Strength the shape inference for aten.arange-like op by
1. registering aten.sub and aten.ceil.Scalar op and design folders for them.
2. register a new constant-like op: Torch::ConstantNumberOp and design canonicalizer for it.
This PR adds an `AllowedInModuleInitializer` trait to keep track of ops that are permitted in the module initializer. We have a handful of such ops that are produced by the IValue importer, and so this change avoids maintaining a list of ops in `TorchOps.cpp` that could lead to spurious merge conflicts, and help us integrate torch-mlir in our downstream compiler better. Please let me know if you'd prefer a better name for the trait itself. Feedback is welcome!
The Torch dialect has an include to `mlir/Dialect/Func/IR/FuncOps.h` and
should therefore have a CMake dependency to the MLIRFuncDialect.
Otherwise, the build can fail since it may occur that
`mlir/Dialect/Func/IR/FuncOps.h.inc` isn't generated yet.
Summary of changes:
- Switch to C++17 (similar to https://reviews.llvm.org/D131348)
- Update MHLO to build with LLVM commit hash 061e0189
- Replace deprecated `hasValue()` and `getValue()` with `has_value()`
and `value()` respectively (https://reviews.llvm.org/D131349)
- Use `TypedAttr` (https://reviews.llvm.org/D130092)
- Use updated assembly format of `mhlo.compare` op (commit
d03ef01e70fbf9afd0fa1976fbb7ed31838929b3 in MHLO repo)
Rather than a per-global-slot initializer region, we now have one for
the whole module. For example, it might look like this:
```
torch.global_slot "private" @tensor : !torch.tensor
torch.global_slot "private" @list : !torch.list<tensor>
torch.global_slot.module_initializer {
%0 = torch.tensor.literal(dense<0.0> : tensor<f32>) : !torch.tensor
%1 = torch.prim.ListConstruct %0 : (!torch.tensor) -> !torch.list<tensor>
torch.initialize.global_slots [
@tensor(%0 : !torch.tensor)
@list(%1 : !torch.list<tensor>)
]
}
```
This new structure allows GlobalizeObjectGraph to create the initializer in a
much simpler way, avoiding the need to reason about whether different slots
alias each other. Reasoning about whether slots alias each other now is the
responsibility of InlineGlobalSlots, which has to do a much more complicated
analysis, implemented using MLIR's dataflow analysis framework.
Recommended review order:
- Check out the new IR constructs in the .mlir files of various passes
- Op definitions (*.td)
- Changes to GlobalizeObjectGraph pass.
- InlineGlobalSlots pass (~total rewrite)
- Misc changes:
- Moving torchMlirAdjustStaticInformation for sharing with C++ code.
- EraseModuleInitializer pass
To make this a bit nicer, it would be good to have a `torch.module` op
with an initializer region attached. That would be more invasive though.
This change has highlighted certain aspects of our project layering
which are worth calling out. None of our backends can handle global
slots, so we enforce that there are no global slots before backend
lowering. At an earlier stage in the project, we had aspirations of
transparently handling mutable global state and such, but for reasons
described below, that is no longer a goal. So really global slots should
be seen as a progressive lowering step as part of inlining all the
IValue's in the original program (GlobalizeObjectGraph is also one such
step).
Over time, with insights from work like IREE-JAX, it has become clear
that there isn't a reliable programming model we can compile for users
where we just transparently handle mutable global state (and some other
things, like lists and dictionaries). There is a need for an "outer
program" that orchestrates more restricted subroutines of the kind we
can handle in our compile flow here. The benefit of that is that it
decouples considerations like shapes, dtypes, etc. from the program
constructs used in the outer program. As long as the outer program can
efficiently invoke (pipelining/async/etc.) high-performance
data-parallel numerical subroutines of the kind we compile in our flow
here, then there is a complete programming model. This is also
consistent with the direction of upstream PyTorch which is becoming more
tracing-based (which inherently loses a lot of program structure, which
then has to be applied back with an "outer program" orchestrating the
traced subroutines).
