Abstract Spatial audio rendering in large‐scale urban scenes is crucial for immersive VR/AR, digital twins, and mobility‐centered simulations. Yet, accurately modeling wave propagation through dynamic media remains a significant challenge in visual computing, particularly when complex urban wind fields, non‐uniform shear flows, and voxelized geometries interact. While geometric acoustics inherently struggles to explicitly resolve low‐frequency diffraction and interference, conventional wave‐based solvers—including standard Lattice Boltzmann Methods (LBM)—often suffer from numerical dispersion or severe instability under strong convection. In this paper, we present the V‐LLBM framework, a robust volumetric wave‐rendering pipeline tailored for dynamic media. By recasting the D3Q27‐Cumulant collision operator into a graphics‐oriented acoustic context, our framework helps suppress spurious non‐physical modes through central‐moment relaxation in a co‐moving reference frame. We evaluate the solver hierarchically, progressing from canonical propagation and rigid‐boundary scattering to demanding aerodynamic stress tests involving high‐subsonic shear and broadband pulses in strong crosswinds. Across these scenarios, the cumulant formulation maintains the baseline acoustic fidelity of established reference models while providing improved robustness where standard collision models fail to maintain coherent wavefields. Furthermore, we scale our method to a 140‐million‐voxel real‐world urban canyon topology. By employing a two‐pass volumetric baking strategy, we successfully isolate delicate acoustic perturbations from orders‐of‐magnitude larger aerodynamic background pressures. This enables the extraction of wind‐aware receiver‐side temporal responses, quantitatively capturing critical physical wavefront distortions such as convective Doppler shifts and significant amplification in peak acoustic pressure. Ultimately, our approach establishes a reliable offline pre‐computation pipeline, providing physically grounded spatial audio assets and high‐fidelity data priors for downstream data‐driven representations, such as Neural Acoustic Fields (NAF).

We propose the SaSpGEMM: a parallel sparse general matrix-matrix multiplication (SpGEMM) to avoid the overhead of sorting. The typical workflow of SpGEMM contains: size prediction, memory allocation, numeric calculation and sorting. However, sorting has always been overlooked as a bottleneck in the performance of SpGEMM. It constitutes an average of 30% in HASH and 55% in ESC during the calculation stage. The key idea behind SaSpGEMM is to leverage the compressed sparse row (CSR) storage format’s feature of increasing index order of elements in the same row which is sorted during preprocessing, preserving intermediate products ordered consistently. To achieve this, we introduce a linked list-based accumulator (LLA) designed for batch insertion while maintaining order with a low time complexity. We provide a comprehensive empirical evidence showing that SaSpGEMM outperforms other methods based on time complexity analysis. Compared to three state-of-the-art methods ESC, SPA, and HASH on both x86 (Intel Xeon Gold 6348) and ARM (Phytium2000+) architectures, our method achieves an average speedup of 2.82x, 5.24x, 1.16x (with a maximum speedup of 34.4x, 195x, 23.3x) on the Intel Xeon Gold 6348. On the Phytium2000+, it achieves an average speedup of 2.21x, 4,65x, 1.05x (with a maximum speedup of 40.93x, 146.7x, 9.03x).