Ilmu Komputer & AI editorial
Open AccessOA2026
Memory-efficient GPU pipelines for real-time non-line-of-sight reconstruction
Fused kernels, analytic ring kernels, and FP16 storage deliver up to 42× speedup and 2.5% memory footprint for wave-based NLOS imaging
Alfonso López-Ruiz; Diego Royo· 2026· DOI 10.48550/arXiv.2608.28183
The core problem
Non-line-of-sight (NLOS) imaging recovers scenes hidden around a corner by analyzing indirect light captured with a single-photon avalanche diode (SPAD). Each reconstruction is a large inverse problem: billions of photon timestamps must be binned, moved through memory, transformed, and inverted. As SPAD arrays increase acquisition throughput, reconstruction has become the limiting stage. This work rebuilds the GPU execution of two established wave-based algorithms—f-k migration and phasor-fields—for both streaming and offline processing. The central question is not algorithmic accuracy but execution efficiency: how to reorganize memory movement, kernel fusion, and precision to keep pace with modern SPAD sensors. The authors report up to 42× faster streaming reconstruction than the reference pipeline and up to 14× faster than the fastest published GPU baseline, while reducing memory use to as little as 2.5% of prior requirements. The result enables vastly larger and finer reconstructions on the same hardware, or comparable reconstructions within a much lower memory budget. The paper also proposes three denoising strategies enabled by the resulting frame budget for next-generation NL
Innovation
The rebuilt pipelines achieve up to 42× faster streaming reconstruction than the reference streaming pipeline and up to 14× faster than the fastest published GPU baseline. Memory use drops to as little as 2.5% of prior requirements. These gains are consistent across both f-k migration and phasor-fields, and across streaming and offline modes. The ablation isolates the contribution of each choice: fused kernels, warp-level photon binning, batched transforms, CUDA graph replay, and selective FP16 storage. The analytic ring kernel for phasor-fields eliminates dense kernel storage at runtime, reducing both memory and bandwidth. The practical effect is that vastly larger and finer reconstructions become feasible on the same hardware, or comparable reconstructions fit within a much lower memory budget. The authors also propose three denoising strategies enabled by the resulting frame budget, pointing toward next-generation NLOS video processing.
Non-line-of-sight (NLOS) imaging recovers scenes hidden around a corner by analyzing indirect light captured with a single-photon avalanche diode (SPAD). Each reconstruction is a large inverse problem: billions of photon timestamps must be binned, moved through memory, transformed, and inverted. As SPAD arrays increase acquisition throughput, reconstruction has become the limiting stage. This work rebuilds the GPU execution of two established wave-based algorithms—f-k migration and phasor-fields—for both streaming and offline processing. The central question is not algorithmic accuracy but execution efficiency: how to reorganize memory movement, kernel fusion, and precision to keep pace with modern SPAD sensors. The authors report up to 42× faster streaming reconstruction than the reference pipeline and up to 14× faster than the fastest published GPU baseline, while reducing memory use to as little as 2.5% of prior requirements. The result enables vastly larger and finer reconstructions on the same hardware, or comparable reconstructions within a much lower memory budget. The paper also proposes three denoising strategies enabled by the resulting frame budget for next-generation NLOS video processing.
The authors reorganize the GPU pipeline of both algorithms around fused kernels, warp-level photon binning, batched transforms, CUDA graph replay, and FP16 storage applied only where it reduces the actual bottleneck. For phasor-fields, they assemble the ring-and-radius kernels once and offline using the analytic Fourier transform of a ring, so the propagation kernel never exists in dense form at runtime. This analytic construction reduces both memory and bandwidth. The pipeline can be expressed as a sequence of transformations:
Why it matters
The results reframe NLOS reconstruction as a memory-bandwidth problem rather than a compute-bound one. By fusing kernels and batching transforms, the pipeline reduces redundant global memory traffic. Warp-level photon binning keeps timestamp processing close to the data, while CUDA graph replay removes launch overhead. The analytic ring kernel is particularly effective because it replaces a dense propagation kernel with a closed-form expression, so the kernel never exists in dense form at runtime. Selective FP16 storage is applied only where it reduces the actual bottleneck, avoiding precision loss where it would not help. The 2.5% memory footprint is the most striking outcome: it means that reconstructions previously limited by GPU memory can now scale in size and resolution. The proposed denoising strategies exploit the frame budget created by faster reconstruction, suggesting a path to real-time NLOS video. Limitations include the need for careful profiling to choose FP16 placement and the assumption that the analytic ring kernel applies to the phasor-fields formulation. Overall, the work demonstrates that execution-level reorganization can deliver order-of-magnitude gains without changing the underlying algorithms.
Who should read this
CS practitioners and researchers
Opening member content…