floating point optimizations in vc7/7.1 
Author Message
 floating point optimizations in vc7/7.1

Hi all,

We are considering migrating from vc6 to vc7.  Our code has a lot of
floating point computations.  Has anyone run into problems with vc7
produced optimized floating point code?  I remember very well the very
painful migration from 4.2 to 6.  It took quite a while to determine
that the first release of vc6 had a number of diffulties in this
regard and we were not able to use it until SP3.

I understand that vc7 has a completely different back end.  Hopefully
it would produce better fp optimizations than vc6 (which seems rather
bad in this regard), but if it has the occasional problem we probably
can't use it (it's just too painful to track these problems down and
work around them).

And if people have come across problems in vc7, do you know if they
have been fixed in vc7.1?

Thanks in advance,

Eric



Sat, 16 Apr 2005 12:52:12 GMT  
 floating point optimizations in vc7/7.1

Quote:

>We are considering migrating from vc6 to vc7.  Our code has a lot of
>floating point computations.  Has anyone run into problems with vc7
>produced optimized floating point code?

Yes.

http://groups.google.com/groups?q=float+bug&btnG=Google+Search&meta=g...

--
Mike



Sun, 17 Apr 2005 23:56:36 GMT  
 floating point optimizations in vc7/7.1

Quote:


> >We are considering migrating from vc6 to vc7.  Our code has a lot of
> >floating point computations.  Has anyone run into problems with vc7
> >produced optimized floating point code?

> Yes.

> http://groups.google.com/groups?q=float+bug&btnG=Google+Search&meta=g...

Thanks for the pointer.  There is one thread in particular that causes
me to worry.  Specifically:

http://groups.google.com/groups?hl=en&lr=&ie=UTF-8&threadm=uTr9TN6wBH...

with title "VC7 Compiler Bug!!??".

Does anyone at MS know whether this particular code gen problem is
fixed with the upcoming 7.1 release?

Having such fp code gen problems lurking makes it impossible to
migrate to 7.0 as we (as the original poster of the bug) have fp
intensive code, and like him will have to stick with VC6 until such
problems are resolved.

Thanks in advance,

-Eric Twietmeyer



Tue, 19 Apr 2005 00:12:36 GMT  
 floating point optimizations in vc7/7.1
if u are looking for fp optimizations, look into icl [ Intel C++ compiler ]
or gcc 2.95 MingW. I had 3-4 times improvement for a large with a lot of
inlined code.


Quote:
> Hi all,

> We are considering migrating from vc6 to vc7.  Our code has a lot of
> floating point computations.  Has anyone run into problems with vc7
> produced optimized floating point code?  I remember very well the very
> painful migration from 4.2 to 6.  It took quite a while to determine
> that the first release of vc6 had a number of diffulties in this
> regard and we were not able to use it until SP3.

> I understand that vc7 has a completely different back end.  Hopefully
> it would produce better fp optimizations than vc6 (which seems rather
> bad in this regard), but if it has the occasional problem we probably
> can't use it (it's just too painful to track these problems down and
> work around them).

> And if people have come across problems in vc7, do you know if they
> have been fixed in vc7.1?

> Thanks in advance,

> Eric



Tue, 19 Apr 2005 07:14:41 GMT  
 floating point optimizations in vc7/7.1
My team has worked hard targeting float code specifically in the 7.0 to 7.1
time frame.  I would highly encourage you to try it.

The specific bug you mentioned was an error in loop boundary cases and had
nothing to do with float.  It is fixed.

The other bugs I have looked at in that thread have also been fixed
(although sadly some of the newer reports have been fixed post beta -
however I expect the betas to be refreshed reasonably soon).

I can't guarantee you smooth sailing; but I can promise you we took float
bugs very seriously.

We still have some optimization issues where we are aggressive in float
usage.  I believe you find these same issues and work arounds in competing
products.  While the /Op switch is heavy-handed; keep in mind that it can be
used for specific function with #pragma optimize() to help control the perf
hit.

Finally, we have also implemented /arch:SSE2 to generate native float/double
code using the SSE2 instructions if so desired.  Our perf in this area is
comparable (in my opinion) to our competitors.

Thanks,
Joe


Quote:
> Hi all,

> We are considering migrating from vc6 to vc7.  Our code has a lot of
> floating point computations.  Has anyone run into problems with vc7
> produced optimized floating point code?  I remember very well the very
> painful migration from 4.2 to 6.  It took quite a while to determine
> that the first release of vc6 had a number of diffulties in this
> regard and we were not able to use it until SP3.

> I understand that vc7 has a completely different back end.  Hopefully
> it would produce better fp optimizations than vc6 (which seems rather
> bad in this regard), but if it has the occasional problem we probably
> can't use it (it's just too painful to track these problems down and
> work around them).

> And if people have come across problems in vc7, do you know if they
> have been fixed in vc7.1?

> Thanks in advance,

> Eric



Wed, 20 Apr 2005 08:20:47 GMT  
 floating point optimizations in vc7/7.1
Hi Joe,

Thanks for the candid information.  We will try 7.1 once it is
available.  I hope you can get the issues that have been resolved post
beta into the release when it occurs!  I look forward to trying it
out.

-Eric Twietmeyer


Quote:
> My team has worked hard targeting float code specifically in the 7.0 to 7.1
> time frame.  I would highly encourage you to try it.

> The specific bug you mentioned was an error in loop boundary cases and had
> nothing to do with float.  It is fixed.

> The other bugs I have looked at in that thread have also been fixed
> (although sadly some of the newer reports have been fixed post beta -
> however I expect the betas to be refreshed reasonably soon).

> I can't guarantee you smooth sailing; but I can promise you we took float
> bugs very seriously.

> We still have some optimization issues where we are aggressive in float
> usage.  I believe you find these same issues and work arounds in competing
> products.  While the /Op switch is heavy-handed; keep in mind that it can be
> used for specific function with #pragma optimize() to help control the perf
> hit.

> Finally, we have also implemented /arch:SSE2 to generate native float/double
> code using the SSE2 instructions if so desired.  Our perf in this area is
> comparable (in my opinion) to our competitors.

> Thanks,
> Joe



Fri, 22 Apr 2005 08:52:28 GMT  
 
 [ 6 post ] 

 Relevant Pages 

1. Floating point optimization "annoyance"

2. MSVC V5.0 Floating-point Optimization problem

3. floating point rounding or floating point printing ?

4. Implicit decimal points in floating-point reads

5. Converting floating point to scaled fixed point

6. Converting Microsoft Floating Point type to the IEEE floating

7. float to string to float - decimal points

8. float to string to float - decimal points

9. Poor Optimization in VC7.1

10. VC7 optimization question

11. optimization settings in vc7

12. VC7 optimization?

 

 
Powered by phpBB® Forum Software