[async] Optimization stage 1 #994
Conversation
| irpass::typecheck(new_root.get()); | ||
| irpass::typecheck(new_root.get(), kernel); |
There was a problem hiding this comment.
Actually this pass is probably unnecessary...
There was a problem hiding this comment.
...since all ret_types are cloned.
There was a problem hiding this comment.
Merging this in for now and let's get rid of all kernel parameters and this type check after #988 is done.
| // We do not fuse serial kernels for now since they can be SNode accessors | ||
| bool are_both_serial = task_a->task_type == OffloadedStmt::serial && | ||
| task_b->task_type == OffloadedStmt::serial; | ||
| bool same_kernel = task_queue[i].kernel == task_queue[i + 1].kernel; |
There was a problem hiding this comment.
BTW, thanks to #915, I realized that kernel fusion has some preconditions.. For example, i don't think clear and listgen are easily fusable, because the latter depends on the fact that all threads would observe the result of clear.
While I don't think this change would fuse those kernels, I'd like to confirm that: kernel fusion would only be done if the later kernel doesn't rely on the kernel boundary being a barrier?
Related issue = #742
[Click here for the format server]