[Carpet] type-cast warnings

[Erik Schnetter]

On Oct 17, 2006, at 09:13:54, Bela Szilagyi wrote:

> While compiling Carpet on  our AEI cluster "lagavulin" I got the  
> following
> warnings:
>
> /home/szilagyi/Cactus/configs/lharm/build/CarpetReduce/reduce.cc 
> (459): warning
> #                           810: conversion from "double" to
> "CarpetReduce::typeconv<int>::goodtype= 
> {int}"                           m ay
> lose significant bits
>           mycnt += w;
>                 ^
>           detected during instantiation of "void
> CarpetReduce::reduce<T,OP>(cons                           t int *,  
> const int
> *, const int *, const std::vector<const void *, std::allocat
> o r<const void *>> &, const std::vector<double,  
> std::allocator<double>> &,
> void                           *,  void *, const double *, double)  
> [with
> T=CarpetReduce::typeconv<int>::goodtyp                          e= 
> { int},
> OP=CarpetReduce::minimum::op<CarpetReduce::typeconv<int>::goodtype={in
> t}>] "
>
>
> What is worrisome is that the resulting value of (mycnt) could  
> depend "w"
> being a roundoff quantity above or below a certain integer.  In  
> case one runs
> the application across a non-homogeneus architecture cluster, it is  
> very
> likely that on certain nodes mycnt will have values different from  
> other
> nodes.
>
> It is possible that code before this assignment cleans up the value  
> of "w" so
> that the result is architecture independent.  But I was not able to  
> confirm
> this.
>
> The question is then -- should one worry about these kinds of  
> warnings in
> Carpet(Regrid) ?

Béla,

in general you do not need to worry about these warnings in  
CarpetReduce.  CarpetReduce offers a variety of operations to users  
for a variety of types.  Among these are e.g. calculating the average  
of a set of integers and storing the result as integer.  This has,  
obviously, the problem you mention.  However, these operations are  
only performed if someone requests them, and I hope that people won't  
do that unless they know what they are doing.

You can see that this is a generic piece of code by the <T,OP> C++  
template syntax; the same code implements these operations for all  
types.  In other, non-generic code, such warnings would be serious.

There is no good way to define e.g. integer averages of integers.   
The alternative would be to forbid this operation.

-erik

-- 
Erik Schnetter <schnetter at cct.lsu.edu>

My email is as private as my paper mail.  I therefore support encrypting
and signing email messages.  Get my PGP key from www.keyserver.net.



-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 186 bytes
Desc: This is a digitally signed message part
Url : /archives/developers/attachments/20061017/b5ea0f16/attachment.pgp