[Mesa-dev] [PATCH 000/123] Reintroducing NIR, a new IR for mesa

Connor Abbott cwabbott0 at gmail.com
Tue Dec 16 14:52:29 PST 2014


Hi,

On Tue, Dec 16, 2014 at 1:04 AM, Jason Ekstrand <jason at jlekstrand.net> wrote:
> NIR (pronounced "ner") is a new IR (internal representation) for the Mesa
> shader compiler that will sit between the old IR (GLSL IR) and back-end
> compilers.  The primary purpose of NIR is to be more efficient for doing
> optimizations and generate better code for the back-ends.  We have a lot of
> optimizations implemented in GLSL IR right now.  However, they still
> generate fairly bad code primarily because its tree-based structure makes
> writing good optimizations difficult.  For this reason, we have implemented
> a lot of optimizations in the i965 back-end compilers just to fix up the
> code we get from GLSL IR.  The "proper fix" to this is to implement a
> better high-level IR; enter NIR.
>
> Most of the initial work on NIR including setting up common data
> structures, helper methods, and a few basic passes was by Connor Abbot who
> interned with us over the summer.  Connor did a fantastic job, but there is
> still a lot left to be done.  I've spent the last two months trying to fill
> in the pieces that we need in order to get NIR off the ground.  At this
> point, we now have compitent in and out of SSA passes, are at zero piglit
> regressions for i965 SIMD8 fragment shaders, and the shader-db numbers
> aren't terrible.
>
> This is still a bit experimental.  I have been testing only on HSW but it
> should work ok on SNB and later.  Eventually, once we get booleans fixed
> up, it should work fine on older chips as well.  It also doesn't yet
> support SIMD16, so performance won't be that great.  That said, I think we
> are at the point now where we should try and land this and I can stop
> developing in my masive private branch.  Since this isn't quite ready for
> prime-time yet, using it requires setting the INTEL_USE_NIR environment
> variable.
>
> A few key points about NIR:
>
>  1. It is primarily an SSA-based IR.
>  2. It supports source/destination-modifiers and swizzles/*write-masks.
>  3. Standard GPU operations such as sin() and fmad() are first-class ALU
>     operations, not intrinsics.
>  4. GLSL concepts like inputs, outputs, uniforms, etc. are built into the
>     IR so we can do proper analysis on them.
>  5. Even though it's SSA, it still has a concept of registers and
>     write-masks in the core IR data structures.  This means we can generate
>     code that is much closer to what backends want.
>  6. Control flow is structured explicitly in the IR.
>
> (*write-masks are not available for SSA values)
>
> While source/destination modifiers and writemasks/swizzles are not
> particularly useful for optimizations, having them represented in the IR
> gives us the ability to generate more useful code for backends.
>
> A few notes about review:
>
>  1. For those of you who aren't interested in the general compiler, I'm
>     sorry for the patch-bomb.  However, several people have requsted that
>     we maintain the history of the NIR development since connor's original
>     drop at the end of the summer.  Therefore, while I've squashed several
>     things, I've tried to leave the diff of what I've done more-or-less
>     preserved.
>
>  2. No, this is not LLVM.  There was a long-winded discussion about that
>     when Connor dropped his patches that went a whole lot of nowhere as
>     usual.  I would really prefer if we left that debate alone.  If there
>     must be bikeshedding on the topic, please do so on the cover-letter
>     e-mail.
>
>  3. Please keep all bikeshedding about C++, typedefs, etc.  on the core
>     datastructures e-mail.  If we need, we can split that off in its own
>     thread.
>
>  4. While I welcome review, I don't plan to make non-trivial changes to
>     specific patches or squash anything beyond what has already been
>     squashed.  I've tried thus far to more-or-less keep the history and I'd
>     like to continue this if we can.

I know you've said this, but I think there might still be some benefit
from re-arranging a few things. In particular, I think patches 21, 36,
39, 59, and 65 should probably get put first so that we can push them
+ patch 1 right away (with appropriate review), since they're not
NIR-specific. I've reviewed the ones I feel qualified to review (and
that I didn't write!) to help with this. I know I got feedback on a
few of those prep patches that we should wait to commit them until the
things they introduce have users, but I think that since there are now
patches in the list and we want to land them soon-ish it might be a
good idea to commit them earlier in order to reduce the size of this
patch-bomb :) Feel free to disagree, though...

