Re: [hatari-devel] Emulated DSP speed

[ Thread Index | Date Index | More lists.tuxfamily.org/hatari-devel Archives ]


Le 06/11/2016 à 19:46, Eero Tamminen a écrit :
Hi,

On 11/05/2016 01:37 PM, Douglas Little wrote:
\o/   great news. well done making the release.

I have just one real issue with the release so far, and maybe also some
good news hidden inside that....

It is known that Hatari Falcon is somewhat hungry to run on modern PCs
when
the DSP is busy - its a single threaded app and easily saturates one
core.

I have measured the DSP in Hatari 2.00 at 32mips. This is 2x the normal
performance in a Falcon.

But this also means the host PC/Mac is doing 2x as much work as necessary
per second. I think reducing the DSP to the expected 16mips should
alleviate some of the choppyness issues and performance warnings still
visible on 3ghz CPUs.

I've seen some odd performance measurements from the DSP over recent
versions of Hatari but this one reads exactly 32.08mips, which means the
DSP is now getting exactly 2x the combel clock, one way or another.

Put another way - for every 2 cycles the emulated MFP experiences, the
DSP
sees only 1 elapsing, instead of 2.

When I ran Centek's "Centurbo Benchmark", that gave following
Falcon results...

2.0, WinUAE CPU core:
* FPU: 920 MHz
* DSP: 64 MHz
* CPU: 27 MHz

1.9, WinUAE CPU core:
* FPU: 812 MHz
* DSP: 56 MHz
* CPU: 27 MHz

2.0, oldUAE CPU core:
* FPU: 403 MHz
* DSP: 62 MHz
* CPU: 78 MHz

1.9 & 1.8, oldUAE CPU core:
* FPU: 403 MHz
* DSP: 32 MHz
* CPU: 78 MHz


I'll play with it a bit more over the next weeks but this is the only
thing
that stands out so far.

Have you considered adding FPU & CPU measurements against
that clock same too?


Hi

I don't think it's worth spending too much time on these measures for now ; as long as 68030 is not cycle exact in Hatari, you can get many different results.

Nicolas




Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/