_int_t N = 0, n2, lda, ldda;
magma_int_t size[10] =
{ 1024, 2048, 3072, 4032, 5184, 6048, 7200, 8064, 8928, 10560 };
magma_int_t i, info;
magmaFloatComplex mz_one = MAGMA_C_NEG_ONE;
magma_int_t ione = 1;
magma_int_t ISEED[4] = {0,0,0,1};
float work[1], matnorm, diffnorm;
if (argc != 1){
for(i = 1; i<argc; i++){
if (strcmp("-N", argv[i])==0)
N = atoi(argv[++i]);
}
if (N>0) size[0] = size[9] = N;
else exit(1);
}
else {
printf("\nUsage: \n");
printf(" testing_cpotrf_gpu -N %d\n\n", 1024);
}
/* Initialize */
magma_queue_t queue;
magma_device_t device;
int num = 0;
magma_err_t err;
magma_init();
err = magma_get_devices( &device, 1, &num );
if ( err != 0 || num < 1 ) {
fprintf( stderr, "magma_get_devices failed: %d\n", err );
exit(-1);
}
err = magma_queue_create( device, &queue );
if ( err != 0 ) {
fprintf( stderr, "magma_queue_create failed: %d\n", err );
exit(-1);
}
/* Allocate memory for the largest matrix */
N = size[9];
n2 = N * N;
ldda = ((N+31)/32) * 32;
TESTING_MALLOC( hA, magmaFloatComplex, n2 );
TESTING_MALLOC_HOST( hR, magmaFloatComplex, n2 );
TESTING_MALLOC_DEV( dA, magmaFloatComplex, ldda*N );
printf("\n\n");
printf(" N CPU GFlop/s (sec) GPU GFlop/s (sec) ||R_magma-R_lapack||_F / ||R_lapack||_F\n");
printf("========================================================================================\n");
for(i=0; i<10; i++){
N = size[i];
lda = N;
n2 = lda*N;
ldda = ((N+31)/32)*32;
gflops = FLOPS( (float)N ) * 1e-9;
/* Initialize the matrix */
lapackf77_clarnv( &ione, ISEED, &n2, hA );
/* Symmetrize and increase the diagonal */
for( int i = 0; i < N; ++i ) {
MAGMA_C_SET2REAL( hA(i,i), MAGMA_C_REAL(hA(i,i)) + N );
for( int j = 0; j < i; ++j ) {
hA(i, j) = MAGMA_C_CNJG( hA(j,i) );
}
}
lapackf77_clacpy( MagmaFullStr, &N, &N, hA, &lda, hR, &lda );
/* Warm up to measure the performance */
magma_csetmatrix( N, N, hA, 0, lda, dA, 0, ldda, queue );
magma_cpotrf_gpu( MagmaUpper, N, dA, 0, ldda, &info, queue );
/* ====================================================================
Performs operation using MAGMA
=================================================================== */
magma_csetmatrix( N, N, hA, 0, lda, dA, 0, ldda, queue );
gpu_time = get_time();
magma_cpotrf_gpu( MagmaUpper, N, dA, 0, ldda, &info, queue );
gpu_time = get_time() - gpu_time;
if (info != 0)
printf( "magma_cpotrf had error %d.\n", info );
gpu_perf = gflops / gpu_time;
/* =====================================================================
Performs operation using LAPACK
=================================================================== */
cpu_time = get_time();
lapackf77_cpotrf( MagmaUpperStr, &N, hA, &lda, &info );
cpu_time = get_time() - cpu_time;
if (info != 0)
printf( "lapackf77_cpotrf had error %d.\n", info );
cpu_perf = gflops / cpu_time;
/* =====================================================================
Check the result compared to LAPACK
|R_magma - R_lapack| / |R_lapack|
=================================================================== */
magma_cgetmatrix( N, N, dA, 0, ldda, hR, 0, lda, queue );
matnorm = lapackf77_clange("f", &N, &N, hA, &lda, work);
blasf77_caxpy(&n2, &mz_one, hA, &ione, hR, &ione);
diffnorm = lapackf77_clange("f", &N, &N, hR, &lda, work);
printf( "%5d %6.2f (%6.2f) %6.2f (%6.2f) %e\n",
N, cpu_perf, cpu_time, gpu_perf, gpu_time, diffnorm / matnorm );
