Evaluate complete-model training and matched cuFFTDx performance

Determine the effects of Flash-CQT on complete model training and compare its performance with that of a matched cuFFTDx implementation.

Background

The paper evaluates Flash-CQT through a differentiable round trip involving analysis, synthesis, a squared-error loss, and backpropagation, reporting speedups over a CQT baseline. However, these measurements do not assess complete neural-audio-model training, nor do they compare Flash-CQT against an implementation based on cuFFTDx. The authors explicitly leave both evaluations unresolved.

References

Effects on complete model training and comparison with a matched cuFFTDx implementation remain open.

— Exact Factorisation and Fast Computation of Invertible Constant-Q Transforms  (2609.29119 - Franchino et al., 24 Sep 2026) in Section 5.2, “Differentiable round trip”