[Deps] Update qonnx to track https://github.com/iksnagreb/qonnx

[iksnagreb]

• Make gen_finn_dt_tensor consider the numpy type for INT and FIXED types fastmachinelearning/qonnx#118
• [InferDataLayouts] Add fallback for layout propagation fastmachinelearning/qonnx#145
• Run and fix FINN unit tests...
• Run and fix Transformer builds...

Discuss whether to include the following PRs here as well...

• [Opset] Increase preferred opset version to 12 fastmachinelearning/qonnx#140
• [MultiThreshold] Generalize data layouts for node execution fastmachinelearning/qonnx#143
• [Transform] Introduce ComposedTransformation fastmachinelearning/qonnx#169

[fpjentzsch]

I get the following for a Transformer build that previously worked on iksnagreb/qonnx/transformer (+ 2 unrelated commits cherry-picked from main):

Traceback (most recent call last):
1128 File "/usr/lib/python3.10/multiprocessing/pool.py", line 125, in worker
1129 result = (True, func(*args, **kwds))
1130 File "/usr/lib/python3.10/multiprocessing/pool.py", line 48, in mapstar
1131 return list(map(*args))
1132 File "/dev/shm/finn/src/finn/transformation/fpgadataflow/prepare_cppsim.py", line 84, in prepareCppSim_node
1133 _codegen_single_node(node, self.model)
1134 File "/dev/shm/finn/src/finn/transformation/fpgadataflow/prepare_cppsim.py", line 57, in _codegen_single_node
1135 inst.code_generation_cppsim(model)
1136 File "/dev/shm/finn/src/finn/custom_op/fpgadataflow/hlsbackend.py", line 203, in code_generation_cppsim
1137 self.defines("cppsim")
1138 File "/dev/shm/finn/src/finn/custom_op/fpgadataflow/hls/elementwise_binary_hls.py", line 233, in defines
1139 f"using OutType = {self.out_dtype.get_hls_datatype_str()};",
1140 File "/dev/shm/finn/deps/qonnx/src/qonnx/core/datatype.py", line 365, in get_hls_datatype_str
1141 raise Exception("Undefined for ScaledIntType")
1142Exception: Undefined for ScaledIntType

[iksnagreb]

Hm, as the ScaledIntType issue seems to be fixed on some of the range analysis related branches and I will need this stuff for ongoing/upcoming research anyway, I am seriously considering to merge the current state of range analysis into our qonnx dependency. The experimental analysis pass and the (not yet fully integrated) reworked streamlining should not interfere with any of our build flows (these are just optional extras for now), but somewhere among the ~130 commits must be the bugfix for whatever this ScaledIntType issue is... What do you think?

[fpjentzsch]

I still get the ScaledIntType error if I switch to this QONNX commit: iksnagreb/qonnx@31880be

This is the old commit where synthesis and verification still passes: iksnagreb/qonnx@1888fab

[iksnagreb]

Could someone (maybe @fpjentzsch, you have the infrastructure for testing) do some non-trivial (i.e., not just some one layer MLP...) build with opset version 12 (or maybe even higher, to see how far we could push this)? Then we could decide to merge or close the one remaining PR here...

[fpjentzsch]

We can run the test suite afterwards in combination with the other Transformer-related PRs. Feel free to create a separate PR for the opset version 12 if needed.