drjit error message

[owaranainatsu]

The codes are mainly as follows:

@dr.wrap_ad(source='torch', target='drjit')
def render_torch(specular, spp=256, seed=1):
    print(type(specular))
    params[key] = dr.unravel(mi.Color3f, specular.array)
    params.update()
    return mi.render(scene, params, spp=spp, seed=seed, seed_grad=seed+1)

inital_specular = torch.tensor([[0.1, 0.1, 0.1]], dtype=torch.float32).to(device)
image = render_torch(inital_specular, spp=50)

and I got the following errors:

Cell In [18], line 10
      6     return mi.render(scene, params, spp=spp, seed=seed, seed_grad=seed+1)
      8 inital_specular = torch.tensor([[0.1, 0.1, 0.1]], dtype=torch.float32).to(device)
---> 10 image = render_torch(inital_specular, spp=50)

File ~/anaconda3/envs/mitsuba3/lib/python3.9/site-packages/drjit/router.py:5852, in wrap_ad.<locals>.wrapper.<locals>.f(*args, **kwargs)
   5849             del ctx.res_drjit, ctx.args_drjit
   5850             return args_grad
-> 5852     return ToDrJit.apply(*args)
   5854 if target == 'torch':
   5855     class ToTorch(_dr.CustomOp):

File ~/anaconda3/envs/mitsuba3/lib/python3.9/site-packages/drjit/router.py:5837, in wrap_ad.<locals>.wrapper.<locals>.f.<locals>.ToDrJit.forward(ctx, *args)
   5835 ctx.args_drjit = torch_to_drjit(args)
   5836 _dr.enable_grad(ctx.args_drjit)
-> 5837 res = func(*ctx.args_drjit)
   5838 ctx.res_drjit = (res,) if not isinstance(res, tuple) else res
   5839 return drjit_to_torch(res)

Cell In [18], line 4, in render_torch(specular, spp, seed)
      1 @dr.wrap_ad(source='torch', target='drjit')
      2 def render_torch(specular, spp=256, seed=1):
      3     print(type(specular))
...

TypeError: gather_(): incompatible function arguments. The following argument types are supported:
    1. (source: drjit.llvm.ad.Float, index: drjit.llvm.ad.UInt, mask: drjit.llvm.ad.Bool, permute: bool = False) -> drjit.llvm.ad.Float

Invoked with: [0.10000000149011612, 0.10000000149011612, 0.10000000149011612], [0], [True]

The weird thing is this kind of code works in another scene, but I don't know why.

[Speierers]

What is the type printed in render_torch? Could it be that your PyTorch tensor is stored on the GPU while you set the llvm_ad_rgb mode in Mitsuba?

[owaranainatsu]

Exactly! Sorry I forgot to set the 'cuda_ad_rgb' mode.