IEEE-Libs for WinUAE or Amiga 68881/2 FPU

AMIGASYSTEM · 11 June 2018, 06:38

IEEE-libs for WinUAE or AMiGA 68881/2 FPU

IEEE-Libs v45.881 r14: Download

Changes in release 11:
=====================

Always set the rounding to extended precision round to zero.

Safer fallback to the original libs when no FPU is detected.
The fallback is only supported with PatchMSB or RemLib from
Thomas Rapp, but not with LoadModule, LoadResident or a ROM.

Removed the segment list pointer from the resident module to
make it possible to add these libraries to a custom ROM, but
this has not been tested and is not recommended, because if
these libraries are resident it's not possible to fallback to
the original AmigaOS libs when no FPU is found in the system.

Description:
===========

These libraries are translating all floating point function
calls directly into FPU instructions instead of doing integer
calculations as done by the original AmigaOS IEEE libraries.
In case that no FPU is detected, the original libs are loaded.

The standard installation, this is a safe setup:
===============================================

1) Copy c/PatchMSB to C:
2) Make a directory Libs:NoFPU
Copy your AmigaOS mathieeelibs to Libs:NoFPU ; as fallback
Copy the contents of libs/ to Libs:
3) Insert a line below SetPatch in your startup-sequence:

SetPatch QUIET
...
PatchMSB ; installs the same functions as in MSB-library
; remove all other patches for IEEE libraries
4) Reboot

Legal basis:
===========

My libraries and sources are freeware and freely distributable.
I don't give any warranties. You're using this software on your
own risk and you are responsible for any damage it might cause!

PeterK · 11 June 2018, 16:14

That are no news, that's 13 years old ... (just old wine in new bottles).

Romanujan · 11 June 2018, 16:39

Name is confusing - till now I was sure they use some UAE-specific hacks

PeterK · 11 June 2018, 17:45

No, there is only the Pow(a, b)=a^b function left, which is a bit optimized for the PC FPU. Usually you would do e^(b*ln(a)), but the PCs don't have direct support for e^x and ln(x), thus I'm using 2^x and log2(x) instead, and these instructions are also directly available in the Amiga 68881/2 FPU.

All other PC specific "hacks", as once used in release 8, are functions in WinUAE for long. On the Amiga side all IEEE functions are translated into optimized 68881/2 FPU instructions, no other "hacks" are left. So you can use the MathLibsUAE also on a real Amiga with 68881/2 FPU. That was always possible, even with those r8 "hacks".

AMIGASYSTEM · 11 June 2018, 17:53

Peter on AfA OS you advise to use them or I can make it pleasant

N.B. I posted the news because I had not found anything about the forun

PeterK · 11 June 2018, 18:04

Yes, you can install them under AfA_OS as shown in my readme. AmiSh:t is doing that too.

Btw, you've made a nice mathlibs logo on ikir's italian news page, Carlo.

AMIGASYSTEM · 11 June 2018, 19:56

Yes Peter I have already installed the libraries on AFA and they seem to be fine, for the logo or made a simple mixture. I'm not very good with graphics like my two sons now in their thirties, they started with DeluxPaint.

logo on ikir's, also Dual-Png icon

Kalamatee · 11 June 2018, 21:54

Hello PeterK, I am currently fixing up some (soft) float related issues on AROS m68k, and would be interested in integrating something similar to your libraries in our IEEE libs.

Would you be interested in helping or allow the use of your code for such a thing?

PeterK · 11 June 2018, 22:50

Hello Kalamatee,

how can I help you with my simple code? Most of the functions are just mapped to the FPU instructions. Only a few like Pow(), Floor() and Ceil() are based on my personal coding attempts. If you find anything useful for your IEEE libs, feel free to pick out whatever you need. In case that you don't understand some parts of my code or have doubt that it could work correctly, we could discuss the details per email if you like, since this is more a news thread here.

@AMIGASYSTEM
Thanks a lot for the nice icon. Ok, 30000 bytes of stack seems to be a little overkill, but that's on the safe side, for sure. I just don't know if the next release takes another 13 years before I can add your icon to the package.

AMIGASYSTEM · 11 June 2018, 23:28

Quote:

Originally Posted by PeterK

Thanks a lot for the nice icon. Ok, 30000 bytes of stack seems to be a little overkill.

