Ben Ashbaugh

Merging as discussed in the August 6th teleconference.

See comments on the related issue: * https://github.com/KhronosGroup/OpenCL-Docs/issues/900#issuecomment-1588601203 * https://github.com/KhronosGroup/OpenCL-Docs/issues/900#issuecomment-1602054509 In short, even our pure fp16 `dot` implementation isn't failing the new test, but a pure fp16 `cross` implementation is....

Removing "focused review" - we can revisit when the spec issues are resolved.

I found an old version of the non-uniform subgroups spec where we recorded this issue: > In Vulkan, the ballot function returns a uint4, which effectively caps the largest subgroup...

> It's the _rest_ of the subgroup tests that use local work sizes of 200 that are liable to exceed the bounds of the `bitset` in this fashion. I think...

Can we make the change to limit the work-group size (and hence the maximum sub-group size) to be smaller than 128 work-items for all of non-uniform subgroup tests? I haven't...

I'm going to take the "focused review" off of this PR for now because next steps are clear. We can add it back when the changes described above are made....

I'm not comfortable saying it's the right thing to do in all cases, but there are some interesting things that fall out if we only free memory and do not...

> Thus, any command already in flight may fail. Ah, crud, good point. Yes, calls inside layers would have a problem if we're freeing dispatch tables out from under them....

I switched the shutdown order so the objects are freed in the reverse order they are initialized. Do we think the stub layer is valuable (item (2) in my original...