A little secret to rock your YouTube subscribers
Get Free YouTube Subscribers, Views and Likes

Advanced HIP Workshop | Optimising compute with concurrent IO | HipMemcpy3D and asynchronous copies

Follow
Pawsey Supercomputing Research Centre

This video is part of the HIP workshop playlist. View full playlist here:    • Advanced HIP Workshop  

In the concurrent examples you might notice the use of hipMemcpy3DAsync to copy wavefields as planes back to the 3D host memory allocation out_h. This is because using hipMemcpyAsync to copy device memory to an address within a pinned memory allocation failed silently and produced faulty copies. Therefore it is not recommended to perform asychronous copies with pointers derived from within a pinned memory allocation. Only use the pointer returned by hipHostMalloc with asynchronous copy functions.

posted by FiennyFubdj