I did not insert the stack 30000, the system automatically generated it after the conversion of the two images png

AMIGASYSTEM · 28 June 2018, 08:05

IEEE-Libs v45.881 r14 Update

PeterK · 28 June 2018, 18:48

Shock!

I just noticed that accidentally I made a stupid little mistake in my IEEE DP compare function. I don't know wtf has driven me to replace a TST.L D1 instruction with OR.L D3,D1 ???

But then I checked the compare function for possible wrong results caused by this mistake, and I'm very glad now to realize that this "bug" seems to have no negative effects. It's still all ok.

Update: No, this small mistake is not even a bug as supposed, in fact it's a tiny optimization of the code. Some special cases are done a bit faster now.

Code:

Cmp                 TST.L          D2
                    BMI.S          minus
                    BNE.S          plus
                    TST.L          D3
                    BEQ.S          Tst       ; D2/D3=0 -> Tst
plus                SUB.L          D2,D0
                    BLT.S          less
                    BNE.S          more
                    CMP.L          D3,D1
                    BHI.S          more
                    BNE.S          less
                    RTS

                    CNOP           0,4

minus               TST.L          D0
                    BNE.S          skip
                    OR.L           D3,D1
                    BNE.S          more
                    BSET.L         #31,D0    ; D0/D1=0 -> -0
skip                SUB.L          D2,D0
                    BCS.S          more
                    BNE.S          less
                    CMP.L          D3,D1
                    BHI.S          less
                    BEQ.S          same
more                MOVEQ          #1,D0
                    RTS

less                MOVEQ          #-1,D0
                    RTS

                    CNOP           0,4

Tst                 ADD.L          D0,D0
                    BCS.S          lessi
                    OR.L           D1,D0
                    BNE.S          more
same                RTS

lessi               OR.L           D1,D0
                    BNE.S          less
                    RTS

11 June 2018, 19:56	#7
AMIGASYSTEM Registered User Join Date: Aug 2014 Location: Brindisi (Italy) Age: 70 Posts: 8,248	Yes Peter I have already installed the libraries on AFA and they seem to be fine, for the logo or made a simple mixture. I'm not very good with graphics like my two sons now in their thirties, they started with DeluxPaint. logo on ikir's, also Dual-Png icon Last edited by AMIGASYSTEM; 19 July 2018 at 19:36.

11 June 2018, 22:50	#9
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,365	Hello Kalamatee, how can I help you with my simple code? Most of the functions are just mapped to the FPU instructions. Only a few like Pow(), Floor() and Ceil() are based on my personal coding attempts. If you find anything useful for your IEEE libs, feel free to pick out whatever you need. In case that you don't understand some parts of my code or have doubt that it could work correctly, we could discuss the details per email if you like, since this is more a news thread here. @AMIGASYSTEM Thanks a lot for the nice icon. Ok, 30000 bytes of stack seems to be a little overkill, but that's on the safe side, for sure. I just don't know if the next release takes another 13 years before I can add your icon to the package. Last edited by PeterK; 11 June 2018 at 23:15.

28 June 2018, 18:48	#12
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,365	Shock! I just noticed that accidentally I made a stupid little mistake in my IEEE DP compare function. I don't know wtf has driven me to replace a TST.L D1 instruction with OR.L D3,D1 ??? But then I checked the compare function for possible wrong results caused by this mistake, and I'm very glad now to realize that this "bug" seems to have no negative effects. It's still all ok. Update: No, this small mistake is not even a bug as supposed, in fact it's a tiny optimization of the code. Some special cases are done a bit faster now. Code: Cmp TST.L D2 BMI.S minus BNE.S plus TST.L D3 BEQ.S Tst ; D2/D3=0 -> Tst plus SUB.L D2,D0 BLT.S less BNE.S more CMP.L D3,D1 BHI.S more BNE.S less RTS CNOP 0,4 minus TST.L D0 BNE.S skip OR.L D3,D1 BNE.S more BSET.L #31,D0 ; D0/D1=0 -> -0 skip SUB.L D2,D0 BCS.S more BNE.S less CMP.L D3,D1 BHI.S less BEQ.S same more MOVEQ #1,D0 RTS less MOVEQ #-1,D0 RTS CNOP 0,4 Tst ADD.L D0,D0 BCS.S lessi OR.L D1,D0 BNE.S more same RTS lessi OR.L D1,D0 BNE.S less RTS Last edited by PeterK; 28 June 2018 at 19:43.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
68881 FPU FMOVECR test	Toni Wilen	Amiga scene	15	16 January 2021 17:16
WinUAE: Disable JIT/FPU	AMIGASYSTEM	support.WinUAE	11	30 November 2017 14:10
All MUI Libs for ALL flavours of Amiga are dated wrongly	rockape	support.Apps	1	22 December 2013 13:46
Possible WinUAE FPU Problem	Ed Cruse	support.WinUAE	2	01 November 2012 16:46
68881 Fpu	Lord Aga	support.Hardware	2	18 November 2011 21:49