if (argc != 1)
break;
}
/* clean up */
TESTING_FREE( hA );
TESTING_FREE_HOST( hR );
TESTING_FREE_DEV( dA );
magma_queue_destroy( queue );
magma_finalize();
}
magma_int_t magma_clatrsd(
magma_uplo_t uplo, magma_trans_t trans, magma_diag_t diag, magma_bool_t normin,
magma_int_t n, const magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex lambda,
magmaFloatComplex *x,
float *scale, float *cnorm, magma_int_t *info)
{
#define A(i,j) (A + (i) + (j)*lda)
/* constants */
const magma_int_t ione = 1;
const float d_half = 0.5;
const magmaFloatComplex c_zero = MAGMA_C_ZERO;
const magmaFloatComplex c_one = MAGMA_C_ONE;
/* System generated locals */
magma_int_t len;
magmaFloatComplex ztmp;
/* Local variables */
magma_int_t i, j;
float xj, rec, tjj;
magma_int_t jinc;
float xbnd;
magma_int_t imax;
float tmax;
magmaFloatComplex tjjs;
float xmax, grow;
float tscal;
magmaFloatComplex uscal;
magma_int_t jlast;
magmaFloatComplex csumj;
float bignum;
magma_int_t jfirst;
float smlnum;
/* Function Body */
*info = 0;
magma_int_t upper = (uplo == MagmaUpper);
magma_int_t notran = (trans == MagmaNoTrans);
magma_int_t nounit = (diag == MagmaNonUnit);
/* Test the input parameters. */
if ( ! upper && uplo != MagmaLower ) {
*info = -1;
}
else if (! notran &&
trans != MagmaTrans &&
trans != MagmaConjTrans) {
*info = -2;
}
else if ( ! nounit && diag != MagmaUnit ) {
*info = -3;
}
else if ( ! (normin == MagmaTrue) &&
! (normin == MagmaFalse) ) {
*info = -4;
}
else if ( n < 0 ) {
*info = -5;
}
else if ( lda < max(1,n) ) {
*info = -7;
}
if ( *info != 0 ) {
magma_xerbla( __func__, -(*info) );
return *info;
}
/* Quick return if possible */
if ( n == 0 ) {
return *info;
}
/* Determine machine dependent parameters to control overflow. */
smlnum = lapackf77_slamch( "Safe minimum" );
bignum = 1. / smlnum;
lapackf77_slabad( &smlnum, &bignum );
smlnum /= lapackf77_slamch( "Precision" );
bignum = 1. / smlnum;
*scale = 1.;
if ( normin == MagmaFalse ) {
/* Compute the 1-norm of each column, not including the diagonal. */
if ( upper ) {
/* A is upper triangular. */
cnorm[0] = 0.;
for( j = 1; j < n; ++j ) {
cnorm[j] = cblas_scasum( j, A(0,j), ione );
}
}
else {
/* A is lower triangular. */
for( j = 0; j < n-1; ++j ) {
cnorm[j] = cblas_scasum( n-(j+1), A(j+1,j), ione );
}
cnorm[n-1] = 0.;
}
}
/* Scale the column norms by TSCAL if the maximum element in CNORM is */
/* greater than BIGNUM/2. */
imax = blasf77_isamax( &n, &cnorm[0], &ione ) - 1;
tmax = cnorm[imax];
if ( tmax <= bignum * 0.5 ) {
tscal = 1.;
}
else {
tscal = 0.5 / (smlnum * tmax);
blasf77_sscal( &n, &tscal, &cnorm[0], &ione );
}
/* ================================================================= */
/* Compute a bound on the computed solution vector to see if the */
/* Level 2 BLAS routine CTRSV can be used. */
xmax = 0.;
for( j = 0; j < n; ++j ) {
xmax = max( xmax, 0.5*MAGMA_C_ABS1( x[j] ));
}
xbnd = xmax;
if ( notran ) {
/* ---------------------------------------- */
/* Compute the growth in A * x = b. */
if ( upper ) {
jfirst = n-1;
jlast = 0;
jinc = -1;
}
else {
jfirst = 0;
jlast = n;
jinc = 1;
}
if ( tscal != 1. ) {
grow = 0.;
goto L60;
}
/* A is non-unit triangular. */
/* Compute GROW = 1/G(j) and XBND = 1/M(j). */
/* Initially, G(0) = max{x(i), i=1,...,n}. */
grow = 0.5 / max( xbnd, smlnum );
