Linalg has quantized specific operations. We can lower to these
operations when there is a known zeropoint and scale operations. This
allows the `convolution` to occur with lower bitwidth's, improving the
overall performance.
After noticing a number of commits with unrelated formatting changes,
I think something was changed with clang-format at one point and we're
seeing a number of unrelated changes. Doing a refresh can help avoid
this.
The changes made here came from
```
find lib -iname *.h -o -iname *.cpp | xargs clang-format -i --style=llvm
find include -iname *.h -o -iname *.cpp | xargs clang-format -i --style=llvm
find projects -iname *.h -o -iname *.cpp | xargs clang-format -i --style=llvm
```
Torch does not have an equivalent matmul operation for integers. Instead
it sidechannels the information via its quantized types. For this
lowering we setup these sidechannels then invoke `torch.mm`.
This preserves sparsity at the most obvious places of lowering TORCH
tensors to MLIR RankedTensorType tensors. Other places are marked for
audit. With some initial lowering tests.
This includes custom op matching for decomposed operations and fusing
dequantization into dense operations. As a validation we compare
to the dequant+mm torch implementation.
We can plumb the linear matmul into pytorch using its quantized types
with side channel information. To handle the final int8 operation we
dequantize and requantize.
This commit adds mapping from `onnx.pad` op to `torch.pad` op. Currently
it does not support `axes` parameter of `onnx.pad` op.
Signed-off-by: Gaurav Shukla <gaurav.shukla@amd.com>
The logic here is very similar to the conversion for AdaptiveAvgPool1d
#2661 with a few modifications:
1. buffVal = -inf instead of 0
2. the main linalg generic op accumulates a max, instead of a sum, to
the first output tensor
3. avg pooling requires dividing the sum pool by the kernel width, which
we stored as an auxilliary tensor (kSizeTensor). Here, the auxiliary
tensor will be recording the indices. Strangely enough, the only
signature available for this function is to return indices, and it
appears that they must be computed whether the user desires them or not.
See
[pytorch/torch/nn/functional.py](https://github.com/pytorch/pytorch/blob/main/torch/nn/functional.py#L1174).
Before writing other adaptive pooling conversions, the logic of this
decomposition should be rolled into a helper function that will work for
both max and avg pooling ops. Even the auxiliary tensor should likely be
automated. This code was written in a slightly more tedious way than
strictly necessary (often using loops to fill SmallVectors up to rank-2,
which is only two in this case), in order to more easily facilitate the
transition to a helper function.
convolution with [time,batch,channel] ordering, as opposed to the
default [batch, channel, time]. Currently implementing by transposing
the input and output, but may need to get its own implementation in the
future because this is supposed to be an op that gives a speedup. This
is used by fairseq
(https://github.com/facebookresearch/fairseq/issues/172).
(in case you were wondering like me, this is different from transposed
convolution. Transposed convolution has fractional strides).
---------
Co-authored-by: Xida Ren <xida.ren.dev@gmail.com>
Co-authored-by: Frederik Harwath <frederik.harwath@amd.com>
Currently transposed convolution is not handled correctly by
`TorchToTosa`. This PR allows transposed convolutions to pass through
the conversion so that they can be handled by other conversion passes
later in a pipeline.
An example input which produces a compilation error is:
```
func.func @forward(%input: !torch.vtensor<[1,64,1,100],f32>) -> !torch.vtensor<[1,64,2,200],f32> {
%true = torch.constant.bool true
%int1 = torch.constant.int 1
%int2 = torch.constant.int 2
%weight = torch.vtensor.literal(dense<0.0> : tensor<64x64x3x3xf32>) : !torch.vtensor<[64,64,3,3],f32>
%bias = torch.vtensor.literal(dense<0.0> : tensor<64xf32>) : !torch.vtensor<[64],f32>
%stride = torch.prim.ListConstruct %int2, %int2 : (!torch.int, !torch.int) -> !torch.list<int>
%int1x1 = torch.prim.ListConstruct %int1, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
%output = torch.aten.convolution %input, %weight, %bias, %stride, %int1x1, %int1x1, %true, %int1x1, %int1 : !torch.vtensor<[1,64,1,100],f32>, !torch.vtensor<[64,64,3,3],f32>, !torch.vtensor<[64],f32>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int -> !torch.vtensor<[1,64,2,200],f32>
return %output : !torch.vtensor<[1,64,2,200],f32>
}
```
This MLIR produces an error about a cast operation with a size mismatch
when passed through `torch-to-tosa`:
```
error: 'tensor.cast' op operand type 'tensor<1x64x1x50xf32>' and result type 'tensor<1x64x2x200xf32>' are cast incompatible
```
---------
Co-authored-by: Srinath Avadhanula <srinath.avadhanula@getcruise.com>
Introduced in 704cfdaf08 of @wu-s-john
g++ compiler error:
Pooling.cpp:177:13: error: explicit specialization in non-namespace
scope ‘class
Design looks good, g++ is just freaking out for no good reason.
