Revision 0:
==========
Distributed to Anna Tsao, Zhaoujun Bai, and Mehmet at Northwestern U.
It works with the release of LAPACK as of the end of April. 

Revision 1:
============
Fixed a bug in sgeqpf.F. Changed default for nullity to 10.

Revision 2:
==========
o sgeqr2 and sgeqrf were changed to allow for m < n. As a result,
  processing of rejected columns has been simplified.
o more accurate flop counts take ICE and computation and
  updating of column norms into account.
o computes real and 'effective' Mflop rates. Baseline figure
  for QR with pivoting is the Businger/Golub algorithm.
o seed for random number generator was added as an input
  argument


Revision 3:
==========
Processing of rejected columns has been changed to apply column
pivoting until the estimate delivered by ICE is actually above the
threshold for the condition number. This avoids rejecting 'good'
columns that were thrown out as a result of the failure of the column
pivoting strategy.

Distributed to Karen Williamson at Rice 8/1/90.

Revision 5:
==========
o uses the new calling sequence for slatms
o uses the third root of n to scale the maximum column norm
  in computing an estimate for the largest singular value.
o the file structure has been changed to simply link in the
  lapack library from Ed Anderson's directory.

Revision 8:
==========
o restarts ICE at fixed intervals
o consistently uses SLAUC1 to find whether a column is acceptable
o From this point, Greg is doing the work.
o carried out some changes to be capable to compile on the
  sun4 compiler.
o in the unblocked code of sgeqpf.F the call to factorize columns 2:mn 
  lacks off the the parameter MXNM, which has been added.
o added two hyphens needed in the parameter checking part of sgeqpf.F.

Revision 9:
==========
o Removed the part about fixed columns.
o All the stuff to execute inverse iteration every some steps has been
  removed too.
o Adapted to LAPACK vesion 1.0b from LAPACK preliminary version.
o Variable LWSIZE is not allowed to be smaller than KB any more.
o The routine slaic1 is the one included in LAPACK version 1.0b.
o The problem of "Division by zero" in the computation of the Megaflop rate
  when the time is zero has been solved. 

Revision 10:
===========
o The postprocessing is executed after Chris' local pivoting algorithm,
  considered from now the preprocessing.
o Block algorithms are used in the postprocessing.
o The full algorithm is named sGEFUL. sGEQPF is still Chris' preprocessing.

Revision 11:
===========
o The postprocessing algorithm has been changed to improve the worst 
  case, that is, the first type of matrix of sGNTST.
o From now sGEQPF.F is the full algorithm: pre + postprocessing,
  sGEQPR.F is the main routine for chris' preprocessing, and sGEQPO.F is 
  the main routine for the postprocessing.
o Accounting the flops in the postprocessing has beed added.
o Added the part to reveal the right rank from any estimate to the
  postprocessing.
o The flags for the compile have been changed.

Revision 12:
===========
o Modified to generate single and double precision from the same code.
  The new command is: compile s/d machine.

Revision 13:
===========
o Two postprocessing algorithms have been implemented and included 
  in this revision: Chandrasekaran&Ipsen algorithms and Pan&Tang algorithms.

Distributed to Zhaoujun Bai, Kentucky.  May 17, 1994.

Revision 14:
===========
o Chandrasekaran&Ipsen algorithms have been selected.
o Adapted to the final subroutine layout for TOMS and LAPACK.
o Block application of orthogonal transformations to matrix C.
o The transpose of orthogonal transformations (matrix Q') are applied 
  to matrix C from the left if job=2.
o The orthogonal transformations (matrix Q) are applied to matrix C 
  from the right if job=3.

Distributed to George Fann, PNL, Spring 1995.

Revision 15:
============
o Block QR (xGEQRF) is used to factorize rejected columns in the last part 
  of QR with local pivoting.
o Matrix C is completely updated with Blas-3.
o Included again Pan&Tang algorithms.
o Some routine names have changed: xGEQPX, xTRQPX and xTRQXC are the routines 
  to compute Chandra&Ipsen; xGEQPY, xTRQPY and xTRQYC are the routines to 
  compute Pan&Tang; the name xGEQPW (window pivoting) replaces to xGEQP2.
o Fixed some performance and precision problems.
o Prepared for submission to ACM TOMS.
