FPU glitches

[hexaae]

b31
DEMO: The Martini Effect by Flex on 060 emu has some blackscreens not present (smooth effects transitions) with 4.4.0. Visible since very first zoom-in sequence. Rolled back up to b16 but the problem is still reproducible...

EDIT: I was using "64bit host" FPU, using "80bit host" there are no blackscreen moments...

EDIT2: using 64bit host + if I replace OS3.9 LIBS:math#?.library with Peter Keunecke's versions for WinUAE ( http://m68k.aminet.net/package/util/libs/MathLibsUAE ), it also solves blackscreen issues.

[hexaae]

Ikanim (final) http://loonies.dk/productions.asp?group=demos#DEMOS has some performance issues on 060 emu with choppy 3D walking guy animation at 00:21 with WinUAE 4.9.0b31. It runs better with JIT OFF (?!) but still some 3D parts run choppy compared to a very old WinUAE 3.3.0 version which runs equal to youtube videos of this demo on real HW.
A possible issue in 060 cpu/timing emu?

[Toni Wilen]

"Wrong" FPU mode? (80-bit can be much slower, softfloat can be really slow)

[hexaae]

I was using "64bit host", setting "80bit host" will make the demo run defintely better but with many gfx glitches:

[Toni Wilen]

I can't duplicate any glitches or speed issues.
Does same happen when using plain 68020+68882 config? (with "unimplemented FPU emu" ticked)

[hexaae]

After zillion of tests, I've found the following:
1. FPU Host 80bit = horizontal line glitches as in the picture above*
2. when started from a Workbench RTG Z3 P96 3.1.2 (1920x1080x32bit BGRA), even though screen correctly opens in PAL mode from RTG WB, 8/10 times the demo desyncs after the guy walks in from right to left the at the very beggining and demo switches to blurred mode (= walks in ok, then all timings slowdown).
It NEVER happens instead starting the demo from native PAL WB (!), or using screen degraders like "JST EXECUTE IkanimHiRes"
3. 68020+68882 config won't change with FPU Host 80bit.

EDIT:
*re-tested from OS39 fresh install HD setup (using MuLib): glitches reproducible. From OS32 fresh install HD setup: not reproducible (!). Could different mathlibs affect WinUAE 64/80bit Host calcs?

OS3.9 fresh install setup

SysInspector 1.4 - ©1997 by Eric Sauvageau.
===========================================

