cuda stream

发现这个博客越来越像带文本搜索的我的书签了哈哈,不过这样也算合理吧,毕竟高效的搬运经过筛选的有价值的文章比上网搜到一大堆废话要好吧.

要点

1. 现在的GPU一般有3个引擎
   • 传入
   • 传出
   • 计算
   • 这3个可以并发

GPU supporting device overlap possesses the capacity to execute a kernel while performing a copy between device and host memory——————https://cs.wmich.edu/elise/courses/cs626/sp13/savitha/CudaStreams.pptx

2. stream声明的顺序影响其执行时的并发度
   • https://developer.download.nvidia.com/CUDA/training/StreamsAndConcurrencyWebinar.pdf
   • https://devblogs.nvidia.com/how-overlap-data-transfers-cuda-cc/
3. stream更多详细资料
   GTC的讲座:cuda-streams-best-practices-common-pitfalls