- Includes a canonicalizer for `aten.add.t`needed for successfully lowering the shape function
- Only offers support for statically sized index tensors when there is more than one
- Dynamic shape support remains for single indexing tensors
The biggest change here is to upgrade RefineTypes to the new sparse
dataflow framework.
Smaller changes:
- minor changes to type parsing
- suppress warnings in e2e tests
Prior to this patch, the canonicalizers for `AtenSizeOp` and
`AtenSizeIntOp` succeeded only if the tensor operand's type information
included the size of the requested dimension(s). We can extend the set
of optimizable cases by propagating types across operations whose result
type matches the input tensor type.
Specifically, this patch enables the canonicalizers for `AtenSizeOp` and
`AtenSizeIntOp` to see past `tensor_static_info_cast`,
`copy.to_vtensor`, and `copy.to_tensor` ops until it reaches the first
op whose result type contains size information for the requested
dimensions, with a maximum bound of 6 parent lookups to avoid indefinite
compilation times. All other encountered ops cause the canonicalizer to
give up.
This patch replaces the use of raw integers like 6, 4, etc. (that
represent PyTorch's scalar types) with named values from the ScalarType
enum (e.g. `ScalarType::Float`, `ScalarType::Long`, etc.) in code for
folding `prim.dtype` ops into numeric constants.
This patch isn't strictly a non-functional change, since its use of
`Torch::getScalarTypeForType()` implies that the input type has to be
one among the supported types, otherwise compilation will abort, whereas
previously, compilation proceeded without folding the unsupported data
type into a numeric constant.
A prior patch (63538de2) that added support for bfloat16 type did not
add the canonicalization pattern to fold `torch.prim.dtype` operations
on bfloat16 tensors into the integer constant 15. This patch fixes the
problem.
This commit adds lowering of `aten.ge.float`, `aten.ge.float_int`,
`aten.ne.float_int`, `aten.gt.float_int` and `aten.ceil.float` op.
This commit also fixes formatting for the file scalar.py and scalar_comparison.py.
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
This commit decomposes `aten.to.dtype_layout` op into `aten.to.dtype` op.
This commit also fixes the formatting for the file type_conversion.py.
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
This commit adds lowering of `aten.ceil.float` op.
This commit also fixes formatting for the file scalar.py.
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
The updated LLVM code includes a patch to create bfloat16 array
attributes, thus enabling a different patch to torch-mlir to flesh out
support for the bfloat16 type.
The issue was in the canonicalizer for torch.aten.ge.int -- in cases
where the operands were swapped, it would miscompile. This issue is
fixed and folding support generalized to `torch.aten.size.int < 0` as
well.
Fixes#716
A recent PyTorch commit made ConstantPad2d call a helper function with a
`Union[int, float]` type annotated. This commit adds minimal support for
representing and dealing with that.
https://github.com/pytorch/pytorch/pull/73287
Changes:
- Adding support for `!torch.union<T1, T2, T3>`/`Torch::UnionType`,
along with the importer and CAPI code.
- Add support in isValidSubtype for union types.
- Adding a canonicalizer for `torch.derefine` to help simplify some code
that derefines to a UnionType (this also fixes#664).
There is still more work to do for really supporting UnionType well,
such as canonicalizing UnionType's so that they can be compared with
pointer equality.
The reified code to compute the shape of torch.aten.constant_pad_nd
uses negative indices when setting list elements. This was not
converted to a positive offset in one place in SimplifyShapeCalculations
which prevented computation of the static shape.
The `assemblyFormat` stuff (which generates unrolled, per-op C++ code)
was taking up a lot of compile time, and all the ops are essentially
printed with the same logic. So this PR makes them all call the same
helper function. This is done by using
`let hasCustomAssemblyFormat = 1` and then implementing `FooOp::parse`
and `FooOp::print`.
