This patch adds basic support for lowering graphs with per-channel
quantization. Per-channel quantized ops have to be excluded from
`FuseQuantizedOps` for now but can be used in QDQ quantized form.
Using this patch, we're able to import and execute (on the linalg
backend) graphs with per-channel quantization applied using the "new"
PyTorch 2.0 Export Quantization.
The static uneven divisible AdaptiveAvgPool2d means that although the
input size is not an integer multiple of ouput size, but the kernel and
stride size can also be fixed (not dynamic). The derivation logic of
kernel and stride size is consistent with
torch/_decomp/decomposations.py:adaptive_avg_pool2d as described in the
following:
1. Stride Size
Firstly , derive the start index in each reduce operation according to
the output size (`n`), `start_index = ([0, 1, ..., n - 1] * input_size)
// output_size`. For each index `k`, if `k * (input_size % output_size)
< output_size`, then the current and previous stride keeps the same as
`input_size // output_size`. So suppose `(n-1) * (input_size %
output_size) < output_size`, the stride in the whole AdaptiveAvgPool2d
process keeps static, as `input_size // output_size`.
2. Kernel Size
torch/_decomp/decomposations.py:adaptive_avg_pool2d calculates a static
kernel size when the input/output sizes satisfy either of the two
conditions, `input_size % output_size == 0` or `output_size %
(input_size % output_size) == 0`. Here if `input_size % output_size ==
0`, then the kernel size equals `input_size // output_size`, otherwise
`input_size // output_size + 1.`
This patch adds a few misc pad op related changes:
1. Addresses issue <https://github.com/llvm/torch-mlir/issues/3457>
2. Addresses issue <https://github.com/llvm/torch-mlir/issues/3442>
3. Fixes the padding order for asymmetrically padded onnx.Conv ops
4. Enables passing quantization through those onnx.Conv op pre-paddings
5. Modifies the torch-to-linalg lowering of AtenReplicationPad2d op to
enable support for input rank != 4
Unfortunately, even with all of these changes, the e2e tests for the
ReplicationPad2d still fail the onnx config, since the torch export
procedure for rearranging the pad order is complicated enough that the
padding ints end up not being able to fold back to constants.
Register `aten.fake_quantize_per_channel_affine` and
`aten.fake_quantize_per_tensor_affine.tensor_qparams` ops
---------
Co-authored-by: Ze Zhang <ze.zhang@getcruise.com>
Due to the custom operation parser, the print and parser were expecting
two different forms.
One having the dictionary before the value and the other after.
Following the format of the other constants ops, the constant.int will
follow the `value attr-dict` format. Updated the parser accordingly.
In one of our downstreams, we encountered an internal assertion failure
in an intermediate pass from `AtenTensorOp::fold` invocation:
```
external/llvm-project/llvm/include/llvm/Support/Casting.h:650: decltype(auto) llvm::dyn_cast(const From &) [To = mlir::torch::Torch::NonValueTensorType, From = mlir::Type]: Assertion `detail::isPresent(Val) && "dyn_cast on a non-existent value"' failed.
```
for this snippet in the IR:
```
%arg1: !torch.tensor {torch.type_bound = !torch.vtensor<[1,1,15360],f32>}
...
%218 = torch.aten.size %arg1 : !torch.tensor -> !torch.list<int>
%219 = torch.aten.tensor %218, %none, %none, %false : !torch.list<int>, !torch.none, !torch.none, !torch.bool -> !torch.tensor
```
Turns out this was
[fixed](https://github.com/llvm/torch-mlir/pull/3189/files#diff-dc8ed165c207918e606490eee3984b1ad51d7034e6aac36fc046bf47f6f03f4fR3719)
eventually (and we were on an old hash of torch-mlir). This PR submits
just the lit test for test coverage on that specific change:
```c++
OpFoldResult AtenTensorOp::fold(FoldAdaptor adaptor) {
auto resultTy = dyn_cast<ValueTensorType>(getType());
// lit test this
if (!resultTy || !resultTy.hasSizes() || !resultTy.hasDtype())
return nullptr;
...
