This converts a basic list op (torch.prim.ListConstruct) to the IREE
dialect.
```
def forward(self, x: float):
return [x, x]
```
turns into:
```
builtin.func @forward(%arg0: !torch.float) -> !torch.list<!torch.float> {
%0 = torch.prim.ListConstruct %arg0, %arg0 : (!torch.float, !torch.float) -> !torch.list<!torch.float>
return %0 : !torch.list<!torch.float>
}
```
which turns into:
```
builtin.func @forward(%arg0: f64) -> !iree.list<f64> {
%c1 = constant 1 : index
%c0 = constant 0 : index
%c2 = constant 2 : index
%0 = iree.list.create %c2 : !iree.list<f64>
iree.list.set %0[%c0], %arg0 : !iree.list<f64>, f64
iree.list.set %0[%c1], %arg0 : !iree.list<f64>, f64
return %0 : !iree.list<f64>
}
```
As part of doing this, I realized that it was time to formalize the IR
form that we reach right before running TorchTo{Linalg,Std,...}. We now
call it the "Torch backend contract". We then lower the "Torch backend
contract" to the "npcomp backend contract", which involves the new
TorchConversion (`torch_c`) dialect, which holds ops that need to
operate on both the npcomp backend types (e.g. builtin tensors, i1, IREE
list, etc.) and the `!torch` types.
This made more sense, as I realized that if I didn't factor out
`torch_c` then the Torch dialect would have a dependency on IREE
dialect (we previously didn't notice this was an issue because we only
depended on `builtin` types), which seemed wrong to me.
Recommended review order:
- TorchToIREE.cpp / `TorchToIREE/basic.mlir`
- Look at the new structure of createTorchScriptToNpcompBackendPipeline.
It now lives in TorchConversion/Transforms/Passes.cpp and cleanly
calls into `Torch::createTorchScriptToTorchBackendPipeline` for the
frontend lowering to the Torch backend contract.
- Mechanical change extracting
`torch_c.{to,from}_{i1,i64,f64,builtin_tensor,iree_list}` into a new
TorchConversion dialect, and a few passes specific to the lowering
from the Torch backend contract to the npcomp backend contract.
- Minor fixes to TorchToLinalg.cpp to use unconverted operands (now that
we convert lists as part of operand materialization, we need to use
the original operands). Also added test for AtenMaxPool2dOp and fixed
m_TorchConstantIntList.
- TmpDeleteDeadIREELists pass. Temporary pass for deleting dead IREE lists that
are created as part of operand materialization for conv/max pool/avg pool ops
in TorchToLinalg.
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.
- torch.aten.flatten.using_ints to linalg lowering
- torch.aten.max_pool2d to linalg lowering
- Support torch.aten.conv2d for more flexible dilation and strides values
These were legacy concepts that are now superceded by direct Torch to
linalg-on-tensors lowering. These were based on some very early thinking
related to the layering of frontends vs codegen, which is now obsolete
because:
- We expected a lot more centralization at the frontend (TCF) level. It
turns out that frontend needs really vary a lot, and there is no grand
unifying TCF dialect plausible. The additional layer isn't worth it.
- Linalg-on-tensors obsoletes the primary need for TCP. There are still
a few things not representable with linalg-on-tensors, but the support
is growing and the whole "not included in linalg-on-tensors" direction
needs to be rethought. Our TCP dialect didn't cover any of the
actually important things in this space (such as sort, FFT, top-k,
etc.).
See historical [slides](https://drive.google.com/file/d/1iljcpTQ5NPaMfGpoPDFml1XkYxjK_6A4/view) / [recording](https://drive.google.com/file/d/1jSPa8TwPKUt0WuLquGc8OgSUVYJHMvWZ/view)
for more details on the origin story here.
Their presence was confusing users too
[bug](https://github.com/llvm/mlir-npcomp/issues/248).
Also,
- Trim down npcomp-run-mlir testing. It was testing TCF to TCP
lowering for the most part. The essential stuff is retained and
rephrased with linalg-on-tensors. (we should probably rename it
"refback-run" or something, as it is just a way to invoke RefBackend)
- test/Python/Backend/RefJIT/simple_invoke_numpy.py is XFAIL'ed. Our
"anti-framework" direction seems to be the likely future path.
* Adds a minimal setup.py for frontends/pytorch
* Makes npcomp-core export its headers and libraries
* Adds a script to build packages.
* Adds CI step to package and smoke test.
* Will need some more tweaks and coordination prior to deploying (version locking etc).
* Added additional *ToLLVM conversion patterns (they were disaggregated from standard).
* Misc renames.
* Spelling change on ConvNCHW op, and it now expects strides and dilations attributes.
- Build adjustments for `.cpp.inc` dialect files.
- Renaming of `memref.dim` to `tensor.dim` for tensor case.
Minor changes:
- Renaming of `mlir::linalg::ReassociationIndices` to
`mlir::ReassociationIndices`.
- Adjust command line option parsing in npcomp-run-mlir.
We plumb through e2e a fair number of interesting cases:
- unary, binary, ternary elementwise ops
- ops like `torch.aten.add.Tensor` that also take a scalar parameter
- static size-1 broadcasting
We allow the static size-1 broadcasting case, but emit a runtime error
in the case of dynamic size-1 broadcasting. This seems like a sweet spot
subset of things that can be lowered directly to linalg, while not being
overly constraining to users. This is consistent with what IREE is doing
for CHLO->Linalg lowering as well
([code](50bf7a87e4/iree/compiler/InputConversion/MHLO/BroadcastingToLinalgPatterns.cpp (L1))).
