This adds a pattern to MaximizeValueSemantics which does a simple
abstract interpretation within a block, which handles simple cases of
`torch.overwrite_tensor`, enough to remove all the unnecessary uses of
non-value tensors in ResNet right now.
Before/after IR:
[gist](https://gist.github.com/silvasean/a3e1ef625b19dfc63579f73cd3b543b6)
Also,
- Split `torch.copy.tensor` into `torch.copy.to_tensor` and
`torch.copy.to_vtensor` which convert between value and non-value
semantic tensors. This is a much cleaner factorization as they have
very separate use cases and properties (e.g. different side effects)
- Remove the various canonicalization patterns they had, which were
confusing because they resulted in limited forms of maximizing value
semantics throughout the pipeline. We should structure our compilation
pipeline such that only MaximizeValueSemantics should be maximizing
value semantics.
- Adjust pass pipeline to only run MaximizeValueSemantics once.
- Make OverwriteTensorOp `$value` always be a value tensor and
`$overwritten` be a non-value tensor.
1. Added a simplified version of torch.aten.batch_norm which only handles
inference and assumes the weight, bias, running_mean, running_var are not
None.
2. Removed the primitive types check in verifyLinalgCompatibleTypes check
since now we have proper type converter to handle torch types conversion.
The checks for RankedTensorType is kept because the type converter
doesn't guarantee the converted builtin tensor type is ranked. A
separate verification pass to verify the invariant expected by later
passes will need to be added before those can be removed as well.
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 replaces the ad-hoc use of `i64` throughout the Torch layer, and
helps to keep it crystal clear the distinction between `!torch.int`
(which is modeling the Python `int` type) and the various types that
serve as dtypes of tensors, which are a totally different type universe.
Changes:
- `!torch.int` type and C bindings.
- Change `torch.constant.int` parser to not need the `: i64` at the end.
- `m_TorchConstantInt` matcher to aid with matching constants.
- BackendTypeConversion changes for `!torch.int` -> `i64` type
conversion.
- Refactor finalizing patterns in FinalizingBackendTypeConversionPass
(they were getting very repetitive).
- Mechanical rewriting of `!torch.int` to `i64` in all the tests, and
`AnyTorchIntType` to `Torch_IntType` in the `.td` files.
This fixes a "regression" on ResNet where we weren't folding away all
the control flow. For now, our policy is to "optimize hard enough" to
make that control flow go away, because we don't yet have a way to lower
to the backend the stuff guarded by the control flow (RaiseException,
string operations, etc.).
It remains to be seen how much optimization we decide to do at this
level in the fullness of time -- the torch op set is not particularly
well-designed (at least not idiomatically for MLIR) for general
optimization. Ideally, with really good backend support for various
features, all the heavy optimization will happen at that layer on `std`
ops and `scf` control flow. But I have a suspicion we might end up
needing more optimization earlier in the pipeline.
This removes the use of `scf.if`, which required laundering back and
forth between `i1` and `!torch.bool` in the frontend. We will eventually
lower this op to `scf.if`, but this results in a cleaner IR and layering
at the frontend.
This finishes removing the dependence on the basicpy dialect!
Changes:
- Add `!torch.bool` type and replace use of `!basicpy.BoolType` in
Torch-related code.
- Rename BuiltinTensorize to BackendTypeConversion since now it handles
bool conversions (and, when we add !torch.int and !torch.float, it
will handle those as well), and generalize the related utilities (I
also moved them to Torch/Transforms since they aren't really part of
Torch/IR).
- Add `torch.to_i1` and `torch.from_i1` ops for materializations
- [cleanup] Reorganize `torch.constant.*` ops in TorchOps.td
- Remove dependency of `torch` dialect on `basicpy` dialect and also
`std` dialect. For `std`, we use some call related ops, but the
`torch` dialect itself never produces them (we have passes that do
though).
This is fairly mechanical. Recommended review order:
- New stuff in Torch/IR
- New BuiltinTypeConversion files.
- Mechnical fixups elsewhere.
- Make consistent with MLIR Core
- Use `//` or `///` comments.
- Use `bool` type for booleans
- No duplicated comments in .cpp files
- Split types into separate files `{Basicpy,Numpy,Torch}Types.h`
- Add dialect prefix consistently to C API symbols. We have lots of
similarly named types (e.g. "list" type in basicpy and torch).
