Visualizing how advanced GPUs merge threads post-divergence to save cycles.
Set which path each thread takes after divergent Block A. Green means Path B, Purple means Path C.
Baseline SIMT: When warps diverge, they execute each taken path serially. Partially empty warps waste clock cycles.
Dynamic Warp Formation (DWF): Threads from different warps at the same PC are merged into new full warps, dramatically improving throughput and saving execution cycles.