To test the static size-1 case, we added support for the
`torch.aten.unsqueeze` op and lowering for it through
`linalg.tensor_expand_shape`. This involved a generalization of
`MaximizeValueSemantics` able to handle it (the solution there also
works for `torch.aten.flatten.using_ints` which we need for ResNet
anyway)
Also, a few minor additional changes:
- Add `VerifyInvariantsBeforeBackendLowering` pass, which catches a
large class of errors before we get to backend lowering (now that we
are doing dialect conversion, the errors are way nicer if we just emit
them up front rather than in the guts of a random pattern).
- Minor change to RefBackend to allow `linalg.tensor_expand_shape`.
Recommended review order:
- e2e tests in elementwise.py
- `ConvertElementwiseOp` in TorchToLinalg.cpp + elementwise.mlir test
- `ConvertAtenUnsqueezeOp` in TorchToLinalg.cpp + unsqueeze.mlir test
- RefineTypes.cpp + tests
- MaximizeValueSemantics changes + test
- VerifyInvariantsBeforeBackendLowering pass + test
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.
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 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).
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.
These allow users to annotate a known "type bound" on the argument,
which can seed shape/dtype inference. We don't rewrite the function
types as part of the import process (it will happen in a
yet-to-be-written pass) because:
1. We would need to interprocedurally rewrite all calls to keep the IR
consistent. Currently, we have a place after GlobalizeObjectGraph but
before we convert to tensors where this is convenient to do. Ideally,
we would do this on the object graph representation.
1. We don't necessarily know that adjusting the function type is a legal
calling convention change. The pass will have blessed knowledge (by
the pass pipeline author) that adjusting the argument type based on
the type bound is safe (which it frequently is).
2. Note that in principle, a type bound could be a fairly general thing
(such as maximum sizes of dimensions, unions of multiple concrete
types, etc.). The pass will in principle have logic to interpret the
type bounds and to determine a suitable "best" (and legal) argument
type.
- renames of OwningRewritePatternList -> RewritePatternSet
- also `insert` to `add`
- RewritePatternSet holds a context now
- memref dialect split from std
* Adds f32 scalar argument support across the ABI boundary.
* Adds support for passing input type / shape information
across the ABI boundary
* Adds support for parsing / creating input FloatAttr's in
`npcomp-run-mlir`
We already had the `promoteTrailingOutTensor` flag, but weren't using
it. A inplaceVariantKernelName flag needed to be added.
This change is a little dissatisfying, as the conversions done by the
RecognizeKernelsPass are currently non-orthogonal. In particular,
`kDropResultAndAliasArg0` probably won't work as intended if mixed with
these (we probably need to promote kDropResultAndAliasArg0 to not be an
arg-level thing anyway, as we have done with promoteTrailingOutTensor).
This involved adding a new op `numpy.overwrite_array`.
```
numpy.overwrite_array %arg2 overwrites %arg0 : tensor<2x3xf32>, !numpy.ndarray<[2,3]:f32>
```
This models the destructive update behavior. Note that in the above op,
we cannot simply RAUW %arg0 with a suitably conveted %arg2 (for example,
%arg0 might have uses that are not dominated by %arg2, or might have an
alias relation with some other array in the program). In general, we
need a pass analogous to "SSA-formation" which knows how to see through
these to uncover an underlying tensor program.
Also, add tanh_out_e2e.py/div_inplace_e2e.py and fix some bitrot in
refjit.py which is my running example I'm trying to get working.
* Import ATen conv2d conversion and test
This is a first attempt at expanding ATen-to-TCF conversion for the
conv2d operator. Eventually, this will come in use when lowering a
high-level conv-based model.
This happens in practice with e.g. ResNet from torchvision (multiple
instances of the same BatchNorm class).
The key observation is that for this program, and the expected set of
programs, we can convert the program to the same globalized form with a
bit more static analysis and effort to suitably monomorphize the
program. Though what we are doing here is fairly annoying to implement,
it saves any nontrivial later pass from having to do similar analyses
(or worse). E.g. shape inference would need to be object-graph aware,
mutation/lifetime analyses would have to be aware, etc. Additionally, it
would make us front-load what it means to have a !torch.nn.Module type
on an ABI boundary, which we are just not ready to handle.
I'm really, really hoping that in practice we can get away with
this, otherwise it's going to be really rough designing a representation
(and implementing everything to back it) that is convenient to transform
and gracefully scales from full object graph (in the most dynamic case)
down to a fixed set of global slots like we have here (in the most
static case, which we presume a lot of practical programs fall into).
This also involved introducing a
`torch-prepare-for-globalize-object-graph` pass that does a minimal set of
lowerings to simplify the IR into a more orthogonal and analyzable form,
and a `torch-globalize-pipeline` helper.
Recommended review order:
- updated documentation in Passes.td
- new tests in `globalize-object-graph-multiple-instances*.mlir`
- implementation of GlobalizeObjectGraph.cpp
- PrepareForGlobalizeObjectGraph.cpp + prepare-for-globalize-object-graph.mlir
- misc stuff like torch-globalize-pipeline pipeline definition.
With this, we can import, globalize, and inline resnet18 from
torchvision:
https://gist.github.com/silvasean/821586afc19b67d9fb72030b2e0adeb8
This primarily unlocks proper handling of free functions (that is,
functions that are not methods of any torch.nn.Module).
