Add reference implementation for parallel_block feature #1570
Conversation
Signed-off-by: pavelkumbrasev <[email protected]>
Signed-off-by: pavelkumbrasev <[email protected]>
Signed-off-by: pavelkumbrasev <[email protected]>
Signed-off-by: Isaev, Ilya <[email protected]>
Signed-off-by: Isaev, Ilya <[email protected]>
…rallel_block Signed-off-by: Isaev, Ilya <[email protected]>
Signed-off-by: Isaev, Ilya <[email protected]>
isaevil
changed the title
|
@akukanov @aleksei-fedotov @vossmjp Could you please take a look at the PR in terms of implementation and ABI. P.S. Tests are still WIP. |
Signed-off-by: Isaev, Ilya <[email protected]>
Signed-off-by: Isaev, Ilya <[email protected]>
| if (is_worker_should_leave(slot)) { | ||
| if (!governor::hybrid_cpu()) { | ||
| if ( | ||
| #if __TBB_PREVIEW_PARALLEL_BLOCK | ||
| !my_arena.my_thread_leave.should_leave() | ||
| #else | ||
| !governor::hybrid_cpu() | ||
| #endif | ||
| ) |
There was a problem hiding this comment.
A readability issue. These should_leave functions do not convey any additional meaning except for whether thread should leave or not. In this case I would move all the logic into single main should_leave function, which would include a set of probes with more meaningful names the "leaving" algorithm is based on, such as below "my_arena.my_threading_control->is_any_other_client_active()" or "is_recall_requested". In my opinion, the general "should_leave" name is good for the most embracing function of such kind, not a set of smaller functions that check various parts of the system.
|
|
||
| #if __TBB_PREVIEW_PARALLEL_BLOCK | ||
| TBB_EXPORT void __TBB_EXPORTED_FUNC register_parallel_block(d1::task_arena_base&); | ||
| TBB_EXPORT void __TBB_EXPORTED_FUNC unregister_parallel_block(d1::task_arena_base&, bool); |
There was a problem hiding this comment.
Do we anticipate any future extension? In other words, does it make sense to accept an untyped pointer for any future details?
| TBB_EXPORT void __TBB_EXPORTED_FUNC unregister_parallel_block(d1::task_arena_base&, bool); | |
| TBB_EXPORT void __TBB_EXPORTED_FUNC unregister_parallel_block(d1::task_arena_base&, void* /*supplemental details*/); |
Also, I don't think single bit can be passed. At least a word (or even more?) is used anyway. Then, it makes sense to accept at least that width and do not kill opportunity for future extensions:
| TBB_EXPORT void __TBB_EXPORTED_FUNC unregister_parallel_block(d1::task_arena_base&, bool); | |
| TBB_EXPORT void __TBB_EXPORTED_FUNC unregister_parallel_block(d1::task_arena_base&, unsigned short/*supplemental details*/); |
Please double check what size of information is passed.
There was a problem hiding this comment.
Yeah, I also have been thinking about that. Perhaps we should pass std::uintptr_t instead of void*?
Signed-off-by: Isaev, Ilya <[email protected]>
Co-authored-by: Aleksei Fedotov <[email protected]>
Description
Add a comprehensive description of proposed changes
Fixes # - issue number(s) if exists
Type of change
Choose one or multiple, leave empty if none of the other choices apply
Add a respective label(s) to PR if you have permissions
Tests
Documentation
Breaks backward compatibility
Notify the following users
List users with
@to send notificationsOther information