WinUAE QEMU PPC

[Toni Wilen]

Originally Posted by bernd roesch

I have notice with this asm test program that context switch from 68k to PPC and back is very slow. around 40 calls on my 4*2.8 GHZ system. original PPC HW reach 3000 context switches /sec with this amiblitz program. source can work too with a 68k assembler. i reduce loop size to 100 so it take not too long for the attached binary. wy is this so slow ?. on real PPC it is slow because caches of PPC and 68k need flush this cost lots time. I attach shell prog to test.

When I run it, PPC thread starts (so it does something with PPC) but it also exits immediately.

[bernd roesch]

this is a version that do 30000 loops. it start a PPC program and the PPC program quit immidiatly. the problem on real ppc HW was because the switch between 68k to ppc and from PPC to 68k was slow because of cache flushes. but strange wy this is in winuae emulation even more slow. if it can not easy make faster ok, but maybe it is a simple problem or maybe this test do something winuae not like. but it is done as th warpup specs do . i test a little and the PPC assembler in amiblitz work partly. its not fault of winuae. this is a more complex example source. binary i attach is from simple example. for this i currently have no binary

WBStartup
For i.l= 1 To 3000
RunPPc ?ppcprog,?_base,$12345678 ;use warpup (needs wizard.library)
Next
End
_base: Dc.l 1 ;_base is defined in all.bb2
UsePPC
#storereg=8 ;=r8 or fp8
ppcprog
mr r21,r2

lfs f0,dat4-_base(r2)
;la r1,lbl1
;liw r8,$dff180 ;Changes backgroundcolor of aga Screen
;sth r9,0(r8)

lwz r31,ppcdat1-_base(r2) ;same as above
blr
ppcdat1: Dc.l 2
dat2: Dc.s 2 ;float single
fpudat3; Dc.d 3.14 ;float double
dat4: Dc.d 25.25
lbl1: Dc.d 20

[Romanujan]

If it helps, there is a benchmark for measuring PPC/68k context switching, available here: http://cosmosamiga.free.fr/download/measurecon.lha

And here are some screenshots with real hardware results: http://leblogdecosmos.blogspot.com/2...-161beta2.html

In short:

- Cosmos real HW, unoptimized library: 633 microseconds per switch
- my Core i7 4790k, Linux, 64-bit FS-UAE 2.9.0, unoptimized library: around 8500-9200 microseconds per switch

[bernd roesch]

Originally Posted by AMIGASYSTEM

I did not understand what you said, the only library that works on AFA OS PPC is "Wazp3D.library-full".

On AFA OS 68k standard works well "Wazp3D.library-winuae" and "Wazp3D.library-full"

I install newest waz3d. 68k versions crash with ramlib error too(0x8000000A). without PPC enable it work ok. I look and see a ramlib patch that increase stack size to 8192. its http://aminet.net/package/util/sys/StackAid

source is in too. does not help on my system. scout show more stack for ramlib. maybe help on your system. or need more stack. can you tell which stack your ramlib task without this patch ?. if your ramlib stack is larger than 8192 bytes than you need this not

[AMIGASYSTEM]

Ciao bernd, I'm sorry if I'm wrong, I'm not an expert in this


I think I have a more recent Warz3D THIS


I think I have installed well "StackAid" (see startup-sequence attached),
Now I receive a different error, I also attach scout results (Ramlib) see screeenshot

[Toni Wilen]

Originally Posted by Romanujan

- Cosmos real HW, unoptimized library: 633 microseconds per switch
- my Core i7 4790k, Linux, 64-bit FS-UAE 2.9.0, unoptimized library: around 8500-9200 microseconds per switch

I got this:

Starting 10000 context-switch...done!
Ellapsed: 2 sec 985494 microsec, ~298 microsec each.

[AMIGASYSTEM]

If it can be important the demos Warp3D PPC on OS4.1 Emulated they work well !

[bernd roesch]

Originally Posted by Toni Wilen

I got this:

Starting 10000 context-switch...done!
Ellapsed: 2 sec 985494 microsec, ~298 microsec each.

can you attach your winaue config and which warpup libs you use ?. do you use amikit or other aos install ? .maybe there is a setting in winuae that do huge slowdown. with this program is very slow too

Number of context switches to be measured (max. 65535):1000
Starting 1000 context-switch...done!
Ellapsed: 24 sec 679084 microsec, ~24679 microsec each.
12.h0:> "Ram Disk:measurecontextwos"


i have cpu idle for 68k and ppc all full left. when i look with scout CPU load for this test it show around 1.5% so 68k seem wait lots.

[bernd roesch]

Originally Posted by AMIGASYSTEM

Ciao bernd, I'm sorry if I'm wrong, I'm not an expert in this


I think I have a more recent Warz3D THIS


I think I have installed well "StackAid" (see startup-sequence attached),
Now I receive a different error, I also attach scout results (Ramlib) see screeenshot

I have the same wazp version use you use. your ramlib stack is set at 8 kb. strange that only i get the ramlib errors.

can you please do the context switch tests that are here list, to see how fast your ppc context switch. what cyberstorm rom version you have ?. I have the .71

the free memory that is always free from AFA is because the better aros/afa mem handler. it show the program do wrong. with 68k i never get such a requester. so can not wazp problem. free memory that is freed before is not notice in default AOS. but such program mistake is very dangerous it can do bad things, if for example another task alloc memory and get this range now. and if later gearsppc want free this block again, then amiga can do crazy things because gearsppc free the memory and another task can alloc the memory and modify it, so the task that alloc the memory first get then wrong data.