Recommended review order:
- `ivalue_importer.cpp` + `ivalue_import/functions*.py`
- `GlobalizeObjectGraph.cpp` + test case
- misc other stuff
The `torch::jit::CompilationUnit` is basically a backing store or
"context" holding all the possible functions in the program. The
previous code was not explicitly accessing this data structure, since it
just imported the `torch::jit::Function`'s that it saw attached to
methods.
Subtly, any time a TorchScript module called into a free function, the
free function gets incorporated into the torch::jit::CompilationUnit,
but doesn't show up anywhere when dumping the module, except in the
curious pattern:
```
%5 : Function = prim::Constant[name="adaptive_avg_pool2d"]()
%6 : Tensor = prim::CallFunction(%5, %input.1, %4)
```
That is, calls are indirect calls, and are accessed via `prim::Constant`
materializing a function object. Even stranger, the `name` attribute here
doesn't really even tell the full story -- it doesn't correspond to
anything. It turns out that the c10::FunctionType itself actually holds
a pointer to the `torch::jit::Function` in the compilation unit
directly (so there is actually no indirection in prim::CallMethod,
because any two values of the same FunctionType call the same
function!). E.g. when converting the IR to bytecode, the "name" is
ignored [code link](1d6bd15790/torch/csrc/jit/runtime/interpreter.cpp (L937)).
We do import `prim::CallFunction` as a `std.call_indirect` though
because it's more braindead to do it that way (it gets canonicalized to
a direct call easily).
This is a much simpler representation than the ad-hoc initializer
function we had before. It is also less general, but given the rationale
in Passes.td it seems like the right tradeoff right now.
We can probably carry this representation for quite a while, and when we
can't, it likely means that TorchScript has fixed their object identity
bug and we probably need to just upgrade to a more general object graph
modeling (more general than GlobalizeObjectGraph).
In particular, we don't want to deal with defining and carrying around
this initializer function concept until we need it. For example, if we
want to constant-fold the global slots into uses, this is a much better
representation, and it plays better with symbol-dce (the initializer
function counts as a "use" of the symbol).
(the alternative would have been to write a pass that converts the
initializer function to this form when possible, but I realized that
lots of information had been lost which made that fairly annoying -- it
was all self-inflicted anyway, so best to just go to the source
(GlobalizeObjectGraph) before the information is lost)
Now symbol-dce works nicely (no more "training" bools)
```
pt_util ~/tmp/classifier.pt --import --exported-name forward \
| npcomp-opt -torch-globalize-object-graph -inline -symbol-dce
```
IR: https://gist.github.com/silvasean/8abe63d70d24e29d6db9170ccc8d512b
The first use case is to annotate certain program constructs as either
exported or private. In this commit we plumb it down to
GlobalizeObjectGraph which makes use of this information.
Recommended review order:
1. class_annotator.h/.cpp + `test/module_import/annotations/*`
- New abstractions to communicate with Python code and annotate.
2. IR changes in TorchOps.td
- Adding "private" attribute to various things.
3. ivalue_import.cpp changes
- Module + ClassAnnotator = annotated IR
4. GlobalizeObjectGraph.cpp + tests
- use new "private" attributes to create "private" IR.
- also, tweak some of the op deleting mechanics, which was triggering
some memory errors / assertions
With this, we can run the classifier through and inline it as follows:
```
frontends/pytorch/utils/pt_util.py --import --exported-name forward ~/tmp/classifier.pt \
| npcomp-opt -torch-globalize-object-graph -inline
```
IR: https://gist.github.com/silvasean/32dcad9f6270557f412094a77cecdd69
This required restructuring of how we model TorchScript on import. The
main difference is that now we split out a `torch.class_type` that holds
methods and declarations of the types of each slot. This is more
consistent with TorchScript (our previous representation was
"denormalized").
Recommended reading order:
1. check out the description of `torch.class_type` in `TorchOps.td` and
look at `test/Dialect/Torch/ops.mlir` and
`frontends/pytorch/test/module_import/` to familiarize with the new
representation.
- Just look at the new IR. The diff between the old names and new
names is confusing.
2. check out `test/Dialect/Torch/globalize-object-graph*.mlir`
and read along with the pass description in
`include/npcomp/Dialect/Torch/Transforms/Passes.td`
3. Read the code in `GlobalizeObjectGraph.cpp` and miscellaneous changes
in `ivalue_importer.cpp`, `TorchOps.cpp`, etc.
It turns out that this was easiest to structure as a general IValue
importer, since torch module are just one of the possible IValue's.
We import the IValue object graph in a braindead fashion into basicpy
ops and a new `torch.nn_module` op that is used to model the
attributes/methods of a torch::jit::Module IValue. See `Torch/ops.mlir`
for an example, and also check out the .py import tests in
`frontends/pytorch/test/module_import`.
As part of this change, a few housekeeping tasks:
- extract some helpers from graph_importer.cpp
- more helpers around the C API
- misc touchups
This allows building NPCOMP as an external project of LLVM, similar to
how CIRCT can be built: https://github.com/llvm/circt/pull/227.
The CMake options to use this build style look like this:
```
-DLLVM_EXTERNAL_PROJECTS=npcomp \
-DLLVM_EXTERNAL_NPCOMP_SOURCE_DIR=/path/to/mlir-npcomp \
```
- TensorFromElementsOp -> tensor::FromElementsOp
- `cmpi "eq", ...` -> `cmpi eq, ...`. Same for `cmpf`
- syntax change for private func ops
- some changes to the python bindings
* Most updates are mechanical except:
* python/npcomp/__init__.py and python/NpcompModule.cpp: New init/registration bits to replace some automatic things being done in the old bindings. Also an annoying linkage hack that I'll need to triage next.
