AmigaOS News from AmiWest

[Karlos]

I feel that if DMA cannot be allowed/achieved in the A1200 Mediator boards then I think that OS4.1 for Classic A1200 users is realistically a non-starter, doomed, with no hope of success.

Wait, is this the super-fast 132MB/s DMA method that allows other PCI cards to access a block of memory on the video card without CPU intervention?

I mean it's absolutely brilliant as long as the data you've transferred to your graphics card is not actually needed by the CPU at any point. Unfortunately, it almost always is and as soon as it is, you have no DMA when you read the VRAM, you just have the CPU crawling along, waiting for the extremely slow bus as it fetches the data.

Buffering like that does help a bit, but in real world applications, it's not that much of a big deal. I mean, ideally, you'd want to wait until you'd got a few tens of KB of incoming data buffered before you bother tying up the CPU to retrieve it. However, for real applications, this simply isn't feasible. Network traffic, for example, arrives at the rate it arrives. You can't just sit there buffering it until you have a quantity ideally suited for CPU transfer into Fast RAM for the TCP stack to play with.

[Karlos]

I can't say I've ever sat there twiddling my thumbs waiting for the memory/system to react. It usually works efficiently and silently, unnoticed in the background, though I don't ever do a CPU RAM check to see what is being used where and when, and how fast it is being used, etc., but it seems OK for my needs. I also have never noticed anyone on the Mediator forum ever state they have been waiting for the RAM/CPU to catch up with a task, seeing as you are making out that it is so slow, as it just isn't a noticeable problem IMHO.

Though you seem to comprehend what I've said, you seem to have misunderstood the point I was making. The following applies to the Mediator 1200. I can't speak for the 4000 model as I don't own one.

Firstly, I'm not making anything out to be slow, it is slow. I've measured the performance of reading and writing video memory. In fact, when these expansions first appeared, I got lots of people to perform benchmarks for me so that I could compare the memory access speed (from the CPU) of Mediator versus things like BVision. The video ram on my Mediator 1200 / Voodoo 3000 system I have is significantly slower to read and write from the CPU than that my BVision, for example, even though the 040 in my Mediator system is faster than the one in my BVision one.

So, to reiterate, you made an issue that the DMA methods supported in 3.9 should be supported in 4.x or it's all a bit pointless. The truth of the matter is that the DMA methods supported in 3.9 are of limited usefulness even in 3.9 since the limiting factor in a Mediator configuration is not the speed at which different PCI cards can talk to each other and it never was. The limiting factor is how fast any expansion on the Mediator can be read from or written to by the CPU.

In any normal PCI machine, your attached devices don't talk to each other that much (the only obvious exception here would be something like a TV tuner card, where having the ability to DMA write the decoded video into your graphic's card's overlay memory would be perfectly sensible), but instead will perform DMA transfers to and from main memory. However, that's just not possible on the Mediator since it's not part of your accelerator card and has no access to the memory on it. So, what Elbox did, was to use a chunk of ram on the video card as a DMA buffer for data that would be going between PCI cards and the rest of the system. In theory it's not a bad idea, but as I said, the benefit of buffering depends on the specific operation you're doing. Holding on to incoming audio that will be eventually streamed to disk is a case that makes sense. Holding onto incoming network packets is not.

In the end, whether you are reading buffered data from the video card's memory or reading data directly from any of the PCI devices on the bus there's little speed difference, so no matter how fast your cards can DMA into your video card's memory, in the end it doesn't help that much - TV tuner card example excepted.

Elbox's claim about zero CPU intervention DMA transfers is a bit of a red herring when looking at the full picture. Sure, your network or sound card (both of which should have some local buffer memory of their own) can DMA data to your video card without the CPU getting involved, but once that stage of the transfer is done, the CPU has to read that data from the video ram in order to do anything with it. I'm not convinced it's much faster than reading data directly from the source device on the PCI bus. For some devices, it might even be slower.

[Karlos]

These results are from tests in 2003 sometime:

Voodoo 3000 / Mediator 1200 / 68040 @ 28MHz

VRAM Write Bandwidth       : 9293.32 K/s
VRAM Read Bandwidth        : 5604.20 K/s

Permedia 2 / 68040 @ 25MHz

VRAM Write Bandwidth       : 10395.24 K/s
VRAM Read Bandwidth        : 5354.32 K/s

These were using aligned and loop unrolled 32-bit transfers from VRAM < - > CPU register. A similar test was carried out using move16 for cache aligned FastRAM < - > VRAM. On my P2, reading was no faster, but writing was increased a bit (about 25%). On the Voodoo 3000, both reading and writing were a bit slower (about 10%).

Anyway...

Unfortunately DMA seems to be of critical usefulness in OS4.x as DMA is required by OS4.x

Required as of when? OS4 happens to work fine with the Mediator. Admittedly, not every PCI card that has drivers for 3.x has drivers for 4.x, but you're making it sound as if the entire PCI system won't work without Elbox's PCI card <-> VRAM DMA method being employed. Which is patently untrue.

[Karlos]

We are clearly arguing about different things here. You are implying that supporting Elbox's method of inter-device DMA as used in OS3.9 is somehow required to make your expansions work properly under OS4. It isn't. New drivers are.

So the Voodoo stayed within about 90% of the speed of the Permedia/BVision, from the figures you have supplied.

I appreciate the time you have taken to supply this information, thanks.

Frankly though I don't call that slow, it's more like you say yourself - slower, and only a little slower at that, 90%+ is within useful parameters.

It's very slow when you consider the theoretical speed of PCI, or even just comparing it to accessing local memory on your accelerator card. According to Elbox, PCI devices can exchange data at 130MB/s. However, if in the end you are transferring that data to Fast RAM at a paltry 5/MBs, it would have made no difference at all if you'd read it straight from the original card or not.

[Karlos]

I don't consider our discussion an argument, just a friendly exchange of ideas and impressions, I would like to think of what we are saying here in just that manner.

An argument doesn't necessarily imply any hostility anyway.

Oh, by the way, you seem to have forgotten to tell me what other PCI cards you have working in your OS4.0 A1200 Mediator setup, so please don't forget to let me know on this occasion, or you'll be getting me thinking you don't have any other PCI cards in your A1200 PPC powered Mediator working under OS4.0 Classic.

Right now, you'd be right, I have none whatsoever. If you check my signature, you'll see why. The mediator is currently in an OS3.9 machine.