[Mesa-dev] [PATCH 00/21] anv: Do cross-stage link optimizations

[Jason Ekstrand]

On October 29, 2017 21:34:01 Timothy Arceri <tarceri at itsqueeze.com> wrote:

> On 29/10/17 12:58, Jason Ekstrand wrote:
>> On Sat, Oct 28, 2017 at 11:36 AM, Jason Ekstrand <jason at jlekstrand.net
>> <mailto:jason at jlekstrand.net>> wrote:
>>
>>     This series adds support for cross-stage optimizations in anv.
>>     There are a
>>     few patches from Jordan's shader cache series in here that I wanted
>>     because
>>     they made my life easier.  There are also three patches CCd to stable to
>>     fix a but in the i965 cross-stage NIR linking which, as as side-effect,
>>     expose a nice brw_nir_link_shaders helper that we can use in anv.
>>     The bulk
>>     of the series, however, is the annoying refactoring of anv_pipeline.c to
>>     let us work with and cache the shaders an entire pipeline at a time
>>     instead
>>     of having everything be per-stage.  The patch to actually add the
>>     NIR link
>>     optimizations to ANV is almost trivial.
>>
>>     On my thermally throttled (and therefore a bit inconsistent) laptop,
>>     this
>>     seems to help the Aztec Ruins benchmark by 2%.
>>
>>
>> Or not... I'm having trouble reproducing it now.
>
> For what its worth RADV had improvements in the following:
>
> Sascha Willems demo results:
>
>   computecullandlod 39 -> 41 fps
>   pipelines ~6100 -> ~6200 fps
>
> The biggest improvement is with the component packing enabled:
>
> SaschaWillems Vulkan demo tessellation:
>
> ~4300fps -> ~4800fps

Yeah, I tried those out but they were too noisy on my laptop to get good 
data.  I asked Eero to try and get me some better numbers on his perf setup.

>>
>>     Carl Worth (1):
>>        intel/compiler: add new field for storing program size
>>
>>     Jason Ekstrand (17):
>>        anv/pipeline: Rework the parameters to populate_wm_prog_key
>>        anv/pipeline: Add populate_tcs/tes_key helpers
>>        anv/pipline: Add a helper struct for per-stage info
>>        anv/pipeline: Populate keys up-front
>>        anv/pipeline: Hash the entire pipeline in one go
>>        anv/pipeline: Call anv_pipeline_compile_* in a loop
>>        anv/pipeline: Pull shader compilation out into a helper.
>>        anv/pipeline: Drop anv_pipeline_add_compiled_stage
>>        anv/pipeline: Recompile all shaders if any are missing from the cache
>>        anv/pipeline: Compile to NIR in compile_graphics
>>        anv/pipeline: Add a separate "link" stage
>>        anv/pipeline: Pull most of the anv_pipeline_compile_* into common
>>     code
>>        intel/nir: Add a helper for getting the NoIndirect mask
>>        intel/nir: Break the linking code into a helper in brw_nir.c
>>        intel/nir: Use the correct indirect lowering masks in link_shaders
>>        nir/lower_indirect: Bail early if modes == 0
>>        anv/pipeline: Do cross-stage linking optimizations
>>
>>     Jordan Justen (3):
>>        intel/compiler: Add union types for prog_data and prog_key stages
>>        intel/compiler: Add functions to get prog_data and prog_key sizes for
>>          a stage
>>        intel/compiler: Remove final_program_size from brw_compile_*
>>
>>       src/compiler/nir/nir_lower_indirect_derefs.c |   3 +
>>       src/intel/blorp/blorp.c                      |  10 +-
>>       src/intel/blorp/blorp_blit.c                 |   5 +-
>>       src/intel/blorp/blorp_clear.c                |  15 +-
>>       src/intel/blorp/blorp_priv.h                 |   6 +-
>>       src/intel/compiler/brw_compiler.c            |  36 +
>>       src/intel/compiler/brw_compiler.h            |  34 +-
>>       src/intel/compiler/brw_fs.cpp                |   6 +-
>>       src/intel/compiler/brw_nir.c                 |  63 +-
>>       src/intel/compiler/brw_nir.h                 |   4 +
>>       src/intel/compiler/brw_shader.cpp            |  12 +-
>>       src/intel/compiler/brw_vec4.cpp              |   5 +-
>>       src/intel/compiler/brw_vec4_gs_visitor.cpp   |   8 +-
>>       src/intel/compiler/brw_vec4_tcs.cpp          |  12 +-
>>       src/intel/vulkan/anv_pipeline.c              | 971
>>     ++++++++++++++-------------
>>       src/intel/vulkan/anv_private.h               |   2 +-
>>       src/intel/vulkan/genX_pipeline.c             |   2 -
>>       src/mesa/drivers/dri/i965/brw_cs.c           |   5 +-
>>       src/mesa/drivers/dri/i965/brw_gs.c           |   5 +-
>>       src/mesa/drivers/dri/i965/brw_link.cpp       |  38 +-
>>       src/mesa/drivers/dri/i965/brw_tcs.c          |   5 +-
>>       src/mesa/drivers/dri/i965/brw_tes.c          |   5 +-
>>       src/mesa/drivers/dri/i965/brw_vs.c           |  11 +-
>>       src/mesa/drivers/dri/i965/brw_wm.c           |   5 +-
>>       24 files changed, 668 insertions(+), 600 deletions(-)
>>
>>     --
>>     2.5.0.400.gff86faf
>>
>>
>>
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>