Connor

>
>  5. Eric Anholt has also written NIR -> TGSI -> NIR passes which will
>     hopefully get landed soon after NIR initially lands.  Exactly how that
>     all gets hooked up for other gallium drivers beyond vc4 is outside the
>     scope of this series.
>
> I have pushed a branch to my personal freedesktop.org account.  For certain
> types of review, it may be easier to look at the end result rather than the
> patches.  The branch can be found via freedesktop cgit here:
>
> http://cgit.freedesktop.org/~jekstrand/mesa/log/?h=review/nir-v1
>
> Last week, I did a presentation for some of the other Intel people to try
> and help bring them up to speed on NIR concepts quickly.  As part of this,
> I typed up a bunch of notes that provide a decent overview of a lot of NIR
> concepts.  Those notes can be found here:
>
> http://www.jlekstrand.net/jason/projects/mesa/nir-notes/
>
> Happy reviewing!
>
> P.S. Connor, Don't do too much reviewing before your finals are done. :-P
>
> Connor Abbott (22):
>   exec_list: add a list_foreach_typed_reverse() macro
>   nir: add initial README
>   nir: add a simple C wrapper around glsl_types.h
>   nir: add the core datastructures
>   nir: add core helper functions
>   nir: add a printer
>   nir: add a validation pass
>   nir: add a glsl-to-nir pass
>   nir: add a pass to lower variables for scalar backends
>   nir: keep track of the number of input, output, and uniform slots
>   nir: add a pass to remove unused variables
>   nir: add a pass to lower sampler instructions
>   nir: add a pass to lower system value reads
>   nir: add a pass to lower atomics
>   nir: add an optimization to turn global registers into local registers
>   nir: calculate dominance information
>   nir: add a pass to convert to SSA
>   nir: add an SSA-based copy propagation pass
>   nir: add an SSA-based dead code elimination pass
>   i965/fs: make emit_fragcoord_interpolation() more general
>   i965/fs: Don't pass through the coordinate type
>   i965/fs: add a NIR frontend
>
> Jason Ekstrand (101):
>   i965/fs: Only use nir for 8-wide non-fast-clear shaders.
>   i965/fs_nir: Make the sampler register always unsigned
>   i965/fs_nir: Use the correct types for texture inputs
>   i965/fs_nir: Use the correct texture offset immediate
>   Fix what I think are a few NIR typos
>   Fix up varying pull constants
>   i965/fs_nir: Add support for sample_pos and sample_id
>   nir/glsl: Add support for saturate
>   nir: Add fine and coarse derivative opcodes
>   nir/glsl: Add support for coarse and fine derivatives
>   i965/fs_nir: Handle coarse/fine derivatives
>   nir/lower_atomics: Multiply array offsets by ATOMIC_COUNTER_SIZE
>   i965/fs_nir: Add atomic counters support
>   i965/fs: Allow reinterpretation in constant propagation
>   nir: Add NIR_TRUE and NIR_FALSE constants and use them for boolean
>     immediates
>   nir: Add intrinsics to do alternate interpolation on inputs
>   i965/fs: Don't take an ir_variable for emit_general_interpolation
>   i965/fs_nir: Don't duplicate emit_general_interpolation
>   nir: Add a naieve from-SSA pass
>   nir: Add a lower_vec_to_movs pass
>   i965/fs_nir: Convert the shader to/from SSA
>   nir/lower_variables_scalar: Silence a compiler warning
>   nir: Add a basic metadata management system
>   nir: Add an assert
>   nir/foreach_block: Return false if the callback on the last block
>     fails
>   nir: Add a foreach_block_reverse function
>   nir: Add a function to detect if a block is immediately followed by an
>     if
>   nir: Make the nir_index_* functions return the nuber of items
>   nir: Add an SSA-based liveness analysis pass.
>   nir: Add an initialization function for SSA definitions
>   nir: Automatically handle SSA uses when an instruction is inserted
>   nir: Add a function for rewriting all the uses of a SSA def
>   nir: Add a parallel copy instruction type
>   nir: Add a function for comparing two sources
>   nir: Add a better out-of-SSA pass
>   i965/fs_nir: Do retyping for ALU srouces in get_nir_alu_src
>   glsl/list: Fix the exec_list_validate function
>   nir: Validate all lists in the validator
>   nir/print: Don't reindex things
>   nir: Differentiate between signed and unsigned versions of find_msb
>   i965/fs_nir: Validate optimization passes
>   nir/nir: Fix a bug in move_successors
>   glsl/list: Add a foreach_list_typed_safe_reverse macro
