[Mesa-dev] [RFC PATCH 00/16] A new IR for Mesa

[Francisco Jerez]

Connor Abbott <cwabbott0 at gmail.com> writes:

> I know what you might be thinking right now. "Wait, *another* IR? Don't
> we already have like 5 of those, not counting all the driver-specific
> ones? Isn't this stuff complicated enough already?" Well, there are some
> pretty good reasons to start afresh (again...). In the years we've been
> using GLSL IR, we've come to realize that, in fact, it's not what we
> want *at all* to do optimizations on. Ian has done a talk at FOSDEM that
> highlights some of the problems they've run into:
>
> https://video.fosdem.org/2014/H1301_Cornil/Saturday/Three_Years_Experience_with_a_Treelike_Shader_IR.webm
>
> But here's the summary:
>
> * GLSL IR is way too much of a memory hog, since it has to make a new
> variable for each temporary the compiler creates and then each time you
> want to dereference that temporary you need to create an
> ir_dereference_variable that points to it which is also very
> cache-unfriendly ("downright cache-mean!").
>
> * The expression trees were originally added so that we could do
> pattern matching to automatically optimize things, but this turned out
> to be both very difficult to do and not very helpful. Instead, all it
> does is add more complexity to the IR without much benefit - with SSA or
> having proper use-def chains, we could get back what the trees give us
> while also being able to do lots more optimizations.
>
> * We don't have the concept of basic blocks in GLSL IR, which makes a
> lot of optimizations harder because they were originally designed with
> basic blocks in mind - take, for example, my SSA series. I had to map a
> whole lot of concepts that were based on the control flow graph to this
> tree of statements that GLSL IR uses, and the end result wound up
> looking nothing at all like the original paper. This problem gets even
> worse for things like e.g. Global Code Motion that depend upon having
> the dominance tree.
>
> I originally wanted to modify GLSL IR to fix these problems by adding
> new instruction types that would address these issues and then
> converting back and forth between the old and the new form, but I
> realized that fixing all the problems would basically mean a complete
> rewrite - and if that's the case, then why don't we start from scratch?
> So I took Ken's suggestions and started designing, and then at Intel
> over the summer started implementing, a completely new IR which I call
> NIR that's at a lower level than GLSL IR, but still high-level enough to
> be mostly device-independant (different drivers may have different
> passes and different ways of lowering e.g.  matrix multiplies) so that
> we can do generic optimizations on it. Having support for SSA from the
> beginning was also a must, because lots of optimisations that we really
> want for cleaning up DX9-translated games are either a lot easier in or
> made possible by SSA. I also made the decision for it to be typeless,
> because that's what the cool kids are all doing :) and for a
> lower-level, flat IR it seemed like the thing to do (it could have gone
> either way, though). So the key design points of NIR (pronounced either
> like "near" as in "NIR is near!" or to rhyme with "burr") are:
>
> * It's flat (no expression trees)
>
> * It's typeless
>
> * Modifiers (abs, negate, saturate), swizzles, and write masks are part
> of ALU instructions
>
> * It includes enough GLSL-like things (variables that you can load from
> or store to, function calls) to be hardware-agnostic (although we don't
> have a way to represent matrix multiplies right now, but that could
> easily be added) to be able to do optimizations at a high level, while
> having lowering passes that convert variables to registers and
> input/output/uniform loads/stores that will open up more opportunities
> for optimization and save memory while being more hardware-specific.
>
> * Control flow consists of a tree of if statements and loops, like in
> GLSL IR, except the leaves of the tree are now basic blocks instead of
> instructions. Also, each basic block keeps track of its successors and
> predecessors, so the control flow graph is explicit in the IR.
>
> * SSA is natively supported, and SSA uses point directly to the SSA
> definition, which means that the use-def chains are always there, and
> def-use chains are kept by tracking the set of all uses for each
> definition.
>
> * It's written in C.
>
> (see the README in patch 3 and nir.h in patch 4 for more details)
>
> Some things that are missing or could be improved:
>
> * There's currently no alias tracking for inputs, outputs, and uniforms.
> This is especially important for uniforms because we don't pack them
> like we pack inputs and outputs.
>
> * We need a way to represent matrix multiplies so that we can do
> matrix-flipping optimizations in NIR (currently GLSL IR does this for
> us).
>
> * I'm not entirely happy about how we represent loads and stores in the
> IR. Right now, they're intrinsics, but that means we need a different
> intrinsic for each size and combination of arguments (indirect vs. not
> indirect, etc.) and we might run into a combinatorial explosion problem
