Imaging flow cytometry using Fourier light-field microscopy enables high-throughput three-dimensional cellular imaging, capable of capturing thousands of events per second. However, volumetric reconstruction speed remains orders of magnitude slower than the acquisition speed. The current state of art uses Richardson-Lucy algorithm, restricted to just 5-10 reconstructed events per second with GPU acceleration. This limitation hinders real-time applications such as cell sorting and thus has bottlenecked the widespread adoption of 3D imaging flow cytometry. We introduce patch deconvolution, the first training-free algorithm compatible with the Richardson-Lucy framework that significantly accelerates convergence, achieving over 100-200 reconstructions per second on standard GPUs, a 20- to 40-fold improvement over Richardson-Lucy. Validated on both simulated and experimental data sets, patch deconvolution achieves reconstruction quality comparable to Richardson-Lucy in both static and flow data. This supports rapid cell sorting based on spatial features and enables advanced applications, such as detecting rare spatial events in large cell populations, which would otherwise be indistinguishable in traditional flow cytometry.