Warp Shuffles and Reduction and Scan Operations CUDA Slides 计算机科学

Lecture 4: warp shuffles, and reduction / scan operationsProf. Mike Gilesmike.giles@maths.ox.ac.ukOxford University Mathematical InstituteOxford e-Research CentreLecture 4 – p. 1Warp shufflesThe Kepler architecture introduced a new machine instructon: a warp shuffleThis gives a mechanism for moving data between threads in the same warp, without using any shared memory.At present it is only for 32-bit data, but 64-bit data can be handled (in software) as a pair of 32-bit shuffles.Lecture 4 – p. 2