From: Keith W. <ke...@tu...> - 2003-01-28 16:59:11
|
Ian Romanick wrote: > Keith Whitwell wrote: > >> Evgeny Kotsuba wrote: >> >>> On Mon, 27 Jan 2003 14:30:14 -0500 Brad Lyon <bra...@nt...> wrote: >>> >>>> I even did an optimized build first and it didn't seem to matter. >>>> I'm willing to look into this, but I'm not sure where to start. >>> >>> >>> Obviously you should show your small test programm. We aren't >>> telepathes ;-) >> >> >> There's a long standing problem with this in mesa. I'm (slowly) >> addressing it with the work on the Mesa-vtx branch. > > > Can you expound on that a bit? I know I'm going to be working on > ARB_vertex_buffer_object (if nobody beats me to it) sometime, so I'm > gaining interest in the vetex processing path in Mesa. The basic idea is a massively cleaned up radeon_vtxfmt*.* used to construct array-of-struct vertices for immediate-mode and displaylist storage in mesa. There are a number of nice things about this that eliminate a lot of the kruft and wasted memory in the old t_imm_* approach. > I've also been thinking about adding the ability to process > structure-of-array vertex data instead of array-of-structure vertex data > (like INTEL_parallel_arrays) for software TNL. This would be a win for > SSE and (I'm pretty sure) AltiVec architectures. That's a > someday-if-I-have-a-chance thing, though. That might be possible as a parallel implementation. Keith |