[Toni Wilen]

Originally Posted by bernd roesch

can you attach your winaue config and which warpup libs you use ?. do you use amikit or other aos install ? .maybe there is a setting in winuae that do huge slowdown. with this program is very slow too

Number of context switches to be measured (max. 65535):1000
Starting 1000 context-switch...done!
Ellapsed: 24 sec 679084 microsec, ~24679 microsec each.
12.h0:> "Ram Disk:measurecontextwos"


i have cpu idle for 68k and ppc all full left. when i look with scout CPU load for this test it show around 1.5% so 68k seem wait lots.

Attached. It is very basic config, WB is 3.1 with some minimal changes. Above mentioned aminet v40 (which replaced some other version I didn't note when I trying to duplicate your crash)

[AMIGASYSTEM]

Originally Posted by bernd roesch

can you please do the context switch tests that are here list, to see how fast your ppc context switch. what cyberstorm rom version you have ?. I have the .71

My configuration on is the same as in your screenshot

[thellier]

Hello

>On AFA OS 68k standard works well "Wazp3D.library-winuae" and "Wazp3D.library-full"

True
Wazp3D.library-full is named like this as it contain also the renderer in the code (the so called soft3d library) so is pure 68k software only

Wazp3D.library-winuae call an external soft3d library that call an x86 soft3d.dll that use hardware opengl32.dll. So this is 68k/x86/hardware rendering (fast & nice)

Wazp3D.library-os4 is same as Wazp3D.library-full but compiled for OS4/ppc so is pure ppc software only
There is no hardware rendering here as WinUAE-ppc dont support native calls to x86 so Wazp3D works in OS4 but is slow

As you have noticed there is no Wazp3D.library-wos for warpos ppc/os3
It is unclear if there is a way to make warpos ppc/os3 progs works with Wazp3D
(I mean I dont know how but it may be possible)

Voila

[AMIGASYSTEM]

Ciao thellier and thanks for the information, i on the AFA OS 68k Standard do not know any significant difference in speed between "Wazp3D.library-winuae" and "Wazp3D.library-full".

Yes on OS4 Wazp3D, it is much slower, but may depend on the slow emulation of OS4, it is much slower, but may depend on the slow emulation of OS4


Colliding Stars AFA OS3 Vs OS4 1 Test

[ Show youtube player ]

[bernd roesch]

Have you copy the soft3d.dll to your windows native libs folder ?. must be in winuae_dll folder in folder you have winuae

I notice it seem newest winuae log file do not output when a amiga native lib as sosft3d.dll is load. this is usefull to see if it is really load. maybe on my system ppc crash because it use native soft3d.

I look your video again and i notice that the gears demo run on your PC very slow. On my system gears demo in window run smooth with or without soft3.dll

to work fast the JIT need set to direct. but this work only with 68k and not ppc. so 68k run lots faster

when i switch JIT to indirect mode gears demo windows run much slower

[AMIGASYSTEM]

Originally Posted by bernd roesch

Have you copy the soft3d.dll to your windows native libs folder ?. must be in winuae_dll folder in folder you have winuae

Yes, I have the library soft3d.dll in the WinUAE directory

I look your video again and i notice that the gears demo run on your PC very slow. On my system gears demo in window run smooth with or without soft3.dll

calculates that mine is an old DualCore Windows7 32Bit, add slowdown video recording , redone video test with gears demo

[bernd roesch]

Originally Posted by Toni Wilen

Attached. It is very basic config, WB is 3.1 with some minimal changes. Above mentioned aminet v40 (which replaced some other version I didn't note when I trying to duplicate your crash)

thanks for config. i do a quick test without startup-sequence and your config have a very big speed boost. more than 10* faster. in this test come only on first start a message that env is not find. can ignore and then second testrun there come no such message. i do test with same hardfile and with my install too without startup sequence. so can not be a AOS install problem, because no setpatch etc is load. on my config only 1000 times need a little more as 20 sec

so i need find what the reason is for this. i notice you disable JIT complete. but this is not he reason. when i disable JIT it get only slightly 10-20% faster.

see screenshot with the result of your config.

[Romanujan]

Good to know, some testing will be needed to find out what slows the context switching down. First results: on BlizzardPPC 68040 emulation, the 68040.library from AmigaOS 3.9 installation CD is faster than the ROM built-in or the one from Thomas Richter MMU package - my context switching time went down by 1500-2000 microseconds. But still there has to be something more...

[AMIGASYSTEM]

I tried Blizzard 1260 and the system became faster, Gears PPC it did not work but I have not received any error

[bernd roesch]

The rtg board Ram size can not larger as 32 MB and no other Z3 ram or other ram should add. then context switch is more than 20* faster. If set in Toni's config rtg board size to 64 MB then the huge slowdown happen. same happen if Z3 ram is add

[Romanujan]

On my FS-UAE results are slightly different - GFX memory size (even 128M) does not increase the context switch time, 32MB (or less) of Z3 memory is also fine. 64MB or more Z3 memory => slowdown (the more memory, the slower the context switching).

Checked using OS 3.9 BB2 + many patches, BPPC 68040, Thomas Richter 68040 library.