Re: Snappy GUI response
Sven Luther <[email protected]>
| Newsgroups | gmane.comp.xfree86.forum |
|---|---|
| Message-ID | <20030708114303.GB4143@iliana> |
On Fri, Jul 04, 2003 at 12:47:10PM -0700, Mark Vojkovich wrote: > On Fri, 4 Jul 2003, Alan Cox wrote: > > > > Yes, scrolling long documents has finally become somewhat limited by text > > > output performance -- and that is actually more limited by frame buffer > > > read speed than anything else... > > > > Its limited by the fact XFree86 doesn't know how to pull frame buffer > > data properly. On PCI/AGP you have to push data from the card. The card > > knows that, the bus knows that. XFree86 hasn't quite woken up to it. > > > > That's not an XFree86 issue per-se. It's a driver issue. > NVIDIA's binary drivers use DMA for GetImage and PutImage. > The fastest way to do software AA text would be to GetImage the > bounding box, render the text to the image, and PutImage it back. > Even with non-DMA GetImage and PutImage this is probably faster > that alpha blending into the framebuffer since the read-modify- > write is probably thwarting write combining. > > Actually, on a P4 which will blow it's write combine on a cache > miss, this method of drawing SW AA text is probably always faster. That said, getting images from the chip, even if doing it with DMA is an expensive operation. Nowaday, most graphic cards have programmable accels, altough maybe we have not the specs to implement them, and it has not been envisioned to use them outside of DRI, but i guess most AA stuff and other render extensions could well be implemented with one of the shader units in the graphic pipeline, the pixel shaders being the most obvious candidate. That said, the ball is then in the hardware vendor camp, since they would need to release the appropriate specs for it, and we would need better integration of the DRI in the server also. Friendly, Sven Luther