Previous | Next --- Slide 26 of 74
Back to Lecture Thumbnails

CUDA is a data parallel programming model and uses shared address space (e.g. input can be shared across multiple CUDA threads). It is not a message passing model because we do not need to send explicit messages among threads. I believe CUDA threads can communicate with each other via load/store to shared/global memory on the GPU.

CUDA threads are like a ISPC instances because multiple threads can be gathered together for SIMT execution that runs in lock step for each instruction among all threads that are not masked off. The analogy to pthreads is less obvious because pthreads execute instructions at independent points among its hardware threads.


Also just to add, the blocks in CUDA resemble the tasks in ISPC

Please log in to leave a comment.