How to replicate dr.cuda.ArrayXf([1,2,3]) using Mitsuba?

[maxfrei750]

Hi! I'd like to create the same output as the abovementioned command, but using Mitsuba, so that the data type depends on the user-selected backend. How can I do that?

Also: Is there documentation for this really basic vector/array/tensor stuff? I tried to look in the mitsuba and drjit docs, but didn't find a basic explanation on how to use the different data structures.

Thanks!

[njroussel]

Hi @maxfrei750 !!

It looks like ArrayXf was never bound/aliased in Mitsuba. I wonder if this was intentional or if we just never needed it (you can usually use TensorXf to accomplish the same goal). Until I figure that out, you should be able to at least monkey patch it with Python.

We definitely still need to and want to improve the documentation: Tensors and textures aren't even documented in Dr.Jit. Is there something specific that you feel like is lacking? "basic vector/array/tensor stuff" what were you expecting?

[maxfrei750]

@njroussel For the time being, do you have any advice on how to realize an equivalent output using mi.TensorXf? Thanks!

[njroussel]

Thanks for the feedback!

I'm a bit surprised that you're using .Shape 🤯. This is mostly meant for internal uses, whereas dr.shape(var) or var.shape is what you'd usually want. Can I ask what you're trying to use it for? The Shape of a TensorXf will indeed always be uni-dimensional.

[maxfrei750]

Thanks for the hint! That's so helpful. The reason, I went for .Shape is rather dumb: it's what shows up in VSCode's debugger, and I wasn't aware of dr.shape(). 🤦

var.shape doesn't work for me:

>> dr.cuda.ArrayXf(1,2,3).shape
AttributeError: 'drjit.cuda.ArrayXf' object has no attribute 'shape'

Was that property added recently? Then maybe I need to update... It would be a nice feature, because it is more convenient and would also show up in the debugger.

Using dr.shape, I was able to get a much better understanding of my problem, which might be a little much to unpack here (or might even be a bug), so I'll open another discussion, to keep things tidy. Thanks a ton and see you in the next discussion. 😅

[njroussel]

No, it's always been there I think 😅

We use a script to automatically generate files that are use by your IDE's auto-complete system. It's very likely this doesn't wrotk that well.

[maxfrei750]

No, it's always been there I think sweat_smile

I think there was a misunderstanding on my behalf. When referring to var.shape, you were exclusively talking about mi.TensorXf (which indeed has the shape attribute), correct ? I thought that it was also supposed to work for e.g. Float and Array types (since dr.shape can be used on them).

That being said, IMHO, it would be convenient to have the shape attribute for the other types as well.