[Mesa-dev] [PATCH v2 00/15] Make LOAD_PAYLOAD sane again
kenneth at whitecape.org
Tue May 5 18:50:43 PDT 2015
On Tuesday, May 05, 2015 06:28:03 PM Jason Ekstrand wrote:
> This is a v2 of the series that I sent out a month or two ago to fix up the
> LOAD_PAYLOAD instruction in the i965 FS backend compiler. This new version
> incorporates comments from a variety of people.
> The last patch in the series is one that I think at least 3 of us have
> independantly written to enable copy-propagation from the ATTR register
> file. This helps the old fs_visitor code some but it helps NIR a *lot*.
> The end result is that NIR is now on par from a shader-db perspective with
> the old fs_visitor code.
> Shader-db results across all but the last patch:
> total instructions in shared programs: 7440514 -> 7440378 (-0.00%)
> instructions in affected programs: 11954 -> 11818 (-1.14%)
> helped: 68
> HURT: 0
> GAINED: 3
> LOST: 0
> All 68 helped shaders are in unigen heaven ultra 4xmsaa. It appears that
> it helps us to copy-propagate uses of omask. The omask is weird because
> it's a single-register header that's a 16-wide UW type. The previous
> guessing code was actually getting it wrong and copying it with an 8-wide
> UD move that was not force_writemask all. Now it is getting copied with
> force_writemask_all set correctly (because it's a header source). It seems
> like copy propagation can now come through and clean it up. (At least
> that's my theory; I didn't look at it that hard.)
Did you verify that Heaven actually works still? :)
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
except for my patch of course, since that wouldn't make any sense...
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 819 bytes
Desc: This is a digitally signed message part.
More information about the mesa-dev