dlaqp3rk.f

Go to the documentation of this file.

*> \brief \b DLAQP3RK computes a step of truncated QR factorization with column pivoting of a real m-by-n matrix A using Level 3 BLAS and overwrites a real m-by-nrhs matrix B with Q**T * B.
*
*  =========== DOCUMENTATION ===========
*
* Online html documentation available at
*            http://www.netlib.org/lapack/explore-html/
*
*> Download DLAQP3RK + dependencies
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dlaqp3rk.f">
*> [TGZ]</a>
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dlaqp3rk.f">
*> [ZIP]</a>
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dlaqp3rk.f">
*> [TXT]</a>
*
*  Definition:
*  ===========
*
*      SUBROUTINE DLAQP3RK( M, N, NRHS, IOFFSET, NB, ABSTOL,
*     $                     RELTOL, KP1, MAXC2NRM, A, LDA, DONE, KB,
*     $                     MAXC2NRMK, RELMAXC2NRMK, JPIV, TAU,
*     $                     VN1, VN2, AUXV, F, LDF, IWORK, INFO )
*      IMPLICIT NONE
*      LOGICAL            DONE
*      INTEGER            INFO, IOFFSET, KB, KP1, LDA, LDF, M, N,
*     $                   NB, NRHS
*      DOUBLE PRECISION   ABSTOL, MAXC2NRM, MAXC2NRMK, RELMAXC2NRMK,
*     $                   RELTOL
*
*     .. Scalar Arguments ..
*      LOGICAL            DONE
*      INTEGER            KB, LDA, LDF, M, N, NB, NRHS, IOFFSET
*      DOUBLE PRECISION   ABSTOL, MAXC2NRM, MAXC2NRMK, RELMAXC2NRMK,
*     $                   RELTOL
*     ..
*     .. Array Arguments ..
*      INTEGER            IWORK( * ), JPIV( * )
*      DOUBLE PRECISION   A( LDA, * ), AUXV( * ), F( LDF, * ), TAU( * ),
*     $                   VN1( * ), VN2( * )
*     ..
*
*
*> \par Purpose:
*  =============
*>
*> \verbatim
*>
*> DLAQP3RK computes a step of truncated QR factorization with column
*> pivoting of a real M-by-N matrix A block A(IOFFSET+1:M,1:N)
*> by using Level 3 BLAS as
*>
*>   A * P(KB) = Q(KB) * R(KB).
*>
*> The routine tries to factorize NB columns from A starting from
*> the row IOFFSET+1 and updates the residual matrix with BLAS 3
*> xGEMM. The number of actually factorized columns is returned
*> is smaller than NB.
*>
*> Block A(1:IOFFSET,1:N) is accordingly pivoted, but not factorized.
*>
*> The routine also overwrites the right-hand-sides B matrix stored
*> in A(IOFFSET+1:M,1:N+1:N+NRHS) with Q(KB)**T * B.
*>
*> Cases when the number of factorized columns KB < NB:
*>
*> (1) In some cases, due to catastrophic cancellations, it cannot
*> factorize all NB columns and need to update the residual matrix.
*> Hence, the actual number of factorized columns in the block returned
*> in KB is smaller than NB. The logical DONE is returned as FALSE.
*> The factorization of the whole original matrix A_orig must proceed
*> with the next block.
*>
*> (2) Whenever the stopping criterion ABSTOL or RELTOL is satisfied,
*> the factorization of the whole original matrix A_orig is stopped,
*> the logical DONE is returned as TRUE. The number of factorized
*> columns which is smaller than NB is returned in KB.
*>
*> (3) In case both stopping criteria ABSTOL or RELTOL are not used,
*> and when the residual matrix is a zero matrix in some factorization
*> step KB, the factorization of the whole original matrix A_orig is
*> stopped, the logical DONE is returned as TRUE. The number of
*> factorized columns which is smaller than NB is returned in KB.
*>
*> (4) Whenever NaN is detected in the matrix A or in the array TAU,
*> the factorization of the whole original matrix A_orig is stopped,
*> the logical DONE is returned as TRUE. The number of factorized
*> columns which is smaller than NB is returned in KB. The INFO
*> parameter is set to the column index of the first NaN occurrence.
*>
*> \endverbatim
*
*  Arguments:
*  ==========
*
*> \param[in] M
*> \verbatim
*>          M is INTEGER
*>          The number of rows of the matrix A. M >= 0.