> in the future, so we might need to make separate load/store instructions
> like what I did for textures.
>
> * Right now, we only have a pass that lowers variables for scalar
> backends. We need to write a similar pass for vector backends that uses
> std140 packing or something similar, as well as porting
> lower_ubo_reference to NIR and changing it to output offsets in the
> hardware-native units instead of bytes.
>
> * We'll need to write a pass that splits up vector expressions for
> scalar backends.
>
> The first two patches are preperatory patches that I already sent to the
> list, but I'm re-sending them as part of the series as they haven't
> landed yet. Right now, the series only has code to convert GLSL IR to
> NIR, but no way to actually hook it up to a backend in order to generate
> code from it, and it also doesn't do anything with the SSA part of the
> IR. I have a branch on my Github that does the conversion to SSA and a
> few simple SSA-based optimizations, which hasn't been tested as much
> (since I haven't written a pass to get out of SSA or a backend that uses
> SSA):
>
> https://github.com/cwabbott0/mesa/tree/nir
>
> and an experimental backend for i965 fs that I hope to combine with
> Matt's SSA work; right now, there are only a few piglit regressions and
> most of them are because of the hacky way I changed boolean true to be
> 0xFFFFFF instead of 1 (with Matt's series to do the same thing in a
> better way, they should go away) or because of unimplemented features
> (atomics and some system values):
>
> https://github.com/cwabbott0/mesa/tree/nir-i965-fs
>
> NIR has been what I've worked on for my entire summer internship at
> Intel, and before I go off to my freshman year at college, I'd like to
> thank the other Intel folks for the knowledge they've given me and the
> many interesting discussions that made this go from an idea to a reality
> - I'll miss you guys!
>
> Connor
>
> Connor Abbott (16):
>   exec_list: add a list_foreach_typed_reverse() macro
>   glsl/linker: pass through the is_intrinsic flag
>   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
>
>  src/glsl/Makefile.sources                 |   18 +-
>  src/glsl/link_functions.cpp               |    2 +
>  src/glsl/list.h                           |    6 +
>  src/glsl/nir/README                       |  118 ++
>  src/glsl/nir/glsl_to_nir.cpp              | 1759 +++++++++++++++++++++++++++++
>  src/glsl/nir/glsl_to_nir.h                |   40 +
>  src/glsl/nir/nir.c                        | 1717 ++++++++++++++++++++++++++++
>  src/glsl/nir/nir.h                        | 1270 +++++++++++++++++++++
>  src/glsl/nir/nir_intrinsics.c             |   49 +
>  src/glsl/nir/nir_intrinsics.h             |  158 +++
>  src/glsl/nir/nir_lower_atomics.c          |  127 +++
>  src/glsl/nir/nir_lower_samplers.cpp       |  170 +++
>  src/glsl/nir/nir_lower_system_values.c    |  106 ++
>  src/glsl/nir/nir_lower_variables_scalar.c | 1243 ++++++++++++++++++++
>  src/glsl/nir/nir_opcodes.c                |   46 +
>  src/glsl/nir/nir_opcodes.h                |  346 ++++++
>  src/glsl/nir/nir_opt_global_to_local.c    |  103 ++
>  src/glsl/nir/nir_print.c                  |  916 +++++++++++++++
>  src/glsl/nir/nir_remove_dead_variables.c  |  138 +++
>  src/glsl/nir/nir_types.cpp                |  155 +++
>  src/glsl/nir/nir_types.h                  |   78 ++
>  src/glsl/nir/nir_validate.c               |  798 +++++++++++++
>  22 files changed, 9362 insertions(+), 1 deletion(-)
>  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_intrinsics.c
>  create mode 100644 src/glsl/nir/nir_intrinsics.h
>  create mode 100644 src/glsl/nir/nir_lower_atomics.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_variables_scalar.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_global_to_local.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_types.cpp
>  create mode 100644 src/glsl/nir/nir_types.h
>  create mode 100644 src/glsl/nir/nir_validate.c
>

After reading through your patch series, even though I agree with many
of your design decisions (e.g. the IR being SSA, flat, structured in
BBs), I'm still mildly skeptical about this.  I don't see any real
benefits over e.g. LLVM IR that would justify inventing a new IR from
scratch and rewriting the world.  LLVM IR already has a mature and
extensive set of algorithms and optimization passes, and a few drivers
already support it as input in some way (radeon has the r600/amdgpu
back-end, intel has at least beignet's "back-end") or are planning
support for it (nouveau).  There's also previous work on translating
GLSL IR into LLVM IR, and a Khronos standard based on LLVM IR that we
could use as starting point if unstable APIs are a concern.

It seems that we'll have to keep dealing with LLVM IR anyway because of
OpenCL, and the current way we handle this sucks (maintain two separate
compilers, one for graphics and another for compute).  It would be great
if we could come up with a solution that addresses all the weak points
of GLSL IR we have found along the way and at the same time gets rid of
the duplicated compilers...
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 212 bytes
Desc: not available
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20140819/ef47367d/attachment-0001.sig>