Shuffle Tips and Tricks Slides GPU Tech Conf 2013 (S3174) 计算机科学

Shuffle: Tips and Tricks Julien Demouth, NVIDIAhttp://www.gputechconf.com/page/home.htmlGlossary Warp— Implicitly synchronized group of threads (32 on current HW) Warp ID (warpid)— Identifier of the warp in a block: threadIdx.x / 32 Lane ID (laneid)— Coordinate of the thread in a warp: threadIdx.x % 32— Special register (available from PTX): %laneidShuffle (SHFL) Instruction to exchange data in a warp Threads can “read” other threads’ registers No shared memory is needed It is available