In music, unlike general audio data, there are characteristics according to progression such as structural information and BPM. However, in many MIR models, audio is often preprocessed under fixed information in consideration of compatibility with existing models. This work investigates BPM-aware and structure-based segmentation for music genre classification. We compare three approaches: 30-second whole track, fixed 5-second segments, and musically-informed 2-bar segments on GTZAN dataset. Our experiments demonstrate the feasibility and trade-offs of structure-aware preprocessing. The codes are available online.