[Mesa-dev] i RESEND More NIR linking optimisations

[Dieter Nützel]

For the series:

Tested-by: Dieter Nützel <Dieter at nuetzel-hh.de>

on RX580 with F1 2017
with Mesa-git taken from 2 days ago.
As there isn't much tessellation I do NOT see (any) much improvements.

Full HD 1920x1080, all _max_ (!)

before 3 times

min      24 27 25
average  41 40 40
max      57 56 56

after 3 times

min      24 28 24
average  40 41 40
max      58 57 57

Intel Xeon X3470 (Nehalem), 2.93 GHz, 4/8, 24 GB
but with (poor) PCIe 2.0, Width x8, Speed 5GT/s

Dieter

Am 15.11.2017 00:35, schrieb Timothy Arceri:
> This is a resend/rebase of the series I sent a couple of weeks ago,
> its rebased on some of Jason's changes to i965 NIR linking that
> landed in master.
> 
> This series adds a varying array splitting pass as well a previous
> component packing series I sent out previously. This allows avoiding 
> the
> workaround of calling gather shader info twice since we can more easily
> keep the input/output bitmasks in sync now that we don't need to worry
> about partial marking of arrays.
> 
> Remaining improvements include adding a pass to compact varyings into
> consecutive
> slots rather than leaving empty slots when removing dead varyings.
> 
> Shader-db results for serires on i965 (BDW):
> 
> total instructions in shared programs: 13298718 -> 13191284 (-0.81%)
> instructions in affected programs: 2315180 -> 2207746 (-4.64%)
> helped: 14956
> HURT: 390
> 
> total cycles in shared programs: 540151400 -> 539397048 (-0.14%)
> cycles in affected programs: 297905258 -> 297150906 (-0.25%)
> helped: 25231
> HURT: 13033
> 
> total loops in shared programs: 3807 -> 3804 (-0.08%)
> loops in affected programs: 3 -> 0
> helped: 3
> HURT: 0
> 
> total spills in shared programs: 86577 -> 86640 (0.07%)
> spills in affected programs: 1380 -> 1443 (4.57%)
> helped: 7
> HURT: 15
> 
> total fills in shared programs: 90871 -> 90946 (0.08%)
> fills in affected programs: 1728 -> 1803 (4.34%)
> helped: 16
> HURT: 9
> 
> LOST:   4
> GAINED: 15
> 
> The spill hurt is all in dolphin uber shaders (as is most of the spill
> improvements).
> 
> Two of the lost programs are SIMD16 programs are from CS: GO because 
> 80% of the
> shaders get optimised away when we remove dead varying components, 
> these are
> also the shaders where the 3 loops go away.
> 
> Please review.
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev