stephenswat
commented
7 months ago Agreed the naming could be improved; I see a few functions that we could implement which would need a name:
- Implicitly insert $n$ elements, not calling their constructors for performance reasons, but the objects are well-formed because they are implicit lifetime objects.
- Implicitly insert $n$ elements, not calling their constructors for performance reasons, in an unsafe way because they are not implicit lifetime objects.
- Insert $n$ elements, default-constructing each of them.
- Insert $n$ elements, copy-constructing them from a given value.
The first two are implemented, I agree the second two could also be useful; but they need names.
krasznaa
A very common pattern in massively parallel computing is to reserve space for an entire thread block at once and to insert those values later. This commit adds a
reserveandreserve_unsafemethod todevice_vector, which allows clients to do this.The semantics of
reserveare fully compliant with the C++ standard starting from C++20; C++ standards below that will work just fine. Any program which calls this function and compiles should be well formed, as defined in P0593R6 and P2590R2.The
reserve_unsafemethod does the same, but technically leaves the program in an ill-defined state. Note that this form of ill-defined memory has been used for donkeys years and is actually safe. It's just technically not standards compliant.