Summary: 42 instances, 29 unique

Text	Count
// TODO Handle multiple tensors.	1
// FIXME Instead of hardcoding the next three values, we could use	2
// FIXME In principle we could just exclude the GPUs that violate this check	1
// FIXME Instead of throwing away the error and setting a bool, we should have	2
// FIXME Avoid this anonymous namespace and use inline variables in C++-17.	2
// FIXME In C++-17 perhaps a global static inline variable would be better?	10
// FIXME Consider using a (ordered) map, because keys are IDs which are	1
// FIXME O_TMPFILE is also not that old, and some users have reported issues due	1
// FIXME This may actually not be true, as channels could for example be	1
// FIXME Escape the contexts' domain descriptors in case they contain a colon?	1
// TODO: If we can guarantee the order in which the accept() calls happen,	1
// TODO: Lazy initialization of cuda loop.	1
// FIXME Can we assume that the two endpoints will see the same list of LSMs,	1
# TODO: Default to ON if CUDA available.	1
// FIXME We'd want this to return a std::vector<CudaEvent>, but CudaEvents	1
// FIXME Rename this to OnDemandDeferredExecutor?	1
// FIXME: Once we go C++17, use std::launder on the returned pointer.	2
// FIXME It would be nice if this could be done by the thread itself just	1
// FIXME Once we've fixed the viability (by having a factory function return	1
// FIXME Instead of re-querying the device, have the caller provide it.	1
// TODO Implement setIdImpl to propagate the ID to the connections	1
# TODO Add CUDA to this list, in order to fix the TODO below	1
// FIXME Should we use a proper hash combiner? We can copy Boost's one.	1
// TODO Split this into a armConnectionCallback, a listenStart and a	1
// FIXME Can we support a "range" API (i.e., a begin() and end() method) so	1
// FIXME We're copying the args here...	1
// TODO: Store the PID upon channel/context instantiation.	1
// FIXME: Make this configurable similarly to the port number.	1
// FIXME It's not really great to change a global property of the	1