[Mesa-dev] i965: slow glTexSubImage2D because of Y tiling

[Tomash Brechko]

Hello,

I have an application that on each frame fills a full-window pixel buffer
and then calls glTexSubImage2D() and maps the texture to a screen by
drawing full-window rectangle (my hardware is Intel(R) Core(TM) i7-2670QM
CPU @ 2.20GHz (brw->gen == 6), OpenGL 3.0, direct glDrawPixels() is much
slower and is not used).

However in
src/mesa/drivers/dri/i965/intel_tex_subimage.c:intel_texsubimage_tiled_memcpy()
fast path is implemented only for I915_TILING_X, and
src/mesa/drivers/dri/i965/intel_mipmap_tree.c:intel_miptree_choose_tiling()
returns (I915_TILING_Y | I915_TILING_X) in my case, which makes
intel_miptree_create() in the same file call intel_region_alloc() with
I915_TILING_Y.

By forcing intel_region_alloc() to use I915_TILING_X I get 50% FPS rise
(for glxgears 2400FPS -> 3600FPS in default window, 108FPS -> 175FPS for
1920x1080 fullscreen).  I do this with

  gdb -x i965-texture-tiling-x.gdb --args myapp

where i965-texture-tiling-x.gdb script contains:
----------
  set breakpoint pending on
  break intel_miptree_create
  break intel_region_alloc
  disable 2
  comm 1
    silent
    enable 2
    cont
  end
  comm 2
    silent
    set tiling = 1
    disable 2
    cont
  end
  run
---------

So I wonder if there are strong reasons to prefer Y tiling for textures, at
least on gen == 6?  Is Y tiling preffered because of faster texel lookups
(assuming glTexSubImage2D() is not called on each frame)?


-- 
  Tomash Brechko
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20131227/061d8451/attachment.html>