Re: [hatari-devel] relative perf between 030 / DSP / CPU?

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


Le 24/06/2025 à 12:43, Eero Tamminen a écrit :
Hi,

On 24.6.2025 12.39, Nicolas Pomarède wrote:
No, 040/060 is not cycle accurate at all and will likely never be (seeing how difficult it is to get CE mode on 68020/30 without any available doc of the inner working of the cpu, there's no chance we will ever have it on 040/060)

For now, IIRC all 040/060 instructions are considered as taking 1 cpu cycle, which gives the expected speed boost compared to 020/030 but is of course not accurate at all.

Ok, that's good to know!

Btw. Have you planned any improvements to the FPU emulation accuracy?

Centurbo Bench reports following numbers for emulated 16Mhz Falcon with Hatari Git version, in 640x400@2-bit and 640x4@16-bit Falcon resolutions (under EmuTOS v1.4).

With no MMU:
CPU: 96-119 Mhz
DSP: 60-63 MHZ
FPU: 588-796 Mhz

With MMU:
CPU: 112-116 Mhz
DSP: 63-65 MHZ
FPU: 269-275 Mhz

(Lowest numbers are when moving mouse, and both CPU & FPU numbers fluctuate a lot more without MMU.)

For some reason it reports DSP running at nearly half of the CPU speed, instead of the expected 2x speed (32Mhz DSP vs 16Mhz CPU). And there's the rather radical 2x-3x FPU speed difference caused by MMU.

Any idea why these might happen?


Hi

I would need to have a closer look for DSP, but for FPU Toni/WinUAE told that it was not cycle accurate at all ; cycles are not known so I guess it uses a rough estimate that will be ok in some cases and not in others

As to why FPU would be slower when using MMU, either it's a different code path in WinUAE or there's sthg wrong in Hatari.

Note that MMU emulation has no cycle exact behaviour, so if you want to compare with/without MMU, be sure to not use "cycle exact" mode in Hatari


Nicolas



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