Effectively, this mode works by compiling op by op as the NN is eagerly executed by PyTorch. Entailed in that compilation is building a representation of the op that can be `torch.jit.script`ed, importing using `ModuleBuilder`, and then executing (e.g., with `RefBackendLinalgOnTensorsBackend`). This mode includes a fallback to conventional PyTorch if anything in the torch-mlir compilation process fails (e.g., unsupported op).
Currently, all e2e tests pass execpt for two that involve an upstream PyTorch bug (https://github.com/pytorch/pytorch/issues/74400).
High priority next steps:
1. A compile cache in order to speed up reruns of the same NN.
2. Integration with IREE (though not in this repo).
3. Integration with `torch.distributed`.
Torch FX was never really a different path, since all FX modules are
actually valid TorchScript modules. Instead, replace it with the new
torch.dispatch work that we are building.
- This commit adds decomposition of `aten.dropout` op. It also covers the
training mode of the same op.
- It also adds lowering of `aten.sub.float` op.
Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
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 the op `ValsemVariantAtenCopyOp` that represents
`AtenCopy_Op` without the underscore. This is needed to make sure
that the `ReduceOpVariants` pass turns the in-place op into an op
that takes value tensors as inputs, otherwise the
`MaximizeValueSemantics` pass will not be able to add value
semantics correctly.
This commit also adds the lowering of `ValsemVariantAtenCopyOp`.
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
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.
This commit fixes the 2nd and 3rd return types of the `aten.native_layer_norm`.
Previously the mean and rSTD were returned with reduction dims removed.
This commit fixes this and keeps the reduction dims of the results.
Signed-Off-By: Prateek Gupta <prateek@nord-labs.com>
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 commit adds the op `ValsemVariantAtenIndexPutImplOp` that represents
`Aten_IndexPutImpl_Op` without the underscore. This is needed to
make sure that the `ReduceOpVariants` pass turns the in-place op
into an op that takes value tensors as inputs, otherwise the
`MaximizeValueSemantics` pass will not be able to add value
semantics correctly.
This commit also adds the lowering of `ValsemVariantAtenIndexPutImplOp` op.
This commit also updates the `torch.bincount` op test cases.
The term "pseudo" is very vague and was getting confusing (I felt I had
to explain it in every comment referencing it). Instead, rework the
"pseudo" ops to instead be named:
- MLIR Syntax: `torch.valsem.*`
- C++ / ODS: `ValsemVariant*Op`
This makes it clear what the concept is, and avoids confusion with other
things that might be called "pseudo", since these are very specific and
should be 100% consistently named w.r.t. the non-valsem-variant ops that
they correspond to.
This is code that we always want to treat as "foreign" and not get too
comfortable using in many functions. One way to accomplish that is to
make it a bit clunkier to use.
Also, fix Utils.cpp to match the LLVM/MLIR coding conventions (don't
define functions inside namespaces -- prefer `using` and explicit
qualification).
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.
This was an aspirational goal at an earlier stage in the project where
the focus was heavily on programs with state, control flow, and
lists/dicts. We will circle back to such programs likely 2022H2 at some
point, but for now, having this test doesn't add much, since basically
nothing works or is being worked on.
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 helps keep things organized and also exposes more parallelism to
the build system. It seems though that most of the compile time is
actually spent in the headers though, so the wall time doesn't decrease
as much as I had hoped (and now that the headers are being included
multiple times, the cpu time actually increases a lot, sadly -- will try
to dig into this).
This commit replaces the two rewrite patterns of
maximize-value-semantics with a single pattern that captures the
behavior of both as well as other edge cases previously not
supported. The new pattern works by first performing alias analysis on
a subgraph to see if pattern is applicable, then rewriting all
non-value tensors to value tensors in a single go.
This pass is added to lower ops, which can not be lowered
via the TorchToLinalg pass, such as `torch.bincount` op.
This pass also uses torch-mlir's TMTensor Dialect to lower the
complex ops.
Also add torch.bincount op lowering with the help of TMTensor dialect
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
This commit moves the helper function which are common across
different torch-mlir conversion passes into a common directory
Utils.
