[Mesa-dev] [PATCH 0/9] NIR dead control-flow removal

[Jason Ekstrand]

On Fri, May 8, 2015 at 11:15 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> On Fri, May 8, 2015 at 10:03 PM, Connor Abbott <cwabbott0 at gmail.com> wrote:
>> This series implements a bunch of related optimizations that run at
>> once as part of the same pass to eliminate control flow that is
>> useless. This includes both unreachable code and useless loops, i.e.
>> loops that don't compute a result used by the rest of the code. To do
>> this, I needed to change some low-level things that change the behavior
>> of nir_instr_remove() (see patch 2 for details), but I've tested the
>> series on piglit and there are no regressions, and there are no
>> shader-db changes either.
>
> I tried to get you shader-db results, but it crashes part-way through.
> In particular, it looks like unigen heaven on high will crash it.
> Either that or a couple of synmark shaders but I doubt those are the
> culprets.

That patch seems to fix the crashes.  Here's some shader-db results for you:

Shader-db results for Broadwell:

total FS instructions in shared programs: 7182550 -> 7182379 (-0.00%)
FS instructions in affected programs:     3055 -> 2884 (-5.60%)
helped:                                    29
HURT:                                     0

total NIR instructions in shared programs: 5413353 -> 5443233 (0.55%)
NIR instructions in affected programs:     1453327 -> 1483207 (2.06%)
helped:                                    4
HURT:                                     6758

I have no ide why it *added* NIR instructions but aparently it did.
However, they were things that the back-end was getting rid of anyway.
Any idea why?
--Jason

> I think it's also worth noting that piglit probably isn't a good test
> of whether or not this works.   I doubt there are many piglit shaders
> with dead control-flow and I think GLSL IR can get rid of it in some
> of the common cases.  If you wanted to "generate NIR" to test with,
> you could always hack up my SPIRV -> NIR branch to run an optimization
> loop after converting.  Then you can see what happens without GLSL IR
> in the way.
> --Jason
>
>> This is rebased on Jason's series to use lists for use-def sets; it
>> turns out that the changes required were pretty minimal.
>>
>> The series is also available at:
>>
>> git://people.freedesktop.org/~cwabbott0/mesa nir-dead-cf-v3
>>
>> Connor Abbott (9):
>>   nir/vars_to_ssa: don't rewrite removed instructions
>>   nir: insert ssa_undef instructions when cleanup up defs/uses
>>   nir: add an optimization for removing dead control flow
>>   nir/validate: validate successors at the end of a loop
>>   nir/dead_cf: delete code that's unreachable due to jumps
>>   nir: add nir_block_get_following_loop() helper
>>   nir: add a helper for iterating over blocks in a cf node
>>   nir/dead_cf: add support for removing useless loops
>>   i965/fs/nir: enable the dead control flow optimization
>>
>>  src/glsl/Makefile.sources            |   1 +
>>  src/glsl/nir/nir.c                   |  94 ++++++++--
>>  src/glsl/nir/nir.h                   |  12 ++
>>  src/glsl/nir/nir_lower_vars_to_ssa.c |   3 +-
>>  src/glsl/nir/nir_opt_dead_cf.c       | 332 +++++++++++++++++++++++++++++++++++
>>  src/glsl/nir/nir_validate.c          |  21 +++
>>  src/mesa/drivers/dri/i965/brw_nir.c  |   2 +
>>  7 files changed, 452 insertions(+), 13 deletions(-)
>>  create mode 100644 src/glsl/nir/nir_opt_dead_cf.c
>>
>> --
>> 2.1.0
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev