search

schoolpost / PiDNG

Create Adobe DNG RAW files using Python. Works with any Bayer RAW Data including native support for Raspberry Pi cameras.
MIT License
199 stars 37 forks source link

Illegal Instruction #52

Closed eat-sleep-code closed 2 years ago

eat-sleep-code commented 2 years ago

@schoolpost, well this is a weird one.

I am running into an issue with PiDNG, but only on one Rasperry Pi.

This is a Raspberry Pi Zero W with an Arducam IMX477 HQ Camera module.

I ran pip3 install PiDNG and it says it successfully installed:

Successfully installed PiDNG-3.4.7 exifread-2.3.2 numpy-1.21.5

But then my Python script fails on this line with an error of Illegal instruction:

from pidng.core import RPICAM2DNG

This worked fine on another Raspberry Pi 4 with a Raspberry Pi-branded HQ camera. So, not sure if this is related to the Pi Zero W or the Arducam camera. There is nothing dumped to the screen other than Illegal instruction so not much helpful debugging info.

schoolpost commented 2 years ago

Can you post the image you are trying to convert?

eat-sleep-code commented 2 years ago

@schoolpost, I have tracked the root cause of this issue...

NumPy 1.21.5 was released on December 20, 2021. It appears to have introduced breaking changes on 32-bit ARMV6 processors.

As your install script has the dependency restriction to be any version under 2.0 (Collecting numpy<2.0,>=1.14 (from PiDNG)) it grabs 1.21.5 during the install. This works fine on the 64-bit ARMV8 processors, but does not work on the "older" Raspberry Pi Zero W.

It appears NumPy 1.21.5 was a maintenance release, so not sure why it ended up causing this breaking change. It also appears 1.22.0 dropped support for 32-bit processors so all future versions will likely not work on the original Pi Zero W.

To "resolve" this, I ran the following command after installing PiDNG. This reverts NumPy to use the older -- working -- 1.21.4 version.

sudo pip3 uninstall -y numpy && sudo pip3 install numpy==1.21.4

eat-sleep-code commented 2 years ago

I have opened an issue with NumPy on this to see if they will resolve, or if support for the original Pi Zero W is considered "dead" by them.

eat-sleep-code commented 2 years ago

Just as an update...

I worked quite a bit with the developers of NumPy to find what the root cause of it was.

After much digging, this was traced to an issue with PiWheels affecting all Python 3.9 wheels that were touched in the last several weeks.

The creator of PiWheels posted that this is currently being researched: https://github.com/piwheels/packages/issues/269#issuecomment-1007273390

As of the morning of January 13, 2022 there is no resolution to this.