Zoom Logo

HPCSE II, Spring Semester 2021 - Shared screen with speaker view
Marie-Louise Achart
34:54
what about the numthreads ? Should we write myKernel<<<grid,numthreads,shmem,stream>>>(param) ?
Fabian Wermelinger
35:08
yes
Fabian Wermelinger
38:05
This is the general kernel launch syntax: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#device-side-kernel-launch
Marie-Louise Achart
44:08
I don’t understand the first exception : most kernel launches in the default stream. could you say few more words on it ?
Marie-Louise Achart
44:30
How is this asynchronous ?
Fabian Wermelinger
01:36:59
either global memory or atomics. There are higher-level primitives for launch-grid level synchronization called cooperative groups. We have not talked about them in class
Fabian Wermelinger
01:39:44
See this appendix for cooperative groups: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#cooperative-groups
Fabian Wermelinger
01:40:13
and this for grid synchronization in particular: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#grid-synchronization-cg
Severin Fritschi
01:42:52
via shared memory?
Marie-Louise Achart
01:54:54
Maybe few minutes on this butterfly reduction ? :)
Yannick Schubert
01:59:58
5
Marie-Louise Achart
02:03:44
Thank you