[llvm] [Offload] Implement the remaining initial Offload API (PR #122106)

[Joseph Huber via llvm-commits]

================
@@ -104,3 +104,15 @@ def : Function {
     Return<"OL_ERRC_INVALID_DEVICE">
   ];
 }
+
+def : Function {
+  let name = "olGetHostDevice";
----------------
jhuber6 wrote:

I prefer the second, since I think we should really think of this as abstracting over devices, not vendors. However it definitely makes some things more difficulit, because then we need to imagine how to handle things that are incompatible between devices. For example, you can't make an async object for an AMD kernel and then try to schedule an NVIDIA kernel on it.

A possible solution is to allow iterating devices with a filter on it, that's kind of what HSA does. You iterate all the devices and check if it's a GPU or a CPU or whatever.

https://github.com/llvm/llvm-project/pull/122106