Robustness and GPU hangs
If we hit a GPU hang, the kernel will reset our context to a known-good state - the default state. Iris emits very minimal batches (even smaller than i965), which inherit a lot of state from one to the next. So, getting reset to the default state means that our expectations for subsequent batches will be completely, and utterly wrong.
In practical terms, this means that if you hit a GPU hang, the following batches will all GPU hang as well. We need to land Chris Wilson's "ban my context after one hang" kernel patches, detect that we've been banned, create a new (non-banned) GEM context, re-flag all state as dirty, and possibly re-try the draw. (Or for robust contexts, just report device lost.)