Rather than a per-global-slot initializer region, we now have one for
the whole module. For example, it might look like this:
```
torch.global_slot "private" @tensor : !torch.tensor
torch.global_slot "private" @list : !torch.list<tensor>
torch.global_slot.module_initializer {
%0 = torch.tensor.literal(dense<0.0> : tensor<f32>) : !torch.tensor
%1 = torch.prim.ListConstruct %0 : (!torch.tensor) -> !torch.list<tensor>
torch.initialize.global_slots [
@tensor(%0 : !torch.tensor)
@list(%1 : !torch.list<tensor>)
]
}
```
This new structure allows GlobalizeObjectGraph to create the initializer in a
much simpler way, avoiding the need to reason about whether different slots
alias each other. Reasoning about whether slots alias each other now is the
responsibility of InlineGlobalSlots, which has to do a much more complicated
analysis, implemented using MLIR's dataflow analysis framework.
Recommended review order:
- Check out the new IR constructs in the .mlir files of various passes
- Op definitions (*.td)
- Changes to GlobalizeObjectGraph pass.
- InlineGlobalSlots pass (~total rewrite)
- Misc changes:
- Moving torchMlirAdjustStaticInformation for sharing with C++ code.
- EraseModuleInitializer pass
To make this a bit nicer, it would be good to have a `torch.module` op
with an initializer region attached. That would be more invasive though.
This change has highlighted certain aspects of our project layering
which are worth calling out. None of our backends can handle global
slots, so we enforce that there are no global slots before backend
lowering. At an earlier stage in the project, we had aspirations of
transparently handling mutable global state and such, but for reasons
described below, that is no longer a goal. So really global slots should
be seen as a progressive lowering step as part of inlining all the
IValue's in the original program (GlobalizeObjectGraph is also one such
step).
Over time, with insights from work like IREE-JAX, it has become clear
that there isn't a reliable programming model we can compile for users
where we just transparently handle mutable global state (and some other
things, like lists and dictionaries). There is a need for an "outer
program" that orchestrates more restricted subroutines of the kind we
can handle in our compile flow here. The benefit of that is that it
decouples considerations like shapes, dtypes, etc. from the program
constructs used in the outer program. As long as the outer program can
efficiently invoke (pipelining/async/etc.) high-performance
data-parallel numerical subroutines of the kind we compile in our flow
here, then there is a complete programming model. This is also
consistent with the direction of upstream PyTorch which is becoming more
tracing-based (which inherently loses a lot of program structure, which
then has to be applied back with an "outer program" orchestrating the
traced subroutines).
The updated LLVM code includes a patch to create bfloat16 array
attributes, thus enabling a different patch to torch-mlir to flesh out
support for the bfloat16 type.
This creates the `external/torch-mlir` directory as an
LLVM_EXTERNAL_PROJECTS-compatible project (analogous to
`iree-dialects`) and completes movement/rename of all pure MLIR C/C++
compiler code into there. The next step will be to move all the Python
code / code that links/includes PyTorch C++ code (which currently lives
in `frontends/pytorch`) into a subdirectory here.
I call this "earthmoving" because it is mostly mechanical changes and
renames. As a quick summary (we can change this down the road easily)
- C++ `mlir::NPCOMP::Torch -> mlir::torch::Torch`
- CAPI `npcompTorchListTypeGet -> torchMlirTorchListTypeGet`
- preprocessor `#ifndef NPCOMP_ -> #ifndef TORCHMLIR_`
- CMake `NPCOMPFoo -> TorchMLIRFoo`
The goal of this is to create a standalone project creating a center of
mass for entry into the MLIR ecosystem from PyTorch, suitable in scope
for eventual inclusion/ownership in PyTorch. The idea is that
`external/torch-mlir` will some day be pulled out into its own
repository, and then npcomp will simply pull it in as a submodule.
Layering-wise, what lives in `torch-mlir` lowers code from PyTorch
(currently TorchScript, but TorchFX or pytorch/xla-style tracing are
possible extensions) down to what we have been calling the "Torch
backend contract" which is cleaned up IR (inlining, simplifcation,
conversion to value tensors, ...) entirely in the `torch` dialect. This
is the branching off point for further lowering, of which npcomp takes
one opinion (outside `torch-mlir` of course!), namely the
`TorchConversion` dialect/transforms which lower to IR suitable for IREE
and other linalg-on-tensors based lower-level compilers.
Summary of changes:
- move `{include,lib,test}/Dialect/Torch` into `torch-mlir`
- move relevant parts of CAPI into `torch-mlir`.
