Is it true that one stream can submit work to multiple different HSA queues?

[ipe-zhangyz]

In an article Exploring AMD GPU scheduling details by experimenting with “worst practices” (https://link.springer.com/article/10.1007/s11241-022-09381-y) there is a statement: ROCclr’s software queues internally share a pool of HSA queues: one software queue may submit work to multiple diferent HSA queues, and each HSA queue may contain work from multiple software queues. There is no doubt that one HSA queue can receive work from multiple different streams, but according to the relative source code, it seems that with the DIRECT DISPATCH feature since rocm 4.5, one stream may only submit to one HSA queue that it is bound to. I want to know if I understand the relative code correctly. Thanks in advance!