The original design for the dtype functions outlined in
https://github.com/llvm/torch-mlir/issues/1462 was unable to properly
handle ops that take optional tensors as an input when the optional
tensor has a value of None. By the time the op gets imported into
torch-mlir, if an optional value is None, all information about the
original type is lost from the op type signature, preventing
torch-mlir from knowing if a value of None was from an optional tensor
or not, which was crucial in the original design since each tensor
argument must be turned into two separate arguments for the dtype
function.
This commit changes the interface to dtype functions such that each
tensor turns into a tuple of two ints, the first representing the rank
of the tensor and the second the dtype of the tensor. Since now there
is a one-to-one correspondence between the operands of an op and the
operands of its dtype function, there is no ambiguity about which
operand of the op corresponds with which operand of the dtype
function.
To test the implementation, this commit defines dtype function for
convolution op, which takes one optional tensor as an argument.
* LowerToBackendContract: Explicitly error out on unimplemented operator
But only reject torch.operator when results are invalid.
Otherwise it might be a custom op that the backend supports.
Currently, the op `torch.tensor_static_info_cast` will not get
canonicalized away if the result type has any shape or dtype
information. This is because `isValidSubtype` only returns true when
the tensor types being compared are exactly the same or the supertype
has no shape and dtype information. Being unable to canonicalize away
the `torch.tensor_static_info_cast` gets in the way of further
optimizations, such as shape propagation.
This commit improves `isValidSubtype` by adding logic that compares
the shapes and dtypes of the two tensor types to determine of one type
is indeed a valid subtype of the other.
Fixes https://github.com/llvm/torch-mlir/issues/1926
The data-flow analysis does not always propagate information to the
entire graph. This results in some lattice elements being
uninitialized. Currently the lattice elements are not checked to see
if they are uninitialized before rewriting the graph, potentially
resulting in invalid IR (see
https://github.com/llvm/torch-mlir/issues/1896).
This commit adds handling for uninitialized lattice elements.
This commit replaces the `tanh` dtype function, which was being used
to test the implementation of dtype functions in
a710237437, with a dtype function for
`expm1`. The dtype function for `expm1` is identical to the `tanh`
one, so the same level of testing is maintained.
Currently, there are ops getting dtype information from the
`RefineTypes` pass and ops getting dtype information from the
`TorchDtypeRefinementPipeline`. Since each pass can only propagete
dtype information for the ops it knows how to handle, some models with
many ops handled in both passes require the two dtype propagation
passes to execute many times, reaching the iteration limit set in the
`LowerToBackendContractPass`. To temporarily avoid this issue while
the migration to `TorchDtypeRefinementPipeline` is finished, this
commit switches `tanh` to `expm1`, since the latter is used a lot less
in large models.
In order to verify if a given IR satisfies the backend contract, the
verifier needs to know if decompositions took place, and if so, which
ops were decomposed and which were not.
This commit adds two arguments to `verifyBackendContractPass` to
specify if decompositions took place and which ops to consider backend
legal, similar to the arguments of `LowerToBackendContractPass`.
* [custom op] Generalize shape library logic to work with dtypes
This commit generalizes the shape library logic, so that dtype rules
for ops can also be expressed using the same mechanism. In other
words, each op can now have a shape function and a dtype function
specified in Python that is imported during lowering to calculate the
shapes and dtypes throught a program. For more information about how
to specify a dtype function, see the updated
`docs/adding_a_shape_and_dtype_function.md`.
For those not familiar with how the shape library works, the file
`docs/calculations_lib.md` provides an overview.
The current implementation of `DecomposeComplexOps` fails if an op
expected to be decomposed does not get decomposed in the first
iteration of the `createTorchSimplificationPipeline` in
`LowerToBackendContractPass`. However, some graphs require multiple
iterations of `createTorchSimplificationPipeline` to fully propagate
all statically knowable information, such as dtypes and shapes, to the
entire graph, sometimes resulting in the need to run
`DecomposeComplexOps` more than once.
This commit changes `DecomposeComplexOps` to use a greedy algorithm
for pattern application and moves the legalization check of ops to the
`LowerToBackendContractPass` to allow for the `DecomposeComplexOps` to
run more than once.
- Support for non-prefixed accessors has been removed. See:
https://reviews.llvm.org/D136727
- Rename `operands` to `methodOperands` in `prim.CallMethod` since the
name `operands` overlaps with a builtin method name. See:
https://reviews.llvm.org/D136727
- Add passes in refbackend to lower memref.subview. See:
https://reviews.llvm.org/D136377
- Replace `CopyToValueTensorOps` first in `RewriteViewLikeSubgraph` in
maximize-value-semantics.
The current implementation of the `RewriteViewLikeSubgraph` pass in
maximize-value-semantics creates temporarily invalid IR. In
particular, given a forward slice starting from a
`CopyToNonValueTensorOp` and ending in `CopyToValueTensorOp`s, the
pass first replaces all uses of the `CopyToNonValueTensorOp` with
its operand, which results in all the `CopyToValueTensorOp` users
having their operand have type `!torch.vtensor`, which is invalid.
