|Re: [hatari-devel] DSP performance|
[ Thread Index |
| More lists.tuxfamily.org/hatari-devel Archives
Hi, I think we should re explain correctly how instructions cycles work for the FALCON to be sure eveybody speaks of the same things ;)
1) The falcon bus is a 16 Mhz bus, so all long instructions takes 2 access to reach a data.
2) the memory access is 4 cycles
3) the cycles values given in the motorola doc are given for a 32 bits bus and a 2 cycles memory access
So, there's a mathematic computing to do (easy one) to adapt the cycles to the Falcon.
I did it in a openoffice "excel" table to help me when I wrote the static table.
I can share it if needed.
One must be careful, there a 4 different ways to compute the FALCON values according to the type of instructions (some of them use one formula whereas some other need another one). I'll have to dig again into the motorola doc to refresh my mind here.
So, here, we deal with 2 problems :
1) using the correct cycles values for the falcon
2) adjust a CPU *1 DSP *2 ratio
The 2 problems are independant and can be treated separatly.
I'll rewrite about this later tonight and give you all the formulas.
----- Mail original -----
De: "Miro Kropáček" <miro.kropacek@xxxxxxxxx>
Envoyé: Mercredi 1 Juillet 2015 09:37:16
Objet: Re: [hatari-devel] DSP performance
On Wed, Jul 1, 2015 at 9:15 AM, Toni Wilen < twilen@xxxxxxxxxx > wrote:
4 cycles because it has 16bit bus and does 2 16-bit fetches? (2x2) or does it really take 2x4 cycles to fetch 32-bit word?
Yes, it's both :) Two 16-bit prefetches and each of them takes 4 cycles.
MiKRO / Mystic Bytes