*> \endverbatim
*>
*> \param[in] N
*> \verbatim
*>          N is INTEGER
*>          The number of columns of the matrix A. N >= 0
*> \endverbatim
*>
*> \param[in] NRHS
*> \verbatim
*>          NRHS is INTEGER
*>          The number of right hand sides, i.e., the number of
*>          columns of the matrix B. NRHS >= 0.
*> \endverbatim
*>
*> \param[in] IOFFSET
*> \verbatim
*>          IOFFSET is INTEGER
*>          The number of rows of the matrix A that must be pivoted
*>          but not factorized. IOFFSET >= 0.
*>
*>          IOFFSET also represents the number of columns of the whole
*>          original matrix A_orig that have been factorized
*>          in the previous steps.
*> \endverbatim
*>
*> \param[in] NB
*> \verbatim
*>          NB is INTEGER
*>          Factorization block size, i.e the number of columns
*>          to factorize in the matrix A. 0 <= NB
*>
*>          If NB = 0, then the routine exits immediately.
*>             This means that the factorization is not performed,
*>             the matrices A and B and the arrays TAU, IPIV
*>             are not modified.
*> \endverbatim
*>
*> \param[in] ABSTOL
*> \verbatim
*>          ABSTOL is DOUBLE PRECISION, cannot be NaN.
*>
*>          The absolute tolerance (stopping threshold) for
*>          maximum column 2-norm of the residual matrix.
*>          The algorithm converges (stops the factorization) when
*>          the maximum column 2-norm of the residual matrix
*>          is less than or equal to ABSTOL.
*>
*>          a) If ABSTOL < 0.0, then this stopping criterion is not
*>                used, the routine factorizes columns depending
*>                on NB and RELTOL.
*>                This includes the case ABSTOL = -Inf.
*>
*>          b) If 0.0 <= ABSTOL then the input value
*>                of ABSTOL is used.
*> \endverbatim
*>
*> \param[in] RELTOL
*> \verbatim
*>          RELTOL is DOUBLE PRECISION, cannot be NaN.
*>
*>          The tolerance (stopping threshold) for the ratio of the
*>          maximum column 2-norm of the residual matrix to the maximum
*>          column 2-norm of the original matrix A_orig. The algorithm
*>          converges (stops the factorization), when this ratio is
*>          less than or equal to RELTOL.
*>
*>          a) If RELTOL < 0.0, then this stopping criterion is not
*>                used, the routine factorizes columns depending
*>                on NB and ABSTOL.
*>                This includes the case RELTOL = -Inf.
*>
*>          d) If 0.0 <= RELTOL then the input value of RELTOL
*>                is used.
*> \endverbatim
*>
*> \param[in] KP1
*> \verbatim
*>          KP1 is INTEGER
*>          The index of the column with the maximum 2-norm in
*>          the whole original matrix A_orig determined in the
*>          main routine DGEQP3RK. 1 <= KP1 <= N_orig.
*> \endverbatim
*>
*> \param[in] MAXC2NRM
*> \verbatim
*>          MAXC2NRM is DOUBLE PRECISION
*>          The maximum column 2-norm of the whole original
*>          matrix A_orig computed in the main routine DGEQP3RK.
*>          MAXC2NRM >= 0.
*> \endverbatim
*>
*> \param[in,out] A
*> \verbatim
*>          A is DOUBLE PRECISION array, dimension (LDA,N+NRHS)
*>          On entry:
*>              the M-by-N matrix A and M-by-NRHS matrix B, as in
*>
*>                                  N     NRHS
*>              array_A   =   M  [ mat_A, mat_B ]
*>
*>          On exit:
*>          1. The elements in block A(IOFFSET+1:M,1:KB) below
*>             the diagonal together with the array TAU represent
*>             the orthogonal matrix Q(KB) as a product of elementary
*>             reflectors.
*>          2. The upper triangular block of the matrix A stored
*>             in A(IOFFSET+1:M,1:KB) is the triangular factor obtained.
*>          3. The block of the matrix A stored in A(1:IOFFSET,1:N)
*>             has been accordingly pivoted, but not factorized.
*>          4. The rest of the array A, block A(IOFFSET+1:M,KB+1:N+NRHS).
*>             The left part A(IOFFSET+1:M,KB+1:N) of this block
*>             contains the residual of the matrix A, and,
*>             if NRHS > 0, the right part of the block
*>             A(IOFFSET+1:M,N+1:N+NRHS) contains the block of
*>             the right-hand-side matrix B. Both these blocks have been
*>             updated by multiplication from the left by Q(KB)**T.
*> \endverbatim
*>
*> \param[in] LDA
*> \verbatim
*>          LDA is INTEGER
*>          The leading dimension of the array A. LDA >= max(1,M).