- Add `torch.constant.none` op to construct it (naming is chosen to be
analogous to Torch's representation of a prim::Constant with
NoneType, rather than using the "singleton" terminology of Basicpy).
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/`
This now gives [much nicer output](https://gist.github.com/silvasean/f048e0f37b04542dae6469b86802bb3e).
Embarrassingly, we previously couldn't even report failures for two
different tests, and weren't able to report on compilation failures
(besides just crashing).
This is enough to import the program and get it through the compilation
pipeline. It of course fails at the VerifyBackendContract pass since
there is a lot missing, but the final IR for a simple quantized MLP is
looking pretty decent already:
[IR](https://gist.github.com/silvasean/f76bccd76e9b193d396cfb2f9a11f54d)
Main changes:
- Add support for importing torch quantized tensors, including
`torch.per_tensor_affine.create` op and `!torch.qint8` element type.
- Add support for importing `LinearPackedParamsBase` (basically a weight
+ optional bias, but requires `torch.linear_params.create` op +
`!torch.LinearParams` type to model it). This was less painful than I
expected, as it has the necessary methods to opaquely unpack itself. I
factored things so it should be easy to extend to other custom classes
like `ConvPackedParamsBase`.
- Add minimal boilerplate for importing `quantized::*` ops, with
`quantized::linear` being a motivating example.
- Add e2e test with simple quantized MLP (courtesy of @phoenix-meadowlark).
This is somewhat of an abuse of `!numpy.ndarray` / `tensor`, as
really the proper semantics of `!torch.qint8` dtype on a Torch tensor is
"check the quantizer object of the tensor for side data (scale/offset,
possibly per-channel) that defines the full semantics of the tensor". We
don't have any such notion of "side data" for `!numpy.ndarray` /
`tensor`, let alone anything that would have the associated behavior of
keying off the dtype to determine if the side data is present.
This will be fixed by a proper `!torch.tensor` type.
This is a really major and invasive restructuring of the way we get
torch operators (`torch::jit::Operator` / `c10::OperatorHandle`) into
MLIR. Please forgive the challenging review, but due to the sheer
invasiveness, it wasn't really practical do do it in sane smaller
pieces.
This fully replaces everything that was already working on the
TorchScript path (actually, more -- we added tanh support to
TorchToLinalg in order to delete the older code paths). Additionally,
I've kept the lights on for the acap path too, including what little e2e
stuff was working before (for expediency I made a few tiny compromises
along the way that will be easy to undo when we give that path proper
attention).
Overview of the new design:
- The torch operator `somens::someunqualname.someoverloadname` is
imported as `torch.somens.someunqualname.someoverloadname` (skip the
last dotted part if the overload name is empty), OR, if we don't have
such an op registered, it is imported as
`torch.operator "somens.someunqualname.someoverloadname" (...) : ...`.
- The addition of the "overload name" is a critical element here, as
the `(ns,unqual,overload)` triple is unique, which solves a lot of
problems we were having.
- This involves having separate MLIR ops for the `trailing_` and
`.out` variants and all the different overloads. This seemed
necessary, because the set of overloads is so wild and varied and
unstructured. The previous design was leaning into some underlying
structure that just isn't there -- the default situation is
the "random overload that we want to manage on the MLIR side",
rather than that being an exception. E.g. `aten::ne` (not-equal)
has 21 overloads, only 4 of which are c10 dispatcher ops see
[gist](https://gist.github.com/silvasean/190ba918c550c956260e21254e1b8aa1),
and the "out" variant is really called `.Tensor_out` instead of
`.out` as it frequently is for other ops.
- Rationale for all being in `torch` namespace: the set of operators
are so varied and unstructured that "dialect per namespace"
doesn't result in anything resembling the typical MLIR dialect
boundary expectations. We could maybe draw the boundary at
dispatcher ops vs non-dispatcher ops, but that doesn't seem to
really result in very much useful structure at this point in time.
- Note: within the torch operator registry, we effectively have a
mini-basicpy subdialect (already type-resolved), which is reasonably
structured.
- The existing Torch op interfaces are also removed -- now that we
track the overload name, we can losslessly find the original
operator.
- Instead of `ATenRecognizeKernelsPass`, we now have a
`ReduceOpVariantsPass` that keys off certain traits (and perhaps
eventually interfaces) to reduce variants of ops to a smaller set,
ideally operating on immutable tensors and using surrounding ops to
model the mutability/aliasing aspects.
