This series implements an API that lets us use a hash set in the CSE pass, which fixes the performance problem while getting most of the code for Global Value Numbering (GVN) on top of GCM into the tree and actively used. The first three patches are taken from Jason's rebase of my GVN series, and patch 4 is modified from the original patch 4 to expose a slightly lower-level API (details in the patch itself). Patch 5 then rewrites the CSE pass to use the API introduced in patch 4. Note that the API could be just as easily adapted to implement nir_value_number() from the original patch 4.
Perf numbers are in patch 5, but for those that prefer pretty pictures, I've uploaded a few graphs: http://people.freedesktop.org/~cwabbott0/cse-series/perf-before.svg http://people.freedesktop.org/~cwabbott0/cse-series/perf-after.svg where you can clearly see how CSE goes from taking up almost all the time spent in nir_optimize() to less than half (but still a significant amount). This series is also available at git://people.freedesktop.org/~cwabbott0/mesa nir-cse-hash Connor Abbott (5): nir: split out instruction comparison functions nir: constify nir_ssa_alu_instr_src_components() nir: constify instruction comparison functions nir: add a helper for finding duplicate instructions nir/cse: use the instr_hash helper src/glsl/Makefile.sources | 3 + src/glsl/nir/nir.c | 27 ----- src/glsl/nir/nir.h | 5 +- src/glsl/nir/nir_instr_compare.c | 179 +++++++++++++++++++++++++++ src/glsl/nir/nir_instr_hash.c | 255 +++++++++++++++++++++++++++++++++++++++ src/glsl/nir/nir_instr_hash.h | 36 ++++++ src/glsl/nir/nir_opt_cse.c | 255 ++++----------------------------------- 7 files changed, 500 insertions(+), 260 deletions(-) create mode 100644 src/glsl/nir/nir_instr_compare.c create mode 100644 src/glsl/nir/nir_instr_hash.c create mode 100644 src/glsl/nir/nir_instr_hash.h -- 2.1.0 _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
