Similar to the (old) addEventIfNecessary we now allow to add an event
and wait for it if that is necessary. This allows a thread to put in
a wait before the event was created yet. Since this is not yet used this
patch won't change anything.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
openmp/libomptarget/include/device.h | ||
---|---|---|
141 | I don't understand what the problem is. Could you say what about it is confusing, maybe suggest an alternative wording or at least describe how that could look like? | |
openmp/libomptarget/src/device.cpp | ||
46 | I do not understand what you mean here either. | |
62 | Is it? I looked at the CUDA API manual and I did not get that idea. Do you happen to have the passage that says so or is it from experimentation? |
openmp/libomptarget/include/device.h | ||
---|---|---|
141 | recordEventIfNecessary. It sounds like under certain circumstance defined as necessary, an event will be recorded. Otherwise, it is no-op. What is "necessary"? It can be much cleaner to have 3 APIs. I also don't get why waitEventIfNecessary is a good API. Should waitEvent just do no-op when there was no event rather than create one? | |
openmp/libomptarget/src/device.cpp | ||
62 | I found the concrete word. It is safe. |
Seems to controversial for now. Will revisit if it comes up later.
openmp/libomptarget/include/device.h | ||
---|---|---|
141 |
While there is some truth to this, you simply ignored the "necessary" part in your new "design". |
both recordEventIfNecessary and waitEventIfNecessary are confusing names "IfNecessary" part.