- Note: `torch.ns.unqual.overload` ops allow both immutable and
mutable tensors (unlike the previous hard distinction in the common
case). This is a premonition for a future change that will introduce a
bona fide `!torch.tensor` type that will clean up a bunch of stuff.
- `TorchToLinalg` / `TorchToStd` supercede the existing
"ATen->TCF->TCP->Linalg" path.
- The new `torch_ods_gen.py` supercedes `torch_signature_ods_gen.py`.
It should look somewhat familiar, but the benefit of hindsight has
allowed a lot of simplifications.
The overall trend seems to be to make the `torch` dialect a nice layer
independent of anything else. It feels like as a natural result of
various future changes we will be removing the reliance on basicpy+numpy
dialects and have a nice self-contained type system too that properly
models the TorchScript type system (including proper subtyping,
mutable/immutable tensors, optional dtype, etc.).
Recommended review order:
- Start at some of the new import IR, e.g. in
`frontends/pytorch/test/node_import/prim.py`,
`frontends/pytorch/test/acap_export/test_export_add3.py`, and other
tests.
- `frontends/pytorch/python/torch_mlir_utils/codegen/torch_ods_gen.py`
and associated generated files:
- `include/npcomp/Dialect/Torch/IR/GeneratedAtenOps.td`
- `include/npcomp/Dialect/Torch/IR/GeneratedPrimOps.td`
- Inspect `ReduceOpVariants.cpp` / `reduce-op-variants.mlir` and the new
traits in `include/npcomp/Dialect/Torch/IR/TorchTraits.h`
- Various code changes in the import path in
`frontends/pytorch/csrc/builder`. Probably most interesting is the new
code in `torch_to_mlir_utils.cpp` that has the logic to create the
`torch.operator` ops or `torch.ns.unqual.overload` ops.
This is the [new ResNet IR](https://gist.github.com/silvasean/5407aafb710d07612b7b5b92eabecebe),
just to be able to look at a substantial sample of IR in the new style.
- aten::relu_, aten::max_pool2d, aten::adaptive_avg_pool2d, aten::batch_norm, aten::conv2d
No aten-to-linalg conversion for the latter ones, as they are fairly
substantial. At this point, I'm trying to get shape inference and stuff
working for them and the IR cleaned up.
This trait lets us model the semantics of various aten/torch/numpy ops
that are insensitive to type refinements. This replaces
hardcoded/inconsistent checks for this property.
To show usage of this new trait, we fix up some old uses, and improve
RefineTypes to be smarter about rewriting with this trait.
Interestingly, TorchScript has its own op (`torch::jit::Operator`)
registry separate from the dispatcher (it is a superset of the
dispatcher).
This is where the "prim" ops and some "aten" ops (that should probably
be renamed to "prim") live. In particular, `aten::__is__` is in that
latter category of "aten but really prim". This registry is also the
source of truth for what the TorchScript interpreter calls into when it
executes.
The bulk of the "not part of the dispatcher" ops live in
09feb5f579/torch/csrc/jit/runtime/register_prim_ops.cpp (L82)
And the registry itself lives in:
09feb5f579/torch/csrc/jit/runtime/operator.cpp (L196)
This fold further reduces the IR of ResNet by folding away some
more not-taken branches. These not-taken branches in ResNet require
first-class handling of the list type which we don't yet have on any
backend.
This is the start of a push to getting ResNet running.
This involves throwing in the towel on an O0 pipelinie for now. See note
in the code. We keep an options struct with `optimize` flag, but it
default to true for now.
This removes the need for defining all of the custom propagation logic,
and also adds support for propagating value knowledge across branches,
through regions, and across calls.
These tests pass on the reference backend.
- Add aten.linear op + shape xfer function + ATen->Linalg lowering.
- Note: this needs to be more automated, and needs to cover more cases.
- Current not implemented caveats:
- size-1 broadcasting for bias vector (either static-size-1 or ? case)
- higher-rank aten.linear ops (not produced by torch.nn.Linear though)
- type promotion (still don't even know the exact rules here)
- Add folder for torch.derefine op. Now the inliner can clean it up as
it inlines. (call boundaries are a main place we need to insert
torch.derefine) This is brittle -- the other important case is control
flow which will need to be handled via an extension to
RefineTypes.cpp (as will more robust call handling). River has an
in-flight patch to update it to the new dataflow framework so I didn't
want to do anything intrusive here.