--- Libraries ---

Address      Name                                Version  Pri  Open Count
-----------------------------------------------------------------------------
$600B6C60    68060.library                       46.6     0     1     
$602AE7E4    asl.library                         45.4     0     1     
$602AB270    bevel.image                         45.7     0     12    
$600B2C0C    bsdsocket.library                   4.1      -50   0     
$602ABE48    button.gadget                       45.6     0     4     
$601A1A48    cgxvideo.library                    42.1     -30   0     
$603A0548    checkbox.gadget                     45.3     0     1     
$602ACC4C    chooser.gadget                      45.7     0     5     
$602ACDA0    clicktab.gadget                     45.3     0     1     
$6020429C    commodities.library                 44.18    0     6     
$6019F958    cybergraphics.library               42.7     -30   5     
$60010EC0    datatypes.library                   44.47    0     0     
$601F3504    diskfont.library                    45.7     0     0     
$6001099C    dos.library                         40.3     0     54    
$602AB358    drawlist.image                      45.0     0     2     
$602685F4    emulation.library                   41.459   -30   1     
$600008D4    exec.library                        45.20    -100  1     
$00004BF4    expansion.library                   40.2     -20   7     
$60161CF0    gadtools.library                    40.4     0     11    
$602AA290    glyph.image                         45.6     0     6     
$600044D4    graphics.library                    40.24    0     1     
$600B7FD0    icon.library                        45.1     0     14    
$601020F0    iffparse.library                    40.1     0     6     
$6041D258    integer.gadget                      45.5     0     1     
$6000A9DC    intuition.library                   40.85    0     1     
$60160860    italiano.language                   38.4     -100  1     
$60249880    jpeg.datatype                       44.4     0     4     
$6000A4A0    keymap.library                      40.4     0     1     
$6039B638    label.image                         45.4     0     3     
$60008C90    layers.library                      40.1     0     1     
$602AEC30    layout.gadget                       45.9     0     4     
$602AA3F8    listbrowser.gadget                  45.12    0     5     
$600BB0E8    locale.library                      44.6     0     0     
$60162298    mathieeedoubbas.library             45.3     0     4     
$602A8C50    mathieeedoubtrans.library           45.5     0     1     
$600DD368    mathieeesingbas.library             40.4     0     1     
$600DCC20    mmu.library                         46.22    0     1     
$600BCFBC    muiconfig.library                   19.20    0     0     
$60204544    muigfx.library                      19.9     0     2     
$60231BA0    muilowlevel.library                 19.9     0     4     
$601F9444    muimaster.library                   19.35    0     2     
$60242688    muirexx.library                     19.9     0     0     
$603AB810    penmap.image                        45.4     0     3     
$60249B2C    picture.datatype                    45.17    0     1     
$602AE67C    radiobutton.gadget                  45.4     0     1     
$6041CFA8    requester.class                     42.10    0     1     
$602A9ED0    resource.library                    44.103   0     1     
$601BAC04    rexxsyslib.library                  44.1     127   3     
$60197914    rtg.library                         42.747   -70   0     
$60231E3C    screennotify.library                1.0      0     1     
$602ABDD0    scroller.gadget                     45.9     0     2     
$60396680    space.gadget                        45.1     0     1     
$602AE538    speedbar.gadget                     45.1     0     1     
$602AC880    string.gadget                       45.16    0     4     
$603A3018    titlebar.image                      40.14    0     1     
$6000E904    uaegfx.card                         3.4      -50   1     
$60000410    utility.library                     40.1     0     59    
$601A1B44    vilintuisup.library                 2.9      -30   0     
$60247430    Virtgroup.mui                       19.10    0     1     
$602AEB8C    window.class                        45.16    0     3     
$602A8BA8    wizard.library                      42.100   0     1     
$601EDC94    workbench.library                   45.127   0     8     

62 libraries.
-----------------------------------------------------------------------------

P.S. Ignore the assumption: 80bit Host = runs well, 64bit runs slow I said above. It was just random (90% running slow) starting from RTG WB.
Instead, when started from WB PAL, even in 64bit host mode runs fine (and without 80bit host glitch).

[hexaae]

This is interesting: if I replace OS3.9 LIBS:math#?.library with Peter Keunecke's versions for WinUAE ( http://m68k.aminet.net/package/util/libs/MathLibsUAE ), it solves 80bit Host FPU gfx glitches (visible in msg #3 of this thread) with IkanimHires!

SysInspector 1.4 - ©1997 by Eric Sauvageau.
===========================================