Additionally, the `Generated*Ops.td` files are all collapsed into just
`GeneratedTorchOps.td` (there is no reason to have the files separate,
since the files are very large anyway so one is always having to search
within them -- editors don't care that the file to search is now a bit
bigger :) ).
This reduces TorchOpsODSGenerated.cpp compile time (which is now
GeneratedTorchOps.cpp) from 39 to 31 seconds on my machine. This is
actually less than I expected, but this PR is an overall cleanup to the
code anyway. The next step will be to introduce (better) functionality
upstream for sharding the TorchOps.cpp.inc file, so that we can truly
parallelize the O(#ops) costs. This is also necessary, because after
this PR, TorchDialect.cpp is now the slowest file to compile, due to the
`addOperations<... all the ops ...>` call, which needs to be shareded
too.
This commit adds support for type refinement when
`torch.tensor_static_info_cast`s are involved, even when there are
users of the casted tensor that don't allow type refinements.
Originally the canonicalization pattern for
`torch.tensor_static_info_cast` would check if all the users of the
casted tensor allowed type refinements before making any changes. This
means that if at least one of the users did not allow type
refinements, the pattern would fail. This becomes an issue when doing
shape calculations because the calculations need the shape information
of each input tensor to be available before the calculation can be
simplified.
The ODS-generated code included via the `TorchOps.cpp.inc` file takes a
very long time to compile. This PR isolates it into its own file so that
the build system can cache it.
This PR creates a new file `TorchOpsODSGenerated.cpp` just to include
the `TorchOps.cpp.inc` file. Doing so required moving to the "new" way
to define verifiers, since the static `verify` free functions in
TorchOps.cpp weren't accessible from the .inc file after it was moved to
`TorchOpsODSGenerated.cpp`.
On my machine, this drops the build time of TorchOps.cpp (such as when
iterating on a canonicalizer) from >40 seconds to <10 seconds.
10 seconds still isn't great though, but at least it isn't "go get a
coffee" type of waiting.
This leads to much more succinct types in many cases:
```
!torch.list<!torch.int>
!torch.list<int>
!torch.tuple<!torch.list<!torch.int>, !torch.list<!torch.int>>
!torch.tuple<list<int>, list<int>>
!torch.optional<!torch.list<!torch.int>>
!torch.optional<list<int>>
!torch.list<list<list<tensor>>>
!torch.list<!torch.list<!torch.list<!torch.tensor>>>
```
I would like to take this further and allow omitting the `!torch.`
prefix in all cases, but that's harder -- for example, we currently use
`FuncOp` for functions, and so I don't think we can customize the
printing there. It seems like it will be a longer road to getting that
level of customization.
See the documentation in `docs/shape_lib.md` and
`docs/adding_a_shape_function.md` for an overview of the system.
This completely overhauls how we represent shape functions. In
particular, RefineTypes does not infer shapes anymore (only dtypes).
Shape functions are now written in (TorchScript'able) Python.
Recommended review order:
1. Read `docs/shape_lib.md` and `docs/adding_a_shape_function.md`.
1. Code and tests for ReifyShapeCalculations, DropShapeCalculations.
1. Code and tests for SimplifyShapeCalculations.
1. shape_lib_gen.py
1. Code and tests for new RefineTypes pass.
1. Random folders/canonicalizers in TorchOps.cpp and associated test in
`canonicalize.mlir`.
1. New ReadOnly trait inferred from the registry.
1. Any miscellaneous remaining stuff.
Example `-print-ir-after-all` for ElementwiseUnaryModule:
[IR lowering dump](https://gist.github.com/silvasean/e4dc8cbc8d00aac7819602e3cbd8e212).
Example `-print-ir-after-all` for ElementwiseBinaryModule:
[IR lowering dump](https://gist.github.com/silvasean/daf6860ecced732af3568af6b1899113).
- This commit adds lowering of `aten.Bool.Tensor` and
`aten.Float.Tensor` op as a part of `convert-torch-to-linalg` pass.
- It also adds support for returning bool types.
