Missing script to reproduce (ResNet-18) results from the paper

[agostini01]

Thank you for sharing this work with the community.

We are trying to reproduce the results for ResNet-18 presented in Table 8 of the HIDA paper (https://arxiv.org/pdf/2311.03379)

---

We have followed the example, but we are facing difficulties on generating a design that produces a correct output and/or provide similar 45 samples/s throughput as reported in the paper.

We have tried variations of:

scalehls-opt resnet18.mlir \
    -hida-pytorch-pipeline="top-func=forward loop-tile-size=8 loop-unroll-factor=4" \
    | scalehls-translate -scalehls-emit-hlscpp -emit-vitis-directives > resnet18.cpp

Could you please share a script that reproduces the results?

Thanks in advance

[dsandall]

Hey agostini,

What kind of throughput are you achieving?
Here's one data point for you to possibly compare: I just ran the same command on my build (except with loop-tile-size and loop-unroll-factor set to 1, as scalehls-opt otherwise hangs), and ran it through v++ (Vitis 2025.1) and was able to get an expected throughput of .05 samples/s on an "xcvu9p-flga2104-2-i".

Does this line up with your methodology/results? Thanks

[dsandall]

Just wanted to follow up - I was eventually able to get results that are much closer to the paper's reported figures for Resnet18, using the same Vitis HLS C++ file as earlier, but with Vitis 2022 (as stated in the author's paper...). Just because it compiles and synthesizes on the new v++ based HLS toolchain does NOT mean you will get similar results.