Reading back to Scalar types floods cache dir with kernels

[tszirr]

Depending on how I read back values from wide device arrays to scalar arrays, I find a weird issue with unexpected kernel regeneration on every read: Using .numpy() to access values generates new kernels every time, whereas dr.slice(...) does not result in additional kernels (only one kernel cached in ~/.drjit).

Interestingly, this behavior is only observed for more complex types such as Array3f, Matrix4f, Transform4f, but not for Float. Are there some kind of uncached reshuffling kernels involved in the multi-dimensional .numpy() accessors?

Repro

Tested on Ubuntu 20.04 LTS with Python 3.8:

import drjit as dr
from drjit.scalar import Array3f as ScalarArray3f
from drjit.cuda.ad import Array3f

device = Array3f(3.0)

for i in range(100):
    host = ScalarArray3f(device.numpy()) # generates many kernels in ~/.drjit
    #host = ScalarArray3f(dr.slice(device)) # generates one kernel in ~/.drjit
    host[0] += i
    device.assign(host)

print(device)