- leave a few things related to the `torch-mlir` Python build commented
out, which should be resolved in a subsequent change.
With the following changes the compilation can continue until
RefineTypes pass:
- Add operators without ODS into `torch_ods_gen.py`
- Add some new optional and list types in `TorchTypes.td`
- Add some folders for aten int type comparator ops
- Modify GlobalizeObjectGraph.cpp. For global slots that's not used,
dont check if an aliased value is stored in more than one of global
slots. This can work around a failure where the same tensor is stored
in multiple "version" slots which are not used.
This op is much better behaved than the `torch.tensor.literal` op
(which is the new name of the `torch.tensor` op). In particular
`torch.tensor.literal`:
- always has a maximally refined type.
- always has value semantics.
- can be constant folded / CSE'd.
ReduceOpVariants is changed to perform the transformation from
`torch.tensor.literal` to `torch.vtensor.literal` (which in general
involves static information casts and copies.
This new op also allowed tightening up `torch.tensor.literal` to only
accept NonValueTensorType (instead of any tensor type).
This new ".literal" name is more descriptive. It was getting too
confusing seeing an op called just `torch.tensor` (we originally called
it that because that's the name of the similar function in the Torch
Python API, but it just doesn't fit here).
This removes the dependence of the `torch` dialect on the low-level
builtin types.
Now the `torch` dialect is a standalone layer, suitable for targeting
from higher-level Python abstractions without any premature lowering to
primitive types.
This removes our reliance on the numpy dialect and avoids our off-label
use of the builtin tnesor type for modeling unknown dtypes. The
`!torch.vtensor` (`ValueTensorType`) type is a value-semantic tensor.
The `!torch.tensor` (`NonValueTensorType`) type is a non-value-semantic
tensor. The new types look as follows syntactically:
```
// Least-static-information, non-value-semantic tensor.
!torch.tensor
// Explicit form of least-static-information variant.
!torch.tensor<*,unk>
// Least-static-information, value-semantic tensor.
!torch.vtensor
// Explicit form of least-static-information variant.
!torch.vtensor<*,unk>
// Fixed-set of allowable element types, with first-class support for
// Torch's frontend signedness semantics.
!torch.tensor<*,si32>
// First-class support for unknown dtypes.
!torch.tensor<[?,?,?],unk>
// Standard MLIR representation of `?` for unknown dimensions.
!torch.tensor<[?,2,?,4],unk>
// Statically shaped / dtyped example.
!torch.vtensor<[1,2,3,4],f32>
```
This required fairly significant changes throughout the compiler, but
overall it is a big cleanup. We now have a much clearer layering of "the
Torch frontend lowering" vs "lowering to std + linalg + etc.".
At the C++ level, there is `ValueTensorType`, `NonValueTensorType`.
We also have a helper `BaseTensorType` (kind of like ShapedType) which
interoperates with those two.
Included changes:
- New `torch.tensor(dense<0.0> : tensor<5xf32>) : !torch.tensor` op for
creating torch tensor literals in the frontend.
- Consistently use signedness for the types (except i1 which I didn't
touch -- we need to sort out the situation with !basicpy.BoolType
there anyway so will be attending to that soon)
- Frontend can annotate whether an argument to the function has value
semantics. We currently require this, as our backend contract does not
currently allow us to even model the non-value-semantic case. Before,
the value-semantic assumption was randomly injected in the middle of
the pass pipeline.
- Move ArrayToTensor (now called MaximizeValueSemantics) and
RefinePublicReturn passes to torch dialect.
- The TorchToStd and TorchToLinalg passes are now type conversions from
`!torch.vtensor` to `tensor` and use the dialect conversion infra.
The overall conversion pipeline is set up following the best practices
of the "Type Conversions the Not-So-Hard Way" talk. This required
introducing `torch-func-builtin-tensorize` and
`torch-finalizing-builtin-tensorize` passes analogous to the upstream
bufferization passes with the corresponding names (mostly just
copypasta from there).
- Misc Torch-level canonicalizations -- we now cleanly layer the
lowering to std later in the pipeline, so we are gradually lessening
our reliance on random std constant folding before we get to that
point.
Recommended review order:
- New types in TorchTypes.td/TorchTypes.h/TorchDialect.cpp
- New ops in TorchOps.td / TorchOps.cpp
- Less important / more mechanical stuff
- Frontend changes.
- Pass changes/additions in `Torch/Transforms` and `Conversion/`
Sean Silva
Ashay Rane
Yi Zhang