Jump to content

  • Log In with Google      Sign In   
  • Create Account


FPU mode


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
12 replies to this topic

#1 Laval B   Crossbones+   -  Reputation: 4346

Like
2Likes
Like

Posted 15 July 2012 - 06:05 PM

Hello everyone.

I need to to make sure that the fpu mode, mostly the rounding mode and the floating point pricision, are set correctly. With Microsoft's compiler i can use the intrinsic _controlfp which is very handy.

With gcc though, i didn't find such an intrinsic. The only way i found was using inline assembler. Is there such an intrinsic for gcc ? I know with Microsoft compiler you can't use inline assembler, is that also true for gcc.
We think in generalities, but we live in details.
 
Alfred North Whitehead

Sponsor:

#2 frob   Moderators   -  Reputation: 18832

Like
1Likes
Like

Posted 16 July 2012 - 09:34 AM

Beware of doing that.

Many libraries, such as D3D, rely on specific settings and will either not operate correctly or will give poor performance when you start messing with floating point settings.

There was a time twenty years ago when you needed to mess with floating point options. These days it is a rare thing.
Check out my personal indie blog at bryanwagstaff.com.

#3 NightCreature83   Crossbones+   -  Reputation: 2669

Like
0Likes
Like

Posted 16 July 2012 - 05:47 PM

MSVC++ does allow inline assembler btw just need to add:
__asm{
	 mov eax, 0; Your assembler code here, rember ';' is comment in assembler
}

The code I provide is just an example and only sets register eax to 0.

Using intrinsics is a far simpler way to express what you want the compiler to create in assmbler however and far more encouraged, remember the compiler is far better at optimising assembler code then you are.(Unless you are a true coding guru)

Edited by NightCreature83, 16 July 2012 - 05:51 PM.

Worked on titles: CMR:DiRT2, DiRT 3, DiRT: Showdown, GRID 2, Mad Max

#4 mark ds   Members   -  Reputation: 1070

Like
1Likes
Like

Posted 16 July 2012 - 06:04 PM

MSVS in 64 bit windows does not allow inline assambler.

To the OP:

The only reason you'd wan't to contol fp mode would be to sync the CPU and GPU. However. this isn't possilbe. The GPU relies on FMA instructions, which have no CPU counterparts (yet), so it's a pretty futile excersise. You cannot get the same results from a GPU and CPU. (at least since gtx400 GPUs).

#5 ApochPiQ   Moderators   -  Reputation: 14252

Like
2Likes
Like

Posted 16 July 2012 - 06:09 PM

To the OP:

The only reason you'd wan't to contol fp mode would be to sync the CPU and GPU. However. this isn't possilbe. The GPU relies on FMA instructions, which have no CPU counterparts (yet), so it's a pretty futile excersise. You cannot get the same results from a GPU and CPU. (at least since gtx400 GPUs).


Not entirely true... for example, the outputs of SSE2 instructions can be made to converge to (close to) FPU instructions with appropriate settings. This might be useful if, say, one wants a 64-bit app to generate similar results to a 32-bit app, within some tolerance.

I do agree that it's rare to need to do this, and any twiddling of the FPU control bits should be viewed with ample suspicion in this day and age. But it's not entirely useless.

#6 Laval B   Crossbones+   -  Reputation: 4346

Like
1Likes
Like

Posted 17 July 2012 - 03:28 AM

Not entirely true... for example, the outputs of SSE2 instructions can be made to converge to (close to) FPU instructions with appropriate settings. This might be useful if, say, one wants a 64-bit app to generate similar results to a 32-bit app, within some tolerance.


Thank you, this is the case.

Edited by Laval B, 17 July 2012 - 03:42 AM.

We think in generalities, but we live in details.
 
Alfred North Whitehead

#7 NightCreature83   Crossbones+   -  Reputation: 2669

Like
2Likes
Like

Posted 17 July 2012 - 04:08 AM

MSVS in 64 bit windows does not allow inline assambler.

To the OP:

The only reason you'd wan't to contol fp mode would be to sync the CPU and GPU. However. this isn't possilbe. The GPU relies on FMA instructions, which have no CPU counterparts (yet), so it's a pretty futile excersise. You cannot get the same results from a GPU and CPU. (at least since gtx400 GPUs).

While that is true the OP didnt state wheter he was using x64 or not, so I assumed not especially since he talked about inline assmbler anyway. However you can still use assmbler if you want to but you have to add an assembler source file and add a reference to the assembler function to the C++ code.

See: http://www.youtube.com/watch?v=guru397zg2g&list=PL0C5C980A28FEE68D&index=1&feature=plpp_video he has more videos on how to write x64 ASM and they are good imo.

Edited by NightCreature83, 17 July 2012 - 04:10 AM.

Worked on titles: CMR:DiRT2, DiRT 3, DiRT: Showdown, GRID 2, Mad Max

#8 scniton   Members   -  Reputation: 252

Like
1Likes
Like

Posted 17 July 2012 - 07:34 AM

Look in the source code in the link, it will likely help: http://www.christian...rojekte/fpmath/
EDIT: The source is buried at the bottom, here is the direct link: http://www.christian-seiler.de/projekte/fpmath/xpfpa.h

Edited by scniton, 17 July 2012 - 07:39 AM.

Stop twiddling your bits and use them already!

#9 Laval B   Crossbones+   -  Reputation: 4346

Like
1Likes
Like

Posted 17 July 2012 - 10:59 AM

EDIT: The source is buried at the bottom, here is the direct link: http://www.christian.../fpmath/xpfpa.h


Very interesting and useful. It explains and clarifies alot of things.

danke schön
We think in generalities, but we live in details.
 
Alfred North Whitehead

#10 ApochPiQ   Moderators   -  Reputation: 14252

Like
1Likes
Like

Posted 17 July 2012 - 12:34 PM

The linked article is indeed informative, but not necessarily up to date or complete.

I've done a fair bit of work on floating-point determinism so if you have any specific questions or problems feel free to ask!

#11 Laval B   Crossbones+   -  Reputation: 4346

Like
0Likes
Like

Posted 17 July 2012 - 02:24 PM

I've done a fair bit of work on floating-point determinism so if you have any specific questions or problems feel free to ask!


Thank you very much, it's really appreciated. I'm still reading the link above.

Edited by Laval B, 17 July 2012 - 02:25 PM.

We think in generalities, but we live in details.
 
Alfred North Whitehead

#12 web383   Members   -  Reputation: 734

Like
0Likes
Like

Posted 17 July 2012 - 03:14 PM

There was a time twenty years ago when you needed to mess with floating point options. These days it is a rare thing.


Isn't controlling the fp mode required for deterministic game simulations and lock-step networking? Or am I thinking of something else?

#13 scniton   Members   -  Reputation: 252

Like
0Likes
Like

Posted 20 July 2012 - 10:35 AM

Isn't controlling the fp mode required for deterministic game simulations and lock-step networking? Or am I thinking of something else?

It's not necessarily required, it is one of the methods used to achieve determinism. Some games use integers only for the parts that require determinism, other games set the FPU control state and sprinkle their code with FPU control state checks to make sure it doesn't get changed (I think they also typically mandate SSE2.)
Stop twiddling your bits and use them already!




Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS