[Mesa-dev] [PATCH v3 0/7] mesa/st: glsl_to_tgsi: improved temp-reg lifetime estimation
Gert Wollny
gw.fossdev at gmail.com
Sun Jun 18 21:54:28 UTC 2017
It seems some of the tests I added require a a too long life-time of
the registers so that the actual aim of the patch is no longer
achieved.
Comments on the code and algorithm structure are welcome, but I have to
rethink the tests, because they seem to be too focused on corner-cases
that may actually border on undefined behavior, and I will have to
review that.
@Dieter: if you want to test the algorithm try v2, it was far from
ready to be applied to mesa, but you could see whether the algorithms
gives you useful results.
Sorry for the noise,
Gert
Am Sonntag, den 18.06.2017, 19:42 +0200 schrieb Gert Wollny:
> Dear all,
>
> following the comments of Emil and Nicolai I've updated the patch
> set.
>
> Changes with respect to the old version are:
>
> - split the changes into more patches
> - correct formatting errors
> - remove the use of the STL with one exception though:
> since in st_glsl_to_tgsi.cpp std::sort is already used and its run-
> time
> performance is significantly better than qsort. It is used in the
> register
> rename mapping evaluation. It can be disabled by commenting out the
> define
> USE_STL_SORT in st_glsl_to_tgsi_temprename.cpp.
> - add more tests and improve the life-time evaluation accordingly
> - further reduce memory allocations
>
> The algorithms is the same as described before, with the little
> exception that
> now initially a dry run over the instructions is used to count the
> numbers of
> scopes. The run-time overhead of this operation can be neglected.
>
> In order to make it easier to transition to the new code and test it
> I tied it
> in parallel to the old code. It can be enabled by setting the
> environment
> variable MESA_GLSL_TO_TGSI_NEW_MERGE.
>
> piglit run on the "shader" test set doesn't show any changes. The
> additional
> passing test of I reported for v2 no longer passes, probably because
> of the
> more conservative life-time estimation required to make the new
> (valid) tests
> pass, but as I wrote before, the problem with this shader
> spec at glsl-1.50@execution at variable-indexing@gs-input-array-vec2-index-
> rd
> (and its sister *vec3*) is, IMHO not solvable by better register-
> renaming.
>
> The performance numbers estimated by running the shader-db are given
> in the
> commit message of the last patch, the trend is the same like reported
> before.
>
> Many thanks for any commenst,
> Gert
>
>
> Gert Wollny (7):
> mesa/st: glsl_to_tgsi move some helper classes to extra files
> mesa: Propagate c++11 CXXFLAGS from LLVM_CXXFLAGS to mesa/
> mesa/st: glsl_to_tgsi: implement new temporary register lifetime
> tracker
> mesa/st: glsl_to_tgsi: add tests for the new temporary lifetime
> tracker
> mesa/st: glsl_to_tgsi: add register renamame mapping evaluator
> mesa/st: glsl_to_tgsi: Add test set for evaluation of rename
> mapping
> mesa/st: glsl_to_tgsi: tie in new temporary register merge approach
>
> configure.ac | 1 +
> src/mesa/Makefile.am | 4 +-
> src/mesa/Makefile.sources | 4 +
> src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 319 +-----
> src/mesa/state_tracker/st_glsl_to_tgsi_private.cpp | 207 ++++
> src/mesa/state_tracker/st_glsl_to_tgsi_private.h | 165 ++++
> .../state_tracker/st_glsl_to_tgsi_temprename.cpp | 752
> ++++++++++++++
> .../state_tracker/st_glsl_to_tgsi_temprename.h | 36 +
> src/mesa/state_tracker/tests/Makefile.am | 41 +
> .../tests/test_glsl_to_tgsi_lifetime.cpp | 1042
> ++++++++++++++++++++
> 10 files changed, 2277 insertions(+), 294 deletions(-)
> create mode 100644
> src/mesa/state_tracker/st_glsl_to_tgsi_private.cpp
> create mode 100644 src/mesa/state_tracker/st_glsl_to_tgsi_private.h
> create mode 100644
> src/mesa/state_tracker/st_glsl_to_tgsi_temprename.cpp
> create mode 100644
> src/mesa/state_tracker/st_glsl_to_tgsi_temprename.h
> create mode 100644 src/mesa/state_tracker/tests/Makefile.am
> create mode 100644
> src/mesa/state_tracker/tests/test_glsl_to_tgsi_lifetime.cpp
>
More information about the mesa-dev
mailing list