|Re: [hatari-devel] Long mail : new cpu cores + 2 questions.|
[ Thread Index |
| More lists.tuxfamily.org/hatari-devel Archives
In complement : cpuemu_21 is really not completly 'cycle_exact'.
Example (taken from Eko_system):
cpu video_cyc= 132 132@ 0 : 0005C6C6 6708 BEQ.B
#$00000008 == $0005C6D0 (T)
cpu video_cyc= 134 134@ 0 : 0005C6D0 4ab8 04be TST.L
cpu video_cyc= 150 150@ 0 : 0005C6D4 6706 BEQ.B
#$00000006 == $0005C6DC (F)
cpu video_cyc= 150 150@ 0 : 0005C6D6 4eb0 01e1 04be JSR.L
(D0.W*1+1214)+0 == $0011a1bc
cpu video_cyc= 168 168@ 0 : 0011A1BC 48e7 fffe MOVEM.L
Here, BEQ is 0 cycles (when condition is false).
There are many examples like this in winuae new core CPU.
Should I correct them ? (and fork from the original winuae cores ?)
Le 31/10/2011 12:04, Laurent Sallafranque a écrit :
I've played around with the new winuae cores for the last days.
There are 6 cores available :
cpuemu_0 : a generic 68x00 cpu (used when no prefetch nor cycle
exact config in the GUI), both for 68000 and 68030 emu
cpuemu_11 : same as before, but for prefetch mode and 68000 only
cpuemu_12 : this one is for the 68000 cycle exact mode only
cpuemu_20 : this one is for the 68020 emulation prefetch + cycle exact
cpuemu_21 : this one is for the 68030 emulation prefetch + cycle exact
cpuemu_31 : this one is for 68040 and 68060 emulation
Actually, I can let run quite correctly the following cores:
cpuemu_21 (I'll have to send a patch soon for this one)
2 cores are still not running :
cpuemu_12 (very strange code behing it)
cpu_emu31+MMU activated (TRY/CATCH/RETRY) to recode first
First : Core updating.
I've tried to upgrade the cores to latest winuae source, but I broke
There are few changes in the cores between the 2 versions, but
patching must be done carefully.
I'll try it another time later.
ST emulation :
For ST emulation, the closest core to our actual oldcpu core are
cpuemu_0 or cpuemu_11 (very close to each other, only prefetch differs)..
If we want to retrieve the actual old cpu compatibility, we have 2
- using cpuemu_11 as the main 68000 core, and patch gencpu.c to
rollback the different patchs Nicolas added there.
For example, getDivs68kCycles and getDivu68kCycles are only called
with cpuemu_12, so, we'll have to reintroduce them into gencpu.c
- let the cpuemu12 running (it may be harder, and I don't understand
I think we should only concentrate on 1 cpu and forget the others, as
Nicolas has proved that the cpuemu11 can reach a good level in accuracy..
I can do this job if you want to try to get as close as possible to
the actual CPU ( I don't think it'll be that long and difficult, as
we've got the old gencpu as example).
Falcon emulation :
2 cores are running :
cpuemu_0 : the generic 68000 core is used for 68030 and 68030 +
prefetch emulation. It runs 68030 data and register caches and 68030
the problem with this core is that is uses the 68000 cycles. I get the
best results with this one actually (I run it with prefetch).
cpuemu_21 : the cycle exact core. I've got some results with it now,
but it seems to be less accurate than the cpuemu_0 core. Some demo
freeze with this one, but runs well with cpuemu_0.
It's not completly "cycle exact", as for example the divs, divu and
divl cycles are computed with the worst cycles value instead of the
It let works 68030 caches and 68030 MMU too.
My suggest: for Hatari emulation, we could keep only 3 cores :
cpuemu_11 for ATARI 68000 core and repatch it to reach the actual
level of compatibility.
cpuemu_0 for FALCON emulation with a lot of changes in it (cycles,
...) to let it be closer to a 68030 or cpuemu_21 (but the code is much
more complex and accuracy is lower for now).
We would only run hatari in accuracy mode for the ST, in best accuracy
for the 68030 Falcon and generic accuracy 68060 for later.
No more GUI with a lot of options like prefetch, ... (I don't think
the speed difference is noteable with or without prefetch nowadays)
No more 68000 running with a MMU or a DSP.
We could only choose ST, ST, TT, FALCON (and later accelerated falcon)
and choose the with/without copro or DSP.
Or we can continue to offer many parameters in the GUI and in the
emulator, but 6 cores are a great effort to maintain and cpuemu_0 is
used both for generic 68000 and Falcon, wich means a lot of "IF" to
add everywhere in gencpu.c to keep accuracy for all cpus.
One day or the other, we'll have to make some choice with the new core
and make some improvements and changes into it.
If I keep it "compatible" with winuae without changing anything,
falcon emu won't improve anymore.
But in the same time, I think I reach the limits of my knowledges here
with Motorola CPU.
Falcon gets quite well emulated, but there are still a lot of little
problems here and there (accuracy, sound noise, ...)
I think DSP is running quite well, everything is emulated. Crossbar is
I lack of documentation (as I had for the DSP for example).
And I'm not such an expert of 68x00 processors.
I can patch the 68030 cycles according to motorola's docs (or better,
Mikro's doc when he finished it) to help.
Maybe you could give me ideas of things to do in hatari to improve it.
Other point : videl emulation:
I've done some work with videl emu (I still have the code here on my
But separating video.c and videl.c implies huge changes I don't want
to do lonely.
video.c in integrated everywhere in hatari (psg.c, mfp.c, ...)
It would be nice to have an interface to call in all parts of hatari
that would run either with video.c or videl.c.
I would have to reintegrate timer A, B, C, D, HBL, VBL into videl.c
(I've already done some of them).
I don't want to add "if FALCON ..." everywhere in the code (as I've
done into my actual code).
It was easy to separate the sound (crossbar VS stedma sound), but
video is too glued everywhere in ST mode.
Last point : cpu speed.
Is dividing the cycles by 2 (or 4) to let the 68000 run at 16 or 32
Mhz is the only way to do it ?
My 68030 falcon should run natively at 16 Mhz (1 instruction cycles
shoudn't be divided).
I should have more cycles for each VBL, crossbar should run with the
68030 cycles (not divided by 2 nor 4), same for DSP.
Isn't there a "more beautiful" way to do the frequency emulation
(keeping the cycles intact and having twice more cycles for each VBL
Maybe this point is perfect and doesn't need any change, but it always
"annoys" me to have the 68030 cycles divided by 2.
I stop here.
Thanks to read it to the end (and sorry for my approximate english)
hatari-devel mailing list