--- Libraries ---

Address      Name                                Version  Pri  Open Count
-----------------------------------------------------------------------------
$60096A78    68060.library                       46.5     0     1     
$6040DB40    akJFIF.datatype                     45.95    0     1     
$6069A808    akPNG.datatype                      45.83    0     1     
$60475E78    arexx.class                         44.1     0     1     
$6048C09C    arp.library                         40.2     0     1     
$601148B4    asl.library                         45.4     0     5     
$604169AC    asyncio.library                     40.6     0     2     
$6047D250    bevel.image                         45.7     0     12    
$600918B4    bsdsocket.library                   4.1      -50   0     
$60481308    button.gadget                       45.6     0     5     
$60368A58    cgxvideo.library                    42.1     -30   0     
$60601720    checkbox.gadget                     45.3     0     2     
$60600284    chooser.gadget                      45.7     0     6     
$6037DF10    clicktab.gadget                     45.3     0     1     
$6011469C    commodities.library                 44.18    0     16    
$603BA760    cybergraphics.library               42.7     -30   7     
$603BF478    datatypes.library                   44.47    0     0     
$601149F4    diskfont.library                    45.7     0     0     
$600108E4    dos.library                         40.3     0     97    
$60471BB8    drawlist.image                      45.0     0     2     
$60276EF4    emulation.library                   41.465   -30   1     
$6048C13C    environment                         2.0      -128  0     
$600008A0    exec.library                        40.10    -100  5     
$00003674    expansion.library                   40.2     -20   6     
$6032D6C0    expat.library                       4.1      0     0     
$60115558    gadtools.library                    40.4     0     21    
$60482A88    glyph.image                         45.6     0     6     
$60004574    graphics.library                    40.24    0     1     
$602650AC    icon.library                        51.4     0     23    
$601194E0    iffparse.library                    40.1     0     11    
$600023D0    ilbm.datatype                       47.4     0     2     
$6048C458    integer.gadget                      45.5     0     2     
$6000AA7C    intuition.library                   40.85    0     1     
$6009CCB0    italiano.language                   38.4     -100  1     
$6000A540    keymap.library                      40.4     0     1     
$605D83F8    label.image                         45.4     0     4     
$60008D24    layers.library                      45.30    0     1     
$60472FA8    layout.gadget                       45.9     0     5     
$605EEEC0    listbrowser.gadget                  45.12    0     6     
$6025BAB8    locale.library                      44.6     0     0     
$6069A8B8    mathffp.library                     40.1     0     1     
$603FD390    mathieeedoubbas.library             45.881   0     2     
$605D21F0    mathieeedoubtrans.library           45.881   0     1     
$6009DB30    mathieeesingbas.library             45.881   0     1     
$6032D75C    mcpgfx.library                      4.4      0     1     
$600F8674    mcpsupport2.library                 7.1      0     1     
$6009D260    mmu.library                         47.1     0     1     
$60601544    muiconfig.library                   19.20    0     0     
$604645D4    muigfx.library                      19.9     0     2     
$605D8A08    muilowlevel.library                 19.9     0     4     
$604716FC    muimaster.library                   19.35    0     2     
$605C9EF0    muirexx.library                     19.9     0     0     
$60692518    penmap.image                        45.4     0     3     
$60416524    picture.datatype                    45.17    0     3     
$606EAB84    powerpacker.library                 37.7     0     0     
$6037E3D4    radiobutton.gadget                  45.4     0     1     
$605EA190    req.library                         2.7      0     1     
$6025C088    reqtools.library                    39.3     0     0     
$606EBB28    requester.class                     42.10    0     1     
$6048C238    resource.library                    44.103   0     1     
$603FD5D0    rexxsyslib.library                  45.7     100   7     
$602774A4    rtg.library                         42.797   -70   0     
$605DD554    screennotify.library                1.0      0     1     
$6047E858    scroller.gadget                     45.9     0     2     
$606316C8    space.gadget                        45.1     0     1     
$6037DFF8    speedbar.gadget                     45.1     0     1     
$605F9FC0    string.gadget                       45.16    0     5     
$603BF6D8    svx.datatype                        41.6     0     1     
$603C02C0    tapedeck.gadget                     40.3     0     1     
$605C9DB8    titlebar.image                      40.14    0     2     
$6000E84C    uaegfx.card                         3.4      -50   1     
$60003D08    utility.library                     40.1     0     88    
$603C0168    v41sound.datatype                   42.0     0     1     
$60368B54    vilintuisup.library                 2.9      -30   0     
$6048C668    Virtgroup.mui                       19.10    0     1     
$6071F624    vnc.library                         42.103   0     0     
$6046D27C    window.class                        45.16    0     4     
$604694C8    wizard.library                      42.100   0     1     
$602ACA1C    workbench.library                   45.132   0     11    

79 libraries.
-----------------------------------------------------------------------------

[Toni Wilen]

Does same glitches also appear in softfloat mode?

[hexaae]

Can't reproduce anymore with latest public b32. Host 80bit has no glitches, tested 10-20 times.

