Code style: capitalize sentences and end with dot.

I think this function needs a mutex lock? In case multiple Cycles sessions execute it at the same time.

Would slightly prefer make_unique and then std::move.

Also, add #include "util/unique_ptr.h" and remove the std:: namespace prefix.

Is there a reason this one is using std::lock_guard and other places use std::unique_lock.

In most Cycles code we use the types defined in util/thread.h. Unless there is a specific to reason something else, it seems best to use those.

This name is missing the GPU architecture, which matters when there are multiple GPUs with different architectures on the device (like an external GPU).

At least I imagine the cache can't be shared between those.

Code style:

/* Inline everything for Apple GPUs.
 * This gives ~1.1x speedup and 10% spill reduction for integator_shade_surface
 * at the cost of longer compile times (~4.5 minutes on M1 Max). */

I think separate caches per devices indeed required. The number of cache entries for different versions should not be reduced when there are more devices.

For HIP and CUDA we simply add the architecture in the name (for example cycles_kernel_sm_86_1FB56B1136836BDB5920C12047C066D4.cubin). Could we do the same here? The device name may need to be hashed or normalized in some way if it contains characters that don't work in a filename.

We could cache the kernels in separate folders per device, but I don't think it makes a practical difference right now.