Bug 24281

I found that I get fewer artifacts using the 2.6.32-rc1 + drm_next kernels than if I use just the drm_next kernel. And those kernels should be quite close with regards to r600. 

Created attachment 30129 [details]
Same video but running with drm-next kernel vs the linux-2.6 + drm-next kernel

There seems to be much more noise in the drm-next kernel. Also the 2d and 3d performance of the system is MUCH better with drm-next, so I'm not sure why linux-2.6 is so different.

Below is the fragment program it uses:
Currently we don't set barrier bit on any of these tex intructions as they 
don't depend on any prev ins. However they write to same dest, so I guess the result may get overwritten sometimes...

# Fragment Program/Shader
  0: TEX TEMP[0].y, INPUT[4], texture[0], 2D;
  1: ADD TEMP[0].x, TEMP[0].yyyy, CONST[0].-z-z-z-z;
  2: TEX TEMP[0].y, INPUT[4], texture[2], 2D;
  3: ADD TEMP[0].z, TEMP[0].yyyy, CONST[1].-x-x-x-x;
  4: MUL TEMP[0].x, TEMP[0], CONST[0].yyyy;
  5: MAD OUTPUT[1].z, TEMP[0], CONST[0].wwww, TEMP[0].xxxx;
  6: TEX TEMP[0].y, INPUT[4], texture[1], 2D;
  7: ADD TEMP[0].y, TEMP[0], CONST[1].-x-x-x-x;
  8: MAD TEMP[0].z, TEMP[0].-x-y-z-w, CONST[1].yyyy, TEMP[0].xxxx;
  9: MAD OUTPUT[1].y, TEMP[0].-x-y-z-w, CONST[1].zzzz, TEMP[0].zzzz;
 10: MAD OUTPUT[1].x, TEMP[0].yyyy, CONST[1].wwww, TEMP[0];
 11: MOV OUTPUT[1].w, CONST[0].xxxx;
 12: END

As a workarond for now, you can try:

diff --git a/src/mesa/drivers/dri/r600/r700_assembler.c b/src/mesa/drivers/dri/r600/r700_assembler.c
index 903b696..139a122 100644
--- a/src/mesa/drivers/dri/r600/r700_assembler.c
+++ b/src/mesa/drivers/dri/r600/r700_assembler.c
@@ -686,7 +686,7 @@ GLboolean add_tex_instruction(r700_AssemblerBase*     pAsm,
 
     // If this clause constains any TEX instruction that is dependent on a previous instruction, 
     // set the barrier bit
-    if( pAsm->pInstDeps[pAsm->uiCurInst].nDstDep > (-1) || pAsm->need_tex_barrier == GL_TRUE )
+    if(1 || pAsm->pInstDeps[pAsm->uiCurInst].nDstDep > (-1) || pAsm->need_tex_barrier == GL_TRUE )
     {
         pAsm->cf_current_tex_clause_ptr->m_Word1.f.barrier = 0x1;  
     }

Andre,

That patch did make things work much better. Can we get something along this merged into master. Is there any way to detect if the texture is in use and if so, set the bit?

Should be fixed in 869e20bcb7db9c6540eb6b538104303df738d302
7_6_branch currently

commit f8bee0e412ef1e8d5aed884561999fd1bd182494
Merge: 5e77b61 1f5b568
Author: Alex Deucher <alexdeucher@gmail.com>
Date:   Thu Oct 22 12:29:36 2009 -0400

    Merge branch 'master' of git+ssh://agd5f@git.freedesktop.org/git/mesa/mesa


Appears to have fixed this issue... Thank you...