EDIT:
Sorry, I'm afraid it was a nightly-build-only bug: I re-downloaded public b31 to be sure and wasn't reproducible even there.

[hexaae]

Originally Posted by hexaae

b31
DEMO: The Martini Effect by Flex on 060 emu has some blackscreens not present (smooth effects transitions) with 4.4.0. Visible since very first zoom-in sequence. Rolled back up to b16 but the problem is still reproducible...

EDIT: I was using "64bit host" FPU, using "80bit host" there are no blackscreen moments...

EDIT2: using 64bit host + if I replace OS3.9 LIBS:math#?.library with Peter Keunecke's versions for WinUAE ( http://m68k.aminet.net/package/util/libs/MathLibsUAE ), it also solves blackscreen issues.

b32 still affected using WinUAE 060 64bit Host (+ Advanced JIT settings: Cache, FPU Support=On) with OS39 mathlibs:




No problems with The Martini Effect (Flex) demo when using same WinUAE settings as above + mathlibs from WB3.1* (!!) or from Peter Keunecke in a OS39 environment.

*
mathieeedoubbas.library 38.2 (24-Gen-1992)
mathieeedoubtrans.library 37.1 (21-Gen-1991)
mathieeesingtrans.library 37.1 (21-Gen-1991)
mathtrans.library 37.1 (22-Gen-1991)

Switching to '80bit host' also solves the issue with the demo, whatever mathlibs I use.

[Toni Wilen]

When did it start? There should not have been any FPU or CPU changes since long time..

Originally Posted by hexaae

b32 still affected using WinUAE 060 64bit Host (+ Advanced JIT settings: Cache, FPU Support=On) with OS39 mathlibs:




No problems with The Martini Effect (Flex) demo when using same WinUAE settings as above + mathlibs from WB3.1* (!!) or from Peter Keunecke in a OS39 environment.

*
mathieeedoubbas.library 38.2 (24-Gen-1992)
mathieeedoubtrans.library 37.1 (21-Gen-1991)
mathieeesingtrans.library 37.1 (21-Gen-1991)
mathtrans.library 37.1 (22-Gen-1991)

Switching to '80bit host' also solves the issue with the demo, whatever mathlibs I use.

[hexaae]

I went back to versions older than b16 and... it happens since b1 in this thread.
4.4.0 not affected:


P.S.
Tested b32 with JIT Off and was still reproducible.

[Toni Wilen]

Originally Posted by hexaae

I went back to versions older than b16 and... it happens since b1 in this thread.
4.4.0 not affected:


P.S.
Tested b32 with JIT Off and was still reproducible.

Lets try with less variables.. b1 only has JIT update. Could you do following test:

JIT off

4.4, b1, b32: 64-bit host, 80-bit host, softfloat. Which combination causes glitches, which does not.

[hexaae]

Results are always the same:

OS3.9 settings:
mathieeedoubbas.library 45.3 (23-Nov-2000)
mathieeedoubtrans.library 45.3 (23-Nov-2000)
mathieeesingtrans.library 45.3 (23-Nov-2000)
mathtrans.library 37.1 (22-Gen-1991)
Always JIT OFF for all tests.

WinUAE executables:
4.4.0: affected only 64bit host (I was wrong here: must have been in 80bit host mode when it worked, or I was using OS31 mathlibs)
b1: affected only 64bit host
b32: : affected only 64bit host

Not reproducible only with 80bit host or 80bit softloat (and Cache FPU OFF in this case), or when I use PK's or 3.1's mathlibs (yes, even 64bit in this case causes no issues).

Seems to have started prior to 4.4.0...
EDIT: just reproduced even with 4.3.0: 64bit mode + OS3.9 mathlibs. 80bit (soft or not) or using other mathlibs not affected.
EDIT2: 4.1.0 and 4.0.1 also affected in 64bit host mode! Must be there from beginning of new FPU 64bit host maths?

