Test failures on big-endian architecture (s390x)

[musicinmybrain]

While trying to update the python-glymur package in Fedora Linux from 0.9.6, I found there are some test failures related to the new tiff2jp2k tool on s390x—Fedora Linux’s sole big-endian architecture.

=========================== short test summary info ============================
FAILED tests/test_tiff2jp2.py::TestSuite::test_minisblack_3strip_to_2x2 - Zer...
FAILED tests/test_tiff2jp2.py::TestSuite::test_partial_last_strip - ZeroDivis...
FAILED tests/test_tiff2jp2.py::TestSuite::test_partial_strip_and_partial_tiles
FAILED tests/test_tiff2jp2.py::TestSuite::test_psnr - AssertionError: False i...
FAILED tests/test_tiff2jp2.py::TestSuiteNoScikitImage::test_rgb_stripped - Ze...
FAILED tests/test_tiff2jp2.py::TestSuiteNoScikitImage::test_rgb_stripped_bottom_of_tile_coincides_with_bottom_of_strip
FAILED tests/test_tiff2jp2.py::TestSuiteNoScikitImage::test_stripped_logging
============= 7 failed, 525 passed, 1 skipped in 168.62s (0:02:48) =============

It looks like, in each of these, rps (read from the RowsPerStrip field), is zero. I think this is because the field is actually 32-bit instead of 16-bit in this test (“LONG”) instead of (“SHORT”), and reading a 16-bit value is getting only the top half of the word on big-endian platforms.

It seems like either type is allowed for this field, and a library should use libtiff’s TIFFFieldDataType to see which is in use.

This should probably be handled automatically and generally in glymur.lib.tiff.getFieldDefaulted(). I’ll probably offer a PR if I have some time to work on it.

[musicinmybrain]

As a quick hack, I confirmed that hard-coding

TAGS["RowsPerStrip"]["type"] = ctypes.c_uint32

at the end of glymur/lib/tiff.py makes all but one of these tests pass on all platforms. It’s not a correct fix, because RowsPerStrip could be 16-bit, but it helps confirm the diagnosis.

The exception is tests/test_tiff2jp2.py::TestSuite::test_psnr, which appears to be failing due to a separate issue that I haven’t looked into at all.

[quintusdias]

Thanks for the report. I'll get to work on it.

[musicinmybrain]

Thanks! Please let me know if I can do anything to help.

[quintusdias]

Using TIFFFieldDatatype is tougher than it looks. I elected on doing this another way, and just bypass LibTIFF when retrieving the tag values. I already had a platform-independent way of getting all the tags for the purposes of creating a UUID box to preserve the original value of all the tags, and that should work here as well.

[musicinmybrain]

Thanks! This seems logical, and I can confirm that it fixes this bug. I still have one test failure on s390x, which I’ll report separately since it seems to be unrelated.