Re: machar and ifort

Posted by Valera on June 11, 2007 at 12:36:52:

In Reply to: machar and ifort posted by Michal Tarana on June 09, 2007 at 17:40:05:

Hi Michal,
if you have a look at src/util/machar.f you will find out that for some platforms these numbers are hardcoded, but for others - are computed..

And I can give a simple explanation - why (sometimes) it is impossible to use "standard" way of computing machine dependent constants.
Let's look at the simple statement:
IF(eps+1.gt.1)
A very 'smart' compiler (with -O2) can simplify the condition to
IF(eps.gt.0)

So. The only way to avoid this problem is to hardcode these constants..

I also believe that the problem you reported in another message is related to the fact that on AMD64 machine dependent constants are computed (and wrong), so you have a convergence problem.

Follow Ups:

Re: machar and ifort Michal Tarana 13:13:44 6/11/07 (1)
- Re: machar and ifort Valera 13:32:42 6/11/07 (0)

Post a Followup

Name:
E-Mail:

Subject:

if B is 1s22s22p1, what is Li?

Passfield:

Comments:
: Hi Michal, : if you have a look at src/util/machar.f you will find out that for some platforms these numbers are hardcoded, but for others - are computed.. : And I can give a simple explanation - why (sometimes) it is impossible to use "standard" way of computing machine dependent constants. : Let's look at the simple statement: : IF(eps+1.gt.1) : A very 'smart' compiler (with -O2) can simplify the condition to : IF(eps.gt.0) : So. The only way to avoid this problem is to hardcode these constants.. : I also believe that the problem you reported in another message is related to the fact that on AMD64 machine dependent constants are computed (and wrong), so you have a convergence problem.

[ Follow Ups ] [ Post Followup ] [ Molcas user's WWWBoard ]