On Wed, Feb 19, 2020 at 8:28 PM BALATON Zoltan <
address@hidden> wrote:
On Wed, 19 Feb 2020, Howard Spoelstra wrote:
> I tested with the current ppc-for-5.0 branch and with v1 of the hardfloat
> patches applied on top of that. There is a noticeable speed improvement in
> Linux and OSX hosts. Windows 10 host doesn't seem to be impressed at all. I
> saw no obvious glitches so far. The fpu performance on OSX hosts seems very
> slow. This was not always the case in the past, when it was on par with
> Linux performance.
Interesting, thanks for the measurements.
> Below are my results.
>
> Best,
> Howard
>
> Host Linux (Fedora 31):
> Mac OS tests: 9.2 with MacBench 5.0
> Baseline(100%): G3 300Mhz
> 5.0 branch + hardfloat patches: cpu 193%, fpu 86%
> 5.0 branch: cpu 188%, fpu 57%
Here there's a difference in cpu value before and after patch which I
can't explain (only changed FPU stuff so it should not change others) but
also not seen in other measurements so this could be some external
influence such as something else using CPU while running test? Unless this
happens consistently I'd put it down to measurement error.
Yes, I would put that cpu value down to some fluctuation in the test
> Mac OSX tests: 10.5 with Skidmarks 4.0 test
> Baseline(100%): G4 1.0Ghz.
> 5.0 branch + hardfloat patches: Int:131 FP:11 Vec:15
> 5.0 branch: Int:131 FP:9 Vec:11
>
> Host OSX Sierra:
> Mac OS tests: 9.2 with MacBench 5.0
> Baseline(100%): G3 300Mhz
> 5.0 branch + hardfloat patches: cpu 199%, fpu 66%
> 5.0 branch: cpu 199%, fpu 40%
> Mac OSX tests: 10.5 with Skidmarks 4.0 test
> Baseline(100%): G4 1.0Ghz.
> 5.0 branch + hardfloat patches: Int:129 PF:11 Vec:14
These values seem to match Linux measurement above so don't seem slower
although MacOS9 seems to be slower (66 vs. 86) so either this depends on
the ops used or something else.
Yes, the baseline speed for the fpu in Mac OS 9.2 is relatively low.
> 5.0 branch: Int:129 FP:8 Vec:9
>
> Host Windows 10:
> Mac OS tests: 9.2 with MacBench 5.0
> Baseline(100%): G3 300Mhz
> 5.0 branch + hardfloat patches: cpu 180%, fpu 54%
new run
5.0 branch + hardfloat patches: cpu 184%, fpu 54%
> 5.0 branch: cpu 199%, fpu 40%
new run
5.0 branch: cpu 184%, fpu 56%
It seems I misreported (copy/past without changing the values) the earlier Windows-based results with Mac OS 9.2
guest. As said above (and this now seems to confirm) Windows is not impressed at
all and perhaps a bit slower even.
Windows builds are particularly sensitive to any other activity on the system. Moving the qemu window drops performance considerably. Perhaps due to SDL not running in its own thread?
Here there's again difference in cpu value but the other way so maybe if
the cause is external CPU usage then this again may be an outlying
measurement? You could retake these two to verify if you get same numbers
again. The fpu value does seem to improve just not as much as the others
and it's also lower to start with. I wonder why.
> Mac OSX tests: 10.5 with Skidmarks 4.0 test
> Baseline(100%): G4 1.0Ghz.
> 5.0 branch + hardfloat patches: Int:130 FP:9 Vec:10
> 5.0 branch: Int:130 FP:10 Vec:11
>
> All tests done on the same host with v1 of the hardfloat patches
> Intel i7-4770K at 3.50Ghz. 32Gb memory
> All guests set to 1024x768 and "thousands" of colors.
Does it mean this host machine were rebooted into these OSes or these were
run in a VM. In case using VM, were all three running in VM or one was on
host (I'd guess OSX host with Linux and Windows VMs).
> Linux and OSX (with brew) use default compilers.
> Windows build cross-compiled from Fedora with x86_64-win64-mingw32
I assume Linux and OSX were 64 bit builds, is Windows 32 bit or 64 bit
exe?
No virtualisation. I run all on the same bare metal, so booted into
these three separately from separate SSDs. You might guess OSX Sierra is
running on less "official" hardware and you would be right. All qemu
builds were 64 bit.
Best,
Howard