*> \endverbatim
*>
*> \param[out] DONE
*> \verbatim
*>          DONE is LOGICAL
*>          TRUE: a) if the factorization completed before processing
*>                   all min(M-IOFFSET,NB,N) columns due to ABSTOL
*>                   or RELTOL criterion,
*>                b) if the factorization completed before processing
*>                   all min(M-IOFFSET,NB,N) columns due to the
*>                   residual matrix being a ZERO matrix.
*>                c) when NaN was detected in the matrix A
*>                   or in the array TAU.
*>          FALSE: otherwise.
*> \endverbatim
*>
*> \param[out] KB
*> \verbatim
*>          KB is INTEGER
*>          Factorization rank of the matrix A, i.e. the rank of
*>          the factor R, which is the same as the number of non-zero
*>          rows of the factor R.  0 <= KB <= min(M-IOFFSET,NB,N).
*>
*>          KB also represents the number of non-zero Householder
*>          vectors.
*> \endverbatim
*>
*> \param[out] MAXC2NRMK
*> \verbatim
*>          MAXC2NRMK is DOUBLE PRECISION
*>          The maximum column 2-norm of the residual matrix,
*>          when the factorization stopped at rank KB. MAXC2NRMK >= 0.
*> \endverbatim
*>
*> \param[out] RELMAXC2NRMK
*> \verbatim
*>          RELMAXC2NRMK is DOUBLE PRECISION
*>          The ratio MAXC2NRMK / MAXC2NRM of the maximum column
*>          2-norm of the residual matrix (when the factorization
*>          stopped at rank KB) to the maximum column 2-norm of the
*>          original matrix A_orig. RELMAXC2NRMK >= 0.
*> \endverbatim
*>
*> \param[out] JPIV
*> \verbatim
*>          JPIV is INTEGER array, dimension (N)
*>          Column pivot indices, for 1 <= j <= N, column j
*>          of the matrix A was interchanged with column JPIV(j).
*> \endverbatim
*>
*> \param[out] TAU
*> \verbatim
*>          TAU is DOUBLE PRECISION array, dimension (min(M-IOFFSET,N))
*>          The scalar factors of the elementary reflectors.
*> \endverbatim
*>
*> \param[in,out] VN1
*> \verbatim
*>          VN1 is DOUBLE PRECISION array, dimension (N)
*>          The vector with the partial column norms.
*> \endverbatim
*>
*> \param[in,out] VN2
*> \verbatim
*>          VN2 is DOUBLE PRECISION array, dimension (N)
*>          The vector with the exact column norms.
*> \endverbatim
*>
*> \param[out] AUXV
*> \verbatim
*>          AUXV is DOUBLE PRECISION array, dimension (NB)
*>          Auxiliary vector.
*> \endverbatim
*>
*> \param[out] F
*> \verbatim
*>          F is DOUBLE PRECISION array, dimension (LDF,NB)
*>          Matrix F**T = L*(Y**T)*A.
*> \endverbatim
*>
*> \param[in] LDF
*> \verbatim
*>          LDF is INTEGER
*>          The leading dimension of the array F. LDF >= max(1,N+NRHS).
*> \endverbatim
*>
*> \param[out] IWORK
*> \verbatim
*>          IWORK is INTEGER array, dimension (N-1).
*>          Is a work array. ( IWORK is used to store indices
*>          of "bad" columns for norm downdating in the residual
*>          matrix ).
*> \endverbatim
*>
*> \param[out] INFO
*> \verbatim
*>          INFO is INTEGER
*>          1) INFO = 0: successful exit.
*>          2) If INFO = j_1, where 1 <= j_1 <= N, then NaN was
*>             detected and the routine stops the computation.
*>             The j_1-th column of the matrix A or the j_1-th
*>             element of array TAU contains the first occurrence
*>             of NaN in the factorization step KB+1 ( when KB columns
*>             have been factorized ).
*>
*>             On exit:
*>             KB                  is set to the number of
*>                                    factorized columns without
*>                                    exception.
*>             MAXC2NRMK           is set to NaN.
*>             RELMAXC2NRMK        is set to NaN.
*>             TAU(KB+1:min(M,N))     is not set and contains undefined
*>                                    elements. If j_1=KB+1, TAU(KB+1)
*>                                    may contain NaN.
*>          3) If INFO = j_2, where N+1 <= j_2 <= 2*N, then no NaN
*>             was detected, but +Inf (or -Inf) was detected and
*>             the routine continues the computation until completion.
*>             The (j_2-N)-th column of the matrix A contains the first
*>             occurrence of +Inf (or -Inf) in the actorization
*>             step KB+1 ( when KB columns have been factorized ).