- It also fixes lowering of the `aten.Int.Tensor` op for non-zero rank
input tensors.
- If a scalar number is converted to a 0-d tensor and passed on to the
`aten.Float.Tensor` op, it folds to the scalar number.
Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
The added e2e maxpool testcase from #545 was not getting a static shape
due to an unfolded prim.If when RefineTypes was called. This was because
of unfolded torch.iaten.__is__ and torch.prim.unchecked_cast operators
with torch.derefine operands.
Note that to enable folding of the code coming from an example
like the ConstantPad2dStaticModule e2e test, support for other
operations had to be added/improved:
- aten::neg.int
- aten::eq.float
- aten::eq.str
- prim::Uninitialized
- It folds `aten.to.dtype` when the input tensor type and result type
are exactly same.
- It folds `aten.view` when the rank of both the input tensor type and
result type is unity.
Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
This commit adds lowering of `aten.squeeze.dim` op into
`linalg.TensorCollapseShape` op. Here, the dim(th) dimension of the
input tensor is not supposed to be dynamic.
Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
This commit adds lowering of `aten.Squeeze` op into
`linalg.TensorCollapseShape` op. The size 1 dynamic dimensions are not
handled as a part of this commit.
Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
This is to fold the common pattern from Bert inference like:
```
%111 = torch.prim.NumToTensor.Scalar %110 : !torch.int ->
!torch.vtensor<[],si64>
%112 = torch.aten.Int.Tensor %111 : !torch.vtensor<[],si64> ->
!torch.int
```
This change is to unblock the work of some backprop ops returning more
than one tensors. We will need to think of a more scalable approach
in the future if more flexible return types combinations are needed.
- Added a DecomposeComplexOps pass to decompose complex torchOps.
- Refactored `visitAtenArgmaxOp` and `visitAtenAnyDimOp` to
`visitReductionAlongDimIntOp`.
- Moved some helper functions into
torch-mlir/Dialect/Torch/Utils/Utils.h to be shared by multiple files.
- Added support for f64 tensor as argument and return types.
This commit (with approval from all contributors) dual licenses
the torch-mlir project under both the standard LLVM license and the
standard PyTorch license. This will facilitate moving code between
torch-mlir and the two upstream projects.
The standard file comment is now:
```
// This file is licensed under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// Also available under a BSD-style license. See LICENSE.
```
See `LICENSE` in the project root for the terms of both licenses.
This creates the `external/torch-mlir` directory as an
LLVM_EXTERNAL_PROJECTS-compatible project (analogous to
`iree-dialects`) and completes movement/rename of all pure MLIR C/C++
compiler code into there. The next step will be to move all the Python
code / code that links/includes PyTorch C++ code (which currently lives
in `frontends/pytorch`) into a subdirectory here.
I call this "earthmoving" because it is mostly mechanical changes and
renames. As a quick summary (we can change this down the road easily)
- C++ `mlir::NPCOMP::Torch -> mlir::torch::Torch`
- CAPI `npcompTorchListTypeGet -> torchMlirTorchListTypeGet`
- preprocessor `#ifndef NPCOMP_ -> #ifndef TORCHMLIR_`
- CMake `NPCOMPFoo -> TorchMLIRFoo`
The goal of this is to create a standalone project creating a center of
mass for entry into the MLIR ecosystem from PyTorch, suitable in scope
for eventual inclusion/ownership in PyTorch. The idea is that
`external/torch-mlir` will some day be pulled out into its own
repository, and then npcomp will simply pull it in as a submodule.
Layering-wise, what lives in `torch-mlir` lowers code from PyTorch
(currently TorchScript, but TorchFX or pytorch/xla-style tracing are
possible extensions) down to what we have been calling the "Torch
backend contract" which is cleaned up IR (inlining, simplifcation,
conversion to value tensors, ...) entirely in the `torch` dialect. This
is the branching off point for further lowering, of which npcomp takes
one opinion (outside `torch-mlir` of course!), namely the
`TorchConversion` dialect/transforms which lower to IR suitable for IREE
and other linalg-on-tensors based lower-level compilers.
