|
|
|
|
|
|
by PixelOfDeath
1472 days ago
|
|
|
It's also about prefetchability of the next task. If all cores e.g. do an atomic add on a single counter variable to get their next task-id just in time, then there is no chance for the cores to do any reasonable prefetching. Because most of the time another core will come in and write to this variable that all prefetching is based on. With short tasks, you can end up with a full pipeline stall each time a new task is started. |
|
|