Re: Isochronous transmission via OHCI

[Ken O. <ko...@ar...>]

Gord Peters wrote:

> Of course, video1394 will likely be better for slower computers since it
> directly maps a chunk of memory from the appliction into the kernel
> (avoiding memory copies between the kernel and the application) and uses
> that for it's isochronous data (application feeds iso data into the mmap()ed
> memory and driver reads data from this memory).  Whereas libraw1394 copies
> data on a per packet basis (incurring performance hits in both memory copies
> and extra function calls to send data on a packet by packet basis).

That was my thinking as well.  The DMA attribute is particularly
attractive for high-data-rate applications.  We're trying to drink from
a fire-hose (or is that a fire-wire-fire-hose?) ;)

Currently I'm getting a lowly transfer rate of 4.1 MBytes/sec using my
hacked-up telemetry driver (clone o' video1394).  I'm guessing its
because I need to nuke some synchronization bit somewhere or convince
the driver to take up more bandwidth on the bus.

In the routine initialize_dma_it_prg, I found the following line:

		it_prg[i].data[0] = 
			(DMA_SPEED_100 << 16) 
			| (/* tag */ 1 << 14)
			| (d->channel << 8) 
			| (TCODE_ISO_DATA << 4);

I changed the DMA_SPEED_100 to DMA_SPEED_400, but I didn't see any
change in performance.  I also tried commenting out the following line:

		if (i==0) it_prg[i].data[0] |= sync_tag;

But that stopped data transmission altogether. I gather I need to do
more groking... :(

-- 
______________________________________________________________
Kenneth Ray Offer, II
Sr. Systems Analyst
Applied Research Laboratories/SDD
University of Texas at Austin
Office (512) 835-3859   FAX (512) 835-3259

Any opinions expressed above are mine and not necessarily
those of my employer.