On 02/02/2008, Axel Dörfler <axeld@xxxxxxxxxxxxxxxx> wrote: > Bruno Albuquerque <bga@xxxxxxxxxxxxx> wrote: > > Axel Dörfler escreveu: > > > The new AGP bus manager does not support a settings file at all at > > > this > > > point; I'm no friend of settings files for kernel drivers if they > > > aren't strictly necessary (and they usually aren't). > > Well, the usual solution to settings file is to make the driver aware > > of > > all probblems in chipsets and disable features or enable workarounds > > when needed. As we probably don't have enough hardware to test every > > possible combination, settings files are a good workaround. > > Sure, but only if your aim is to fix all those bugs before deployment : > -) > It's always interesting to see how other platforms deal with these > issues, and what work-arounds they have in their code. Now you can add > a work-around for your hardware, yay ;-)) > If you know some options are problematic, it's better to only enable it > on certain (tested) hardware by default. AGP fast writes seem to be > such a case. > > > > But since it does not hang as before, I guess there might have been > > > a > > > bug in the old AGP bus manager that caused this hang? Does it > > > report > > > AGP 2x or 4x for you now? > > Correction, now it started to hang (the drawing engine hangs and do > > not > > come back anymore). Sometimes it hangs but come bacl after 5 or 6 > > seconds. > > > > I tried to figure out if it was using AGP 4X or 2X but funny enough > > there was no debug output whatsoever from the nvidia driver or > > agp_gart > > driver to the syslog. Looking at the code it seems you are matching > > chipset and graphics card capabilities and, in this case, it would > > enable AGP 4X (as both report it). > > The nvidia driver should have some debug output in it. I might have > forgotten about it in the AGP driver (update: confirmed, fixed in r). > > > I have dma_acc set to false in the nevidia settings file. So I guess > > it > > is not enabled. Funny enough, DMA works if AGP is disabled but the > > graphics operations seem to be slower than using PIO mode. > > Strange, it should always be faster than with PIO. But I guess Rudolf > could tell us more about it. > Maybe you should always disable DMA for that card in the driver. > Or is anyone out there who hasn't got any problems with that one? > > > > You could try to disable fast-writes and see if that helps. To do > > > that, > > > just remove the AGP_FAST_WRITE bit in the set_agp_mode() function > > > call > > > in the nVidia driver. > > I disabled fastwrites on the BIOS and now it seems to work. In fact, > > even DMA seems to be working. Some interesting stuff: > > > > 1 - I tried GLTeapot. It seems it is drawing outside of its window > > when > > you move the window around. > > That's more or less what happens on other systems, too, at least there > seem to be some problems left in that code (I only looked shortly at > that, but I couldn't even find any code that would stop all direct > windows when moving a window). > > > 2 - If DMA acc is enabled, not only it will draw outside its windows > > as > > it will also crash GLTeapot somewhere in > > MesaSoftwareRenderer::SwapBuffers(). > > That shouldn't really have any impact on that, sounds strange. > > > Having this in minda, any ideas? How much of a performance boost is > > fatswrites? > > I have no idea, I just know that the Radeon driver always disables it. > I guess Rudolf has some numbers somewhere (maybe in his blog or the > update/read-me files of his nVidia/AGP releases). > I don't think it'll be a huge problem. > > Bye, > Axel. > > Most older radeon cards have historically disliked fastwrites so it's normal for fastwrites to be disabled there. Nvidia cards prefer it enabled however. Just my 2 cents.