Add a new op with shape/dtypes and decompose into
`fake_quantize_per_tensor_affine` when the second result is unused.
The xfail_set change is on ONNX because torch cannot export this op to
ONNX.
Resolves#3384.
Many ONNX operators are defined by functions and therefore could be
expanded into simpler ONNX operations during importing, avoiding the
need for tools downstream to support these operators directly.
This commit adds this capability to onnx_importer.py. When importing a
node, the schema for the node's operator is retrieved. If the schema
provides a function for the operator, a specialized version for the
node's types and attributes will be created and imported as an MLIR
function with private visibility. An MLIR function call will then be
emitted, instead of a normal operator node. Caching is used to avoid
generating redundant functions within the same module.
In order to avoid a disruptive change to the importer output for a
large number of operators that already have TorchOnnxToTorch support,
an allowlist strategy is used by default. With this commit, only one
operator is allowlisted for expansion, MeanVarianceNormalization.
However, many other operators can be correctly expanded by the current
code, so hopefully the allowlist can be gradually extended. It is
possible to disable the allowlist in the configuration, in which case
all functions are expanded (useful for testing).
Tools downstream of the importer may now need to do inlining when
consuming the output of the importer, e.g.:
cat imported.mlir | torch-mlir-opt --inline --convert-onnx-to-torch
Explanations for subtle code changes:
- Looking up the correct schema and function for an operator requires
knowing the opset version. NodeImporter retrieves this from the
opset imports on the ModelProto retained by the GraphInfo. Previously,
the model_proto field on GraphInfo was None when importing a subgraph
in import_regions, but this conflicts with the new need for opset
version info. Since the apparent purpose of setting it to None was to
control how GraphInfo generates its input map, a new flag is added to
GraphInfo (is_subgraph) to control this behavior, so that the actual
ModelProto can now be provided without breaking this. This also turned
out to be useful for getting the Config via ModelInfo via GraphInfo.
- Some operators' functions are context-dependent, which means the
function definition depends on the types of the inputs. Therefore node
importing now needs to look up the types of a node's inputs, not just
its outputs as was the case previously. Consequently the operand to
find_type_proto_for_name() may now be a graph input or initializer in
some cases, so it has to be updated.
Issues was found here https://github.com/nod-ai/SHARK-Turbine/issues/643
- [ONNX] Fix padding attributes for onnx.AveragePool
- [Linalg] Add countIncludePad false support for AtenAvgPool1/2dOp
- [Linalg] Add an avg_pool2d countIncludePad False e2e tests
- [Linalg] Fix conflict with AtenAvgPool3dOp
- [Linalg] Fix e2e crash with AtenAvgPool1dOp
- [Linalg] Add dynamic dim support for AtenAvgPool2dOp
- [Linalg] Fix AvgPool2dDivisorOverrideModule crash
This commit also adds the Torch declaration for aten.max_unpool2d and
aten.max_unpool3d op. The TorchToLinalg lowering for the same will be
added in a follow-up commit.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
This addresses 7 of the model failures I'm seeing in the test suite. See
[Shark-Turbine issue
#566](https://github.com/nod-ai/SHARK-Turbine/issues/566).
Need the op ```linalg.conv_2d_ngchw_gfchw_q``` to be added upstream
before merging this. See [llvm-project PR #92136
](https://github.com/llvm/llvm-project/pull/92136).
A small additional expansion to operand quantization is included in this
patch to address a model failure that occurs when unblocking the
quantized group convolutions in one of these onnx models.
Updates:
- some unsupported modes are now going to report a match failure for
unsupported coordinate transformation modes.
- fixes a bug that was introduced in the last patch for resize (my
bad...)
- uses actual x and y coordinates for computing weights in bilinear
interpolation (rather than eps modified values)
- slightly simplifies the bilinear interpolation payload for readability
and performance
- passes coordinate transformation mode information from an onnx.Resize
op to the mode string for the aten._interpolate op. This allows us to
perform custom logic in the torch->linalg lowering to support
onnx.Resize options without losing the default behaviors of the
interpolate op.
