On Sat, 2009-03-21 at 18:02 -0700, Dave Airlie wrote:
> Hi all,
>
> So I had to drop texture tiling when I did the radeon-rewrite but I'd like
> to bring them back.
>
> Now with traditional drivers, we have the mesa copy of the texture and the
> card copy, and we usually texture from VRAM only, so we can upload to VRAM
> and tile on the way, and if we hit a sw fallback we just use the textures
> from the mesa fallback.
>
> In the bufmgr world in theory we don't keep two copies of textures around,
> and on radeons we can at least enable texture tiling on GART textures. So
> if I store the texture tiled for hw use, we don't have anything like span
> access to textures from what I can see for sw use. If the texture was in
> VRAM, in theory for sw use we could use a surface, however it leave tiled
> in GART busted.
>
> I'm just wondering if anyone has already tackled this in any driver, and
> how this could work best. Does Gallium provide surface accessors for
> textures like spans?
>
> I'm sort of thinking I keep two buffers objects, one sw only and one for
> hw to use and use the GPU to blit between them when updates are needed.
This is the kind of scenario that pipe_transfer was created for (from
p_screen.h):
/** Get a transfer object for transferring data to/from a texture */
struct pipe_transfer *(*get_tex_transfer)(struct pipe_screen *,
struct pipe_texture *texture,
unsigned face, unsigned level,
unsigned zslice,
enum pipe_transfer_usage usage,
unsigned x, unsigned y,
unsigned w, unsigned h);
void (*tex_transfer_destroy)(struct pipe_transfer *);
void *(*transfer_map)( struct pipe_screen *,
struct pipe_transfer *transfer );
void (*transfer_unmap)( struct pipe_screen *,
struct pipe_transfer *transfer );
It provides the necessary hooks and context for the pipe driver to
untile/retile the texture. For example, you could keep a linear copy
(temporary or not) in cached memory, and the tiled gpu copy. In
get_tex_transfer or transfer_map you would untile from the gpu copy if
the cpu copy was dirty. On transfer_unmap/tex_transfer_destroy you would
re-tile, or even better, delay the tiling until the texture is bound to
a context and a draw emmited. When rendering to a texture you'll mark
the cpu copy dirty. When mapping the texture for write you'll mark the
gpu copy dirty. etc.
It is not as fine grained as span accessors but it is known to work very
well.
Jose
------------------------------------------------------------------------------
Apps built with the Adobe(R) Flex(R) framework and Flex Builder(TM) are
powering Web 2.0 with engaging, cross-platform capabilities. Quickly and
easily build your RIAs with Flex Builder, the Eclipse(TM)based development
software that enables intelligent coding and step-through debugging.
Download the free 60 day trial. http://p.sf.net/sfu/www-adobe-com
_______________________________________________
Mesa3d-dev mailing list
Mesa3d-dev@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/mesa3d-dev