*> \endverbatim
*
*  Authors:
*  ========
*
*> \author Univ. of Tennessee
*> \author Univ. of California Berkeley
*> \author Univ. of Colorado Denver
*> \author NAG Ltd.
*
*> \ingroup laqp3rk
*
*> \par References:
*  ================
*> [1] A Level 3 BLAS QR factorization algorithm with column pivoting developed in 1996.
*> G. Quintana-Orti, Depto. de Informatica, Universidad Jaime I, Spain.
*> X. Sun, Computer Science Dept., Duke University, USA.
*> C. H. Bischof, Math. and Comp. Sci. Div., Argonne National Lab, USA.
*> A BLAS-3 version of the QR factorization with column pivoting.
*> LAPACK Working Note 114
*> <a href="https://www.netlib.org/lapack/lawnspdf/lawn114.pdf">https://www.netlib.org/lapack/lawnspdf/lawn114.pdf</a>
*> and in
*> SIAM J. Sci. Comput., 19(5):1486-1494, Sept. 1998.
*> <a href="https://doi.org/10.1137/S1064827595296732">https://doi.org/10.1137/S1064827595296732</a>
*>
*> [2] A partial column norm updating strategy developed in 2006.
*> Z. Drmac and Z. Bujanovic, Dept. of Math., University of Zagreb, Croatia.
*> On the failure of rank revealing QR factorization software – a case study.
*> LAPACK Working Note 176.
*> <a href="http://www.netlib.org/lapack/lawnspdf/lawn176.pdf">http://www.netlib.org/lapack/lawnspdf/lawn176.pdf</a>
*> and in
*> ACM Trans. Math. Softw. 35, 2, Article 12 (July 2008), 28 pages.
*> <a href="https://doi.org/10.1145/1377612.1377616">https://doi.org/10.1145/1377612.1377616</a>
*
*> \par Contributors:
*  ==================
*>
*> \verbatim
*>
*>  November  2023, Igor Kozachenko, James Demmel,
*>                  EECS Department,
*>                  University of California, Berkeley, USA.
*>
*> \endverbatim
*
*  =====================================================================
      SUBROUTINE dlaqp3rk( M, N, NRHS, IOFFSET, NB, ABSTOL,
     $                     RELTOL, KP1, MAXC2NRM, A, LDA, DONE, KB,
     $                     MAXC2NRMK, RELMAXC2NRMK, JPIV, TAU,
     $                     VN1, VN2, AUXV, F, LDF, IWORK, INFO )
      IMPLICIT NONE
*
*  -- LAPACK auxiliary routine --
*  -- LAPACK is a software package provided by Univ. of Tennessee,    --
*  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
*
*     .. Scalar Arguments ..
      LOGICAL            DONE
      INTEGER            INFO, IOFFSET, KB, KP1, LDA, LDF, M, N,
     $                   NB, NRHS
      DOUBLE PRECISION   ABSTOL, MAXC2NRM, MAXC2NRMK, RELMAXC2NRMK,
     $                   reltol
*     ..
*     .. Array Arguments ..
      INTEGER            IWORK( * ), JPIV( * )
      DOUBLE PRECISION   A( LDA, * ), AUXV( * ), F( LDF, * ), TAU( * ),
     $                   VN1( * ), VN2( * )
*     ..
*
*  =====================================================================
*
*     .. Parameters ..
      DOUBLE PRECISION   ZERO, ONE
      PARAMETER          ( ZERO = 0.0d+0, one = 1.0d+0 )
*     ..
*     .. Local Scalars ..
      INTEGER            ITEMP, J, K, MINMNFACT, MINMNUPDT,
     $                   LSTICC, KP, I, IF
      DOUBLE PRECISION   AIK, HUGEVAL, TEMP, TEMP2, TOL3Z
*     ..
*     .. External Subroutines ..
      EXTERNAL           dgemm, dgemv, dlarfg, dswap
*     ..
*     .. Intrinsic Functions ..
      INTRINSIC          abs, max, min, sqrt
*     ..
*     .. External Functions ..
      LOGICAL            DISNAN
      INTEGER            IDAMAX
      DOUBLE PRECISION   DLAMCH, DNRM2
      EXTERNAL           disnan, dlamch, idamax, dnrm2
*     ..
*     .. Executable Statements ..
*
*     Initialize INFO
*
      info = 0
*
*     MINMNFACT in the smallest dimension of the submatrix
*     A(IOFFSET+1:M,1:N) to be factorized.
