On 02.05.2006, RaeNye wrote:

>>> (The next totalliy unrelated topic would be DMA for ATA, but
>>> I leave this for later)
>> Yes, ATA DMA might speed things up a little, but not as much
>> as you might
> think. The assembler optimized ATA driver is pretty fast.
> Again I think it's a don't-burden-cpu-for-no-reason thing. DMA
> is there for a reason; let the CPU spend its cycles in SWCODEC
> (or GUI, whatever).

ATA DMA might be equally fast, or even a bit faster,than the asm
optimised transfer routines. I am planning to try that on irver.
However, even with DMA, ATA transfers will still slow down the
CPU unless the CPU is *exclusively* working from cache + IRAM.
SDRAM access is *the* major bottleneck, and CPU and DMA
controller share the same bus. Furthermore, ATA DMA needs
special casing, since one of the IRAM blocks isn't DMA capable.

Regards, Jens
Received on Wed May 3 00:00:15 2006