There is currently no int16 quantization support in torch. This patch
adds a new mlir type to correspond to the missing "torch.qint16" type,
and enables lowering of quantization-related onnx ops using int16 types.
In follow-up patches, custom quantization logic for ops like
aten.matmul/aten.mm/aten.convolution may need to be revisited to allow
support for qint16. The passes in FuseQuantizedOps.cpp may also need
slight modifications.
Supports asymmetric padding by performing a torch.nn.functional.pad on
the input before performing the convolution.
Signed-off-by: Suraj Sudhir <suraj.sudhir@arm.com>
All e2e iree tests compiled, but they have the run issue of mismatch of
dtype like the following
```
expected:
1x1x2x2xsi32=[[[12 16][24 28]]]
actual:
1x1x2x2xi32=[[[12 16][24 28]]]
```
See the related issues here:
[SHARK-Turbine#556](https://github.com/nod-ai/SHARK-Turbine/issues/556)
1. Adds uint8 casting to onnx.Cast op
2. Fixes an issue with onnx.DequantizeLinear when the scale comes with
shape [1].
3. Adds support for unsigned types in an AtenItemOp folder
4. Adds a simpler quantized model for easier debugging
5. Adds a fusion pass to convert [quant -> dequant -> transpose -> mm]
patterns to [transpose -> quant -> mm].
6. Moved some xfails that are still not passing, but for different
reasons than onnx.cast failures.
This adds support for converting DynamicQuantizeLinear from torch-onnx
to torch.
I could not get an e2e test to pass, since there seems to be some issues
with uint8 casting somewhere lower in the pipeline. For example
compiling with IREE for llvm-cpu, I would get either the correct zero
point (if zp < 128) or the correct zero-point minus 256 (if zp >= 128).
The output tensor seems to always return a tensor of zeros, which also
occurs when running uint8 examples through QuantizeLinear.
Edit: the first problem can be resolved by casting the output back to
uint8 on output, the second problem is resolved with PR #3018
The only difference between version 7 and newer versions is support for
different data types. We should allow this pattern to match as early as
7. Earlier versions have a more manual broadcast specification through
attributes, so I did not include those versions.
See: [onnx.Div
docs](https://onnx.ai/onnx/operators/onnx__Div.html#l-onnx-doc-divl)
If the broadcast shape is length-1 at a dim while `?` in the input dim
then we need to broadcast to the dynamic dim. This is equivalent to
taking a max of two dimensions.
`getRawBuffer` expects a densely packed vector of `i1` values however
`onnx` does not densely pack the values. Include code to handle the
packing / unpacking.
There is no reason to treat `ConstantOfShape` as a specialized import
any as there exists a onnx-to-torch equivalent. Dropping the import
coding and adding support for resource conversion substantially
increases test coverage for dynamically shaped tests.
This commit adds the OnnxToTorch lowering for cosh, acosh, asin, asinh,
and atanh op.
This commit also adds the TorchToLinalg lowering for acosh, asin, asinh,
and atanh op.
Signed-Off By: Vivek Khandelwal <vivekkhandelwal1424@gmail.com>
This PR contains three commits to update the validation checks in the
ONNX -> Torch conversion pass for the AveragePool, Pad, and Slice operators:
> onnx: fix preconditions for lowering AveragePool ops
>
> The `pads` attribute of the AveragePool operator specifies the value to
> pad at both the beginning as well as the end of the axis (see
> https://onnx.ai/onnx/operators/onnx__AveragePool.html#attributes), so
> the size of this attribute should be twice the rank of the input tensor.
> However, our TorchOnnxToTorch bails out early since it incorrectly
> compares the pads attribute with the rank (not twice the rank) of the
> input tensor.
>
> This patch fixes the code to match the spec and adds a lit test.
> onnx: allow optional constant value for Pad operator
>
> The `constant_value` input of the onnx.Pad operator is optional (see
> https://onnx.ai/onnx/operators/onnx__Pad.html#inputs), but the
existing
> logic for lowering the operator into the Torch dialect assumes that it
> is mandatory.
>
> This patch makes the attribute optional and constructs a default value
> (a list of zeros the size of the input tensor) if the attribute was not
> specified.
> onnx: fix checks for axes and steps inputs of Slice operator
>
> The ONNX Spec for the Slice operator allows the `starts` and `ends`
> inputs to have fewer indices that the dimensions of the `data` tensor
> (see https://onnx.ai/onnx/operators/onnx__Slice.html), but our code
> expects these inputs to be as many as the `data` tensor's dimensions.
>
> More precisely, the spec requires that the `starts` and `ends` inputs
> are only as long as the `axes` input, but since the `axes` input is
> optional, the default type for the `axes` input has to match the type
> for the `starts` and `ends` inputs. Moreover, the number of indices in
> the `steps` input also has to match those in the `axes` inputs (instad
> of matching the dimensions of the `data` input).
>
> This patch fixes the checks in the TorchOnnxToTorch conversion so that
> they match the ONNX spec.
Leaning on the QDQ functionality in torch we can support the QLinearConv
operation by piggybacking through `torch.Convolution`. This includes
some changes such as allowing the `onnx` rewriter to run recursively.
Doing so allows `QLinearConv` to decopmose to `onnx.Convolution` which
is then lowered to `torch`.
The existing `flatten` lowering did not define what the intermediate
shape was. This could result in failures to lower further to linalg as
the intermediate shape was unknown. Added a shape refinement section.
So that the CumSum Op in OPT can get the constant that it requires to be lowered to TMTensor
---------
Co-authored-by: Rob Suderman <rob.suderman@gmail.com>
Co-authored-by: Xida Ren <xida.ren.dev@gmail.com>
`onnx` explicitly specifies that `raw_data` is stored in `little-endian`
layout. While converting
to `torch` we need to convert from a known endian format to an internal
format of consistent
layout. This means endianness must be correct during the import of
`onnx.Constant`.
---------
Co-authored-by: Xida Ren (Cedar) <cedar.ren@gmail.com>
We can make the per-tensor version of the operation to the dequantize
operation via marking with the make quantized tensor component. This
introductions the `qint*` and `quint*` tensor type that can be lowered
to teh appropriate dequantization behavior during the torch-to-linalg
conversion.
zjgarvey
Vivek Khandelwal
Yuanqiang Liu
Suraj Sudhir
Xida Ren (Cedar)
jinchen
Vinayak Dev
Rob Suderman
penguin_wwy
Xinan Jiang(姜曦楠)
aldesilv
Scott Todd
Ashay Rane
Gaurav Shukla
Xida Ren (Cedar)
Phaneesh Barwaria
Chi_Liu
James Newling
Ramiro Leal-Cavazos
Dave Liddell