The correct way to do things is to first replace all the
`CopyToValueTensorOp`s with their operand, and then replace all uses
of the `CopyToNonValueTensorOp` with its operand.
This only started failing now because the generated accessor
`getOperand` for the `CopyToValueTensorOp` now returns a
`TypedValue<NonValueTensorType>`, which has an assert checking that
the value returned is of the expected type.
Summary of changes:
- Replace call to `MemoryEffectOpInterface::hasNoEffect`
with `isMemoryEffectFree`.
- Make fix for the dynamic dims, since
`kDynamicSize` value changed to
`std::numeric_limits<int64_t>::min()` from `-1` in llvm
- `makeShapeLLVMCompatible` and `makeShapeTorchCompatible`
utilities convert shapes in order to remain consistent
with the Torch and MLIR semantics.
- Update tags
llvm: 147fe9de29dc13c14835127b35280c4d95c8e8ba
mhlo: 1944b5fa6062ec4c065d726c9c5d64f1487ee8c5
Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
This commit fixes the aten.mean and aten.mean.dim op decomposition
for supporting large-sized inputs.
This commit also fixes the formatting for the file stats.py
Signed-Off By: Vivek Khandelwal<vivek@nod-labs.com>
This commit removes almost all of the valsem ops, since the value
semantics version of the ops now exist in PyTorch. The only op missing
is `aten.bernoulli_.float`. In addition, this commit also simplifies
the implementation of `aten.fill.Scalar` by moving it to the pattern
that converts elementwise ops.
This commit makes the following changes needed to update bump LLVM:
- Replace `linalg.init_tensor` with `tensor.empty` (see:
https://reviews.llvm.org/D135129)
- Replace `NoSideEffect` with `Pure` (see
https://reviews.llvm.org/D135505)
- Replace `body` region accessor for `ReduceOp` and `ReduceWindowOp`
with `getBody`
- Fix incorrect use of `tosa::ReduceSumOp` in `AtenNativeLayerNormOp`
conversion pattern. The result type of `tosa::ReduceSumOp` must have
the same rank as the input type. (see:
https://www.mlplatform.org/tosa/tosa_spec.html#_reduce_sum)
Co-authored-by: Ashay Rane <ashay@users.noreply.github.com>
Co-authored-by: Ashay Rane <ashay@users.noreply.github.com>
This commit adds lowering of `aten.div.int` and `aten.bitwise_or.Tensor`
ops. Both these ops are required in order to support bloom_560m model.
Signed-Off-by: Gaurav Shukla <gaurav@nod-labs.com>
We were already hitting many cases where backends different in terms of
the legal ops that they wanted. This caused unnecessary coupling between
the backends. Examples:
- https://github.com/llvm/torch-mlir/pull/1161
- https://github.com/llvm/torch-mlir/pull/862
This PR centralizes all compilation to go through `torch_mlir.compile`
so that we can keep the logic centralized there. We should move these
lists closer to each backend. Especially cases like
https://github.com/llvm/torch-mlir/pull/862 where blocking a
decomposition is necessary to avoid a crash emphasize that the set of
decompositions is tightly coupled to the backend, and should be
"controlled by the backend" and not something arbitrarily tweakable.
Also:
- Fix a small bug in the way we passed through the backendLegalOps
option.
- Add better error messages in `torch_mlir.compile` for import errors.
One of the simplifications made by the pass `RefinePublicReturn`
currently only happens if the tensor in question only has one
user. However, the current method of checking this does not correctly
handle the case of a user having multiple uses of the same
tensor. This commit makes sure only unique users are considered.
This is a first step towards formalizing the set of ops in our backend
contract. The goal is to eventually formalize `torch` dialect ops into 3
categories:
1. Legal in backend contract
2. Illegal in backend contract
3. Conditionally legal in backend contract
The "conditionally legal" set are the ops that we can optionally
decompose for backends.
This patch adds relevant pass options for this throughout the compiler,
in preparation for a new set of traits which will formalize this
classification.
This introduces a new pass LowerToBackendContract (better name very
welcome) which performs the bulk of the simplifications that we do,
such as
- shape refinement
- dtype refinement
- maximizing value semantics
- inlining global slots
- decomposing complex ops
The key difference from before is that it iterates the set of
transformations, which can help to break a number of "catch-22" issues
where one simplification depends on another, the latest example being
here:
https://github.com/llvm/torch-mlir/issues/1131
This also exposed that RefineTypes was sometimes crashing/asserting for
certain inputs. This commit hardens it a bit.
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).
Ramiro Leal-Cavazos
Matthias Gehre
Ziheng Jiang
Zhekun Zhang
Yuanqiang Liu
Gleb Kazantaev
ataheridezfouli-groq
Ahmed S. Taei
Vivek Khandelwal
Tanyo Kwok
Daniel Ellis
Xiafei Qiu
Gaurav Shukla
武家伟
gpetters94
Sean Silva
PhaneeshB