xbnd = grow;
for( j = jfirst; (jinc < 0 ? j >= jlast : j < jlast); j += jinc ) {
/* Exit the loop if the growth factor is too small. */
if ( grow <= smlnum ) {
goto L60;
}
if ( nounit ) {
tjjs = *A(j,j) - lambda;
}
else {
tjjs = c_one - lambda;
}
tjj = MAGMA_C_ABS1( tjjs );
if ( tjj >= smlnum ) {
/* M(j) = G(j-1) / abs(A(j,j)) */
xbnd = min( xbnd, min(1.,tjj)*grow );
}
else {
/* M(j) could overflow, set XBND to 0. */
xbnd = 0.;
}
if ( tjj + cnorm[j] >= smlnum ) {
/* G(j) = G(j-1)*( 1 + CNORM(j) / abs(A(j,j)) ) */
grow *= (tjj / (tjj + cnorm[j]));
}
else {
/* G(j) could overflow, set GROW to 0. */
grow = 0.;
}
}
grow = xbnd;
L60:
;
}
else {
/* ---------------------------------------- */
/* Compute the growth in A**T * x = b or A**H * x = b. */
if ( upper ) {
jfirst = 0;
jlast = n;
jinc = 1;
}
else {
jfirst = n-1;
jlast = 0;
jinc = -1;
}
if ( tscal != 1. ) {
grow = 0.;
goto L90;
}
/* A is non-unit triangular. */
/* Compute GROW = 1/G(j) and XBND = 1/M(j). */
/* Initially, M(0) = max{x(i), i=1,...,n}. */
grow = 0.5 / max( xbnd, smlnum );
xbnd = grow;
for( j = jfirst; (jinc < 0 ? j >= jlast : j < jlast); j += jinc ) {
/* Exit the loop if the growth factor is too small. */
if ( grow <= smlnum ) {
goto L90;
}
/* G(j) = max( G(j-1), M(j-1)*( 1 + CNORM(j) ) ) */
xj = 1. + cnorm[j];
grow = min( grow, xbnd / xj );
if ( nounit ) {
tjjs = *A(j,j) - lambda;
}
else {
tjjs = c_one - lambda;
}
tjj = MAGMA_C_ABS1( tjjs );
if ( tjj >= smlnum ) {
/* M(j) = M(j-1)*( 1 + CNORM(j) ) / abs(A(j,j)) */
if ( xj > tjj ) {
xbnd *= (tjj / xj);
}
}
else {
/* M(j) could overflow, set XBND to 0. */
xbnd = 0.;
}
}
grow = min( grow, xbnd );
L90:
;
}
/* ================================================================= */
/* Due to modified diagonal, we can't use regular BLAS ctrsv. */
/* Use a Level 1 BLAS solve, scaling intermediate results. */
if ( xmax > bignum * 0.5 ) {
/* Scale X so that its components are less than or equal to */
/* BIGNUM in absolute value. */
*scale = (bignum * 0.5) / xmax;
blasf77_csscal( &n, scale, &x[0], &ione );
xmax = bignum;
}
else {
xmax *= 2.;
}
if ( notran ) {
/* ---------------------------------------- */
/* Solve A * x = b */
for( j = jfirst; (jinc < 0 ? j >= jlast : j < jlast); j += jinc ) {
/* Compute x(j) = b(j) / A(j,j), scaling x if necessary. */
xj = MAGMA_C_ABS1( x[j] );
if ( nounit ) {
tjjs = (*A(j,j) - lambda ) * tscal;
}
else {
tjjs = (c_one - lambda) * tscal;
if ( tscal == 1. ) {
goto L110;
}
}
tjj = MAGMA_C_ABS1( tjjs );
if ( tjj > smlnum ) {
/* abs(A(j,j)) > SMLNUM: */
if ( tjj < 1. ) {
if ( xj > tjj * bignum ) {
/* Scale x by 1/b(j). */
rec = 1. / xj;
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
}
x[j] = x[j] / tjjs;
xj = MAGMA_C_ABS1( x[j] );
}
else if ( tjj > 0. ) {
/* 0 < abs(A(j,j)) <= SMLNUM: */
if ( xj > tjj * bignum ) {
/* Scale x by (1/abs(x(j)))*abs(A(j,j))*BIGNUM */
/* to avoid overflow when dividing by A(j,j). */
rec = (tjj * bignum) / xj;
if ( cnorm[j] > 1. ) {
/* Scale by 1/CNORM(j) to avoid overflow when */
/* multiplying x(j) times column j. */
rec /= cnorm[j];
}
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
x[j] = x[j] / tjjs;
xj = MAGMA_C_ABS1( x[j] );
}
else {
/* A(j,j) = 0: Set x(1:n) = 0, x(j) = 1, and */
/* scale = 0, and compute a solution to A*x = 0. */
for( i = 0; i < n; ++i ) {
