Getting nan's from straightforward math operations.

[mark-at-hypothetic]

Hi! I'm trying to simply add to a tensor, but I'm getting nan's as a result.

Here's a simplified piece of code that gets me the nan's. Also, the number of nan's is variable each time I run it, but it's always divisible by 256.

import drjit as dr
import mitsuba as mi
mi.set_variant('llvm_ad_rgb')


base_color_val = mi.TensorXf(np.ones((1,1,3)))   # TensorXf(shape=(1, 1, 3))
tex = mi.TensorXf(np.zeros((256,256,3)))            # TensorXf(shape=(256, 256, 3))
cc = tex + base_color_val                                      # TensorXf(shape=(256, 256, 3))

print(dr.sum(dr.isnan(cc))) --> out:  1792
print(dr.sum(dr.isfinite(cc))) --> out: 194816
print(256*256*3) --> out: 196608

but when I make base_color_val of shape (1,1,1) I get no nan's

base_color_val = mi.TensorXf(np.ones((1,1,3)))   # TensorXf(shape=(1, 1, 3))
tex = mi.TensorXf(np.zeros((256,256,3)))            # TensorXf(shape=(256, 256, 3))
cc = tex + base_color_val[:,:,0]                             # TensorXf(shape=(256, 256, 3))

print(dr.sum(dr.isnan(cc))) --> out: 0
print(dr.sum(dr.isfinite(cc))) --> out: 196608
print(256*256*3) --> out: 196608

obviously something is going wrong with slicing, but I guess it should've given an error instead of returning nan's?

Running
mitsuba 3.3.0
drjit 0.4.2
ubuntu, python 3.10

[njroussel]

Hi @mark-at-hypothetic

I'm unable to reproduce this.

You can "inspect" cc by converting it back to numpy for example. At least we might get some understanding of what is happening with its values.

As a side note, drjit has a dr.ones operation and a dr.zeros operations too.

[njroussel]

I'm unable to reproduce this behavior at all. The fact that is only seems to happen with scalar and LLVM types leaves me to believe that there is some weird host system-level operation going on. The construction of the TensorXf object, is only a single memcpy normally.

Feel free to open an issue.

[mark-at-hypothetic]

Strange. Thanks for looking into it. I'll open an issue

[tomas16]

I wonder if this is related to my issue #890. I'm using llvm_ad_rgb variant on macOS.

[njroussel]

Did either of you make any progress in identifying this issue?

[tomas16]

No, it'll be at least another month before I can continue working on this.