Summary of changes:
- move `{include,lib,test}/Dialect/Torch` into `torch-mlir`
- move relevant parts of CAPI into `torch-mlir`.
- leave a few things related to the `torch-mlir` Python build commented
out, which should be resolved in a subsequent change.
This contains the following changes:
- Fix optional knowledge propagation. The initial knowledge should
always be NotNone for the operations we implemented.
- Add Folder for `prim.dtype`
- builder.getSymbolRefAttr is gone.
- OpAsmOpInterface's getAsmResultNames method needs explicit override
- a bunch of churn for builtin.func needing to be made explicit (and
sometimes implicit?)
- operation printers no longer need to print the operation name
themselves.
- snuck in beneficial trivial addition to TmpDeleteDeadIREEListsPass to
test a particular upstream change e2e with my local patchset.
- Add `!torch.optional` knowledge tracking
- Changes to improve type propagation for branches and terminators. See
examples in `refine-types-branch.mlir`
- Refator to separate handling of different ops from `visitOperation`
- Add refine types for a few new ops
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.
This includes the following changes to import MT model into MLIR. There
are still a lot of work to for actual compilation.
- Add `torch.dict<>`, `torch.any`, `torch.number` types
- Add `torch.prim.DictConstruct` op
- Fix `torch.prim.TupleConstruct` op assembly format to include resulting types
- 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.
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.
- 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 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.
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.
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.
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
In terms of IR structure, TorchScript allows types to vary in many
circumstances where MLIR requires pointer-identical types. In particular,
it is valid to pass any subtype in place of a type. For example, if an
`Optional[int]` is required somewhere in the IR, it is legal to pass a
value of just `int` (but not the other way around; see
`torch.prim.unchecked_cast`). In effect, every *use* can have a different
type.
We introduce a new op `torch.derefine` that models that impedance
mismatch. This op allows casting a value from one type to a type that it
is a subtype of to model this behavior.
Recommended review order:
- TorchOps.td for new torch.derefine (and updated docs for
`torch.prim.unchecked_cast`)
- new test code in if.py, loop.py, function-derefine.py
- new code in node_importer.cpp for handling derefinement insertion
- function_importer.cpp and utils changes in torch_to_mlir_utils.cpp
Properly handling derefinement on function boundaries required
relayering the code so that graph_importer.cpp/.h is now
function_importer.cpp/.h because only the `torch::jit::Function`
(actually the `c10::FunctionSchema` it holds) knows the derefined types that are
actually needed at the boundary (see `function-derefine.py` for a test).
Annoyingly, this churns all the functions which are now prefixed with
`__torch__.` but that is more correct anyway (that is their linkage name
in the `torch::jit::CompilationUnit`; the previous `mb.import_function`
was actually buggy in the case of functions calling each other as it
would reference their unqualified name).
With this change, we can import `resnet18` from `torchvision` :)
IR: https://gist.github.com/silvasean/6426a5272d8a6c7caae533fce05ab704
With this + manually setting private visibility on everything, a simple
classifier can be reduced to this IR, which is looking pretty lean and
mean:
https://gist.github.com/silvasean/19e7e2e21a61ff197aeac0dd864d188f
Also, include a utility script for importing `.pt` models.
```
pt_util.py --import classifier.pt | npcomp-opt -torch-globalize-object-graph
```
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
* 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.
* 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.
武家伟
Sambhav Jain
gpetters94
Sean Silva
Marius Brehler
Ashay Rane
Quinn Dawkins
Vivek Khandelwal
Quinn Dawkins
JakopinA
erman-gurses
Maksim Levental
Liam Fitzpatrick
Qiang Fu
Vigilans
Ramiro Leal-Cavazos
Nirvedh
Gaurav Shukla
Yi Zhang
Liam Fitzpatrick
stephenneuendorffer
Prashant Kumar
dan
Stella Laurenzo