*
      minmnfact = min( m-ioffset, n )
      minmnupdt = min( m-ioffset, n+nrhs )
      nb = min( nb, minmnfact )
      tol3z = sqrt( dlamch( 'Epsilon' ) )
      hugeval = dlamch( 'Overflow' )
*
*     Compute factorization in a while loop over NB columns,
*     K is the column index in the block A(1:M,1:N).
*
      k = 0
      lsticc = 0
      done = .false.
*
      DO WHILE ( k.LT.nb .AND. lsticc.EQ.0 )
         k = k + 1
         i = ioffset + k
*
         IF( i.EQ.1 ) THEN
*
*           We are at the first column of the original whole matrix A_orig,
*           therefore we use the computed KP1 and MAXC2NRM from the
*           main routine.
*
            kp = kp1
*
         ELSE
*
*           Determine the pivot column in K-th step, i.e. the index
*           of the column with the maximum 2-norm in the
*           submatrix A(I:M,K:N).
*
            kp = ( k-1 ) + idamax( n-k+1, vn1( k ), 1 )
*
*           Determine the maximum column 2-norm and the relative maximum
*           column 2-norm of the submatrix A(I:M,K:N) in step K.
*
            maxc2nrmk = vn1( kp )
*
*           ============================================================
*
*           Check if the submatrix A(I:M,K:N) contains NaN, set
*           INFO parameter to the column number, where the first NaN
*           is found and return from the routine.
*           We need to check the condition only if the
*           column index (same as row index) of the original whole
*           matrix is larger than 1, since the condition for whole
*           original matrix is checked in the main routine.
*
            IF( disnan( maxc2nrmk ) ) THEN
*
               done = .true.
*
*              Set KB, the number of factorized partial columns
*                      that are non-zero in each step in the block,
*                      i.e. the rank of the factor R.
*              Set IF, the number of processed rows in the block, which
*                      is the same as the number of processed rows in
*                      the original whole matrix A_orig.
*
               kb = k - 1
               IF = i - 1
               info = kb + kp
*
*              Set RELMAXC2NRMK to NaN.
*
               relmaxc2nrmk = maxc2nrmk
*
*              There is no need to apply the block reflector to the
*              residual of the matrix A stored in A(KB+1:M,KB+1:N),
*              since the submatrix contains NaN and we stop
*              the computation.
*              But, we need to apply the block reflector to the residual
*              right hand sides stored in A(KB+1:M,N+1:N+NRHS), if the
*              residual right hand sides exist.  This occurs
*              when ( NRHS != 0 AND KB <= (M-IOFFSET) ):
*
*              A(I+1:M,N+1:N+NRHS) := A(I+1:M,N+1:N+NRHS) -
*                               A(I+1:M,1:KB) * F(N+1:N+NRHS,1:KB)**T.
 
               IF( nrhs.GT.0 .AND. kb.LT.(m-ioffset) ) THEN
                  CALL dgemm( 'No transpose', 'Transpose',
     $                  m-IF, nrhs, kb, -one, a( if+1, 1 ), lda,
     $                  f( n+1, 1 ), ldf, one, a( if+1, n+1 ), lda )
               END IF
*
*              There is no need to recompute the 2-norm of the
*              difficult columns, since we stop the factorization.
*
*              Array TAU(KF+1:MINMNFACT) is not set and contains
*              undefined elements.
*
*              Return from the routine.
*
               RETURN
            END IF
*
*           Quick return, if the submatrix A(I:M,K:N) is
*           a zero matrix. We need to check it only if the column index
*           (same as row index) is larger than 1, since the condition
*           for the whole original matrix A_orig is checked in the main
*           routine.
*
            IF( maxc2nrmk.EQ.zero ) THEN
*
               done = .true.
*
*              Set KB, the number of factorized partial columns
*                      that are non-zero in each step in the block,
*                      i.e. the rank of the factor R.
*              Set IF, the number of processed rows in the block, which
*                      is the same as the number of processed rows in
*                      the original whole matrix A_orig.
*
               kb = k - 1
               IF = i - 1
               relmaxc2nrmk = zero
*
*              There is no need to apply the block reflector to the
*              residual of the matrix A stored in A(KB+1:M,KB+1:N),
*              since the submatrix is zero and we stop the computation.
*              But, we need to apply the block reflector to the residual
*              right hand sides stored in A(KB+1:M,N+1:N+NRHS), if the
*              residual right hand sides exist.  This occurs
*              when ( NRHS != 0 AND KB <= (M-IOFFSET) ):
*
*              A(I+1:M,N+1:N+NRHS) := A(I+1:M,N+1:N+NRHS) -
*                               A(I+1:M,1:KB) * F(N+1:N+NRHS,1:KB)**T.