* NpcompModule.cpp: New python helpers for custom types and other hard to reach items (for the new bindings).
* PybindUtils.h: Extended type casting so that the local extension can directly exchange Mlir* C types.
* python/npcomp/dialects/*: Build support and ODS bindings for local dialects.
* mlir_utils.py: Defines an ImportContext to replace the old/bad "Helper" class that tracked locations, and insertion points. This has a number of methods on it that would be good candidates to think about better ways to do them upstream.
* Also hoisted a few stand-alone samples to dedicated unit tests as they covered important things.
* More cleanup can be done, but keeping this patch as mechanical as possible to stay in NFC land (this is big enough).
Changes:
- linalg init tensor change (outs+init -> just outs)
- IntegerType::get and other builtin types now take the context as the
first arg
- LLVMType::* is gone. Now LLVM Types are just regular Type's.
Best as I can tell (e.g. from LeakSanitizer), this fixes all the leaks
except for those due to buffers created internally to the codegenned
code itself (up next I'll add the buffer deallocation pass to fix
those).
The main change is that instead of attempting to pass `refbackrt::Tensor`
to the codegenned function directly, we make all the ABI types be
UnrankedMemRef which gets passed awkwardly (but workably) as a
`{size_t rank, void *ptrToDescriptor}` on the ABI. The reason why
refbackrt::Tensor wasn't workable is that is that MLIR doesn't really
have a way to deal with the lifetime of unranked memref descriptors that
happen inside the function, which is inevitably what would happen in the
old code that would emit runtime calls to
`refbackrt.to_memref/refbackrt.from_memref` to convert back and forth to
`refbackrt::Tensor` inside the codegenned code.
So, instead of the `refbackrt.to_memref/refbackrt.from_memref` with no
real sound basis for valid lifetime management, we now have a lovely
piece of code in `refbackrt::invoke` in `Runtime.cpp` that just barely
seems to be sound. We rely on the codegenned code having these
properties, which it seems to have:
- it won't free memref descriptors or their backing buffer for arguments
of UnrankedMemRef type.
- it will allocate a separate memref descriptor for each result
UnrankedMemRef (which is ensured by having a separate memref_cast for
each)
- we can sniff the `allocatedPtr`'s (i.e. the backing buffer pointers)
to avoid double-freeing in the case of aliasing of the backing buffer
(including backing buffers for arguments feeding into results)
- to catch the case of statically allocated data (which we need to avoid
passing to `free`) , check if the `allocatedPtr` is (no joke) equal to
`0xDEADBEEF`, because there is otherwise no way to distinguish
statically allocated from malloc'ed data... (std.global_memref lowering
to LLVM by happenstance sets the allocatedPtr equal to `0xDEADBEEF`,
presumably mainly as a debugging thing)
Even with all this, we *still* need to (internally to refbackrt::invoke)
make copies of all inputs/outputs! And the details of how the LLVM-level
ABI gets laid out for e.g. function arguments/returns is still super
tricky.
This really highlights how deficient memref is as the general runtime
type for our use case. It's stewing in my mind how best to improve the
situation. My general gut feeling is that IREE's abstractions for this
are "right", but I need to think more how to distill those aspects of
IREE's design in a "reference" way for RefBackend.
Some implementation notes:
- In terms of how this is implemented, this did catch a bug in our ABI
wrapper functions in LowerToLLVM.cpp, which I had to fix (it happened to
work before through some combination of npcomprt::Tensor being passed as
a single pointer + probably me infinite-monkey-ing it until it worked)
- This actually removes 2 out of the 3 compiler runtime functions (the
only one left is "abort_if". (most of the memref descriptor code moved
from CopmilerRuntime.cpp to Runtime.cpp)
- this also means deleting `refbackrt.from_memref` and
`refbackrt.to_memref`
* Going through TODOs on the PyTorch side, this is a big cause of them (not being able to have constants for signed/unsigned).
* Added complex while in here since we're at the phase where it is better to just have things complete than partially done.
* Organizes the BasicPyOps.td file by function.
* Renamed `to_boolean` -> `as_predicate_value` (trying to consistently use "predicate" to refer to i1/low-level types and Bool/Boolean to refer to Python bool types).
It was annoying that we were creating shape.get_extent in the middle of
the bufferization pipeline, as it required running convert-shape-to-std
at an awkward place. To make that cleaner, just open-code the
extract_element ops that shape.get_extent expands into.
This is a little gross, but it helps with the macroscopic pipeline
ordering issues. Anyway, the train is long-gone of trying to treat
shapes as some special data type that should only be operated on with
shape ops.
Also,
- reorder tensor constant bufferize (which is a module pass) to bracket
all the bufferization function passes, to make the parallelism
opportunities there clearer. Now we have a very clean little
bufferization segment of our pipeline construction.
* IREE doesn't have proper install support, so there is some temporary hoaky hacking in our CMakeLists.txt to shuttle some symlinks around.
* Reworked the original numpy e2e with IREE test to pipe through iree-translate.
* Removed all of the C++-level dependencies.
* Will generalize and apply to the PyTorch backend in a followup.
This vastly simplifies our code, allowing deleting multiple ops,
simplifying multiple passes, and removing a whole pass.
Now `refback` dialect is down to one op (refback.alloc_memref, which
simplifies allocations to just take a shape instead of individual
extents).
