home bbs files messages ]

Forums before death by AOL, social media and spammers... "We can't have nice things"

   comp.arch      Apparently more than just beeps & boops      131,241 messages   

[   << oldest   |   < older   |   list   |   newer >   |   newest >>   ]

   Message 130,735 of 131,241   
   Anton Ertl to Stephen Fuld   
   Re: A typical non-loop use case for SIMD   
   01 Jan 26 22:21:36   
   
   From: anton@mips.complang.tuwien.ac.at   
      
   Stephen Fuld  writes:   
   >Note also that this works well for min (and max), but for add, etc. if   
   >the data is unsigned only if you don't care where in the sequence an   
   >overflow occurred, but but if the data is signed, only if you don't even   
   >care if overflow occurred.   
      
   Addition with trapping on overflow or with overflow detection is not   
   associative.  Addition modulo 2^n is.   
      
   If you have a sticky overflow bit (Power has something in that   
   direction), or if you trap on overflow (MIPS and Alpha have such   
   instructions for signed addition), you will certainly notice if an   
   overflow occured for signed additions for the particular evaluation   
   used on the machine.  If there are evaluations with and without   
   overflow, I think that those without are preferable (although probably   
   not enough to pay the performance cost for ensuring that).   
      
   Which programming language are you thinking of where reduction uses   
   some addition with overflow detection.   
      
   - anton   
   --   
   'Anyone trying for "industrial quality" ISA should avoid undefined behavior.'   
     Mitch Alsup,    
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)   

[   << oldest   |   < older   |   list   |   newer >   |   newest >>   ]


(c) 1994,  bbs@darkrealms.ca