*
               IF( nrhs.GT.0 .AND. kb.LT.(m-ioffset) ) THEN
                  CALL dgemm( 'No transpose', 'Transpose',
     $                  m-IF, nrhs, kb, -one, a( if+1, 1 ), lda,
     $                  f( n+1, 1 ), ldf, one, a( if+1, n+1 ), lda )
               END IF
*
*              There is no need to recompute the 2-norm of the
*              difficult columns, since we stop the factorization.
*
*              Set TAUs corresponding to the columns that were not
*              factorized to ZERO, i.e. set TAU(KB+1:MINMNFACT) = ZERO,
*              which is equivalent to seting TAU(K:MINMNFACT) = ZERO.
*
               DO j = k, minmnfact
                  tau( j ) = zero
               END DO
*
*              Return from the routine.
*
               RETURN
*
            END IF
*
*           ============================================================
*
*           Check if the submatrix A(I:M,K:N) contains Inf,
*           set INFO parameter to the column number, where
*           the first Inf is found plus N, and continue
*           the computation.
*           We need to check the condition only if the
*           column index (same as row index) of the original whole
*           matrix is larger than 1, since the condition for whole
*           original matrix is checked in the main routine.
*
            IF( info.EQ.0 .AND. maxc2nrmk.GT.hugeval ) THEN
               info = n + k - 1 + kp
            END IF
*
*           ============================================================
*
*           Test for the second and third tolerance stopping criteria.
*           NOTE: There is no need to test for ABSTOL.GE.ZERO, since
*           MAXC2NRMK is non-negative. Similarly, there is no need
*           to test for RELTOL.GE.ZERO, since RELMAXC2NRMK is
*           non-negative.
*           We need to check the condition only if the
*           column index (same as row index) of the original whole
*           matrix is larger than 1, since the condition for whole
*           original matrix is checked in the main routine.
*
            relmaxc2nrmk =  maxc2nrmk / maxc2nrm
*
            IF( maxc2nrmk.LE.abstol .OR. relmaxc2nrmk.LE.reltol ) THEN
*
               done = .true.
*
*              Set KB, the number of factorized partial columns
*                      that are non-zero in each step in the block,
*                      i.e. the rank of the factor R.
*              Set IF, the number of processed rows in the block, which
*                      is the same as the number of processed rows in
*                      the original whole matrix A_orig;
*
                  kb = k - 1
                  IF = i - 1
*
*              Apply the block reflector to the residual of the
*              matrix A and the residual of the right hand sides B, if
*              the residual matrix and and/or the residual of the right
*              hand sides exist,  i.e. if the submatrix
*              A(I+1:M,KB+1:N+NRHS) exists.  This occurs when
*                 KB < MINMNUPDT = min( M-IOFFSET, N+NRHS ):
*
*              A(IF+1:M,K+1:N+NRHS) := A(IF+1:M,KB+1:N+NRHS) -
*                             A(IF+1:M,1:KB) * F(KB+1:N+NRHS,1:KB)**T.
*
               IF( kb.LT.minmnupdt ) THEN
                  CALL dgemm( 'No transpose', 'Transpose',
     $                  m-IF, n+nrhs-kb, kb,-one, a( if+1, 1 ), lda,
     $                  f( kb+1, 1 ), ldf, one, a( if+1, kb+1 ), lda )
               END IF
*
*              There is no need to recompute the 2-norm of the
*              difficult columns, since we stop the factorization.
*
*              Set TAUs corresponding to the columns that were not
*              factorized to ZERO, i.e. set TAU(KB+1:MINMNFACT) = ZERO,
*              which is equivalent to seting TAU(K:MINMNFACT) = ZERO.
*
               DO j = k, minmnfact
                  tau( j ) = zero
               END DO
*
*              Return from the routine.
*
               RETURN
*
            END IF
*
*           ============================================================
*
*           End ELSE of IF(I.EQ.1)
*
         END IF
*
*        ===============================================================
*
*        If the pivot column is not the first column of the
*        subblock A(1:M,K:N):
*        1) swap the K-th column and the KP-th pivot column
*           in A(1:M,1:N);
*        2) swap the K-th row and the KP-th row in F(1:N,1:K-1)
*        3) copy the K-th element into the KP-th element of the partial
*           and exact 2-norm vectors VN1 and VN2. (Swap is not needed
*           for VN1 and VN2 since we use the element with the index
*           larger than K in the next loop step.)
*        4) Save the pivot interchange with the indices relative to the
*           the original matrix A_orig, not the block A(1:M,1:N).