- Also adjust torch.derefine syntax to use the keyword `to` instead of
`->`, as most type-only, cast-like ops do.
This inlines global slots if possible. This allows them to participate
in folding, canonicalization, shape inference, etc.
Example use cases:
- inlining weights and biases that are readonly during inference
- inlining the "training" bool to allow stuff to fold away
For training use cases (especially internal training loop), we will need
something smarter to get good performance. That would look like an "SSA
formation" which promotes the global slots to tensors in the program,
flushing them back to the slots at the minimal number of necessary
places. We might want to let backends do that transformation though.
This also interacts with shape inference (type bounds on the slots to
even lower them to backends in the first place).
- Move frontend lowering pipelines to c++ (this helps with reproducing
failures in npcomp-opt)
- Add debugging printouts when compilation fails on RefBackendTestConfig
The experience now when a test fails during MLIR lowering is now like this:
```
NPCOMP TorchScript Object Graph IR -> NPCOMP Backend IR lowering failed with the following diagnostics:
failed to legalize operation 'torch.global_slot'
Module does not conform to npcomp's backend contract. See dialect conversion legality information above.
Error can be reproduced with:
$ npcomp-opt -torchscript-to-npcomp-backend-pipeline /tmp/ResNet18Module.mlir
```
And when TorchScript->MLIR import fails it looks like this:
```
PyTorch TorchScript module -> NPCOMP Object Graph IR import failed with the following diagnostics:
unhandled prim operation: %18 : int = prim::min(%17) # /usr/local/google/home/silvasean/.local/lib/python3.9/site-packages/torch/nn/functional.py:4532:4
```
Also,
- Add `--filter=<regex>` to e2e test harness to filter tests.
- Add a few prim ops that were needed to import ResNet18
- Fix torch.prim.Loop.condition assemblyFormat (it previously would not
round-trip in the case of no loop-carried variables)
As described in the code comment:
```
When we import TorchScript IR, we import their entire "compilation unit",
which can contain numerous functions unrelated to the current program,
which breaks torch-globalization-pipeline; for example, there can be
random functions referencing types that haven't been imported
as part of the root `torch.nn.Module` we imported. Those will
be unreferenced private functions which symbol-dce will clean up nicely.
```
This situation is really easy to hit in jupyter notebooks, where the
same cell is evaluated multiple times. That results in the same
class name (at the Python level, e.g. class `Foo` in the top-level
main module). Internally to PyTorch, it handles this situation by
mangling in a unique number to the names of ClassType's and such. When
we import the new ClassType's, we see not just the new
torch::jit::Function's in the CompilationUnit, but, also all the old
ones, which reference ClassType's that are not reachable from the
`torch.nn.Module` that we imported.
Note: there is no way to avoid importing the whole CompilationUnit
(including these old remnants) without doing a fairly complicated call
graph reachability analysis of which functions are reachable from the
methods of the ClassType's we imported. It turns out that once we are
inside MLIR, we model visibility correctly so that `symbol-dce`
"Just Works" for this use case. That is to say, this is not a quick
hack, but rather seems like a totally palatable long-term solution.
This pass verifies that a given module satisfies the contract that we
have for backends. This is phrased as an "allowlist", because we want to
keep this interface tight. Also, this gives much better diagnostics than
a backend randomly crashing or failing to compile would (though they
could still be improved).
This was especially painful because if we had
`tensor<?x!numpy.any_dtype>` slip through, at some point RefBackend
would convert it to a memref type and trip the "verify type invariants"
assertion which gives no location or anything and crashed the process,
which was very unpleasant.
We implement this with the dialect conversion framework, which works
reasonably well and was quick to put together and familiar, but is still
very "op oriented". We probably want to make this hand-rolled
eventually, especially the error reporting (the most useful kind of
error for a dialect conversion user is not necessarily the best for this
use case). Also, in production, these error will go to users, and need
to be surfaced carefully such as "the compiler needs a type annotation
on this function parameter" which in general requires some special
analysis, wordsmithing, and overall awareness of the e2e use case (such
as how much we can lean into certain source locations) to provide a
meaningful user-level diagnostic.
Also, add `inline` to the current frontend lowering pass pipeline to
allow slightly more complicated programs that otherwise would fail on
shape inference.
This is our first op with error semantics, and stresses the system.
