On Tue, Apr 5, 2016 at 3:25 PM, Marek Olšák <mar...@gmail.com> wrote: > Hi Rob, > > I have 2 questions. > > 1) Do you have any data about performance of shader compilation between: > > GLSL -> TGSI -> driver IR -> bytecode > -and- > GLSL -> NIR -> driver IR -> bytecode >
I can't compare this directly, since for me it is either glsl->tgsi->nir->driver or glsl->nir->driver.. I dropped the ir3 tgsi f/e a while back. But from memory, NIR does have better constant folding, for example. There are some cases, like matrix column access, where by skipping tgsi I should be able to do a bit better (ie. not do if/else ladder), although I don't do this yet. (driver_location is still in units of vec4, which is a bit of a legacy of tgsi->nir..) I should mention, that part of my goal for this is to avoid having to plumb half-precision support through TGSI.. and for gles apps, half/mediump can be a big win, at least on mobile hw. Also, not loosing the input/output/uniform names in the glsl->nir->driver path is nice for debugging ;-) I do need to make some more formal measurements of shader compile time, but "seat of the pants" it feels faster. > 2) What's the quality of the driver IR when it comes out of NIR? Is it > optimized well? Most of the silly/suboptimal things in resulting shader binary are ir3 issues, not NIR issues. (And I patches I pushed yesterday and today fix the worst of those.) I did notice the other day that NIR is missing out on some algebraic opts in some cases, like loop unrolling that results in foo = ((bar + 1) + 1) + 1) doesn't become foo = bar + 3.. Jason did say he had some patches for that somewhere. BR, -R > Thanks, > Marek > > > On Sat, Mar 26, 2016 at 10:02 PM, Rob Clark <robdcl...@gmail.com> wrote: >> From: Rob Clark <robcl...@freedesktop.org> >> >> Ok, hopefully (close to the) last iteration of this patchset. Quite a >> lot of it has r-b's by now. There a few new nir bits (in particular, >> the "lower-io-types" pass and some updates to the "lower-outputs-to- >> temporaries" pass. Some of that will have some conflicts with stuff >> that Jason has in-flight.. I can either rebase on top of his patchset >> or rebase his patchset on top of mine, depending on what lands first. >> >> There is also a new mesa/st pass to lower builtin uniforms, to address >> issues that I mentioned with the previous round of patches (ie. built- >> ins getting packed differently from normal struct uniforms). >> >> There is also a new nir_lower_io_types pass that splits up complex >> inputs/outputs to vec4 slots so that we can deal with galliums VBO >> slot assignment (ie. it will already only assign VBO space to VS >> varyings which are actually read). >> >> Btw, I wouldn't mind starting to push at least parts of this patchset >> which have r-b's and don't depend on other parts that don't have r-b. >> But I've held off for now to avoid having a bunch of nir passes that >> aren't used upstream. >> >> Rob Clark (16): >> gallium: refactor pipe_shader_state to support multiple IR's >> gallium: add NIR as a possible IR >> nir: add lowering pass for y-transform >> nir: add lowering pass for glDrawPixels >> nir: add lowering pass for glBitmap >> nir: clamp-color-output support >> nir: passthrough-edgeflags support >> nir: lower-io-types pass >> nir: move callsite of lower_outputs_to_temporaries >> nir: rename lower_outputs_to_temporaries -> lower_io_to_temporaries >> nir/lower-io: split out some helper fxns >> nir/lower-io: add support for lowering inputs >> glsl: export accessor for builtin-uniform descriptors >> mesa/st: split the type_size calculation into it's own file >> mesa/st: add nir pass for lowering builtin uniforms >> mesa/st: add support for NIR as possible driver IR >> >> src/compiler/Makefile.sources | 8 +- >> src/compiler/glsl/builtin_variables.cpp | 24 +- >> src/compiler/glsl/ir.h | 3 + >> src/compiler/nir/glsl_to_nir.cpp | 2 - >> src/compiler/nir/nir.h | 40 +- >> src/compiler/nir/nir_builder.h | 2 +- >> src/compiler/nir/nir_lower_bitmap.c | 166 ++++++++ >> src/compiler/nir/nir_lower_clamp_color_outputs.c | 117 ++++++ >> src/compiler/nir/nir_lower_drawpixels.c | 252 ++++++++++++ >> src/compiler/nir/nir_lower_io_to_temporaries.c | 185 +++++++++ >> src/compiler/nir/nir_lower_io_types.c | 178 +++++++++ >> .../nir/nir_lower_outputs_to_temporaries.c | 133 ------- >> src/compiler/nir/nir_lower_passthrough_edgeflags.c | 82 ++++ >> src/compiler/nir/nir_lower_wpos_ytransform.c | 310 +++++++++++++++ >> src/gallium/auxiliary/hud/hud_context.c | 9 +- >> src/gallium/auxiliary/postprocess/pp_run.c | 3 +- >> src/gallium/auxiliary/tgsi/tgsi_ureg.c | 4 +- >> src/gallium/auxiliary/util/u_simple_shaders.c | 18 +- >> src/gallium/auxiliary/util/u_tests.c | 3 +- >> .../drivers/freedreno/ir3/ir3_compiler_nir.c | 8 + >> src/gallium/include/pipe/p_defines.h | 13 +- >> src/gallium/include/pipe/p_screen.h | 9 + >> src/gallium/include/pipe/p_state.h | 32 +- >> src/mesa/Makefile.sources | 5 + >> src/mesa/drivers/dri/i965/brw_nir.c | 1 + >> src/mesa/state_tracker/st_glsl_to_nir.cpp | 425 >> +++++++++++++++++++++ >> src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 104 ++--- >> src/mesa/state_tracker/st_glsl_to_tgsi.h | 5 + >> src/mesa/state_tracker/st_glsl_types.cpp | 98 +++++ >> src/mesa/state_tracker/st_glsl_types.h | 44 +++ >> src/mesa/state_tracker/st_nir.h | 60 +++ >> src/mesa/state_tracker/st_nir_lower_builtin.c | 242 ++++++++++++ >> src/mesa/state_tracker/st_program.c | 148 ++++++- >> src/mesa/state_tracker/st_program.h | 6 + >> 34 files changed, 2487 insertions(+), 252 deletions(-) >> create mode 100644 src/compiler/nir/nir_lower_bitmap.c >> create mode 100644 src/compiler/nir/nir_lower_clamp_color_outputs.c >> create mode 100644 src/compiler/nir/nir_lower_drawpixels.c >> create mode 100644 src/compiler/nir/nir_lower_io_to_temporaries.c >> create mode 100644 src/compiler/nir/nir_lower_io_types.c >> delete mode 100644 src/compiler/nir/nir_lower_outputs_to_temporaries.c >> create mode 100644 src/compiler/nir/nir_lower_passthrough_edgeflags.c >> create mode 100644 src/compiler/nir/nir_lower_wpos_ytransform.c >> create mode 100644 src/mesa/state_tracker/st_glsl_to_nir.cpp >> create mode 100644 src/mesa/state_tracker/st_glsl_types.cpp >> create mode 100644 src/mesa/state_tracker/st_glsl_types.h >> create mode 100644 src/mesa/state_tracker/st_nir.h >> create mode 100644 src/mesa/state_tracker/st_nir_lower_builtin.c >> >> -- >> 2.5.5 >> >> _______________________________________________ >> mesa-dev mailing list >> mesa-dev@lists.freedesktop.org >> https://lists.freedesktop.org/mailman/listinfo/mesa-dev _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
