Missing optimization of DequantizeLinear ∘ Flatten ∘ QuantizeLinear?

[mcollinswisc]

It looks like ONNXRuntime will optimize DequantizeLinear ∘ Reshape ∘ QuantizeLinear to only the Reshape, eliminating the quantization/de-quantization, if the scales & zero points are the same.

However, an equivalent Flatten is not optimized. Is this likely to be just a missing optimization, or is there some reason the qdq would be preserved in this case?

Tested out in:
https://gist.github.com/mcollinswisc/d1cd9d13b4e5fbad01c75dca5c9ca576
with ONNXRuntime 1.18.0

[mcollinswisc]

Perhaps a matter of including Flatten here:

onnxruntime/onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc

Line 62 in 4743803

std::unique_ptr<NodeSelector> selector = std::make_unique<QDQ::DropQDQNodesSelector>(true);

[mcollinswisc]

Other operators that may make sense to include in that list:

• Expand
• Tile
• Slice
• Split [Uncertain if this selector can handle multiple outputs]
• GatherElements
• ScatterElements [Multiple inputs]
• DepthToSpace
• SpaceToDepth

The selector that excludes 16 bit also optimizes around MaxPool:

onnxruntime/onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_selector_action_transformer.cc

Line 57 in 4743803

{{"MaxPool", {12}},

This perhaps assumes that the quantization scale is strictly positive? (Or maybe this is explicitly checked somewhere...) By the same assumption, I'd guess the following can also be optimized:

• Max
• Min
• ReduceMax
• ReduceMin

[mcollinswisc]

Addressing Q on whether the quantization scale is positive with #21182

[mcollinswisc]

DepthToSpace and SpaceToDepth can't be included because there's no integer implementations:
#21287