[hexaae]

Argh! Experimenting with the 64/80bit host issue I was able to reproduce this glitch with Ikanim and 4.9.0b32!
I probably forgot in the message above I had 3.1 mathlibs in LIBS (so was not affected)*

Seems affected in 80bit mode only when using OS3.9/3.2 mathlibs... Will do some more tests....

EDIT:
* or I switched 64-80bit runtime, so the glitch was not reproducible, see msg #10 below

[hexaae]

IkanimHires:

Restart + 80bit host = glitches (notice: if you F12 + switch from 64bit host to 80bit host, it won't happen. You have to Restart and run the emu from the start with 80bit host)
64bit host and 80bit soft = no glitches

Tried with both OS31 and OS39 mathlibs, always the same results.
Ikanim log with glitches: https://pastebin.com/m7MGsjr3

[Toni Wilen]

Thanks.

It is starting to sound more like demo bug (bad FPU initialization) than emulator problem, if it also happens in softfloat mode.

Originally Posted by hexaae

Results are always the same:

OS3.9 settings:
mathieeedoubbas.library 45.3 (23-Nov-2000)
mathieeedoubtrans.library 45.3 (23-Nov-2000)
mathieeesingtrans.library 45.3 (23-Nov-2000)
mathtrans.library 37.1 (22-Gen-1991)
Always JIT OFF for all tests.

WinUAE executables:
4.4.0: affected only 64bit host (I was wrong here: must have been in 80bit host mode when it worked, or I was using OS31 mathlibs)
b1: affected only 64bit host
b32: : affected only 64bit host

Not reproducible only with 80bit host or 80bit softloat (and Cache FPU OFF in this case), or when I use PK's or 3.1's mathlibs (yes, even 64bit in this case causes no issues).

Seems to have started prior to 4.4.0...
EDIT: just reproduced even with 4.3.0: 64bit mode + OS3.9 mathlibs. 80bit (soft or not) or using other mathlibs not affected.
EDIT2: 4.1.0 and 4.0.1 also affected in 64bit host mode! Must be there from beginning of new FPU 64bit host maths?

[hexaae]

80bit softfloat is not affected, as well as in 80bit host (this one causes glitches in Ikanim though ;( )

[PeterK]

Originally Posted by hexaae

...if I replace OS3.9 LIBS:math#?.library with Peter Keunecke's versions for WinUAE, it solves 80bit Host FPU gfx glitches... From OS32 fresh install HD setup: not reproducible (!). Could different mathlibs affect WinUAE 64/80bit Host calcs?

Yes, the mathlibs are setting the precision and the rounding mode globally for a task when it opens these libs. And there are indeed differences between all available versions of mathlibs. Most of these libs set the precision to 32 bit Float for the mathieeesingXXX.libs and to 64 bit Double for the mathieeedoubXXX.libs, older libs may use "round to zero", OS 3.2 uses "round to nearest even" now, I guess. My libs always set the precision to 80 bit Extended or LongDouble, and always to "round to nearest", only the Fix() function uses "round to zero", but restores "round to nearest" afterwards.

Usually, demos are using inline FPU instructions instead of calling mathlibs functions for their main routines. But sometimes these tasks are calling the mathlibs for some other reasons, too. This can happen when they've used a compiler for the GUI or configuration page or whatever or it also could be another library that is used by the demo task, which opens the mathlibs. Then it also depends on which mathlibs are opened last; i.e. the mathieeesingXXX.libs could reduce the precision to 32 bit Float and round to zero in the worst case for all the following FPU instructions of the demo task.

If you are starting the demo from a shell then even the shell or console or another instruction that was executed before could have opened the mathlibs already (maybe ?) and the demo will be the same task later with the same FPU configuration.

Check your demo files with Multiview and the exe.datatype or with a hex viewer for the mathieee libs.

[hexaae]

Hard to see in demos using exotic cyrpted crunchers not found by usual C:Xfddecrunch... but I've found indeed mathieeedoubbas traces in The Martini Effect, affected with a similar issue