view libcruft/amos/cuchk.f @ 8964:f4f4d65faaa0

Implement sparse * diagonal and diagonal * sparse operations, double-prec only. Date: Sun, 8 Mar 2009 16:28:18 -0400 These preserve sparsity, so eye(5) * sprand (5, 5, .2) is *sparse* and not dense. This may affect people who use multiplication by eye() rather than full(). The liboctave routines do *not* check if arguments are scalars in disguise. There is a type problem with checking at that level. I suspect we want diag * "sparse scalar" to stay diagonal, but we have to return a sparse matrix at the liboctave. Rather than worrying about that in liboctave, we cope with it when binding to Octave and return the correct higher-level type. The implementation is in Sparse-diag-op-defs.h rather than Sparse-op-defs.h to limit recompilation. And the implementations are templates rather than macros to produce better compiler errors and debugging information.
author Jason Riedy <jason@acm.org>
date Mon, 09 Mar 2009 17:49:13 -0400
parents 82be108cc558
children
line wrap: on
line source

      SUBROUTINE CUCHK(Y, NZ, ASCLE, TOL)
C***BEGIN PROLOGUE  CUCHK
C***REFER TO CSERI,CUOIK,CUNK1,CUNK2,CUNI1,CUNI2,CKSCL
C
C      Y ENTERS AS A SCALED QUANTITY WHOSE MAGNITUDE IS GREATER THAN
C      EXP(-ALIM)=ASCLE=1.0E+3*R1MACH(1)/TOL. THE TEST IS MADE TO SEE
C      IF THE MAGNITUDE OF THE REAL OR IMAGINARY PART WOULD UNDER FLOW
C      WHEN Y IS SCALED (BY TOL) TO ITS PROPER VALUE. Y IS ACCEPTED
C      IF THE UNDERFLOW IS AT LEAST ONE PRECISION BELOW THE MAGNITUDE
C      OF THE LARGEST COMPONENT; OTHERWISE THE PHASE ANGLE DOES NOT HAVE
C      ABSOLUTE ACCURACY AND AN UNDERFLOW IS ASSUMED.
C
C***ROUTINES CALLED  (NONE)
C***END PROLOGUE  CUCHK
C
      COMPLEX Y
      REAL ASCLE, SS, ST, TOL, YR, YI
      INTEGER NZ
      NZ = 0
      YR = REAL(Y)
      YI = AIMAG(Y)
      YR = ABS(YR)
      YI = ABS(YI)
      ST = AMIN1(YR,YI)
      IF (ST.GT.ASCLE) RETURN
      SS = AMAX1(YR,YI)
      ST=ST/TOL
      IF (SS.LT.ST) NZ = 1
      RETURN
      END