>   nir/nir: Use safe iterators when iterating over the CFG
>   nir/nir: Patch up phi predecessors in move_successors
>   nir: Add a peephole select optimization
>   i965/fs_nir: Turn on the peephole select optimization
>   nir: Validate that the SSA def and register indices are unique
>   nir: Add a fused multiply-add peephole
>   nir: Add a basic CSE pass
>   i965/fs_nir: Add the CSE pass and actually run in a loop
>   i965/fs_nir: Use an array rather than a hash table for register lookup
>   i965/fs_nir: Handle SSA constants
>   i965/fs_nir: Properly saturate multiplies
>   nir: Add a helper for rewriting an instruction source
>   nir/lower_samplers: Use the nir_instr_rewrite_src function
>   nir: Clean up nir_deref helper functions
>   nir: Make array deref direct vs. indirect an enum
>   nir: Add a concept of a wildcard array dereference
>   nir: Use an integer index for specifying structure fields
>   nir: Don't require a function in ssa_def_init
>   nir/copy_propagate: Don't cause size mismatches on phi node sources
>   nir: Validate that the sources of a phi have the same size as the
>     destination
>   nir/glsl: Don't allocate a state_slots array for 0 state slots
>   i965/fs_nir: Don't dump the shader.
>   nir: Use the enum for the variable mode
>   nir: Automatically update SSA if uses
>   nir: Add a copy splitting pass
>   nir: Add a pass to lower local variable accesses to SSA values
>   nir: Add a pass to lower local variables to registers
>   nir: Add a pass for lowering input/output loads/stores
>   nir: Add a pass to lower global variables to local variables
>   nir/glsl: Generate SSA NIR
>   i965/fs_nir: Use the new variable lowering code
>   nir/validate: Ensure that outputs are write-only and inputs are
>     read-only
>   nir: Remove the old variable lowering code
>   nir: Vectorize intrinsics
>   nir/validate: Validate intrinsic source/destination sizes
>   nir: Add gpu_shader5 interpolation intrinsics
>   nir/glsl: Add support for gpu_shader5 interpolation instrinsics
>   nir: Add a helper for getting a constant value from an SSA source
>   i965/fs_nir: Add a has_indirect flag and clean up some of the
>     input/output code
>   i965/fs_nir: Implement the ARB_gpu_shader5 interpolation intrinsics
>   nir: Add neg, abs, and sat opcodes
>   nir: Add a lowering pass for adding source modifiers where possible
>   nir: Make the type casting operations static inline functions
>   nir/glsl: Emit abs, neg, and sat operations instead of source
>     modifiers
>   nir: Add an expression matching framework
>   nir: Add infastructure for generating algebraic transformation passes
>   nir: Add an algebraic optimization pass
>   nir: Add a basic constant folding pass
>   nir: Remove the ffma peephole
>   nir: Make texture instruction names more consistent
>   nir: Constant fold array indirects
>   nir: Use a source for uniform buffer indices instead of an index
>   nir: Add a sampler index indirect to nir_tex_instr
>   nir: Rework the way samplers are lowered
>   i965/fs_nir: Add support for indirect texture arrays
>   nir/metadata: Rename metadata_dirty to metadata_preserve
>   nir: Call nir_metadata_preserve more places
>   nir: Make bcsel a fully vector operation
>
>  src/glsl/Makefile.am                               |   10 +-
>  src/glsl/Makefile.sources                          |   39 +-
>  src/glsl/list.h                                    |   19 +-
>  src/glsl/nir/README                                |  118 ++
>  src/glsl/nir/glsl_to_nir.cpp                       | 1825 +++++++++++++++++
>  src/glsl/nir/glsl_to_nir.h                         |   40 +
>  src/glsl/nir/nir.c                                 | 2042 ++++++++++++++++++++
>  src/glsl/nir/nir.h                                 | 1433 ++++++++++++++
>  src/glsl/nir/nir_algebraic.py                      |  249 +++
>  src/glsl/nir/nir_dominance.c                       |  298 +++
>  src/glsl/nir/nir_from_ssa.c                        |  859 ++++++++
>  src/glsl/nir/nir_intrinsics.c                      |   49 +
>  src/glsl/nir/nir_intrinsics.h                      |  140 ++
>  src/glsl/nir/nir_live_variables.c                  |  282 +++
>  src/glsl/nir/nir_lower_atomics.c                   |  146 ++
>  src/glsl/nir/nir_lower_global_vars_to_local.c      |  107 +
>  src/glsl/nir/nir_lower_io.c                        |  324 ++++
>  src/glsl/nir/nir_lower_locals_to_regs.c            |  308 +++