x[i] = c_zero;
}
x[j] = c_one;
xj = 1.;
*scale = 0.;
xmax = 0.;
}
L110:
/* Scale x if necessary to avoid overflow when adding a */
/* multiple of column j of A. */
if ( xj > 1. ) {
rec = 1. / xj;
if ( cnorm[j] > (bignum - xmax) * rec ) {
/* Scale x by 1/(2*abs(x(j))). */
rec *= 0.5;
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
}
}
else if ( xj * cnorm[j] > bignum - xmax ) {
/* Scale x by 1/2. */
blasf77_csscal( &n, &d_half, &x[0], &ione );
*scale *= 0.5;
}
if ( upper ) {
if ( j > 0 ) {
/* Compute the update */
/* x(1:j-1) := x(1:j-1) - x(j) * A(1:j-1,j) */
len = j;
ztmp = -tscal * x[j];
blasf77_caxpy( &len, &ztmp, A(0,j), &ione, &x[0], &ione );
i = blasf77_icamax( &len, &x[0], &ione ) - 1;
xmax = MAGMA_C_ABS1( x[i] );
}
}
else {
if ( j < n-1 ) {
/* Compute the update */
/* x(j+1:n) := x(j+1:n) - x(j) * A(j+1:n,j) */
len = n - (j+1);
ztmp = -tscal * x[j];
blasf77_caxpy( &len, &ztmp, A(j+1,j), &ione, &x[j + 1], &ione );
i = j + blasf77_icamax( &len, &x[j + 1], &ione );
xmax = MAGMA_C_ABS1( x[i] );
}
}
}
}
else if ( trans == MagmaTrans ) {
/* ---------------------------------------- */
/* Solve A**T * x = b */
for( j = jfirst; (jinc < 0 ? j >= jlast : j < jlast); j += jinc ) {
/* Compute x(j) = b(j) - sum A(k,j)*x(k). */
/* k<>j */
xj = MAGMA_C_ABS1( x[j] );
uscal = MAGMA_C_MAKE( tscal, 0. );
rec = 1. / max( xmax, 1. );
if ( cnorm[j] > (bignum - xj) * rec ) {
/* If x(j) could overflow, scale x by 1/(2*XMAX). */
rec *= 0.5;
if ( nounit ) {
tjjs = (*A(j,j) - lambda) * tscal;
}
else {
tjjs = (c_one - lambda) * tscal;
}
tjj = MAGMA_C_ABS1( tjjs );
if ( tjj > 1. ) {
/* Divide by A(j,j) when scaling x if A(j,j) > 1. */
rec = min( 1., rec * tjj );
uscal = uscal / tjjs;
}
if ( rec < 1. ) {
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
}
csumj = c_zero;
if ( uscal == c_one ) {
/* If the scaling needed for A in the dot product is 1, */
/* call ZDOTU to perform the dot product. */
if ( upper ) {
cblas_cdotu_sub( j, A(0,j), ione, &x[0], ione, &csumj );
}
else if ( j < n-1 ) {
cblas_cdotu_sub( n-(j+1), A(j+1,j), ione, &x[j + 1], ione, &csumj );
}
}
else {
/* Otherwise, use in-line code for the dot product. */
if ( upper ) {
for( i = 0; i < j; ++i ) {
csumj += (*A(i,j) * uscal) * x[i];
}
}
else if ( j < n-1 ) {
for( i = j+1; i < n; ++i ) {
csumj += (*A(i,j) * uscal) * x[i];
}
}
}
if ( uscal == MAGMA_C_MAKE( tscal, 0. )) {
/* Compute x(j) := ( x(j) - CSUMJ ) / A(j,j) if 1/A(j,j) */
/* was not used to scale the dotproduct. */
x[j] -= csumj;
xj = MAGMA_C_ABS1( x[j] );
if ( nounit ) {
tjjs = (*A(j,j) - lambda) * tscal;
}
else {
tjjs = (c_one - lambda) * tscal;
if ( tscal == 1. ) {
goto L160;
}
}
/* Compute x(j) = x(j) / A(j,j), scaling if necessary. */
tjj = MAGMA_C_ABS1( tjjs );
if ( tjj > smlnum ) {
/* abs(A(j,j)) > SMLNUM: */
if ( tjj < 1. ) {
if ( xj > tjj * bignum ) {
/* Scale X by 1/abs(x(j)). */
rec = 1. / xj;
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
}
x[j] = x[j] / tjjs;
}
else if ( tjj > 0. ) {
/* 0 < abs(A(j,j)) <= SMLNUM: */
if ( xj > tjj * bignum ) {
/* Scale x by (1/abs(x(j)))*abs(A(j,j))*BIGNUM. */
rec = (tjj * bignum) / xj;
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
x[j] = x[j] / tjjs;
}
else {
/* A(j,j) = 0: Set x(1:n) = 0, x(j) = 1, and */
/* scale = 0 and compute a solution to A**T *x = 0. */