11 June 2018, 06:38	#1
AMIGASYSTEM Registered User Join Date: Aug 2014 Location: Brindisi (Italy) Age: 70 Posts: 8,248	IEEE-Libs for WinUAE or Amiga 68881/2 FPU IEEE-libs for WinUAE or AMiGA 68881/2 FPU IEEE-Libs v45.881 r14: Download Changes in release 11: ===================== Always set the rounding to extended precision round to zero. Safer fallback to the original libs when no FPU is detected. The fallback is only supported with PatchMSB or RemLib from Thomas Rapp, but not with LoadModule, LoadResident or a ROM. Removed the segment list pointer from the resident module to make it possible to add these libraries to a custom ROM, but this has not been tested and is not recommended, because if these libraries are resident it's not possible to fallback to the original AmigaOS libs when no FPU is found in the system. Description: =========== These libraries are translating all floating point function calls directly into FPU instructions instead of doing integer calculations as done by the original AmigaOS IEEE libraries. In case that no FPU is detected, the original libs are loaded. The standard installation, this is a safe setup: =============================================== 1) Copy c/PatchMSB to C: 2) Make a directory Libs:NoFPU Copy your AmigaOS mathieeelibs to Libs:NoFPU ; as fallback Copy the contents of libs/ to Libs: 3) Insert a line below SetPatch in your startup-sequence: SetPatch QUIET ... PatchMSB ; installs the same functions as in MSB-library ; remove all other patches for IEEE libraries 4) Reboot Legal basis: =========== My libraries and sources are freeware and freely distributable. I don't give any warranties. You're using this software on your own risk and you are responsible for any damage it might cause! Last edited by AMIGASYSTEM; 28 June 2018 at 08:04.

11 June 2018, 16:14	#2
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,365	That are no news, that's 13 years old ... (just old wine in new bottles).

11 June 2018, 16:39	#3
Romanujan Registered User Join Date: Dec 2007 Location: Szczecin/Poland Posts: 424	Name is confusing - till now I was sure they use some UAE-specific hacks

11 June 2018, 17:45	#4
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,365	No, there is only the Pow(a, b)=a^b function left, which is a bit optimized for the PC FPU. Usually you would do e^(b*ln(a)), but the PCs don't have direct support for e^x and ln(x), thus I'm using 2^x and log2(x) instead, and these instructions are also directly available in the Amiga 68881/2 FPU. All other PC specific "hacks", as once used in release 8, are functions in WinUAE for long. On the Amiga side all IEEE functions are translated into optimized 68881/2 FPU instructions, no other "hacks" are left. So you can use the MathLibsUAE also on a real Amiga with 68881/2 FPU. That was always possible, even with those r8 "hacks".

11 June 2018, 17:53	#5
AMIGASYSTEM Registered User Join Date: Aug 2014 Location: Brindisi (Italy) Age: 70 Posts: 8,248	Peter on AfA OS you advise to use them or I can make it pleasant N.B. I posted the news because I had not found anything about the forun

11 June 2018, 18:04	#6
PeterK Registered User Join Date: Apr 2005 Location: digital hell, Germany, after 1984, but worse Posts: 3,365	Yes, you can install them under AfA_OS as shown in my readme. AmiSh:t is doing that too. Btw, you've made a nice mathlibs logo on ikir's italian news page, Carlo.

11 June 2018, 21:54	#8
Kalamatee Registered User Join Date: May 2017 Location: Scotland Posts: 53	Hello PeterK, I am currently fixing up some (soft) float related issues on AROS m68k, and would be interested in integrating something similar to your libraries in our IEEE libs. Would you be interested in helping or allow the use of your code for such a thing?

28 June 2018, 08:05	#11
AMIGASYSTEM Registered User Join Date: Aug 2014 Location: Brindisi (Italy) Age: 70 Posts: 8,248	IEEE-Libs v45.881 r14 Update

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)