Using FPU registers?

[oRBIT]

I'm running out of registers in my application! However I was thinking of beginning to use FPU registers, just for temporary storage. Imagine this scenario:
fmove d0,fp0
-- do calculations --
fmove fp0,d0

Is the above faster than using a simple variable in RAM? Like
move.l d0,temp
--- do calculations --
move.l temp,d0

[StingRay]

move.l d0,-(a7)
move.l (a7)+,d0

'nuff said.

[pmc]

LOL.

Never one to mince words is Stinger.

[oRBIT]

@Stingray: Yeah but is it really faster? Your way involves a memory read/write, which the fmove doesn't (?)

[coze]

well, why don't you just test and see with a simple code ?

[oRBIT]

My environment is currently consisting of WinUAE only (sadly) so..

[StingRay]

Originally Posted by oRBIT

@Stingray: Yeah but is it really faster? Your way involves a memory read/write, which the fmove doesn't (?)

First of all, using the FPU for temporary storage is really really stupid, your program will require an FPU only because you used it to store registers. Also, writing to the FPU requires a fair amount of time (convert to 80bit, convert to 32 bit integer again etc).

[oRBIT]

Yes of course it will require an FPU. I was just curious if there was any obvious "performance-improvement" compared to storing data in RAM...

[matthey]

The integer <-> fp conversion is too slow. Even if the data were single precision fp in the integer register it wouldn't be worth using the fp register. If you don't need fp operations then don't use the floating point registers. Operate directly on variables in memory or use the stack.

[StingRay]

Originally Posted by oRBIT

I was just curious if there was any obvious "performance-improvement" compared to storing data in RAM...

There isn't. And it should actually be obvious why. Anyway, I made some quick tests on my A4060, results:

15000 x fmove d0,fp0 fmove fp0,d0: : 147 raster lines (143 raster lines for fmove.s)
15000 x move.l d0,-(a7) move.l (a7)+,d0: 126 raster lines

It'll be even worse on 030/FPU systems.

Originally Posted by matthey

The integer <-> fp conversion is too slow. Even if the data were single precision fp in the integer register it wouldn't be worth using the fp register. If you don't need fp operations then don't use the floating point registers. Operate directly on variables in memory or use the stack.

That's basically exactly what I said.

[seuden]

If you don't listen to Sting you'll only get stung!

[Photon]

Well, he could have been using the FPU already in another part of the program. No need to call coders stupid, especially those coding advanced things are usually highly intelligent. And it makes sense to not rely on WinUAE timings. So.

But when you get bitten by the coding bug (again) I think you should get a real Amiga, then you can test things like this easily.

btw, orbit=orbiter?

[StingRay]

Originally Posted by Photon

Well, he could have been using the FPU already in another part of the program. No need to call coders stupid, especially those coding advanced things are usually highly intelligent. And it makes sense to not rely on WinUAE timings. So.

Dear Photon, please learn to read, thank you! I said using FPU registers for temporary storage is stupid and not that oRBIT is stupid!

[Photon]

Yes, you're correct. But I also think there's no need to accuse someone of doing stupid things before you know the reason why. He had two good reasons why he even asked about it, and since the stack and FPU method were so close, speed-wise it was not obvious that one was faster than the other. What if, for example, the FPU register method had been the fastest by 1 scanline for the AGA-060 platform? Then he would have found a faster method which all AGA 060 coders could use, and as we all know, "if it's faster it's faster" , and this 'standard demo platform' has a built in FPU.

Whether it feels correct to do this, well, I agree. But if it had been slightly faster, then all coders would have used this new method (where applicable). So.

[StingRay]

Originally Posted by Photon

But I also think there's no need to accuse someone of doing stupid things before you know the reason why. He had two good reasons why he even asked about it, and since the stack and FPU method were so close, speed-wise it was not obvious that one was faster than the other.

He asked, I answered. As simple as that. EOD.

[cosmicfrog]

heheheh

fight fight

don`t think anyone got called stupid, idea may have been shot down in flames though with proof to back it up

[Don_Adan]

Originally Posted by oRBIT

I'm running out of registers in my application! However I was thinking of beginning to use FPU registers, just for temporary storage. Imagine this scenario:
fmove d0,fp0
-- do calculations --
fmove fp0,d0

Is the above faster than using a simple variable in RAM? Like
move.l d0,temp
--- do calculations --
move.l temp,d0

How many registers you need?
You can use up to 17 Motorola registers, and if no longword values you can split 8 data registers via SWAP command, this is enough for almost all routines.