*
         IF( kp.NE.k ) THEN
            CALL dswap( m, a( 1, kp ), 1, a( 1, k ), 1 )
            CALL dswap( k-1, f( kp, 1 ), ldf, f( k, 1 ), ldf )
            vn1( kp ) = vn1( k )
            vn2( kp ) = vn2( k )
            itemp = jpiv( kp )
            jpiv( kp ) = jpiv( k )
            jpiv( k ) = itemp
         END IF
*
*        Apply previous Householder reflectors to column K:
*        A(I:M,K) := A(I:M,K) - A(I:M,1:K-1)*F(K,1:K-1)**T.
*
         IF( k.GT.1 ) THEN
            CALL dgemv( 'No transpose', m-i+1, k-1, -one, a( i, 1 ),
     $                  lda, f( k, 1 ), ldf, one, a( i, k ), 1 )
         END IF
*
*        Generate elementary reflector H(k) using the column A(I:M,K).
*
         IF( i.LT.m ) THEN
            CALL dlarfg( m-i+1, a( i, k ), a( i+1, k ), 1, tau( k ) )
         ELSE
            tau( k ) = zero
         END IF
*
*        Check if TAU(K) contains NaN, set INFO parameter
*        to the column number where NaN is found and return from
*        the routine.
*        NOTE: There is no need to check TAU(K) for Inf,
*        since DLARFG cannot produce TAU(K) or Householder vector
*        below the diagonal containing Inf. Only BETA on the diagonal,
*        returned by DLARFG can contain Inf, which requires
*        TAU(K) to contain NaN. Therefore, this case of generating Inf
*        by DLARFG is covered by checking TAU(K) for NaN.
*
         IF( disnan( tau(k) ) ) THEN
*
            done = .true.
*
*           Set KB, the number of factorized partial columns
*                   that are non-zero in each step in the block,
*                   i.e. the rank of the factor R.
*           Set IF, the number of processed rows in the block, which
*                   is the same as the number of processed rows in
*                   the original whole matrix A_orig.
*
            kb = k - 1
            IF = i - 1
            info = k
*
*           Set MAXC2NRMK and  RELMAXC2NRMK to NaN.
*
            maxc2nrmk = tau( k )
            relmaxc2nrmk = tau( k )
*
*           There is no need to apply the block reflector to the
*           residual of the matrix A stored in A(KB+1:M,KB+1:N),
*           since the submatrix contains NaN and we stop
*           the computation.
*           But, we need to apply the block reflector to the residual
*           right hand sides stored in A(KB+1:M,N+1:N+NRHS), if the
*           residual right hand sides exist.  This occurs
*           when ( NRHS != 0 AND KB <= (M-IOFFSET) ):
*
*           A(I+1:M,N+1:N+NRHS) := A(I+1:M,N+1:N+NRHS) -
*                            A(I+1:M,1:KB) * F(N+1:N+NRHS,1:KB)**T.
*
            IF( nrhs.GT.0 .AND. kb.LT.(m-ioffset) ) THEN
               CALL dgemm( 'No transpose', 'Transpose',
     $               m-IF, nrhs, kb, -one, a( if+1, 1 ), lda,
     $               f( n+1, 1 ), ldf, one, a( if+1, n+1 ), lda )
            END IF
*
*           There is no need to recompute the 2-norm of the
*           difficult columns, since we stop the factorization.
*
*           Array TAU(KF+1:MINMNFACT) is not set and contains
*           undefined elements.
*
*           Return from the routine.
*
            RETURN
         END IF
*
*        ===============================================================
*
         aik = a( i, k )
         a( i, k ) = one
*
*        ===============================================================
*
*        Compute the current K-th column of F:
*          1) F(K+1:N,K) := tau(K) * A(I:M,K+1:N)**T * A(I:M,K).
*
         IF( k.LT.n+nrhs ) THEN
            CALL dgemv( 'Transpose', m-i+1, n+nrhs-k,
     $                  tau( k ), a( i, k+1 ), lda, a( i, k ), 1,
     $                  zero, f( k+1, k ), 1 )
         END IF
*
*           2) Zero out elements above and on the diagonal of the
*              column K in matrix F, i.e elements F(1:K,K).
*
         DO j = 1, k
            f( j, k ) = zero
         END DO
*
*         3) Incremental updating of the K-th column of F:
*        F(1:N,K) := F(1:N,K) - tau(K) * F(1:N,1:K-1) * A(I:M,1:K-1)**T
*                    * A(I:M,K).
