Originally posted by icyh
View Post
Announcement
Collapse
No announcement yet.
Kernel Mode-Setting, GEM, DRI Progresses On FreeBSD
Collapse
X
-
-
Originally posted by Shining Arcanine View PostThe other option is to modify the Radeon and Nouveau drivers to use GEM.
- Likes 1
Comment
-
I'm not a DRM developer so this may be completely wrong, but I think most of this is accurate. I'm sure I'll be corrected if it isn't.
Originally posted by michael-vb View Postwhy do modern graphics cards need such complicated heap management, rather than just say telling the kernel which areas of VRAM are available for textures and things and letting the kernel set up a heap there itself?
Originally posted by michael-vb View PostOr for that matter what makes GEM a better fit to some cards and TTM a better fit to others?
Let the corrections begin.
Comment
-
Originally posted by michael-vb View PostUsing this as an excuse to ask what is probably a rather basic question - why do modern graphics cards need such complicated heap management, rather than just say telling the kernel which areas of VRAM are available for textures and things and letting the kernel set up a heap there itself? Or for that matter what makes GEM a better fit to some cards and TTM a better fit to others?
Comment
-
Almost all modern GPUs can address several types of memory all with their own set of limitations. Some examples:
- VRAM. This is local memory attached to the GPU. It's fast for the GPU to access, but relatively slow for the CPU (especially for reads). Additionally, the CPU cannot necessarily access the entire amount of VRAM, only the GPU can. As such, if a buffer is in a non-CPU accessible region of VRAM, it has to be migrated by the GPU to a region of VRAM that can be accessed by the CPU or to anther memory type that the CPU can access.
- AGP GART memory. These are pinned system pages that are mapped into a contiguous aperture provided by the AGP GART (Graphics Address Remapping Table) mechanism on the northbridge. Unfortunately, lots of AGP chipsets are buggy and AGP pages need to be uncached, so they are slow to access with the CPU.
- GPU GART memory. These are pinned system pages that are mapped into a contiguous aperture provided by GART hardware on GPU. Usually the GPU GART can support both cached and uncached paged by doing a snooped request for cached pages. Both cached and uncached pages have their advantages. Cached pages are faster for the CPU to access, but slower for the GPU. Uncached pages are the opposite. Both kinds are slower than VRAM from the GPU's perspective, but faster for the CPU.
Taken together the acceleration drivers have to decide what types of memory to are best for the specific task at hand. They make requests to the memory manager to fulfill those requests, but the pools are of limited sizes, so the memory manager has to do it's best fulfill those requests by migrating data around to keep everything running as optimally as possible which is not an easy task.
- Likes 1
Comment
-
?????????? ??
?????????? ?????. ????? ? ??????? ???? ? ?????? ? ?????? ?????????, ?? ???????? ???????? ?? ???????????? ???????.
????? ????? ???? ???????? ??????, ?????? ????????? ???????? ? ???????? ????????????! ???? ??? ?? ??????? ???? ? ?????.
?????????? ??? ??? ??? ????????.?
????????:
????????????. ???? ???????? ??????????? ????? ?????. ?, ? ?????????, ?????? ??? ????? ????? ??????.
????????????, ??? ?? ????? - ??? ??????? ?? ??? ? ???????????. ??? ??? ??????? ??? ??? ??????.
? ? ????? ????? ???!
Comment
Comment