for( i = 0; i < n; ++i ) {
x[i] = c_zero;
}
x[j] = c_one;
*scale = 0.;
xmax = 0.;
}
L160:
;
}
else {
/* Compute x(j) := x(j) / A(j,j) - CSUMJ if the dot */
/* product has already been divided by 1/A(j,j). */
x[j] = (x[j] / tjjs) - csumj;
}
xmax = max( xmax, MAGMA_C_ABS1( x[j] ));
}
}
else {
/* ---------------------------------------- */
/* Solve A**H * x = b */
for( j = jfirst; (jinc < 0 ? j >= jlast : j < jlast); j += jinc ) {
/* Compute x(j) = b(j) - sum A(k,j)*x(k). */
/* k<>j */
xj = MAGMA_C_ABS1( x[j] );
uscal = MAGMA_C_MAKE( tscal, 0. );
rec = 1. / max(xmax, 1.);
if ( cnorm[j] > (bignum - xj) * rec ) {
/* If x(j) could overflow, scale x by 1/(2*XMAX). */
rec *= 0.5;
if ( nounit ) {
tjjs = MAGMA_C_CNJG( *A(j,j) - lambda ) * tscal;
}
else {
tjjs = (c_one - lambda) * tscal;
}
tjj = MAGMA_C_ABS1( tjjs );
if ( tjj > 1. ) {
/* Divide by A(j,j) when scaling x if A(j,j) > 1. */
rec = min( 1., rec * tjj );
uscal = uscal / tjjs;
}
if ( rec < 1. ) {
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
}
csumj = c_zero;
if ( uscal == c_one ) {
/* If the scaling needed for A in the dot product is 1, */
/* call CDOTC to perform the dot product. */
if ( upper ) {
cblas_cdotc_sub( j, A(0,j), ione, &x[0], ione, &csumj );
}
else if ( j < n-1 ) {
cblas_cdotc_sub( n-(j+1), A(j+1,j), ione, &x[j + 1], ione, &csumj );
}
}
else {
/* Otherwise, use in-line code for the dot product. */
if ( upper ) {
for( i = 0; i < j; ++i ) {
csumj += (MAGMA_C_CNJG( *A(i,j) ) * uscal) * x[i];
}
}
else if ( j < n-1 ) {
for( i = j + 1; i < n; ++i ) {
csumj += (MAGMA_C_CNJG( *A(i,j) ) * uscal) * x[i];
}
}
}
if ( uscal == tscal ) {
/* Compute x(j) := ( x(j) - CSUMJ ) / A(j,j) if 1/A(j,j) */
/* was not used to scale the dotproduct. */
x[j] -= csumj;
xj = MAGMA_C_ABS1( x[j] );
if ( nounit ) {
tjjs = MAGMA_C_CNJG( *A(j,j) - lambda ) * tscal;
}
else {
tjjs = (c_one - lambda) * tscal;
if ( tscal == 1. ) {
goto L210;
}
}
/* Compute x(j) = x(j) / A(j,j), scaling if necessary. */
tjj = MAGMA_C_ABS1( tjjs );
if ( tjj > smlnum ) {
/* abs(A(j,j)) > SMLNUM: */
if ( tjj < 1. ) {
if ( xj > tjj * bignum ) {
/* Scale X by 1/abs(x(j)). */
rec = 1. / xj;
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
}
x[j] = x[j] / tjjs;
}
else if ( tjj > 0. ) {
/* 0 < abs(A(j,j)) <= SMLNUM: */
if ( xj > tjj * bignum ) {
/* Scale x by (1/abs(x(j)))*abs(A(j,j))*BIGNUM. */
rec = (tjj * bignum) / xj;
blasf77_csscal( &n, &rec, &x[0], &ione );
*scale *= rec;
xmax *= rec;
}
x[j] = x[j] / tjjs;
}
else {
/* A(j,j) = 0: Set x(1:n) = 0, x(j) = 1, and */
/* scale = 0 and compute a solution to A**H *x = 0. */
for( i = 0; i < n; ++i ) {
x[i] = c_zero;
}
x[j] = c_one;
*scale = 0.;
xmax = 0.;
}
L210:
;
}
else {
/* Compute x(j) := x(j) / A(j,j) - CSUMJ if the dot */
/* product has already been divided by 1/A(j,j). */
x[j] = (x[j] / tjjs) - csumj;
}
xmax = max( xmax, MAGMA_C_ABS1( x[j] ));
}
}
*scale /= tscal;
/* Scale the column norms by 1/TSCAL for return. */
if ( tscal != 1. ) {
float d = 1. / tscal;
blasf77_sscal( &n, &d, &cnorm[0], &ione );
}
return *info;
} /* end clatrsd */
/**
Purpose
-------
CLAQPS computes a step of QR factorization with column pivoting
of a complex M-by-N matrix A by using Blas-3. It tries to factorize
NB columns from A starting from the row OFFSET+1, and updates all
of the matrix with Blas-3 xGEMM.
