tag

#cuda

2 pages tagged

Wiki › Algorithms › 01. Searching and Sorting › 11. Parallel, Distributed, and GPU Sorting ›

CUDA Warp Sort

CUDA Warp Sort CUDA warp sort sorts a small group of values within one warp. A warp is a fixed group of GPU threads that execute together. On NVIDIA GPUs, a warp commonly contains 32 lanes. The algorithm is usually implemented with compare and exchange steps using warp shuffle instructions. Since all lanes execute the same instruction stream, warp sort works best with regular sorting networks such as bitonic sort...

Wiki › Algorithms › 01. Searching and Sorting › 11. Parallel, Distributed, and GPU Sorting ›

CUDA Block Sort

CUDA Block Sort CUDA block sort sorts a tile of data using all threads in a block. It extends warp level sorting to larger sizes by using shared memory and synchronization within a thread block. Each block loads a chunk of the array into shared memory, sorts it cooperatively, then writes it back. This is often used as a building block for full GPU sorting algorithms such as merge sort...