This series contains several optimization-related patches that are all focussed around being able to do better at emitting MAD instructions.
Final shader-db results for the whole series on Haswell: total NIR instructions in shared programs: 2183472 -> 2152911 (-1.40%) NIR instructions in affected programs: 1299664 -> 1269103 (-2.35%) helped: 7106 HURT: 321 total FS instructions in shared programs: 4419884 -> 4378433 (-0.94%) FS instructions in affected programs: 1895564 -> 1854113 (-2.19%) helped: 10025 HURT: 221 GAINED: 81 LOST: 8 There is one notable place where this hurts us. There are three shaders in Kerbal Space Program that now spill. However, this does not seem to be the fault of NIR. See patch 8 for more details. Jason Ekstrand (9): nir/algebraic: #define around structure definitions nir/algebraic: Add a seperate section for "late" optimizations i965/nir: Use the late optimizations nir: Move the compare-with-zero optimizations to the late section nir: Add a dedicated ffma peephole optimization i965/nir: Use the dedicated ffma peephole nir/peephole_ffma: Be less agressive about fusing multiply-adds i965/nir: Run the ffma peephole after the rest of the optimizations i965: Use the same nir options for all gens src/glsl/Makefile.sources | 1 + src/glsl/nir/nir.h | 1 + src/glsl/nir/nir_algebraic.py | 17 +- src/glsl/nir/nir_opt_algebraic.py | 20 ++- src/glsl/nir/nir_opt_peephole_ffma.c | 261 +++++++++++++++++++++++++++++++ src/mesa/drivers/dri/i965/brw_context.c | 11 +- src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 11 ++ 7 files changed, 300 insertions(+), 22 deletions(-) create mode 100644 src/glsl/nir/nir_opt_peephole_ffma.c -- 2.3.3 _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