In some cases, due to catastrophic cancellations, it cannot
factorize NB columns. Hence, the actual number of factorized
columns is returned in KB.
Block A(1:OFFSET,1:N) is accordingly pivoted, but not factorized.
Arguments
---------
@param[in]
m INTEGER
The number of rows of the matrix A. M >= 0.
@param[in]
n INTEGER
The number of columns of the matrix A. N >= 0
@param[in]
offset INTEGER
The number of rows of A that have been factorized in
previous steps.
@param[in]
nb INTEGER
The number of columns to factorize.
@param[out]
kb INTEGER
The number of columns actually factorized.
@param[in,out]
A COMPLEX array, dimension (LDA,N)
On entry, the M-by-N matrix A.
On exit, block A(OFFSET+1:M,1:KB) is the triangular
factor obtained and block A(1:OFFSET,1:N) has been
accordingly pivoted, but no factorized.
The rest of the matrix, block A(OFFSET+1:M,KB+1:N) has
been updated.
@param[in]
lda INTEGER
The leading dimension of the array A. LDA >= max(1,M).
@param[in,out]
jpvt INTEGER array, dimension (N)
JPVT(I) = K <==> Column K of the full matrix A has been
permuted into position I in AP.
@param[out]
tau COMPLEX array, dimension (KB)
The scalar factors of the elementary reflectors.
@param[in,out]
vn1 REAL array, dimension (N)
The vector with the partial column norms.
@param[in,out]
vn2 REAL array, dimension (N)
The vector with the exact column norms.
@param[in,out]
auxv COMPLEX array, dimension (NB)
Auxiliar vector.
@param[in,out]
F COMPLEX array, dimension (LDF,NB)
Matrix F' = L*Y'*A.
@param[in]
ldf INTEGER
The leading dimension of the array F. LDF >= max(1,N).
@ingroup magma_cgeqp3_aux
********************************************************************/
extern "C" magma_int_t
magma_claqps(magma_int_t m, magma_int_t n, magma_int_t offset,
magma_int_t nb, magma_int_t *kb,
magmaFloatComplex *A, magma_int_t lda,
magmaFloatComplex *dA, magma_int_t ldda,
magma_int_t *jpvt, magmaFloatComplex *tau, float *vn1, float *vn2,
magmaFloatComplex *auxv,
magmaFloatComplex *F, magma_int_t ldf,
magmaFloatComplex *dF, magma_int_t lddf)
{
#define A(i, j) (A + (i) + (j)*(lda ))
#define dA(i, j) (dA + (i) + (j)*(ldda))
#define F(i, j) (F + (i) + (j)*(ldf ))
#define dF(i, j) (dF + (i) + (j)*(lddf))
magmaFloatComplex c_zero = MAGMA_C_MAKE( 0.,0.);
magmaFloatComplex c_one = MAGMA_C_MAKE( 1.,0.);
magmaFloatComplex c_neg_one = MAGMA_C_MAKE(-1.,0.);
magma_int_t ione = 1;
magma_int_t i__1, i__2;
float d__1;
magmaFloatComplex z__1;
magma_int_t j, k, rk;
magmaFloatComplex Akk;
magma_int_t pvt;
float temp, temp2, tol3z;
magma_int_t itemp;
magma_int_t lsticc;
magma_int_t lastrk;
lastrk = min( m, n + offset );
tol3z = magma_ssqrt( lapackf77_slamch("Epsilon"));
magma_queue_t stream;
magma_queue_create( &stream );
lsticc = 0;
k = 0;
while( k < nb && lsticc == 0 ) {
rk = offset + k;
/* Determine ith pivot column and swap if necessary */
// subtract 1 from Fortran isamax; pvt, k are 0-based.