This commit fixes the onnx.MaxPool op lowering which was lacking the
indices result support.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
* not to decompose `aten.amax` on `stablehlo` backend. Because it could
be lowering to `stablehlo.reduce` directly.
* lowering `aten.max.dim` to `stablehlo.reduce apply max` when
`AtenMaxDimOp.getIndices()` doesn't have users. It's more simple.
…cation and sparse tensors.
**NOTE**: This PR _doges_ the issue in buffer-deallocation pass instead
of resolving it. In the future, we need to fix the bug in
buffer-deallocation pass when handling code generated by sparse
compiler.
While playing with TorchDynamo on ResNet18. I notice following issues:
- `prims.convert_element_type` can’t be canonicalized even if the input
and the output share the same type
- `aten.max_pool2d_with_indices` is always used instead of
`aten.max_pool2d`, even if the second returned output (indices) has no
user
This PR fixes above issues by adding a folder to the
PrimsConvertElementTypeOp and a canonicalizer to the
AtenMaxPool2dWithIndicesOp
Lit test:
`cmake --build build --target check-torch-mlir-all`
---------
Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
This is probably a decent PR for learning about blocks and regions.
If you're here to learn about that, consider also looking at
lib/Conversion/TorchToSCF/TorchToSCF.cpp
While this doesn't include an e2e test, it is tested downstream in
https://github.com/nod-ai/SHARK-TestSuite/blob/main/e2eshark/onnx/operators/If/model.py
---------
Co-authored-by: Xida Ren <xida.ren.dev@gmail.com>
I spent a little while debugging numerics issues with some tests similar
to the ones in quantized_models.py, only to find that pytorch's
quantized conv transpose is catastrophically inaccurate. I'll upstream
the issue and only leave the tests here which are of the form quantize
-> dequantize -> op.
For some sparse programs (and I am sure other not-seen corner cases for
dense), some passes were missing in the reference pipeline, eventually
resulting in e.g. a unresolved unrealized cast issue. This PR adds some
very obvious missing passes to avoid this situation.
This is a large change because prior to this point, Python files in the
project were not consistently formatted. This reformats them all with
black defaults.
Based on experience with prior projects, if you have a dev/long-term
branch with Python patches, you can minimize merge conflicts prior to
rebasing to include this commit by running `black` on your modified
Python files, squashing, and then rebasing/merging.
This is part 1 of ~3, formatting all miscellaneous text files and CPP files matched by a first run of pre-commit. These tend to be low change-traffic and are likely not disruptive.
Subsequent patches will format Python files and remaining CPP files.
A choice was made to quantize the return type of Relu with a scale and
zero point copied from the input's quantization scheme. With this
choice, the torch-to-linalg conversion of quantized Relu essentially
computes max(input, zeroPoint) in the elementwise payload.
* promote input to output element-type when lowering to stablehlo, so
that it could satisfy stablehlo's type constraints.
* split promote-to-fp unary ops from fp-only unary ops.
This commit also cleans up the OnnxToTorch lowering for the Squeeze and
Unsqueeze op and adds the support for handling edge cases.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
Version number was set too high. Lowered to support more cases allows
more tests to pass.
Co-authored-by: Robert Suderman <rsuderman@Roberts-MacBook-Pro.local>
Previous implementation erroneously mixed up num_outputs with
slice_size. New version correctly computs the slice size and directly
performs slicing rather than leveraging `aten.split.tensor`. This is due
to `onnx` supporting a fixed number of splits making the size
computation more easily computeable when lowering to `aten` rather than
deferring to `aten.split.tensor`.
---------
Co-authored-by: Robert Suderman <rsuderman@Roberts-MacBook-Pro.local>
Matthias Gehre
Vivek Khandelwal
Xinyu Yang
Branko Trifkovic
ptrifunovic98
Andrea 🦈
Wu Yuan
Phaneesh Barwaria
Vinayak Dev
Chi_Liu
Sambhav Jain
penguin_wwy
Yuanqiang Liu
zjgarvey
Suraj Sudhir
Peiming Liu
Andreas Falkenberg
NeverRaR
aldesilv
Jiawei Wu
Ze Zhang
Prashant Kumar
Xida Ren (Cedar)
Aart Bik
Stella Laurenzo
Rob Suderman