Preserve onnx tensor dtype when inserting FIFOs

Xilinx / finn

Dataflow compiler for QNN inference on FPGAs

https://xilinx.github.io/finn

BSD 3-Clause "New" or "Revised" License

753 stars 243 forks source link

Preserve onnx tensor dtype when inserting FIFOs #1192

Closed auphelia closed 2 months ago

auphelia commented 2 months ago

Before, the onnx tensor data type wasn't preserved and defaulted to float. While the FINN dtype was preserved (so no errors in the resulting hw), this could lead to errors for the executability of an onnx graph after the InsertFIFO transformation, see https://github.com/Xilinx/finn/issues/1113

This PR contains changes to keep the onnx tensor dtype when inserting FIFOs.