Resources can not be compared or used in containers

[MrBurmark]

Currently resources are not comparable of hashable so they can't be used as keys in unordered containers. It would also generally be useful if they were comparable so it would be possible to know if two represented the same resource, for example the same cuda stream.

[trws]

I’ll be happy to make this work. The two wrinkles to figure out: 1. Hashing of differently typed resources, presumably we want resources with the same initial number but different types to hash to different values. 2. Comparing resources: Do we want the “logical” result or the actual result. Meaning, if you put 10 in for one, 1000 in for another, but they happen to map to the same stream on the current scheme, should they compare the same? I’m not sure how I feel about it, in cuda I think they kinda should because it’s more representative of the cuda behavior. On OpenMP though the reverse is true. Thoughts?

…

Sent from my iPhone On Apr 27, 2021, at 9:45 PM, Jason Burmark ***@***.***> wrote:  Currently resources are not comparable of hashable so they can't be used as keys in unordered containers. It would also generally be useful if they were comparable so it would be possible to know if two represented the same resource, for example the same cuda stream. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<https://urldefense.us/v3/__https://github.com/LLNL/camp/issues/55__;!!G2kpM7uM-TzIFchu!k3w0eg3i2Auzjhn8BjsDlfoElHECbTW3AO8-dTDAK20Ng0nrtASdHeiG34pVRAe8cw$>, or unsubscribe<https://urldefense.us/v3/__https://github.com/notifications/unsubscribe-auth/AAFBFNPNBYUV7Z4V5UM7FB3TK6HETANCNFSM43WI7DCQ__;!!G2kpM7uM-TzIFchu!k3w0eg3i2Auzjhn8BjsDlfoElHECbTW3AO8-dTDAK20Ng0nrtASdHeiG34roup2MHw$>.

[MrBurmark]

1. That makes sense to me, and presumably this only really matters for the type erased resource. Would combining the number and the platform in some way provide a reasonable hash?
2. That's a good point. My first instinct is to honor the actual behavior/result, but I can see why it would also make sense to depend on the logical behavior. I can imagine wanting to put thing in an associative container by the behavior of the resource and by the number that was used. Perhaps comparison should use the number and a separate function can distinguish between resources with equivalent behavior? Then we could at least make a comparator that uses equivalence.

[trws]

That seems like a good compromise to me.

The hash we can probably hack together by just taking the number that goes in and ORing it with the shifted platform ID or something, since they have essentially unique numbers to begin with that would be a nearly-perfect hash essentially for free.