This involved adding a `tcp.splatted` op to splat a dynamically sized
init tensor. See rationale in TCPOps.td docs.
One interesting observation is that when lowering tcf.matmul to
linalg.matmul, we need to both 1) create the error checks and 2)
calculate a shape transfer function to create the init tensors.
Previously, 2) was deferred to bufferizing tcp.matmul later. I'm not
sure if this is a conflation of concerns or not. For now, it's not a big
burden.
* Conversions are very simple, suporting mul, maximum and add (alpha=1 only).
* Example added with pass pipeline needed to run.
* Much missing off of the golden path but sufficient for such simple cases.
* convolution, convolution_backward, _log_softmax, _log_softmax_backward_data, nll_loss_forward, nll_loss_backward, nll_loss2d_forward, nll_loss2d_backward, copy_
* Extends the recognition logic and metadata for handling inplace transformations, optional tensors, ints, lists and dropped args.
* The kernel_calls generated by test_conv_nllloss_grads.py now convert to ATen.
* The result *almost* comes out as a pure tensor program with the exception of the copy_ op, which I will do some followup work to deal with.
* More progress on #97
Now, the only bufferization we have left is lowering tensor constants to
memref, which will hopefully proceed soon after Rahul's new
std.global_memref lands + the lowering to LLVM IR. Then I'll port
LowerConstantTensorsToMemref to upstream and we'll be 100% upstream
bufferization, except for our local TCP dialect (which will probably go
away and be replaced by std elementwise + linalg named ops on tensors :)
).
* Deletes prior code generator from previous attempt (moved some of it into this one).
* Renames old generated tablegen source to "Legacy".
* Generates ODS and import rules for most binary and unary arithmetic ops.
* Removes old generated ops and integration tests that were testing details of the prior setup.
Register the following for the multiply op:
- tcf.mul
- tcp.mul
- TCP->TCP lowering
- Shape transfer, broadcasted multiplicands
- Lower to standard `MulFOp` op
* Two op interfaces, one for querying instance metadata and one for getting static data needed to construct an op from a generic form.
* For torch.generic_kernel ops, metadata is splatted in during capture from Torch (it comes from the op registry, which will work for either device capture or graph import).
* Moved the 'add' out of the generated set so I can experiment on it. It implements the TorchBuildableKernelOpInterface interface which provides its metadata.
* The ATenRecognizeKernelsPass pass generically lowers from a torch.generic_kernel to recognized ops that implement the TorchBuildableKernelOpInterface, handling the various types of transformations that we allow at this stage.
* Adds Basicpy List, Tuple, Dict types and plumbs through C API.
* Started debugging the issues around aten::conv2d capture, but a PyTorch bug is suspected.
* Was able to manually verify that the basic conv2d forward test captures correctly with a workaround.
* Need to resolve some printing issues upstream and move these tests to an integration test target (they take ~seconds to run).
The time has come for BypassShapes/LowerShapedResultsToMemref to go away :(
For the reference backend, being consistent with upstream conventions is
the name of the game now.
This is a step down in a number of ways, e.g. test clarity and
separation of concerns. But it is fewer files and fewer tests, and
*does* address the "TODO: This is really fragile". It also eliminates two
more ops from the refback dialect (sadly, they are the
shaped_results/yield that we were getting kind of fond of, but alas).
Now that it has grown source/target materialization capabilities
(spelled with ops tensor_load/tensor_to_memref), we can use it. We can
also now delete refback.memref_to_tensor/refback.tensor_to_memref.
This is also a first step to reducing the downstream functionality
needed in the refback dialect.
* Also adds two lit tests to verify that all of our extensions load without fireworks, which is a good indication that the shared library deps are sane.
* Bumps llvm-project version to use D89167.
Now the reference backend is cleanly accepts "TCP"+scalar ops.
We introduce tcf-refback-lowering-pipeline which also does TCF->TCP
conversion for convenience until we have a "target interface".
* Need to have a dag of shared library deps in order to interop across python extensions (as presented in ODM).
* Introduced add_npcomp_library and friends to mirror the MLIR setup.
* Adds a libNPCOMP.so shared library.
* Redirects tools and extensions to link against libNPCOMP.so (instead of static libs).
* Moves all libraries to lib/, all binaries to bin/ and all python extensions to python/. The invariant is that the rpaths are setup to have a one level directory structure.
* Reworks the _torch_mlir extension to build like the others (still need to come up with a consolidated rule to do this instead of open coded).
* Includes an upstream version bump to pick up needed changes.
Sizes with dynamic linking (stripped, release, asserts enabled):
libNPCOMP.so: 43M (includes much of the underlying LLVM codegen deps)
libMLIR.so: 31M
_npcomp.so: 1.6M (python extension)
_torch_mlir.so: 670K (python extension)
npcomp-capi-ir-test: 6.3K
npcomp-opt: 351K
npcomp-run-mlir: 461K
mnist-playground: 530K
Still more can be done to normalize and optimize but this gets us structurally to the starting point.
I now realize that VerboseCamelCase is not the best choice for dialect
directory/file names and C++ identifiers (take e.g. "Linalg", "Basicpy",
etc. as prior art here; not LinearAlgebra or BasicPython). If I had to
name the convention it seems to be "Shortword" (or of course just
acronym dialects like LLVM, SCF, etc.).
This rename also has the side benefit of differentiating RefBackend
directories, which now refer to the actual backend itself, from
Refback/Refbackrt, which are the dialects which happen to be used by
that backend.