Un-nesting the template classes fixes the error.
We don't have g++ CI. This hopefully happens infrequently enough that we
can just fix manually. My service to those folks who really like
building with g++... :)
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.
We can map the per_tensor case to the `torch.aten.quantize_per_linear`
operation. In this case we extract the `scale` and `zeropoint` values
and directly invoke the quantization, then return the integer
representation value.
Implemented ONNX.Range. The spec says the data type for start, limit,
delta are 0-D can be double, float, int16, int32, int64, All int types
mapped to !torch.int and all float types mapped to !torch.float
---------
Co-authored-by: Kumar Deepak <kumar@xilinx.com>
Handles the multiple cases of `onnx` constant values and converts them
to `torch` literal tensors. This can include splats with a single
integer or floating point value, a set of explicit integer values, or
an elements array attr of values.
Handle both `torch.dequantize` and `torch.quantize_per_tensor` including
the op based quantization parameter tracking. This includes adding
`qint32` to torch types as it was missing during the initial type
inclusion.
For testing we only have `torch.int8` and `torch.float` types on
function boundaries as the `qint8` types require passing the scale
and zero point quantization information which is not supported yet.
This PR updates the torch-to-tosa conversion with following changes:
- Support torch.none as min/max input argument for tosa.clamp op
- Support negative value as start index for tosa.slice op
- Add tosa.logical_or lowering support
e2e test:
python -m e2e_testing.main --config=tosa
LIT tests:
cmake --build build --target tools/torch-mlir/all
---------
Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
Adaptive pooling ops can only be decomposed into their non-adaptive
counterparts in trivial cases.
For example, the current decomposition for AtenAdaptiveAvgPool1dOp in
DecomposeComplexOps.cpp supports outSize = inSize (i.e., do literally
nothing), and outSize = 1 (i.e., do a batched average).
The reason adaptive pooling ops are difficult to lower to linalg is that
they are not constantly strided. They are computed by taking an input
tensor of shape (N, C, Hin), and an output size Hout, and computing the
output tensor at position (n,c, h) in the following way:
1. compute st(h) = (h*Hin)//Hout
2. compute en(h) = 1 + ((h+1)*Hin -1)//Hout
3. apply a computation (max or avg) to the slice: INPUT[n, c,
st(h):en(h)]
The provided sample implementation (for ConvertAtenAdaptiveAvgPool1dOp)
uses tensor.extract to access the input tensor inside the payload of a
linalg generic op. This is likely an unattractive use of linalg generic
ops, which is why I am asking for some more targeted feedback on the
validity of this approach before attempting to support the many other
adaptive pooling ops.
Specifically:
- Is the performance of this implementation bad enough to warrant
targeting different dialects entirely? e.g. TMtensor/linalg ext/ etc.
- If the provided implementation is of acceptable performance to the
community, then is it permissable to remove the Adaptive pooling
decompositions from DecomposeComplexOps.cpp? Based on the current
structure of the -torch-decompose-complex-ops pass, it does not seem
possible to only decompose the adaptive ops in special cases (it seems
to get stuck in an infinite loop on a match failure). I would be happy
to instead incorporate the case logic into the conversion directly, and
remove the decompositions once they are rendered completely obsolete.
As long as this approach is acceptable, I can clean up the
implementation with some helper functions, and quickly add support for
each of the remaining Adaptive pooling ops.
Adds a lowering to Linalg for reflection_pad1d. Based on ideas/code from draft PR
https://github.com/llvm/torch-mlir/pull/2693.
---------
Co-authored-by: Kumar Deepak <kumar@xilinx.com>
The expression for HardSigmoid in Onnx
(https://onnx.ai/onnx/operators/onnx__HardSigmoid.html): max(0, min(1,
alpha * x + beta))
is inherently different from HardSigmoid in Torch
(https://pytorch.org/docs/stable/generated/torch.nn.Hardsigmoid.html)
which is: if x < -3 -> 0
elif x > 3 -> 1
else x/6 + 1/2
That being said, it was just better to compute out the entire expression
when translating the Onnx expression to Torch mlir, which is done in
this PR. Some of the logic is shared from the files in
`DecomposeComplexOps`. Therefore, refactored some shared logic between
`DecomposeComplexOps` and `DefaultDomainGToP` and put it in a `Utils`
file.
The three remaining compare operations
onnx.Greater
onnx.Less
onnx.GreaterOrEqual
Are also added with this push request.
This concludes a set of basic tensor compare functions.
Lowerings for `transpose` from ONNX to `aten`. Implementation depends on
making multiple `aten.transpose` operations swapping pairs of dimensions.
