GPUPrefixSums: A portable collection of GPU prefix sum algorithms for CUDA, D3D12, Unity, and WGPU

Summary

GPUPrefixSums is an open-source GitHub project that implements a comprehensive collection of prefix sum algorithms for GPUs, supporting CUDA, D3D12, Unity, and WGPU. It aims to make state-of-the-art GPU prefix sum techniques portable across different compute shader environments. The project introduces a novel technique called "Decoupled Fallback" for Chained Scan with Decoupled Lookback, designed to enable devices without forward thread progress guarantees to perform scans without crashing. The D3D12 implementation includes an extensive survey of GPU prefix sums from warp to device level, all utilizing wave/warp/subgroup level parallelism.