This is the first in a patch series that is refactoring the
constellation of things variously called or associated with "E2E",
"RefE2E", "npcomprt", and "TCP" into a more cleanly layered result.
Concretely, this first patch fixes the fact that TCP was basically
acting like a dumping ground needed by the reference backend. This
splits it out, which is fairly mechanical, but touches a lot of lines of
code (basically replacing `tcp` with `refback` and `TCP` with
`RefBackend).
Now, the RefBackend dialect is that dumping ground, which
is slighly better, as it starts allowing TCP to become a nice clean
middle layer that is not related per se to the reference backend.
The previous name RefE2E or "reference e2e flow" was super confusing.
Now that we are seeing more clearly where the "backend" distinction
lies, the [RefBackend] commit tag is born :)
I was seeing some miscompiles due to the uninitialized data read here
before. Interestingly, this was masked in some of our previous test
cases, since the uninitialized data "always" was so small that it would
present as a rounding error for the 1.0-10.0 sized values that the
matmul was computing on.
* Adds at::Tensor -> MlirValue tracking.
* Adds conversions for tensor and scalar types to MLIR types.
* Adds npcomp C APIs for constructing custom types.
* Reworks pybind include so as to get Torch pybind helpers (needed to pass at::Tensor type from Python->C++).
Now that we upstreamed our pass, we can remove it.
The final pass that landed upstream doesn't do the shape.assuming
canonicalization to legalize that op away, so added a
restricted-canonicalizer pass that allowed to run just shape dialect
canonicalizations, which deletes the shape.assuming.
The pass ended up kind of ugly. See the TODO's on it for some potential
cleaner directions.
Date: Fri Sep 18 13:55:52 2020 -0700
- Update to linalg syntax
- New generated builders are better. Custom builder for
tcp.shaped_results is now redundant.
This cleans up the lowering pipeline to easily allow extending to
multiple binary ops. It looks fairly repetitive at multiple levels, but
I don't want to prematurely generalize. I think that in principle we
could derive a large swatch of TCF + TCP from a single linalg-style
specification. Another direction is to use an OpInterface (something
like "buildLinalgGenericBody"). I'm keeping my eye on it.
In a subsequent commit, I'll mechanically add a set of binary ops
modeled off of the std arithmetic ops.
It was previously going through this awkward route that prematurely
created linalg.generic ops, which was an annoying layering problem since
we can't compute a shape transfer function for linalg.generic in the
general case. Now we pass it through the same path as tcp.matmul, with
the shape transfer function being defined for tcp.add.
This also removed the need for TCPToLinalg (now deleted). The equivalent
of that is happening in lower-shaped-results-to-memref. One interesting
outcome of this: we're basically using linalg as a "Buffer TCP". We
might want to look into using named structured ops for more of TCP, but
that would be a big velocity hit since then any change to the ODS /
verification for those ops would be a change to the upstream structured
op ODS generator. After we have more experience defining this manually,
we should re-evaluate rebasing TCP on generated named linalg ops.
I'm pretty happy with how this turned out. It looks pretty much like it
should -- one change at each layer. This particular op bottoms out on
linalg which takes care of the rest.
- Add tcf.matmul
- Add tcp.matmul
- Add TCF->TCP lowering
- Add tcp.matmul shape transfer function (BypassShapes.cpp)
- Add tcp.matmul -> linalg.matmul lowering (LowerShapedResultsToMemref.cpp)
- Add support to LowerShapeConstraints for lowering the new
shape.cstr_require
This matmul op is pretty limited in its capabilities. There is no
batching and no multidimensional contraction. Certainly more design work
will be needed to find the right abstractions that aren't too general
but also help to canonicalize many cases from frontends. This is mainly
to show that adding a new op needn't be very "scary" once we have the
e2e infra in place.
Also,
- this clears out some exploratory cruft from the TCF dialect now that
this is starting to become real.
This now gets the overall "RefE2E" compilation stack to a point that I'm
fairly happy with. We simplify it by mostly embracing the "descriptor"
view of the world.
The overall flow is best understood by reading through the
createE2ELoweringPipeline function in lib/E2E/E2E.cpp
That function creates a pass pipeline that lowers from "TCF" (which is
~numpy level of abstraction) down to LLVM IR.
A brief high-level summary of what happens there:
1. TCF to TCP conversion. This involves reifying error handling in the
form of shape constraints. See test/Conversion/TCFToTCP/basic.mlir
2. Lowering shape constraints. This converts shape constraints into
eager error-handling code. See test/E2E/lower-shape-constraints.mlir
This pass will soon go upstream.
Because this lowers to std.assert, some later passes like
LowerToNpcomprtABI and LowerToLLVM are updated to properly plumb this
through e2e.
See test/npcomp-run-mlir/invalid-broadcast.mlir for an execution test
that properly aborts in case of an error.
3. Lowering tensors to memrefs. This is done via a series of passes
rather than an single mega conversion. Unlike the previous code that
mixed in the npcomprt ABI stuff here, it's now a very clean "pure
memref" conversion.
See test/E2E/lower-*-to-memref.mlir and
lib/E2E/TensorToMemref/
Most of the changes are concentrated here.
4. As part of the above, we use the upstream ConvertShapeToStandard for
lowering shapes.
5. We lower linalg to loops and lower loops to CFG using upstream
passes.
6. Rewrite the "ABI" boundaries of the program to npcomprt data
structures (LowerToNpcomprtABI). This mainly affects ABI boundaries and
how global tensor constants are represented. One of the major
improvements in this commit is that now it's a very clean rewrite that
just replaces memrefs on ABI boundaries with !npcomprt.tensor (before
there was a get_extent function that is not needed).