As `onnx.transpose` can swap around any dimensions it may require
constructing multiple `aten.transpose`.
This replaces the lowering of aten.cat with tensor.concat, allowing more
efficient handling of concatenations in downstream flows. The refbackend
populates concat decomposition patterns that can be used to recover the
previous lowering.
This commit adds the OnnxToTorch support for Reciprocal, Round,
ScatterElements, Sigmoid, Sin, Tanh, Sqrt, Sub, Sum, Where, Xor,
Squeeze, Unsqueeze ops.
For reviewers, the ops that weren't trivial and probably require extra
review are Sum, Squeeze, and Unsqueeze.
Lowerings for `selu` lowerings for ONNX to the corresponding torch
implementations. Torch's `selu` implementation has fewer features so
we use the a generalized `elu` with the input scale set to `1.0`.
This commit adds the OnnxToTorch support for BitwiseXor, BitwiseOr, Div, Equal, Cast,
Ceil, Floor, Cos, and Clip op.
This commit also adds the TorchToLinalg support for aten.clamp.Tensor and aten.clamp_min.Tensor op.
Signed-Off By: vivekkhandelwal1424@gmail.com
The linalg Op `linalg.conv_2d_ngchw_fgchw` had a bug where
1. Weights were accessed as G,F,C,H,W instead of as F,G,C,H,W
2. Output was accessed as N,F,G,H,W instead of as N,G,F,H,W
Now this has been fixed in
https://github.com/llvm/llvm-project/pull/73855 which broke the
torch-mlir lowering to that Op.
This patch switches lowering in torch-mlir to the newly introduced
`linalg.conv_2d_ngchw_gfchw` op which accesses weights in an order that
is compatible with PyTorch's memory layout.
Fix https://github.com/llvm/torch-mlir/issues/2622
llvm-project: bbd2b08b95fe76bea138c1b03c1cd42ed3ee04df
stablehlo: ab709fe48de88c67717abfbd7ef17425eb95ddaf
These commits were chosen in order to account for an MLIR API break from
3dbac2c007
which required a patch to stablehlo. We integrate a bit beyond that
commit to deal with some revert/reapply cycles in the intervening range
which were discovered in another downstream.
Further, it requires adaptation to the stablehlo API breaks introduced
from https://github.com/openxla/stablehlo/pull/1872 which are along for
the ride.
Since some stablehlo builders were changed to directly take int64_t
array refs, also traced that up some call stacks to eliminate some
signed/unsigned mismatches that result.
Also adds a few TOSA tests to the passing set that seem to work now.
Despite aten.mm requiring the input and output types match, we still opt
to maintain signedness semantics in case later passes try to do any sort
of integer type narrowing.
The function `getTypeForScalarType` currently takes an argument to
specify the signedness of integer types. This is leakage of backend
specific requirements into the torch dialect world. Because
`getTypeForScalarType` is a utility function for the torch dialect, it
should only produce types that match the sign conventions used by
PyTorch (regular integers are signed and unsigned integers are
unsigned).
This commit removes the signedness argument from
`getTypeForScalarType`, and moves the backend specific handling of
integer types to the backend code.
This commit adds the OnnxToTorch support for Atan, Bitshift, BitwiseAnd,
and BitwiseNot op.
This commit also adds the TorchToLinalg support for AtenBitwiseLeftShiftTensorOp.
Signed-Off By: vivekkhandelwal@nod-labs.com
Adds a pipeline to convert custom ops and metadata represented as
`torch.operator` custom ops to corresponding `torch` ops where possible.
This is part of a multi-part approach for building ONNX import in as a
regular feature of torch-mlir. It is focused on the conversions vs the
infra. We will end up maintaining a [pure-python
importer](https://github.com/nod-ai/SHARK-Turbine/blob/main/python/shark_turbine/importers/onnx_importer.py)
to go with this in torch-mlir, and we will also maintain test case
generation utilities derived from it.
I have left substantial documentation in the README of the conversion
directory, including the recommended approach that we will take to keep
building this out.
(note that this organizes the code to coincide with the refactoring in
#2442 versus the current flat arrangement)
Rob Suderman
aldesilv
Quinn Dawkins
Aart Bik
Vivek Khandelwal
Phaneesh Barwaria
lonely eagle
Gaurav Shukla
zjgarvey
Xida Ren (Cedar)
Chi_Liu
James Newling
Ramiro Leal-Cavazos
Srinath Avadhanula
Franz Haniel
Dave Liddell
John Wu
Ilija Kalinić
Andreas Falkenberg
Ze Zhang
kumardeepakamd
lisaliu1
Han-Chung Wang
Frederik Harwath
Ben Vanik
saienduri
Stella Laurenzo
Sungsoon Cho
Gaurav Shukla
Felix Schneider
Frederik Harwath
Quinn Dawkins