Re: [hatari-devel] Adding cache support for the MegaSTE

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 13/07/2024 à 22:37, Nicolas Pomarède a écrit :
> Le 13/07/2024 à 15:29, Christian Zietz a écrit :
> > Christian Zietz schrieb:
> >
> > > Results attached.
> >
> > I struggle to understand the difference between 8 MHz and 16 MHz + 
> > 100% cache hits, though. The come out at $19E = 414 HBLs and $A7 = 167 
> > HBLs, respectively, which is a 414/167 ≈ 2.5(!) speed increase.
> >
> > You do a MOVE.W (Ax)+,Dx; DBRA Dy,xxx. The MOVE is 8 cycles, the DBRA 
> > is 10 cycles. In 8 MHz mode the DBRA incurs two waitstates (because of 
> > unaligned prefetch from ST RAM). Hence, in 8 MHz mode the loop should 
> > take 8+10+2=20 cycles, equivalent to 40 16-MHz cycles.
> >
> > Assuming 100% cache hit, in 16-MHz mode the same loop ought to take 
> > 8+10 = 18 cycles.
> >
> > But that only would explain a 2.22 (40/18) speed increase. I wonder if 
> > the rest is due to overhead in interrupt handling in 8-MHz mode.
> >
> > Do you see a way to modify your test program to run with interrupts 
> > disabled, using, e.g., one of MFP timers for time measurement?
>
> In my case where I want to compare Hatari with MegaSTE, it's not too 
> much of a problem if counting takes more time, as long as both Hatari 
> and MegaSTE give the same result
>
> but it's true that an HBL interrupt takes 112 cycles, even with just a 
> "add" in the interrupt. And when you run in monochrome an HBL is 
> triggered every 224 cycles at 8 MHz, so half of the time is spent in the 
> hbl, which is a lot compared to color mode at 50 Hz for example.
>
> and even at 16 MHz you loose a lot of cpu just for the hbl int, which 
> makes it harder to really measure cache alone with as little overhead as 
> possible.
>
> I will change the measure mode to use a timer and send a new version.

Here's an updated version that uses timer A running at div=100  24576 Hz 
instead of HBL. Overhead should be very low compared to HBL.

Attachment: mste_cac.prg
Description: Binary data