See test/E2E/lower-to-npcomprt-abi.mlir
7. Lower to LLVM with upstream mlir patterns + some patterns for the
npcomprt lowerings.
One aspect here that is still a remnant of a non-descriptor-based tensor
to memref flow is the BypassShapes + LowerShapedResultsToMemref.
BypassShapes wraps the "tensor compute" ops in a tcp.shaped_results
(basically a "tie_shape" kind of op), and then
LowerShapedResultsToMemref uses those annotations to allocate output
buffers while lowering the "tensor compute ops". Note that there are
very few "tensor compute" ops currently supported (tcp.add +
tcp.broadcast_to), so we just hardcode them in both passes.
Realistically, I expect this to go away as we fully embrace the
descriptor-based approach for simplicity, so don't look too deep into
it.
* llvm-project: b5924a8e27536d19dd5c4d302db29fb6163d5faa
* mhlo: 848ca244d20f045b7921da55a98a04d95ef94f0e
* Multiple breakages that need to be fixed.
Fixes:
* Refactor dialect registration
* Remove all kindof methods (Casting functionality has been added upstream and is implicitly
available, see https://llvm.discourse.group/t/removing-kinds-from-attributes-and-types/1547.)
* Update dialect registration to comply with https://reviews.llvm.org/D85495.
* Remove type kinds and update some changed dialect signatures.
* Upgrade ATen dialect to match upstream needs.
* Move dialect registration to tablegen.
* Register the ListType in tablegen.
* Change dialect initialization signature.
* Use TypeSwitch in MlirIr location printer.
* Remove global registry depends from npcomp-opt.
* Change LowerToLLVM to pass an MLIRContext vs an LLVMDialect for type creation.
* Remove dep on MLIREDSCInterface that is removed upstream.
* Thread through the DialectRegistry for opt and python-like tools.
* Modernize pass registration (This was forced because the GEN_PASS_REGISTRATION code now generates inline functions vs literal pass registration statements)
Co-authored-by: Marius Brehler <marius.brehler@iml.fraunhofer.de>
This patch adds a dialect intended to be used as a frontend dialect
to facilitate lowering from "A Tensor Library" in torch/pytorch.
This patch includes several passes that are useful in conjuction with the
dialect:
--aten-layer-name: Generates layer names for each operation, which are not
present in the original pytorch.
--aten-to-std: Lower the ATen dialect into standard dialect function calls.
--return-elimination-pass: convert functions (primarily the toplevel function)
to pass return values by reference. This simplifies pytorch integration.
--aten-op-report: generate a textual report about the model
--liveness-report
Future patches will implement actual integration with the pytorch jit to
intercept and generates MLIR in this dialect, then lower the resulting MLIR
into function calls through aten-layer-name -> aten-to-std ->
return-elimination -> std-to-llvm. The result would then jitted using the LLVM
jit, linked against a runtime library which makes calls back into pytorch to
implement all the layers.
Co-authored-by: Jeff Fifield <jeff.fifield@xilinx.com>
Co-authored-by: Jeff Fifield <jeff.fifield@xilinx.com>
* These were separated originally for layering reasons that no longer apply.
* Most of the python extension code is under lib/ with just the module setup in python/.
* Uses local configs and unsupported annotation to disable optional tests.
* This separation was just an artifact of having initial trouble getting lit setup.
* Primarily, the upstream shape dialect now uses tensor<?xindex> for non-erroring, immediate shape calculations (and will return this for shape_of of a tensor or memref).
* In addition, upstream passes do not yet exist for fully lowering to standard ops, so the passes here need to be extended to handle this new convention.
* This should be seen as an intermediate state, necessary to integrate a new LLVM version and needs more work and cleanup for generality.
* There is a good deal of awkwardness in these conversions. The hope is that additional upstream work will yield better defined conversion paths once out of this intermediate state.
This required making module descriptors hold a FuncDescriptor* instead
of a pointer to array of FuncDescriptors as it previously did, which is
innocuous (just requires an llvm.bitcast after the llvm.mlir.addressof).
* Rewrites public function signatures to operate on tensors (vs ndarray).
* Most of our backends presume immutable tensors at public function boundaries.
Also, the previous code had a special case for deleting this op when it
had no uses. This is subsumed by the change in this commit since now
shape.const_shape is properly lowered.
With this change, the included test case with multiple serially
dependent ops works!
This specific issue was related to the scalar argument to that
function. We needed to compute a broadcast of a scalar shape (which is a
shape.const_shape) with another shape.
This ~totally reworks the existing "runtime" stuff to be more
principled and usable, such as from Python. It's still not fully
production-quality, mainly in the department of memory management (e.g.
it currently leaks memory; we need to figure out "who frees memrefs" +
the analysis and transformation needed to do that (maybe use upstream
buffer allocation pass?)).
The user API is in include/npcomp/runtime/UserAPI.h, though
include/npcomp/JITRuntime/JITModule.h is a friendlier wrapper.
The stuff under {include,lib}/runtime is totally firewalled from the
compiler and tiny (<6kB, though no attention has gone into optimizing
that size). For example, we don't link in libSupport into the runtime,
instead having our own bare bones replacements for basics like ArrayRef
(the JITRuntime helps with bridging that gap, since it *can* depend on
all common LLVM utilities).