>  src/glsl/nir/nir_lower_samplers.cpp                |  181 ++
>  src/glsl/nir/nir_lower_system_values.c             |  107 +
>  src/glsl/nir/nir_lower_to_source_mods.c            |  181 ++
>  src/glsl/nir/nir_lower_variables.c                 | 1046 ++++++++++
>  src/glsl/nir/nir_lower_vec_to_movs.c               |   96 +
>  src/glsl/nir/nir_metadata.c                        |   54 +
>  src/glsl/nir/nir_opcodes.c                         |   46 +
>  src/glsl/nir/nir_opcodes.h                         |  356 ++++
>  src/glsl/nir/nir_opt_algebraic.py                  |   67 +
>  src/glsl/nir/nir_opt_constant_folding.c            |  355 ++++
>  src/glsl/nir/nir_opt_copy_propagate.c              |  325 ++++
>  src/glsl/nir/nir_opt_cse.c                         |  269 +++
>  src/glsl/nir/nir_opt_dce.c                         |  186 ++
>  src/glsl/nir/nir_opt_global_to_local.c             |  103 +
>  src/glsl/nir/nir_opt_peephole_select.c             |  214 ++
>  src/glsl/nir/nir_print.c                           |  948 +++++++++
>  src/glsl/nir/nir_remove_dead_variables.c           |  138 ++
>  src/glsl/nir/nir_search.c                          |  337 ++++
>  src/glsl/nir/nir_search.h                          |   80 +
>  src/glsl/nir/nir_split_var_copies.c                |  225 +++
>  src/glsl/nir/nir_to_ssa.c                          |  660 +++++++
>  src/glsl/nir/nir_types.cpp                         |  143 ++
>  src/glsl/nir/nir_types.h                           |   75 +
>  src/glsl/nir/nir_validate.c                        |  912 +++++++++
>  src/mesa/drivers/dri/i965/Makefile.sources         |    1 +
>  src/mesa/drivers/dri/i965/brw_fs.cpp               |   74 +-
>  src/mesa/drivers/dri/i965/brw_fs.h                 |   57 +-
>  .../drivers/dri/i965/brw_fs_copy_propagation.cpp   |    4 +-
>  src/mesa/drivers/dri/i965/brw_fs_fp.cpp            |   32 +-
>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp           | 1778 +++++++++++++++++
>  src/mesa/drivers/dri/i965/brw_fs_visitor.cpp       |   39 +-
>  src/mesa/main/bitset.h                             |    1 +
>  50 files changed, 17301 insertions(+), 77 deletions(-)
>  create mode 100644 src/glsl/nir/README
>  create mode 100644 src/glsl/nir/glsl_to_nir.cpp
>  create mode 100644 src/glsl/nir/glsl_to_nir.h
>  create mode 100644 src/glsl/nir/nir.c
>  create mode 100644 src/glsl/nir/nir.h
>  create mode 100644 src/glsl/nir/nir_algebraic.py
>  create mode 100644 src/glsl/nir/nir_dominance.c
>  create mode 100644 src/glsl/nir/nir_from_ssa.c
>  create mode 100644 src/glsl/nir/nir_intrinsics.c
>  create mode 100644 src/glsl/nir/nir_intrinsics.h
>  create mode 100644 src/glsl/nir/nir_live_variables.c
>  create mode 100644 src/glsl/nir/nir_lower_atomics.c
>  create mode 100644 src/glsl/nir/nir_lower_global_vars_to_local.c
>  create mode 100644 src/glsl/nir/nir_lower_io.c
>  create mode 100644 src/glsl/nir/nir_lower_locals_to_regs.c
>  create mode 100644 src/glsl/nir/nir_lower_samplers.cpp
>  create mode 100644 src/glsl/nir/nir_lower_system_values.c
>  create mode 100644 src/glsl/nir/nir_lower_to_source_mods.c
>  create mode 100644 src/glsl/nir/nir_lower_variables.c
>  create mode 100644 src/glsl/nir/nir_lower_vec_to_movs.c
>  create mode 100644 src/glsl/nir/nir_metadata.c
>  create mode 100644 src/glsl/nir/nir_opcodes.c
>  create mode 100644 src/glsl/nir/nir_opcodes.h
>  create mode 100644 src/glsl/nir/nir_opt_algebraic.py
>  create mode 100644 src/glsl/nir/nir_opt_constant_folding.c
>  create mode 100644 src/glsl/nir/nir_opt_copy_propagate.c
>  create mode 100644 src/glsl/nir/nir_opt_cse.c
>  create mode 100644 src/glsl/nir/nir_opt_dce.c
>  create mode 100644 src/glsl/nir/nir_opt_global_to_local.c
>  create mode 100644 src/glsl/nir/nir_opt_peephole_select.c
>  create mode 100644 src/glsl/nir/nir_print.c
>  create mode 100644 src/glsl/nir/nir_remove_dead_variables.c
>  create mode 100644 src/glsl/nir/nir_search.c
>  create mode 100644 src/glsl/nir/nir_search.h
>  create mode 100644 src/glsl/nir/nir_split_var_copies.c
>  create mode 100644 src/glsl/nir/nir_to_ssa.c
>  create mode 100644 src/glsl/nir/nir_types.cpp
>  create mode 100644 src/glsl/nir/nir_types.h
>  create mode 100644 src/glsl/nir/nir_validate.c
>  create mode 100644 src/mesa/drivers/dri/i965/brw_fs_nir.cpp
>
> --
> 2.2.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev


More information about the mesa-dev mailing list