```
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.
This PR fixes the bugs for `Torch::AtenOneHotOp` by:
1) Using `Torch::kUnknownSize` as the default value for `numClasses` in
the pattern matching stage in `DecomposeAtenOneHotOp`
2) Adding `AtenIntScalarOp` to the patterns in `TorchToArith`
3) Handling both `int` and `float` types for `off` and `on` values in
`TorchOnnxToTorch` conversion
It also includes:
1) A new test in `TorchToArith/basic.mlir`, for `torch.aten.Int.Scalar`,
and
2) A new test in `decompose-complex-ops.mlir`, for `torch.aten.one_hot`
**Dependencies**
This PR is dependent on #3334.
This change enables more customization with operand quantization, and
generalizes the patterns QuantizeOperands and QuantizeTransposeOperands
to QuantizeOperandsPastCommutingOps.
This allows for passing quantization through operations which are
functionally unaffected by quantization, such as view-like ops. The
purpose of this change is to address a myriad of quantization issues
seen in quantized onnx models that have some reshape-like operations
sandwiched in between a dequant and something like a matmul (whose other
operand is immediately quantizable).
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 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.
Previously, it could only handle the situations where outputsize == (1,
1) or outputsize == (input_H, input_W). Now it supports all situations
where input_H % output_H== 0 && input_W % output_W == 0
Reshaping tensors depend on directly matching individual dimensions to
their corresponding dim in the `torch.view` reshape dimensions. This
involves decoupling dynamic dimensions from their static counterparts
and support cleanup / canonicalization.
This folds small version of the tensor-scalar comparison operators as
they are commonly used for shape computations. This includes le, lt, ge,
gt, eq, and ne.
A handful of operations are commonly used in shape calculations (slice,
concat, broadcast). Added these additional folders to better propagate
simple shape computations.
A bunch of small fixes are interlinked and trigger crashes if not
addressed as a group. This includes:
- aten view when expand from a rank-0 tensor
- slice folder with negative indices
- `aten._shape_as_tensor` folder on a rank-0 tensor
- `aten.cat` of a tensor with a length-0 tensor
Simple folder for limited size aten tensor operations. This is primarily
useful for shape computation folding as they unfortunately can use
`aten` operators. Add, sub, mul are common examples of these folders.
Some operations include a backend matcher for specialized operations. We
map these back to generics so they appropriately match to the high
performance versions. This is done for the attention operation.
Folds aten::index_select ops under the following conditions:
1. If the input and output are the same shape, the indexing operation is
a NOP, so just return the input.
2. If the input has shape <1x1x...xNx...x1> (all 1's except for one
dim), and the output shape is <1x1x...x1> (all 1's), then there is a
single index, so extract the single element value and return a tensor
with that value.
---------
Co-authored-by: Dave Liddell <dliddell@xilinx.com>
If a tensor is initialized by a list with a single constant integer,
this folder turns it into a torch.vtensor.literal
---------
Co-authored-by: Dave Liddell <dliddell@xilinx.com>
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`.
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>
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.
We were seeing some assertion failures after some checks around folders
were tightened up in LLVM:
https://github.com/llvm/llvm-project/pull/75887 . This PR essentially
moves the logic that used to be applied at the LLVM level into the
folder, which seems to be the suggested fix.
I'm not sure if the IR that caused issues for us _should_ be valid?
```
%1 = torch.aten.detach %arg0 : !torch.tensor<[1],f32> -> !torch.tensor
```
A better fix might be to create a verifier ensuring the result of
`aten.detach` has the same type as its operand.
---------
Co-authored-by: aaron-stgeorge <aaron.stgeorge@getcruise.com>
yyp0
Felix Schneider
Rob Suderman
Yuanqiang Liu
Ze Zhang
zjgarvey
Christopher McGirr
Sambhav Jain
Peiming Liu
Angel Zhang
Aaron St George
Vivek Khandelwal
Xida Ren (Cedar)
Stella Laurenzo
Xinyu Yang
Rob Suderman
Dave Liddell
Xida Ren (Cedar)
Aart Bik