Want to make sure I enable the cache in the right way

[Zamuel_a]

I am testing a 3d program I made and get very strange results when I compare it on a real Falcon and in Hatari (I wrote about it in the Hatari section). With everything turned on in Hatari for the most accurate emulation I get around 18fps on my program (both what I measure with the 200Hz timer and it "feels" like 18fps).
On my real machine I get 4 fps, so not really close.
I wonder what the problem might be. If it's something related to the cache on the real machine? Are the caches turned on by default? I guess so since when I turned it on in my program I didn't see any difference.

To set the different cache flags I used:

Code: Select all

;---------------------------------------------------------------------
; Set CacheControlRegister
setCACR:
	movec	cacr,d1
	or.w	d0,d1
	movec	d1,cacr
	rts
	
clearCACR:
	movec	cacr,d1
	eor.w	#$FFFF,d0
	and.w	d0,d1
	movec	d1,cacr
	rts

I call this from Pure C and use these defines

Code: Select all

#define WA	8192		/* 13 -Write Allocate */
#define DBE	4096		/* 12 - Data Cache Burst Enable - Not supported by the Falcon030! */
#define CD	2048		/* 11 - Clear Data Cache */
#define CED	1024		/* 10 - Clear Entry in Data Cache */
#define FD	512		/* 9 - Freeze Data Cache */
#define ED	256		/* 8 - Enable Data Cache */

#define IBE	16		/* 4 - Instruction Cache Burst Enable - Not supported by the Falcon030! */
#define CI	8		/* 3 - Clear Instruction Cache */
#define CEI	4		/* 2 - Clear Entry in Instruction Cache */
#define FI	2		/* 1 - Freeze Instruction Cache */
#define EI	1		/* 0 - Enable Instruction Cache */

And in the beginning of my program I turn Instruction cache on and data cache off

Code: Select all

	clearCACR(ED);
	setCACR(EI | CI);

This is done in the same way on Hatari as on a real machine, but why I get 4 times the speed in Hatari is very strange.

[mikro]

What about those cache freeze bits? That looks like the culprit to me.

[ThorstenOtto]

And are you sure you have are emulating a 030 in Hatari? for 040/060, the CACR is different.

[Zamuel_a]

And are you sure you have are emulating a 030 in Hatari? for 040/060, the CACR is different.

Yes Hatari says
16MHz/030(CE)-14Mb Falcon TOS 4.04 VGA 50Hz

[Cyprian]

VGA 50Hz

VGA modes slow down CPU significantly and I'm not sure whether Hatari properly emulates Videl's bus accesses

[czietz]

VGA modes slow down CPU significantly and I'm not sure whether Hatari properly emulates Videl's bus accesses

Good point! I ran CoreMark under Hatari 2.3.1 Falcon emulation in ST High and in Truecolor mode and it gave the same result: 9.73 it/s (Hatari settings: "16MHz/030(CE)"). On a real Falcon the benchmark results are significantly less and decrease further for Truecolor video mode, as shown here: https://github.com/czietz/coremark/wiki/Results

[Eero Tamminen]

VGA modes slow down CPU significantly and I'm not sure whether Hatari properly emulates Videl's bus accesses

Good point! I ran CoreMark under Hatari 2.3.1 Falcon emulation in ST High and in Truecolor mode and it gave the same result: 9.73 it/s (Hatari settings: "16MHz/030(CE)"). On a real Falcon the benchmark results are significantly less and decrease further for Truecolor video mode, as shown here: https://github.com/czietz/coremark/wiki/Results

What kind of speed differences there are with different TrueColor resolutions, for example "40 cols, interlace off" and "80 cols, interlace on"?

[czietz]

What kind of speed differences there are with different TrueColor resolutions, for example "40 cols, interlace off" and "80 cols, interlace on"?

I can't test this right now as my Falcon (mainboard) is in storage. Maybe someone else can provide these numbers by running CoreMark? Generally, the higher the memory bandwidth for video, the slower the Falcon becomes as Videl and CPU have to share access to RAM.

[Cyprian]

What kind of speed differences there are with different TrueColor resolutions, for example "40 cols, interlace off" and "80 cols, interlace on"?

there you can find some figures for different video modes:
Falcon and NemBench
I have also some additional Falcon's results for RGB.

Mikro wrote nice article about Videl's bus usage: https://mikro.naprvyraz.sk/docs/mikro/030_stram.html and that one https://mikro.naprvyraz.sk/docs/mikro/videl.html

[Zamuel_a]

I tried to run my program on a RGB monitor and gained 1 fps so it's still not close to Hatari (got 5 in RGB mode, 4 in VGA and 18 on Hatari)

[Rustynutt]

Deleted.
An epiphany worked out