Unable to directly compute texture gradient on SceneParameters object

[dvicini]

It used to be possible to call dr.enable_grad on any differentiable scene parameter, call params.update(), render and compute gradients of the respective parameters.

This is currently no longer possible, the offending commit seems to be af5aaa6. I haven't looked into the details yet, but it might be related to the enabling of packet instructions for textures and the associated padding. Something in that logic prevents gradients from being accumulated at the original value in the scene parameters object.

The following reproducer now prints False, but prior to af5aaa6 prints True:

import drjit as dr
import mitsuba as mi
import numpy as np

mi.set_variant('llvm_ad_rgb')

scene = mi.load_dict({
    'type': 'scene',
    'integrator': {'type': 'direct'},
    'sensor': {
        'type': 'perspective',
        'to_world': mi.ScalarTransform4f().look_at(
            [0, 0, 5], [0, 0, 0], [0, 1, 0]
        ),
    },
    'emitter': {'type': 'constant'},
    'mesh': {
        'type': 'cube',
        'bsdf': {
            'type': 'diffuse',
            'reflectance': {
                'type': 'bitmap',
                'bitmap': mi.Bitmap(np.random.rand(16, 16, 3).astype(np.float32)),
                "format": "variant"
            },
        },
    },
})

scene_parameters = mi.traverse(scene)
key = 'mesh.bsdf.reflectance.data'
scene_parameters.keep([key])
dr.enable_grad(scene_parameters[key])
scene_parameters.update()

dr.backward(dr.mean(mi.render(scene, params=scene_parameters, spp=4)))
grad = dr.grad(scene_parameters[key])
print("Non-zero gradient? ", dr.any(grad != 0))

It's possible to work around this problem by creating a new tensor and re-assigning it:

texture = mi.TensorXf(scene_parameters[key])
dr.enable_grad(texture)
scene_parameters[key] = texture
...
grad = dr.grad(texture)

However, it's still confusing that directly enabling gradients on the SceneParameters object does not work for textures, but keeps working for all other parameters.