The overall features of npcomprt is that it exposes a module that
with multiple function entry points. Each function has arguments and
results that are tensor-valued, and npcomprt::Tensor is the runtime type
that is used to interact with that (and a npcomprt::Ref<T>
reference-counting wrapper is provided to wrap npcomprt::Tensor in the
common case).
From an implementation perspective, an npcomprt module at the
LLVM/object/binary level exposes a single module descriptor struct that
has pointers to other metadata (currently just a list of function
metadata descriptors). All interactions with the npcomp runtime are
keyed off of that module descriptor, including function lookups and
dispatching. This is done to dodge platform ABI issues and also allow
enough reflection to e.g. verify provided arguments.
Most of the compiler-side work here was in LowerToNpcomprtABI and
LowerToLLVM.
Also,
- Rename npcomp_rt/NpcompRt to npcomprt/Npcomprt; it was getting
annoying to type the underscores/caps.
- misc improvements to bash_helpers.sh
* This elides the very common code the compiler adds for chaining otherwise tensor-related numpy ops together.
* More aggressive canonicalizations would require more advanced analysis.
* This starts to lay down the infra for reasoning about calls
* Adds the importer code to generate IR for function calls of compiler recognized static functions.
* Adds python bindings for invoking flow, HAL, and VM lowering pipelines.
* Adds pythong bindings for translating to VM module flatbuffer.
* Adds a new backend_test/iree directory and configure lit to find the IREE python rt bindings.
* Open code a simple_invoke.py that exercises the whole pipeline (need real APIs for a lot of this).
* Fails when invoking the function because I never implemented argument marshaling for scalars :(
* Plenty of stuff to do tomorrow.
The secret here is LLVM_ENABLE_WARNINGS=ON.
I also fixed a couple warnings, which gets us to be warning-clean.
I noticed also that npcomp-run-mlir/basic.mlir seems to be failing.
Maybe something since the latest integrate. My next commit (introduce
npcomp mini runtime) will largely rewrite it though, so it'll get fixed
then.
The code isn't super clean, but is a useful incremental step
establishing most of the boilerplate for future enhancements.
We can't print or return tensors yet so correctness TBD, but I've
stepped into the running code in the debugger so I know it definitely is
running.
This is the first step to building out an npcomp mini-runtime. The
mini-runtime doesn't have to be fancy or complex, but it should at least
be layered nicely (which this code and the current compiler interaction
with the "runtime" code is not). Now that we have boilerplate for e2e
execution in some form, we can build that out.
With this commit, we finish conversion to LLVM dialect, and should be
ready for subsequent commits to convert to an LLVM module and let LLVM
codegen to native machine code.
This required a custom "lower to LLVM" pass to support lowering
tcp.abort_if to a runtime call. In the future, this pass will grow to do
type conversions for our own runtime types as we add those.
This more clearly captures its semantics as a structural "observer" of
code that we currently mark as NoSideEffect but eventually lowers to
eager error handling code.
Also, update LowerRankedShapes to erase it, now that the layering here
is clear. That pass reifies the eager error handling code, so the need
for the dummy op to keep things alive isn't needed.
With this change, we are now ready to start lowering to LLVM!
This is the current print-ir-after-all from e2e-lowering-pipeline:
https://reviews.llvm.org/P8221
Specifically, we use unranked memrefs which get passed as a fixed-size
set of arguments/returns. One big caveat about this is that returning
results isn't going to work. See TODO in LowerTensorLoadOp.
This is far from enough runtime-wise, but it starts to demarcate a
plausible layering. Notice for example how this removes the
runtime-dependence from LowerRankedShapes.
Eventually, we want to have an `npcomp_rt` or `npcomp_hal` dialect with
its own set of runtime types that will supercede this.
See comments in LowerTensorLoadOp for more direction about where this is
going to evolve.
The idea was half-baked and after some deep thought felt like a solution
looking for a problem. What we had here (and is removed in this patch)
just wasn't pulling its weight.
I cannot think of anything we would want to do with tcp.island as it is
removed here beyond just sinking and merging them within a basic block,
such that the witness argument is kind of pointless (only matters for
hoisting).
TCP compute ops like tcp.add and tcp.broadcast_to have the strong
invariant of "pure or undefined behavior", which means they are always
safe to sink. The island concept as removed here conferred no benefit.
Also, I'll note that "islands" are a trick you can only play once in a
system (unless they strictly nest). I have some early-stage thoughs on
having an island concept that helps with modeling tensor shapes
robustly which seems promising (the island would serve a similar role as
tie_shape).
- Make rank1.mlir be the new "basic.mlir", as it is really the simplest
case.
- Move basic.mlir to mixed-ranks.mlir
- Delete starting-from-linalg.mlir, it wasn't really useful anymore.
This uses an approach inspired by what is done in IREE. See comments on
LowerRankedShapes.cpp for how it works.
The basic gist is that we have an op that creates a !shape.shape from a
set of SSA values representing the extents, and then iteratively replace
any op producing a !shape.shape with instances of that op.
There's a lot of details to flesh out here, but the basic approach seems
promising (see comments in createE2ELoweringPipeline).
This approach will be put to the test when we try to do our first
fusions since that tickles some of the nasty phase ordering issues
involved here.
But we're not there yet.
* This is intended to provide low-level modeling for built-in objects.
* It is now possible to trace slice tuples (which are tuples of NoneType|EllipsisType|SlotObjectType<slice, ...>).
Sean Silva
Yi Zhang
Stella Laurenzo
River Riddle
Bryce Arden
Aaron Arthurs
Aaron J Arthurs
mikeurbach
stephenneuendorffer