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
This commit adds support for integer type inputs for
`AtenMaxOp`, `AtenSumOp`, `AtenSumDimIntListOp`.
Signed-Off By: Vivek Khandelwal <vivek@nod-labs.com>
- This commit adds E2E support for `aten.rand_like` and
`aten.bernoulli_.Tensor` ops.
- The `aten.bernoulli(x)` was implemented as:
`aten.bernoulli(x) = rand_like(x) < 0.5`, assuming 0.5 as default
probability, whereas according to the pytorch documentation:
https://pytorch.org/docs/stable/generated/torch.bernoulli.html#torch.bernoulli
the input x in `aten.bernoulli(x)` is itself a tensor containing
probabilities to be used for drawing the binary random number.
- So this commit fixes the `aten.bernoulli(x)` implementation as:
`aten.bernoulli(x) = rand_like(x) < x`.
- It also fixes the case where the input to `aten.bernoulli_.float` is
an integer tensor. In this case the input must be casted to float type
before passing it as operand to `aten.rand_like` op.
`aten.bernoulli_.float(x, p) = rand_like(float(x)) < p`.
Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
The pass is mostly borrowed from the BufferizeAnyLinalgOp pass in mlir
upstream with some minor changes. At a high level, it's a naive partial
bufferization pass which allocate new buffers for all the output
tensors. The initial value of an output buffer is copied from the
original buffer if there are uses of the original value.
One difference from linalg bufferization pass is the way to tell if
the loop body uses the init value of output operand. For TMTensor ops,
it differs from op to op because the payload region doesn't represent
the entire loop body.
The view op allows for the new shape argument to have a -1 value for
one of the dimensions, and the op is expected to deduce the size of
that dimension by looking at the sizes of the other dimensions and
comparing it to the total number of elements in the original
tensor. This commit adds this functionality.
This commit does a couple of things. First, it fixes a bug in the
`linalg.generic` body of the `nll_loss_forward` lowering where the
`ignoreIndex` was being compared with the loop index rather than the
current element of the `target` tensor. This was not being caught by
the tests because they were not testing the case where `ingnoreIndex`
actually corresponds to a value in `target`. This has been fixed.
Second, this commit adds support for the `reduction` argument in
`torch.nll_loss_forward` as well as support for 1-D inputs. In order
to simplify the lowering code, I've refactored the code that creates
the `linalg.generic` ops for elementwise and reduction ops into static
functions, to avoid having boilerplate code for indexing maps, etc
that can be very error prone.
Note: The function `convertScalarToDtype` was moved to before all the
conversion patterns, but nothing in it was modified.
Its unclear to me what the right layering is here: Are you expecting
torch-mlir-dialects to always get built with LLVM? This is pretty
breaking for us, if so.
This should be set elsewhere depending on the build configuration.
In particular, we need to be careful when cross-compiling to pick
up the host mlir-tblgen.
This commit adds the invariant to the op `torch.overwrite.tensor.contents` that
both of its operands have the same shape and size. In order to
maintain the invariant, special handling of this op is added to the
`RefineTypes` pass.
There are cases where the op may update the same indices multiple
times. In this context, we can not parallelize updates. Instead,
we have to execute them sequentially. Adding a boolean attribute
to control the behavior.
Also adding test cases for invalid IR.
This commit adds handling to the `maximize-value-semantics` pass for
the case where a view-like op depends on a tensor that has been
overwritten by a value tensor. The approach for removing the
dependency is to change the input to the view-like op to be a copy of
the value tensor that is being used to overwrite.
This commit also removes `AtenFill_ScalarOp` and
`AtenBernoulli_FloatOp` from the list of view-like ops, since these
ops now have a corresponding op with value semantics into which they
get converted in the `reduce-op-variants` pass.
max
Ahmed Taei
Sean Silva
Gaurav Shukla
Vivek Khandelwal
Vigilans
Ramiro Leal-Cavazos
Prateek Gupta
Yi Zhang
Prashant Kumar
Stephen Neuendorffer
Jae Hoon (Antonio) Kim