*
         IF( k.GT.1 ) THEN
            CALL dgemv( 'Transpose', m-i+1, k-1, -tau( k ),
     $                  a( i, 1 ), lda, a( i, k ), 1, zero,
     $                  auxv( 1 ), 1 )
*
            CALL dgemv( 'No transpose', n+nrhs, k-1, one,
     $                  f( 1, 1 ), ldf, auxv( 1 ), 1, one,
     $                  f( 1, k ), 1 )
         END IF
*
*        ===============================================================
*
*        Update the current I-th row of A:
*        A(I,K+1:N+NRHS) := A(I,K+1:N+NRHS)
*                         - A(I,1:K)*F(K+1:N+NRHS,1:K)**T.
*
         IF( k.LT.n+nrhs ) THEN
            CALL dgemv( 'No transpose', n+nrhs-k, k, -one,
     $                  f( k+1, 1 ), ldf, a( i, 1 ), lda, one,
     $                  a( i, k+1 ), lda )
         END IF
*
         a( i, k ) = aik
*
*        Update the partial column 2-norms for the residual matrix,
*        only if the residual matrix A(I+1:M,K+1:N) exists, i.e.
*        when K < MINMNFACT = min( M-IOFFSET, N ).
*
         IF( k.LT.minmnfact ) THEN
*
            DO j = k + 1, n
               IF( vn1( j ).NE.zero ) THEN
*
*                 NOTE: The following lines follow from the analysis in
*                 Lapack Working Note 176.
*
                  temp = abs( a( i, j ) ) / vn1( j )
                  temp = max( zero, ( one+temp )*( one-temp ) )
                  temp2 = temp*( vn1( j ) / vn2( j ) )**2
                  IF( temp2.LE.tol3z ) THEN
*
*                    At J-index, we have a difficult column for the
*                    update of the 2-norm. Save the index of the previous
*                    difficult column in IWORK(J-1).
*                    NOTE: ILSTCC > 1, threfore we can use IWORK only
*                    with N-1 elements, where the elements are
*                    shifted by 1 to the left.
*
                     iwork( j-1 ) = lsticc
*
*                    Set the index of the last difficult column LSTICC.
*
                     lsticc = j
*
                  ELSE
                     vn1( j ) = vn1( j )*sqrt( temp )
                  END IF
               END IF
            END DO
*
         END IF
*
*        End of while loop.
*
      END DO
*
*     Now, afler the loop:
*        Set KB, the number of factorized columns in the block;
*        Set IF, the number of processed rows in the block, which
*                is the same as the number of processed rows in
*                the original whole matrix A_orig, IF = IOFFSET + KB.
*
      kb = k
      IF = i
*
*     Apply the block reflector to the residual of the matrix A
*     and the residual of the right hand sides B, if the residual
*     matrix and and/or the residual of the right hand sides
*     exist,  i.e. if the submatrix A(I+1:M,KB+1:N+NRHS) exists.
*     This occurs when KB < MINMNUPDT = min( M-IOFFSET, N+NRHS ):
*
*     A(IF+1:M,K+1:N+NRHS) := A(IF+1:M,KB+1:N+NRHS) -
*                         A(IF+1:M,1:KB) * F(KB+1:N+NRHS,1:KB)**T.
*
      IF( kb.LT.minmnupdt ) THEN
         CALL dgemm( 'No transpose', 'Transpose',
     $         m-IF, n+nrhs-kb, kb, -one, a( if+1, 1 ), lda,
     $         f( kb+1, 1 ), ldf, one, a( if+1, kb+1 ), lda )
      END IF
*
*     Recompute the 2-norm of the difficult columns.
*     Loop over the index of the difficult columns from the largest
*     to the smallest index.
*
      DO WHILE( lsticc.GT.0 )
*
*        LSTICC is the index of the last difficult column is greater
*        than 1.
*        ITEMP is the index of the previous difficult column.
*
         itemp = iwork( lsticc-1 )
*
*        Compute the 2-norm explicilty for the last difficult column and
*        save it in the partial and exact 2-norm vectors VN1 and VN2.
*
*        NOTE: The computation of VN1( LSTICC ) relies on the fact that
*        DNRM2 does not fail on vectors with norm below the value of
*        SQRT(DLAMCH('S'))
*
         vn1( lsticc ) = dnrm2( m-IF, a( if+1, lsticc ), 1 )
         vn2( lsticc ) = vn1( lsticc )
*
*        Downdate the index of the last difficult column to
*        the index of the previous difficult column.
*
         lsticc = itemp
*
      END DO
*
      RETURN
*
*     End of DLAQP3RK
*
      END