There are a few design notes of special interest:
- RefineTypes.cpp's note about shape inference in the presence of code
that dynamically produces and error, and it is provable statically.
- ATenToLinalg.cpp's notes about future automation of the ATen->linalg
path.
- The notes in Passes.td about using low-tech `std.assert` ops instead
of `shape.assuming`.
Note: Doesn't work on IREE yet due to the `std.assert` op (needs to be
lowered to `vm.fail` on the IREE side).
Recommended review order:
- Changes in frontends/pytorch/examples/
- Changes in python/npcomp/compiler/pytorch/backend/
- Boilerplate for the `npcomp-iree-backend-lower-linkage` pass.
This change separates out a
`npcomp.compiler.pytorch.backend.frontend_lowering` module that does the
common lowering for all backends. The individual compiler backends
`npcomp.compiler.pytorch.backend.{refjit,iree}` now accept a loosely
defined "TCP + scalar code" IR mix that will be formalized in the
future as the interface to codegen backends.
This also required adding a small pass
`npcomp-iree-backend-lower-linkage` which adds `iree.module.export` onto
functions, and layering that into the frontend flow. The pass doesn't
require a C++-level dependency on IREE, which is nice for now. TBD how
we are going to handle lists (we hope we can get away with sneakerneting
some td files and relying on loose IR compatibility).
Running through IREE requires the ability to import `iree.compiler` and
`iree.runtime`, which can be obtained as follows:
```
python3 -m pip install iree-compiler-snapshot iree-runtime-snapshot -f https://github.com/google/iree/releases/tag/snapshot-20210406.200
PYTHONPATH="${PYTHONPATH}:${MY_IREE_BUILD}/bindings/python/"
```
This patch makes it painfully clear that we don't have any e2e testing
harness to really plug into, and also don't have a usable Python API to
our compiler stack (something usable in a jupyter notebook).
That will be addressed in subsequent commits. We've been flying by the
seat of our pants with this `examples` directory that isn't subject to
any kind of testing or real usability concerns.
This pass allows shape information to be propagated to return types,
which is nontrivial and cannot be cleanly put anywhere else as it
changes the public ABI, which is a concern that we want to keep
concentrated in one place.
Currently implemented as a simple intraprocedural dataflow analysis over
a standard ShapedType lattice (hasRank, sizes, and elementType).
It currently hardcodes a few key pieces of information:
- shape transfer functions
- whether it is legal to update the operand type of an op
This needs to be made pluggable obviously and the core propagation logic
moved somewhere agnostic.
The current implementation is just sufficient to do a unary aten.tanh
from the e2e spike, and just applies some local rewrite patterns. I've
sketched out the more full explanation of where this pass eventually
need to go in the pass docs.
Adding this required adding `numpy.tensor_static_info_cast`, which is
the tensor analog of `numpy.static_info_cast`. This op encapsulates the
same numpy-specific "no runtime code" casting semantics, in particular
the interpretation of `!numpy.any_dtype`. The
`numpy.tensor_static_info_cast` I see in practice now are "information
erasing" and will be removed by a later pass that exploits the fact that
aten ops are agnostic to the static info in the operand types (so
substituting a type with more static info is fine).
Side note: we *need* to do dtype and rank inference before aten->tcf
(which will eventually mostly be aten->linalg+guards), because each aten
op is idiosyncratically overloaded based on dtype and rank. Without
copying that idiosyncratic overloading into lower layers (layering
violation), we cannot really lower it to anything until we do that.
This pass incorporates torch.type_bound info and also removes NoneType
returns (eventually it will rewrite tuple types too, but can't yet
because !basicpy.TupleType doesn't track element types).
Recommend looking at adjust-calling-conventions.mlir first to see what
it is doing, and holding your nose for the implementation of the pass.
I decided to implement this with the conversion framework, because it
gives us *some* goodies for type conversion -- mainly avoiding large
amounts of tricky RAUW dances. Unfortunately, the conversion framework
isn't a perfect fit for a couple reasons:
- the incorporation of torch.type_bound is a context-sensitive rewrite
(requires looking at the arg attr, not just the type).
- NoneType conversion is 1->0, which requires some special handling
- (not implemented yet) 1->N tuple type conversions require special
handling.
It's a little bit scary, but on balance doing it the other way would
have its own downsides.
Yi Zhang
Sean Silva
River Riddle
Aaron J Arthurs