i__1 = n-k;
pvt = k + blasf77_isamax( &i__1, &vn1[k], &ione ) - 1;
if (pvt != k) {
if (pvt >= nb) {
/* 1. Start copy from GPU */
magma_cgetmatrix_async( m - offset - nb, 1,
dA(offset + nb, pvt), ldda,
A (offset + nb, pvt), lda, stream );
}
/* F gets swapped so F must be sent at the end to GPU */
i__1 = k;
blasf77_cswap( &i__1, F(pvt,0), &ldf, F(k,0), &ldf );
itemp = jpvt[pvt];
jpvt[pvt] = jpvt[k];
jpvt[k] = itemp;
vn1[pvt] = vn1[k];
vn2[pvt] = vn2[k];
if (pvt < nb) {
/* no need of transfer if pivot is within the panel */
blasf77_cswap( &m, A(0, pvt), &ione, A(0, k), &ione );
}
else {
/* 1. Finish copy from GPU */
magma_queue_sync( stream );
/* 2. Swap as usual on CPU */
blasf77_cswap(&m, A(0, pvt), &ione, A(0, k), &ione);
/* 3. Restore the GPU */
magma_csetmatrix_async( m - offset - nb, 1,
A (offset + nb, pvt), lda,
dA(offset + nb, pvt), ldda, stream);
}
}
/* Apply previous Householder reflectors to column K:
A(RK:M,K) := A(RK:M,K) - A(RK:M,1:K-1)*F(K,1:K-1)'.
Optimization: multiply with beta=0; wait for vector and subtract */
if (k > 0) {
#if defined(PRECISION_c) || defined(PRECISION_z)
for (j = 0; j < k; ++j) {
*F(k,j) = MAGMA_C_CNJG( *F(k,j) );
}
#endif
i__1 = m - rk;
i__2 = k;
blasf77_cgemv( MagmaNoTransStr, &i__1, &i__2,
&c_neg_one, A(rk, 0), &lda,
F(k, 0), &ldf,
&c_one, A(rk, k), &ione );
#if defined(PRECISION_c) || defined(PRECISION_z)
for (j = 0; j < k; ++j) {
*F(k,j) = MAGMA_C_CNJG( *F(k,j) );
}
#endif
}
/* Generate elementary reflector H(k). */
if (rk < m-1) {
i__1 = m - rk;
lapackf77_clarfg( &i__1, A(rk, k), A(rk + 1, k), &ione, &tau[k] );
} else {
lapackf77_clarfg( &ione, A(rk, k), A(rk, k), &ione, &tau[k] );
}
Akk = *A(rk, k);
*A(rk, k) = c_one;
/* Compute Kth column of F:
Compute F(K+1:N,K) := tau(K)*A(RK:M,K+1:N)'*A(RK:M,K) on the GPU */
if (k < n-1) {
i__1 = m - rk;
i__2 = n - k - 1;
/* Send the vector to the GPU */
magma_csetmatrix( i__1, 1, A(rk, k), lda, dA(rk,k), ldda );
/* Multiply on GPU */
// was CALL CGEMV( 'Conjugate transpose', M-RK+1, N-K,
// TAU( K ), A( RK, K+1 ), LDA,
// A( RK, K ), 1,
// CZERO, F( K+1, K ), 1 )
magma_int_t i__3 = nb-k-1;
magma_int_t i__4 = i__2 - i__3;
magma_int_t i__5 = nb-k;
magma_cgemv( MagmaConjTrans, i__1 - i__5, i__2 - i__3,
tau[k], dA(rk +i__5, k+1+i__3), ldda,
dA(rk +i__5, k ), ione,
c_zero, dF(k+1+i__3, k ), ione );
magma_cgetmatrix_async( i__2-i__3, 1,
dF(k + 1 +i__3, k), i__2,
F (k + 1 +i__3, k), i__2, stream );
blasf77_cgemv( MagmaConjTransStr, &i__1, &i__3,
&tau[k], A(rk, k+1), &lda,
A(rk, k ), &ione,
&c_zero, F(k+1, k ), &ione );
magma_queue_sync( stream );
blasf77_cgemv( MagmaConjTransStr, &i__5, &i__4,
&tau[k], A(rk, k+1+i__3), &lda,
A(rk, k ), &ione,
