[petsc-dev] default SNES rtol for single precision builds

[Jose E. Roman]

El 19/09/2011, a las 19:59, Barry Smith escribió:

> 
>  In petscmath.h we have 
> 
> #if defined(PETSC_USE_REAL_SINGLE)
> #if defined(MAXFLOAT)
> #  define PETSC_MAX_REAL                 MAXFLOAT
> #else
> #  define PETSC_MAX_REAL                1.e30
> #endif
> #  define PETSC_MIN_REAL                -PETSC_MAX_REAL
> #  define PETSC_MACHINE_EPSILON         1.e-7
> #  define PETSC_SQRT_MACHINE_EPSILON    3.e-4
> #  define PETSC_SMALL                   1.e-5
> #elif defined(PETSC_USE_REAL_DOUBLE)
> #  define PETSC_MAX_REAL                1.e300
> #  define PETSC_MIN_REAL                -PETSC_MAX_REAL
> #  define PETSC_MACHINE_EPSILON         1.e-14
> #  define PETSC_SQRT_MACHINE_EPSILON    1.e-7
> #  define PETSC_SMALL                   1.e-10
> #elif defined(PETSC_USE_REAL_LONG_DOUBLE)
> #  define PETSC_MAX_REAL                1.e4900L
> #  define PETSC_MIN_REAL                -PETSC_MAX_REAL
> #  define PETSC_MACHINE_EPSILON         1.e-18
> #  define PETSC_SQRT_MACHINE_EPSILON    1.e-9
> #  define PETSC_SMALL                   1.e-13
> #elif defined(PETSC_USE_REAL___FLOAT128)
> #  define PETSC_MAX_REAL                FLT128_MAX
> #  define PETSC_MIN_REAL                -FLT128_MAX
> #  define PETSC_MACHINE_EPSILON         FLT128_EPSILON
> #  define PETSC_SQRT_MACHINE_EPSILON    1.38777878078e-17
> #  define PETSC_SMALL                   1.e-20
> #endif
> 
>  these are a bit ad-hoc and there maybe more items we wish to add. If possible we could try to use values here to set the various default tolerances we have in various places 
> 
>   In particular the e-8 for SNES is very ad-hoc and has no justification, I originally used that number because I wanted something smaller than what we use for KSP but not too small. Does anybody have suggestions for a self-consistent justified set of rules for defining all our default tolerances?

Could use precision-independent expressions, such as
  rtol_KSP = PETSC_SQRT_MACHINE_EPSILON
  rtol_SNES = 0.1*PETSC_SQRT_MACHINE_EPSILON
but I don't have a justification why these values and not others.

By the way, it always seemed strange to me that PETSC_MACHINE_EPSILON (in double) is defined to be 1.e-14 instead of the actual IEEE value 1.11022e-16 (or 2.22045e-16). Its square root is 1.0537e-08.

Also, it could be possible to set these values by calling Lapack's _lamch at configure.

Jose

> 
>   Barry
> 
> 
> On Sep 19, 2011, at 12:27 PM, Lisandro Dalcin wrote:
> 
>> After debugging a failing petsc4py testcase using single precision
>> builds, I realized the default SNES relative tolerance is 1e-8. What
>> should we do?
>> 
>> -- 
>> Lisandro Dalcin
>> ---------------
>> CIMEC (INTEC/CONICET-UNL)
>> Predio CONICET-Santa Fe
>> Colectora RN 168 Km 472, Paraje El Pozo
>> 3000 Santa Fe, Argentina
>> Tel: +54-342-4511594 (ext 1011)
>> Tel/Fax: +54-342-4511169
>