`bad_alloc` error when loading NMX nexus file

[jokasimr]

See https://git.esss.dk/dmsc-nightly/dmsc-nightly/-/jobs/53623

On scippnexus 24.11.1:

In [1]: import scippnexus as snx

In [2]: with snx.File('994160_00043336.hdf', "r", locking=False) as f:   dg = f[()]

MemoryError: std::bad_alloc
...

on scippnexus 24.3.1 this does not happen.

[jl-wynen]

This is an NMX file, not DREAM.

The offending commit is 3a97343
The problem seems to be that in the new code in the commit, we have

group_dims = ('x_pixel_offset', 'y_pixel_offset', 'z_pixel_offset')
self._signal.shape = (1280, 1280, 8097)

which gets propagated to Field.__getitem__ for detector_number which then raises bad_alloc because the data volume is too large.

The detector_number in the file has shape (1280, 1280). And 8097 is the number of events, not a z-offset. Even though the shape matches the comment in the linked commit, this does not seem right to me. Why would we want to allocate such an array in the first place?

[SimonHeybrock]

I presume this is related to the NMX now having (or still not having?) an axis attribute for the pixel offset fields?

[jokasimr]

I presume this is related to the NMX now having (or still not having?) an axis attribute for the pixel offset fields?

NMX files now have an axis attribute on the pixel offset fields

[jokasimr]

And 8097 is the number of events, not a z-offset.

Looks like 8097 is the number of pulses in detector panel 0, not the number of events.
The number of events is considerably larger, the file size is 740MB.

The other panels have no pulses or events.

[SimonHeybrock]

Maybe z_pixel_offset has a bad value in it's axis? If detector_number is 2-D but we have axis=2, ScippNexus might think it refers to some other data dimension?

[jokasimr]

Maybe z_pixel_offset has a bad value in it's axis? If detector_number is 2-D but we have axis=2, ScippNexus might think it refers to some other data dimension?

Yes this is definitely part of the issue.
z_pixel_offset has axis=3 (x=1, y=2).

But the length of z_pixel_offset is 1 so I don't know why it picks up 8097 as the size of that dimension.

[SimonHeybrock]

I think it happens the other way round: This claims to be the third axis of the data/detector_number, therefore ScippNexus assigns it the dim label of the third axis.

[jokasimr]

Is it just a typo in the commit @jl-wynen referenced?

Should

                self._signal._grouping.sizes = dict(
                    zip(group_dims, self._signal.shape, strict=False)

be

                self._signal._grouping.sizes = dict(
                    zip(group_dims, self._signal._grouping.shape, strict=False)

?