&c_one, F(k+1+i__3, k ), &ione );
}
/* Padding F(1:K,K) with zeros. */
for (j = 0; j < k; ++j) {
*F(j, k) = c_zero;
}
/* Incremental updating of F:
F(1:N,K) := F(1:N,K) - tau(K)*F(1:N,1:K-1)*A(RK:M,1:K-1)'*A(RK:M,K). */
if (k > 0) {
i__1 = m - rk;
i__2 = k;
z__1 = MAGMA_C_NEGATE( tau[k] );
blasf77_cgemv( MagmaConjTransStr, &i__1, &i__2,
&z__1, A(rk, 0), &lda,
A(rk, k), &ione,
&c_zero, auxv, &ione );
i__1 = k;
blasf77_cgemv( MagmaNoTransStr, &n, &i__1,
&c_one, F(0,0), &ldf,
auxv, &ione,
&c_one, F(0,k), &ione );
}
/* Optimization: On the last iteration start sending F back to the GPU */
/* Update the current row of A:
A(RK,K+1:N) := A(RK,K+1:N) - A(RK,1:K)*F(K+1:N,1:K)'. */
if (k < n-1) {
i__1 = n - k - 1;
i__2 = k + 1;
blasf77_cgemm( MagmaNoTransStr, MagmaConjTransStr, &ione, &i__1, &i__2,
&c_neg_one, A(rk, 0 ), &lda,
F(k+1,0 ), &ldf,
&c_one, A(rk, k+1), &lda );
}
/* Update partial column norms. */
if (rk < lastrk) {
for (j = k + 1; j < n; ++j) {
if (vn1[j] != 0.) {
/* NOTE: The following 4 lines follow from the analysis in
Lapack Working Note 176. */
temp = MAGMA_C_ABS( *A(rk,j) ) / vn1[j];
temp = max( 0., ((1. + temp) * (1. - temp)) );
d__1 = vn1[j] / vn2[j];
temp2 = temp * (d__1 * d__1);
if (temp2 <= tol3z) {
vn2[j] = (float) lsticc;
lsticc = j;
} else {
vn1[j] *= magma_ssqrt(temp);
}
}
}
}
*A(rk, k) = Akk;
++k;
}
// leave k as the last column done
--k;
*kb = k + 1;
rk = offset + *kb - 1;
/* Apply the block reflector to the rest of the matrix:
A(OFFSET+KB+1:M,KB+1:N) := A(OFFSET+KB+1:M,KB+1:N) - A(OFFSET+KB+1:M,1:KB)*F(KB+1:N,1:KB)' */
if (*kb < min(n, m - offset)) {
i__1 = m - rk - 1;
i__2 = n - *kb;
/* Send F to the GPU */
magma_csetmatrix( i__2, *kb,
F (*kb, 0), ldf,
dF(*kb, 0), i__2 );
magma_cgemm( MagmaNoTrans, MagmaConjTrans, i__1, i__2, *kb,
c_neg_one, dA(rk+1, 0 ), ldda,
dF(*kb, 0 ), i__2,
c_one, dA(rk+1, *kb), ldda );
}
/* Recomputation of difficult columns. */
while( lsticc > 0 ) {
itemp = (magma_int_t)(vn2[lsticc] >= 0. ? floor(vn2[lsticc] + .5) : -floor(.5 - vn2[lsticc]));
i__1 = m - rk - 1;
if (lsticc <= nb)
vn1[lsticc] = magma_cblas_scnrm2( i__1, A(rk+1,lsticc), ione );
else {
/* Where is the data, CPU or GPU ? */
float r1, r2;
r1 = magma_cblas_scnrm2( nb-k, A(rk+1,lsticc), ione );
r2 = magma_scnrm2(m-offset-nb, dA(offset + nb + 1, lsticc), ione);
//vn1[lsticc] = magma_scnrm2(i__1, dA(rk + 1, lsticc), ione);
vn1[lsticc] = magma_ssqrt(r1*r1 + r2*r2);
}
/* NOTE: The computation of VN1( LSTICC ) relies on the fact that
SNRM2 does not fail on vectors with norm below the value of SQRT(SLAMCH('S')) */
vn2[lsticc] = vn1[lsticc];
lsticc = itemp;
}
magma_queue_destroy( stream );
return MAGMA_SUCCESS;
} /* magma_claqps */